asterisk/res/res_rtp_asterisk.c

10347 lines
342 KiB
C
Raw Normal View History

/*
* Asterisk -- An open source telephony toolkit.
*
* Copyright (C) 1999 - 2008, Digium, Inc.
*
* Mark Spencer <markster@digium.com>
*
* See http://www.asterisk.org for more information about
* the Asterisk project. Please do not directly contact
* any of the maintainers of this project for assistance;
* the project provides a web site, mailing lists and IRC
* channels for your use.
*
* This program is free software, distributed under the terms of
* the GNU General Public License Version 2. See the LICENSE file
* at the top of the source tree.
*/
/*!
* \file
*
* \brief Supports RTP and RTCP with Symmetric RTP support for NAT traversal.
*
* \author Mark Spencer <markster@digium.com>
*
* \note RTP is defined in RFC 3550.
*
* \ingroup rtp_engines
*/
/*** MODULEINFO
<use type="external">openssl</use>
<use type="external">pjproject</use>
<support_level>core</support_level>
***/
#include "asterisk.h"
#include <arpa/nameser.h>
#include "asterisk/dns_core.h"
#include "asterisk/dns_internal.h"
#include "asterisk/dns_recurring.h"
#include <sys/time.h>
#include <signal.h>
#include <fcntl.h>
#include <math.h>
#ifdef HAVE_OPENSSL
#include <openssl/opensslconf.h>
#include <openssl/opensslv.h>
#if !defined(OPENSSL_NO_SRTP) && (OPENSSL_VERSION_NUMBER >= 0x10001000L)
#include <openssl/ssl.h>
#include <openssl/err.h>
#include <openssl/bio.h>
#if !defined(OPENSSL_NO_ECDH) && (OPENSSL_VERSION_NUMBER >= 0x10000000L)
#include <openssl/bn.h>
#endif
#ifndef OPENSSL_NO_DH
#include <openssl/dh.h>
#endif
#endif
#endif
#ifdef HAVE_PJPROJECT
#include <pjlib.h>
#include <pjlib-util.h>
#include <pjnath.h>
#include <ifaddrs.h>
#endif
#include "asterisk/conversions.h"
PJPROJECT logging: Made easier to get available logging levels. Use of the new logging is as simple as issuing the new CLI command or setting the new pjproject.conf option. Other options that can affect the logging are how you have the pjproject log levels mapped to Asterisk log types in pjproject.conf and if you have configured Asterisk to log the DEBUG type messages. Altering the pjproject.conf level mapping shouldn't be necessary for most installations as the default mapping is sensible. Configuring Asterisk to log the DEBUG message type is standard practice for collecting debug information. * Added CLI "pjproject set log level" command to dynamically adjust the maximum pjproject log message level. * Added CLI "pjproject show log level" command to see the currently set maximum pjproject log message level. * Added pjproject.conf startup section "log_level" option to set the initial maximum pjproject log message level so all messages could be captured from initialization. * Set PJ_LOG_MAX_LEVEL to 6 to compile in all defined logging levels into bundled pjproject. Pjproject will use the currently set run time log level to determine if a log message is generated just like Asterisk verbose and debug logging levels. * In log_forwarder(), made always log enabled and mapped pjproject log messages. DEBUG mapped log messages are no longer gated by the current Asterisk debug logging level. * Removed RAII_VAR() from res_pjproject.c:get_log_level(). ASTERISK-26630 #close Change-Id: I6dca12979f482ffb0450aaf58db0fe0f6d2e5389
2016-11-24 00:27:54 +00:00
#include "asterisk/options.h"
#include "asterisk/logger_category.h"
#include "asterisk/stun.h"
#include "asterisk/pbx.h"
#include "asterisk/frame.h"
media formats: re-architect handling of media for performance improvements In the old times media formats were represented using a bit field. This was fast but had a few limitations. 1. Asterisk was limited in how many formats it could handle. 2. Formats, being a bit field, could not include any attribute information. A format was strictly its type, e.g., "this is ulaw". This was changed in Asterisk 10 (see https://wiki.asterisk.org/wiki/display/AST/Media+Architecture+Proposal for notes on that work) which led to the creation of the ast_format structure. This structure allowed Asterisk to handle attributes and bundle information with a format. Additionally, ast_format_cap was created to act as a container for multiple formats that, together, formed the capability of some entity. Another mechanism was added to allow logic to be registered which performed format attribute negotiation. Everywhere throughout the codebase Asterisk was changed to use this strategy. Unfortunately, in software, there is no free lunch. These new capabilities came at a cost. Performance analysis and profiling showed that we spend an inordinate amount of time comparing, copying, and generally manipulating formats and their related structures. Basic prototyping has shown that a reasonably large performance improvement could be made in this area. This patch is the result of that project, which overhauled the media format architecture and its usage in Asterisk to improve performance. Generally, the new philosophy for handling formats is as follows: * The ast_format structure is reference counted. This removed a large amount of the memory allocations and copying that was done in prior versions. * In order to prevent race conditions while keeping things performant, the ast_format structure is immutable by convention and lock-free. Violate this tenet at your peril! * Because formats are reference counted, codecs are also reference counted. The Asterisk core generally provides built-in codecs and caches the ast_format structures created to represent them. Generally, to prevent inordinate amounts of module reference bumping, codecs and formats can be added at run-time but cannot be removed. * All compatibility with the bit field representation of codecs/formats has been moved to a compatibility API. The primary user of this representation is chan_iax2, which must continue to maintain its bit-field usage of formats for interoperability concerns. * When a format is negotiated with attributes, or when a format cannot be represented by one of the cached formats, a new format object is created or cloned from an existing format. That format may have the same codec underlying it, but is a different format than a version of the format with different attributes or without attributes. * While formats are reference counted objects, the reference count maintained on the format should be manipulated with care. Formats are generally cached and will persist for the lifetime of Asterisk and do not explicitly need to have their lifetime modified. An exception to this is when the user of a format does not know where the format came from *and* the user may outlive the provider of the format. This occurs, for example, when a format is read from a channel: the channel may have a format with attributes (hence, non-cached) and the user of the format may last longer than the channel (if the reference to the channel is released prior to the format's reference). For more information on this work, see the API design notes: https://wiki.asterisk.org/wiki/display/AST/Media+Format+Rewrite Finally, this work was the culmination of a large number of developer's efforts. Extra thanks goes to Corey Farrell, who took on a large amount of the work in the Asterisk core, chan_sip, and was an invaluable resource in peer reviews throughout this project. There were a substantial number of patches contributed during this work; the following issues/patch names simply reflect some of the work (and will cause the release scripts to give attribution to the individuals who work on them). Reviews: https://reviewboard.asterisk.org/r/3814 https://reviewboard.asterisk.org/r/3808 https://reviewboard.asterisk.org/r/3805 https://reviewboard.asterisk.org/r/3803 https://reviewboard.asterisk.org/r/3801 https://reviewboard.asterisk.org/r/3798 https://reviewboard.asterisk.org/r/3800 https://reviewboard.asterisk.org/r/3794 https://reviewboard.asterisk.org/r/3793 https://reviewboard.asterisk.org/r/3792 https://reviewboard.asterisk.org/r/3791 https://reviewboard.asterisk.org/r/3790 https://reviewboard.asterisk.org/r/3789 https://reviewboard.asterisk.org/r/3788 https://reviewboard.asterisk.org/r/3787 https://reviewboard.asterisk.org/r/3786 https://reviewboard.asterisk.org/r/3784 https://reviewboard.asterisk.org/r/3783 https://reviewboard.asterisk.org/r/3778 https://reviewboard.asterisk.org/r/3774 https://reviewboard.asterisk.org/r/3775 https://reviewboard.asterisk.org/r/3772 https://reviewboard.asterisk.org/r/3761 https://reviewboard.asterisk.org/r/3754 https://reviewboard.asterisk.org/r/3753 https://reviewboard.asterisk.org/r/3751 https://reviewboard.asterisk.org/r/3750 https://reviewboard.asterisk.org/r/3748 https://reviewboard.asterisk.org/r/3747 https://reviewboard.asterisk.org/r/3746 https://reviewboard.asterisk.org/r/3742 https://reviewboard.asterisk.org/r/3740 https://reviewboard.asterisk.org/r/3739 https://reviewboard.asterisk.org/r/3738 https://reviewboard.asterisk.org/r/3737 https://reviewboard.asterisk.org/r/3736 https://reviewboard.asterisk.org/r/3734 https://reviewboard.asterisk.org/r/3722 https://reviewboard.asterisk.org/r/3713 https://reviewboard.asterisk.org/r/3703 https://reviewboard.asterisk.org/r/3689 https://reviewboard.asterisk.org/r/3687 https://reviewboard.asterisk.org/r/3674 https://reviewboard.asterisk.org/r/3671 https://reviewboard.asterisk.org/r/3667 https://reviewboard.asterisk.org/r/3665 https://reviewboard.asterisk.org/r/3625 https://reviewboard.asterisk.org/r/3602 https://reviewboard.asterisk.org/r/3519 https://reviewboard.asterisk.org/r/3518 https://reviewboard.asterisk.org/r/3516 https://reviewboard.asterisk.org/r/3515 https://reviewboard.asterisk.org/r/3512 https://reviewboard.asterisk.org/r/3506 https://reviewboard.asterisk.org/r/3413 https://reviewboard.asterisk.org/r/3410 https://reviewboard.asterisk.org/r/3387 https://reviewboard.asterisk.org/r/3388 https://reviewboard.asterisk.org/r/3389 https://reviewboard.asterisk.org/r/3390 https://reviewboard.asterisk.org/r/3321 https://reviewboard.asterisk.org/r/3320 https://reviewboard.asterisk.org/r/3319 https://reviewboard.asterisk.org/r/3318 https://reviewboard.asterisk.org/r/3266 https://reviewboard.asterisk.org/r/3265 https://reviewboard.asterisk.org/r/3234 https://reviewboard.asterisk.org/r/3178 ASTERISK-23114 #close Reported by: mjordan media_formats_translation_core.diff uploaded by kharwell (License 6464) rb3506.diff uploaded by mjordan (License 6283) media_format_app_file.diff uploaded by kharwell (License 6464) misc-2.diff uploaded by file (License 5000) chan_mild-3.diff uploaded by file (License 5000) chan_obscure.diff uploaded by file (License 5000) jingle.diff uploaded by file (License 5000) funcs.diff uploaded by file (License 5000) formats.diff uploaded by file (License 5000) core.diff uploaded by file (License 5000) bridges.diff uploaded by file (License 5000) mf-codecs-2.diff uploaded by file (License 5000) mf-app_fax.diff uploaded by file (License 5000) mf-apps-3.diff uploaded by file (License 5000) media-formats-3.diff uploaded by file (License 5000) ASTERISK-23715 rb3713.patch uploaded by coreyfarrell (License 5909) rb3689.patch uploaded by mjordan (License 6283) ASTERISK-23957 rb3722.patch uploaded by mjordan (License 6283) mf-attributes-3.diff uploaded by file (License 5000) ASTERISK-23958 Tested by: jrose rb3822.patch uploaded by coreyfarrell (License 5909) rb3800.patch uploaded by jrose (License 6182) chan_sip.diff uploaded by mjordan (License 6283) rb3747.patch uploaded by jrose (License 6182) ASTERISK-23959 #close Tested by: sgriepentrog, mjordan, coreyfarrell sip_cleanup.diff uploaded by opticron (License 6273) chan_sip_caps.diff uploaded by mjordan (License 6283) rb3751.patch uploaded by coreyfarrell (License 5909) chan_sip-3.diff uploaded by file (License 5000) ASTERISK-23960 #close Tested by: opticron direct_media.diff uploaded by opticron (License 6273) pjsip-direct-media.diff uploaded by file (License 5000) format_cap_remove.diff uploaded by opticron (License 6273) media_format_fixes.diff uploaded by opticron (License 6273) chan_pjsip-2.diff uploaded by file (License 5000) ASTERISK-23966 #close Tested by: rmudgett rb3803.patch uploaded by rmudgetti (License 5621) chan_dahdi.diff uploaded by file (License 5000) ASTERISK-24064 #close Tested by: coreyfarrell, mjordan, opticron, file, rmudgett, sgriepentrog, jrose rb3814.patch uploaded by rmudgett (License 5621) moh_cleanup.diff uploaded by opticron (License 6273) bridge_leak.diff uploaded by opticron (License 6273) translate.diff uploaded by file (License 5000) rb3795.patch uploaded by rmudgett (License 5621) tls_fix.diff uploaded by mjordan (License 6283) fax-mf-fix-2.diff uploaded by file (License 5000) rtp_transfer_stuff uploaded by mjordan (License 6283) rb3787.patch uploaded by rmudgett (License 5621) media-formats-explicit-translate-format-3.diff uploaded by file (License 5000) format_cache_case_fix.diff uploaded by opticron (License 6273) rb3774.patch uploaded by rmudgett (License 5621) rb3775.patch uploaded by rmudgett (License 5621) rtp_engine_fix.diff uploaded by opticron (License 6273) rtp_crash_fix.diff uploaded by opticron (License 6273) rb3753.patch uploaded by mjordan (License 6283) rb3750.patch uploaded by mjordan (License 6283) rb3748.patch uploaded by rmudgett (License 5621) media_format_fixes.diff uploaded by opticron (License 6273) rb3740.patch uploaded by mjordan (License 6283) rb3739.patch uploaded by mjordan (License 6283) rb3734.patch uploaded by mjordan (License 6283) rb3689.patch uploaded by mjordan (License 6283) rb3674.patch uploaded by coreyfarrell (License 5909) rb3671.patch uploaded by coreyfarrell (License 5909) rb3667.patch uploaded by coreyfarrell (License 5909) rb3665.patch uploaded by mjordan (License 6283) rb3625.patch uploaded by coreyfarrell (License 5909) rb3602.patch uploaded by coreyfarrell (License 5909) format_compatibility-2.diff uploaded by file (License 5000) core.diff uploaded by file (License 5000) git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@419044 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-07-20 22:06:33 +00:00
#include "asterisk/format_cache.h"
#include "asterisk/channel.h"
#include "asterisk/acl.h"
#include "asterisk/config.h"
#include "asterisk/lock.h"
#include "asterisk/utils.h"
#include "asterisk/cli.h"
#include "asterisk/manager.h"
#include "asterisk/unaligned.h"
#include "asterisk/module.h"
#include "asterisk/rtp_engine.h"
media formats: re-architect handling of media for performance improvements In the old times media formats were represented using a bit field. This was fast but had a few limitations. 1. Asterisk was limited in how many formats it could handle. 2. Formats, being a bit field, could not include any attribute information. A format was strictly its type, e.g., "this is ulaw". This was changed in Asterisk 10 (see https://wiki.asterisk.org/wiki/display/AST/Media+Architecture+Proposal for notes on that work) which led to the creation of the ast_format structure. This structure allowed Asterisk to handle attributes and bundle information with a format. Additionally, ast_format_cap was created to act as a container for multiple formats that, together, formed the capability of some entity. Another mechanism was added to allow logic to be registered which performed format attribute negotiation. Everywhere throughout the codebase Asterisk was changed to use this strategy. Unfortunately, in software, there is no free lunch. These new capabilities came at a cost. Performance analysis and profiling showed that we spend an inordinate amount of time comparing, copying, and generally manipulating formats and their related structures. Basic prototyping has shown that a reasonably large performance improvement could be made in this area. This patch is the result of that project, which overhauled the media format architecture and its usage in Asterisk to improve performance. Generally, the new philosophy for handling formats is as follows: * The ast_format structure is reference counted. This removed a large amount of the memory allocations and copying that was done in prior versions. * In order to prevent race conditions while keeping things performant, the ast_format structure is immutable by convention and lock-free. Violate this tenet at your peril! * Because formats are reference counted, codecs are also reference counted. The Asterisk core generally provides built-in codecs and caches the ast_format structures created to represent them. Generally, to prevent inordinate amounts of module reference bumping, codecs and formats can be added at run-time but cannot be removed. * All compatibility with the bit field representation of codecs/formats has been moved to a compatibility API. The primary user of this representation is chan_iax2, which must continue to maintain its bit-field usage of formats for interoperability concerns. * When a format is negotiated with attributes, or when a format cannot be represented by one of the cached formats, a new format object is created or cloned from an existing format. That format may have the same codec underlying it, but is a different format than a version of the format with different attributes or without attributes. * While formats are reference counted objects, the reference count maintained on the format should be manipulated with care. Formats are generally cached and will persist for the lifetime of Asterisk and do not explicitly need to have their lifetime modified. An exception to this is when the user of a format does not know where the format came from *and* the user may outlive the provider of the format. This occurs, for example, when a format is read from a channel: the channel may have a format with attributes (hence, non-cached) and the user of the format may last longer than the channel (if the reference to the channel is released prior to the format's reference). For more information on this work, see the API design notes: https://wiki.asterisk.org/wiki/display/AST/Media+Format+Rewrite Finally, this work was the culmination of a large number of developer's efforts. Extra thanks goes to Corey Farrell, who took on a large amount of the work in the Asterisk core, chan_sip, and was an invaluable resource in peer reviews throughout this project. There were a substantial number of patches contributed during this work; the following issues/patch names simply reflect some of the work (and will cause the release scripts to give attribution to the individuals who work on them). Reviews: https://reviewboard.asterisk.org/r/3814 https://reviewboard.asterisk.org/r/3808 https://reviewboard.asterisk.org/r/3805 https://reviewboard.asterisk.org/r/3803 https://reviewboard.asterisk.org/r/3801 https://reviewboard.asterisk.org/r/3798 https://reviewboard.asterisk.org/r/3800 https://reviewboard.asterisk.org/r/3794 https://reviewboard.asterisk.org/r/3793 https://reviewboard.asterisk.org/r/3792 https://reviewboard.asterisk.org/r/3791 https://reviewboard.asterisk.org/r/3790 https://reviewboard.asterisk.org/r/3789 https://reviewboard.asterisk.org/r/3788 https://reviewboard.asterisk.org/r/3787 https://reviewboard.asterisk.org/r/3786 https://reviewboard.asterisk.org/r/3784 https://reviewboard.asterisk.org/r/3783 https://reviewboard.asterisk.org/r/3778 https://reviewboard.asterisk.org/r/3774 https://reviewboard.asterisk.org/r/3775 https://reviewboard.asterisk.org/r/3772 https://reviewboard.asterisk.org/r/3761 https://reviewboard.asterisk.org/r/3754 https://reviewboard.asterisk.org/r/3753 https://reviewboard.asterisk.org/r/3751 https://reviewboard.asterisk.org/r/3750 https://reviewboard.asterisk.org/r/3748 https://reviewboard.asterisk.org/r/3747 https://reviewboard.asterisk.org/r/3746 https://reviewboard.asterisk.org/r/3742 https://reviewboard.asterisk.org/r/3740 https://reviewboard.asterisk.org/r/3739 https://reviewboard.asterisk.org/r/3738 https://reviewboard.asterisk.org/r/3737 https://reviewboard.asterisk.org/r/3736 https://reviewboard.asterisk.org/r/3734 https://reviewboard.asterisk.org/r/3722 https://reviewboard.asterisk.org/r/3713 https://reviewboard.asterisk.org/r/3703 https://reviewboard.asterisk.org/r/3689 https://reviewboard.asterisk.org/r/3687 https://reviewboard.asterisk.org/r/3674 https://reviewboard.asterisk.org/r/3671 https://reviewboard.asterisk.org/r/3667 https://reviewboard.asterisk.org/r/3665 https://reviewboard.asterisk.org/r/3625 https://reviewboard.asterisk.org/r/3602 https://reviewboard.asterisk.org/r/3519 https://reviewboard.asterisk.org/r/3518 https://reviewboard.asterisk.org/r/3516 https://reviewboard.asterisk.org/r/3515 https://reviewboard.asterisk.org/r/3512 https://reviewboard.asterisk.org/r/3506 https://reviewboard.asterisk.org/r/3413 https://reviewboard.asterisk.org/r/3410 https://reviewboard.asterisk.org/r/3387 https://reviewboard.asterisk.org/r/3388 https://reviewboard.asterisk.org/r/3389 https://reviewboard.asterisk.org/r/3390 https://reviewboard.asterisk.org/r/3321 https://reviewboard.asterisk.org/r/3320 https://reviewboard.asterisk.org/r/3319 https://reviewboard.asterisk.org/r/3318 https://reviewboard.asterisk.org/r/3266 https://reviewboard.asterisk.org/r/3265 https://reviewboard.asterisk.org/r/3234 https://reviewboard.asterisk.org/r/3178 ASTERISK-23114 #close Reported by: mjordan media_formats_translation_core.diff uploaded by kharwell (License 6464) rb3506.diff uploaded by mjordan (License 6283) media_format_app_file.diff uploaded by kharwell (License 6464) misc-2.diff uploaded by file (License 5000) chan_mild-3.diff uploaded by file (License 5000) chan_obscure.diff uploaded by file (License 5000) jingle.diff uploaded by file (License 5000) funcs.diff uploaded by file (License 5000) formats.diff uploaded by file (License 5000) core.diff uploaded by file (License 5000) bridges.diff uploaded by file (License 5000) mf-codecs-2.diff uploaded by file (License 5000) mf-app_fax.diff uploaded by file (License 5000) mf-apps-3.diff uploaded by file (License 5000) media-formats-3.diff uploaded by file (License 5000) ASTERISK-23715 rb3713.patch uploaded by coreyfarrell (License 5909) rb3689.patch uploaded by mjordan (License 6283) ASTERISK-23957 rb3722.patch uploaded by mjordan (License 6283) mf-attributes-3.diff uploaded by file (License 5000) ASTERISK-23958 Tested by: jrose rb3822.patch uploaded by coreyfarrell (License 5909) rb3800.patch uploaded by jrose (License 6182) chan_sip.diff uploaded by mjordan (License 6283) rb3747.patch uploaded by jrose (License 6182) ASTERISK-23959 #close Tested by: sgriepentrog, mjordan, coreyfarrell sip_cleanup.diff uploaded by opticron (License 6273) chan_sip_caps.diff uploaded by mjordan (License 6283) rb3751.patch uploaded by coreyfarrell (License 5909) chan_sip-3.diff uploaded by file (License 5000) ASTERISK-23960 #close Tested by: opticron direct_media.diff uploaded by opticron (License 6273) pjsip-direct-media.diff uploaded by file (License 5000) format_cap_remove.diff uploaded by opticron (License 6273) media_format_fixes.diff uploaded by opticron (License 6273) chan_pjsip-2.diff uploaded by file (License 5000) ASTERISK-23966 #close Tested by: rmudgett rb3803.patch uploaded by rmudgetti (License 5621) chan_dahdi.diff uploaded by file (License 5000) ASTERISK-24064 #close Tested by: coreyfarrell, mjordan, opticron, file, rmudgett, sgriepentrog, jrose rb3814.patch uploaded by rmudgett (License 5621) moh_cleanup.diff uploaded by opticron (License 6273) bridge_leak.diff uploaded by opticron (License 6273) translate.diff uploaded by file (License 5000) rb3795.patch uploaded by rmudgett (License 5621) tls_fix.diff uploaded by mjordan (License 6283) fax-mf-fix-2.diff uploaded by file (License 5000) rtp_transfer_stuff uploaded by mjordan (License 6283) rb3787.patch uploaded by rmudgett (License 5621) media-formats-explicit-translate-format-3.diff uploaded by file (License 5000) format_cache_case_fix.diff uploaded by opticron (License 6273) rb3774.patch uploaded by rmudgett (License 5621) rb3775.patch uploaded by rmudgett (License 5621) rtp_engine_fix.diff uploaded by opticron (License 6273) rtp_crash_fix.diff uploaded by opticron (License 6273) rb3753.patch uploaded by mjordan (License 6283) rb3750.patch uploaded by mjordan (License 6283) rb3748.patch uploaded by rmudgett (License 5621) media_format_fixes.diff uploaded by opticron (License 6273) rb3740.patch uploaded by mjordan (License 6283) rb3739.patch uploaded by mjordan (License 6283) rb3734.patch uploaded by mjordan (License 6283) rb3689.patch uploaded by mjordan (License 6283) rb3674.patch uploaded by coreyfarrell (License 5909) rb3671.patch uploaded by coreyfarrell (License 5909) rb3667.patch uploaded by coreyfarrell (License 5909) rb3665.patch uploaded by mjordan (License 6283) rb3625.patch uploaded by coreyfarrell (License 5909) rb3602.patch uploaded by coreyfarrell (License 5909) format_compatibility-2.diff uploaded by file (License 5000) core.diff uploaded by file (License 5000) git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@419044 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-07-20 22:06:33 +00:00
#include "asterisk/smoother.h"
#include "asterisk/uuid.h"
#include "asterisk/test.h"
#include "asterisk/data_buffer.h"
#ifdef HAVE_PJPROJECT
#include "asterisk/res_pjproject.h"
#include "asterisk/security_events.h"
#endif
#define MAX_TIMESTAMP_SKEW 640
#define RTP_SEQ_MOD (1<<16) /*!< A sequence number can't be more than 16 bits */
#define RTCP_DEFAULT_INTERVALMS 5000 /*!< Default milli-seconds between RTCP reports we send */
#define RTCP_MIN_INTERVALMS 500 /*!< Min milli-seconds between RTCP reports we send */
#define RTCP_MAX_INTERVALMS 60000 /*!< Max milli-seconds between RTCP reports we send */
#define DEFAULT_RTP_START 5000 /*!< Default port number to start allocating RTP ports from */
#define DEFAULT_RTP_END 31000 /*!< Default maximum port number to end allocating RTP ports at */
#define MINIMUM_RTP_PORT 1024 /*!< Minimum port number to accept */
#define MAXIMUM_RTP_PORT 65535 /*!< Maximum port number to accept */
#define DEFAULT_TURN_PORT 3478
#define TURN_STATE_WAIT_TIME 2000
#define DEFAULT_RTP_SEND_BUFFER_SIZE 250 /*!< The initial size of the RTP send buffer */
#define MAXIMUM_RTP_SEND_BUFFER_SIZE (DEFAULT_RTP_SEND_BUFFER_SIZE + 200) /*!< Maximum RTP send buffer size */
#define DEFAULT_RTP_RECV_BUFFER_SIZE 20 /*!< The initial size of the RTP receiver buffer */
#define MAXIMUM_RTP_RECV_BUFFER_SIZE (DEFAULT_RTP_RECV_BUFFER_SIZE + 20) /*!< Maximum RTP receive buffer size */
#define OLD_PACKET_COUNT 1000 /*!< The number of previous packets that are considered old */
#define MISSING_SEQNOS_ADDED_TRIGGER 2 /*!< The number of immediate missing packets that will trigger an immediate NACK */
#define SEQNO_CYCLE_OVER 65536 /*!< The number after the maximum allowed sequence number */
/*! Full INTRA-frame Request / Fast Update Request (From RFC2032) */
#define RTCP_PT_FUR 192
/*! Sender Report (From RFC3550) */
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
#define RTCP_PT_SR AST_RTP_RTCP_SR
/*! Receiver Report (From RFC3550) */
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
#define RTCP_PT_RR AST_RTP_RTCP_RR
/*! Source Description (From RFC3550) */
#define RTCP_PT_SDES 202
/*! Goodbye (To remove SSRC's from tables) (From RFC3550) */
#define RTCP_PT_BYE 203
/*! Application defined (From RFC3550) */
#define RTCP_PT_APP 204
/* VP8: RTCP Feedback */
/*! Payload Specific Feed Back (From RFC4585 also RFC5104) */
#define RTCP_PT_PSFB AST_RTP_RTCP_PSFB
#define RTP_MTU 1200
#define DTMF_SAMPLE_RATE_MS 8 /*!< DTMF samples per millisecond */
#define DEFAULT_DTMF_TIMEOUT (150 * (8000 / 1000)) /*!< samples */
#define ZFONE_PROFILE_ID 0x505a
#define DEFAULT_LEARNING_MIN_SEQUENTIAL 4
/*!
* \brief Calculate the min learning duration in ms.
*
* \details
* The min supported packet size represents 10 ms and we need to account
* for some jitter and fast clocks while learning. Some messed up devices
* have very bad jitter for a small packet sample size. Jitter can also
* be introduced by the network itself.
*
* So we'll allow packets to come in every 9ms on average for fast clocking
* with the last one coming in 5ms early for jitter.
*/
#define CALC_LEARNING_MIN_DURATION(count) (((count) - 1) * 9 - 5)
#define DEFAULT_LEARNING_MIN_DURATION CALC_LEARNING_MIN_DURATION(DEFAULT_LEARNING_MIN_SEQUENTIAL)
#define SRTP_MASTER_KEY_LEN 16
#define SRTP_MASTER_SALT_LEN 14
#define SRTP_MASTER_LEN (SRTP_MASTER_KEY_LEN + SRTP_MASTER_SALT_LEN)
#define RTP_DTLS_ESTABLISHED -37
enum strict_rtp_state {
STRICT_RTP_OPEN = 0, /*! No RTP packets should be dropped, all sources accepted */
STRICT_RTP_LEARN, /*! Accept next packet as source */
STRICT_RTP_CLOSED, /*! Drop all RTP packets not coming from source that was learned */
};
enum strict_rtp_mode {
STRICT_RTP_NO = 0, /*! Don't adhere to any strict RTP rules */
STRICT_RTP_YES, /*! Strict RTP that restricts packets based on time and sequence number */
STRICT_RTP_SEQNO, /*! Strict RTP that restricts packets based on sequence number */
};
/*!
* \brief Strict RTP learning timeout time in milliseconds
*
* \note Set to 5 seconds to allow reinvite chains for direct media
* to settle before media actually starts to arrive. There may be a
* reinvite collision involved on the other leg.
*/
#define STRICT_RTP_LEARN_TIMEOUT 5000
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
#define DEFAULT_STRICT_RTP STRICT_RTP_YES /*!< Enabled by default */
#define DEFAULT_SRTP_REPLAY_PROTECTION 1
#define DEFAULT_ICESUPPORT 1
#define DEFAULT_STUN_SOFTWARE_ATTRIBUTE 1
#define DEFAULT_DTLS_MTU 1200
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
/*!
* Because both ends usually don't start sending RTP
* at the same time, some of the calculations like
* rtt and jitter will probably be unstable for a while
* so we'll skip some received packets before starting
* analyzing. This just affects analyzing; we still
* process the RTP as normal.
*/
#define RTP_IGNORE_FIRST_PACKETS_COUNT 15
extern struct ast_srtp_res *res_srtp;
extern struct ast_srtp_policy_res *res_srtp_policy;
static int dtmftimeout = DEFAULT_DTMF_TIMEOUT;
static int rtpstart = DEFAULT_RTP_START; /*!< First port for RTP sessions (set in rtp.conf) */
static int rtpend = DEFAULT_RTP_END; /*!< Last port for RTP sessions (set in rtp.conf) */
static int rtcpstats; /*!< Are we debugging RTCP? */
static int rtcpinterval = RTCP_DEFAULT_INTERVALMS; /*!< Time between rtcp reports in millisecs */
static struct ast_sockaddr rtpdebugaddr; /*!< Debug packets to/from this host */
static struct ast_sockaddr rtcpdebugaddr; /*!< Debug RTCP packets to/from this host */
static int rtpdebugport; /*!< Debug only RTP packets from IP or IP+Port if port is > 0 */
static int rtcpdebugport; /*!< Debug only RTCP packets from IP or IP+Port if port is > 0 */
#ifdef SO_NO_CHECK
static int nochecksums;
#endif
static int strictrtp = DEFAULT_STRICT_RTP; /*!< Only accept RTP frames from a defined source. If we receive an indication of a changing source, enter learning mode. */
static int learning_min_sequential = DEFAULT_LEARNING_MIN_SEQUENTIAL; /*!< Number of sequential RTP frames needed from a single source during learning mode to accept new source. */
static int learning_min_duration = DEFAULT_LEARNING_MIN_DURATION; /*!< Lowest acceptable timeout between the first and the last sequential RTP frame. */
static int srtp_replay_protection = DEFAULT_SRTP_REPLAY_PROTECTION;
#if defined(HAVE_OPENSSL) && (OPENSSL_VERSION_NUMBER >= 0x10001000L) && !defined(OPENSSL_NO_SRTP)
static int dtls_mtu = DEFAULT_DTLS_MTU;
#endif
#ifdef HAVE_PJPROJECT
static int icesupport = DEFAULT_ICESUPPORT;
static int stun_software_attribute = DEFAULT_STUN_SOFTWARE_ATTRIBUTE;
static struct sockaddr_in stunaddr;
static pj_str_t turnaddr;
static int turnport = DEFAULT_TURN_PORT;
static pj_str_t turnusername;
static pj_str_t turnpassword;
static struct stasis_subscription *acl_change_sub = NULL;
static struct ast_sockaddr lo6 = { .len = 0 };
/*! ACL for ICE addresses */
static struct ast_acl_list *ice_acl = NULL;
static ast_rwlock_t ice_acl_lock = AST_RWLOCK_INIT_VALUE;
/*! ACL for STUN requests */
static struct ast_acl_list *stun_acl = NULL;
static ast_rwlock_t stun_acl_lock = AST_RWLOCK_INIT_VALUE;
/*! stunaddr recurring resolution */
static ast_rwlock_t stunaddr_lock = AST_RWLOCK_INIT_VALUE;
static struct ast_dns_query_recurring *stunaddr_resolver = NULL;
/*! \brief Pool factory used by pjlib to allocate memory. */
static pj_caching_pool cachingpool;
/*! \brief Global memory pool for configuration and timers */
static pj_pool_t *pool;
/*! \brief Global timer heap */
static pj_timer_heap_t *timer_heap;
/*! \brief Thread executing the timer heap */
static pj_thread_t *timer_thread;
/*! \brief Used to tell the timer thread to terminate */
static int timer_terminate;
/*! \brief Structure which contains ioqueue thread information */
struct ast_rtp_ioqueue_thread {
/*! \brief Pool used by the thread */
pj_pool_t *pool;
/*! \brief The thread handling the queue and timer heap */
pj_thread_t *thread;
/*! \brief Ioqueue which polls on sockets */
pj_ioqueue_t *ioqueue;
/*! \brief Timer heap for scheduled items */
pj_timer_heap_t *timerheap;
/*! \brief Termination request */
int terminate;
/*! \brief Current number of descriptors being waited on */
unsigned int count;
/*! \brief Linked list information */
AST_LIST_ENTRY(ast_rtp_ioqueue_thread) next;
};
/*! \brief List of ioqueue threads */
static AST_LIST_HEAD_STATIC(ioqueues, ast_rtp_ioqueue_thread);
/*! \brief Structure which contains ICE host candidate mapping information */
struct ast_ice_host_candidate {
struct ast_sockaddr local;
struct ast_sockaddr advertised;
unsigned int include_local;
AST_RWLIST_ENTRY(ast_ice_host_candidate) next;
};
/*! \brief List of ICE host candidate mappings */
static AST_RWLIST_HEAD_STATIC(host_candidates, ast_ice_host_candidate);
static char *generate_random_string(char *buf, size_t size);
#endif
#define FLAG_3389_WARNING (1 << 0)
#define FLAG_NAT_ACTIVE (3 << 1)
#define FLAG_NAT_INACTIVE (0 << 1)
#define FLAG_NAT_INACTIVE_NOWARN (1 << 1)
#define FLAG_NEED_MARKER_BIT (1 << 3)
#define FLAG_DTMF_COMPENSATE (1 << 4)
#define FLAG_REQ_LOCAL_BRIDGE_BIT (1 << 5)
#define TRANSPORT_SOCKET_RTP 0
#define TRANSPORT_SOCKET_RTCP 1
#define TRANSPORT_TURN_RTP 2
#define TRANSPORT_TURN_RTCP 3
/*! \brief RTP learning mode tracking information */
struct rtp_learning_info {
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
struct ast_sockaddr proposed_address; /*!< Proposed remote address for strict RTP */
struct timeval start; /*!< The time learning mode was started */
struct timeval received; /*!< The time of the first received packet */
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
int max_seq; /*!< The highest sequence number received */
int packets; /*!< The number of remaining packets before the source is accepted */
/*! Type of media stream carried by the RTP instance */
enum ast_media_type stream_type;
};
#if defined(HAVE_OPENSSL) && (OPENSSL_VERSION_NUMBER >= 0x10001000L) && !defined(OPENSSL_NO_SRTP)
struct dtls_details {
SSL *ssl; /*!< SSL session */
BIO *read_bio; /*!< Memory buffer for reading */
BIO *write_bio; /*!< Memory buffer for writing */
enum ast_rtp_dtls_setup dtls_setup; /*!< Current setup state */
enum ast_rtp_dtls_connection connection; /*!< Whether this is a new or existing connection */
int timeout_timer; /*!< Scheduler id for timeout timer */
};
#endif
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
#ifdef HAVE_PJPROJECT
/*! An ao2 wrapper protecting the PJPROJECT ice structure with ref counting. */
struct ice_wrap {
pj_ice_sess *real_ice; /*!< ICE session */
};
#endif
/*! \brief Structure used for mapping an incoming SSRC to an RTP instance */
struct rtp_ssrc_mapping {
/*! \brief The received SSRC */
unsigned int ssrc;
/*! True if the SSRC is available. Otherwise, this is a placeholder mapping until the SSRC is set. */
unsigned int ssrc_valid;
/*! \brief The RTP instance this SSRC belongs to*/
struct ast_rtp_instance *instance;
};
/*! \brief Packet statistics (used for transport-cc) */
struct rtp_transport_wide_cc_packet_statistics {
/*! The transport specific sequence number */
unsigned int seqno;
/*! The time at which the packet was received */
struct timeval received;
/*! The delta between this packet and the previous */
int delta;
};
/*! \brief Statistics information (used for transport-cc) */
struct rtp_transport_wide_cc_statistics {
/*! A vector of packet statistics */
AST_VECTOR(, struct rtp_transport_wide_cc_packet_statistics) packet_statistics; /*!< Packet statistics, used for transport-cc */
/*! The last sequence number received */
unsigned int last_seqno;
/*! The last extended sequence number */
unsigned int last_extended_seqno;
/*! How many feedback packets have gone out */
unsigned int feedback_count;
/*! How many cycles have occurred for the sequence numbers */
unsigned int cycles;
/*! Scheduler id for periodic feedback transmission */
int schedid;
};
typedef struct {
unsigned int ts;
unsigned char is_set;
} optional_ts;
/*! \brief RTP session description */
struct ast_rtp {
int s;
/*! \note The f.subclass.format holds a ref. */
struct ast_frame f;
unsigned char rawdata[8192 + AST_FRIENDLY_OFFSET];
unsigned int ssrc; /*!< Synchronization source, RFC 3550, page 10. */
unsigned int ssrc_orig; /*!< SSRC used before native bridge activated */
unsigned char ssrc_saved; /*!< indicates if ssrc_orig has a value */
char cname[AST_UUID_STR_LEN]; /*!< Our local CNAME */
unsigned int themssrc; /*!< Their SSRC */
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
unsigned int themssrc_valid; /*!< True if their SSRC is available. */
unsigned int lastts;
unsigned int lastividtimestamp;
unsigned int lastovidtimestamp;
unsigned int lastitexttimestamp;
unsigned int lastotexttimestamp;
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
int prevrxseqno; /*!< Previous received packeted sequence number, from the network */
int lastrxseqno; /*!< Last received sequence number, from the network */
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
int expectedrxseqno; /*!< Next expected sequence number, from the network */
2018-05-10 18:11:06 +00:00
AST_VECTOR(, int) missing_seqno; /*!< A vector of sequence numbers we never received */
int expectedseqno; /*!< Next expected sequence number, from the core */
unsigned short seedrxseqno; /*!< What sequence number did they start with?*/
unsigned int rxcount; /*!< How many packets have we received? */
unsigned int rxoctetcount; /*!< How many octets have we received? should be rxcount *160*/
unsigned int txcount; /*!< How many packets have we sent? */
unsigned int txoctetcount; /*!< How many octets have we sent? (txcount*160)*/
unsigned int cycles; /*!< Shifted count of sequence number cycles */
media formats: re-architect handling of media for performance improvements In the old times media formats were represented using a bit field. This was fast but had a few limitations. 1. Asterisk was limited in how many formats it could handle. 2. Formats, being a bit field, could not include any attribute information. A format was strictly its type, e.g., "this is ulaw". This was changed in Asterisk 10 (see https://wiki.asterisk.org/wiki/display/AST/Media+Architecture+Proposal for notes on that work) which led to the creation of the ast_format structure. This structure allowed Asterisk to handle attributes and bundle information with a format. Additionally, ast_format_cap was created to act as a container for multiple formats that, together, formed the capability of some entity. Another mechanism was added to allow logic to be registered which performed format attribute negotiation. Everywhere throughout the codebase Asterisk was changed to use this strategy. Unfortunately, in software, there is no free lunch. These new capabilities came at a cost. Performance analysis and profiling showed that we spend an inordinate amount of time comparing, copying, and generally manipulating formats and their related structures. Basic prototyping has shown that a reasonably large performance improvement could be made in this area. This patch is the result of that project, which overhauled the media format architecture and its usage in Asterisk to improve performance. Generally, the new philosophy for handling formats is as follows: * The ast_format structure is reference counted. This removed a large amount of the memory allocations and copying that was done in prior versions. * In order to prevent race conditions while keeping things performant, the ast_format structure is immutable by convention and lock-free. Violate this tenet at your peril! * Because formats are reference counted, codecs are also reference counted. The Asterisk core generally provides built-in codecs and caches the ast_format structures created to represent them. Generally, to prevent inordinate amounts of module reference bumping, codecs and formats can be added at run-time but cannot be removed. * All compatibility with the bit field representation of codecs/formats has been moved to a compatibility API. The primary user of this representation is chan_iax2, which must continue to maintain its bit-field usage of formats for interoperability concerns. * When a format is negotiated with attributes, or when a format cannot be represented by one of the cached formats, a new format object is created or cloned from an existing format. That format may have the same codec underlying it, but is a different format than a version of the format with different attributes or without attributes. * While formats are reference counted objects, the reference count maintained on the format should be manipulated with care. Formats are generally cached and will persist for the lifetime of Asterisk and do not explicitly need to have their lifetime modified. An exception to this is when the user of a format does not know where the format came from *and* the user may outlive the provider of the format. This occurs, for example, when a format is read from a channel: the channel may have a format with attributes (hence, non-cached) and the user of the format may last longer than the channel (if the reference to the channel is released prior to the format's reference). For more information on this work, see the API design notes: https://wiki.asterisk.org/wiki/display/AST/Media+Format+Rewrite Finally, this work was the culmination of a large number of developer's efforts. Extra thanks goes to Corey Farrell, who took on a large amount of the work in the Asterisk core, chan_sip, and was an invaluable resource in peer reviews throughout this project. There were a substantial number of patches contributed during this work; the following issues/patch names simply reflect some of the work (and will cause the release scripts to give attribution to the individuals who work on them). Reviews: https://reviewboard.asterisk.org/r/3814 https://reviewboard.asterisk.org/r/3808 https://reviewboard.asterisk.org/r/3805 https://reviewboard.asterisk.org/r/3803 https://reviewboard.asterisk.org/r/3801 https://reviewboard.asterisk.org/r/3798 https://reviewboard.asterisk.org/r/3800 https://reviewboard.asterisk.org/r/3794 https://reviewboard.asterisk.org/r/3793 https://reviewboard.asterisk.org/r/3792 https://reviewboard.asterisk.org/r/3791 https://reviewboard.asterisk.org/r/3790 https://reviewboard.asterisk.org/r/3789 https://reviewboard.asterisk.org/r/3788 https://reviewboard.asterisk.org/r/3787 https://reviewboard.asterisk.org/r/3786 https://reviewboard.asterisk.org/r/3784 https://reviewboard.asterisk.org/r/3783 https://reviewboard.asterisk.org/r/3778 https://reviewboard.asterisk.org/r/3774 https://reviewboard.asterisk.org/r/3775 https://reviewboard.asterisk.org/r/3772 https://reviewboard.asterisk.org/r/3761 https://reviewboard.asterisk.org/r/3754 https://reviewboard.asterisk.org/r/3753 https://reviewboard.asterisk.org/r/3751 https://reviewboard.asterisk.org/r/3750 https://reviewboard.asterisk.org/r/3748 https://reviewboard.asterisk.org/r/3747 https://reviewboard.asterisk.org/r/3746 https://reviewboard.asterisk.org/r/3742 https://reviewboard.asterisk.org/r/3740 https://reviewboard.asterisk.org/r/3739 https://reviewboard.asterisk.org/r/3738 https://reviewboard.asterisk.org/r/3737 https://reviewboard.asterisk.org/r/3736 https://reviewboard.asterisk.org/r/3734 https://reviewboard.asterisk.org/r/3722 https://reviewboard.asterisk.org/r/3713 https://reviewboard.asterisk.org/r/3703 https://reviewboard.asterisk.org/r/3689 https://reviewboard.asterisk.org/r/3687 https://reviewboard.asterisk.org/r/3674 https://reviewboard.asterisk.org/r/3671 https://reviewboard.asterisk.org/r/3667 https://reviewboard.asterisk.org/r/3665 https://reviewboard.asterisk.org/r/3625 https://reviewboard.asterisk.org/r/3602 https://reviewboard.asterisk.org/r/3519 https://reviewboard.asterisk.org/r/3518 https://reviewboard.asterisk.org/r/3516 https://reviewboard.asterisk.org/r/3515 https://reviewboard.asterisk.org/r/3512 https://reviewboard.asterisk.org/r/3506 https://reviewboard.asterisk.org/r/3413 https://reviewboard.asterisk.org/r/3410 https://reviewboard.asterisk.org/r/3387 https://reviewboard.asterisk.org/r/3388 https://reviewboard.asterisk.org/r/3389 https://reviewboard.asterisk.org/r/3390 https://reviewboard.asterisk.org/r/3321 https://reviewboard.asterisk.org/r/3320 https://reviewboard.asterisk.org/r/3319 https://reviewboard.asterisk.org/r/3318 https://reviewboard.asterisk.org/r/3266 https://reviewboard.asterisk.org/r/3265 https://reviewboard.asterisk.org/r/3234 https://reviewboard.asterisk.org/r/3178 ASTERISK-23114 #close Reported by: mjordan media_formats_translation_core.diff uploaded by kharwell (License 6464) rb3506.diff uploaded by mjordan (License 6283) media_format_app_file.diff uploaded by kharwell (License 6464) misc-2.diff uploaded by file (License 5000) chan_mild-3.diff uploaded by file (License 5000) chan_obscure.diff uploaded by file (License 5000) jingle.diff uploaded by file (License 5000) funcs.diff uploaded by file (License 5000) formats.diff uploaded by file (License 5000) core.diff uploaded by file (License 5000) bridges.diff uploaded by file (License 5000) mf-codecs-2.diff uploaded by file (License 5000) mf-app_fax.diff uploaded by file (License 5000) mf-apps-3.diff uploaded by file (License 5000) media-formats-3.diff uploaded by file (License 5000) ASTERISK-23715 rb3713.patch uploaded by coreyfarrell (License 5909) rb3689.patch uploaded by mjordan (License 6283) ASTERISK-23957 rb3722.patch uploaded by mjordan (License 6283) mf-attributes-3.diff uploaded by file (License 5000) ASTERISK-23958 Tested by: jrose rb3822.patch uploaded by coreyfarrell (License 5909) rb3800.patch uploaded by jrose (License 6182) chan_sip.diff uploaded by mjordan (License 6283) rb3747.patch uploaded by jrose (License 6182) ASTERISK-23959 #close Tested by: sgriepentrog, mjordan, coreyfarrell sip_cleanup.diff uploaded by opticron (License 6273) chan_sip_caps.diff uploaded by mjordan (License 6283) rb3751.patch uploaded by coreyfarrell (License 5909) chan_sip-3.diff uploaded by file (License 5000) ASTERISK-23960 #close Tested by: opticron direct_media.diff uploaded by opticron (License 6273) pjsip-direct-media.diff uploaded by file (License 5000) format_cap_remove.diff uploaded by opticron (License 6273) media_format_fixes.diff uploaded by opticron (License 6273) chan_pjsip-2.diff uploaded by file (License 5000) ASTERISK-23966 #close Tested by: rmudgett rb3803.patch uploaded by rmudgetti (License 5621) chan_dahdi.diff uploaded by file (License 5000) ASTERISK-24064 #close Tested by: coreyfarrell, mjordan, opticron, file, rmudgett, sgriepentrog, jrose rb3814.patch uploaded by rmudgett (License 5621) moh_cleanup.diff uploaded by opticron (License 6273) bridge_leak.diff uploaded by opticron (License 6273) translate.diff uploaded by file (License 5000) rb3795.patch uploaded by rmudgett (License 5621) tls_fix.diff uploaded by mjordan (License 6283) fax-mf-fix-2.diff uploaded by file (License 5000) rtp_transfer_stuff uploaded by mjordan (License 6283) rb3787.patch uploaded by rmudgett (License 5621) media-formats-explicit-translate-format-3.diff uploaded by file (License 5000) format_cache_case_fix.diff uploaded by opticron (License 6273) rb3774.patch uploaded by rmudgett (License 5621) rb3775.patch uploaded by rmudgett (License 5621) rtp_engine_fix.diff uploaded by opticron (License 6273) rtp_crash_fix.diff uploaded by opticron (License 6273) rb3753.patch uploaded by mjordan (License 6283) rb3750.patch uploaded by mjordan (License 6283) rb3748.patch uploaded by rmudgett (License 5621) media_format_fixes.diff uploaded by opticron (License 6273) rb3740.patch uploaded by mjordan (License 6283) rb3739.patch uploaded by mjordan (License 6283) rb3734.patch uploaded by mjordan (License 6283) rb3689.patch uploaded by mjordan (License 6283) rb3674.patch uploaded by coreyfarrell (License 5909) rb3671.patch uploaded by coreyfarrell (License 5909) rb3667.patch uploaded by coreyfarrell (License 5909) rb3665.patch uploaded by mjordan (License 6283) rb3625.patch uploaded by coreyfarrell (License 5909) rb3602.patch uploaded by coreyfarrell (License 5909) format_compatibility-2.diff uploaded by file (License 5000) core.diff uploaded by file (License 5000) git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@419044 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-07-20 22:06:33 +00:00
struct ast_format *lasttxformat;
struct ast_format *lastrxformat;
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
/*
* RX RTP Timestamp and Jitter calculation.
*/
double rxstart; /*!< RX time of the first packet in the session in seconds since EPOCH. */
double rxstart_stable; /*!< RX time of the first packet after RTP_IGNORE_FIRST_PACKETS_COUNT */
unsigned int remote_seed_rx_rtp_ts; /*!< RTP timestamp of first RX packet. */
unsigned int remote_seed_rx_rtp_ts_stable; /*!< RTP timestamp of first packet after RTP_IGNORE_FIRST_PACKETS_COUNT */
unsigned int last_transit_time_samples; /*!< The last transit time in samples */
double rxjitter; /*!< Last calculated Interarrival jitter in seconds. */
double rxjitter_samples; /*!< Last calculated Interarrival jitter in samples. */
double rxmes; /*!< Media Experince Score at the moment to be reported */
/* DTMF Reception Variables */
Handle extremely out of order RFC 2833 DTMF The current implementation of RFC 2833 DTMF handling in res_rtp_asterisk will, if a packet arrives out of order, drop the packet. This is to prevent duplicate ton generation in the Asterisk core. Since the RTP layer does not buffer data itself, this is the only option the RTP layer currently has for handling packets that arrive out of order. For the most part, this doesn't matter. For a particular digit, so long as a BEGIN packet arrives before the first END packet, the digit will be produced. If subsequent BEGIN packets arrive interleaved with the ENDs, they will be dropped; likewise, if the BEGIN or END packets themselves are out of order, those packets are dropped but sufficient information is conveyed to the Asterisk core to produce the appropriate digit. For certain sequences of DTMF packets - most notably when, for a particular digit, an END packet arrives before any BEGIN packet for that digit - this is a real problem. When an END arrives before any BEGINs, the END packet is dropped - but at the same time, it causes subsequent BEGIN packets for that digit to be ignored. When the next in order END packet arrives, it too is dropped - Asterisk believes that there was no initial BEGIN. The solution this patch provides is to trust the END packet to convey the information needed for the Asterisk core to produce the DTMF digit. If we receive an END packet, and it: * Has a timestamp greater then the last timestamp received from an END packet * Does not have the same sequence number as the last received sequence number (and is thus not an END packet retransmission) Then we send the END frame up to the Asterisk core. It contains enough DTMF information for Asterisk to produce the digit. On the other hand, if we receive a BEGIN or continuation packet that occurs with a timestamp equal to or less then the last END timestamp, then we've received something out of order - but we already have received enough information to produce the digit. These packets are dropped. Much thanks goes to Olle Johansson (oej) for providing the idea for this solution. Review: https://reviewboard.asterisk.org/r/2033/ (closes issue ASTERISK-18404) Reported by: Stephane Chazelas Tested by: Matt Jordan ........ Merged revisions 370252 from http://svn.asterisk.org/svn/asterisk/branches/1.8 ........ Merged revisions 370271 from http://svn.asterisk.org/svn/asterisk/branches/10 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@370272 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2012-07-19 21:45:20 +00:00
char resp; /*!< The current digit being processed */
unsigned int last_seqno; /*!< The last known sequence number for any DTMF packet */
optional_ts last_end_timestamp; /*!< The last known timestamp received from an END packet */
Handle extremely out of order RFC 2833 DTMF The current implementation of RFC 2833 DTMF handling in res_rtp_asterisk will, if a packet arrives out of order, drop the packet. This is to prevent duplicate ton generation in the Asterisk core. Since the RTP layer does not buffer data itself, this is the only option the RTP layer currently has for handling packets that arrive out of order. For the most part, this doesn't matter. For a particular digit, so long as a BEGIN packet arrives before the first END packet, the digit will be produced. If subsequent BEGIN packets arrive interleaved with the ENDs, they will be dropped; likewise, if the BEGIN or END packets themselves are out of order, those packets are dropped but sufficient information is conveyed to the Asterisk core to produce the appropriate digit. For certain sequences of DTMF packets - most notably when, for a particular digit, an END packet arrives before any BEGIN packet for that digit - this is a real problem. When an END arrives before any BEGINs, the END packet is dropped - but at the same time, it causes subsequent BEGIN packets for that digit to be ignored. When the next in order END packet arrives, it too is dropped - Asterisk believes that there was no initial BEGIN. The solution this patch provides is to trust the END packet to convey the information needed for the Asterisk core to produce the DTMF digit. If we receive an END packet, and it: * Has a timestamp greater then the last timestamp received from an END packet * Does not have the same sequence number as the last received sequence number (and is thus not an END packet retransmission) Then we send the END frame up to the Asterisk core. It contains enough DTMF information for Asterisk to produce the digit. On the other hand, if we receive a BEGIN or continuation packet that occurs with a timestamp equal to or less then the last END timestamp, then we've received something out of order - but we already have received enough information to produce the digit. These packets are dropped. Much thanks goes to Olle Johansson (oej) for providing the idea for this solution. Review: https://reviewboard.asterisk.org/r/2033/ (closes issue ASTERISK-18404) Reported by: Stephane Chazelas Tested by: Matt Jordan ........ Merged revisions 370252 from http://svn.asterisk.org/svn/asterisk/branches/1.8 ........ Merged revisions 370271 from http://svn.asterisk.org/svn/asterisk/branches/10 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@370272 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2012-07-19 21:45:20 +00:00
unsigned int dtmf_duration; /*!< Total duration in samples since the digit start event */
unsigned int dtmf_timeout; /*!< When this timestamp is reached we consider END frame lost and forcibly abort digit */
unsigned int dtmfsamples;
Handle extremely out of order RFC 2833 DTMF The current implementation of RFC 2833 DTMF handling in res_rtp_asterisk will, if a packet arrives out of order, drop the packet. This is to prevent duplicate ton generation in the Asterisk core. Since the RTP layer does not buffer data itself, this is the only option the RTP layer currently has for handling packets that arrive out of order. For the most part, this doesn't matter. For a particular digit, so long as a BEGIN packet arrives before the first END packet, the digit will be produced. If subsequent BEGIN packets arrive interleaved with the ENDs, they will be dropped; likewise, if the BEGIN or END packets themselves are out of order, those packets are dropped but sufficient information is conveyed to the Asterisk core to produce the appropriate digit. For certain sequences of DTMF packets - most notably when, for a particular digit, an END packet arrives before any BEGIN packet for that digit - this is a real problem. When an END arrives before any BEGINs, the END packet is dropped - but at the same time, it causes subsequent BEGIN packets for that digit to be ignored. When the next in order END packet arrives, it too is dropped - Asterisk believes that there was no initial BEGIN. The solution this patch provides is to trust the END packet to convey the information needed for the Asterisk core to produce the DTMF digit. If we receive an END packet, and it: * Has a timestamp greater then the last timestamp received from an END packet * Does not have the same sequence number as the last received sequence number (and is thus not an END packet retransmission) Then we send the END frame up to the Asterisk core. It contains enough DTMF information for Asterisk to produce the digit. On the other hand, if we receive a BEGIN or continuation packet that occurs with a timestamp equal to or less then the last END timestamp, then we've received something out of order - but we already have received enough information to produce the digit. These packets are dropped. Much thanks goes to Olle Johansson (oej) for providing the idea for this solution. Review: https://reviewboard.asterisk.org/r/2033/ (closes issue ASTERISK-18404) Reported by: Stephane Chazelas Tested by: Matt Jordan ........ Merged revisions 370252 from http://svn.asterisk.org/svn/asterisk/branches/1.8 ........ Merged revisions 370271 from http://svn.asterisk.org/svn/asterisk/branches/10 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@370272 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2012-07-19 21:45:20 +00:00
enum ast_rtp_dtmf_mode dtmfmode; /*!< The current DTMF mode of the RTP stream */
/* DTMF Transmission Variables */
unsigned int lastdigitts;
char sending_digit; /*!< boolean - are we sending digits */
char send_digit; /*!< digit we are sending */
int send_payload;
int send_duration;
unsigned int flags;
struct timeval rxcore;
struct timeval txcore;
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
struct timeval dtmfmute;
struct ast_smoother *smoother;
unsigned short seqno; /*!< Sequence number, RFC 3550, page 13. */
struct ast_sched_context *sched;
struct ast_rtcp *rtcp;
unsigned int asymmetric_codec; /*!< Indicate if asymmetric send/receive codecs are allowed */
struct ast_rtp_instance *bundled; /*!< The RTP instance we are bundled to */
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
/*!
* \brief The RTP instance owning us (used for debugging purposes)
* We don't hold a reference to the instance because it created
* us in the first place. It can't go away.
*/
struct ast_rtp_instance *owner;
int stream_num; /*!< Stream num for this RTP instance */
AST_VECTOR(, struct rtp_ssrc_mapping) ssrc_mapping; /*!< Mappings of SSRC to RTP instances */
struct ast_sockaddr bind_address; /*!< Requested bind address for the sockets */
enum strict_rtp_state strict_rtp_state; /*!< Current state that strict RTP protection is in */
struct ast_sockaddr strict_rtp_address; /*!< Remote address information for strict RTP purposes */
/*
* Learning mode values based on pjmedia's probation mode. Many of these values are redundant to the above,
* but these are in place to keep learning mode sequence values sealed from their normal counterparts.
*/
struct rtp_learning_info rtp_source_learn; /* Learning mode track for the expected RTP source */
struct rtp_red *red;
struct ast_data_buffer *send_buffer; /*!< Buffer for storing sent packets for retransmission */
2018-05-10 18:11:06 +00:00
struct ast_data_buffer *recv_buffer; /*!< Buffer for storing received packets for retransmission */
struct rtp_transport_wide_cc_statistics transport_wide_cc; /*!< Transport-cc statistics information */
#ifdef HAVE_PJPROJECT
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ast_cond_t cond; /*!< ICE/TURN condition for signaling */
struct ice_wrap *ice; /*!< ao2 wrapped ICE session */
enum ast_rtp_ice_role role; /*!< Our role in ICE negotiation */
pj_turn_sock *turn_rtp; /*!< RTP TURN relay */
pj_turn_sock *turn_rtcp; /*!< RTCP TURN relay */
pj_turn_state_t turn_state; /*!< Current state of the TURN relay session */
unsigned int passthrough:1; /*!< Bit to indicate that the received packet should be passed through */
unsigned int rtp_passthrough:1; /*!< Bit to indicate that TURN RTP should be passed through */
unsigned int rtcp_passthrough:1; /*!< Bit to indicate that TURN RTCP should be passed through */
unsigned int ice_port; /*!< Port that ICE was started with if it was previously started */
struct ast_sockaddr rtp_loop; /*!< Loopback address for forwarding RTP from TURN */
struct ast_sockaddr rtcp_loop; /*!< Loopback address for forwarding RTCP from TURN */
struct ast_rtp_ioqueue_thread *ioqueue; /*!< The ioqueue thread handling us */
char remote_ufrag[256]; /*!< The remote ICE username */
char remote_passwd[256]; /*!< The remote ICE password */
char local_ufrag[256]; /*!< The local ICE username */
char local_passwd[256]; /*!< The local ICE password */
struct ao2_container *ice_local_candidates; /*!< The local ICE candidates */
struct ao2_container *ice_active_remote_candidates; /*!< The remote ICE candidates */
struct ao2_container *ice_proposed_remote_candidates; /*!< Incoming remote ICE candidates for new session */
struct ast_sockaddr ice_original_rtp_addr; /*!< rtp address that ICE started on first session */
unsigned int ice_num_components; /*!< The number of ICE components */
unsigned int ice_media_started:1; /*!< ICE media has started, either on a valid pair or on ICE completion */
#endif
#if defined(HAVE_OPENSSL) && (OPENSSL_VERSION_NUMBER >= 0x10001000L) && !defined(OPENSSL_NO_SRTP)
SSL_CTX *ssl_ctx; /*!< SSL context */
enum ast_rtp_dtls_verify dtls_verify; /*!< What to verify */
enum ast_srtp_suite suite; /*!< SRTP crypto suite */
enum ast_rtp_dtls_hash local_hash; /*!< Local hash used for the fingerprint */
char local_fingerprint[160]; /*!< Fingerprint of our certificate */
enum ast_rtp_dtls_hash remote_hash; /*!< Remote hash used for the fingerprint */
unsigned char remote_fingerprint[EVP_MAX_MD_SIZE]; /*!< Fingerprint of the peer certificate */
unsigned int rekey; /*!< Interval at which to renegotiate and rekey */
int rekeyid; /*!< Scheduled item id for rekeying */
struct dtls_details dtls; /*!< DTLS state information */
#endif
};
/*!
* \brief Structure defining an RTCP session.
*
* The concept "RTCP session" is not defined in RFC 3550, but since
* this structure is analogous to ast_rtp, which tracks a RTP session,
* it is logical to think of this as a RTCP session.
*
* RTCP packet is defined on page 9 of RFC 3550.
*
*/
struct ast_rtcp {
int rtcp_info;
int s; /*!< Socket */
struct ast_sockaddr us; /*!< Socket representation of the local endpoint. */
struct ast_sockaddr them; /*!< Socket representation of the remote endpoint. */
unsigned int soc; /*!< What they told us */
unsigned int spc; /*!< What they told us */
unsigned int themrxlsr; /*!< The middle 32 bits of the NTP timestamp in the last received SR*/
struct timeval rxlsr; /*!< Time when we got their last SR */
struct timeval txlsr; /*!< Time when we sent or last SR*/
unsigned int expected_prior; /*!< no. packets in previous interval */
unsigned int received_prior; /*!< no. packets received in previous interval */
int schedid; /*!< Schedid returned from ast_sched_add() to schedule RTCP-transmissions*/
unsigned int rr_count; /*!< number of RRs we've sent, not including report blocks in SR's */
unsigned int sr_count; /*!< number of SRs we've sent */
unsigned int lastsrtxcount; /*!< Transmit packet count when last SR sent */
double accumulated_transit; /*!< accumulated a-dlsr-lsr */
double rtt; /*!< Last reported rtt */
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
double reported_jitter; /*!< The contents of their last jitter entry in the RR in seconds */
unsigned int reported_lost; /*!< Reported lost packets in their RR */
double reported_maxjitter; /*!< Maximum reported interarrival jitter */
double reported_minjitter; /*!< Minimum reported interarrival jitter */
double reported_normdev_jitter; /*!< Mean of reported interarrival jitter */
double reported_stdev_jitter; /*!< Standard deviation of reported interarrival jitter */
unsigned int reported_jitter_count; /*!< Reported interarrival jitter count */
double reported_maxlost; /*!< Maximum reported packets lost */
double reported_minlost; /*!< Minimum reported packets lost */
double reported_normdev_lost; /*!< Mean of reported packets lost */
double reported_stdev_lost; /*!< Standard deviation of reported packets lost */
unsigned int reported_lost_count; /*!< Reported packets lost count */
double rxlost; /*!< Calculated number of lost packets since last report */
double maxrxlost; /*!< Maximum calculated lost number of packets between reports */
double minrxlost; /*!< Minimum calculated lost number of packets between reports */
double normdev_rxlost; /*!< Mean of calculated lost packets between reports */
double stdev_rxlost; /*!< Standard deviation of calculated lost packets between reports */
unsigned int rxlost_count; /*!< Calculated lost packets sample count */
double maxrxjitter; /*!< Maximum of calculated interarrival jitter */
double minrxjitter; /*!< Minimum of calculated interarrival jitter */
double normdev_rxjitter; /*!< Mean of calculated interarrival jitter */
double stdev_rxjitter; /*!< Standard deviation of calculated interarrival jitter */
unsigned int rxjitter_count; /*!< Calculated interarrival jitter count */
double maxrtt; /*!< Maximum of calculated round trip time */
double minrtt; /*!< Minimum of calculated round trip time */
double normdevrtt; /*!< Mean of calculated round trip time */
double stdevrtt; /*!< Standard deviation of calculated round trip time */
unsigned int rtt_count; /*!< Calculated round trip time count */
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
double reported_mes; /*!< The calculated MES from their last RR */
double reported_maxmes; /*!< Maximum reported mes */
double reported_minmes; /*!< Minimum reported mes */
double reported_normdev_mes; /*!< Mean of reported mes */
double reported_stdev_mes; /*!< Standard deviation of reported mes */
unsigned int reported_mes_count; /*!< Reported mes count */
double maxrxmes; /*!< Maximum of calculated mes */
double minrxmes; /*!< Minimum of calculated mes */
double normdev_rxmes; /*!< Mean of calculated mes */
double stdev_rxmes; /*!< Standard deviation of calculated mes */
unsigned int rxmes_count; /*!< mes count */
/* VP8: sequence number for the RTCP FIR FCI */
int firseq;
#if defined(HAVE_OPENSSL) && (OPENSSL_VERSION_NUMBER >= 0x10001000L) && !defined(OPENSSL_NO_SRTP)
struct dtls_details dtls; /*!< DTLS state information */
#endif
/* Cached local address string allows us to generate
* RTCP stasis messages without having to look up our
* own address every time
*/
char *local_addr_str;
enum ast_rtp_instance_rtcp type;
/* Buffer for frames created during RTCP interpretation */
unsigned char frame_buf[512 + AST_FRIENDLY_OFFSET];
};
struct rtp_red {
struct ast_frame t140; /*!< Primary data */
struct ast_frame t140red; /*!< Redundant t140*/
unsigned char pt[AST_RED_MAX_GENERATION]; /*!< Payload types for redundancy data */
unsigned char ts[AST_RED_MAX_GENERATION]; /*!< Time stamps */
unsigned char len[AST_RED_MAX_GENERATION]; /*!< length of each generation */
int num_gen; /*!< Number of generations */
int schedid; /*!< Timer id */
int ti; /*!< How long to buffer data before send */
unsigned char t140red_data[64000];
unsigned char buf_data[64000]; /*!< buffered primary data */
int hdrlen;
long int prev_ts;
};
/*! \brief Structure for storing RTP packets for retransmission */
struct ast_rtp_rtcp_nack_payload {
size_t size; /*!< The size of the payload */
unsigned char buf[0]; /*!< The payload data */
};
Recorded merge of revisions 254452 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r254452 | mmichelson | 2010-03-25 10:59:56 -0500 (Thu, 25 Mar 2010) | 44 lines Several fixes regarding RFC2833 DTMF detection. Here is a copy and paste of the details from my request on reviewboard that dealt with these changes: Fix 1. The first change in place is to fix Mantis issue 15811, which deals with a situation where Asterisk will incorrectly interpret out of order RFC2833 frames as duplicate DTMF digits. For instance, we would receive a sequence like: seqno 1: DTMF 1 seqno 2: DTMF 1 seqno 3: DTMF 1 seqno 4: DTMF 1 seqno 6: DTMF 1 (end) seqno 5: DTMF 1 seqno 7: DTMF 1 (end) seqno 8: DTMF 1 (end) Prior to this patch when we received the frame with seqno 5, we would interpret this as a new DTMF 1. With this patch, we will check the seqno of the incoming digit and not process the frame if the seqno is lower than the last recorded seqno. Note that we do not record the seqno of the dropped DTMF frame for future processing. While the above situation is what was designed to be fixed, the patch is written in such a way that the following would also be fixed too: seqno 9: DTMF 1 seqno 10: DTMF 1 (end) seqno 11: DTMF 1 (end) seqno 13: DTMF 2 seqno 12: DTMF 1 (end) seqno 14: DTMF 2 seqno 15: DTMF 2 (end) seqno 16: DTMF 2 (end) seqno 17: DTMF 2 (end) In this second situation, the beginning of the DTMF 2 arrives before the final end frame of the DTMF 1. With the patch, seqno 12 is no processed and thus we properly interpret the DTMF. Fix 2. The second change in place is to fix an issue like the following: seqno 1: DTMF 1 seqno 2: DTMF 1 seqno 3: DTMF 1 (end) *packet lost* seqno 4: DTMF 1 (end) *packet lost* seqno 5: DTMF 1 (end) *packet lost* seqno 6: DTMF 2 When we receive seqno 6, we had code in place that was supposed to properly end the previously unended DTMF 1. The problem was that the code was essentially a no-op. The code would set up an end frame for the DTMF 1 but would immediately overwrite the frame with the begin for DTMF 2. I changed process_dtmf_rfc2833() so that instead of returning a single frame, it is given as an output parameter a list of frames. Each frame that needs to be returned is appended to this list. Fix 3. The final change is a minor one where an AST_CONTROL_SRCCHANGE frame could get lost. If we process a cisco DTMF or an RFC 3389 frame and no frame was returned, then we would return &ast_null_frame. The problem is that earlier in the function, we may have generated an AST_CONTROL_SRCCHANGE frame and put it in the list of frames we wish to return. This frame would be lost in such a case. The patch fixes this problem ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@254454 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2010-03-25 16:04:48 +00:00
AST_LIST_HEAD_NOLOCK(frame_list, ast_frame);
/* Forward Declarations */
static int ast_rtp_new(struct ast_rtp_instance *instance, struct ast_sched_context *sched, struct ast_sockaddr *addr, void *data);
static int ast_rtp_destroy(struct ast_rtp_instance *instance);
static int ast_rtp_dtmf_begin(struct ast_rtp_instance *instance, char digit);
static int ast_rtp_dtmf_end(struct ast_rtp_instance *instance, char digit);
static int ast_rtp_dtmf_end_with_duration(struct ast_rtp_instance *instance, char digit, unsigned int duration);
static int ast_rtp_dtmf_mode_set(struct ast_rtp_instance *instance, enum ast_rtp_dtmf_mode dtmf_mode);
static enum ast_rtp_dtmf_mode ast_rtp_dtmf_mode_get(struct ast_rtp_instance *instance);
static void ast_rtp_update_source(struct ast_rtp_instance *instance);
static void ast_rtp_change_source(struct ast_rtp_instance *instance);
static int ast_rtp_write(struct ast_rtp_instance *instance, struct ast_frame *frame);
static struct ast_frame *ast_rtp_read(struct ast_rtp_instance *instance, int rtcp);
static void ast_rtp_prop_set(struct ast_rtp_instance *instance, enum ast_rtp_property property, int value);
static int ast_rtp_fd(struct ast_rtp_instance *instance, int rtcp);
static void ast_rtp_remote_address_set(struct ast_rtp_instance *instance, struct ast_sockaddr *addr);
static int rtp_red_init(struct ast_rtp_instance *instance, int buffer_time, int *payloads, int generations);
static int rtp_red_buffer(struct ast_rtp_instance *instance, struct ast_frame *frame);
static int ast_rtp_local_bridge(struct ast_rtp_instance *instance0, struct ast_rtp_instance *instance1);
static int ast_rtp_get_stat(struct ast_rtp_instance *instance, struct ast_rtp_instance_stats *stats, enum ast_rtp_instance_stat stat);
static int ast_rtp_dtmf_compatible(struct ast_channel *chan0, struct ast_rtp_instance *instance0, struct ast_channel *chan1, struct ast_rtp_instance *instance1);
static void ast_rtp_stun_request(struct ast_rtp_instance *instance, struct ast_sockaddr *suggestion, const char *username);
static void ast_rtp_stop(struct ast_rtp_instance *instance);
static int ast_rtp_qos_set(struct ast_rtp_instance *instance, int tos, int cos, const char* desc);
static int ast_rtp_sendcng(struct ast_rtp_instance *instance, int level);
static unsigned int ast_rtp_get_ssrc(struct ast_rtp_instance *instance);
static const char *ast_rtp_get_cname(struct ast_rtp_instance *instance);
static void ast_rtp_set_remote_ssrc(struct ast_rtp_instance *instance, unsigned int ssrc);
static void ast_rtp_set_stream_num(struct ast_rtp_instance *instance, int stream_num);
static int ast_rtp_extension_enable(struct ast_rtp_instance *instance, enum ast_rtp_extension extension);
static int ast_rtp_bundle(struct ast_rtp_instance *child, struct ast_rtp_instance *parent);
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
static void update_reported_mes_stats(struct ast_rtp *rtp);
static void update_local_mes_stats(struct ast_rtp *rtp);
#if defined(HAVE_OPENSSL) && (OPENSSL_VERSION_NUMBER >= 0x10001000L) && !defined(OPENSSL_NO_SRTP)
static int ast_rtp_activate(struct ast_rtp_instance *instance);
static void dtls_srtp_start_timeout_timer(struct ast_rtp_instance *instance, struct ast_rtp *rtp, int rtcp);
static void dtls_srtp_stop_timeout_timer(struct ast_rtp_instance *instance, struct ast_rtp *rtp, int rtcp);
static int dtls_bio_write(BIO *bio, const char *buf, int len);
static long dtls_bio_ctrl(BIO *bio, int cmd, long arg1, void *arg2);
static int dtls_bio_new(BIO *bio);
static int dtls_bio_free(BIO *bio);
#ifndef HAVE_OPENSSL_BIO_METHOD
static BIO_METHOD dtls_bio_methods = {
.type = BIO_TYPE_BIO,
.name = "rtp write",
.bwrite = dtls_bio_write,
.ctrl = dtls_bio_ctrl,
.create = dtls_bio_new,
.destroy = dtls_bio_free,
};
#else
static BIO_METHOD *dtls_bio_methods;
#endif
#endif
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
static int __rtp_sendto(struct ast_rtp_instance *instance, void *buf, size_t size, int flags, struct ast_sockaddr *sa, int rtcp, int *via_ice, int use_srtp);
#ifdef HAVE_PJPROJECT
static void stunaddr_resolve_callback(const struct ast_dns_query *query);
static int store_stunaddr_resolved(const struct ast_dns_query *query);
#endif
#if defined(HAVE_OPENSSL) && (OPENSSL_VERSION_NUMBER >= 0x10001000L) && !defined(OPENSSL_NO_SRTP)
static int dtls_bio_new(BIO *bio)
{
#ifdef HAVE_OPENSSL_BIO_METHOD
BIO_set_init(bio, 1);
BIO_set_data(bio, NULL);
BIO_set_shutdown(bio, 0);
#else
bio->init = 1;
bio->ptr = NULL;
bio->flags = 0;
#endif
return 1;
}
static int dtls_bio_free(BIO *bio)
{
/* The pointer on the BIO is that of the RTP instance. It is not reference counted as the BIO
* lifetime is tied to the instance, and actions on the BIO are taken by the thread handling
* the RTP instance - not another thread.
*/
#ifdef HAVE_OPENSSL_BIO_METHOD
BIO_set_data(bio, NULL);
#else
bio->ptr = NULL;
#endif
return 1;
}
static int dtls_bio_write(BIO *bio, const char *buf, int len)
{
#ifdef HAVE_OPENSSL_BIO_METHOD
struct ast_rtp_instance *instance = BIO_get_data(bio);
#else
struct ast_rtp_instance *instance = bio->ptr;
#endif
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
int rtcp = 0;
struct ast_sockaddr remote_address = { {0, } };
int ice;
int bytes_sent;
/* OpenSSL can't tolerate a packet not being sent, so we always state that
* we sent the packet. If it isn't then retransmission will occur.
*/
if (rtp->rtcp && rtp->rtcp->dtls.write_bio == bio) {
rtcp = 1;
ast_sockaddr_copy(&remote_address, &rtp->rtcp->them);
} else {
ast_rtp_instance_get_remote_address(instance, &remote_address);
}
if (ast_sockaddr_isnull(&remote_address)) {
return len;
}
bytes_sent = __rtp_sendto(instance, (char *)buf, len, 0, &remote_address, rtcp, &ice, 0);
if (bytes_sent > 0 && ast_debug_dtls_packet_is_allowed) {
ast_debug(0, "(%p) DTLS - sent %s packet to %s%s (len %-6.6d)\n",
instance, rtcp ? "RTCP" : "RTP", ast_sockaddr_stringify(&remote_address),
ice ? " (via ICE)" : "", bytes_sent);
}
return len;
}
static long dtls_bio_ctrl(BIO *bio, int cmd, long arg1, void *arg2)
{
switch (cmd) {
case BIO_CTRL_FLUSH:
return 1;
case BIO_CTRL_DGRAM_QUERY_MTU:
return dtls_mtu;
case BIO_CTRL_WPENDING:
case BIO_CTRL_PENDING:
return 0L;
default:
return 0;
}
}
#endif
#ifdef HAVE_PJPROJECT
/*! \brief Helper function which clears the ICE host candidate mapping */
static void host_candidate_overrides_clear(void)
{
struct ast_ice_host_candidate *candidate;
AST_RWLIST_WRLOCK(&host_candidates);
AST_RWLIST_TRAVERSE_SAFE_BEGIN(&host_candidates, candidate, next) {
AST_RWLIST_REMOVE_CURRENT(next);
ast_free(candidate);
}
AST_RWLIST_TRAVERSE_SAFE_END;
AST_RWLIST_UNLOCK(&host_candidates);
}
/*! \brief Helper function which updates an ast_sockaddr with the candidate used for the component */
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
static void update_address_with_ice_candidate(pj_ice_sess *ice, enum ast_rtp_ice_component_type component,
struct ast_sockaddr *cand_address)
{
char address[PJ_INET6_ADDRSTRLEN];
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
if (component < 1 || !ice->comp[component - 1].valid_check) {
return;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ast_sockaddr_parse(cand_address,
pj_sockaddr_print(&ice->comp[component - 1].valid_check->rcand->addr, address,
sizeof(address), 0), 0);
ast_sockaddr_set_port(cand_address,
pj_sockaddr_get_port(&ice->comp[component - 1].valid_check->rcand->addr));
}
/*! \brief Destructor for locally created ICE candidates */
static void ast_rtp_ice_candidate_destroy(void *obj)
{
struct ast_rtp_engine_ice_candidate *candidate = obj;
if (candidate->foundation) {
ast_free(candidate->foundation);
}
if (candidate->transport) {
ast_free(candidate->transport);
}
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static void ast_rtp_ice_set_authentication(struct ast_rtp_instance *instance, const char *ufrag, const char *password)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
int ice_attrb_reset = 0;
if (!ast_strlen_zero(ufrag)) {
if (!ast_strlen_zero(rtp->remote_ufrag) && strcmp(ufrag, rtp->remote_ufrag)) {
ice_attrb_reset = 1;
}
ast_copy_string(rtp->remote_ufrag, ufrag, sizeof(rtp->remote_ufrag));
}
if (!ast_strlen_zero(password)) {
if (!ast_strlen_zero(rtp->remote_passwd) && strcmp(password, rtp->remote_passwd)) {
ice_attrb_reset = 1;
}
ast_copy_string(rtp->remote_passwd, password, sizeof(rtp->remote_passwd));
}
/* If the remote ufrag or passwd changed, local ufrag and passwd need to regenerate */
if (ice_attrb_reset) {
generate_random_string(rtp->local_ufrag, sizeof(rtp->local_ufrag));
generate_random_string(rtp->local_passwd, sizeof(rtp->local_passwd));
}
}
static int ice_candidate_cmp(void *obj, void *arg, int flags)
{
struct ast_rtp_engine_ice_candidate *candidate1 = obj, *candidate2 = arg;
if (strcmp(candidate1->foundation, candidate2->foundation) ||
candidate1->id != candidate2->id ||
candidate1->type != candidate2->type ||
ast_sockaddr_cmp(&candidate1->address, &candidate2->address)) {
return 0;
}
return CMP_MATCH | CMP_STOP;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static void ast_rtp_ice_add_remote_candidate(struct ast_rtp_instance *instance, const struct ast_rtp_engine_ice_candidate *candidate)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
struct ast_rtp_engine_ice_candidate *remote_candidate;
/* ICE sessions only support UDP candidates */
if (strcasecmp(candidate->transport, "udp")) {
return;
}
if (!rtp->ice_proposed_remote_candidates) {
rtp->ice_proposed_remote_candidates = ao2_container_alloc_list(
AO2_ALLOC_OPT_LOCK_MUTEX, 0, NULL, ice_candidate_cmp);
if (!rtp->ice_proposed_remote_candidates) {
return;
}
}
/* If this is going to exceed the maximum number of ICE candidates don't even add it */
if (ao2_container_count(rtp->ice_proposed_remote_candidates) == PJ_ICE_MAX_CAND) {
return;
}
if (!(remote_candidate = ao2_alloc(sizeof(*remote_candidate), ast_rtp_ice_candidate_destroy))) {
return;
}
remote_candidate->foundation = ast_strdup(candidate->foundation);
remote_candidate->id = candidate->id;
remote_candidate->transport = ast_strdup(candidate->transport);
remote_candidate->priority = candidate->priority;
ast_sockaddr_copy(&remote_candidate->address, &candidate->address);
ast_sockaddr_copy(&remote_candidate->relay_address, &candidate->relay_address);
remote_candidate->type = candidate->type;
ast_debug_ice(2, "(%p) ICE add remote candidate\n", instance);
ao2_link(rtp->ice_proposed_remote_candidates, remote_candidate);
ao2_ref(remote_candidate, -1);
}
AST_THREADSTORAGE(pj_thread_storage);
/*! \brief Function used to check if the calling thread is registered with pjlib. If it is not it will be registered. */
static void pj_thread_register_check(void)
{
pj_thread_desc *desc;
pj_thread_t *thread;
if (pj_thread_is_registered() == PJ_TRUE) {
return;
}
desc = ast_threadstorage_get(&pj_thread_storage, sizeof(pj_thread_desc));
if (!desc) {
ast_log(LOG_ERROR, "Could not get thread desc from thread-local storage. Expect awful things to occur\n");
return;
}
pj_bzero(*desc, sizeof(*desc));
if (pj_thread_register("Asterisk Thread", *desc, &thread) != PJ_SUCCESS) {
ast_log(LOG_ERROR, "Coudln't register thread with PJLIB.\n");
}
return;
}
static int ice_create(struct ast_rtp_instance *instance, struct ast_sockaddr *addr,
int port, int replace);
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static void ast_rtp_ice_stop(struct ast_rtp_instance *instance)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
struct ice_wrap *ice;
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ice = rtp->ice;
rtp->ice = NULL;
if (ice) {
/* Release the instance lock to avoid deadlock with PJPROJECT group lock */
ao2_unlock(instance);
ao2_ref(ice, -1);
ao2_lock(instance);
ast_debug_ice(2, "(%p) ICE stopped\n", instance);
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*!
* \brief ao2 ICE wrapper object destructor.
*
* \param vdoomed Object being destroyed.
*
* \note The associated struct ast_rtp_instance object must not
* be locked when unreffing the object. Otherwise we could
* deadlock trying to destroy the PJPROJECT ICE structure.
*/
static void ice_wrap_dtor(void *vdoomed)
{
struct ice_wrap *ice = vdoomed;
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
if (ice->real_ice) {
pj_thread_register_check();
pj_ice_sess_destroy(ice->real_ice);
}
}
static void ast2pj_rtp_ice_role(enum ast_rtp_ice_role ast_role, enum pj_ice_sess_role *pj_role)
{
switch (ast_role) {
case AST_RTP_ICE_ROLE_CONTROLLED:
*pj_role = PJ_ICE_SESS_ROLE_CONTROLLED;
break;
case AST_RTP_ICE_ROLE_CONTROLLING:
*pj_role = PJ_ICE_SESS_ROLE_CONTROLLING;
break;
}
}
static void pj2ast_rtp_ice_role(enum pj_ice_sess_role pj_role, enum ast_rtp_ice_role *ast_role)
{
switch (pj_role) {
case PJ_ICE_SESS_ROLE_CONTROLLED:
*ast_role = AST_RTP_ICE_ROLE_CONTROLLED;
return;
case PJ_ICE_SESS_ROLE_CONTROLLING:
*ast_role = AST_RTP_ICE_ROLE_CONTROLLING;
return;
case PJ_ICE_SESS_ROLE_UNKNOWN:
/* Don't change anything */
return;
default:
/* If we aren't explicitly handling something, it's a bug */
ast_assert(0);
return;
}
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static int ice_reset_session(struct ast_rtp_instance *instance)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
int res;
ast_debug_ice(3, "(%p) ICE resetting\n", instance);
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
if (!rtp->ice->real_ice->is_nominating && !rtp->ice->real_ice->is_complete) {
ast_debug_ice(3, " (%p) ICE nevermind, not ready for a reset\n", instance);
return 0;
}
ast_debug_ice(3, "(%p) ICE recreating ICE session %s (%d)\n",
instance, ast_sockaddr_stringify(&rtp->ice_original_rtp_addr), rtp->ice_port);
res = ice_create(instance, &rtp->ice_original_rtp_addr, rtp->ice_port, 1);
if (!res) {
/* Use the current expected role for the ICE session */
enum pj_ice_sess_role role = PJ_ICE_SESS_ROLE_UNKNOWN;
ast2pj_rtp_ice_role(rtp->role, &role);
pj_ice_sess_change_role(rtp->ice->real_ice, role);
}
/* If we only have one component now, and we previously set up TURN for RTCP,
* we need to destroy that TURN socket.
*/
if (rtp->ice_num_components == 1 && rtp->turn_rtcp) {
struct timeval wait = ast_tvadd(ast_tvnow(), ast_samp2tv(TURN_STATE_WAIT_TIME, 1000));
struct timespec ts = { .tv_sec = wait.tv_sec, .tv_nsec = wait.tv_usec * 1000, };
rtp->turn_state = PJ_TURN_STATE_NULL;
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/* Release the instance lock to avoid deadlock with PJPROJECT group lock */
ao2_unlock(instance);
pj_turn_sock_destroy(rtp->turn_rtcp);
ao2_lock(instance);
while (rtp->turn_state != PJ_TURN_STATE_DESTROYING) {
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ast_cond_timedwait(&rtp->cond, ao2_object_get_lockaddr(instance), &ts);
}
}
rtp->ice_media_started = 0;
return res;
}
static int ice_candidates_compare(struct ao2_container *left, struct ao2_container *right)
{
struct ao2_iterator i;
struct ast_rtp_engine_ice_candidate *right_candidate;
if (ao2_container_count(left) != ao2_container_count(right)) {
return -1;
}
i = ao2_iterator_init(right, 0);
while ((right_candidate = ao2_iterator_next(&i))) {
struct ast_rtp_engine_ice_candidate *left_candidate = ao2_find(left, right_candidate, OBJ_POINTER);
if (!left_candidate) {
ao2_ref(right_candidate, -1);
ao2_iterator_destroy(&i);
return -1;
}
ao2_ref(left_candidate, -1);
ao2_ref(right_candidate, -1);
}
ao2_iterator_destroy(&i);
return 0;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static void ast_rtp_ice_start(struct ast_rtp_instance *instance)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
pj_str_t ufrag = pj_str(rtp->remote_ufrag), passwd = pj_str(rtp->remote_passwd);
pj_ice_sess_cand candidates[PJ_ICE_MAX_CAND];
struct ao2_iterator i;
struct ast_rtp_engine_ice_candidate *candidate;
int cand_cnt = 0, has_rtp = 0, has_rtcp = 0;
if (!rtp->ice || !rtp->ice_proposed_remote_candidates) {
return;
}
/* Check for equivalence in the lists */
if (rtp->ice_active_remote_candidates &&
!ice_candidates_compare(rtp->ice_proposed_remote_candidates, rtp->ice_active_remote_candidates)) {
ast_debug_ice(2, "(%p) ICE proposed equals active candidates\n", instance);
ao2_cleanup(rtp->ice_proposed_remote_candidates);
rtp->ice_proposed_remote_candidates = NULL;
/* If this ICE session is being preserved then go back to the role it currently is */
pj2ast_rtp_ice_role(rtp->ice->real_ice->role, &rtp->role);
return;
}
/* Out with the old, in with the new */
ao2_cleanup(rtp->ice_active_remote_candidates);
rtp->ice_active_remote_candidates = rtp->ice_proposed_remote_candidates;
rtp->ice_proposed_remote_candidates = NULL;
ast_debug_ice(2, "(%p) ICE start\n", instance);
/* Reset the ICE session. Is this going to work? */
if (ice_reset_session(instance)) {
ast_log(LOG_NOTICE, "(%p) ICE failed to create replacement session\n", instance);
return;
}
pj_thread_register_check();
i = ao2_iterator_init(rtp->ice_active_remote_candidates, 0);
while ((candidate = ao2_iterator_next(&i)) && (cand_cnt < PJ_ICE_MAX_CAND)) {
pj_str_t address;
/* there needs to be at least one rtp and rtcp candidate in the list */
has_rtp |= candidate->id == AST_RTP_ICE_COMPONENT_RTP;
has_rtcp |= candidate->id == AST_RTP_ICE_COMPONENT_RTCP;
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
pj_strdup2(rtp->ice->real_ice->pool, &candidates[cand_cnt].foundation,
candidate->foundation);
candidates[cand_cnt].comp_id = candidate->id;
candidates[cand_cnt].prio = candidate->priority;
pj_sockaddr_parse(pj_AF_UNSPEC(), 0, pj_cstr(&address, ast_sockaddr_stringify(&candidate->address)), &candidates[cand_cnt].addr);
if (!ast_sockaddr_isnull(&candidate->relay_address)) {
pj_sockaddr_parse(pj_AF_UNSPEC(), 0, pj_cstr(&address, ast_sockaddr_stringify(&candidate->relay_address)), &candidates[cand_cnt].rel_addr);
}
if (candidate->type == AST_RTP_ICE_CANDIDATE_TYPE_HOST) {
candidates[cand_cnt].type = PJ_ICE_CAND_TYPE_HOST;
} else if (candidate->type == AST_RTP_ICE_CANDIDATE_TYPE_SRFLX) {
candidates[cand_cnt].type = PJ_ICE_CAND_TYPE_SRFLX;
} else if (candidate->type == AST_RTP_ICE_CANDIDATE_TYPE_RELAYED) {
candidates[cand_cnt].type = PJ_ICE_CAND_TYPE_RELAYED;
}
if (candidate->id == AST_RTP_ICE_COMPONENT_RTP && rtp->turn_rtp) {
ast_debug_ice(2, "(%p) ICE RTP candidate %s\n", instance, ast_sockaddr_stringify(&candidate->address));
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/* Release the instance lock to avoid deadlock with PJPROJECT group lock */
ao2_unlock(instance);
pj_turn_sock_set_perm(rtp->turn_rtp, 1, &candidates[cand_cnt].addr, 1);
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_lock(instance);
} else if (candidate->id == AST_RTP_ICE_COMPONENT_RTCP && rtp->turn_rtcp) {
ast_debug_ice(2, "(%p) ICE RTCP candidate %s\n", instance, ast_sockaddr_stringify(&candidate->address));
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/* Release the instance lock to avoid deadlock with PJPROJECT group lock */
ao2_unlock(instance);
pj_turn_sock_set_perm(rtp->turn_rtcp, 1, &candidates[cand_cnt].addr, 1);
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_lock(instance);
}
cand_cnt++;
ao2_ref(candidate, -1);
}
ao2_iterator_destroy(&i);
if (cand_cnt < ao2_container_count(rtp->ice_active_remote_candidates)) {
ast_log(LOG_WARNING, "(%p) ICE lost %d candidates. Consider increasing PJ_ICE_MAX_CAND in PJSIP\n",
instance, ao2_container_count(rtp->ice_active_remote_candidates) - cand_cnt);
}
if (!has_rtp) {
ast_log(LOG_WARNING, "(%p) ICE no RTP candidates; skipping checklist\n", instance);
}
/* If we're only dealing with one ICE component, then we don't care about the lack of RTCP candidates */
if (!has_rtcp && rtp->ice_num_components > 1) {
ast_log(LOG_WARNING, "(%p) ICE no RTCP candidates; skipping checklist\n", instance);
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
if (rtp->ice && has_rtp && (has_rtcp || rtp->ice_num_components == 1)) {
pj_status_t res;
char reason[80];
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
struct ice_wrap *ice;
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/* Release the instance lock to avoid deadlock with PJPROJECT group lock */
ice = rtp->ice;
ao2_ref(ice, +1);
ao2_unlock(instance);
res = pj_ice_sess_create_check_list(ice->real_ice, &ufrag, &passwd, cand_cnt, &candidates[0]);
if (res == PJ_SUCCESS) {
ast_debug_ice(2, "(%p) ICE successfully created checklist\n", instance);
ast_test_suite_event_notify("ICECHECKLISTCREATE", "Result: SUCCESS");
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
pj_ice_sess_start_check(ice->real_ice);
pj_timer_heap_poll(timer_heap, NULL);
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_ref(ice, -1);
ao2_lock(instance);
rtp->strict_rtp_state = STRICT_RTP_OPEN;
return;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_ref(ice, -1);
ao2_lock(instance);
pj_strerror(res, reason, sizeof(reason));
ast_log(LOG_WARNING, "(%p) ICE failed to create session check list: %s\n", instance, reason);
}
ast_test_suite_event_notify("ICECHECKLISTCREATE", "Result: FAILURE");
/* even though create check list failed don't stop ice as
it might still work */
/* however we do need to reset remote candidates since
this function may be re-entered */
ao2_ref(rtp->ice_active_remote_candidates, -1);
rtp->ice_active_remote_candidates = NULL;
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
if (rtp->ice) {
rtp->ice->real_ice->rcand_cnt = rtp->ice->real_ice->clist.count = 0;
}
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static const char *ast_rtp_ice_get_ufrag(struct ast_rtp_instance *instance)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
return rtp->local_ufrag;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static const char *ast_rtp_ice_get_password(struct ast_rtp_instance *instance)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
return rtp->local_passwd;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static struct ao2_container *ast_rtp_ice_get_local_candidates(struct ast_rtp_instance *instance)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
if (rtp->ice_local_candidates) {
ao2_ref(rtp->ice_local_candidates, +1);
}
return rtp->ice_local_candidates;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static void ast_rtp_ice_lite(struct ast_rtp_instance *instance)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
if (!rtp->ice) {
return;
}
pj_thread_register_check();
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
pj_ice_sess_change_role(rtp->ice->real_ice, PJ_ICE_SESS_ROLE_CONTROLLING);
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static void ast_rtp_ice_set_role(struct ast_rtp_instance *instance, enum ast_rtp_ice_role role)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
if (!rtp->ice) {
ast_debug_ice(3, "(%p) ICE set role failed; no ice instance\n", instance);
return;
}
rtp->role = role;
if (!rtp->ice->real_ice->is_nominating && !rtp->ice->real_ice->is_complete) {
pj_thread_register_check();
ast_debug_ice(2, "(%p) ICE set role to %s\n",
instance, role == AST_RTP_ICE_ROLE_CONTROLLED ? "CONTROLLED" : "CONTROLLING");
pj_ice_sess_change_role(rtp->ice->real_ice, role == AST_RTP_ICE_ROLE_CONTROLLED ?
PJ_ICE_SESS_ROLE_CONTROLLED : PJ_ICE_SESS_ROLE_CONTROLLING);
} else {
ast_debug_ice(2, "(%p) ICE not setting role because state is %s\n",
instance, rtp->ice->real_ice->is_nominating ? "nominating" : "complete");
}
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static void ast_rtp_ice_add_cand(struct ast_rtp_instance *instance, struct ast_rtp *rtp,
unsigned comp_id, unsigned transport_id, pj_ice_cand_type type, pj_uint16_t local_pref,
const pj_sockaddr_t *addr, const pj_sockaddr_t *base_addr, const pj_sockaddr_t *rel_addr,
int addr_len)
{
pj_str_t foundation;
struct ast_rtp_engine_ice_candidate *candidate, *existing;
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
struct ice_wrap *ice;
char address[PJ_INET6_ADDRSTRLEN];
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
pj_status_t status;
if (!rtp->ice) {
return;
}
pj_thread_register_check();
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
pj_ice_calc_foundation(rtp->ice->real_ice->pool, &foundation, type, addr);
if (!rtp->ice_local_candidates) {
rtp->ice_local_candidates = ao2_container_alloc_list(AO2_ALLOC_OPT_LOCK_MUTEX, 0,
NULL, ice_candidate_cmp);
if (!rtp->ice_local_candidates) {
return;
}
}
if (!(candidate = ao2_alloc(sizeof(*candidate), ast_rtp_ice_candidate_destroy))) {
return;
}
candidate->foundation = ast_strndup(pj_strbuf(&foundation), pj_strlen(&foundation));
candidate->id = comp_id;
candidate->transport = ast_strdup("UDP");
ast_sockaddr_parse(&candidate->address, pj_sockaddr_print(addr, address, sizeof(address), 0), 0);
ast_sockaddr_set_port(&candidate->address, pj_sockaddr_get_port(addr));
if (rel_addr) {
ast_sockaddr_parse(&candidate->relay_address, pj_sockaddr_print(rel_addr, address, sizeof(address), 0), 0);
ast_sockaddr_set_port(&candidate->relay_address, pj_sockaddr_get_port(rel_addr));
}
if (type == PJ_ICE_CAND_TYPE_HOST) {
candidate->type = AST_RTP_ICE_CANDIDATE_TYPE_HOST;
} else if (type == PJ_ICE_CAND_TYPE_SRFLX) {
candidate->type = AST_RTP_ICE_CANDIDATE_TYPE_SRFLX;
} else if (type == PJ_ICE_CAND_TYPE_RELAYED) {
candidate->type = AST_RTP_ICE_CANDIDATE_TYPE_RELAYED;
}
if ((existing = ao2_find(rtp->ice_local_candidates, candidate, OBJ_POINTER))) {
ao2_ref(existing, -1);
ao2_ref(candidate, -1);
return;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/* Release the instance lock to avoid deadlock with PJPROJECT group lock */
ice = rtp->ice;
ao2_ref(ice, +1);
ao2_unlock(instance);
status = pj_ice_sess_add_cand(ice->real_ice, comp_id, transport_id, type, local_pref,
&foundation, addr, base_addr, rel_addr, addr_len, NULL);
ao2_ref(ice, -1);
ao2_lock(instance);
if (!rtp->ice || status != PJ_SUCCESS) {
ast_debug_ice(2, "(%p) ICE unable to add candidate: %s, %d\n", instance, ast_sockaddr_stringify(
&candidate->address), candidate->priority);
ao2_ref(candidate, -1);
return;
}
/* By placing the candidate into the ICE session it will have produced the priority, so update the local candidate with it */
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
candidate->priority = rtp->ice->real_ice->lcand[rtp->ice->real_ice->lcand_cnt - 1].prio;
ast_debug_ice(2, "(%p) ICE add candidate: %s, %d\n", instance, ast_sockaddr_stringify(
&candidate->address), candidate->priority);
ao2_link(rtp->ice_local_candidates, candidate);
ao2_ref(candidate, -1);
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/* PJPROJECT TURN callback */
static void ast_rtp_on_turn_rx_rtp_data(pj_turn_sock *turn_sock, void *pkt, unsigned pkt_len, const pj_sockaddr_t *peer_addr, unsigned addr_len)
{
struct ast_rtp_instance *instance = pj_turn_sock_get_user_data(turn_sock);
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
struct ice_wrap *ice;
pj_status_t status;
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_lock(instance);
ice = ao2_bump(rtp->ice);
ao2_unlock(instance);
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
if (ice) {
status = pj_ice_sess_on_rx_pkt(ice->real_ice, AST_RTP_ICE_COMPONENT_RTP,
TRANSPORT_TURN_RTP, pkt, pkt_len, peer_addr, addr_len);
ao2_ref(ice, -1);
if (status != PJ_SUCCESS) {
char buf[100];
pj_strerror(status, buf, sizeof(buf));
ast_log(LOG_WARNING, "(%p) ICE PJ Rx error status code: %d '%s'.\n",
instance, (int)status, buf);
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
return;
}
if (!rtp->rtp_passthrough) {
return;
}
rtp->rtp_passthrough = 0;
}
ast_sendto(rtp->s, pkt, pkt_len, 0, &rtp->rtp_loop);
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/* PJPROJECT TURN callback */
static void ast_rtp_on_turn_rtp_state(pj_turn_sock *turn_sock, pj_turn_state_t old_state, pj_turn_state_t new_state)
{
struct ast_rtp_instance *instance = pj_turn_sock_get_user_data(turn_sock);
struct ast_rtp *rtp;
/* If this is a leftover from an already notified RTP instance just ignore the state change */
if (!instance) {
return;
}
rtp = ast_rtp_instance_get_data(instance);
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_lock(instance);
/* We store the new state so the other thread can actually handle it */
rtp->turn_state = new_state;
ast_cond_signal(&rtp->cond);
if (new_state == PJ_TURN_STATE_DESTROYING) {
pj_turn_sock_set_user_data(rtp->turn_rtp, NULL);
rtp->turn_rtp = NULL;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_unlock(instance);
}
/* RTP TURN Socket interface declaration */
static pj_turn_sock_cb ast_rtp_turn_rtp_sock_cb = {
.on_rx_data = ast_rtp_on_turn_rx_rtp_data,
.on_state = ast_rtp_on_turn_rtp_state,
};
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/* PJPROJECT TURN callback */
static void ast_rtp_on_turn_rx_rtcp_data(pj_turn_sock *turn_sock, void *pkt, unsigned pkt_len, const pj_sockaddr_t *peer_addr, unsigned addr_len)
{
struct ast_rtp_instance *instance = pj_turn_sock_get_user_data(turn_sock);
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
struct ice_wrap *ice;
pj_status_t status;
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_lock(instance);
ice = ao2_bump(rtp->ice);
ao2_unlock(instance);
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
if (ice) {
status = pj_ice_sess_on_rx_pkt(ice->real_ice, AST_RTP_ICE_COMPONENT_RTCP,
TRANSPORT_TURN_RTCP, pkt, pkt_len, peer_addr, addr_len);
ao2_ref(ice, -1);
if (status != PJ_SUCCESS) {
char buf[100];
pj_strerror(status, buf, sizeof(buf));
ast_log(LOG_WARNING, "PJ ICE Rx error status code: %d '%s'.\n",
(int)status, buf);
return;
}
if (!rtp->rtcp_passthrough) {
return;
}
rtp->rtcp_passthrough = 0;
}
ast_sendto(rtp->rtcp->s, pkt, pkt_len, 0, &rtp->rtcp_loop);
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/* PJPROJECT TURN callback */
static void ast_rtp_on_turn_rtcp_state(pj_turn_sock *turn_sock, pj_turn_state_t old_state, pj_turn_state_t new_state)
{
struct ast_rtp_instance *instance = pj_turn_sock_get_user_data(turn_sock);
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
struct ast_rtp *rtp;
/* If this is a leftover from an already destroyed RTP instance just ignore the state change */
if (!instance) {
return;
}
rtp = ast_rtp_instance_get_data(instance);
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_lock(instance);
/* We store the new state so the other thread can actually handle it */
rtp->turn_state = new_state;
ast_cond_signal(&rtp->cond);
if (new_state == PJ_TURN_STATE_DESTROYING) {
pj_turn_sock_set_user_data(rtp->turn_rtcp, NULL);
rtp->turn_rtcp = NULL;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_unlock(instance);
}
/* RTCP TURN Socket interface declaration */
static pj_turn_sock_cb ast_rtp_turn_rtcp_sock_cb = {
.on_rx_data = ast_rtp_on_turn_rx_rtcp_data,
.on_state = ast_rtp_on_turn_rtcp_state,
};
/*! \brief Worker thread for ioqueue and timerheap */
static int ioqueue_worker_thread(void *data)
{
struct ast_rtp_ioqueue_thread *ioqueue = data;
while (!ioqueue->terminate) {
const pj_time_val delay = {0, 10};
pj_ioqueue_poll(ioqueue->ioqueue, &delay);
pj_timer_heap_poll(ioqueue->timerheap, NULL);
}
return 0;
}
/*! \brief Destroyer for ioqueue thread */
static void rtp_ioqueue_thread_destroy(struct ast_rtp_ioqueue_thread *ioqueue)
{
if (ioqueue->thread) {
ioqueue->terminate = 1;
pj_thread_join(ioqueue->thread);
pj_thread_destroy(ioqueue->thread);
}
if (ioqueue->pool) {
/* This mimics the behavior of pj_pool_safe_release
* which was introduced in pjproject 2.6.
*/
pj_pool_t *temp_pool = ioqueue->pool;
ioqueue->pool = NULL;
pj_pool_release(temp_pool);
}
ast_free(ioqueue);
}
/*! \brief Removal function for ioqueue thread, determines if it should be terminated and destroyed */
static void rtp_ioqueue_thread_remove(struct ast_rtp_ioqueue_thread *ioqueue)
{
int destroy = 0;
/* If nothing is using this ioqueue thread destroy it */
AST_LIST_LOCK(&ioqueues);
if ((ioqueue->count -= 2) == 0) {
destroy = 1;
AST_LIST_REMOVE(&ioqueues, ioqueue, next);
}
AST_LIST_UNLOCK(&ioqueues);
if (!destroy) {
return;
}
rtp_ioqueue_thread_destroy(ioqueue);
}
/*! \brief Finder and allocator for an ioqueue thread */
static struct ast_rtp_ioqueue_thread *rtp_ioqueue_thread_get_or_create(void)
{
struct ast_rtp_ioqueue_thread *ioqueue;
pj_lock_t *lock;
AST_LIST_LOCK(&ioqueues);
/* See if an ioqueue thread exists that can handle more */
AST_LIST_TRAVERSE(&ioqueues, ioqueue, next) {
if ((ioqueue->count + 2) < PJ_IOQUEUE_MAX_HANDLES) {
break;
}
}
/* If we found one bump it up and return it */
if (ioqueue) {
ioqueue->count += 2;
goto end;
}
ioqueue = ast_calloc(1, sizeof(*ioqueue));
if (!ioqueue) {
goto end;
}
ioqueue->pool = pj_pool_create(&cachingpool.factory, "rtp", 512, 512, NULL);
/* We use a timer on the ioqueue thread for TURN so that two threads aren't operating
* on a session at the same time
*/
if (pj_timer_heap_create(ioqueue->pool, 4, &ioqueue->timerheap) != PJ_SUCCESS) {
goto fatal;
}
if (pj_lock_create_recursive_mutex(ioqueue->pool, "rtp%p", &lock) != PJ_SUCCESS) {
goto fatal;
}
pj_timer_heap_set_lock(ioqueue->timerheap, lock, PJ_TRUE);
if (pj_ioqueue_create(ioqueue->pool, PJ_IOQUEUE_MAX_HANDLES, &ioqueue->ioqueue) != PJ_SUCCESS) {
goto fatal;
}
if (pj_thread_create(ioqueue->pool, "ice", &ioqueue_worker_thread, ioqueue, 0, 0, &ioqueue->thread) != PJ_SUCCESS) {
goto fatal;
}
AST_LIST_INSERT_HEAD(&ioqueues, ioqueue, next);
/* Since this is being returned to an active session the count always starts at 2 */
ioqueue->count = 2;
goto end;
fatal:
rtp_ioqueue_thread_destroy(ioqueue);
ioqueue = NULL;
end:
AST_LIST_UNLOCK(&ioqueues);
return ioqueue;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static void ast_rtp_ice_turn_request(struct ast_rtp_instance *instance, enum ast_rtp_ice_component_type component,
enum ast_transport transport, const char *server, unsigned int port, const char *username, const char *password)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
pj_turn_sock **turn_sock;
const pj_turn_sock_cb *turn_cb;
pj_turn_tp_type conn_type;
int conn_transport;
pj_stun_auth_cred cred = { 0, };
pj_str_t turn_addr;
struct ast_sockaddr addr = { { 0, } };
pj_stun_config stun_config;
struct timeval wait = ast_tvadd(ast_tvnow(), ast_samp2tv(TURN_STATE_WAIT_TIME, 1000));
struct timespec ts = { .tv_sec = wait.tv_sec, .tv_nsec = wait.tv_usec * 1000, };
pj_turn_session_info info;
struct ast_sockaddr local, loop;
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
pj_status_t status;
pj_turn_sock_cfg turn_sock_cfg;
struct ice_wrap *ice;
ast_rtp_instance_get_local_address(instance, &local);
if (ast_sockaddr_is_ipv4(&local)) {
ast_sockaddr_parse(&loop, "127.0.0.1", PARSE_PORT_FORBID);
} else {
ast_sockaddr_parse(&loop, "::1", PARSE_PORT_FORBID);
}
/* Determine what component we are requesting a TURN session for */
if (component == AST_RTP_ICE_COMPONENT_RTP) {
turn_sock = &rtp->turn_rtp;
turn_cb = &ast_rtp_turn_rtp_sock_cb;
conn_transport = TRANSPORT_TURN_RTP;
ast_sockaddr_set_port(&loop, ast_sockaddr_port(&local));
} else if (component == AST_RTP_ICE_COMPONENT_RTCP) {
turn_sock = &rtp->turn_rtcp;
turn_cb = &ast_rtp_turn_rtcp_sock_cb;
conn_transport = TRANSPORT_TURN_RTCP;
ast_sockaddr_set_port(&loop, ast_sockaddr_port(&rtp->rtcp->us));
} else {
return;
}
if (transport == AST_TRANSPORT_UDP) {
conn_type = PJ_TURN_TP_UDP;
} else if (transport == AST_TRANSPORT_TCP) {
conn_type = PJ_TURN_TP_TCP;
} else {
ast_assert(0);
return;
}
ast_sockaddr_parse(&addr, server, PARSE_PORT_FORBID);
if (*turn_sock) {
rtp->turn_state = PJ_TURN_STATE_NULL;
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/* Release the instance lock to avoid deadlock with PJPROJECT group lock */
ao2_unlock(instance);
pj_turn_sock_destroy(*turn_sock);
ao2_lock(instance);
while (rtp->turn_state != PJ_TURN_STATE_DESTROYING) {
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ast_cond_timedwait(&rtp->cond, ao2_object_get_lockaddr(instance), &ts);
}
}
if (component == AST_RTP_ICE_COMPONENT_RTP && !rtp->ioqueue) {
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*
* We cannot hold the instance lock because we could wait
* for the ioqueue thread to die and we might deadlock as
* a result.
*/
ao2_unlock(instance);
rtp->ioqueue = rtp_ioqueue_thread_get_or_create();
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_lock(instance);
if (!rtp->ioqueue) {
return;
}
}
pj_stun_config_init(&stun_config, &cachingpool.factory, 0, rtp->ioqueue->ioqueue, rtp->ioqueue->timerheap);
if (!stun_software_attribute) {
stun_config.software_name = pj_str(NULL);
}
/* Use ICE session group lock for TURN session to avoid deadlock */
pj_turn_sock_cfg_default(&turn_sock_cfg);
ice = rtp->ice;
if (ice) {
turn_sock_cfg.grp_lock = ice->real_ice->grp_lock;
ao2_ref(ice, +1);
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/* Release the instance lock to avoid deadlock with PJPROJECT group lock */
ao2_unlock(instance);
status = pj_turn_sock_create(&stun_config,
ast_sockaddr_is_ipv4(&addr) ? pj_AF_INET() : pj_AF_INET6(), conn_type,
turn_cb, &turn_sock_cfg, instance, turn_sock);
ao2_cleanup(ice);
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
if (status != PJ_SUCCESS) {
ast_log(LOG_WARNING, "(%p) Could not create a TURN client socket\n", instance);
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_lock(instance);
return;
}
cred.type = PJ_STUN_AUTH_CRED_STATIC;
pj_strset2(&cred.data.static_cred.username, (char*)username);
cred.data.static_cred.data_type = PJ_STUN_PASSWD_PLAIN;
pj_strset2(&cred.data.static_cred.data, (char*)password);
pj_turn_sock_alloc(*turn_sock, pj_cstr(&turn_addr, server), port, NULL, &cred, NULL);
ast_debug_ice(2, "(%p) ICE request TURN %s %s candidate\n", instance,
transport == AST_TRANSPORT_UDP ? "UDP" : "TCP",
component == AST_RTP_ICE_COMPONENT_RTP ? "RTP" : "RTCP");
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_lock(instance);
/*
* Because the TURN socket is asynchronous and we are synchronous we need to
* wait until it is done
*/
while (rtp->turn_state < PJ_TURN_STATE_READY) {
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ast_cond_timedwait(&rtp->cond, ao2_object_get_lockaddr(instance), &ts);
}
/* If a TURN session was allocated add it as a candidate */
if (rtp->turn_state != PJ_TURN_STATE_READY) {
return;
}
pj_turn_sock_get_info(*turn_sock, &info);
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ast_rtp_ice_add_cand(instance, rtp, component, conn_transport,
PJ_ICE_CAND_TYPE_RELAYED, 65535, &info.relay_addr, &info.relay_addr,
&info.mapped_addr, pj_sockaddr_get_len(&info.relay_addr));
if (component == AST_RTP_ICE_COMPONENT_RTP) {
ast_sockaddr_copy(&rtp->rtp_loop, &loop);
} else if (component == AST_RTP_ICE_COMPONENT_RTCP) {
ast_sockaddr_copy(&rtp->rtcp_loop, &loop);
}
}
static char *generate_random_string(char *buf, size_t size)
{
long val[4];
int x;
for (x=0; x<4; x++) {
val[x] = ast_random();
}
snprintf(buf, size, "%08lx%08lx%08lx%08lx", (long unsigned)val[0], (long unsigned)val[1], (long unsigned)val[2], (long unsigned)val[3]);
return buf;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static void ast_rtp_ice_change_components(struct ast_rtp_instance *instance, int num_components)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
/* Don't do anything if ICE is unsupported or if we're not changing the
* number of components
*/
if (!icesupport || !rtp->ice || rtp->ice_num_components == num_components) {
return;
}
ast_debug_ice(2, "(%p) ICE change number of components %u -> %u\n", instance,
rtp->ice_num_components, num_components);
rtp->ice_num_components = num_components;
ice_reset_session(instance);
}
/* ICE RTP Engine interface declaration */
static struct ast_rtp_engine_ice ast_rtp_ice = {
.set_authentication = ast_rtp_ice_set_authentication,
.add_remote_candidate = ast_rtp_ice_add_remote_candidate,
.start = ast_rtp_ice_start,
.stop = ast_rtp_ice_stop,
.get_ufrag = ast_rtp_ice_get_ufrag,
.get_password = ast_rtp_ice_get_password,
.get_local_candidates = ast_rtp_ice_get_local_candidates,
.ice_lite = ast_rtp_ice_lite,
.set_role = ast_rtp_ice_set_role,
.turn_request = ast_rtp_ice_turn_request,
.change_components = ast_rtp_ice_change_components,
};
#endif
#if defined(HAVE_OPENSSL) && (OPENSSL_VERSION_NUMBER >= 0x10001000L) && !defined(OPENSSL_NO_SRTP)
static int dtls_verify_callback(int preverify_ok, X509_STORE_CTX *ctx)
{
/* We don't want to actually verify the certificate so just accept what they have provided */
return 1;
}
static int dtls_details_initialize(struct dtls_details *dtls, SSL_CTX *ssl_ctx,
enum ast_rtp_dtls_setup setup, struct ast_rtp_instance *instance)
{
dtls->dtls_setup = setup;
if (!(dtls->ssl = SSL_new(ssl_ctx))) {
ast_log(LOG_ERROR, "Failed to allocate memory for SSL\n");
goto error;
}
if (!(dtls->read_bio = BIO_new(BIO_s_mem()))) {
ast_log(LOG_ERROR, "Failed to allocate memory for inbound SSL traffic\n");
goto error;
}
BIO_set_mem_eof_return(dtls->read_bio, -1);
#ifdef HAVE_OPENSSL_BIO_METHOD
if (!(dtls->write_bio = BIO_new(dtls_bio_methods))) {
ast_log(LOG_ERROR, "Failed to allocate memory for outbound SSL traffic\n");
goto error;
}
BIO_set_data(dtls->write_bio, instance);
#else
if (!(dtls->write_bio = BIO_new(&dtls_bio_methods))) {
ast_log(LOG_ERROR, "Failed to allocate memory for outbound SSL traffic\n");
goto error;
}
dtls->write_bio->ptr = instance;
#endif
SSL_set_bio(dtls->ssl, dtls->read_bio, dtls->write_bio);
if (dtls->dtls_setup == AST_RTP_DTLS_SETUP_PASSIVE) {
SSL_set_accept_state(dtls->ssl);
} else {
SSL_set_connect_state(dtls->ssl);
}
dtls->connection = AST_RTP_DTLS_CONNECTION_NEW;
return 0;
error:
if (dtls->read_bio) {
BIO_free(dtls->read_bio);
dtls->read_bio = NULL;
}
if (dtls->write_bio) {
BIO_free(dtls->write_bio);
dtls->write_bio = NULL;
}
if (dtls->ssl) {
SSL_free(dtls->ssl);
dtls->ssl = NULL;
}
return -1;
}
static int dtls_setup_rtcp(struct ast_rtp_instance *instance)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
if (!rtp->ssl_ctx || !rtp->rtcp) {
return 0;
}
ast_debug_dtls(3, "(%p) DTLS RTCP setup\n", instance);
return dtls_details_initialize(&rtp->rtcp->dtls, rtp->ssl_ctx, rtp->dtls.dtls_setup, instance);
}
static const SSL_METHOD *get_dtls_method(void)
{
#if OPENSSL_VERSION_NUMBER < 0x10002000L
return DTLSv1_method();
#else
return DTLS_method();
#endif
}
struct dtls_cert_info {
EVP_PKEY *private_key;
X509 *certificate;
};
static int apply_dh_params(SSL_CTX *ctx, BIO *bio)
{
int res = 0;
#if OPENSSL_VERSION_NUMBER >= 0x30000000L
EVP_PKEY *dhpkey = PEM_read_bio_Parameters(bio, NULL);
if (dhpkey && EVP_PKEY_is_a(dhpkey, "DH")) {
res = SSL_CTX_set0_tmp_dh_pkey(ctx, dhpkey);
}
if (!res) {
/* A successful call to SSL_CTX_set0_tmp_dh_pkey() means
that we lost ownership of dhpkey and should not free
it ourselves */
EVP_PKEY_free(dhpkey);
}
#else
DH *dh = PEM_read_bio_DHparams(bio, NULL, NULL, NULL);
if (dh) {
res = SSL_CTX_set_tmp_dh(ctx, dh);
}
DH_free(dh);
#endif
return res;
}
static void configure_dhparams(const struct ast_rtp *rtp, const struct ast_rtp_dtls_cfg *dtls_cfg)
{
#if !defined(OPENSSL_NO_ECDH) && (OPENSSL_VERSION_NUMBER >= 0x10000000L) && (OPENSSL_VERSION_NUMBER < 0x10100000L)
EC_KEY *ecdh;
#endif
#ifndef OPENSSL_NO_DH
if (!ast_strlen_zero(dtls_cfg->pvtfile)) {
BIO *bio = BIO_new_file(dtls_cfg->pvtfile, "r");
if (bio) {
if (apply_dh_params(rtp->ssl_ctx, bio)) {
long options = SSL_OP_CIPHER_SERVER_PREFERENCE |
SSL_OP_SINGLE_DH_USE | SSL_OP_SINGLE_ECDH_USE;
options = SSL_CTX_set_options(rtp->ssl_ctx, options);
ast_verb(2, "DTLS DH initialized, PFS enabled\n");
}
BIO_free(bio);
}
}
#endif /* !OPENSSL_NO_DH */
#if !defined(OPENSSL_NO_ECDH) && (OPENSSL_VERSION_NUMBER >= 0x10000000L) && (OPENSSL_VERSION_NUMBER < 0x10100000L)
/* enables AES-128 ciphers, to get AES-256 use NID_secp384r1 */
ecdh = EC_KEY_new_by_curve_name(NID_X9_62_prime256v1);
if (ecdh) {
if (SSL_CTX_set_tmp_ecdh(rtp->ssl_ctx, ecdh)) {
#ifndef SSL_CTRL_SET_ECDH_AUTO
#define SSL_CTRL_SET_ECDH_AUTO 94
#endif
/* SSL_CTX_set_ecdh_auto(rtp->ssl_ctx, on); requires OpenSSL 1.0.2 which wraps: */
if (SSL_CTX_ctrl(rtp->ssl_ctx, SSL_CTRL_SET_ECDH_AUTO, 1, NULL)) {
ast_verb(2, "DTLS ECDH initialized (automatic), faster PFS enabled\n");
} else {
ast_verb(2, "DTLS ECDH initialized (secp256r1), faster PFS enabled\n");
}
}
EC_KEY_free(ecdh);
}
#endif /* !OPENSSL_NO_ECDH */
}
#if !defined(OPENSSL_NO_ECDH) && (OPENSSL_VERSION_NUMBER >= 0x10000000L)
static int create_ephemeral_ec_keypair(EVP_PKEY **keypair)
{
#if OPENSSL_VERSION_NUMBER >= 0x30000000L
*keypair = EVP_EC_gen(SN_X9_62_prime256v1);
return *keypair ? 0 : -1;
#else
EC_KEY *eckey = NULL;
EC_GROUP *group = NULL;
group = EC_GROUP_new_by_curve_name(NID_X9_62_prime256v1);
if (!group) {
goto error;
}
EC_GROUP_set_asn1_flag(group, OPENSSL_EC_NAMED_CURVE);
EC_GROUP_set_point_conversion_form(group, POINT_CONVERSION_UNCOMPRESSED);
eckey = EC_KEY_new();
if (!eckey) {
goto error;
}
if (!EC_KEY_set_group(eckey, group)) {
goto error;
}
if (!EC_KEY_generate_key(eckey)) {
goto error;
}
*keypair = EVP_PKEY_new();
if (!*keypair) {
goto error;
}
EVP_PKEY_assign_EC_KEY(*keypair, eckey);
EC_GROUP_free(group);
return 0;
error:
EC_KEY_free(eckey);
EC_GROUP_free(group);
return -1;
#endif
}
/* From OpenSSL's x509 command */
#define SERIAL_RAND_BITS 159
static int create_ephemeral_certificate(EVP_PKEY *keypair, X509 **certificate)
{
X509 *cert = NULL;
BIGNUM *serial = NULL;
X509_NAME *name = NULL;
cert = X509_new();
if (!cert) {
goto error;
}
if (!X509_set_version(cert, 2)) {
goto error;
}
/* Set the public key */
X509_set_pubkey(cert, keypair);
/* Generate a random serial number */
if (!(serial = BN_new())
|| !BN_rand(serial, SERIAL_RAND_BITS, -1, 0)
|| !BN_to_ASN1_INTEGER(serial, X509_get_serialNumber(cert))) {
BN_free(serial);
goto error;
}
BN_free(serial);
/*
* Validity period - Current Chrome & Firefox make it 31 days starting
* with yesterday at the current time, so we will do the same.
*/
#if OPENSSL_VERSION_NUMBER < 0x10100000L
if (!X509_time_adj_ex(X509_get_notBefore(cert), -1, 0, NULL)
|| !X509_time_adj_ex(X509_get_notAfter(cert), 30, 0, NULL)) {
goto error;
}
#else
if (!X509_time_adj_ex(X509_getm_notBefore(cert), -1, 0, NULL)
|| !X509_time_adj_ex(X509_getm_notAfter(cert), 30, 0, NULL)) {
goto error;
}
#endif
/* Set the name and issuer */
if (!(name = X509_get_subject_name(cert))
|| !X509_NAME_add_entry_by_NID(name, NID_commonName, MBSTRING_ASC,
(unsigned char *) "asterisk", -1, -1, 0)
|| !X509_set_issuer_name(cert, name)) {
goto error;
}
/* Sign it */
if (!X509_sign(cert, keypair, EVP_sha256())) {
goto error;
}
*certificate = cert;
return 0;
error:
X509_free(cert);
return -1;
}
static int create_certificate_ephemeral(struct ast_rtp_instance *instance,
const struct ast_rtp_dtls_cfg *dtls_cfg,
struct dtls_cert_info *cert_info)
{
/* Make sure these are initialized */
cert_info->private_key = NULL;
cert_info->certificate = NULL;
if (create_ephemeral_ec_keypair(&cert_info->private_key)) {
ast_log(LOG_ERROR, "Failed to create ephemeral ECDSA keypair\n");
goto error;
}
if (create_ephemeral_certificate(cert_info->private_key, &cert_info->certificate)) {
ast_log(LOG_ERROR, "Failed to create ephemeral X509 certificate\n");
goto error;
}
return 0;
error:
X509_free(cert_info->certificate);
EVP_PKEY_free(cert_info->private_key);
return -1;
}
#else
static int create_certificate_ephemeral(struct ast_rtp_instance *instance,
const struct ast_rtp_dtls_cfg *dtls_cfg,
struct dtls_cert_info *cert_info)
{
ast_log(LOG_ERROR, "Your version of OpenSSL does not support ECDSA keys\n");
return -1;
}
#endif /* !OPENSSL_NO_ECDH */
static int create_certificate_from_file(struct ast_rtp_instance *instance,
const struct ast_rtp_dtls_cfg *dtls_cfg,
struct dtls_cert_info *cert_info)
{
FILE *fp;
BIO *certbio = NULL;
EVP_PKEY *private_key = NULL;
X509 *cert = NULL;
char *private_key_file = ast_strlen_zero(dtls_cfg->pvtfile) ? dtls_cfg->certfile : dtls_cfg->pvtfile;
fp = fopen(private_key_file, "r");
if (!fp) {
ast_log(LOG_ERROR, "Failed to read private key from file '%s': %s\n", private_key_file, strerror(errno));
goto error;
}
if (!PEM_read_PrivateKey(fp, &private_key, NULL, NULL)) {
ast_log(LOG_ERROR, "Failed to read private key from PEM file '%s'\n", private_key_file);
fclose(fp);
goto error;
}
if (fclose(fp)) {
ast_log(LOG_ERROR, "Failed to close private key file '%s': %s\n", private_key_file, strerror(errno));
goto error;
}
certbio = BIO_new(BIO_s_file());
if (!certbio) {
ast_log(LOG_ERROR, "Failed to allocate memory for certificate fingerprinting on RTP instance '%p'\n",
instance);
goto error;
}
if (!BIO_read_filename(certbio, dtls_cfg->certfile)
|| !(cert = PEM_read_bio_X509(certbio, NULL, 0, NULL))) {
ast_log(LOG_ERROR, "Failed to read certificate from file '%s'\n", dtls_cfg->certfile);
goto error;
}
cert_info->private_key = private_key;
cert_info->certificate = cert;
BIO_free_all(certbio);
return 0;
error:
X509_free(cert);
BIO_free_all(certbio);
EVP_PKEY_free(private_key);
return -1;
}
static int load_dtls_certificate(struct ast_rtp_instance *instance,
const struct ast_rtp_dtls_cfg *dtls_cfg,
struct dtls_cert_info *cert_info)
{
if (dtls_cfg->ephemeral_cert) {
return create_certificate_ephemeral(instance, dtls_cfg, cert_info);
} else if (!ast_strlen_zero(dtls_cfg->certfile)) {
return create_certificate_from_file(instance, dtls_cfg, cert_info);
} else {
return -1;
}
}
/*! \pre instance is locked */
static int ast_rtp_dtls_set_configuration(struct ast_rtp_instance *instance, const struct ast_rtp_dtls_cfg *dtls_cfg)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
struct dtls_cert_info cert_info = { 0 };
int res;
if (!dtls_cfg->enabled) {
return 0;
}
ast_debug_dtls(3, "(%p) DTLS RTP setup\n", instance);
if (!ast_rtp_engine_srtp_is_registered()) {
ast_log(LOG_ERROR, "SRTP support module is not loaded or available. Try loading res_srtp.so.\n");
return -1;
}
if (rtp->ssl_ctx) {
return 0;
}
rtp->ssl_ctx = SSL_CTX_new(get_dtls_method());
if (!rtp->ssl_ctx) {
return -1;
}
SSL_CTX_set_read_ahead(rtp->ssl_ctx, 1);
configure_dhparams(rtp, dtls_cfg);
rtp->dtls_verify = dtls_cfg->verify;
SSL_CTX_set_verify(rtp->ssl_ctx, (rtp->dtls_verify & AST_RTP_DTLS_VERIFY_FINGERPRINT) || (rtp->dtls_verify & AST_RTP_DTLS_VERIFY_CERTIFICATE) ?
SSL_VERIFY_PEER | SSL_VERIFY_FAIL_IF_NO_PEER_CERT : SSL_VERIFY_NONE, !(rtp->dtls_verify & AST_RTP_DTLS_VERIFY_CERTIFICATE) ?
dtls_verify_callback : NULL);
if (dtls_cfg->suite == AST_AES_CM_128_HMAC_SHA1_80) {
SSL_CTX_set_tlsext_use_srtp(rtp->ssl_ctx, "SRTP_AES128_CM_SHA1_80");
} else if (dtls_cfg->suite == AST_AES_CM_128_HMAC_SHA1_32) {
SSL_CTX_set_tlsext_use_srtp(rtp->ssl_ctx, "SRTP_AES128_CM_SHA1_32");
} else {
ast_log(LOG_ERROR, "Unsupported suite specified for DTLS-SRTP on RTP instance '%p'\n", instance);
return -1;
}
rtp->local_hash = dtls_cfg->hash;
if (!load_dtls_certificate(instance, dtls_cfg, &cert_info)) {
const EVP_MD *type;
unsigned int size, i;
unsigned char fingerprint[EVP_MAX_MD_SIZE];
char *local_fingerprint = rtp->local_fingerprint;
if (!SSL_CTX_use_certificate(rtp->ssl_ctx, cert_info.certificate)) {
ast_log(LOG_ERROR, "Specified certificate for RTP instance '%p' could not be used\n",
instance);
return -1;
}
if (!SSL_CTX_use_PrivateKey(rtp->ssl_ctx, cert_info.private_key)
|| !SSL_CTX_check_private_key(rtp->ssl_ctx)) {
ast_log(LOG_ERROR, "Specified private key for RTP instance '%p' could not be used\n",
instance);
return -1;
}
if (rtp->local_hash == AST_RTP_DTLS_HASH_SHA1) {
type = EVP_sha1();
} else if (rtp->local_hash == AST_RTP_DTLS_HASH_SHA256) {
type = EVP_sha256();
} else {
ast_log(LOG_ERROR, "Unsupported fingerprint hash type on RTP instance '%p'\n",
instance);
return -1;
}
if (!X509_digest(cert_info.certificate, type, fingerprint, &size) || !size) {
ast_log(LOG_ERROR, "Could not produce fingerprint from certificate for RTP instance '%p'\n",
instance);
return -1;
}
for (i = 0; i < size; i++) {
sprintf(local_fingerprint, "%02hhX:", fingerprint[i]);
local_fingerprint += 3;
}
*(local_fingerprint - 1) = 0;
EVP_PKEY_free(cert_info.private_key);
X509_free(cert_info.certificate);
}
if (!ast_strlen_zero(dtls_cfg->cipher)) {
if (!SSL_CTX_set_cipher_list(rtp->ssl_ctx, dtls_cfg->cipher)) {
ast_log(LOG_ERROR, "Invalid cipher specified in cipher list '%s' for RTP instance '%p'\n",
dtls_cfg->cipher, instance);
return -1;
}
}
if (!ast_strlen_zero(dtls_cfg->cafile) || !ast_strlen_zero(dtls_cfg->capath)) {
if (!SSL_CTX_load_verify_locations(rtp->ssl_ctx, S_OR(dtls_cfg->cafile, NULL), S_OR(dtls_cfg->capath, NULL))) {
ast_log(LOG_ERROR, "Invalid certificate authority file '%s' or path '%s' specified for RTP instance '%p'\n",
S_OR(dtls_cfg->cafile, ""), S_OR(dtls_cfg->capath, ""), instance);
return -1;
}
}
rtp->rekey = dtls_cfg->rekey;
rtp->suite = dtls_cfg->suite;
res = dtls_details_initialize(&rtp->dtls, rtp->ssl_ctx, dtls_cfg->default_setup, instance);
if (!res) {
dtls_setup_rtcp(instance);
}
return res;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static int ast_rtp_dtls_active(struct ast_rtp_instance *instance)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
return !rtp->ssl_ctx ? 0 : 1;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static void ast_rtp_dtls_stop(struct ast_rtp_instance *instance)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
SSL *ssl = rtp->dtls.ssl;
ast_debug_dtls(3, "(%p) DTLS stop\n", instance);
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_unlock(instance);
dtls_srtp_stop_timeout_timer(instance, rtp, 0);
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_lock(instance);
if (rtp->ssl_ctx) {
SSL_CTX_free(rtp->ssl_ctx);
rtp->ssl_ctx = NULL;
}
if (rtp->dtls.ssl) {
SSL_free(rtp->dtls.ssl);
rtp->dtls.ssl = NULL;
}
if (rtp->rtcp) {
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_unlock(instance);
dtls_srtp_stop_timeout_timer(instance, rtp, 1);
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_lock(instance);
if (rtp->rtcp->dtls.ssl) {
if (rtp->rtcp->dtls.ssl != ssl) {
SSL_free(rtp->rtcp->dtls.ssl);
}
rtp->rtcp->dtls.ssl = NULL;
}
}
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static void ast_rtp_dtls_reset(struct ast_rtp_instance *instance)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
if (SSL_is_init_finished(rtp->dtls.ssl)) {
SSL_shutdown(rtp->dtls.ssl);
rtp->dtls.connection = AST_RTP_DTLS_CONNECTION_NEW;
}
if (rtp->rtcp && SSL_is_init_finished(rtp->rtcp->dtls.ssl)) {
SSL_shutdown(rtp->rtcp->dtls.ssl);
rtp->rtcp->dtls.connection = AST_RTP_DTLS_CONNECTION_NEW;
}
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static enum ast_rtp_dtls_connection ast_rtp_dtls_get_connection(struct ast_rtp_instance *instance)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
return rtp->dtls.connection;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static enum ast_rtp_dtls_setup ast_rtp_dtls_get_setup(struct ast_rtp_instance *instance)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
return rtp->dtls.dtls_setup;
}
static void dtls_set_setup(enum ast_rtp_dtls_setup *dtls_setup, enum ast_rtp_dtls_setup setup, SSL *ssl)
{
enum ast_rtp_dtls_setup old = *dtls_setup;
switch (setup) {
case AST_RTP_DTLS_SETUP_ACTIVE:
*dtls_setup = AST_RTP_DTLS_SETUP_PASSIVE;
break;
case AST_RTP_DTLS_SETUP_PASSIVE:
*dtls_setup = AST_RTP_DTLS_SETUP_ACTIVE;
break;
case AST_RTP_DTLS_SETUP_ACTPASS:
/* We can't respond to an actpass setup with actpass ourselves... so respond with active, as we can initiate connections */
if (*dtls_setup == AST_RTP_DTLS_SETUP_ACTPASS) {
*dtls_setup = AST_RTP_DTLS_SETUP_ACTIVE;
}
break;
case AST_RTP_DTLS_SETUP_HOLDCONN:
*dtls_setup = AST_RTP_DTLS_SETUP_HOLDCONN;
break;
default:
/* This should never occur... if it does exit early as we don't know what state things are in */
return;
}
/* If the setup state did not change we go on as if nothing happened */
if (old == *dtls_setup) {
return;
}
/* If they don't want us to establish a connection wait until later */
if (*dtls_setup == AST_RTP_DTLS_SETUP_HOLDCONN) {
return;
}
if (*dtls_setup == AST_RTP_DTLS_SETUP_ACTIVE) {
SSL_set_connect_state(ssl);
} else if (*dtls_setup == AST_RTP_DTLS_SETUP_PASSIVE) {
SSL_set_accept_state(ssl);
} else {
return;
}
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static void ast_rtp_dtls_set_setup(struct ast_rtp_instance *instance, enum ast_rtp_dtls_setup setup)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
if (rtp->dtls.ssl) {
dtls_set_setup(&rtp->dtls.dtls_setup, setup, rtp->dtls.ssl);
}
if (rtp->rtcp && rtp->rtcp->dtls.ssl) {
dtls_set_setup(&rtp->rtcp->dtls.dtls_setup, setup, rtp->rtcp->dtls.ssl);
}
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static void ast_rtp_dtls_set_fingerprint(struct ast_rtp_instance *instance, enum ast_rtp_dtls_hash hash, const char *fingerprint)
{
char *tmp = ast_strdupa(fingerprint), *value;
int pos = 0;
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
if (hash != AST_RTP_DTLS_HASH_SHA1 && hash != AST_RTP_DTLS_HASH_SHA256) {
return;
}
rtp->remote_hash = hash;
while ((value = strsep(&tmp, ":")) && (pos != (EVP_MAX_MD_SIZE - 1))) {
sscanf(value, "%02hhx", &rtp->remote_fingerprint[pos++]);
}
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static enum ast_rtp_dtls_hash ast_rtp_dtls_get_fingerprint_hash(struct ast_rtp_instance *instance)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
return rtp->local_hash;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static const char *ast_rtp_dtls_get_fingerprint(struct ast_rtp_instance *instance)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
return rtp->local_fingerprint;
}
/* DTLS RTP Engine interface declaration */
static struct ast_rtp_engine_dtls ast_rtp_dtls = {
.set_configuration = ast_rtp_dtls_set_configuration,
.active = ast_rtp_dtls_active,
.stop = ast_rtp_dtls_stop,
.reset = ast_rtp_dtls_reset,
.get_connection = ast_rtp_dtls_get_connection,
.get_setup = ast_rtp_dtls_get_setup,
.set_setup = ast_rtp_dtls_set_setup,
.set_fingerprint = ast_rtp_dtls_set_fingerprint,
.get_fingerprint_hash = ast_rtp_dtls_get_fingerprint_hash,
.get_fingerprint = ast_rtp_dtls_get_fingerprint,
};
#endif
#ifdef TEST_FRAMEWORK
static size_t get_recv_buffer_count(struct ast_rtp_instance *instance)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
if (rtp && rtp->recv_buffer) {
return ast_data_buffer_count(rtp->recv_buffer);
}
return 0;
}
static size_t get_recv_buffer_max(struct ast_rtp_instance *instance)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
if (rtp && rtp->recv_buffer) {
return ast_data_buffer_max(rtp->recv_buffer);
}
return 0;
}
static size_t get_send_buffer_count(struct ast_rtp_instance *instance)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
if (rtp && rtp->send_buffer) {
return ast_data_buffer_count(rtp->send_buffer);
}
return 0;
}
static void set_rtp_rtcp_schedid(struct ast_rtp_instance *instance, int id)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
if (rtp && rtp->rtcp) {
rtp->rtcp->schedid = id;
}
}
static struct ast_rtp_engine_test ast_rtp_test = {
.packets_to_drop = 0,
.send_report = 0,
.sdes_received = 0,
.recv_buffer_count = get_recv_buffer_count,
.recv_buffer_max = get_recv_buffer_max,
.send_buffer_count = get_send_buffer_count,
.set_schedid = set_rtp_rtcp_schedid,
};
#endif
/* RTP Engine Declaration */
static struct ast_rtp_engine asterisk_rtp_engine = {
.name = "asterisk",
.new = ast_rtp_new,
.destroy = ast_rtp_destroy,
.dtmf_begin = ast_rtp_dtmf_begin,
.dtmf_end = ast_rtp_dtmf_end,
.dtmf_end_with_duration = ast_rtp_dtmf_end_with_duration,
.dtmf_mode_set = ast_rtp_dtmf_mode_set,
.dtmf_mode_get = ast_rtp_dtmf_mode_get,
.update_source = ast_rtp_update_source,
.change_source = ast_rtp_change_source,
.write = ast_rtp_write,
.read = ast_rtp_read,
.prop_set = ast_rtp_prop_set,
.fd = ast_rtp_fd,
.remote_address_set = ast_rtp_remote_address_set,
.red_init = rtp_red_init,
.red_buffer = rtp_red_buffer,
.local_bridge = ast_rtp_local_bridge,
.get_stat = ast_rtp_get_stat,
.dtmf_compatible = ast_rtp_dtmf_compatible,
.stun_request = ast_rtp_stun_request,
.stop = ast_rtp_stop,
.qos = ast_rtp_qos_set,
.sendcng = ast_rtp_sendcng,
#ifdef HAVE_PJPROJECT
.ice = &ast_rtp_ice,
#endif
#if defined(HAVE_OPENSSL) && (OPENSSL_VERSION_NUMBER >= 0x10001000L) && !defined(OPENSSL_NO_SRTP)
.dtls = &ast_rtp_dtls,
.activate = ast_rtp_activate,
#endif
.ssrc_get = ast_rtp_get_ssrc,
.cname_get = ast_rtp_get_cname,
.set_remote_ssrc = ast_rtp_set_remote_ssrc,
.set_stream_num = ast_rtp_set_stream_num,
.extension_enable = ast_rtp_extension_enable,
.bundle = ast_rtp_bundle,
#ifdef TEST_FRAMEWORK
.test = &ast_rtp_test,
#endif
};
#if defined(HAVE_OPENSSL) && (OPENSSL_VERSION_NUMBER >= 0x10001000L) && !defined(OPENSSL_NO_SRTP)
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static void dtls_perform_handshake(struct ast_rtp_instance *instance, struct dtls_details *dtls, int rtcp)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
ast_debug_dtls(3, "(%p) DTLS perform handshake - ssl = %p, setup = %d\n",
res_rtp_asterisk: bad audio (static) due to incomplete dtls/srtp setup There was a race condition between client initiated DTLS setup, and handling of server side ice completion that caused the underlying SSL object to get cleared during DTLS initialization. If this happened Asterisk would be left in a partial DTLS setup state. RTP packets were sent and received, but were not being encrypted and decrypted. This resulted in no audio, or static. Specifically, this occurred when '__rtp_recvfrom' was processing the handshake sequence from the client to the server, and then 'ast_rtp_on_ice_complete' gets called from another thread and clears the SSL object when calling the 'dtls_perform_setup' function. The timing had to be just right in the sense that from the external SSL library perspective SSL initialization completed (rtp recv), Asterisk clears/resets the SSL object (ice done), and then checks to see if SSL is intialized (rtp recv). Since it was cleared, Asterisk thinks it is not finished, thus not completing 'dtls_srtp_setup'. This patch removes calls to 'dtls_perform_setup', which clears the SSL object, in 'ast_rtp_on_ice_complete'. When ice completes, there is no reason to clear the underlying SSL object. If an ice candidate changes a full protocol level renegotiation occurs. Also, in the case of bundled ICE candidates are reused when a stream is added. So no real reason to have to clear, and reset in this instance. Also, this patch adds a bit of extra logging to aid in diagnosis of any future problems. ASTERISK-28742 #close Change-Id: I34c9e6bad5a39b087164646e2836e3e48fe6892f
2020-02-13 21:08:10 +00:00
rtp, dtls->ssl, dtls->dtls_setup);
/* If we are not acting as a client connecting to the remote side then
* don't start the handshake as it will accomplish nothing and would conflict
* with the handshake we receive from the remote side.
*/
if (!dtls->ssl || (dtls->dtls_setup != AST_RTP_DTLS_SETUP_ACTIVE)) {
return;
}
SSL_do_handshake(dtls->ssl);
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*
* A race condition is prevented between this function and __rtp_recvfrom()
* because both functions have to get the instance lock before they can do
* anything. Without holding the instance lock, this function could start
* the SSL handshake above in one thread and the __rtp_recvfrom() function
* called by the channel thread could read the response and stop the timeout
* timer before we have a chance to even start it.
*/
dtls_srtp_start_timeout_timer(instance, rtp, rtcp);
}
#endif
#if defined(HAVE_OPENSSL) && (OPENSSL_VERSION_NUMBER >= 0x10001000L) && !defined(OPENSSL_NO_SRTP)
static void dtls_perform_setup(struct dtls_details *dtls)
{
if (!dtls->ssl || !SSL_is_init_finished(dtls->ssl)) {
return;
}
SSL_clear(dtls->ssl);
if (dtls->dtls_setup == AST_RTP_DTLS_SETUP_PASSIVE) {
SSL_set_accept_state(dtls->ssl);
} else {
SSL_set_connect_state(dtls->ssl);
}
dtls->connection = AST_RTP_DTLS_CONNECTION_NEW;
res_rtp_asterisk: bad audio (static) due to incomplete dtls/srtp setup There was a race condition between client initiated DTLS setup, and handling of server side ice completion that caused the underlying SSL object to get cleared during DTLS initialization. If this happened Asterisk would be left in a partial DTLS setup state. RTP packets were sent and received, but were not being encrypted and decrypted. This resulted in no audio, or static. Specifically, this occurred when '__rtp_recvfrom' was processing the handshake sequence from the client to the server, and then 'ast_rtp_on_ice_complete' gets called from another thread and clears the SSL object when calling the 'dtls_perform_setup' function. The timing had to be just right in the sense that from the external SSL library perspective SSL initialization completed (rtp recv), Asterisk clears/resets the SSL object (ice done), and then checks to see if SSL is intialized (rtp recv). Since it was cleared, Asterisk thinks it is not finished, thus not completing 'dtls_srtp_setup'. This patch removes calls to 'dtls_perform_setup', which clears the SSL object, in 'ast_rtp_on_ice_complete'. When ice completes, there is no reason to clear the underlying SSL object. If an ice candidate changes a full protocol level renegotiation occurs. Also, in the case of bundled ICE candidates are reused when a stream is added. So no real reason to have to clear, and reset in this instance. Also, this patch adds a bit of extra logging to aid in diagnosis of any future problems. ASTERISK-28742 #close Change-Id: I34c9e6bad5a39b087164646e2836e3e48fe6892f
2020-02-13 21:08:10 +00:00
ast_debug_dtls(3, "DTLS perform setup - connection reset\n");
}
#endif
#ifdef HAVE_PJPROJECT
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
static void rtp_learning_start(struct ast_rtp *rtp);
/* Handles start of media during ICE negotiation or completion */
static void ast_rtp_ice_start_media(pj_ice_sess *ice, pj_status_t status)
{
struct ast_rtp_instance *instance = ice->user_data;
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_lock(instance);
if (status == PJ_SUCCESS) {
struct ast_sockaddr remote_address;
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ast_sockaddr_setnull(&remote_address);
update_address_with_ice_candidate(ice, AST_RTP_ICE_COMPONENT_RTP, &remote_address);
if (!ast_sockaddr_isnull(&remote_address)) {
/* Symmetric RTP must be disabled for the remote address to not get overwritten */
ast_rtp_instance_set_prop(instance, AST_RTP_PROPERTY_NAT, 0);
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ast_rtp_instance_set_remote_address(instance, &remote_address);
}
if (rtp->rtcp) {
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
update_address_with_ice_candidate(ice, AST_RTP_ICE_COMPONENT_RTCP, &rtp->rtcp->them);
}
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
#if defined(HAVE_OPENSSL) && (OPENSSL_VERSION_NUMBER >= 0x10001000L) && !defined(OPENSSL_NO_SRTP)
/* If we've already started media, no need to do all of this again */
if (rtp->ice_media_started) {
ao2_unlock(instance);
return;
}
ast_debug_category(2, AST_DEBUG_CATEGORY_ICE | AST_DEBUG_CATEGORY_DTLS,
"(%p) ICE starting media - perform DTLS - (%p)\n", instance, rtp);
res_rtp_asterisk: bad audio (static) due to incomplete dtls/srtp setup There was a race condition between client initiated DTLS setup, and handling of server side ice completion that caused the underlying SSL object to get cleared during DTLS initialization. If this happened Asterisk would be left in a partial DTLS setup state. RTP packets were sent and received, but were not being encrypted and decrypted. This resulted in no audio, or static. Specifically, this occurred when '__rtp_recvfrom' was processing the handshake sequence from the client to the server, and then 'ast_rtp_on_ice_complete' gets called from another thread and clears the SSL object when calling the 'dtls_perform_setup' function. The timing had to be just right in the sense that from the external SSL library perspective SSL initialization completed (rtp recv), Asterisk clears/resets the SSL object (ice done), and then checks to see if SSL is intialized (rtp recv). Since it was cleared, Asterisk thinks it is not finished, thus not completing 'dtls_srtp_setup'. This patch removes calls to 'dtls_perform_setup', which clears the SSL object, in 'ast_rtp_on_ice_complete'. When ice completes, there is no reason to clear the underlying SSL object. If an ice candidate changes a full protocol level renegotiation occurs. Also, in the case of bundled ICE candidates are reused when a stream is added. So no real reason to have to clear, and reset in this instance. Also, this patch adds a bit of extra logging to aid in diagnosis of any future problems. ASTERISK-28742 #close Change-Id: I34c9e6bad5a39b087164646e2836e3e48fe6892f
2020-02-13 21:08:10 +00:00
/*
* Seemingly no reason to call dtls_perform_setup here. Currently we'll do a full
* protocol level renegotiation if things do change. And if bundled is being used
* then ICE is reused when a stream is added.
*
* Note, if for some reason in the future dtls_perform_setup does need to done here
* be aware that creates a race condition between the call here (on ice completion)
* and potential DTLS handshaking when receiving RTP. What happens is the ssl object
* can get cleared (SSL_clear) during that handshaking process (DTLS init). If that
* happens then Asterisk won't complete DTLS initialization. RTP packets are still
* sent/received but won't be encrypted/decrypted.
*/
dtls_perform_handshake(instance, &rtp->dtls, 0);
if (rtp->rtcp && rtp->rtcp->type == AST_RTP_INSTANCE_RTCP_STANDARD) {
dtls_perform_handshake(instance, &rtp->rtcp->dtls, 1);
}
#endif
rtp->ice_media_started = 1;
if (!strictrtp) {
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_unlock(instance);
return;
}
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
ast_verb(4, "%p -- Strict RTP learning after ICE completion\n", rtp);
rtp_learning_start(rtp);
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_unlock(instance);
}
#ifdef HAVE_PJPROJECT_ON_VALID_ICE_PAIR_CALLBACK
/* PJPROJECT ICE optional callback */
static void ast_rtp_on_valid_pair(pj_ice_sess *ice)
{
ast_debug_ice(2, "(%p) ICE valid pair, start media\n", ice->user_data);
ast_rtp_ice_start_media(ice, PJ_SUCCESS);
}
#endif
/* PJPROJECT ICE callback */
static void ast_rtp_on_ice_complete(pj_ice_sess *ice, pj_status_t status)
{
ast_debug_ice(2, "(%p) ICE complete, start media\n", ice->user_data);
ast_rtp_ice_start_media(ice, status);
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/* PJPROJECT ICE callback */
static void ast_rtp_on_ice_rx_data(pj_ice_sess *ice, unsigned comp_id, unsigned transport_id, void *pkt, pj_size_t size, const pj_sockaddr_t *src_addr, unsigned src_addr_len)
{
struct ast_rtp_instance *instance = ice->user_data;
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
/* Instead of handling the packet here (which really doesn't work with our architecture) we set a bit to indicate that it should be handled after pj_ice_sess_on_rx_pkt
* returns */
if (transport_id == TRANSPORT_SOCKET_RTP || transport_id == TRANSPORT_SOCKET_RTCP) {
rtp->passthrough = 1;
} else if (transport_id == TRANSPORT_TURN_RTP) {
rtp->rtp_passthrough = 1;
} else if (transport_id == TRANSPORT_TURN_RTCP) {
rtp->rtcp_passthrough = 1;
}
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/* PJPROJECT ICE callback */
static pj_status_t ast_rtp_on_ice_tx_pkt(pj_ice_sess *ice, unsigned comp_id, unsigned transport_id, const void *pkt, pj_size_t size, const pj_sockaddr_t *dst_addr, unsigned dst_addr_len)
{
struct ast_rtp_instance *instance = ice->user_data;
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
pj_status_t status = PJ_EINVALIDOP;
pj_ssize_t _size = (pj_ssize_t)size;
if (transport_id == TRANSPORT_SOCKET_RTP) {
/* Traffic is destined to go right out the RTP socket we already have */
status = pj_sock_sendto(rtp->s, pkt, &_size, 0, dst_addr, dst_addr_len);
/* sendto on a connectionless socket should send all the data, or none at all */
ast_assert(_size == size || status != PJ_SUCCESS);
} else if (transport_id == TRANSPORT_SOCKET_RTCP) {
/* Traffic is destined to go right out the RTCP socket we already have */
if (rtp->rtcp) {
status = pj_sock_sendto(rtp->rtcp->s, pkt, &_size, 0, dst_addr, dst_addr_len);
/* sendto on a connectionless socket should send all the data, or none at all */
ast_assert(_size == size || status != PJ_SUCCESS);
} else {
status = PJ_SUCCESS;
}
} else if (transport_id == TRANSPORT_TURN_RTP) {
/* Traffic is going through the RTP TURN relay */
if (rtp->turn_rtp) {
status = pj_turn_sock_sendto(rtp->turn_rtp, pkt, size, dst_addr, dst_addr_len);
}
} else if (transport_id == TRANSPORT_TURN_RTCP) {
/* Traffic is going through the RTCP TURN relay */
if (rtp->turn_rtcp) {
status = pj_turn_sock_sendto(rtp->turn_rtcp, pkt, size, dst_addr, dst_addr_len);
}
}
return status;
}
/* ICE Session interface declaration */
static pj_ice_sess_cb ast_rtp_ice_sess_cb = {
#ifdef HAVE_PJPROJECT_ON_VALID_ICE_PAIR_CALLBACK
.on_valid_pair = ast_rtp_on_valid_pair,
#endif
.on_ice_complete = ast_rtp_on_ice_complete,
.on_rx_data = ast_rtp_on_ice_rx_data,
.on_tx_pkt = ast_rtp_on_ice_tx_pkt,
};
/*! \brief Worker thread for timerheap */
static int timer_worker_thread(void *data)
{
pj_ioqueue_t *ioqueue;
if (pj_ioqueue_create(pool, 1, &ioqueue) != PJ_SUCCESS) {
return -1;
}
while (!timer_terminate) {
const pj_time_val delay = {0, 10};
pj_timer_heap_poll(timer_heap, NULL);
pj_ioqueue_poll(ioqueue, &delay);
}
return 0;
}
#endif
static inline int rtp_debug_test_addr(struct ast_sockaddr *addr)
{
if (!ast_debug_rtp_packet_is_allowed) {
return 0;
}
if (!ast_sockaddr_isnull(&rtpdebugaddr)) {
if (rtpdebugport) {
return (ast_sockaddr_cmp(&rtpdebugaddr, addr) == 0); /* look for RTP packets from IP+Port */
} else {
return (ast_sockaddr_cmp_addr(&rtpdebugaddr, addr) == 0); /* only look for RTP packets from IP */
}
}
return 1;
}
static inline int rtcp_debug_test_addr(struct ast_sockaddr *addr)
{
if (!ast_debug_rtcp_packet_is_allowed) {
return 0;
}
if (!ast_sockaddr_isnull(&rtcpdebugaddr)) {
if (rtcpdebugport) {
return (ast_sockaddr_cmp(&rtcpdebugaddr, addr) == 0); /* look for RTCP packets from IP+Port */
} else {
return (ast_sockaddr_cmp_addr(&rtcpdebugaddr, addr) == 0); /* only look for RTCP packets from IP */
}
}
return 1;
}
#if defined(HAVE_OPENSSL) && (OPENSSL_VERSION_NUMBER >= 0x10001000L) && !defined(OPENSSL_NO_SRTP)
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static int dtls_srtp_handle_timeout(struct ast_rtp_instance *instance, int rtcp)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
struct dtls_details *dtls = !rtcp ? &rtp->dtls : &rtp->rtcp->dtls;
struct timeval dtls_timeout;
ast_debug_dtls(3, "(%p) DTLS srtp - handle timeout - rtcp=%d\n", instance, rtcp);
DTLSv1_handle_timeout(dtls->ssl);
/* If a timeout can't be retrieved then this recurring scheduled item must stop */
if (!DTLSv1_get_timeout(dtls->ssl, &dtls_timeout)) {
dtls->timeout_timer = -1;
return 0;
}
return dtls_timeout.tv_sec * 1000 + dtls_timeout.tv_usec / 1000;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/* Scheduler callback */
static int dtls_srtp_handle_rtp_timeout(const void *data)
{
struct ast_rtp_instance *instance = (struct ast_rtp_instance *)data;
int reschedule;
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_lock(instance);
reschedule = dtls_srtp_handle_timeout(instance, 0);
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_unlock(instance);
if (!reschedule) {
ao2_ref(instance, -1);
}
return reschedule;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/* Scheduler callback */
static int dtls_srtp_handle_rtcp_timeout(const void *data)
{
struct ast_rtp_instance *instance = (struct ast_rtp_instance *)data;
int reschedule;
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_lock(instance);
reschedule = dtls_srtp_handle_timeout(instance, 1);
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_unlock(instance);
if (!reschedule) {
ao2_ref(instance, -1);
}
return reschedule;
}
static void dtls_srtp_start_timeout_timer(struct ast_rtp_instance *instance, struct ast_rtp *rtp, int rtcp)
{
struct dtls_details *dtls = !rtcp ? &rtp->dtls : &rtp->rtcp->dtls;
struct timeval dtls_timeout;
if (DTLSv1_get_timeout(dtls->ssl, &dtls_timeout)) {
int timeout = dtls_timeout.tv_sec * 1000 + dtls_timeout.tv_usec / 1000;
ast_assert(dtls->timeout_timer == -1);
ao2_ref(instance, +1);
if ((dtls->timeout_timer = ast_sched_add(rtp->sched, timeout,
!rtcp ? dtls_srtp_handle_rtp_timeout : dtls_srtp_handle_rtcp_timeout, instance)) < 0) {
ao2_ref(instance, -1);
ast_log(LOG_WARNING, "Scheduling '%s' DTLS retransmission for RTP instance [%p] failed.\n",
!rtcp ? "RTP" : "RTCP", instance);
} else {
ast_debug_dtls(3, "(%p) DTLS srtp - scheduled timeout timer for '%d'\n", instance, timeout);
}
}
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre Must not be called with the instance locked. */
static void dtls_srtp_stop_timeout_timer(struct ast_rtp_instance *instance, struct ast_rtp *rtp, int rtcp)
{
struct dtls_details *dtls = !rtcp ? &rtp->dtls : &rtp->rtcp->dtls;
AST_SCHED_DEL_UNREF(rtp->sched, dtls->timeout_timer, ao2_ref(instance, -1));
ast_debug_dtls(3, "(%p) DTLS srtp - stopped timeout timer'\n", instance);
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/* Scheduler callback */
static int dtls_srtp_renegotiate(const void *data)
{
struct ast_rtp_instance *instance = (struct ast_rtp_instance *)data;
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_lock(instance);
ast_debug_dtls(3, "(%p) DTLS srtp - renegotiate'\n", instance);
SSL_renegotiate(rtp->dtls.ssl);
SSL_do_handshake(rtp->dtls.ssl);
if (rtp->rtcp && rtp->rtcp->dtls.ssl && rtp->rtcp->dtls.ssl != rtp->dtls.ssl) {
SSL_renegotiate(rtp->rtcp->dtls.ssl);
SSL_do_handshake(rtp->rtcp->dtls.ssl);
}
rtp->rekeyid = -1;
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_unlock(instance);
ao2_ref(instance, -1);
return 0;
}
static int dtls_srtp_add_local_ssrc(struct ast_rtp *rtp, struct ast_rtp_instance *instance, int rtcp, unsigned int ssrc, int set_remote_policy)
{
unsigned char material[SRTP_MASTER_LEN * 2];
unsigned char *local_key, *local_salt, *remote_key, *remote_salt;
struct ast_srtp_policy *local_policy, *remote_policy = NULL;
int res = -1;
struct dtls_details *dtls = !rtcp ? &rtp->dtls : &rtp->rtcp->dtls;
ast_debug_dtls(3, "(%p) DTLS srtp - add local ssrc - rtcp=%d, set_remote_policy=%d'\n",
instance, rtcp, set_remote_policy);
/* Produce key information and set up SRTP */
if (!SSL_export_keying_material(dtls->ssl, material, SRTP_MASTER_LEN * 2, "EXTRACTOR-dtls_srtp", 19, NULL, 0, 0)) {
ast_log(LOG_WARNING, "Unable to extract SRTP keying material from DTLS-SRTP negotiation on RTP instance '%p'\n",
instance);
return -1;
}
/* Whether we are acting as a server or client determines where the keys/salts are */
if (rtp->dtls.dtls_setup == AST_RTP_DTLS_SETUP_ACTIVE) {
local_key = material;
remote_key = local_key + SRTP_MASTER_KEY_LEN;
local_salt = remote_key + SRTP_MASTER_KEY_LEN;
remote_salt = local_salt + SRTP_MASTER_SALT_LEN;
} else {
remote_key = material;
local_key = remote_key + SRTP_MASTER_KEY_LEN;
remote_salt = local_key + SRTP_MASTER_KEY_LEN;
local_salt = remote_salt + SRTP_MASTER_SALT_LEN;
}
if (!(local_policy = res_srtp_policy->alloc())) {
return -1;
}
if (res_srtp_policy->set_master_key(local_policy, local_key, SRTP_MASTER_KEY_LEN, local_salt, SRTP_MASTER_SALT_LEN) < 0) {
ast_log(LOG_WARNING, "Could not set key/salt information on local policy of '%p' when setting up DTLS-SRTP\n", rtp);
goto error;
}
if (res_srtp_policy->set_suite(local_policy, rtp->suite)) {
ast_log(LOG_WARNING, "Could not set suite to '%u' on local policy of '%p' when setting up DTLS-SRTP\n", rtp->suite, rtp);
goto error;
}
res_srtp_policy->set_ssrc(local_policy, ssrc, 0);
if (set_remote_policy) {
if (!(remote_policy = res_srtp_policy->alloc())) {
goto error;
}
if (res_srtp_policy->set_master_key(remote_policy, remote_key, SRTP_MASTER_KEY_LEN, remote_salt, SRTP_MASTER_SALT_LEN) < 0) {
ast_log(LOG_WARNING, "Could not set key/salt information on remote policy of '%p' when setting up DTLS-SRTP\n", rtp);
goto error;
}
if (res_srtp_policy->set_suite(remote_policy, rtp->suite)) {
ast_log(LOG_WARNING, "Could not set suite to '%u' on remote policy of '%p' when setting up DTLS-SRTP\n", rtp->suite, rtp);
goto error;
}
res_srtp_policy->set_ssrc(remote_policy, 0, 1);
}
if (ast_rtp_instance_add_srtp_policy(instance, remote_policy, local_policy, rtcp)) {
ast_log(LOG_WARNING, "Could not set policies when setting up DTLS-SRTP on '%p'\n", rtp);
goto error;
}
res = 0;
error:
/* policy->destroy() called even on success to release local reference to these resources */
res_srtp_policy->destroy(local_policy);
if (remote_policy) {
res_srtp_policy->destroy(remote_policy);
}
return res;
}
static int dtls_srtp_setup(struct ast_rtp *rtp, struct ast_rtp_instance *instance, int rtcp)
{
struct dtls_details *dtls = !rtcp ? &rtp->dtls : &rtp->rtcp->dtls;
int index;
ast_debug_dtls(3, "(%p) DTLS setup SRTP rtp=%p'\n", instance, rtp);
res_rtp_asterisk: bad audio (static) due to incomplete dtls/srtp setup There was a race condition between client initiated DTLS setup, and handling of server side ice completion that caused the underlying SSL object to get cleared during DTLS initialization. If this happened Asterisk would be left in a partial DTLS setup state. RTP packets were sent and received, but were not being encrypted and decrypted. This resulted in no audio, or static. Specifically, this occurred when '__rtp_recvfrom' was processing the handshake sequence from the client to the server, and then 'ast_rtp_on_ice_complete' gets called from another thread and clears the SSL object when calling the 'dtls_perform_setup' function. The timing had to be just right in the sense that from the external SSL library perspective SSL initialization completed (rtp recv), Asterisk clears/resets the SSL object (ice done), and then checks to see if SSL is intialized (rtp recv). Since it was cleared, Asterisk thinks it is not finished, thus not completing 'dtls_srtp_setup'. This patch removes calls to 'dtls_perform_setup', which clears the SSL object, in 'ast_rtp_on_ice_complete'. When ice completes, there is no reason to clear the underlying SSL object. If an ice candidate changes a full protocol level renegotiation occurs. Also, in the case of bundled ICE candidates are reused when a stream is added. So no real reason to have to clear, and reset in this instance. Also, this patch adds a bit of extra logging to aid in diagnosis of any future problems. ASTERISK-28742 #close Change-Id: I34c9e6bad5a39b087164646e2836e3e48fe6892f
2020-02-13 21:08:10 +00:00
/* If a fingerprint is present in the SDP make sure that the peer certificate matches it */
if (rtp->dtls_verify & AST_RTP_DTLS_VERIFY_FINGERPRINT) {
X509 *certificate;
if (!(certificate = SSL_get_peer_certificate(dtls->ssl))) {
ast_log(LOG_WARNING, "No certificate was provided by the peer on RTP instance '%p'\n", instance);
return -1;
}
/* If a fingerprint is present in the SDP make sure that the peer certificate matches it */
if (rtp->remote_fingerprint[0]) {
const EVP_MD *type;
unsigned char fingerprint[EVP_MAX_MD_SIZE];
unsigned int size;
if (rtp->remote_hash == AST_RTP_DTLS_HASH_SHA1) {
type = EVP_sha1();
} else if (rtp->remote_hash == AST_RTP_DTLS_HASH_SHA256) {
type = EVP_sha256();
} else {
ast_log(LOG_WARNING, "Unsupported fingerprint hash type on RTP instance '%p'\n", instance);
return -1;
}
if (!X509_digest(certificate, type, fingerprint, &size) ||
!size ||
memcmp(fingerprint, rtp->remote_fingerprint, size)) {
X509_free(certificate);
ast_log(LOG_WARNING, "Fingerprint provided by remote party does not match that of peer certificate on RTP instance '%p'\n",
instance);
return -1;
}
}
X509_free(certificate);
}
if (dtls_srtp_add_local_ssrc(rtp, instance, rtcp, ast_rtp_instance_get_ssrc(instance), 1)) {
res_rtp_asterisk: bad audio (static) due to incomplete dtls/srtp setup There was a race condition between client initiated DTLS setup, and handling of server side ice completion that caused the underlying SSL object to get cleared during DTLS initialization. If this happened Asterisk would be left in a partial DTLS setup state. RTP packets were sent and received, but were not being encrypted and decrypted. This resulted in no audio, or static. Specifically, this occurred when '__rtp_recvfrom' was processing the handshake sequence from the client to the server, and then 'ast_rtp_on_ice_complete' gets called from another thread and clears the SSL object when calling the 'dtls_perform_setup' function. The timing had to be just right in the sense that from the external SSL library perspective SSL initialization completed (rtp recv), Asterisk clears/resets the SSL object (ice done), and then checks to see if SSL is intialized (rtp recv). Since it was cleared, Asterisk thinks it is not finished, thus not completing 'dtls_srtp_setup'. This patch removes calls to 'dtls_perform_setup', which clears the SSL object, in 'ast_rtp_on_ice_complete'. When ice completes, there is no reason to clear the underlying SSL object. If an ice candidate changes a full protocol level renegotiation occurs. Also, in the case of bundled ICE candidates are reused when a stream is added. So no real reason to have to clear, and reset in this instance. Also, this patch adds a bit of extra logging to aid in diagnosis of any future problems. ASTERISK-28742 #close Change-Id: I34c9e6bad5a39b087164646e2836e3e48fe6892f
2020-02-13 21:08:10 +00:00
ast_log(LOG_ERROR, "Failed to add local source '%p'\n", rtp);
return -1;
}
for (index = 0; index < AST_VECTOR_SIZE(&rtp->ssrc_mapping); ++index) {
struct rtp_ssrc_mapping *mapping = AST_VECTOR_GET_ADDR(&rtp->ssrc_mapping, index);
if (dtls_srtp_add_local_ssrc(rtp, instance, rtcp, ast_rtp_instance_get_ssrc(mapping->instance), 0)) {
return -1;
}
}
if (rtp->rekey) {
ao2_ref(instance, +1);
if ((rtp->rekeyid = ast_sched_add(rtp->sched, rtp->rekey * 1000, dtls_srtp_renegotiate, instance)) < 0) {
ao2_ref(instance, -1);
return -1;
}
}
return 0;
}
#endif
2018-05-10 18:11:06 +00:00
/*! \brief Helper function to compare an elem in a vector by value */
static int compare_by_value(int elem, int value)
{
return elem - value;
}
/*! \brief Helper function to find an elem in a vector by value */
static int find_by_value(int elem, int value)
{
return elem == value;
}
static int rtcp_mux(struct ast_rtp *rtp, const unsigned char *packet)
{
uint8_t version;
uint8_t pt;
uint8_t m;
if (!rtp->rtcp || rtp->rtcp->type != AST_RTP_INSTANCE_RTCP_MUX) {
return 0;
}
version = (packet[0] & 0XC0) >> 6;
if (version == 0) {
/* version 0 indicates this is a STUN packet and shouldn't
* be interpreted as a possible RTCP packet
*/
return 0;
}
/* The second octet of a packet will be one of the following:
* For RTP: The marker bit (1 bit) and the RTP payload type (7 bits)
* For RTCP: The payload type (8)
*
* RTP has a forbidden range of payload types (64-95) since these
* will conflict with RTCP payload numbers if the marker bit is set.
*/
m = packet[1] & 0x80;
pt = packet[1] & 0x7F;
if (m && pt >= 64 && pt <= 95) {
return 1;
}
return 0;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static int __rtp_recvfrom(struct ast_rtp_instance *instance, void *buf, size_t size, int flags, struct ast_sockaddr *sa, int rtcp)
{
int len;
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
#if defined(HAVE_OPENSSL) && (OPENSSL_VERSION_NUMBER >= 0x10001000L) && !defined(OPENSSL_NO_SRTP)
char *in = buf;
#endif
#ifdef HAVE_PJPROJECT
struct ast_sockaddr *loop = rtcp ? &rtp->rtcp_loop : &rtp->rtp_loop;
#endif
#ifdef TEST_FRAMEWORK
struct ast_rtp_engine_test *test = ast_rtp_instance_get_test(instance);
#endif
if ((len = ast_recvfrom(rtcp ? rtp->rtcp->s : rtp->s, buf, size, flags, sa)) < 0) {
return len;
}
#ifdef TEST_FRAMEWORK
if (test && test->packets_to_drop > 0) {
test->packets_to_drop--;
return 0;
}
#endif
#if defined(HAVE_OPENSSL) && (OPENSSL_VERSION_NUMBER >= 0x10001000L) && !defined(OPENSSL_NO_SRTP)
/* If this is an SSL packet pass it to OpenSSL for processing. RFC section for first byte value:
* https://tools.ietf.org/html/rfc5764#section-5.1.2 */
if ((*in >= 20) && (*in <= 63)) {
struct dtls_details *dtls = !rtcp ? &rtp->dtls : &rtp->rtcp->dtls;
int res = 0;
/* If no SSL session actually exists terminate things */
if (!dtls->ssl) {
ast_log(LOG_ERROR, "Received SSL traffic on RTP instance '%p' without an SSL session\n",
instance);
return -1;
}
ast_debug_dtls(3, "(%p) DTLS - __rtp_recvfrom rtp=%p - Got SSL packet '%d'\n", instance, rtp, *in);
res_rtp_asterisk: bad audio (static) due to incomplete dtls/srtp setup There was a race condition between client initiated DTLS setup, and handling of server side ice completion that caused the underlying SSL object to get cleared during DTLS initialization. If this happened Asterisk would be left in a partial DTLS setup state. RTP packets were sent and received, but were not being encrypted and decrypted. This resulted in no audio, or static. Specifically, this occurred when '__rtp_recvfrom' was processing the handshake sequence from the client to the server, and then 'ast_rtp_on_ice_complete' gets called from another thread and clears the SSL object when calling the 'dtls_perform_setup' function. The timing had to be just right in the sense that from the external SSL library perspective SSL initialization completed (rtp recv), Asterisk clears/resets the SSL object (ice done), and then checks to see if SSL is intialized (rtp recv). Since it was cleared, Asterisk thinks it is not finished, thus not completing 'dtls_srtp_setup'. This patch removes calls to 'dtls_perform_setup', which clears the SSL object, in 'ast_rtp_on_ice_complete'. When ice completes, there is no reason to clear the underlying SSL object. If an ice candidate changes a full protocol level renegotiation occurs. Also, in the case of bundled ICE candidates are reused when a stream is added. So no real reason to have to clear, and reset in this instance. Also, this patch adds a bit of extra logging to aid in diagnosis of any future problems. ASTERISK-28742 #close Change-Id: I34c9e6bad5a39b087164646e2836e3e48fe6892f
2020-02-13 21:08:10 +00:00
/*
* If ICE is in use, we can prevent a possible DOS attack
* by allowing DTLS protocol messages (client hello, etc)
* only from sources that are in the active remote
* candidates list.
*/
#ifdef HAVE_PJPROJECT
if (rtp->ice) {
int pass_src_check = 0;
int ix = 0;
/*
* You'd think that this check would cause a "deadlock"
* because ast_rtp_ice_start_media calls dtls_perform_handshake
* before it sets ice_media_started = 1 so how can we do a
* handshake if we're dropping packets before we send them
* to openssl. Fortunately, dtls_perform_handshake just sets
* up openssl to do the handshake and doesn't actually perform it
* itself and the locking prevents __rtp_recvfrom from
* running before the ice_media_started flag is set. So only
* unexpected DTLS packets can get dropped here.
*/
if (!rtp->ice_media_started) {
ast_log(LOG_WARNING, "%s: DTLS packet from %s dropped. ICE not completed yet.\n",
ast_rtp_instance_get_channel_id(instance),
ast_sockaddr_stringify(sa));
return 0;
}
/*
* If we got this far, then there have to be candidates.
* We have to use pjproject's rcands because they may have
* peer reflexive candidates that our ice_active_remote_candidates
* won't.
*/
for (ix = 0; ix < rtp->ice->real_ice->rcand_cnt; ix++) {
pj_ice_sess_cand *rcand = &rtp->ice->real_ice->rcand[ix];
if (ast_sockaddr_pj_sockaddr_cmp(sa, &rcand->addr) == 0) {
pass_src_check = 1;
break;
}
}
if (!pass_src_check) {
ast_log(LOG_WARNING, "%s: DTLS packet from %s dropped. Source not in ICE active candidate list.\n",
ast_rtp_instance_get_channel_id(instance),
ast_sockaddr_stringify(sa));
return 0;
}
}
#endif
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*
* A race condition is prevented between dtls_perform_handshake()
* and this function because both functions have to get the
* instance lock before they can do anything. The
* dtls_perform_handshake() function needs to start the timer
* before we stop it below.
*/
/* Before we feed data into OpenSSL ensure that the timeout timer is either stopped or completed */
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_unlock(instance);
dtls_srtp_stop_timeout_timer(instance, rtp, rtcp);
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_lock(instance);
/* If we don't yet know if we are active or passive and we receive a packet... we are obviously passive */
if (dtls->dtls_setup == AST_RTP_DTLS_SETUP_ACTPASS) {
dtls->dtls_setup = AST_RTP_DTLS_SETUP_PASSIVE;
SSL_set_accept_state(dtls->ssl);
}
BIO_write(dtls->read_bio, buf, len);
len = SSL_read(dtls->ssl, buf, len);
if ((len < 0) && (SSL_get_error(dtls->ssl, len) == SSL_ERROR_SSL)) {
unsigned long error = ERR_get_error();
ast_log(LOG_ERROR, "DTLS failure occurred on RTP instance '%p' due to reason '%s', terminating\n",
instance, ERR_reason_error_string(error));
return -1;
}
if (SSL_is_init_finished(dtls->ssl)) {
/* Any further connections will be existing since this is now established */
dtls->connection = AST_RTP_DTLS_CONNECTION_EXISTING;
/* Use the keying material to set up key/salt information */
if ((res = dtls_srtp_setup(rtp, instance, rtcp))) {
return res;
}
/* Notify that dtls has been established */
res = RTP_DTLS_ESTABLISHED;
res_rtp_asterisk: bad audio (static) due to incomplete dtls/srtp setup There was a race condition between client initiated DTLS setup, and handling of server side ice completion that caused the underlying SSL object to get cleared during DTLS initialization. If this happened Asterisk would be left in a partial DTLS setup state. RTP packets were sent and received, but were not being encrypted and decrypted. This resulted in no audio, or static. Specifically, this occurred when '__rtp_recvfrom' was processing the handshake sequence from the client to the server, and then 'ast_rtp_on_ice_complete' gets called from another thread and clears the SSL object when calling the 'dtls_perform_setup' function. The timing had to be just right in the sense that from the external SSL library perspective SSL initialization completed (rtp recv), Asterisk clears/resets the SSL object (ice done), and then checks to see if SSL is intialized (rtp recv). Since it was cleared, Asterisk thinks it is not finished, thus not completing 'dtls_srtp_setup'. This patch removes calls to 'dtls_perform_setup', which clears the SSL object, in 'ast_rtp_on_ice_complete'. When ice completes, there is no reason to clear the underlying SSL object. If an ice candidate changes a full protocol level renegotiation occurs. Also, in the case of bundled ICE candidates are reused when a stream is added. So no real reason to have to clear, and reset in this instance. Also, this patch adds a bit of extra logging to aid in diagnosis of any future problems. ASTERISK-28742 #close Change-Id: I34c9e6bad5a39b087164646e2836e3e48fe6892f
2020-02-13 21:08:10 +00:00
ast_debug_dtls(3, "(%p) DTLS - __rtp_recvfrom rtp=%p - established'\n", instance, rtp);
} else {
/* Since we've sent additional traffic start the timeout timer for retransmission */
dtls_srtp_start_timeout_timer(instance, rtp, rtcp);
}
return res;
}
#endif
#ifdef HAVE_PJPROJECT
if (!ast_sockaddr_isnull(loop) && !ast_sockaddr_cmp(loop, sa)) {
/* ICE traffic will have been handled in the TURN callback, so skip it but update the address
* so it reflects the actual source and not the loopback
*/
if (rtcp) {
ast_sockaddr_copy(sa, &rtp->rtcp->them);
} else {
ast_rtp_instance_get_remote_address(instance, sa);
}
} else if (rtp->ice) {
pj_str_t combined = pj_str(ast_sockaddr_stringify(sa));
pj_sockaddr address;
pj_status_t status;
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
struct ice_wrap *ice;
pj_thread_register_check();
pj_sockaddr_parse(pj_AF_UNSPEC(), 0, &combined, &address);
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/* Release the instance lock to avoid deadlock with PJPROJECT group lock */
ice = rtp->ice;
ao2_ref(ice, +1);
ao2_unlock(instance);
status = pj_ice_sess_on_rx_pkt(ice->real_ice,
rtcp ? AST_RTP_ICE_COMPONENT_RTCP : AST_RTP_ICE_COMPONENT_RTP,
rtcp ? TRANSPORT_SOCKET_RTCP : TRANSPORT_SOCKET_RTP, buf, len, &address,
pj_sockaddr_get_len(&address));
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_ref(ice, -1);
ao2_lock(instance);
if (status != PJ_SUCCESS) {
char err_buf[100];
pj_strerror(status, err_buf, sizeof(err_buf));
ast_log(LOG_WARNING, "PJ ICE Rx error status code: %d '%s'.\n",
(int)status, err_buf);
return -1;
}
if (!rtp->passthrough) {
/* If a unidirectional ICE negotiation occurs then lock on to the source of the
* ICE traffic and use it as the target. This will occur if the remote side only
* wants to receive media but never send to us.
*/
if (!rtp->ice_active_remote_candidates && !rtp->ice_proposed_remote_candidates) {
if (rtcp) {
ast_sockaddr_copy(&rtp->rtcp->them, sa);
} else {
ast_rtp_instance_set_remote_address(instance, sa);
}
}
return 0;
}
rtp->passthrough = 0;
}
#endif
return len;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static int rtcp_recvfrom(struct ast_rtp_instance *instance, void *buf, size_t size, int flags, struct ast_sockaddr *sa)
{
return __rtp_recvfrom(instance, buf, size, flags, sa, 1);
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static int rtp_recvfrom(struct ast_rtp_instance *instance, void *buf, size_t size, int flags, struct ast_sockaddr *sa)
{
return __rtp_recvfrom(instance, buf, size, flags, sa, 0);
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static int __rtp_sendto(struct ast_rtp_instance *instance, void *buf, size_t size, int flags, struct ast_sockaddr *sa, int rtcp, int *via_ice, int use_srtp)
{
int len = size;
void *temp = buf;
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
struct ast_rtp_instance *transport = rtp->bundled ? rtp->bundled : instance;
struct ast_rtp *transport_rtp = ast_rtp_instance_get_data(transport);
struct ast_srtp *srtp = ast_rtp_instance_get_srtp(transport, rtcp);
int res;
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
*via_ice = 0;
if (use_srtp && res_srtp && srtp && res_srtp->protect(srtp, &temp, &len, rtcp) < 0) {
return -1;
}
#ifdef HAVE_PJPROJECT
if (transport_rtp->ice) {
enum ast_rtp_ice_component_type component = rtcp ? AST_RTP_ICE_COMPONENT_RTCP : AST_RTP_ICE_COMPONENT_RTP;
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
pj_status_t status;
struct ice_wrap *ice;
/* If RTCP is sharing the same socket then use the same component */
if (rtcp && rtp->rtcp->s == rtp->s) {
component = AST_RTP_ICE_COMPONENT_RTP;
}
pj_thread_register_check();
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/* Release the instance lock to avoid deadlock with PJPROJECT group lock */
ice = transport_rtp->ice;
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_ref(ice, +1);
if (instance == transport) {
ao2_unlock(instance);
}
status = pj_ice_sess_send_data(ice->real_ice, component, temp, len);
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_ref(ice, -1);
if (instance == transport) {
ao2_lock(instance);
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
if (status == PJ_SUCCESS) {
*via_ice = 1;
return len;
}
}
#endif
res = ast_sendto(rtcp ? transport_rtp->rtcp->s : transport_rtp->s, temp, len, flags, sa);
if (res > 0) {
ast_rtp_instance_set_last_tx(instance, time(NULL));
}
return res;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static int rtcp_sendto(struct ast_rtp_instance *instance, void *buf, size_t size, int flags, struct ast_sockaddr *sa, int *ice)
{
return __rtp_sendto(instance, buf, size, flags, sa, 1, ice, 1);
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static int rtp_sendto(struct ast_rtp_instance *instance, void *buf, size_t size, int flags, struct ast_sockaddr *sa, int *ice)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
int hdrlen = 12;
int res;
if ((res = __rtp_sendto(instance, buf, size, flags, sa, 0, ice, 1)) > 0) {
rtp->txcount++;
rtp->txoctetcount += (res - hdrlen);
}
return res;
}
static unsigned int ast_rtcp_calc_interval(struct ast_rtp *rtp)
{
unsigned int interval;
/*! \todo XXX Do a more reasonable calculation on this one
* Look in RFC 3550 Section A.7 for an example*/
interval = rtcpinterval;
return interval;
}
static void calc_mean_and_standard_deviation(double new_sample, double *mean, double *std_dev, unsigned int *count)
{
double delta1;
double delta2;
/* First convert the standard deviation back into a sum of squares. */
double last_sum_of_squares = (*std_dev) * (*std_dev) * (*count ?: 1);
if (++(*count) == 0) {
/* Avoid potential divide by zero on an overflow */
*count = 1;
}
/*
* Below is an implementation of Welford's online algorithm [1] for calculating
* mean and variance in a single pass.
*
* [1] https://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
*/
delta1 = new_sample - *mean;
*mean += (delta1 / *count);
delta2 = new_sample - *mean;
/* Now calculate the new variance, and subsequent standard deviation */
*std_dev = sqrt((last_sum_of_squares + (delta1 * delta2)) / *count);
}
static int create_new_socket(const char *type, int af)
{
int sock = ast_socket_nonblock(af, SOCK_DGRAM, 0);
if (sock < 0) {
ast_log(LOG_WARNING, "Unable to allocate %s socket: %s\n", type, strerror(errno));
return sock;
}
#ifdef SO_NO_CHECK
if (nochecksums) {
setsockopt(sock, SOL_SOCKET, SO_NO_CHECK, &nochecksums, sizeof(nochecksums));
}
#endif
return sock;
}
/*!
* \internal
* \brief Initializes sequence values and probation for learning mode.
* \note This is an adaptation of pjmedia's pjmedia_rtp_seq_init function.
*
* \param info The learning information to track
* \param seq sequence number read from the rtp header to initialize the information with
*/
static void rtp_learning_seq_init(struct rtp_learning_info *info, uint16_t seq)
{
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
info->max_seq = seq;
info->packets = learning_min_sequential;
memset(&info->received, 0, sizeof(info->received));
}
/*!
* \internal
* \brief Updates sequence information for learning mode and determines if probation/learning mode should remain in effect.
* \note This function was adapted from pjmedia's pjmedia_rtp_seq_update function.
*
* \param info Structure tracking the learning progress of some address
* \param seq sequence number read from the rtp header
* \retval 0 if probation mode should exit for this address
* \retval non-zero if probation mode should continue
*/
static int rtp_learning_rtp_seq_update(struct rtp_learning_info *info, uint16_t seq)
{
if (seq == (uint16_t) (info->max_seq + 1)) {
/* packet is in sequence */
info->packets--;
} else {
/* Sequence discontinuity; reset */
info->packets = learning_min_sequential - 1;
info->received = ast_tvnow();
}
/* Only check time if strictrtp is set to yes. Otherwise, we only needed to check seqno */
if (strictrtp == STRICT_RTP_YES) {
switch (info->stream_type) {
case AST_MEDIA_TYPE_UNKNOWN:
case AST_MEDIA_TYPE_AUDIO:
/*
* Protect against packet floods by checking that we
* received the packet sequence in at least the minimum
* allowed time.
*/
if (ast_tvzero(info->received)) {
info->received = ast_tvnow();
} else if (!info->packets
&& ast_tvdiff_ms(ast_tvnow(), info->received) < learning_min_duration) {
/* Packet flood; reset */
info->packets = learning_min_sequential - 1;
info->received = ast_tvnow();
}
break;
case AST_MEDIA_TYPE_VIDEO:
case AST_MEDIA_TYPE_IMAGE:
case AST_MEDIA_TYPE_TEXT:
case AST_MEDIA_TYPE_END:
break;
}
}
info->max_seq = seq;
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
return info->packets;
}
/*!
* \brief Start the strictrtp learning mode.
*
* \param rtp RTP session description
*/
static void rtp_learning_start(struct ast_rtp *rtp)
{
rtp->strict_rtp_state = STRICT_RTP_LEARN;
memset(&rtp->rtp_source_learn.proposed_address, 0,
sizeof(rtp->rtp_source_learn.proposed_address));
rtp->rtp_source_learn.start = ast_tvnow();
rtp_learning_seq_init(&rtp->rtp_source_learn, (uint16_t) rtp->lastrxseqno);
}
#ifdef HAVE_PJPROJECT
static void acl_change_stasis_cb(void *data, struct stasis_subscription *sub, struct stasis_message *message);
/*!
* \internal
* \brief Resets and ACL to empty state.
*/
static void rtp_unload_acl(ast_rwlock_t *lock, struct ast_acl_list **acl)
{
ast_rwlock_wrlock(lock);
*acl = ast_free_acl_list(*acl);
ast_rwlock_unlock(lock);
}
/*!
* \internal
* \brief Checks an address against the ICE blacklist
* \note If there is no ice_blacklist list, always returns 0
*
* \param address The address to consider
* \retval 0 if address is not ICE blacklisted
* \retval 1 if address is ICE blacklisted
*/
static int rtp_address_is_ice_blacklisted(const struct ast_sockaddr *address)
{
int result = 0;
ast_rwlock_rdlock(&ice_acl_lock);
result |= ast_apply_acl_nolog(ice_acl, address) == AST_SENSE_DENY;
ast_rwlock_unlock(&ice_acl_lock);
return result;
}
/*!
* \internal
* \brief Checks an address against the STUN blacklist
* \since 13.16.0
*
* \note If there is no stun_blacklist list, always returns 0
*
* \param addr The address to consider
*
* \retval 0 if address is not STUN blacklisted
* \retval 1 if address is STUN blacklisted
*/
static int stun_address_is_blacklisted(const struct ast_sockaddr *addr)
{
int result = 0;
ast_rwlock_rdlock(&stun_acl_lock);
result |= ast_apply_acl_nolog(stun_acl, addr) == AST_SENSE_DENY;
ast_rwlock_unlock(&stun_acl_lock);
return result;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static void rtp_add_candidates_to_ice(struct ast_rtp_instance *instance, struct ast_rtp *rtp, struct ast_sockaddr *addr, int port, int component,
int transport)
{
unsigned int count = 0;
struct ifaddrs *ifa, *ia;
struct ast_sockaddr tmp;
pj_sockaddr pjtmp;
struct ast_ice_host_candidate *candidate;
int af_inet_ok = 0, af_inet6_ok = 0;
struct sockaddr_in stunaddr_copy;
if (ast_sockaddr_is_ipv4(addr)) {
af_inet_ok = 1;
} else if (ast_sockaddr_is_any(addr)) {
af_inet_ok = af_inet6_ok = 1;
} else {
af_inet6_ok = 1;
}
if (getifaddrs(&ifa) < 0) {
/* If we can't get addresses, we can't load ICE candidates */
ast_log(LOG_ERROR, "(%p) ICE Error obtaining list of local addresses: %s\n",
instance, strerror(errno));
} else {
ast_debug_ice(2, "(%p) ICE add system candidates\n", instance);
/* Iterate through the list of addresses obtained from the system,
* until we've iterated through all of them, or accepted
* PJ_ICE_MAX_CAND candidates */
for (ia = ifa; ia && count < PJ_ICE_MAX_CAND; ia = ia->ifa_next) {
/* Interface is either not UP or doesn't have an address assigned,
* eg, a ppp that just completed LCP but no IPCP yet */
if (!ia->ifa_addr || (ia->ifa_flags & IFF_UP) == 0) {
continue;
}
/* Filter out non-IPvX addresses, eg, link-layer */
if (ia->ifa_addr->sa_family != AF_INET && ia->ifa_addr->sa_family != AF_INET6) {
continue;
}
ast_sockaddr_from_sockaddr(&tmp, ia->ifa_addr);
if (ia->ifa_addr->sa_family == AF_INET) {
const struct sockaddr_in *sa_in = (struct sockaddr_in*)ia->ifa_addr;
if (!af_inet_ok) {
continue;
}
/* Skip 127.0.0.0/8 (loopback) */
/* Don't use IFF_LOOPBACK check since one could assign usable
* publics to the loopback */
if ((sa_in->sin_addr.s_addr & htonl(0xFF000000)) == htonl(0x7F000000)) {
continue;
}
/* Skip 0.0.0.0/8 based on RFC1122, and from pjproject */
if ((sa_in->sin_addr.s_addr & htonl(0xFF000000)) == 0) {
continue;
}
} else { /* ia->ifa_addr->sa_family == AF_INET6 */
if (!af_inet6_ok) {
continue;
}
/* Filter ::1 */
if (!ast_sockaddr_cmp_addr(&lo6, &tmp)) {
continue;
}
}
/* Pull in the host candidates from [ice_host_candidates] */
AST_RWLIST_RDLOCK(&host_candidates);
AST_LIST_TRAVERSE(&host_candidates, candidate, next) {
if (!ast_sockaddr_cmp(&candidate->local, &tmp)) {
/* candidate->local matches actual assigned, so check if
* advertised is blacklisted, if not, add it to the
* advertised list. Not that it would make sense to remap
* a local address to a blacklisted address, but honour it
* anyway. */
if (!rtp_address_is_ice_blacklisted(&candidate->advertised)) {
ast_sockaddr_to_pj_sockaddr(&candidate->advertised, &pjtmp);
pj_sockaddr_set_port(&pjtmp, port);
ast_rtp_ice_add_cand(instance, rtp, component, transport,
PJ_ICE_CAND_TYPE_HOST, 65535, &pjtmp, &pjtmp, NULL,
pj_sockaddr_get_len(&pjtmp));
++count;
}
if (!candidate->include_local) {
/* We don't want to advertise the actual address */
ast_sockaddr_setnull(&tmp);
}
break;
}
}
AST_RWLIST_UNLOCK(&host_candidates);
/* we had an entry in [ice_host_candidates] that matched, and
* didn't have include_local_address set. Alternatively, adding
* that match resulted in us going to PJ_ICE_MAX_CAND */
if (ast_sockaddr_isnull(&tmp) || count == PJ_ICE_MAX_CAND) {
continue;
}
if (rtp_address_is_ice_blacklisted(&tmp)) {
continue;
}
ast_sockaddr_to_pj_sockaddr(&tmp, &pjtmp);
pj_sockaddr_set_port(&pjtmp, port);
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ast_rtp_ice_add_cand(instance, rtp, component, transport,
PJ_ICE_CAND_TYPE_HOST, 65535, &pjtmp, &pjtmp, NULL,
pj_sockaddr_get_len(&pjtmp));
++count;
}
freeifaddrs(ifa);
}
ast_rwlock_rdlock(&stunaddr_lock);
memcpy(&stunaddr_copy, &stunaddr, sizeof(stunaddr));
ast_rwlock_unlock(&stunaddr_lock);
/* If configured to use a STUN server to get our external mapped address do so */
if (stunaddr_copy.sin_addr.s_addr && !stun_address_is_blacklisted(addr) &&
(ast_sockaddr_is_ipv4(addr) || ast_sockaddr_is_any(addr)) &&
count < PJ_ICE_MAX_CAND) {
struct sockaddr_in answer;
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
int rsp;
ast_debug_category(3, AST_DEBUG_CATEGORY_ICE | AST_DEBUG_CATEGORY_STUN,
"(%p) ICE request STUN %s %s candidate\n", instance,
transport == AST_TRANSPORT_UDP ? "UDP" : "TCP",
component == AST_RTP_ICE_COMPONENT_RTP ? "RTP" : "RTCP");
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*
* The instance should not be locked because we can block
* waiting for a STUN respone.
*/
ao2_unlock(instance);
rsp = ast_stun_request(component == AST_RTP_ICE_COMPONENT_RTCP
? rtp->rtcp->s : rtp->s, &stunaddr_copy, NULL, &answer);
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_lock(instance);
if (!rsp) {
struct ast_rtp_engine_ice_candidate *candidate;
pj_sockaddr ext, base;
pj_str_t mapped = pj_str(ast_strdupa(ast_inet_ntoa(answer.sin_addr)));
int srflx = 1, baseset = 0;
struct ao2_iterator i;
pj_sockaddr_init(pj_AF_INET(), &ext, &mapped, ntohs(answer.sin_port));
/*
* If the returned address is the same as one of our host
* candidates, don't send the srflx. At the same time,
* we need to set the base address (raddr).
*/
i = ao2_iterator_init(rtp->ice_local_candidates, 0);
while (srflx && (candidate = ao2_iterator_next(&i))) {
if (!baseset && ast_sockaddr_is_ipv4(&candidate->address)) {
baseset = 1;
ast_sockaddr_to_pj_sockaddr(&candidate->address, &base);
}
if (!pj_sockaddr_cmp(&candidate->address, &ext)) {
srflx = 0;
}
ao2_ref(candidate, -1);
}
ao2_iterator_destroy(&i);
if (srflx && baseset) {
pj_sockaddr_set_port(&base, port);
ast_rtp_ice_add_cand(instance, rtp, component, transport,
PJ_ICE_CAND_TYPE_SRFLX, 65535, &ext, &base, &base,
pj_sockaddr_get_len(&ext));
}
}
}
/* If configured to use a TURN relay create a session and allocate */
if (pj_strlen(&turnaddr)) {
ast_rtp_ice_turn_request(instance, component, AST_TRANSPORT_TCP, pj_strbuf(&turnaddr), turnport,
pj_strbuf(&turnusername), pj_strbuf(&turnpassword));
}
}
#endif
/*!
* \internal
* \brief Calculates the elapsed time from issue of the first tx packet in an
* rtp session and a specified time
*
* \param rtp pointer to the rtp struct with the transmitted rtp packet
* \param delivery time of delivery - if NULL or zero value, will be ast_tvnow()
*
* \return time elapsed in milliseconds
*/
static unsigned int calc_txstamp(struct ast_rtp *rtp, struct timeval *delivery)
{
struct timeval t;
long ms;
if (ast_tvzero(rtp->txcore)) {
rtp->txcore = ast_tvnow();
rtp->txcore.tv_usec -= rtp->txcore.tv_usec % 20000;
}
t = (delivery && !ast_tvzero(*delivery)) ? *delivery : ast_tvnow();
if ((ms = ast_tvdiff_ms(t, rtp->txcore)) < 0) {
ms = 0;
}
rtp->txcore = t;
return (unsigned int) ms;
}
#ifdef HAVE_PJPROJECT
/*!
* \internal
* \brief Creates an ICE session. Can be used to replace a destroyed ICE session.
*
* \param instance RTP instance for which the ICE session is being replaced
* \param addr ast_sockaddr to use for adding RTP candidates to the ICE session
* \param port port to use for adding RTP candidates to the ICE session
* \param replace 0 when creating a new session, 1 when replacing a destroyed session
*
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
* \pre instance is locked
*
* \retval 0 on success
* \retval -1 on failure
*/
static int ice_create(struct ast_rtp_instance *instance, struct ast_sockaddr *addr,
int port, int replace)
{
pj_stun_config stun_config;
pj_str_t ufrag, passwd;
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
pj_status_t status;
struct ice_wrap *ice_old;
struct ice_wrap *ice;
pj_ice_sess *real_ice = NULL;
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
ao2_cleanup(rtp->ice_local_candidates);
rtp->ice_local_candidates = NULL;
ast_debug_ice(2, "(%p) ICE create%s\n", instance, replace ? " and replace" : "");
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ice = ao2_alloc_options(sizeof(*ice), ice_wrap_dtor, AO2_ALLOC_OPT_LOCK_NOLOCK);
if (!ice) {
ast_rtp_ice_stop(instance);
return -1;
}
pj_thread_register_check();
pj_stun_config_init(&stun_config, &cachingpool.factory, 0, NULL, timer_heap);
if (!stun_software_attribute) {
stun_config.software_name = pj_str(NULL);
}
ufrag = pj_str(rtp->local_ufrag);
passwd = pj_str(rtp->local_passwd);
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/* Release the instance lock to avoid deadlock with PJPROJECT group lock */
ao2_unlock(instance);
/* Create an ICE session for ICE negotiation */
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
status = pj_ice_sess_create(&stun_config, NULL, PJ_ICE_SESS_ROLE_UNKNOWN,
rtp->ice_num_components, &ast_rtp_ice_sess_cb, &ufrag, &passwd, NULL, &real_ice);
ao2_lock(instance);
if (status == PJ_SUCCESS) {
/* Safely complete linking the ICE session into the instance */
real_ice->user_data = instance;
ice->real_ice = real_ice;
ice_old = rtp->ice;
rtp->ice = ice;
if (ice_old) {
ao2_unlock(instance);
ao2_ref(ice_old, -1);
ao2_lock(instance);
}
/* Add all of the available candidates to the ICE session */
rtp_add_candidates_to_ice(instance, rtp, addr, port, AST_RTP_ICE_COMPONENT_RTP,
TRANSPORT_SOCKET_RTP);
/* Only add the RTCP candidates to ICE when replacing the session and if
* the ICE session contains more than just an RTP component. New sessions
* handle this in a separate part of the setup phase */
if (replace && rtp->rtcp && rtp->ice_num_components > 1) {
rtp_add_candidates_to_ice(instance, rtp, &rtp->rtcp->us,
ast_sockaddr_port(&rtp->rtcp->us), AST_RTP_ICE_COMPONENT_RTCP,
TRANSPORT_SOCKET_RTCP);
}
return 0;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*
* It is safe to unref this while instance is locked here.
* It was not initialized with a real_ice pointer.
*/
ao2_ref(ice, -1);
ast_rtp_ice_stop(instance);
return -1;
}
#endif
static int rtp_allocate_transport(struct ast_rtp_instance *instance, struct ast_rtp *rtp)
{
int x, startplace, i, maxloops;
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
rtp->strict_rtp_state = (strictrtp ? STRICT_RTP_CLOSED : STRICT_RTP_OPEN);
/* Create a new socket for us to listen on and use */
if ((rtp->s =
create_new_socket("RTP",
ast_sockaddr_is_ipv4(&rtp->bind_address) ? AF_INET :
ast_sockaddr_is_ipv6(&rtp->bind_address) ? AF_INET6 : -1)) < 0) {
ast_log(LOG_WARNING, "Failed to create a new socket for RTP instance '%p'\n", instance);
return -1;
}
/* Now actually find a free RTP port to use */
x = (ast_random() % (rtpend - rtpstart)) + rtpstart;
x = x & ~1;
startplace = x;
/* Protection against infinite loops in the case there is a potential case where the loop is not broken such as an odd
start port sneaking in (even though this condition is checked at load.) */
maxloops = rtpend - rtpstart;
for (i = 0; i <= maxloops; i++) {
ast_sockaddr_set_port(&rtp->bind_address, x);
/* Try to bind, this will tell us whether the port is available or not */
if (!ast_bind(rtp->s, &rtp->bind_address)) {
ast_debug_rtp(1, "(%p) RTP allocated port %d\n", instance, x);
ast_rtp_instance_set_local_address(instance, &rtp->bind_address);
ast_test_suite_event_notify("RTP_PORT_ALLOCATED", "Port: %d", x);
break;
}
x += 2;
if (x > rtpend) {
x = (rtpstart + 1) & ~1;
}
/* See if we ran out of ports or if the bind actually failed because of something other than the address being in use */
if (x == startplace || (errno != EADDRINUSE && errno != EACCES)) {
ast_log(LOG_ERROR, "Oh dear... we couldn't allocate a port for RTP instance '%p'\n", instance);
Fix a variety of memory leaks This patch addresses a number of memory leaks in a variety of modules that were found by a static analysis tool. A brief summary of the changes: * app_minivm: free ast_str objects on off nominal paths * app_page: free the ast_dial object if the requested channel technology cannot be appended to the dialing structure * app_queue: if a penalty rule failed to match any existing rule list names, the created rule would not be inserted and its memory would be leaked * app_read: dispose of the created silence detector in the presence of off nominal circumstances * app_voicemail: dispose of an allocated unique ID field for MWI event un-subscribe requests in off nominal paths; dispose of configuration objects when using the secret.conf option * chan_dahdi: dispose of the allocated frame produced by ast_dsp_process * chan_iax2: properly unref peer in CLI command "iax2 unregister" * chan_sip: dispose of the allocated frame produced by sip_rtp_read's call of ast_dsp_process; free memory in parse unit tests * func_dialgroup: properly deref ao2 object grhead in nominal path of dialgroup_read * func_odbc: free resultset in off nominal paths of odbc_read * cli: free match_list in off nominal paths of CLI match completion * config: free comment_buffer/list_buffer when configuration file load is unchanged; free the same buffers any time they were created and config files were processed * data: free XML nodes in various places * enum: free context buffer in off nominal paths * features: free ast_call_feature in off nominal paths of applicationmap config processing * netsock2: users of ast_sockaddr_resolve pass in an ast_sockaddr struct that is allocated by the method. Failures in ast_sockaddr_resolve could result in the users of the method not knowing whether or not the buffer was allocated. The method will now not allocate the ast_sockaddr struct if it will return failure. * pbx: cleanup hash table traversals in off nominal paths; free ignore pattern buffer if it already exists for the specified context * xmldoc: cleanup various nodes when we no longer need them * main/editline: various cleanup of pointers not being freed before being assigned to other memory, cleanup along off nominal paths * menuselect/mxml: cleanup of value buffer for an attribute when that attribute did not specify a value * res_calendar*: responses are allocated via the various *_request method returns and should not be allocated in the various write_event methods; ensure attendee buffer is freed if no data exists in the parsed node; ensure that calendar objects are de-ref'd appropriately * res_jabber: free buffer in off nominal path * res_musiconhold: close the DIR* object in off nominal paths * res_rtp_asterisk: if we run out of ports, close the rtp socket object and free the rtp object * res_srtp: if we fail to create the session in libsrtp, destroy the temporary ast_srtp object (issue ASTERISK-19665) Reported by: Matt Jordan Review: https://reviewboard.asterisk.org/r/1922 ........ Merged revisions 366880 from http://svn.asterisk.org/svn/asterisk/branches/1.8 ........ Merged revisions 366881 from http://svn.asterisk.org/svn/asterisk/branches/10 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@366917 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2012-05-18 14:43:44 +00:00
close(rtp->s);
rtp->s = -1;
return -1;
}
}
#ifdef HAVE_PJPROJECT
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/* Initialize synchronization aspects */
ast_cond_init(&rtp->cond, NULL);
generate_random_string(rtp->local_ufrag, sizeof(rtp->local_ufrag));
generate_random_string(rtp->local_passwd, sizeof(rtp->local_passwd));
/* Create an ICE session for ICE negotiation */
if (icesupport) {
rtp->ice_num_components = 2;
ast_debug_ice(2, "(%p) ICE creating session %s (%d)\n", instance,
ast_sockaddr_stringify(&rtp->bind_address), x);
if (ice_create(instance, &rtp->bind_address, x, 0)) {
ast_log(LOG_NOTICE, "(%p) ICE failed to create session\n", instance);
} else {
rtp->ice_port = x;
ast_sockaddr_copy(&rtp->ice_original_rtp_addr, &rtp->bind_address);
}
}
#endif
#if defined(HAVE_OPENSSL) && (OPENSSL_VERSION_NUMBER >= 0x10001000L) && !defined(OPENSSL_NO_SRTP)
rtp->rekeyid = -1;
rtp->dtls.timeout_timer = -1;
#endif
return 0;
}
static void rtp_deallocate_transport(struct ast_rtp_instance *instance, struct ast_rtp *rtp)
{
int saved_rtp_s = rtp->s;
#ifdef HAVE_PJPROJECT
struct timeval wait = ast_tvadd(ast_tvnow(), ast_samp2tv(TURN_STATE_WAIT_TIME, 1000));
struct timespec ts = { .tv_sec = wait.tv_sec, .tv_nsec = wait.tv_usec * 1000, };
#endif
#if defined(HAVE_OPENSSL) && (OPENSSL_VERSION_NUMBER >= 0x10001000L) && !defined(OPENSSL_NO_SRTP)
ast_rtp_dtls_stop(instance);
#endif
/* Close our own socket so we no longer get packets */
if (rtp->s > -1) {
close(rtp->s);
rtp->s = -1;
}
/* Destroy RTCP if it was being used */
if (rtp->rtcp && rtp->rtcp->s > -1) {
if (saved_rtp_s != rtp->rtcp->s) {
close(rtp->rtcp->s);
}
rtp->rtcp->s = -1;
}
#ifdef HAVE_PJPROJECT
pj_thread_register_check();
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*
* The instance lock is already held.
*
* Destroy the RTP TURN relay if being used
*/
if (rtp->turn_rtp) {
rtp->turn_state = PJ_TURN_STATE_NULL;
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/* Release the instance lock to avoid deadlock with PJPROJECT group lock */
ao2_unlock(instance);
pj_turn_sock_destroy(rtp->turn_rtp);
ao2_lock(instance);
while (rtp->turn_state != PJ_TURN_STATE_DESTROYING) {
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ast_cond_timedwait(&rtp->cond, ao2_object_get_lockaddr(instance), &ts);
}
rtp->turn_rtp = NULL;
}
/* Destroy the RTCP TURN relay if being used */
if (rtp->turn_rtcp) {
rtp->turn_state = PJ_TURN_STATE_NULL;
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/* Release the instance lock to avoid deadlock with PJPROJECT group lock */
ao2_unlock(instance);
pj_turn_sock_destroy(rtp->turn_rtcp);
ao2_lock(instance);
while (rtp->turn_state != PJ_TURN_STATE_DESTROYING) {
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ast_cond_timedwait(&rtp->cond, ao2_object_get_lockaddr(instance), &ts);
}
rtp->turn_rtcp = NULL;
}
ast_debug_ice(2, "(%p) ICE RTP transport deallocating\n", instance);
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/* Destroy any ICE session */
ast_rtp_ice_stop(instance);
/* Destroy any candidates */
if (rtp->ice_local_candidates) {
ao2_ref(rtp->ice_local_candidates, -1);
rtp->ice_local_candidates = NULL;
}
if (rtp->ice_active_remote_candidates) {
ao2_ref(rtp->ice_active_remote_candidates, -1);
rtp->ice_active_remote_candidates = NULL;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
if (rtp->ice_proposed_remote_candidates) {
ao2_ref(rtp->ice_proposed_remote_candidates, -1);
rtp->ice_proposed_remote_candidates = NULL;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
if (rtp->ioqueue) {
/*
* We cannot hold the instance lock because we could wait
* for the ioqueue thread to die and we might deadlock as
* a result.
*/
ao2_unlock(instance);
rtp_ioqueue_thread_remove(rtp->ioqueue);
ao2_lock(instance);
rtp->ioqueue = NULL;
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
}
#endif
}
/*! \pre instance is locked */
static int ast_rtp_new(struct ast_rtp_instance *instance,
struct ast_sched_context *sched, struct ast_sockaddr *addr,
void *data)
{
struct ast_rtp *rtp = NULL;
/* Create a new RTP structure to hold all of our data */
if (!(rtp = ast_calloc(1, sizeof(*rtp)))) {
return -1;
}
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
rtp->owner = instance;
/* Set default parameters on the newly created RTP structure */
rtp->ssrc = ast_random();
ast_uuid_generate_str(rtp->cname, sizeof(rtp->cname));
rtp->seqno = ast_random() & 0x7fff;
2018-05-10 18:11:06 +00:00
rtp->expectedrxseqno = -1;
rtp->expectedseqno = -1;
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
rtp->rxstart = -1;
rtp->sched = sched;
ast_sockaddr_copy(&rtp->bind_address, addr);
/* Transport creation operations can grab the RTP data from the instance, so set it */
ast_rtp_instance_set_data(instance, rtp);
if (rtp_allocate_transport(instance, rtp)) {
return -1;
}
if (AST_VECTOR_INIT(&rtp->ssrc_mapping, 1)) {
return -1;
}
if (AST_VECTOR_INIT(&rtp->transport_wide_cc.packet_statistics, 0)) {
return -1;
}
rtp->transport_wide_cc.schedid = -1;
rtp->f.subclass.format = ao2_bump(ast_format_none);
rtp->lastrxformat = ao2_bump(ast_format_none);
rtp->lasttxformat = ao2_bump(ast_format_none);
rtp->stream_num = -1;
return 0;
}
/*!
* \brief SSRC mapping comparator for AST_VECTOR_REMOVE_CMP_UNORDERED()
*
* \param elem Element to compare against
* \param value Value to compare with the vector element.
*
* \retval 0 if element does not match.
* \retval Non-zero if element matches.
*/
#define SSRC_MAPPING_ELEM_CMP(elem, value) ((elem).instance == (value))
/*! \pre instance is locked */
static int ast_rtp_destroy(struct ast_rtp_instance *instance)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
if (rtp->bundled) {
struct ast_rtp *bundled_rtp;
/* We can't hold our instance lock while removing ourselves from the parent */
ao2_unlock(instance);
ao2_lock(rtp->bundled);
bundled_rtp = ast_rtp_instance_get_data(rtp->bundled);
AST_VECTOR_REMOVE_CMP_UNORDERED(&bundled_rtp->ssrc_mapping, instance, SSRC_MAPPING_ELEM_CMP, AST_VECTOR_ELEM_CLEANUP_NOOP);
ao2_unlock(rtp->bundled);
ao2_lock(instance);
ao2_ref(rtp->bundled, -1);
}
rtp_deallocate_transport(instance, rtp);
/* Destroy the smoother that was smoothing out audio if present */
if (rtp->smoother) {
ast_smoother_free(rtp->smoother);
}
/* Destroy RTCP if it was being used */
if (rtp->rtcp) {
/*
* It is not possible for there to be an active RTCP scheduler
* entry at this point since it holds a reference to the
* RTP instance while it's active.
*/
ast_free(rtp->rtcp->local_addr_str);
ast_free(rtp->rtcp);
}
/* Destroy RED if it was being used */
if (rtp->red) {
ao2_unlock(instance);
AST_SCHED_DEL(rtp->sched, rtp->red->schedid);
ao2_lock(instance);
ast_free(rtp->red);
rtp->red = NULL;
}
/* Destroy the send buffer if it was being used */
if (rtp->send_buffer) {
ast_data_buffer_free(rtp->send_buffer);
}
2018-05-10 18:11:06 +00:00
/* Destroy the recv buffer if it was being used */
if (rtp->recv_buffer) {
ast_data_buffer_free(rtp->recv_buffer);
}
AST_VECTOR_FREE(&rtp->transport_wide_cc.packet_statistics);
media formats: re-architect handling of media for performance improvements In the old times media formats were represented using a bit field. This was fast but had a few limitations. 1. Asterisk was limited in how many formats it could handle. 2. Formats, being a bit field, could not include any attribute information. A format was strictly its type, e.g., "this is ulaw". This was changed in Asterisk 10 (see https://wiki.asterisk.org/wiki/display/AST/Media+Architecture+Proposal for notes on that work) which led to the creation of the ast_format structure. This structure allowed Asterisk to handle attributes and bundle information with a format. Additionally, ast_format_cap was created to act as a container for multiple formats that, together, formed the capability of some entity. Another mechanism was added to allow logic to be registered which performed format attribute negotiation. Everywhere throughout the codebase Asterisk was changed to use this strategy. Unfortunately, in software, there is no free lunch. These new capabilities came at a cost. Performance analysis and profiling showed that we spend an inordinate amount of time comparing, copying, and generally manipulating formats and their related structures. Basic prototyping has shown that a reasonably large performance improvement could be made in this area. This patch is the result of that project, which overhauled the media format architecture and its usage in Asterisk to improve performance. Generally, the new philosophy for handling formats is as follows: * The ast_format structure is reference counted. This removed a large amount of the memory allocations and copying that was done in prior versions. * In order to prevent race conditions while keeping things performant, the ast_format structure is immutable by convention and lock-free. Violate this tenet at your peril! * Because formats are reference counted, codecs are also reference counted. The Asterisk core generally provides built-in codecs and caches the ast_format structures created to represent them. Generally, to prevent inordinate amounts of module reference bumping, codecs and formats can be added at run-time but cannot be removed. * All compatibility with the bit field representation of codecs/formats has been moved to a compatibility API. The primary user of this representation is chan_iax2, which must continue to maintain its bit-field usage of formats for interoperability concerns. * When a format is negotiated with attributes, or when a format cannot be represented by one of the cached formats, a new format object is created or cloned from an existing format. That format may have the same codec underlying it, but is a different format than a version of the format with different attributes or without attributes. * While formats are reference counted objects, the reference count maintained on the format should be manipulated with care. Formats are generally cached and will persist for the lifetime of Asterisk and do not explicitly need to have their lifetime modified. An exception to this is when the user of a format does not know where the format came from *and* the user may outlive the provider of the format. This occurs, for example, when a format is read from a channel: the channel may have a format with attributes (hence, non-cached) and the user of the format may last longer than the channel (if the reference to the channel is released prior to the format's reference). For more information on this work, see the API design notes: https://wiki.asterisk.org/wiki/display/AST/Media+Format+Rewrite Finally, this work was the culmination of a large number of developer's efforts. Extra thanks goes to Corey Farrell, who took on a large amount of the work in the Asterisk core, chan_sip, and was an invaluable resource in peer reviews throughout this project. There were a substantial number of patches contributed during this work; the following issues/patch names simply reflect some of the work (and will cause the release scripts to give attribution to the individuals who work on them). Reviews: https://reviewboard.asterisk.org/r/3814 https://reviewboard.asterisk.org/r/3808 https://reviewboard.asterisk.org/r/3805 https://reviewboard.asterisk.org/r/3803 https://reviewboard.asterisk.org/r/3801 https://reviewboard.asterisk.org/r/3798 https://reviewboard.asterisk.org/r/3800 https://reviewboard.asterisk.org/r/3794 https://reviewboard.asterisk.org/r/3793 https://reviewboard.asterisk.org/r/3792 https://reviewboard.asterisk.org/r/3791 https://reviewboard.asterisk.org/r/3790 https://reviewboard.asterisk.org/r/3789 https://reviewboard.asterisk.org/r/3788 https://reviewboard.asterisk.org/r/3787 https://reviewboard.asterisk.org/r/3786 https://reviewboard.asterisk.org/r/3784 https://reviewboard.asterisk.org/r/3783 https://reviewboard.asterisk.org/r/3778 https://reviewboard.asterisk.org/r/3774 https://reviewboard.asterisk.org/r/3775 https://reviewboard.asterisk.org/r/3772 https://reviewboard.asterisk.org/r/3761 https://reviewboard.asterisk.org/r/3754 https://reviewboard.asterisk.org/r/3753 https://reviewboard.asterisk.org/r/3751 https://reviewboard.asterisk.org/r/3750 https://reviewboard.asterisk.org/r/3748 https://reviewboard.asterisk.org/r/3747 https://reviewboard.asterisk.org/r/3746 https://reviewboard.asterisk.org/r/3742 https://reviewboard.asterisk.org/r/3740 https://reviewboard.asterisk.org/r/3739 https://reviewboard.asterisk.org/r/3738 https://reviewboard.asterisk.org/r/3737 https://reviewboard.asterisk.org/r/3736 https://reviewboard.asterisk.org/r/3734 https://reviewboard.asterisk.org/r/3722 https://reviewboard.asterisk.org/r/3713 https://reviewboard.asterisk.org/r/3703 https://reviewboard.asterisk.org/r/3689 https://reviewboard.asterisk.org/r/3687 https://reviewboard.asterisk.org/r/3674 https://reviewboard.asterisk.org/r/3671 https://reviewboard.asterisk.org/r/3667 https://reviewboard.asterisk.org/r/3665 https://reviewboard.asterisk.org/r/3625 https://reviewboard.asterisk.org/r/3602 https://reviewboard.asterisk.org/r/3519 https://reviewboard.asterisk.org/r/3518 https://reviewboard.asterisk.org/r/3516 https://reviewboard.asterisk.org/r/3515 https://reviewboard.asterisk.org/r/3512 https://reviewboard.asterisk.org/r/3506 https://reviewboard.asterisk.org/r/3413 https://reviewboard.asterisk.org/r/3410 https://reviewboard.asterisk.org/r/3387 https://reviewboard.asterisk.org/r/3388 https://reviewboard.asterisk.org/r/3389 https://reviewboard.asterisk.org/r/3390 https://reviewboard.asterisk.org/r/3321 https://reviewboard.asterisk.org/r/3320 https://reviewboard.asterisk.org/r/3319 https://reviewboard.asterisk.org/r/3318 https://reviewboard.asterisk.org/r/3266 https://reviewboard.asterisk.org/r/3265 https://reviewboard.asterisk.org/r/3234 https://reviewboard.asterisk.org/r/3178 ASTERISK-23114 #close Reported by: mjordan media_formats_translation_core.diff uploaded by kharwell (License 6464) rb3506.diff uploaded by mjordan (License 6283) media_format_app_file.diff uploaded by kharwell (License 6464) misc-2.diff uploaded by file (License 5000) chan_mild-3.diff uploaded by file (License 5000) chan_obscure.diff uploaded by file (License 5000) jingle.diff uploaded by file (License 5000) funcs.diff uploaded by file (License 5000) formats.diff uploaded by file (License 5000) core.diff uploaded by file (License 5000) bridges.diff uploaded by file (License 5000) mf-codecs-2.diff uploaded by file (License 5000) mf-app_fax.diff uploaded by file (License 5000) mf-apps-3.diff uploaded by file (License 5000) media-formats-3.diff uploaded by file (License 5000) ASTERISK-23715 rb3713.patch uploaded by coreyfarrell (License 5909) rb3689.patch uploaded by mjordan (License 6283) ASTERISK-23957 rb3722.patch uploaded by mjordan (License 6283) mf-attributes-3.diff uploaded by file (License 5000) ASTERISK-23958 Tested by: jrose rb3822.patch uploaded by coreyfarrell (License 5909) rb3800.patch uploaded by jrose (License 6182) chan_sip.diff uploaded by mjordan (License 6283) rb3747.patch uploaded by jrose (License 6182) ASTERISK-23959 #close Tested by: sgriepentrog, mjordan, coreyfarrell sip_cleanup.diff uploaded by opticron (License 6273) chan_sip_caps.diff uploaded by mjordan (License 6283) rb3751.patch uploaded by coreyfarrell (License 5909) chan_sip-3.diff uploaded by file (License 5000) ASTERISK-23960 #close Tested by: opticron direct_media.diff uploaded by opticron (License 6273) pjsip-direct-media.diff uploaded by file (License 5000) format_cap_remove.diff uploaded by opticron (License 6273) media_format_fixes.diff uploaded by opticron (License 6273) chan_pjsip-2.diff uploaded by file (License 5000) ASTERISK-23966 #close Tested by: rmudgett rb3803.patch uploaded by rmudgetti (License 5621) chan_dahdi.diff uploaded by file (License 5000) ASTERISK-24064 #close Tested by: coreyfarrell, mjordan, opticron, file, rmudgett, sgriepentrog, jrose rb3814.patch uploaded by rmudgett (License 5621) moh_cleanup.diff uploaded by opticron (License 6273) bridge_leak.diff uploaded by opticron (License 6273) translate.diff uploaded by file (License 5000) rb3795.patch uploaded by rmudgett (License 5621) tls_fix.diff uploaded by mjordan (License 6283) fax-mf-fix-2.diff uploaded by file (License 5000) rtp_transfer_stuff uploaded by mjordan (License 6283) rb3787.patch uploaded by rmudgett (License 5621) media-formats-explicit-translate-format-3.diff uploaded by file (License 5000) format_cache_case_fix.diff uploaded by opticron (License 6273) rb3774.patch uploaded by rmudgett (License 5621) rb3775.patch uploaded by rmudgett (License 5621) rtp_engine_fix.diff uploaded by opticron (License 6273) rtp_crash_fix.diff uploaded by opticron (License 6273) rb3753.patch uploaded by mjordan (License 6283) rb3750.patch uploaded by mjordan (License 6283) rb3748.patch uploaded by rmudgett (License 5621) media_format_fixes.diff uploaded by opticron (License 6273) rb3740.patch uploaded by mjordan (License 6283) rb3739.patch uploaded by mjordan (License 6283) rb3734.patch uploaded by mjordan (License 6283) rb3689.patch uploaded by mjordan (License 6283) rb3674.patch uploaded by coreyfarrell (License 5909) rb3671.patch uploaded by coreyfarrell (License 5909) rb3667.patch uploaded by coreyfarrell (License 5909) rb3665.patch uploaded by mjordan (License 6283) rb3625.patch uploaded by coreyfarrell (License 5909) rb3602.patch uploaded by coreyfarrell (License 5909) format_compatibility-2.diff uploaded by file (License 5000) core.diff uploaded by file (License 5000) git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@419044 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-07-20 22:06:33 +00:00
ao2_cleanup(rtp->lasttxformat);
ao2_cleanup(rtp->lastrxformat);
ao2_cleanup(rtp->f.subclass.format);
AST_VECTOR_FREE(&rtp->ssrc_mapping);
2018-05-10 18:11:06 +00:00
AST_VECTOR_FREE(&rtp->missing_seqno);
/* Finally destroy ourselves */
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
rtp->owner = NULL;
ast_free(rtp);
return 0;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static int ast_rtp_dtmf_mode_set(struct ast_rtp_instance *instance, enum ast_rtp_dtmf_mode dtmf_mode)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
rtp->dtmfmode = dtmf_mode;
return 0;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static enum ast_rtp_dtmf_mode ast_rtp_dtmf_mode_get(struct ast_rtp_instance *instance)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
return rtp->dtmfmode;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static int ast_rtp_dtmf_begin(struct ast_rtp_instance *instance, char digit)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
struct ast_sockaddr remote_address = { {0,} };
int hdrlen = 12, res = 0, i = 0, payload = 101;
char data[256];
unsigned int *rtpheader = (unsigned int*)data;
ast_rtp_instance_get_remote_address(instance, &remote_address);
/* If we have no remote address information bail out now */
if (ast_sockaddr_isnull(&remote_address)) {
return -1;
}
/* Convert given digit into what we want to transmit */
if ((digit <= '9') && (digit >= '0')) {
digit -= '0';
} else if (digit == '*') {
digit = 10;
} else if (digit == '#') {
digit = 11;
} else if ((digit >= 'A') && (digit <= 'D')) {
digit = digit - 'A' + 12;
} else if ((digit >= 'a') && (digit <= 'd')) {
digit = digit - 'a' + 12;
} else {
ast_log(LOG_WARNING, "Don't know how to represent '%c'\n", digit);
return -1;
}
/* Grab the payload that they expect the RFC2833 packet to be received in */
rtp_engine.c: Initial split of payload types into rx and tx mappings. There are numerous problems with the current implementation of the RTP payload type mapping in Asterisk. It uses only one mapping structure to associate payload types to codecs. The single mapping is overkill if all of the payload type values are well known values. Dynamic payload type mappings do not work as well with the single mapping because RFC3264 allows each side of the link to negotiate different dynamic mappings for what they want to receive. Not only could you have the same codec mapped for sending and receiving on different payload types you could wind up with the same payload type mapped to different codecs for each direction. 1) An independent payload type mapping is needed for sending and receiving. 2) The receive mapping needs to keep track of previous mappings because of the slack to when negotiation happens and current packets in flight using the old mapping arrive. 3) The transmit mapping only needs to keep track of the current negotiated values since we are sending the packets and know when the switchover takes place. * Needed to create ast_rtp_codecs_payload_code_tx() and make some callers use the new function because ast_rtp_codecs_payload_code() was used for mappings in both directions. * Needed to create ast_rtp_codecs_payloads_xover() for cases where we need to pass preferred codec mappings to the peer channel for early media bridging or when we need to prefer the offered mapping that RFC3264 says we SHOULD use. * ast_rtp_codecs_payloads_xover() and ast_rtp_codecs_payload_code_tx() are the only new public functions created. All the others were only used for the tx or rx mapping direction so the function doxygen now reflects which direction the function operates. * chan_mgcp.c: Removed call to ast_rtp_codecs_payloads_clear() as doing that makes no sense when processing an incoming SDP. We would be wiping out any mappings that we set for the possible outgoing SDP we sent earlier. ASTERISK-25166 Reported by: Kevin Harwell ASTERISK-17410 Reported by: Boris Fox Change-Id: Iaf6c227bca68cb7c414cf2fd4108a8ac98bd45ac
2015-07-24 00:24:04 +00:00
payload = ast_rtp_codecs_payload_code_tx(ast_rtp_instance_get_codecs(instance), 0, NULL, AST_RTP_DTMF);
rtp->dtmfmute = ast_tvadd(ast_tvnow(), ast_tv(0, 500000));
rtp->send_duration = 160;
rtp->lastts += calc_txstamp(rtp, NULL) * DTMF_SAMPLE_RATE_MS;
rtp->lastdigitts = rtp->lastts + rtp->send_duration;
/* Create the actual packet that we will be sending */
rtpheader[0] = htonl((2 << 30) | (1 << 23) | (payload << 16) | (rtp->seqno));
rtpheader[1] = htonl(rtp->lastdigitts);
rtpheader[2] = htonl(rtp->ssrc);
/* Actually send the packet */
for (i = 0; i < 2; i++) {
int ice;
rtpheader[3] = htonl((digit << 24) | (0xa << 16) | (rtp->send_duration));
res = rtp_sendto(instance, (void *) rtpheader, hdrlen + 4, 0, &remote_address, &ice);
if (res < 0) {
ast_log(LOG_ERROR, "RTP Transmission error to %s: %s\n",
ast_sockaddr_stringify(&remote_address),
strerror(errno));
}
if (rtp_debug_test_addr(&remote_address)) {
ast_verbose("Sent RTP DTMF packet to %s%s (type %-2.2d, seq %-6.6d, ts %-6.6u, len %-6.6d)\n",
ast_sockaddr_stringify(&remote_address),
ice ? " (via ICE)" : "",
payload, rtp->seqno, rtp->lastdigitts, res - hdrlen);
}
rtp->seqno++;
rtp->send_duration += 160;
rtpheader[0] = htonl((2 << 30) | (payload << 16) | (rtp->seqno));
}
/* Record that we are in the process of sending a digit and information needed to continue doing so */
rtp->sending_digit = 1;
rtp->send_digit = digit;
rtp->send_payload = payload;
return 0;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static int ast_rtp_dtmf_continuation(struct ast_rtp_instance *instance)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
struct ast_sockaddr remote_address = { {0,} };
int hdrlen = 12, res = 0;
char data[256];
unsigned int *rtpheader = (unsigned int*)data;
int ice;
ast_rtp_instance_get_remote_address(instance, &remote_address);
/* Make sure we know where the other side is so we can send them the packet */
if (ast_sockaddr_isnull(&remote_address)) {
return -1;
}
/* Actually create the packet we will be sending */
rtpheader[0] = htonl((2 << 30) | (rtp->send_payload << 16) | (rtp->seqno));
rtpheader[1] = htonl(rtp->lastdigitts);
rtpheader[2] = htonl(rtp->ssrc);
rtpheader[3] = htonl((rtp->send_digit << 24) | (0xa << 16) | (rtp->send_duration));
/* Boom, send it on out */
res = rtp_sendto(instance, (void *) rtpheader, hdrlen + 4, 0, &remote_address, &ice);
if (res < 0) {
ast_log(LOG_ERROR, "RTP Transmission error to %s: %s\n",
ast_sockaddr_stringify(&remote_address),
strerror(errno));
}
if (rtp_debug_test_addr(&remote_address)) {
ast_verbose("Sent RTP DTMF packet to %s%s (type %-2.2d, seq %-6.6d, ts %-6.6u, len %-6.6d)\n",
ast_sockaddr_stringify(&remote_address),
ice ? " (via ICE)" : "",
rtp->send_payload, rtp->seqno, rtp->lastdigitts, res - hdrlen);
}
/* And now we increment some values for the next time we swing by */
rtp->seqno++;
rtp->send_duration += 160;
rtp->lastts += calc_txstamp(rtp, NULL) * DTMF_SAMPLE_RATE_MS;
return 0;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static int ast_rtp_dtmf_end_with_duration(struct ast_rtp_instance *instance, char digit, unsigned int duration)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
struct ast_sockaddr remote_address = { {0,} };
int hdrlen = 12, res = -1, i = 0;
char data[256];
unsigned int *rtpheader = (unsigned int*)data;
unsigned int measured_samples;
ast_rtp_instance_get_remote_address(instance, &remote_address);
/* Make sure we know where the remote side is so we can send them the packet we construct */
if (ast_sockaddr_isnull(&remote_address)) {
goto cleanup;
}
/* Convert the given digit to the one we are going to send */
if ((digit <= '9') && (digit >= '0')) {
digit -= '0';
} else if (digit == '*') {
digit = 10;
} else if (digit == '#') {
digit = 11;
} else if ((digit >= 'A') && (digit <= 'D')) {
digit = digit - 'A' + 12;
} else if ((digit >= 'a') && (digit <= 'd')) {
digit = digit - 'a' + 12;
} else {
ast_log(LOG_WARNING, "Don't know how to represent '%c'\n", digit);
goto cleanup;
}
rtp->dtmfmute = ast_tvadd(ast_tvnow(), ast_tv(0, 500000));
if (duration > 0 && (measured_samples = duration * ast_rtp_get_rate(rtp->f.subclass.format) / 1000) > rtp->send_duration) {
ast_debug_rtp(2, "(%p) RTP adjusting final end duration from %d to %u\n",
instance, rtp->send_duration, measured_samples);
rtp->send_duration = measured_samples;
}
/* Construct the packet we are going to send */
rtpheader[1] = htonl(rtp->lastdigitts);
rtpheader[2] = htonl(rtp->ssrc);
rtpheader[3] = htonl((digit << 24) | (0xa << 16) | (rtp->send_duration));
rtpheader[3] |= htonl((1 << 23));
/* Send it 3 times, that's the magical number */
for (i = 0; i < 3; i++) {
int ice;
rtpheader[0] = htonl((2 << 30) | (rtp->send_payload << 16) | (rtp->seqno));
res = rtp_sendto(instance, (void *) rtpheader, hdrlen + 4, 0, &remote_address, &ice);
if (res < 0) {
ast_log(LOG_ERROR, "RTP Transmission error to %s: %s\n",
ast_sockaddr_stringify(&remote_address),
strerror(errno));
}
if (rtp_debug_test_addr(&remote_address)) {
ast_verbose("Sent RTP DTMF packet to %s%s (type %-2.2d, seq %-6.6d, ts %-6.6u, len %-6.6d)\n",
ast_sockaddr_stringify(&remote_address),
ice ? " (via ICE)" : "",
rtp->send_payload, rtp->seqno, rtp->lastdigitts, res - hdrlen);
}
rtp->seqno++;
}
res = 0;
/* Oh and we can't forget to turn off the stuff that says we are sending DTMF */
rtp->lastts += calc_txstamp(rtp, NULL) * DTMF_SAMPLE_RATE_MS;
/* Reset the smoother as the delivery time stored in it is now out of date */
if (rtp->smoother) {
ast_smoother_free(rtp->smoother);
rtp->smoother = NULL;
}
cleanup:
rtp->sending_digit = 0;
rtp->send_digit = 0;
/* Re-Learn expected seqno */
rtp->expectedseqno = -1;
return res;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static int ast_rtp_dtmf_end(struct ast_rtp_instance *instance, char digit)
{
return ast_rtp_dtmf_end_with_duration(instance, digit, 0);
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static void ast_rtp_update_source(struct ast_rtp_instance *instance)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
/* We simply set this bit so that the next packet sent will have the marker bit turned on */
ast_set_flag(rtp, FLAG_NEED_MARKER_BIT);
ast_debug_rtp(3, "(%p) RTP setting the marker bit due to a source update\n", instance);
return;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static void ast_rtp_change_source(struct ast_rtp_instance *instance)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
struct ast_srtp *srtp = ast_rtp_instance_get_srtp(instance, 0);
struct ast_srtp *rtcp_srtp = ast_rtp_instance_get_srtp(instance, 1);
unsigned int ssrc = ast_random();
if (rtp->lastts) {
/* We simply set this bit so that the next packet sent will have the marker bit turned on */
ast_set_flag(rtp, FLAG_NEED_MARKER_BIT);
}
ast_debug_rtp(3, "(%p) RTP changing ssrc from %u to %u due to a source change\n",
instance, rtp->ssrc, ssrc);
if (srtp) {
ast_debug_rtp(3, "(%p) RTP changing ssrc for SRTP from %u to %u\n",
instance, rtp->ssrc, ssrc);
res_srtp->change_source(srtp, rtp->ssrc, ssrc);
if (rtcp_srtp != srtp) {
res_srtp->change_source(rtcp_srtp, rtp->ssrc, ssrc);
}
}
rtp->ssrc = ssrc;
Merged revisions 221086 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r221086 | twilson | 2009-09-30 09:49:11 -0500 (Wed, 30 Sep 2009) | 25 lines Change the SSRC by default when our media stream changes Be default, change SSRC when doing an audio stream changes Asterisk doesn't honor marker bit when reinvited to already-bridged RTP streams,resulting in far-end stack discarding packets with "old" timestamps that areactually part of a new stream. This patch sends AST_CONTROL_SRCUPDATE whenever there is a reinvite, unless the 'constantssrc' is set to true in sip.conf. The original issue reported to Digium support detailed the following situation: ITSP <-> Asterisk 1.4.26.2 <-> SIP-based Application Server Call comes in fromITSP, Asterisk dials the app server which sends a re-invite back toAsterisk--not to negotiate to send media directly to the ITSP, but to indicatethat it's changing the stream it's sending to Asterisk. The app servergenerates a new SSRC, sequence numbers, timestamps, and sets the marker bit on the new stream. Asterisk passes through the teimstamp of the new stream, butdoes not reset the SSRC, sequence numbers, or set the marker bit. When the timestamp on the new stream is older than the timestamp on the originalstream, the ITSP (which doesn't know there has been any change) discards the newframes because it thinks they are too old. This patch addresses this by changing the SSRC on a stream update unless constantssrc=true is set in sip.conf. Review: https://reviewboard.asterisk.org/r/374/ ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@221266 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2009-09-30 17:52:30 +00:00
2018-05-10 18:11:06 +00:00
/* Since the source is changing, we don't know what sequence number to expect next */
rtp->expectedrxseqno = -1;
return;
}
static void timeval2ntp(struct timeval tv, unsigned int *msw, unsigned int *lsw)
{
unsigned int sec, usec, frac;
sec = tv.tv_sec + 2208988800u; /* Sec between 1900 and 1970 */
usec = tv.tv_usec;
/*
* Convert usec to 0.32 bit fixed point without overflow.
*
* = usec * 2^32 / 10^6
* = usec * 2^32 / (2^6 * 5^6)
* = usec * 2^26 / 5^6
*
* The usec value needs 20 bits to represent 999999 usec. So
* splitting the 2^26 to get the most precision using 32 bit
* values gives:
*
* = ((usec * 2^12) / 5^6) * 2^14
*
* Splitting the division into two stages preserves all the
* available significant bits of usec over doing the division
* all at once.
*
* = ((((usec * 2^12) / 5^3) * 2^7) / 5^3) * 2^7
*/
frac = ((((usec << 12) / 125) << 7) / 125) << 7;
*msw = sec;
*lsw = frac;
}
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
static void ntp2timeval(unsigned int msw, unsigned int lsw, struct timeval *tv)
{
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
tv->tv_sec = msw - 2208988800u;
/* Reverse the sequence in timeval2ntp() */
tv->tv_usec = ((((lsw >> 7) * 125) >> 7) * 125) >> 12;
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
}
static void calculate_lost_packet_statistics(struct ast_rtp *rtp,
unsigned int *lost_packets,
int *fraction_lost)
{
unsigned int extended_seq_no;
unsigned int expected_packets;
unsigned int expected_interval;
unsigned int received_interval;
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
int lost_interval;
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
/* Compute statistics */
extended_seq_no = rtp->cycles + rtp->lastrxseqno;
expected_packets = extended_seq_no - rtp->seedrxseqno + 1;
if (rtp->rxcount > expected_packets) {
expected_packets += rtp->rxcount - expected_packets;
}
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
*lost_packets = expected_packets - rtp->rxcount;
expected_interval = expected_packets - rtp->rtcp->expected_prior;
received_interval = rtp->rxcount - rtp->rtcp->received_prior;
if (received_interval > expected_interval) {
/* If we receive some late packets it is possible for the packets
* we received in this interval to exceed the number we expected.
* We update the expected so that the packet loss calculations
* show that no packets are lost.
*/
expected_interval = received_interval;
}
lost_interval = expected_interval - received_interval;
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
if (expected_interval == 0 || lost_interval <= 0) {
*fraction_lost = 0;
} else {
*fraction_lost = (lost_interval << 8) / expected_interval;
}
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
/* Update RTCP statistics */
rtp->rtcp->received_prior = rtp->rxcount;
rtp->rtcp->expected_prior = expected_packets;
/*
* While rxlost represents the number of packets lost since the last report was sent, for
* the calculations below it should be thought of as a single sample. Thus min/max are the
* lowest/highest sample value seen, and the mean is the average number of packets lost
* between each report. As such rxlost_count only needs to be incremented per report.
*/
if (lost_interval <= 0) {
rtp->rtcp->rxlost = 0;
} else {
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
rtp->rtcp->rxlost = lost_interval;
}
if (rtp->rtcp->rxlost_count == 0) {
rtp->rtcp->minrxlost = rtp->rtcp->rxlost;
}
if (lost_interval && lost_interval < rtp->rtcp->minrxlost) {
rtp->rtcp->minrxlost = rtp->rtcp->rxlost;
}
if (lost_interval > rtp->rtcp->maxrxlost) {
rtp->rtcp->maxrxlost = rtp->rtcp->rxlost;
}
calc_mean_and_standard_deviation(rtp->rtcp->rxlost, &rtp->rtcp->normdev_rxlost,
&rtp->rtcp->stdev_rxlost, &rtp->rtcp->rxlost_count);
}
2018-05-10 18:11:06 +00:00
static int ast_rtcp_generate_report(struct ast_rtp_instance *instance, unsigned char *rtcpheader,
struct ast_rtp_rtcp_report *rtcp_report, int *sr)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
int len = 0;
struct timeval now;
unsigned int now_lsw;
unsigned int now_msw;
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
unsigned int lost_packets;
int fraction_lost;
struct timeval dlsr = { 0, };
res_hep_rtcp: Add module that sends RTCP information to a Homer Server This patch adds a new module to Asterisk, res_hep_rtcp. The module subscribes to the RTCP topics in Stasis and receives RTCP information back from the message bus. It encodes into HEPv3 packets and sends the information to the res_hep module for transmission. Using this, someone with a Homer server can get live call quality monitoring for all RTP-based channels in their Asterisk 12+ systems. In addition, there were a few bugs in the RTP engine, res_rtp_asterisk, and chan_pjsip that were uncovered by the tests written for the Asterisk Test Suite. This patch fixes the following: 1) chan_pjsip failed to set its channel unique ids on its RTP instance on outbound calls. It now does this in the appropriate location, in the serialized call callback. 2) The rtp_engine was overflowing some values when packed into JSON. Specifically, some longs and unsigned ints can't be be packed into integer values, for obvious reasons. Since libjansson only supports integers, floats, strings, booleans, and objects, we print these values into strings. 3) res_rtp_asterisk had a few problems: (a) it would emit a source IP address of 0.0.0.0 if bound to that IP address. We now use ast_find_ourip to get a better IP address, and properly marshal the result into an ast_strdupa'd string. (b) Reports can be generated with no report bodies. In particular, this occurs when a sender is transmitting information to a receiver (who will send no RTP back to the sender). As such, the sender has no report body for what it received. We now properly handle this case, and the sender will emit SR reports with no body. Likewise, if we receive an RTCP packet with no report body, we will still generate the appropriate events. ASTERISK-24119 #close ........ Merged revisions 419823 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@419825 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-07-31 11:57:51 +00:00
struct ast_rtp_rtcp_report_block *report_block = NULL;
if (!rtp || !rtp->rtcp) {
return 0;
}
2018-05-10 18:11:06 +00:00
if (ast_sockaddr_isnull(&rtp->rtcp->them)) { /* This'll stop rtcp for this rtp session */
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
/* RTCP was stopped. */
return 0;
}
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
if (!rtcp_report) {
return 1;
}
2018-05-10 18:11:06 +00:00
*sr = rtp->txcount > rtp->rtcp->lastsrtxcount ? 1 : 0;
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
/* Compute statistics */
calculate_lost_packet_statistics(rtp, &lost_packets, &fraction_lost);
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
/*
* update_local_mes_stats must be called AFTER
* calculate_lost_packet_statistics
*/
update_local_mes_stats(rtp);
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
gettimeofday(&now, NULL);
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
rtcp_report->reception_report_count = rtp->themssrc_valid ? 1 : 0;
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
rtcp_report->ssrc = rtp->ssrc;
2018-05-10 18:11:06 +00:00
rtcp_report->type = *sr ? RTCP_PT_SR : RTCP_PT_RR;
if (*sr) {
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
rtcp_report->sender_information.ntp_timestamp = now;
rtcp_report->sender_information.rtp_timestamp = rtp->lastts;
rtcp_report->sender_information.packet_count = rtp->txcount;
rtcp_report->sender_information.octet_count = rtp->txoctetcount;
}
res_hep_rtcp: Add module that sends RTCP information to a Homer Server This patch adds a new module to Asterisk, res_hep_rtcp. The module subscribes to the RTCP topics in Stasis and receives RTCP information back from the message bus. It encodes into HEPv3 packets and sends the information to the res_hep module for transmission. Using this, someone with a Homer server can get live call quality monitoring for all RTP-based channels in their Asterisk 12+ systems. In addition, there were a few bugs in the RTP engine, res_rtp_asterisk, and chan_pjsip that were uncovered by the tests written for the Asterisk Test Suite. This patch fixes the following: 1) chan_pjsip failed to set its channel unique ids on its RTP instance on outbound calls. It now does this in the appropriate location, in the serialized call callback. 2) The rtp_engine was overflowing some values when packed into JSON. Specifically, some longs and unsigned ints can't be be packed into integer values, for obvious reasons. Since libjansson only supports integers, floats, strings, booleans, and objects, we print these values into strings. 3) res_rtp_asterisk had a few problems: (a) it would emit a source IP address of 0.0.0.0 if bound to that IP address. We now use ast_find_ourip to get a better IP address, and properly marshal the result into an ast_strdupa'd string. (b) Reports can be generated with no report bodies. In particular, this occurs when a sender is transmitting information to a receiver (who will send no RTP back to the sender). As such, the sender has no report body for what it received. We now properly handle this case, and the sender will emit SR reports with no body. Likewise, if we receive an RTCP packet with no report body, we will still generate the appropriate events. ASTERISK-24119 #close ........ Merged revisions 419823 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@419825 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-07-31 11:57:51 +00:00
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
if (rtp->themssrc_valid) {
res_hep_rtcp: Add module that sends RTCP information to a Homer Server This patch adds a new module to Asterisk, res_hep_rtcp. The module subscribes to the RTCP topics in Stasis and receives RTCP information back from the message bus. It encodes into HEPv3 packets and sends the information to the res_hep module for transmission. Using this, someone with a Homer server can get live call quality monitoring for all RTP-based channels in their Asterisk 12+ systems. In addition, there were a few bugs in the RTP engine, res_rtp_asterisk, and chan_pjsip that were uncovered by the tests written for the Asterisk Test Suite. This patch fixes the following: 1) chan_pjsip failed to set its channel unique ids on its RTP instance on outbound calls. It now does this in the appropriate location, in the serialized call callback. 2) The rtp_engine was overflowing some values when packed into JSON. Specifically, some longs and unsigned ints can't be be packed into integer values, for obvious reasons. Since libjansson only supports integers, floats, strings, booleans, and objects, we print these values into strings. 3) res_rtp_asterisk had a few problems: (a) it would emit a source IP address of 0.0.0.0 if bound to that IP address. We now use ast_find_ourip to get a better IP address, and properly marshal the result into an ast_strdupa'd string. (b) Reports can be generated with no report bodies. In particular, this occurs when a sender is transmitting information to a receiver (who will send no RTP back to the sender). As such, the sender has no report body for what it received. We now properly handle this case, and the sender will emit SR reports with no body. Likewise, if we receive an RTCP packet with no report body, we will still generate the appropriate events. ASTERISK-24119 #close ........ Merged revisions 419823 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@419825 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-07-31 11:57:51 +00:00
report_block = ast_calloc(1, sizeof(*report_block));
if (!report_block) {
return 1;
}
rtcp_report->report_block[0] = report_block;
report_block->source_ssrc = rtp->themssrc;
report_block->lost_count.fraction = (fraction_lost & 0xff);
report_block->lost_count.packets = (lost_packets & 0xffffff);
report_block->highest_seq_no = (rtp->cycles | (rtp->lastrxseqno & 0xffff));
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
report_block->ia_jitter = (unsigned int)rtp->rxjitter_samples;
res_hep_rtcp: Add module that sends RTCP information to a Homer Server This patch adds a new module to Asterisk, res_hep_rtcp. The module subscribes to the RTCP topics in Stasis and receives RTCP information back from the message bus. It encodes into HEPv3 packets and sends the information to the res_hep module for transmission. Using this, someone with a Homer server can get live call quality monitoring for all RTP-based channels in their Asterisk 12+ systems. In addition, there were a few bugs in the RTP engine, res_rtp_asterisk, and chan_pjsip that were uncovered by the tests written for the Asterisk Test Suite. This patch fixes the following: 1) chan_pjsip failed to set its channel unique ids on its RTP instance on outbound calls. It now does this in the appropriate location, in the serialized call callback. 2) The rtp_engine was overflowing some values when packed into JSON. Specifically, some longs and unsigned ints can't be be packed into integer values, for obvious reasons. Since libjansson only supports integers, floats, strings, booleans, and objects, we print these values into strings. 3) res_rtp_asterisk had a few problems: (a) it would emit a source IP address of 0.0.0.0 if bound to that IP address. We now use ast_find_ourip to get a better IP address, and properly marshal the result into an ast_strdupa'd string. (b) Reports can be generated with no report bodies. In particular, this occurs when a sender is transmitting information to a receiver (who will send no RTP back to the sender). As such, the sender has no report body for what it received. We now properly handle this case, and the sender will emit SR reports with no body. Likewise, if we receive an RTCP packet with no report body, we will still generate the appropriate events. ASTERISK-24119 #close ........ Merged revisions 419823 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@419825 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-07-31 11:57:51 +00:00
report_block->lsr = rtp->rtcp->themrxlsr;
/* If we haven't received an SR report, DLSR should be 0 */
if (!ast_tvzero(rtp->rtcp->rxlsr)) {
timersub(&now, &rtp->rtcp->rxlsr, &dlsr);
report_block->dlsr = (((dlsr.tv_sec * 1000) + (dlsr.tv_usec / 1000)) * 65536) / 1000;
}
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
}
timeval2ntp(rtcp_report->sender_information.ntp_timestamp, &now_msw, &now_lsw);
put_unaligned_uint32(rtcpheader + 4, htonl(rtcp_report->ssrc)); /* Our SSRC */
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
len += 8;
2018-05-10 18:11:06 +00:00
if (*sr) {
put_unaligned_uint32(rtcpheader + len, htonl(now_msw)); /* now, MSW. gettimeofday() + SEC_BETWEEN_1900_AND_1970 */
put_unaligned_uint32(rtcpheader + len + 4, htonl(now_lsw)); /* now, LSW */
put_unaligned_uint32(rtcpheader + len + 8, htonl(rtcp_report->sender_information.rtp_timestamp));
put_unaligned_uint32(rtcpheader + len + 12, htonl(rtcp_report->sender_information.packet_count));
put_unaligned_uint32(rtcpheader + len + 16, htonl(rtcp_report->sender_information.octet_count));
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
len += 20;
}
res_hep_rtcp: Add module that sends RTCP information to a Homer Server This patch adds a new module to Asterisk, res_hep_rtcp. The module subscribes to the RTCP topics in Stasis and receives RTCP information back from the message bus. It encodes into HEPv3 packets and sends the information to the res_hep module for transmission. Using this, someone with a Homer server can get live call quality monitoring for all RTP-based channels in their Asterisk 12+ systems. In addition, there were a few bugs in the RTP engine, res_rtp_asterisk, and chan_pjsip that were uncovered by the tests written for the Asterisk Test Suite. This patch fixes the following: 1) chan_pjsip failed to set its channel unique ids on its RTP instance on outbound calls. It now does this in the appropriate location, in the serialized call callback. 2) The rtp_engine was overflowing some values when packed into JSON. Specifically, some longs and unsigned ints can't be be packed into integer values, for obvious reasons. Since libjansson only supports integers, floats, strings, booleans, and objects, we print these values into strings. 3) res_rtp_asterisk had a few problems: (a) it would emit a source IP address of 0.0.0.0 if bound to that IP address. We now use ast_find_ourip to get a better IP address, and properly marshal the result into an ast_strdupa'd string. (b) Reports can be generated with no report bodies. In particular, this occurs when a sender is transmitting information to a receiver (who will send no RTP back to the sender). As such, the sender has no report body for what it received. We now properly handle this case, and the sender will emit SR reports with no body. Likewise, if we receive an RTCP packet with no report body, we will still generate the appropriate events. ASTERISK-24119 #close ........ Merged revisions 419823 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@419825 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-07-31 11:57:51 +00:00
if (report_block) {
put_unaligned_uint32(rtcpheader + len, htonl(report_block->source_ssrc)); /* Their SSRC */
put_unaligned_uint32(rtcpheader + len + 4, htonl((report_block->lost_count.fraction << 24) | report_block->lost_count.packets));
put_unaligned_uint32(rtcpheader + len + 8, htonl(report_block->highest_seq_no));
put_unaligned_uint32(rtcpheader + len + 12, htonl(report_block->ia_jitter));
put_unaligned_uint32(rtcpheader + len + 16, htonl(report_block->lsr));
put_unaligned_uint32(rtcpheader + len + 20, htonl(report_block->dlsr));
res_hep_rtcp: Add module that sends RTCP information to a Homer Server This patch adds a new module to Asterisk, res_hep_rtcp. The module subscribes to the RTCP topics in Stasis and receives RTCP information back from the message bus. It encodes into HEPv3 packets and sends the information to the res_hep module for transmission. Using this, someone with a Homer server can get live call quality monitoring for all RTP-based channels in their Asterisk 12+ systems. In addition, there were a few bugs in the RTP engine, res_rtp_asterisk, and chan_pjsip that were uncovered by the tests written for the Asterisk Test Suite. This patch fixes the following: 1) chan_pjsip failed to set its channel unique ids on its RTP instance on outbound calls. It now does this in the appropriate location, in the serialized call callback. 2) The rtp_engine was overflowing some values when packed into JSON. Specifically, some longs and unsigned ints can't be be packed into integer values, for obvious reasons. Since libjansson only supports integers, floats, strings, booleans, and objects, we print these values into strings. 3) res_rtp_asterisk had a few problems: (a) it would emit a source IP address of 0.0.0.0 if bound to that IP address. We now use ast_find_ourip to get a better IP address, and properly marshal the result into an ast_strdupa'd string. (b) Reports can be generated with no report bodies. In particular, this occurs when a sender is transmitting information to a receiver (who will send no RTP back to the sender). As such, the sender has no report body for what it received. We now properly handle this case, and the sender will emit SR reports with no body. Likewise, if we receive an RTCP packet with no report body, we will still generate the appropriate events. ASTERISK-24119 #close ........ Merged revisions 419823 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@419825 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-07-31 11:57:51 +00:00
len += 24;
}
put_unaligned_uint32(rtcpheader, htonl((2 << 30) | (rtcp_report->reception_report_count << 24)
2018-05-10 18:11:06 +00:00
| ((*sr ? RTCP_PT_SR : RTCP_PT_RR) << 16) | ((len/4)-1)));
2018-05-10 18:11:06 +00:00
return len;
}
2018-05-10 18:11:06 +00:00
static int ast_rtcp_calculate_sr_rr_statistics(struct ast_rtp_instance *instance,
struct ast_rtp_rtcp_report *rtcp_report, struct ast_sockaddr remote_address, int ice, int sr)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
struct ast_rtp_rtcp_report_block *report_block = NULL;
RAII_VAR(struct ast_json *, message_blob, NULL, ast_json_unref);
2018-05-10 18:11:06 +00:00
if (!rtp || !rtp->rtcp) {
return 0;
}
2018-05-10 18:11:06 +00:00
if (ast_sockaddr_isnull(&rtp->rtcp->them)) {
return 0;
}
2018-05-10 18:11:06 +00:00
if (!rtcp_report) {
return -1;
}
report_block = rtcp_report->report_block[0];
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
if (sr) {
rtp->rtcp->txlsr = rtcp_report->sender_information.ntp_timestamp;
rtp->rtcp->sr_count++;
rtp->rtcp->lastsrtxcount = rtp->txcount;
} else {
rtp->rtcp->rr_count++;
}
if (rtcp_debug_test_addr(&rtp->rtcp->them)) {
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
ast_verbose("* Sent RTCP %s to %s%s\n", sr ? "SR" : "RR",
ast_sockaddr_stringify(&remote_address), ice ? " (via ICE)" : "");
ast_verbose(" Our SSRC: %u\n", rtcp_report->ssrc);
if (sr) {
ast_verbose(" Sent(NTP): %u.%06u\n",
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
(unsigned int)rtcp_report->sender_information.ntp_timestamp.tv_sec,
(unsigned int)rtcp_report->sender_information.ntp_timestamp.tv_usec);
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
ast_verbose(" Sent(RTP): %u\n", rtcp_report->sender_information.rtp_timestamp);
ast_verbose(" Sent packets: %u\n", rtcp_report->sender_information.packet_count);
ast_verbose(" Sent octets: %u\n", rtcp_report->sender_information.octet_count);
}
if (report_block) {
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
int rate = ast_rtp_get_rate(rtp->f.subclass.format);
ast_verbose(" Report block:\n");
ast_verbose(" Their SSRC: %u\n", report_block->source_ssrc);
ast_verbose(" Fraction lost: %d\n", report_block->lost_count.fraction);
ast_verbose(" Cumulative loss: %u\n", report_block->lost_count.packets);
ast_verbose(" Highest seq no: %u\n", report_block->highest_seq_no);
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
ast_verbose(" IA jitter (samp): %u\n", report_block->ia_jitter);
ast_verbose(" IA jitter (secs): %.6f\n", ast_samp2sec(report_block->ia_jitter, rate));
ast_verbose(" Their last SR: %u\n", report_block->lsr);
ast_verbose(" DLSR: %4.4f (sec)\n\n", (double)(report_block->dlsr / 65536.0));
}
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
}
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
message_blob = ast_json_pack("{s: s, s: s, s: f}",
"to", ast_sockaddr_stringify(&remote_address),
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
"from", rtp->rtcp->local_addr_str,
"mes", rtp->rxmes);
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
ast_rtp_publish_rtcp_message(instance, ast_rtp_rtcp_sent_type(),
2018-05-10 18:11:06 +00:00
rtcp_report, message_blob);
return 1;
}
static int ast_rtcp_generate_sdes(struct ast_rtp_instance *instance, unsigned char *rtcpheader,
struct ast_rtp_rtcp_report *rtcp_report)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
int len = 0;
uint16_t sdes_packet_len_bytes;
uint16_t sdes_packet_len_rounded;
if (!rtp || !rtp->rtcp) {
return 0;
}
if (ast_sockaddr_isnull(&rtp->rtcp->them)) {
return 0;
}
if (!rtcp_report) {
return -1;
}
sdes_packet_len_bytes =
4 + /* RTCP Header */
4 + /* SSRC */
1 + /* Type (CNAME) */
1 + /* Text Length */
AST_UUID_STR_LEN /* Text and NULL terminator */
;
/* Round to 32 bit boundary */
sdes_packet_len_rounded = (sdes_packet_len_bytes + 3) & ~0x3;
put_unaligned_uint32(rtcpheader, htonl((2 << 30) | (1 << 24) | (RTCP_PT_SDES << 16) | ((sdes_packet_len_rounded / 4) - 1)));
put_unaligned_uint32(rtcpheader + 4, htonl(rtcp_report->ssrc));
rtcpheader[8] = 0x01; /* CNAME */
rtcpheader[9] = AST_UUID_STR_LEN - 1; /* Number of bytes of text */
memcpy(rtcpheader + 10, rtp->cname, AST_UUID_STR_LEN);
len += 10 + AST_UUID_STR_LEN;
/* Padding - Note that we don't set the padded bit on the packet. From
* RFC 3550 Section 6.5:
*
* No length octet follows the null item type octet, but additional null
* octets MUST be included if needd to pad until the next 32-bit
* boundary. Note that this padding is separate from that indicated by
* the P bit in the RTCP header.
*
* These bytes will already be zeroed out during array initialization.
*/
len += (sdes_packet_len_rounded - sdes_packet_len_bytes);
return len;
}
/* Lock instance before calling this if it isn't already
*
* If successful, the overall packet length is returned
* If not, then 0 is returned
*/
static int ast_rtcp_generate_compound_prefix(struct ast_rtp_instance *instance, unsigned char *rtcpheader,
struct ast_rtp_rtcp_report *report, int *sr)
{
int packet_len = 0;
int res;
/* Every RTCP packet needs to be sent out with a SR/RR and SDES prefixing it.
* At the end of this function, rtcpheader should contain both of those packets,
* and will return the length of the overall packet. This can be used to determine
* where further packets can be inserted in the compound packet.
*/
res = ast_rtcp_generate_report(instance, rtcpheader, report, sr);
if (res == 0 || res == 1) {
ast_debug_rtcp(1, "(%p) RTCP failed to generate %s report!\n", instance, sr ? "SR" : "RR");
return 0;
}
packet_len += res;
res = ast_rtcp_generate_sdes(instance, rtcpheader + packet_len, report);
if (res == 0 || res == 1) {
ast_debug_rtcp(1, "(%p) RTCP failed to generate SDES!\n", instance);
return 0;
}
return packet_len + res;
}
2018-05-10 18:11:06 +00:00
static int ast_rtcp_generate_nack(struct ast_rtp_instance *instance, unsigned char *rtcpheader)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
int packet_len;
int blp_index = -1;
2018-05-10 18:11:06 +00:00
int current_seqno;
unsigned int fci = 0;
size_t remaining_missing_seqno;
2018-05-10 18:11:06 +00:00
if (!rtp || !rtp->rtcp) {
return 0;
}
if (ast_sockaddr_isnull(&rtp->rtcp->them)) {
return 0;
}
current_seqno = rtp->expectedrxseqno;
remaining_missing_seqno = AST_VECTOR_SIZE(&rtp->missing_seqno);
2018-05-10 18:11:06 +00:00
packet_len = 12; /* The header length is 12 (version line, packet source SSRC, media source SSRC) */
/* If there are no missing sequence numbers then don't bother sending a NACK needlessly */
if (!remaining_missing_seqno) {
return 0;
}
/* This iterates through the possible forward sequence numbers seeing which ones we
* have no packet for, adding it to the NACK until we are out of missing packets.
*/
while (remaining_missing_seqno) {
2018-05-10 18:11:06 +00:00
int *missing_seqno;
/* On the first entry to this loop blp_index will be -1, so this will become 0
* and the sequence number will be placed into the packet as the PID.
*/
blp_index++;
2018-05-10 18:11:06 +00:00
missing_seqno = AST_VECTOR_GET_CMP(&rtp->missing_seqno, current_seqno,
find_by_value);
if (missing_seqno) {
/* We hit the max blp size, reset */
if (blp_index >= 17) {
put_unaligned_uint32(rtcpheader + packet_len, htonl(fci));
fci = 0;
blp_index = 0;
packet_len += 4;
}
2018-05-10 18:11:06 +00:00
if (blp_index == 0) {
fci |= (current_seqno << 16);
} else {
fci |= (1 << (blp_index - 1));
}
2018-05-10 18:11:06 +00:00
/* Since we've used a missing sequence number, we're down one */
remaining_missing_seqno--;
2018-05-10 18:11:06 +00:00
}
/* Handle cycling of the sequence number */
current_seqno++;
if (current_seqno == SEQNO_CYCLE_OVER) {
current_seqno = 0;
2018-05-10 18:11:06 +00:00
}
}
put_unaligned_uint32(rtcpheader + packet_len, htonl(fci));
packet_len += 4;
/* Length MUST be 2+n, where n is the number of NACKs. Same as length in words minus 1 */
put_unaligned_uint32(rtcpheader, htonl((2 << 30) | (AST_RTP_RTCP_FMT_NACK << 24)
| (AST_RTP_RTCP_RTPFB << 16) | ((packet_len / 4) - 1)));
put_unaligned_uint32(rtcpheader + 4, htonl(rtp->ssrc));
put_unaligned_uint32(rtcpheader + 8, htonl(rtp->themssrc));
return packet_len;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*!
* \brief Write a RTCP packet to the far end
*
* \note Decide if we are going to send an SR (with Reception Block) or RR
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
* RR is sent if we have not sent any rtp packets in the previous interval
*
* Scheduler callback
*/
static int ast_rtcp_write(const void *data)
{
struct ast_rtp_instance *instance = (struct ast_rtp_instance *) data;
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
int res;
2018-05-10 18:11:06 +00:00
int sr = 0;
int packet_len = 0;
int ice;
struct ast_sockaddr remote_address = { { 0, } };
unsigned char *rtcpheader;
unsigned char bdata[AST_UUID_STR_LEN + 128] = ""; /* More than enough */
RAII_VAR(struct ast_rtp_rtcp_report *, rtcp_report, NULL, ao2_cleanup);
Merged revisions 336878 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/10 ................ r336878 | russell | 2011-09-19 20:03:55 -0500 (Mon, 19 Sep 2011) | 43 lines Merged revisions 336877 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.8 ........ r336877 | russell | 2011-09-19 19:56:20 -0500 (Mon, 19 Sep 2011) | 36 lines Fix crashes in ast_rtcp_write(). This patch addresses crashes related to RTCP handling. The backtraces just show a crash in ast_rtcp_write() where it appears that the RTP instance is no longer valid. There is a race condition with scheduled RTCP transmissions and the destruction of the RTP instance. This patch utilizes the fact that ast_rtp_instance is a reference counted object and ensures that it will not get destroyed while a reference is still around due to scheduled RTCP transmissions. RTCP transmissions are scheduled and executed from the chan_sip scheduler context. This scheduler context is processed in the SIP monitor thread. The destruction of an RTP instance occurs when the associated sip_pvt gets destroyed (which happens when the sip_pvt reference count reaches 0). However, the SIP monitor thread is not the only thread that can cause a sip_pvt to get destroyed. The sip_hangup function, executed from a channel thread, also decrements the reference count on a sip_pvt and could cause it to get destroyed. While this is being changed anyway, the patch also removes calling ast_sched_del() from within the RTCP scheduler callback. It's not helpful. Simply returning 0 prevents the callback from being rescheduled. (closes issue ASTERISK-18570) Related issues that look like they are the same problem: (issue ASTERISK-17560) (issue ASTERISK-15406) (issue ASTERISK-15257) (issue ASTERISK-13334) (issue ASTERISK-9977) (issue ASTERISK-9716) Review: https://reviewboard.asterisk.org/r/1444/ ........ ................ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@336879 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2011-09-20 01:11:18 +00:00
if (!rtp || !rtp->rtcp || rtp->rtcp->schedid == -1) {
ao2_ref(instance, -1);
return 0;
Merged revisions 336878 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/10 ................ r336878 | russell | 2011-09-19 20:03:55 -0500 (Mon, 19 Sep 2011) | 43 lines Merged revisions 336877 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.8 ........ r336877 | russell | 2011-09-19 19:56:20 -0500 (Mon, 19 Sep 2011) | 36 lines Fix crashes in ast_rtcp_write(). This patch addresses crashes related to RTCP handling. The backtraces just show a crash in ast_rtcp_write() where it appears that the RTP instance is no longer valid. There is a race condition with scheduled RTCP transmissions and the destruction of the RTP instance. This patch utilizes the fact that ast_rtp_instance is a reference counted object and ensures that it will not get destroyed while a reference is still around due to scheduled RTCP transmissions. RTCP transmissions are scheduled and executed from the chan_sip scheduler context. This scheduler context is processed in the SIP monitor thread. The destruction of an RTP instance occurs when the associated sip_pvt gets destroyed (which happens when the sip_pvt reference count reaches 0). However, the SIP monitor thread is not the only thread that can cause a sip_pvt to get destroyed. The sip_hangup function, executed from a channel thread, also decrements the reference count on a sip_pvt and could cause it to get destroyed. While this is being changed anyway, the patch also removes calling ast_sched_del() from within the RTCP scheduler callback. It's not helpful. Simply returning 0 prevents the callback from being rescheduled. (closes issue ASTERISK-18570) Related issues that look like they are the same problem: (issue ASTERISK-17560) (issue ASTERISK-15406) (issue ASTERISK-15257) (issue ASTERISK-13334) (issue ASTERISK-9977) (issue ASTERISK-9716) Review: https://reviewboard.asterisk.org/r/1444/ ........ ................ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@336879 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2011-09-20 01:11:18 +00:00
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_lock(instance);
2018-05-10 18:11:06 +00:00
rtcpheader = bdata;
rtcp_report = ast_rtp_rtcp_report_alloc(rtp->themssrc_valid ? 1 : 0);
res = ast_rtcp_generate_compound_prefix(instance, rtcpheader, rtcp_report, &sr);
2018-05-10 18:11:06 +00:00
if (res == 0 || res == 1) {
goto cleanup;
}
packet_len += res;
if (rtp->bundled) {
ast_rtp_instance_get_remote_address(instance, &remote_address);
} else {
ast_sockaddr_copy(&remote_address, &rtp->rtcp->them);
}
res = rtcp_sendto(instance, (unsigned int *)rtcpheader, packet_len, 0, &remote_address, &ice);
if (res < 0) {
ast_log(LOG_ERROR, "RTCP %s transmission error to %s, rtcp halted %s\n",
sr ? "SR" : "RR",
ast_sockaddr_stringify(&rtp->rtcp->them),
strerror(errno));
res = 0;
Merged revisions 336878 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/10 ................ r336878 | russell | 2011-09-19 20:03:55 -0500 (Mon, 19 Sep 2011) | 43 lines Merged revisions 336877 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.8 ........ r336877 | russell | 2011-09-19 19:56:20 -0500 (Mon, 19 Sep 2011) | 36 lines Fix crashes in ast_rtcp_write(). This patch addresses crashes related to RTCP handling. The backtraces just show a crash in ast_rtcp_write() where it appears that the RTP instance is no longer valid. There is a race condition with scheduled RTCP transmissions and the destruction of the RTP instance. This patch utilizes the fact that ast_rtp_instance is a reference counted object and ensures that it will not get destroyed while a reference is still around due to scheduled RTCP transmissions. RTCP transmissions are scheduled and executed from the chan_sip scheduler context. This scheduler context is processed in the SIP monitor thread. The destruction of an RTP instance occurs when the associated sip_pvt gets destroyed (which happens when the sip_pvt reference count reaches 0). However, the SIP monitor thread is not the only thread that can cause a sip_pvt to get destroyed. The sip_hangup function, executed from a channel thread, also decrements the reference count on a sip_pvt and could cause it to get destroyed. While this is being changed anyway, the patch also removes calling ast_sched_del() from within the RTCP scheduler callback. It's not helpful. Simply returning 0 prevents the callback from being rescheduled. (closes issue ASTERISK-18570) Related issues that look like they are the same problem: (issue ASTERISK-17560) (issue ASTERISK-15406) (issue ASTERISK-15257) (issue ASTERISK-13334) (issue ASTERISK-9977) (issue ASTERISK-9716) Review: https://reviewboard.asterisk.org/r/1444/ ........ ................ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@336879 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2011-09-20 01:11:18 +00:00
} else {
2018-05-10 18:11:06 +00:00
ast_rtcp_calculate_sr_rr_statistics(instance, rtcp_report, remote_address, ice, sr);
Merged revisions 336878 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/10 ................ r336878 | russell | 2011-09-19 20:03:55 -0500 (Mon, 19 Sep 2011) | 43 lines Merged revisions 336877 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.8 ........ r336877 | russell | 2011-09-19 19:56:20 -0500 (Mon, 19 Sep 2011) | 36 lines Fix crashes in ast_rtcp_write(). This patch addresses crashes related to RTCP handling. The backtraces just show a crash in ast_rtcp_write() where it appears that the RTP instance is no longer valid. There is a race condition with scheduled RTCP transmissions and the destruction of the RTP instance. This patch utilizes the fact that ast_rtp_instance is a reference counted object and ensures that it will not get destroyed while a reference is still around due to scheduled RTCP transmissions. RTCP transmissions are scheduled and executed from the chan_sip scheduler context. This scheduler context is processed in the SIP monitor thread. The destruction of an RTP instance occurs when the associated sip_pvt gets destroyed (which happens when the sip_pvt reference count reaches 0). However, the SIP monitor thread is not the only thread that can cause a sip_pvt to get destroyed. The sip_hangup function, executed from a channel thread, also decrements the reference count on a sip_pvt and could cause it to get destroyed. While this is being changed anyway, the patch also removes calling ast_sched_del() from within the RTCP scheduler callback. It's not helpful. Simply returning 0 prevents the callback from being rescheduled. (closes issue ASTERISK-18570) Related issues that look like they are the same problem: (issue ASTERISK-17560) (issue ASTERISK-15406) (issue ASTERISK-15257) (issue ASTERISK-13334) (issue ASTERISK-9977) (issue ASTERISK-9716) Review: https://reviewboard.asterisk.org/r/1444/ ........ ................ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@336879 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2011-09-20 01:11:18 +00:00
}
2018-05-10 18:11:06 +00:00
cleanup:
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_unlock(instance);
Merged revisions 336878 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/10 ................ r336878 | russell | 2011-09-19 20:03:55 -0500 (Mon, 19 Sep 2011) | 43 lines Merged revisions 336877 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.8 ........ r336877 | russell | 2011-09-19 19:56:20 -0500 (Mon, 19 Sep 2011) | 36 lines Fix crashes in ast_rtcp_write(). This patch addresses crashes related to RTCP handling. The backtraces just show a crash in ast_rtcp_write() where it appears that the RTP instance is no longer valid. There is a race condition with scheduled RTCP transmissions and the destruction of the RTP instance. This patch utilizes the fact that ast_rtp_instance is a reference counted object and ensures that it will not get destroyed while a reference is still around due to scheduled RTCP transmissions. RTCP transmissions are scheduled and executed from the chan_sip scheduler context. This scheduler context is processed in the SIP monitor thread. The destruction of an RTP instance occurs when the associated sip_pvt gets destroyed (which happens when the sip_pvt reference count reaches 0). However, the SIP monitor thread is not the only thread that can cause a sip_pvt to get destroyed. The sip_hangup function, executed from a channel thread, also decrements the reference count on a sip_pvt and could cause it to get destroyed. While this is being changed anyway, the patch also removes calling ast_sched_del() from within the RTCP scheduler callback. It's not helpful. Simply returning 0 prevents the callback from being rescheduled. (closes issue ASTERISK-18570) Related issues that look like they are the same problem: (issue ASTERISK-17560) (issue ASTERISK-15406) (issue ASTERISK-15257) (issue ASTERISK-13334) (issue ASTERISK-9977) (issue ASTERISK-9716) Review: https://reviewboard.asterisk.org/r/1444/ ........ ................ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@336879 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2011-09-20 01:11:18 +00:00
if (!res) {
/*
Merged revisions 336878 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/10 ................ r336878 | russell | 2011-09-19 20:03:55 -0500 (Mon, 19 Sep 2011) | 43 lines Merged revisions 336877 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.8 ........ r336877 | russell | 2011-09-19 19:56:20 -0500 (Mon, 19 Sep 2011) | 36 lines Fix crashes in ast_rtcp_write(). This patch addresses crashes related to RTCP handling. The backtraces just show a crash in ast_rtcp_write() where it appears that the RTP instance is no longer valid. There is a race condition with scheduled RTCP transmissions and the destruction of the RTP instance. This patch utilizes the fact that ast_rtp_instance is a reference counted object and ensures that it will not get destroyed while a reference is still around due to scheduled RTCP transmissions. RTCP transmissions are scheduled and executed from the chan_sip scheduler context. This scheduler context is processed in the SIP monitor thread. The destruction of an RTP instance occurs when the associated sip_pvt gets destroyed (which happens when the sip_pvt reference count reaches 0). However, the SIP monitor thread is not the only thread that can cause a sip_pvt to get destroyed. The sip_hangup function, executed from a channel thread, also decrements the reference count on a sip_pvt and could cause it to get destroyed. While this is being changed anyway, the patch also removes calling ast_sched_del() from within the RTCP scheduler callback. It's not helpful. Simply returning 0 prevents the callback from being rescheduled. (closes issue ASTERISK-18570) Related issues that look like they are the same problem: (issue ASTERISK-17560) (issue ASTERISK-15406) (issue ASTERISK-15257) (issue ASTERISK-13334) (issue ASTERISK-9977) (issue ASTERISK-9716) Review: https://reviewboard.asterisk.org/r/1444/ ........ ................ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@336879 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2011-09-20 01:11:18 +00:00
* Not being rescheduled.
*/
rtp->rtcp->schedid = -1;
ao2_ref(instance, -1);
Merged revisions 336878 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/10 ................ r336878 | russell | 2011-09-19 20:03:55 -0500 (Mon, 19 Sep 2011) | 43 lines Merged revisions 336877 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.8 ........ r336877 | russell | 2011-09-19 19:56:20 -0500 (Mon, 19 Sep 2011) | 36 lines Fix crashes in ast_rtcp_write(). This patch addresses crashes related to RTCP handling. The backtraces just show a crash in ast_rtcp_write() where it appears that the RTP instance is no longer valid. There is a race condition with scheduled RTCP transmissions and the destruction of the RTP instance. This patch utilizes the fact that ast_rtp_instance is a reference counted object and ensures that it will not get destroyed while a reference is still around due to scheduled RTCP transmissions. RTCP transmissions are scheduled and executed from the chan_sip scheduler context. This scheduler context is processed in the SIP monitor thread. The destruction of an RTP instance occurs when the associated sip_pvt gets destroyed (which happens when the sip_pvt reference count reaches 0). However, the SIP monitor thread is not the only thread that can cause a sip_pvt to get destroyed. The sip_hangup function, executed from a channel thread, also decrements the reference count on a sip_pvt and could cause it to get destroyed. While this is being changed anyway, the patch also removes calling ast_sched_del() from within the RTCP scheduler callback. It's not helpful. Simply returning 0 prevents the callback from being rescheduled. (closes issue ASTERISK-18570) Related issues that look like they are the same problem: (issue ASTERISK-17560) (issue ASTERISK-15406) (issue ASTERISK-15257) (issue ASTERISK-13334) (issue ASTERISK-9977) (issue ASTERISK-9716) Review: https://reviewboard.asterisk.org/r/1444/ ........ ................ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@336879 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2011-09-20 01:11:18 +00:00
}
return res;
}
static void put_unaligned_time24(void *p, uint32_t time_msw, uint32_t time_lsw)
{
unsigned char *cp = p;
uint32_t datum;
/* Convert the time to 6.18 format */
datum = (time_msw << 18) & 0x00fc0000;
datum |= (time_lsw >> 14) & 0x0003ffff;
cp[0] = datum >> 16;
cp[1] = datum >> 8;
cp[2] = datum;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static int rtp_raw_write(struct ast_rtp_instance *instance, struct ast_frame *frame, int codec)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
int pred, mark = 0;
unsigned int ms = calc_txstamp(rtp, &frame->delivery);
struct ast_sockaddr remote_address = { {0,} };
int rate = ast_rtp_get_rate(frame->subclass.format) / 1000;
unsigned int seqno;
#ifdef TEST_FRAMEWORK
struct ast_rtp_engine_test *test = ast_rtp_instance_get_test(instance);
#endif
media formats: re-architect handling of media for performance improvements In the old times media formats were represented using a bit field. This was fast but had a few limitations. 1. Asterisk was limited in how many formats it could handle. 2. Formats, being a bit field, could not include any attribute information. A format was strictly its type, e.g., "this is ulaw". This was changed in Asterisk 10 (see https://wiki.asterisk.org/wiki/display/AST/Media+Architecture+Proposal for notes on that work) which led to the creation of the ast_format structure. This structure allowed Asterisk to handle attributes and bundle information with a format. Additionally, ast_format_cap was created to act as a container for multiple formats that, together, formed the capability of some entity. Another mechanism was added to allow logic to be registered which performed format attribute negotiation. Everywhere throughout the codebase Asterisk was changed to use this strategy. Unfortunately, in software, there is no free lunch. These new capabilities came at a cost. Performance analysis and profiling showed that we spend an inordinate amount of time comparing, copying, and generally manipulating formats and their related structures. Basic prototyping has shown that a reasonably large performance improvement could be made in this area. This patch is the result of that project, which overhauled the media format architecture and its usage in Asterisk to improve performance. Generally, the new philosophy for handling formats is as follows: * The ast_format structure is reference counted. This removed a large amount of the memory allocations and copying that was done in prior versions. * In order to prevent race conditions while keeping things performant, the ast_format structure is immutable by convention and lock-free. Violate this tenet at your peril! * Because formats are reference counted, codecs are also reference counted. The Asterisk core generally provides built-in codecs and caches the ast_format structures created to represent them. Generally, to prevent inordinate amounts of module reference bumping, codecs and formats can be added at run-time but cannot be removed. * All compatibility with the bit field representation of codecs/formats has been moved to a compatibility API. The primary user of this representation is chan_iax2, which must continue to maintain its bit-field usage of formats for interoperability concerns. * When a format is negotiated with attributes, or when a format cannot be represented by one of the cached formats, a new format object is created or cloned from an existing format. That format may have the same codec underlying it, but is a different format than a version of the format with different attributes or without attributes. * While formats are reference counted objects, the reference count maintained on the format should be manipulated with care. Formats are generally cached and will persist for the lifetime of Asterisk and do not explicitly need to have their lifetime modified. An exception to this is when the user of a format does not know where the format came from *and* the user may outlive the provider of the format. This occurs, for example, when a format is read from a channel: the channel may have a format with attributes (hence, non-cached) and the user of the format may last longer than the channel (if the reference to the channel is released prior to the format's reference). For more information on this work, see the API design notes: https://wiki.asterisk.org/wiki/display/AST/Media+Format+Rewrite Finally, this work was the culmination of a large number of developer's efforts. Extra thanks goes to Corey Farrell, who took on a large amount of the work in the Asterisk core, chan_sip, and was an invaluable resource in peer reviews throughout this project. There were a substantial number of patches contributed during this work; the following issues/patch names simply reflect some of the work (and will cause the release scripts to give attribution to the individuals who work on them). Reviews: https://reviewboard.asterisk.org/r/3814 https://reviewboard.asterisk.org/r/3808 https://reviewboard.asterisk.org/r/3805 https://reviewboard.asterisk.org/r/3803 https://reviewboard.asterisk.org/r/3801 https://reviewboard.asterisk.org/r/3798 https://reviewboard.asterisk.org/r/3800 https://reviewboard.asterisk.org/r/3794 https://reviewboard.asterisk.org/r/3793 https://reviewboard.asterisk.org/r/3792 https://reviewboard.asterisk.org/r/3791 https://reviewboard.asterisk.org/r/3790 https://reviewboard.asterisk.org/r/3789 https://reviewboard.asterisk.org/r/3788 https://reviewboard.asterisk.org/r/3787 https://reviewboard.asterisk.org/r/3786 https://reviewboard.asterisk.org/r/3784 https://reviewboard.asterisk.org/r/3783 https://reviewboard.asterisk.org/r/3778 https://reviewboard.asterisk.org/r/3774 https://reviewboard.asterisk.org/r/3775 https://reviewboard.asterisk.org/r/3772 https://reviewboard.asterisk.org/r/3761 https://reviewboard.asterisk.org/r/3754 https://reviewboard.asterisk.org/r/3753 https://reviewboard.asterisk.org/r/3751 https://reviewboard.asterisk.org/r/3750 https://reviewboard.asterisk.org/r/3748 https://reviewboard.asterisk.org/r/3747 https://reviewboard.asterisk.org/r/3746 https://reviewboard.asterisk.org/r/3742 https://reviewboard.asterisk.org/r/3740 https://reviewboard.asterisk.org/r/3739 https://reviewboard.asterisk.org/r/3738 https://reviewboard.asterisk.org/r/3737 https://reviewboard.asterisk.org/r/3736 https://reviewboard.asterisk.org/r/3734 https://reviewboard.asterisk.org/r/3722 https://reviewboard.asterisk.org/r/3713 https://reviewboard.asterisk.org/r/3703 https://reviewboard.asterisk.org/r/3689 https://reviewboard.asterisk.org/r/3687 https://reviewboard.asterisk.org/r/3674 https://reviewboard.asterisk.org/r/3671 https://reviewboard.asterisk.org/r/3667 https://reviewboard.asterisk.org/r/3665 https://reviewboard.asterisk.org/r/3625 https://reviewboard.asterisk.org/r/3602 https://reviewboard.asterisk.org/r/3519 https://reviewboard.asterisk.org/r/3518 https://reviewboard.asterisk.org/r/3516 https://reviewboard.asterisk.org/r/3515 https://reviewboard.asterisk.org/r/3512 https://reviewboard.asterisk.org/r/3506 https://reviewboard.asterisk.org/r/3413 https://reviewboard.asterisk.org/r/3410 https://reviewboard.asterisk.org/r/3387 https://reviewboard.asterisk.org/r/3388 https://reviewboard.asterisk.org/r/3389 https://reviewboard.asterisk.org/r/3390 https://reviewboard.asterisk.org/r/3321 https://reviewboard.asterisk.org/r/3320 https://reviewboard.asterisk.org/r/3319 https://reviewboard.asterisk.org/r/3318 https://reviewboard.asterisk.org/r/3266 https://reviewboard.asterisk.org/r/3265 https://reviewboard.asterisk.org/r/3234 https://reviewboard.asterisk.org/r/3178 ASTERISK-23114 #close Reported by: mjordan media_formats_translation_core.diff uploaded by kharwell (License 6464) rb3506.diff uploaded by mjordan (License 6283) media_format_app_file.diff uploaded by kharwell (License 6464) misc-2.diff uploaded by file (License 5000) chan_mild-3.diff uploaded by file (License 5000) chan_obscure.diff uploaded by file (License 5000) jingle.diff uploaded by file (License 5000) funcs.diff uploaded by file (License 5000) formats.diff uploaded by file (License 5000) core.diff uploaded by file (License 5000) bridges.diff uploaded by file (License 5000) mf-codecs-2.diff uploaded by file (License 5000) mf-app_fax.diff uploaded by file (License 5000) mf-apps-3.diff uploaded by file (License 5000) media-formats-3.diff uploaded by file (License 5000) ASTERISK-23715 rb3713.patch uploaded by coreyfarrell (License 5909) rb3689.patch uploaded by mjordan (License 6283) ASTERISK-23957 rb3722.patch uploaded by mjordan (License 6283) mf-attributes-3.diff uploaded by file (License 5000) ASTERISK-23958 Tested by: jrose rb3822.patch uploaded by coreyfarrell (License 5909) rb3800.patch uploaded by jrose (License 6182) chan_sip.diff uploaded by mjordan (License 6283) rb3747.patch uploaded by jrose (License 6182) ASTERISK-23959 #close Tested by: sgriepentrog, mjordan, coreyfarrell sip_cleanup.diff uploaded by opticron (License 6273) chan_sip_caps.diff uploaded by mjordan (License 6283) rb3751.patch uploaded by coreyfarrell (License 5909) chan_sip-3.diff uploaded by file (License 5000) ASTERISK-23960 #close Tested by: opticron direct_media.diff uploaded by opticron (License 6273) pjsip-direct-media.diff uploaded by file (License 5000) format_cap_remove.diff uploaded by opticron (License 6273) media_format_fixes.diff uploaded by opticron (License 6273) chan_pjsip-2.diff uploaded by file (License 5000) ASTERISK-23966 #close Tested by: rmudgett rb3803.patch uploaded by rmudgetti (License 5621) chan_dahdi.diff uploaded by file (License 5000) ASTERISK-24064 #close Tested by: coreyfarrell, mjordan, opticron, file, rmudgett, sgriepentrog, jrose rb3814.patch uploaded by rmudgett (License 5621) moh_cleanup.diff uploaded by opticron (License 6273) bridge_leak.diff uploaded by opticron (License 6273) translate.diff uploaded by file (License 5000) rb3795.patch uploaded by rmudgett (License 5621) tls_fix.diff uploaded by mjordan (License 6283) fax-mf-fix-2.diff uploaded by file (License 5000) rtp_transfer_stuff uploaded by mjordan (License 6283) rb3787.patch uploaded by rmudgett (License 5621) media-formats-explicit-translate-format-3.diff uploaded by file (License 5000) format_cache_case_fix.diff uploaded by opticron (License 6273) rb3774.patch uploaded by rmudgett (License 5621) rb3775.patch uploaded by rmudgett (License 5621) rtp_engine_fix.diff uploaded by opticron (License 6273) rtp_crash_fix.diff uploaded by opticron (License 6273) rb3753.patch uploaded by mjordan (License 6283) rb3750.patch uploaded by mjordan (License 6283) rb3748.patch uploaded by rmudgett (License 5621) media_format_fixes.diff uploaded by opticron (License 6273) rb3740.patch uploaded by mjordan (License 6283) rb3739.patch uploaded by mjordan (License 6283) rb3734.patch uploaded by mjordan (License 6283) rb3689.patch uploaded by mjordan (License 6283) rb3674.patch uploaded by coreyfarrell (License 5909) rb3671.patch uploaded by coreyfarrell (License 5909) rb3667.patch uploaded by coreyfarrell (License 5909) rb3665.patch uploaded by mjordan (License 6283) rb3625.patch uploaded by coreyfarrell (License 5909) rb3602.patch uploaded by coreyfarrell (License 5909) format_compatibility-2.diff uploaded by file (License 5000) core.diff uploaded by file (License 5000) git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@419044 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-07-20 22:06:33 +00:00
if (ast_format_cmp(frame->subclass.format, ast_format_g722) == AST_FORMAT_CMP_EQUAL) {
frame->samples /= 2;
}
if (rtp->sending_digit) {
return 0;
}
#ifdef TEST_FRAMEWORK
if (test && test->send_report) {
test->send_report = 0;
ast_rtcp_write(instance);
return 0;
}
#endif
if (frame->frametype == AST_FRAME_VOICE) {
pred = rtp->lastts + frame->samples;
/* Re-calculate last TS */
rtp->lastts = rtp->lastts + ms * rate;
if (ast_tvzero(frame->delivery)) {
/* If this isn't an absolute delivery time, Check if it is close to our prediction,
and if so, go with our prediction */
if (abs((int)rtp->lastts - pred) < MAX_TIMESTAMP_SKEW) {
rtp->lastts = pred;
} else {
ast_debug_rtp(3, "(%p) RTP audio difference is %d, ms is %u\n",
instance, abs((int)rtp->lastts - pred), ms);
mark = 1;
}
}
} else if (frame->frametype == AST_FRAME_VIDEO) {
media formats: re-architect handling of media for performance improvements In the old times media formats were represented using a bit field. This was fast but had a few limitations. 1. Asterisk was limited in how many formats it could handle. 2. Formats, being a bit field, could not include any attribute information. A format was strictly its type, e.g., "this is ulaw". This was changed in Asterisk 10 (see https://wiki.asterisk.org/wiki/display/AST/Media+Architecture+Proposal for notes on that work) which led to the creation of the ast_format structure. This structure allowed Asterisk to handle attributes and bundle information with a format. Additionally, ast_format_cap was created to act as a container for multiple formats that, together, formed the capability of some entity. Another mechanism was added to allow logic to be registered which performed format attribute negotiation. Everywhere throughout the codebase Asterisk was changed to use this strategy. Unfortunately, in software, there is no free lunch. These new capabilities came at a cost. Performance analysis and profiling showed that we spend an inordinate amount of time comparing, copying, and generally manipulating formats and their related structures. Basic prototyping has shown that a reasonably large performance improvement could be made in this area. This patch is the result of that project, which overhauled the media format architecture and its usage in Asterisk to improve performance. Generally, the new philosophy for handling formats is as follows: * The ast_format structure is reference counted. This removed a large amount of the memory allocations and copying that was done in prior versions. * In order to prevent race conditions while keeping things performant, the ast_format structure is immutable by convention and lock-free. Violate this tenet at your peril! * Because formats are reference counted, codecs are also reference counted. The Asterisk core generally provides built-in codecs and caches the ast_format structures created to represent them. Generally, to prevent inordinate amounts of module reference bumping, codecs and formats can be added at run-time but cannot be removed. * All compatibility with the bit field representation of codecs/formats has been moved to a compatibility API. The primary user of this representation is chan_iax2, which must continue to maintain its bit-field usage of formats for interoperability concerns. * When a format is negotiated with attributes, or when a format cannot be represented by one of the cached formats, a new format object is created or cloned from an existing format. That format may have the same codec underlying it, but is a different format than a version of the format with different attributes or without attributes. * While formats are reference counted objects, the reference count maintained on the format should be manipulated with care. Formats are generally cached and will persist for the lifetime of Asterisk and do not explicitly need to have their lifetime modified. An exception to this is when the user of a format does not know where the format came from *and* the user may outlive the provider of the format. This occurs, for example, when a format is read from a channel: the channel may have a format with attributes (hence, non-cached) and the user of the format may last longer than the channel (if the reference to the channel is released prior to the format's reference). For more information on this work, see the API design notes: https://wiki.asterisk.org/wiki/display/AST/Media+Format+Rewrite Finally, this work was the culmination of a large number of developer's efforts. Extra thanks goes to Corey Farrell, who took on a large amount of the work in the Asterisk core, chan_sip, and was an invaluable resource in peer reviews throughout this project. There were a substantial number of patches contributed during this work; the following issues/patch names simply reflect some of the work (and will cause the release scripts to give attribution to the individuals who work on them). Reviews: https://reviewboard.asterisk.org/r/3814 https://reviewboard.asterisk.org/r/3808 https://reviewboard.asterisk.org/r/3805 https://reviewboard.asterisk.org/r/3803 https://reviewboard.asterisk.org/r/3801 https://reviewboard.asterisk.org/r/3798 https://reviewboard.asterisk.org/r/3800 https://reviewboard.asterisk.org/r/3794 https://reviewboard.asterisk.org/r/3793 https://reviewboard.asterisk.org/r/3792 https://reviewboard.asterisk.org/r/3791 https://reviewboard.asterisk.org/r/3790 https://reviewboard.asterisk.org/r/3789 https://reviewboard.asterisk.org/r/3788 https://reviewboard.asterisk.org/r/3787 https://reviewboard.asterisk.org/r/3786 https://reviewboard.asterisk.org/r/3784 https://reviewboard.asterisk.org/r/3783 https://reviewboard.asterisk.org/r/3778 https://reviewboard.asterisk.org/r/3774 https://reviewboard.asterisk.org/r/3775 https://reviewboard.asterisk.org/r/3772 https://reviewboard.asterisk.org/r/3761 https://reviewboard.asterisk.org/r/3754 https://reviewboard.asterisk.org/r/3753 https://reviewboard.asterisk.org/r/3751 https://reviewboard.asterisk.org/r/3750 https://reviewboard.asterisk.org/r/3748 https://reviewboard.asterisk.org/r/3747 https://reviewboard.asterisk.org/r/3746 https://reviewboard.asterisk.org/r/3742 https://reviewboard.asterisk.org/r/3740 https://reviewboard.asterisk.org/r/3739 https://reviewboard.asterisk.org/r/3738 https://reviewboard.asterisk.org/r/3737 https://reviewboard.asterisk.org/r/3736 https://reviewboard.asterisk.org/r/3734 https://reviewboard.asterisk.org/r/3722 https://reviewboard.asterisk.org/r/3713 https://reviewboard.asterisk.org/r/3703 https://reviewboard.asterisk.org/r/3689 https://reviewboard.asterisk.org/r/3687 https://reviewboard.asterisk.org/r/3674 https://reviewboard.asterisk.org/r/3671 https://reviewboard.asterisk.org/r/3667 https://reviewboard.asterisk.org/r/3665 https://reviewboard.asterisk.org/r/3625 https://reviewboard.asterisk.org/r/3602 https://reviewboard.asterisk.org/r/3519 https://reviewboard.asterisk.org/r/3518 https://reviewboard.asterisk.org/r/3516 https://reviewboard.asterisk.org/r/3515 https://reviewboard.asterisk.org/r/3512 https://reviewboard.asterisk.org/r/3506 https://reviewboard.asterisk.org/r/3413 https://reviewboard.asterisk.org/r/3410 https://reviewboard.asterisk.org/r/3387 https://reviewboard.asterisk.org/r/3388 https://reviewboard.asterisk.org/r/3389 https://reviewboard.asterisk.org/r/3390 https://reviewboard.asterisk.org/r/3321 https://reviewboard.asterisk.org/r/3320 https://reviewboard.asterisk.org/r/3319 https://reviewboard.asterisk.org/r/3318 https://reviewboard.asterisk.org/r/3266 https://reviewboard.asterisk.org/r/3265 https://reviewboard.asterisk.org/r/3234 https://reviewboard.asterisk.org/r/3178 ASTERISK-23114 #close Reported by: mjordan media_formats_translation_core.diff uploaded by kharwell (License 6464) rb3506.diff uploaded by mjordan (License 6283) media_format_app_file.diff uploaded by kharwell (License 6464) misc-2.diff uploaded by file (License 5000) chan_mild-3.diff uploaded by file (License 5000) chan_obscure.diff uploaded by file (License 5000) jingle.diff uploaded by file (License 5000) funcs.diff uploaded by file (License 5000) formats.diff uploaded by file (License 5000) core.diff uploaded by file (License 5000) bridges.diff uploaded by file (License 5000) mf-codecs-2.diff uploaded by file (License 5000) mf-app_fax.diff uploaded by file (License 5000) mf-apps-3.diff uploaded by file (License 5000) media-formats-3.diff uploaded by file (License 5000) ASTERISK-23715 rb3713.patch uploaded by coreyfarrell (License 5909) rb3689.patch uploaded by mjordan (License 6283) ASTERISK-23957 rb3722.patch uploaded by mjordan (License 6283) mf-attributes-3.diff uploaded by file (License 5000) ASTERISK-23958 Tested by: jrose rb3822.patch uploaded by coreyfarrell (License 5909) rb3800.patch uploaded by jrose (License 6182) chan_sip.diff uploaded by mjordan (License 6283) rb3747.patch uploaded by jrose (License 6182) ASTERISK-23959 #close Tested by: sgriepentrog, mjordan, coreyfarrell sip_cleanup.diff uploaded by opticron (License 6273) chan_sip_caps.diff uploaded by mjordan (License 6283) rb3751.patch uploaded by coreyfarrell (License 5909) chan_sip-3.diff uploaded by file (License 5000) ASTERISK-23960 #close Tested by: opticron direct_media.diff uploaded by opticron (License 6273) pjsip-direct-media.diff uploaded by file (License 5000) format_cap_remove.diff uploaded by opticron (License 6273) media_format_fixes.diff uploaded by opticron (License 6273) chan_pjsip-2.diff uploaded by file (License 5000) ASTERISK-23966 #close Tested by: rmudgett rb3803.patch uploaded by rmudgetti (License 5621) chan_dahdi.diff uploaded by file (License 5000) ASTERISK-24064 #close Tested by: coreyfarrell, mjordan, opticron, file, rmudgett, sgriepentrog, jrose rb3814.patch uploaded by rmudgett (License 5621) moh_cleanup.diff uploaded by opticron (License 6273) bridge_leak.diff uploaded by opticron (License 6273) translate.diff uploaded by file (License 5000) rb3795.patch uploaded by rmudgett (License 5621) tls_fix.diff uploaded by mjordan (License 6283) fax-mf-fix-2.diff uploaded by file (License 5000) rtp_transfer_stuff uploaded by mjordan (License 6283) rb3787.patch uploaded by rmudgett (License 5621) media-formats-explicit-translate-format-3.diff uploaded by file (License 5000) format_cache_case_fix.diff uploaded by opticron (License 6273) rb3774.patch uploaded by rmudgett (License 5621) rb3775.patch uploaded by rmudgett (License 5621) rtp_engine_fix.diff uploaded by opticron (License 6273) rtp_crash_fix.diff uploaded by opticron (License 6273) rb3753.patch uploaded by mjordan (License 6283) rb3750.patch uploaded by mjordan (License 6283) rb3748.patch uploaded by rmudgett (License 5621) media_format_fixes.diff uploaded by opticron (License 6273) rb3740.patch uploaded by mjordan (License 6283) rb3739.patch uploaded by mjordan (License 6283) rb3734.patch uploaded by mjordan (License 6283) rb3689.patch uploaded by mjordan (License 6283) rb3674.patch uploaded by coreyfarrell (License 5909) rb3671.patch uploaded by coreyfarrell (License 5909) rb3667.patch uploaded by coreyfarrell (License 5909) rb3665.patch uploaded by mjordan (License 6283) rb3625.patch uploaded by coreyfarrell (License 5909) rb3602.patch uploaded by coreyfarrell (License 5909) format_compatibility-2.diff uploaded by file (License 5000) core.diff uploaded by file (License 5000) git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@419044 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-07-20 22:06:33 +00:00
mark = frame->subclass.frame_ending;
pred = rtp->lastovidtimestamp + frame->samples;
/* Re-calculate last TS */
rtp->lastts = rtp->lastts + ms * 90;
/* If it's close to our prediction, go for it */
if (ast_tvzero(frame->delivery)) {
if (abs((int)rtp->lastts - pred) < 7200) {
rtp->lastts = pred;
rtp->lastovidtimestamp += frame->samples;
} else {
ast_debug_rtp(3, "(%p) RTP video difference is %d, ms is %u (%u), pred/ts/samples %u/%d/%d\n",
instance, abs((int)rtp->lastts - pred), ms, ms * 90, rtp->lastts, pred, frame->samples);
rtp->lastovidtimestamp = rtp->lastts;
}
}
} else {
pred = rtp->lastotexttimestamp + frame->samples;
/* Re-calculate last TS */
rtp->lastts = rtp->lastts + ms;
/* If it's close to our prediction, go for it */
if (ast_tvzero(frame->delivery)) {
if (abs((int)rtp->lastts - pred) < 7200) {
rtp->lastts = pred;
rtp->lastotexttimestamp += frame->samples;
} else {
ast_debug_rtp(3, "(%p) RTP other difference is %d, ms is %u, pred/ts/samples %u/%d/%d\n",
instance, abs((int)rtp->lastts - pred), ms, rtp->lastts, pred, frame->samples);
rtp->lastotexttimestamp = rtp->lastts;
}
}
}
/* If we have been explicitly told to set the marker bit then do so */
if (ast_test_flag(rtp, FLAG_NEED_MARKER_BIT)) {
mark = 1;
ast_clear_flag(rtp, FLAG_NEED_MARKER_BIT);
}
/* If the timestamp for non-digt packets has moved beyond the timestamp for digits, update the digit timestamp */
if (rtp->lastts > rtp->lastdigitts) {
rtp->lastdigitts = rtp->lastts;
}
/* Assume that the sequence number we expect to use is what will be used until proven otherwise */
seqno = rtp->seqno;
/* If the frame contains sequence number information use it to influence our sequence number */
if (ast_test_flag(frame, AST_FRFLAG_HAS_SEQUENCE_NUMBER)) {
if (rtp->expectedseqno != -1) {
/* Determine where the frame from the core is in relation to where we expected */
int difference = frame->seqno - rtp->expectedseqno;
/* If there is a substantial difference then we've either got packets really out
* of order, or the source is RTP and it has cycled. If this happens we resync
* the sequence number adjustments to this frame. If we also have packet loss
* things won't be reflected correctly but it will sort itself out after a bit.
*/
if (abs(difference) > 100) {
difference = 0;
}
/* Adjust the sequence number being used for this packet accordingly */
seqno += difference;
if (difference >= 0) {
/* This frame is on time or in the future */
rtp->expectedseqno = frame->seqno + 1;
rtp->seqno += difference;
}
} else {
/* This is the first frame with sequence number we've seen, so start keeping track */
rtp->expectedseqno = frame->seqno + 1;
}
} else {
rtp->expectedseqno = -1;
}
if (ast_test_flag(frame, AST_FRFLAG_HAS_TIMING_INFO)) {
rtp->lastts = frame->ts * rate;
}
ast_rtp_instance_get_remote_address(instance, &remote_address);
/* If we know the remote address construct a packet and send it out */
if (!ast_sockaddr_isnull(&remote_address)) {
int hdrlen = 12;
int res;
int ice;
int ext = 0;
int abs_send_time_id;
int packet_len;
unsigned char *rtpheader;
/* If the abs-send-time extension has been negotiated determine how much space we need */
abs_send_time_id = ast_rtp_instance_extmap_get_id(instance, AST_RTP_EXTENSION_ABS_SEND_TIME);
if (abs_send_time_id != -1) {
/* 4 bytes for the shared information, 1 byte for identifier, 3 bytes for abs-send-time */
hdrlen += 8;
ext = 1;
}
packet_len = frame->datalen + hdrlen;
rtpheader = (unsigned char *)(frame->data.ptr - hdrlen);
put_unaligned_uint32(rtpheader, htonl((2 << 30) | (ext << 28) | (codec << 16) | (seqno) | (mark << 23)));
put_unaligned_uint32(rtpheader + 4, htonl(rtp->lastts));
put_unaligned_uint32(rtpheader + 8, htonl(rtp->ssrc));
/* We assume right now that we will only ever have the abs-send-time extension in the packet
* which simplifies things a bit.
*/
if (abs_send_time_id != -1) {
unsigned int now_msw;
unsigned int now_lsw;
/* This happens before being placed into the retransmission buffer so that when we
* retransmit we only have to update the timestamp, not everything else.
*/
put_unaligned_uint32(rtpheader + 12, htonl((0xBEDE << 16) | 1));
rtpheader[16] = (abs_send_time_id << 4) | 2;
timeval2ntp(ast_tvnow(), &now_msw, &now_lsw);
put_unaligned_time24(rtpheader + 17, now_msw, now_lsw);
}
/* If retransmissions are enabled, we need to store this packet for future use */
if (rtp->send_buffer) {
struct ast_rtp_rtcp_nack_payload *payload;
payload = ast_malloc(sizeof(*payload) + packet_len);
if (payload) {
payload->size = packet_len;
memcpy(payload->buf, rtpheader, packet_len);
if (ast_data_buffer_put(rtp->send_buffer, rtp->seqno, payload) == -1) {
ast_free(payload);
}
}
}
res = rtp_sendto(instance, (void *)rtpheader, packet_len, 0, &remote_address, &ice);
if (res < 0) {
if (!ast_rtp_instance_get_prop(instance, AST_RTP_PROPERTY_NAT) || (ast_rtp_instance_get_prop(instance, AST_RTP_PROPERTY_NAT) && (ast_test_flag(rtp, FLAG_NAT_ACTIVE) == FLAG_NAT_ACTIVE))) {
ast_debug_rtp(1, "(%p) RTP transmission error of packet %d to %s: %s\n",
instance, rtp->seqno,
ast_sockaddr_stringify(&remote_address),
strerror(errno));
} else if (((ast_test_flag(rtp, FLAG_NAT_ACTIVE) == FLAG_NAT_INACTIVE) || ast_debug_rtp_packet_is_allowed) && !ast_test_flag(rtp, FLAG_NAT_INACTIVE_NOWARN)) {
/* Only give this error message once if we are not RTP debugging */
if (ast_debug_rtp_packet_is_allowed)
ast_debug(0, "(%p) RTP NAT: Can't write RTP to private address %s, waiting for other end to send audio...\n",
instance, ast_sockaddr_stringify(&remote_address));
ast_set_flag(rtp, FLAG_NAT_INACTIVE_NOWARN);
}
} else {
if (rtp->rtcp && rtp->rtcp->schedid < 0) {
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
ast_debug_rtcp(2, "(%s) RTCP starting transmission in %u ms\n",
ast_rtp_instance_get_channel_id(instance), ast_rtcp_calc_interval(rtp));
Merged revisions 336878 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/10 ................ r336878 | russell | 2011-09-19 20:03:55 -0500 (Mon, 19 Sep 2011) | 43 lines Merged revisions 336877 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.8 ........ r336877 | russell | 2011-09-19 19:56:20 -0500 (Mon, 19 Sep 2011) | 36 lines Fix crashes in ast_rtcp_write(). This patch addresses crashes related to RTCP handling. The backtraces just show a crash in ast_rtcp_write() where it appears that the RTP instance is no longer valid. There is a race condition with scheduled RTCP transmissions and the destruction of the RTP instance. This patch utilizes the fact that ast_rtp_instance is a reference counted object and ensures that it will not get destroyed while a reference is still around due to scheduled RTCP transmissions. RTCP transmissions are scheduled and executed from the chan_sip scheduler context. This scheduler context is processed in the SIP monitor thread. The destruction of an RTP instance occurs when the associated sip_pvt gets destroyed (which happens when the sip_pvt reference count reaches 0). However, the SIP monitor thread is not the only thread that can cause a sip_pvt to get destroyed. The sip_hangup function, executed from a channel thread, also decrements the reference count on a sip_pvt and could cause it to get destroyed. While this is being changed anyway, the patch also removes calling ast_sched_del() from within the RTCP scheduler callback. It's not helpful. Simply returning 0 prevents the callback from being rescheduled. (closes issue ASTERISK-18570) Related issues that look like they are the same problem: (issue ASTERISK-17560) (issue ASTERISK-15406) (issue ASTERISK-15257) (issue ASTERISK-13334) (issue ASTERISK-9977) (issue ASTERISK-9716) Review: https://reviewboard.asterisk.org/r/1444/ ........ ................ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@336879 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2011-09-20 01:11:18 +00:00
ao2_ref(instance, +1);
rtp->rtcp->schedid = ast_sched_add(rtp->sched, ast_rtcp_calc_interval(rtp), ast_rtcp_write, instance);
Merged revisions 336878 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/10 ................ r336878 | russell | 2011-09-19 20:03:55 -0500 (Mon, 19 Sep 2011) | 43 lines Merged revisions 336877 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.8 ........ r336877 | russell | 2011-09-19 19:56:20 -0500 (Mon, 19 Sep 2011) | 36 lines Fix crashes in ast_rtcp_write(). This patch addresses crashes related to RTCP handling. The backtraces just show a crash in ast_rtcp_write() where it appears that the RTP instance is no longer valid. There is a race condition with scheduled RTCP transmissions and the destruction of the RTP instance. This patch utilizes the fact that ast_rtp_instance is a reference counted object and ensures that it will not get destroyed while a reference is still around due to scheduled RTCP transmissions. RTCP transmissions are scheduled and executed from the chan_sip scheduler context. This scheduler context is processed in the SIP monitor thread. The destruction of an RTP instance occurs when the associated sip_pvt gets destroyed (which happens when the sip_pvt reference count reaches 0). However, the SIP monitor thread is not the only thread that can cause a sip_pvt to get destroyed. The sip_hangup function, executed from a channel thread, also decrements the reference count on a sip_pvt and could cause it to get destroyed. While this is being changed anyway, the patch also removes calling ast_sched_del() from within the RTCP scheduler callback. It's not helpful. Simply returning 0 prevents the callback from being rescheduled. (closes issue ASTERISK-18570) Related issues that look like they are the same problem: (issue ASTERISK-17560) (issue ASTERISK-15406) (issue ASTERISK-15257) (issue ASTERISK-13334) (issue ASTERISK-9977) (issue ASTERISK-9716) Review: https://reviewboard.asterisk.org/r/1444/ ........ ................ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@336879 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2011-09-20 01:11:18 +00:00
if (rtp->rtcp->schedid < 0) {
ao2_ref(instance, -1);
ast_log(LOG_WARNING, "scheduling RTCP transmission failed.\n");
}
}
}
if (rtp_debug_test_addr(&remote_address)) {
ast_verbose("Sent RTP packet to %s%s (type %-2.2d, seq %-6.6d, ts %-6.6u, len %-6.6d)\n",
ast_sockaddr_stringify(&remote_address),
ice ? " (via ICE)" : "",
codec, rtp->seqno, rtp->lastts, res - hdrlen);
}
}
/* If the sequence number that has been used doesn't match what we expected then this is an out of
* order late packet, so we don't need to increment as we haven't yet gotten the expected frame from
* the core.
*/
if (seqno == rtp->seqno) {
rtp->seqno++;
}
return 0;
}
static struct ast_frame *red_t140_to_red(struct rtp_red *red)
{
unsigned char *data = red->t140red.data.ptr;
int len = 0;
int i;
/* replace most aged generation */
if (red->len[0]) {
for (i = 1; i < red->num_gen+1; i++)
len += red->len[i];
memmove(&data[red->hdrlen], &data[red->hdrlen+red->len[0]], len);
}
/* Store length of each generation and primary data length*/
for (i = 0; i < red->num_gen; i++)
red->len[i] = red->len[i+1];
red->len[i] = red->t140.datalen;
/* write each generation length in red header */
len = red->hdrlen;
for (i = 0; i < red->num_gen; i++) {
len += data[i*4+3] = red->len[i];
}
/* add primary data to buffer */
memcpy(&data[len], red->t140.data.ptr, red->t140.datalen);
red->t140red.datalen = len + red->t140.datalen;
/* no primary data and no generations to send */
if (len == red->hdrlen && !red->t140.datalen) {
return NULL;
}
/* reset t.140 buffer */
red->t140.datalen = 0;
return &red->t140red;
}
static void rtp_write_rtcp_fir(struct ast_rtp_instance *instance, struct ast_rtp *rtp, struct ast_sockaddr *remote_address)
{
unsigned char *rtcpheader;
unsigned char bdata[1024];
int packet_len = 0;
int fir_len = 20;
int ice;
int res;
int sr;
RAII_VAR(struct ast_rtp_rtcp_report *, rtcp_report, NULL, ao2_cleanup);
if (!rtp || !rtp->rtcp) {
return;
}
if (ast_sockaddr_isnull(&rtp->rtcp->them) || rtp->rtcp->schedid < 0) {
/*
* RTCP was stopped.
*/
return;
}
if (!rtp->themssrc_valid) {
/* We don't know their SSRC value so we don't know who to update. */
return;
}
/* Prepare RTCP FIR (PT=206, FMT=4) */
rtp->rtcp->firseq++;
if(rtp->rtcp->firseq == 256) {
rtp->rtcp->firseq = 0;
}
rtcpheader = bdata;
ao2_lock(instance);
rtcp_report = ast_rtp_rtcp_report_alloc(rtp->themssrc_valid ? 1 : 0);
res = ast_rtcp_generate_compound_prefix(instance, rtcpheader, rtcp_report, &sr);
if (res == 0 || res == 1) {
ao2_unlock(instance);
return;
}
packet_len += res;
put_unaligned_uint32(rtcpheader + packet_len + 0, htonl((2 << 30) | (4 << 24) | (RTCP_PT_PSFB << 16) | ((fir_len/4)-1)));
put_unaligned_uint32(rtcpheader + packet_len + 4, htonl(rtp->ssrc));
put_unaligned_uint32(rtcpheader + packet_len + 8, htonl(rtp->themssrc));
put_unaligned_uint32(rtcpheader + packet_len + 12, htonl(rtp->themssrc)); /* FCI: SSRC */
put_unaligned_uint32(rtcpheader + packet_len + 16, htonl(rtp->rtcp->firseq << 24)); /* FCI: Sequence number */
res = rtcp_sendto(instance, (unsigned int *)rtcpheader, packet_len + fir_len, 0, rtp->bundled ? remote_address : &rtp->rtcp->them, &ice);
if (res < 0) {
ast_log(LOG_ERROR, "RTCP FIR transmission error: %s\n", strerror(errno));
} else {
ast_rtcp_calculate_sr_rr_statistics(instance, rtcp_report, rtp->bundled ? *remote_address : rtp->rtcp->them, ice, sr);
}
ao2_unlock(instance);
}
static void rtp_write_rtcp_psfb(struct ast_rtp_instance *instance, struct ast_rtp *rtp, struct ast_frame *frame, struct ast_sockaddr *remote_address)
{
struct ast_rtp_rtcp_feedback *feedback = frame->data.ptr;
unsigned char *rtcpheader;
unsigned char bdata[1024];
int remb_len = 24;
int ice;
int res;
int sr = 0;
int packet_len = 0;
RAII_VAR(struct ast_rtp_rtcp_report *, rtcp_report, NULL, ao2_cleanup);
if (feedback->fmt != AST_RTP_RTCP_FMT_REMB) {
ast_debug_rtcp(1, "(%p) RTCP provided feedback frame of format %d to write, but only REMB is supported\n",
instance, feedback->fmt);
return;
}
if (!rtp || !rtp->rtcp) {
return;
}
/* If REMB support is not enabled don't send this RTCP packet */
if (!ast_rtp_instance_get_prop(instance, AST_RTP_PROPERTY_REMB)) {
ast_debug_rtcp(1, "(%p) RTCP provided feedback REMB report to write, but REMB support not enabled\n",
instance);
return;
}
if (ast_sockaddr_isnull(&rtp->rtcp->them) || rtp->rtcp->schedid < 0) {
/*
* RTCP was stopped.
*/
return;
}
rtcpheader = bdata;
ao2_lock(instance);
rtcp_report = ast_rtp_rtcp_report_alloc(rtp->themssrc_valid ? 1 : 0);
res = ast_rtcp_generate_compound_prefix(instance, rtcpheader, rtcp_report, &sr);
if (res == 0 || res == 1) {
ao2_unlock(instance);
return;
}
packet_len += res;
put_unaligned_uint32(rtcpheader + packet_len + 0, htonl((2 << 30) | (AST_RTP_RTCP_FMT_REMB << 24) | (RTCP_PT_PSFB << 16) | ((remb_len/4)-1)));
put_unaligned_uint32(rtcpheader + packet_len + 4, htonl(rtp->ssrc));
put_unaligned_uint32(rtcpheader + packet_len + 8, htonl(0)); /* Per the draft, this should always be 0 */
put_unaligned_uint32(rtcpheader + packet_len + 12, htonl(('R' << 24) | ('E' << 16) | ('M' << 8) | ('B'))); /* Unique identifier 'R' 'E' 'M' 'B' */
put_unaligned_uint32(rtcpheader + packet_len + 16, htonl((1 << 24) | (feedback->remb.br_exp << 18) | (feedback->remb.br_mantissa))); /* Number of SSRCs / BR Exp / BR Mantissa */
put_unaligned_uint32(rtcpheader + packet_len + 20, htonl(rtp->ssrc)); /* The SSRC this feedback message applies to */
res = rtcp_sendto(instance, (unsigned int *)rtcpheader, packet_len + remb_len, 0, rtp->bundled ? remote_address : &rtp->rtcp->them, &ice);
if (res < 0) {
ast_log(LOG_ERROR, "RTCP PSFB transmission error: %s\n", strerror(errno));
} else {
ast_rtcp_calculate_sr_rr_statistics(instance, rtcp_report, rtp->bundled ? *remote_address : rtp->rtcp->them, ice, sr);
}
ao2_unlock(instance);
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static int ast_rtp_write(struct ast_rtp_instance *instance, struct ast_frame *frame)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
struct ast_sockaddr remote_address = { {0,} };
media formats: re-architect handling of media for performance improvements In the old times media formats were represented using a bit field. This was fast but had a few limitations. 1. Asterisk was limited in how many formats it could handle. 2. Formats, being a bit field, could not include any attribute information. A format was strictly its type, e.g., "this is ulaw". This was changed in Asterisk 10 (see https://wiki.asterisk.org/wiki/display/AST/Media+Architecture+Proposal for notes on that work) which led to the creation of the ast_format structure. This structure allowed Asterisk to handle attributes and bundle information with a format. Additionally, ast_format_cap was created to act as a container for multiple formats that, together, formed the capability of some entity. Another mechanism was added to allow logic to be registered which performed format attribute negotiation. Everywhere throughout the codebase Asterisk was changed to use this strategy. Unfortunately, in software, there is no free lunch. These new capabilities came at a cost. Performance analysis and profiling showed that we spend an inordinate amount of time comparing, copying, and generally manipulating formats and their related structures. Basic prototyping has shown that a reasonably large performance improvement could be made in this area. This patch is the result of that project, which overhauled the media format architecture and its usage in Asterisk to improve performance. Generally, the new philosophy for handling formats is as follows: * The ast_format structure is reference counted. This removed a large amount of the memory allocations and copying that was done in prior versions. * In order to prevent race conditions while keeping things performant, the ast_format structure is immutable by convention and lock-free. Violate this tenet at your peril! * Because formats are reference counted, codecs are also reference counted. The Asterisk core generally provides built-in codecs and caches the ast_format structures created to represent them. Generally, to prevent inordinate amounts of module reference bumping, codecs and formats can be added at run-time but cannot be removed. * All compatibility with the bit field representation of codecs/formats has been moved to a compatibility API. The primary user of this representation is chan_iax2, which must continue to maintain its bit-field usage of formats for interoperability concerns. * When a format is negotiated with attributes, or when a format cannot be represented by one of the cached formats, a new format object is created or cloned from an existing format. That format may have the same codec underlying it, but is a different format than a version of the format with different attributes or without attributes. * While formats are reference counted objects, the reference count maintained on the format should be manipulated with care. Formats are generally cached and will persist for the lifetime of Asterisk and do not explicitly need to have their lifetime modified. An exception to this is when the user of a format does not know where the format came from *and* the user may outlive the provider of the format. This occurs, for example, when a format is read from a channel: the channel may have a format with attributes (hence, non-cached) and the user of the format may last longer than the channel (if the reference to the channel is released prior to the format's reference). For more information on this work, see the API design notes: https://wiki.asterisk.org/wiki/display/AST/Media+Format+Rewrite Finally, this work was the culmination of a large number of developer's efforts. Extra thanks goes to Corey Farrell, who took on a large amount of the work in the Asterisk core, chan_sip, and was an invaluable resource in peer reviews throughout this project. There were a substantial number of patches contributed during this work; the following issues/patch names simply reflect some of the work (and will cause the release scripts to give attribution to the individuals who work on them). Reviews: https://reviewboard.asterisk.org/r/3814 https://reviewboard.asterisk.org/r/3808 https://reviewboard.asterisk.org/r/3805 https://reviewboard.asterisk.org/r/3803 https://reviewboard.asterisk.org/r/3801 https://reviewboard.asterisk.org/r/3798 https://reviewboard.asterisk.org/r/3800 https://reviewboard.asterisk.org/r/3794 https://reviewboard.asterisk.org/r/3793 https://reviewboard.asterisk.org/r/3792 https://reviewboard.asterisk.org/r/3791 https://reviewboard.asterisk.org/r/3790 https://reviewboard.asterisk.org/r/3789 https://reviewboard.asterisk.org/r/3788 https://reviewboard.asterisk.org/r/3787 https://reviewboard.asterisk.org/r/3786 https://reviewboard.asterisk.org/r/3784 https://reviewboard.asterisk.org/r/3783 https://reviewboard.asterisk.org/r/3778 https://reviewboard.asterisk.org/r/3774 https://reviewboard.asterisk.org/r/3775 https://reviewboard.asterisk.org/r/3772 https://reviewboard.asterisk.org/r/3761 https://reviewboard.asterisk.org/r/3754 https://reviewboard.asterisk.org/r/3753 https://reviewboard.asterisk.org/r/3751 https://reviewboard.asterisk.org/r/3750 https://reviewboard.asterisk.org/r/3748 https://reviewboard.asterisk.org/r/3747 https://reviewboard.asterisk.org/r/3746 https://reviewboard.asterisk.org/r/3742 https://reviewboard.asterisk.org/r/3740 https://reviewboard.asterisk.org/r/3739 https://reviewboard.asterisk.org/r/3738 https://reviewboard.asterisk.org/r/3737 https://reviewboard.asterisk.org/r/3736 https://reviewboard.asterisk.org/r/3734 https://reviewboard.asterisk.org/r/3722 https://reviewboard.asterisk.org/r/3713 https://reviewboard.asterisk.org/r/3703 https://reviewboard.asterisk.org/r/3689 https://reviewboard.asterisk.org/r/3687 https://reviewboard.asterisk.org/r/3674 https://reviewboard.asterisk.org/r/3671 https://reviewboard.asterisk.org/r/3667 https://reviewboard.asterisk.org/r/3665 https://reviewboard.asterisk.org/r/3625 https://reviewboard.asterisk.org/r/3602 https://reviewboard.asterisk.org/r/3519 https://reviewboard.asterisk.org/r/3518 https://reviewboard.asterisk.org/r/3516 https://reviewboard.asterisk.org/r/3515 https://reviewboard.asterisk.org/r/3512 https://reviewboard.asterisk.org/r/3506 https://reviewboard.asterisk.org/r/3413 https://reviewboard.asterisk.org/r/3410 https://reviewboard.asterisk.org/r/3387 https://reviewboard.asterisk.org/r/3388 https://reviewboard.asterisk.org/r/3389 https://reviewboard.asterisk.org/r/3390 https://reviewboard.asterisk.org/r/3321 https://reviewboard.asterisk.org/r/3320 https://reviewboard.asterisk.org/r/3319 https://reviewboard.asterisk.org/r/3318 https://reviewboard.asterisk.org/r/3266 https://reviewboard.asterisk.org/r/3265 https://reviewboard.asterisk.org/r/3234 https://reviewboard.asterisk.org/r/3178 ASTERISK-23114 #close Reported by: mjordan media_formats_translation_core.diff uploaded by kharwell (License 6464) rb3506.diff uploaded by mjordan (License 6283) media_format_app_file.diff uploaded by kharwell (License 6464) misc-2.diff uploaded by file (License 5000) chan_mild-3.diff uploaded by file (License 5000) chan_obscure.diff uploaded by file (License 5000) jingle.diff uploaded by file (License 5000) funcs.diff uploaded by file (License 5000) formats.diff uploaded by file (License 5000) core.diff uploaded by file (License 5000) bridges.diff uploaded by file (License 5000) mf-codecs-2.diff uploaded by file (License 5000) mf-app_fax.diff uploaded by file (License 5000) mf-apps-3.diff uploaded by file (License 5000) media-formats-3.diff uploaded by file (License 5000) ASTERISK-23715 rb3713.patch uploaded by coreyfarrell (License 5909) rb3689.patch uploaded by mjordan (License 6283) ASTERISK-23957 rb3722.patch uploaded by mjordan (License 6283) mf-attributes-3.diff uploaded by file (License 5000) ASTERISK-23958 Tested by: jrose rb3822.patch uploaded by coreyfarrell (License 5909) rb3800.patch uploaded by jrose (License 6182) chan_sip.diff uploaded by mjordan (License 6283) rb3747.patch uploaded by jrose (License 6182) ASTERISK-23959 #close Tested by: sgriepentrog, mjordan, coreyfarrell sip_cleanup.diff uploaded by opticron (License 6273) chan_sip_caps.diff uploaded by mjordan (License 6283) rb3751.patch uploaded by coreyfarrell (License 5909) chan_sip-3.diff uploaded by file (License 5000) ASTERISK-23960 #close Tested by: opticron direct_media.diff uploaded by opticron (License 6273) pjsip-direct-media.diff uploaded by file (License 5000) format_cap_remove.diff uploaded by opticron (License 6273) media_format_fixes.diff uploaded by opticron (License 6273) chan_pjsip-2.diff uploaded by file (License 5000) ASTERISK-23966 #close Tested by: rmudgett rb3803.patch uploaded by rmudgetti (License 5621) chan_dahdi.diff uploaded by file (License 5000) ASTERISK-24064 #close Tested by: coreyfarrell, mjordan, opticron, file, rmudgett, sgriepentrog, jrose rb3814.patch uploaded by rmudgett (License 5621) moh_cleanup.diff uploaded by opticron (License 6273) bridge_leak.diff uploaded by opticron (License 6273) translate.diff uploaded by file (License 5000) rb3795.patch uploaded by rmudgett (License 5621) tls_fix.diff uploaded by mjordan (License 6283) fax-mf-fix-2.diff uploaded by file (License 5000) rtp_transfer_stuff uploaded by mjordan (License 6283) rb3787.patch uploaded by rmudgett (License 5621) media-formats-explicit-translate-format-3.diff uploaded by file (License 5000) format_cache_case_fix.diff uploaded by opticron (License 6273) rb3774.patch uploaded by rmudgett (License 5621) rb3775.patch uploaded by rmudgett (License 5621) rtp_engine_fix.diff uploaded by opticron (License 6273) rtp_crash_fix.diff uploaded by opticron (License 6273) rb3753.patch uploaded by mjordan (License 6283) rb3750.patch uploaded by mjordan (License 6283) rb3748.patch uploaded by rmudgett (License 5621) media_format_fixes.diff uploaded by opticron (License 6273) rb3740.patch uploaded by mjordan (License 6283) rb3739.patch uploaded by mjordan (License 6283) rb3734.patch uploaded by mjordan (License 6283) rb3689.patch uploaded by mjordan (License 6283) rb3674.patch uploaded by coreyfarrell (License 5909) rb3671.patch uploaded by coreyfarrell (License 5909) rb3667.patch uploaded by coreyfarrell (License 5909) rb3665.patch uploaded by mjordan (License 6283) rb3625.patch uploaded by coreyfarrell (License 5909) rb3602.patch uploaded by coreyfarrell (License 5909) format_compatibility-2.diff uploaded by file (License 5000) core.diff uploaded by file (License 5000) git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@419044 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-07-20 22:06:33 +00:00
struct ast_format *format;
int codec;
ast_rtp_instance_get_remote_address(instance, &remote_address);
/* If we don't actually know the remote address don't even bother doing anything */
if (ast_sockaddr_isnull(&remote_address)) {
ast_debug_rtp(1, "(%p) RTP no remote address on instance, so dropping frame\n", instance);
return 0;
}
/* VP8: is this a request to send a RTCP FIR? */
if (frame->frametype == AST_FRAME_CONTROL && frame->subclass.integer == AST_CONTROL_VIDUPDATE) {
rtp_write_rtcp_fir(instance, rtp, &remote_address);
return 0;
} else if (frame->frametype == AST_FRAME_RTCP) {
if (frame->subclass.integer == AST_RTP_RTCP_PSFB) {
rtp_write_rtcp_psfb(instance, rtp, frame, &remote_address);
}
return 0;
}
/* If there is no data length we can't very well send the packet */
if (!frame->datalen) {
ast_debug_rtp(1, "(%p) RTP received frame with no data for instance, so dropping frame\n", instance);
return 0;
}
/* If the packet is not one our RTP stack supports bail out */
if (frame->frametype != AST_FRAME_VOICE && frame->frametype != AST_FRAME_VIDEO && frame->frametype != AST_FRAME_TEXT) {
ast_log(LOG_WARNING, "RTP can only send voice, video, and text\n");
return -1;
}
if (rtp->red) {
/* return 0; */
/* no primary data or generations to send */
if ((frame = red_t140_to_red(rtp->red)) == NULL)
return 0;
}
/* Grab the subclass and look up the payload we are going to use */
rtp_engine.c: Initial split of payload types into rx and tx mappings. There are numerous problems with the current implementation of the RTP payload type mapping in Asterisk. It uses only one mapping structure to associate payload types to codecs. The single mapping is overkill if all of the payload type values are well known values. Dynamic payload type mappings do not work as well with the single mapping because RFC3264 allows each side of the link to negotiate different dynamic mappings for what they want to receive. Not only could you have the same codec mapped for sending and receiving on different payload types you could wind up with the same payload type mapped to different codecs for each direction. 1) An independent payload type mapping is needed for sending and receiving. 2) The receive mapping needs to keep track of previous mappings because of the slack to when negotiation happens and current packets in flight using the old mapping arrive. 3) The transmit mapping only needs to keep track of the current negotiated values since we are sending the packets and know when the switchover takes place. * Needed to create ast_rtp_codecs_payload_code_tx() and make some callers use the new function because ast_rtp_codecs_payload_code() was used for mappings in both directions. * Needed to create ast_rtp_codecs_payloads_xover() for cases where we need to pass preferred codec mappings to the peer channel for early media bridging or when we need to prefer the offered mapping that RFC3264 says we SHOULD use. * ast_rtp_codecs_payloads_xover() and ast_rtp_codecs_payload_code_tx() are the only new public functions created. All the others were only used for the tx or rx mapping direction so the function doxygen now reflects which direction the function operates. * chan_mgcp.c: Removed call to ast_rtp_codecs_payloads_clear() as doing that makes no sense when processing an incoming SDP. We would be wiping out any mappings that we set for the possible outgoing SDP we sent earlier. ASTERISK-25166 Reported by: Kevin Harwell ASTERISK-17410 Reported by: Boris Fox Change-Id: Iaf6c227bca68cb7c414cf2fd4108a8ac98bd45ac
2015-07-24 00:24:04 +00:00
codec = ast_rtp_codecs_payload_code_tx(ast_rtp_instance_get_codecs(instance),
1, frame->subclass.format, 0);
media formats: re-architect handling of media for performance improvements In the old times media formats were represented using a bit field. This was fast but had a few limitations. 1. Asterisk was limited in how many formats it could handle. 2. Formats, being a bit field, could not include any attribute information. A format was strictly its type, e.g., "this is ulaw". This was changed in Asterisk 10 (see https://wiki.asterisk.org/wiki/display/AST/Media+Architecture+Proposal for notes on that work) which led to the creation of the ast_format structure. This structure allowed Asterisk to handle attributes and bundle information with a format. Additionally, ast_format_cap was created to act as a container for multiple formats that, together, formed the capability of some entity. Another mechanism was added to allow logic to be registered which performed format attribute negotiation. Everywhere throughout the codebase Asterisk was changed to use this strategy. Unfortunately, in software, there is no free lunch. These new capabilities came at a cost. Performance analysis and profiling showed that we spend an inordinate amount of time comparing, copying, and generally manipulating formats and their related structures. Basic prototyping has shown that a reasonably large performance improvement could be made in this area. This patch is the result of that project, which overhauled the media format architecture and its usage in Asterisk to improve performance. Generally, the new philosophy for handling formats is as follows: * The ast_format structure is reference counted. This removed a large amount of the memory allocations and copying that was done in prior versions. * In order to prevent race conditions while keeping things performant, the ast_format structure is immutable by convention and lock-free. Violate this tenet at your peril! * Because formats are reference counted, codecs are also reference counted. The Asterisk core generally provides built-in codecs and caches the ast_format structures created to represent them. Generally, to prevent inordinate amounts of module reference bumping, codecs and formats can be added at run-time but cannot be removed. * All compatibility with the bit field representation of codecs/formats has been moved to a compatibility API. The primary user of this representation is chan_iax2, which must continue to maintain its bit-field usage of formats for interoperability concerns. * When a format is negotiated with attributes, or when a format cannot be represented by one of the cached formats, a new format object is created or cloned from an existing format. That format may have the same codec underlying it, but is a different format than a version of the format with different attributes or without attributes. * While formats are reference counted objects, the reference count maintained on the format should be manipulated with care. Formats are generally cached and will persist for the lifetime of Asterisk and do not explicitly need to have their lifetime modified. An exception to this is when the user of a format does not know where the format came from *and* the user may outlive the provider of the format. This occurs, for example, when a format is read from a channel: the channel may have a format with attributes (hence, non-cached) and the user of the format may last longer than the channel (if the reference to the channel is released prior to the format's reference). For more information on this work, see the API design notes: https://wiki.asterisk.org/wiki/display/AST/Media+Format+Rewrite Finally, this work was the culmination of a large number of developer's efforts. Extra thanks goes to Corey Farrell, who took on a large amount of the work in the Asterisk core, chan_sip, and was an invaluable resource in peer reviews throughout this project. There were a substantial number of patches contributed during this work; the following issues/patch names simply reflect some of the work (and will cause the release scripts to give attribution to the individuals who work on them). Reviews: https://reviewboard.asterisk.org/r/3814 https://reviewboard.asterisk.org/r/3808 https://reviewboard.asterisk.org/r/3805 https://reviewboard.asterisk.org/r/3803 https://reviewboard.asterisk.org/r/3801 https://reviewboard.asterisk.org/r/3798 https://reviewboard.asterisk.org/r/3800 https://reviewboard.asterisk.org/r/3794 https://reviewboard.asterisk.org/r/3793 https://reviewboard.asterisk.org/r/3792 https://reviewboard.asterisk.org/r/3791 https://reviewboard.asterisk.org/r/3790 https://reviewboard.asterisk.org/r/3789 https://reviewboard.asterisk.org/r/3788 https://reviewboard.asterisk.org/r/3787 https://reviewboard.asterisk.org/r/3786 https://reviewboard.asterisk.org/r/3784 https://reviewboard.asterisk.org/r/3783 https://reviewboard.asterisk.org/r/3778 https://reviewboard.asterisk.org/r/3774 https://reviewboard.asterisk.org/r/3775 https://reviewboard.asterisk.org/r/3772 https://reviewboard.asterisk.org/r/3761 https://reviewboard.asterisk.org/r/3754 https://reviewboard.asterisk.org/r/3753 https://reviewboard.asterisk.org/r/3751 https://reviewboard.asterisk.org/r/3750 https://reviewboard.asterisk.org/r/3748 https://reviewboard.asterisk.org/r/3747 https://reviewboard.asterisk.org/r/3746 https://reviewboard.asterisk.org/r/3742 https://reviewboard.asterisk.org/r/3740 https://reviewboard.asterisk.org/r/3739 https://reviewboard.asterisk.org/r/3738 https://reviewboard.asterisk.org/r/3737 https://reviewboard.asterisk.org/r/3736 https://reviewboard.asterisk.org/r/3734 https://reviewboard.asterisk.org/r/3722 https://reviewboard.asterisk.org/r/3713 https://reviewboard.asterisk.org/r/3703 https://reviewboard.asterisk.org/r/3689 https://reviewboard.asterisk.org/r/3687 https://reviewboard.asterisk.org/r/3674 https://reviewboard.asterisk.org/r/3671 https://reviewboard.asterisk.org/r/3667 https://reviewboard.asterisk.org/r/3665 https://reviewboard.asterisk.org/r/3625 https://reviewboard.asterisk.org/r/3602 https://reviewboard.asterisk.org/r/3519 https://reviewboard.asterisk.org/r/3518 https://reviewboard.asterisk.org/r/3516 https://reviewboard.asterisk.org/r/3515 https://reviewboard.asterisk.org/r/3512 https://reviewboard.asterisk.org/r/3506 https://reviewboard.asterisk.org/r/3413 https://reviewboard.asterisk.org/r/3410 https://reviewboard.asterisk.org/r/3387 https://reviewboard.asterisk.org/r/3388 https://reviewboard.asterisk.org/r/3389 https://reviewboard.asterisk.org/r/3390 https://reviewboard.asterisk.org/r/3321 https://reviewboard.asterisk.org/r/3320 https://reviewboard.asterisk.org/r/3319 https://reviewboard.asterisk.org/r/3318 https://reviewboard.asterisk.org/r/3266 https://reviewboard.asterisk.org/r/3265 https://reviewboard.asterisk.org/r/3234 https://reviewboard.asterisk.org/r/3178 ASTERISK-23114 #close Reported by: mjordan media_formats_translation_core.diff uploaded by kharwell (License 6464) rb3506.diff uploaded by mjordan (License 6283) media_format_app_file.diff uploaded by kharwell (License 6464) misc-2.diff uploaded by file (License 5000) chan_mild-3.diff uploaded by file (License 5000) chan_obscure.diff uploaded by file (License 5000) jingle.diff uploaded by file (License 5000) funcs.diff uploaded by file (License 5000) formats.diff uploaded by file (License 5000) core.diff uploaded by file (License 5000) bridges.diff uploaded by file (License 5000) mf-codecs-2.diff uploaded by file (License 5000) mf-app_fax.diff uploaded by file (License 5000) mf-apps-3.diff uploaded by file (License 5000) media-formats-3.diff uploaded by file (License 5000) ASTERISK-23715 rb3713.patch uploaded by coreyfarrell (License 5909) rb3689.patch uploaded by mjordan (License 6283) ASTERISK-23957 rb3722.patch uploaded by mjordan (License 6283) mf-attributes-3.diff uploaded by file (License 5000) ASTERISK-23958 Tested by: jrose rb3822.patch uploaded by coreyfarrell (License 5909) rb3800.patch uploaded by jrose (License 6182) chan_sip.diff uploaded by mjordan (License 6283) rb3747.patch uploaded by jrose (License 6182) ASTERISK-23959 #close Tested by: sgriepentrog, mjordan, coreyfarrell sip_cleanup.diff uploaded by opticron (License 6273) chan_sip_caps.diff uploaded by mjordan (License 6283) rb3751.patch uploaded by coreyfarrell (License 5909) chan_sip-3.diff uploaded by file (License 5000) ASTERISK-23960 #close Tested by: opticron direct_media.diff uploaded by opticron (License 6273) pjsip-direct-media.diff uploaded by file (License 5000) format_cap_remove.diff uploaded by opticron (License 6273) media_format_fixes.diff uploaded by opticron (License 6273) chan_pjsip-2.diff uploaded by file (License 5000) ASTERISK-23966 #close Tested by: rmudgett rb3803.patch uploaded by rmudgetti (License 5621) chan_dahdi.diff uploaded by file (License 5000) ASTERISK-24064 #close Tested by: coreyfarrell, mjordan, opticron, file, rmudgett, sgriepentrog, jrose rb3814.patch uploaded by rmudgett (License 5621) moh_cleanup.diff uploaded by opticron (License 6273) bridge_leak.diff uploaded by opticron (License 6273) translate.diff uploaded by file (License 5000) rb3795.patch uploaded by rmudgett (License 5621) tls_fix.diff uploaded by mjordan (License 6283) fax-mf-fix-2.diff uploaded by file (License 5000) rtp_transfer_stuff uploaded by mjordan (License 6283) rb3787.patch uploaded by rmudgett (License 5621) media-formats-explicit-translate-format-3.diff uploaded by file (License 5000) format_cache_case_fix.diff uploaded by opticron (License 6273) rb3774.patch uploaded by rmudgett (License 5621) rb3775.patch uploaded by rmudgett (License 5621) rtp_engine_fix.diff uploaded by opticron (License 6273) rtp_crash_fix.diff uploaded by opticron (License 6273) rb3753.patch uploaded by mjordan (License 6283) rb3750.patch uploaded by mjordan (License 6283) rb3748.patch uploaded by rmudgett (License 5621) media_format_fixes.diff uploaded by opticron (License 6273) rb3740.patch uploaded by mjordan (License 6283) rb3739.patch uploaded by mjordan (License 6283) rb3734.patch uploaded by mjordan (License 6283) rb3689.patch uploaded by mjordan (License 6283) rb3674.patch uploaded by coreyfarrell (License 5909) rb3671.patch uploaded by coreyfarrell (License 5909) rb3667.patch uploaded by coreyfarrell (License 5909) rb3665.patch uploaded by mjordan (License 6283) rb3625.patch uploaded by coreyfarrell (License 5909) rb3602.patch uploaded by coreyfarrell (License 5909) format_compatibility-2.diff uploaded by file (License 5000) core.diff uploaded by file (License 5000) git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@419044 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-07-20 22:06:33 +00:00
if (codec < 0) {
ast_log(LOG_WARNING, "Don't know how to send format %s packets with RTP\n",
ast_format_get_name(frame->subclass.format));
return -1;
}
media formats: re-architect handling of media for performance improvements In the old times media formats were represented using a bit field. This was fast but had a few limitations. 1. Asterisk was limited in how many formats it could handle. 2. Formats, being a bit field, could not include any attribute information. A format was strictly its type, e.g., "this is ulaw". This was changed in Asterisk 10 (see https://wiki.asterisk.org/wiki/display/AST/Media+Architecture+Proposal for notes on that work) which led to the creation of the ast_format structure. This structure allowed Asterisk to handle attributes and bundle information with a format. Additionally, ast_format_cap was created to act as a container for multiple formats that, together, formed the capability of some entity. Another mechanism was added to allow logic to be registered which performed format attribute negotiation. Everywhere throughout the codebase Asterisk was changed to use this strategy. Unfortunately, in software, there is no free lunch. These new capabilities came at a cost. Performance analysis and profiling showed that we spend an inordinate amount of time comparing, copying, and generally manipulating formats and their related structures. Basic prototyping has shown that a reasonably large performance improvement could be made in this area. This patch is the result of that project, which overhauled the media format architecture and its usage in Asterisk to improve performance. Generally, the new philosophy for handling formats is as follows: * The ast_format structure is reference counted. This removed a large amount of the memory allocations and copying that was done in prior versions. * In order to prevent race conditions while keeping things performant, the ast_format structure is immutable by convention and lock-free. Violate this tenet at your peril! * Because formats are reference counted, codecs are also reference counted. The Asterisk core generally provides built-in codecs and caches the ast_format structures created to represent them. Generally, to prevent inordinate amounts of module reference bumping, codecs and formats can be added at run-time but cannot be removed. * All compatibility with the bit field representation of codecs/formats has been moved to a compatibility API. The primary user of this representation is chan_iax2, which must continue to maintain its bit-field usage of formats for interoperability concerns. * When a format is negotiated with attributes, or when a format cannot be represented by one of the cached formats, a new format object is created or cloned from an existing format. That format may have the same codec underlying it, but is a different format than a version of the format with different attributes or without attributes. * While formats are reference counted objects, the reference count maintained on the format should be manipulated with care. Formats are generally cached and will persist for the lifetime of Asterisk and do not explicitly need to have their lifetime modified. An exception to this is when the user of a format does not know where the format came from *and* the user may outlive the provider of the format. This occurs, for example, when a format is read from a channel: the channel may have a format with attributes (hence, non-cached) and the user of the format may last longer than the channel (if the reference to the channel is released prior to the format's reference). For more information on this work, see the API design notes: https://wiki.asterisk.org/wiki/display/AST/Media+Format+Rewrite Finally, this work was the culmination of a large number of developer's efforts. Extra thanks goes to Corey Farrell, who took on a large amount of the work in the Asterisk core, chan_sip, and was an invaluable resource in peer reviews throughout this project. There were a substantial number of patches contributed during this work; the following issues/patch names simply reflect some of the work (and will cause the release scripts to give attribution to the individuals who work on them). Reviews: https://reviewboard.asterisk.org/r/3814 https://reviewboard.asterisk.org/r/3808 https://reviewboard.asterisk.org/r/3805 https://reviewboard.asterisk.org/r/3803 https://reviewboard.asterisk.org/r/3801 https://reviewboard.asterisk.org/r/3798 https://reviewboard.asterisk.org/r/3800 https://reviewboard.asterisk.org/r/3794 https://reviewboard.asterisk.org/r/3793 https://reviewboard.asterisk.org/r/3792 https://reviewboard.asterisk.org/r/3791 https://reviewboard.asterisk.org/r/3790 https://reviewboard.asterisk.org/r/3789 https://reviewboard.asterisk.org/r/3788 https://reviewboard.asterisk.org/r/3787 https://reviewboard.asterisk.org/r/3786 https://reviewboard.asterisk.org/r/3784 https://reviewboard.asterisk.org/r/3783 https://reviewboard.asterisk.org/r/3778 https://reviewboard.asterisk.org/r/3774 https://reviewboard.asterisk.org/r/3775 https://reviewboard.asterisk.org/r/3772 https://reviewboard.asterisk.org/r/3761 https://reviewboard.asterisk.org/r/3754 https://reviewboard.asterisk.org/r/3753 https://reviewboard.asterisk.org/r/3751 https://reviewboard.asterisk.org/r/3750 https://reviewboard.asterisk.org/r/3748 https://reviewboard.asterisk.org/r/3747 https://reviewboard.asterisk.org/r/3746 https://reviewboard.asterisk.org/r/3742 https://reviewboard.asterisk.org/r/3740 https://reviewboard.asterisk.org/r/3739 https://reviewboard.asterisk.org/r/3738 https://reviewboard.asterisk.org/r/3737 https://reviewboard.asterisk.org/r/3736 https://reviewboard.asterisk.org/r/3734 https://reviewboard.asterisk.org/r/3722 https://reviewboard.asterisk.org/r/3713 https://reviewboard.asterisk.org/r/3703 https://reviewboard.asterisk.org/r/3689 https://reviewboard.asterisk.org/r/3687 https://reviewboard.asterisk.org/r/3674 https://reviewboard.asterisk.org/r/3671 https://reviewboard.asterisk.org/r/3667 https://reviewboard.asterisk.org/r/3665 https://reviewboard.asterisk.org/r/3625 https://reviewboard.asterisk.org/r/3602 https://reviewboard.asterisk.org/r/3519 https://reviewboard.asterisk.org/r/3518 https://reviewboard.asterisk.org/r/3516 https://reviewboard.asterisk.org/r/3515 https://reviewboard.asterisk.org/r/3512 https://reviewboard.asterisk.org/r/3506 https://reviewboard.asterisk.org/r/3413 https://reviewboard.asterisk.org/r/3410 https://reviewboard.asterisk.org/r/3387 https://reviewboard.asterisk.org/r/3388 https://reviewboard.asterisk.org/r/3389 https://reviewboard.asterisk.org/r/3390 https://reviewboard.asterisk.org/r/3321 https://reviewboard.asterisk.org/r/3320 https://reviewboard.asterisk.org/r/3319 https://reviewboard.asterisk.org/r/3318 https://reviewboard.asterisk.org/r/3266 https://reviewboard.asterisk.org/r/3265 https://reviewboard.asterisk.org/r/3234 https://reviewboard.asterisk.org/r/3178 ASTERISK-23114 #close Reported by: mjordan media_formats_translation_core.diff uploaded by kharwell (License 6464) rb3506.diff uploaded by mjordan (License 6283) media_format_app_file.diff uploaded by kharwell (License 6464) misc-2.diff uploaded by file (License 5000) chan_mild-3.diff uploaded by file (License 5000) chan_obscure.diff uploaded by file (License 5000) jingle.diff uploaded by file (License 5000) funcs.diff uploaded by file (License 5000) formats.diff uploaded by file (License 5000) core.diff uploaded by file (License 5000) bridges.diff uploaded by file (License 5000) mf-codecs-2.diff uploaded by file (License 5000) mf-app_fax.diff uploaded by file (License 5000) mf-apps-3.diff uploaded by file (License 5000) media-formats-3.diff uploaded by file (License 5000) ASTERISK-23715 rb3713.patch uploaded by coreyfarrell (License 5909) rb3689.patch uploaded by mjordan (License 6283) ASTERISK-23957 rb3722.patch uploaded by mjordan (License 6283) mf-attributes-3.diff uploaded by file (License 5000) ASTERISK-23958 Tested by: jrose rb3822.patch uploaded by coreyfarrell (License 5909) rb3800.patch uploaded by jrose (License 6182) chan_sip.diff uploaded by mjordan (License 6283) rb3747.patch uploaded by jrose (License 6182) ASTERISK-23959 #close Tested by: sgriepentrog, mjordan, coreyfarrell sip_cleanup.diff uploaded by opticron (License 6273) chan_sip_caps.diff uploaded by mjordan (License 6283) rb3751.patch uploaded by coreyfarrell (License 5909) chan_sip-3.diff uploaded by file (License 5000) ASTERISK-23960 #close Tested by: opticron direct_media.diff uploaded by opticron (License 6273) pjsip-direct-media.diff uploaded by file (License 5000) format_cap_remove.diff uploaded by opticron (License 6273) media_format_fixes.diff uploaded by opticron (License 6273) chan_pjsip-2.diff uploaded by file (License 5000) ASTERISK-23966 #close Tested by: rmudgett rb3803.patch uploaded by rmudgetti (License 5621) chan_dahdi.diff uploaded by file (License 5000) ASTERISK-24064 #close Tested by: coreyfarrell, mjordan, opticron, file, rmudgett, sgriepentrog, jrose rb3814.patch uploaded by rmudgett (License 5621) moh_cleanup.diff uploaded by opticron (License 6273) bridge_leak.diff uploaded by opticron (License 6273) translate.diff uploaded by file (License 5000) rb3795.patch uploaded by rmudgett (License 5621) tls_fix.diff uploaded by mjordan (License 6283) fax-mf-fix-2.diff uploaded by file (License 5000) rtp_transfer_stuff uploaded by mjordan (License 6283) rb3787.patch uploaded by rmudgett (License 5621) media-formats-explicit-translate-format-3.diff uploaded by file (License 5000) format_cache_case_fix.diff uploaded by opticron (License 6273) rb3774.patch uploaded by rmudgett (License 5621) rb3775.patch uploaded by rmudgett (License 5621) rtp_engine_fix.diff uploaded by opticron (License 6273) rtp_crash_fix.diff uploaded by opticron (License 6273) rb3753.patch uploaded by mjordan (License 6283) rb3750.patch uploaded by mjordan (License 6283) rb3748.patch uploaded by rmudgett (License 5621) media_format_fixes.diff uploaded by opticron (License 6273) rb3740.patch uploaded by mjordan (License 6283) rb3739.patch uploaded by mjordan (License 6283) rb3734.patch uploaded by mjordan (License 6283) rb3689.patch uploaded by mjordan (License 6283) rb3674.patch uploaded by coreyfarrell (License 5909) rb3671.patch uploaded by coreyfarrell (License 5909) rb3667.patch uploaded by coreyfarrell (License 5909) rb3665.patch uploaded by mjordan (License 6283) rb3625.patch uploaded by coreyfarrell (License 5909) rb3602.patch uploaded by coreyfarrell (License 5909) format_compatibility-2.diff uploaded by file (License 5000) core.diff uploaded by file (License 5000) git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@419044 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-07-20 22:06:33 +00:00
/* Note that we do not increase the ref count here as this pointer
* will not be held by any thing explicitly. The format variable is
* merely a convenience reference to frame->subclass.format */
format = frame->subclass.format;
if (ast_format_cmp(rtp->lasttxformat, format) == AST_FORMAT_CMP_NOT_EQUAL) {
/* Oh dear, if the format changed we will have to set up a new smoother */
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
ast_debug_rtp(1, "(%s) RTP ooh, format changed from %s to %s\n",
ast_rtp_instance_get_channel_id(instance),
ast_format_get_name(rtp->lasttxformat),
ast_format_get_name(frame->subclass.format));
media formats: re-architect handling of media for performance improvements In the old times media formats were represented using a bit field. This was fast but had a few limitations. 1. Asterisk was limited in how many formats it could handle. 2. Formats, being a bit field, could not include any attribute information. A format was strictly its type, e.g., "this is ulaw". This was changed in Asterisk 10 (see https://wiki.asterisk.org/wiki/display/AST/Media+Architecture+Proposal for notes on that work) which led to the creation of the ast_format structure. This structure allowed Asterisk to handle attributes and bundle information with a format. Additionally, ast_format_cap was created to act as a container for multiple formats that, together, formed the capability of some entity. Another mechanism was added to allow logic to be registered which performed format attribute negotiation. Everywhere throughout the codebase Asterisk was changed to use this strategy. Unfortunately, in software, there is no free lunch. These new capabilities came at a cost. Performance analysis and profiling showed that we spend an inordinate amount of time comparing, copying, and generally manipulating formats and their related structures. Basic prototyping has shown that a reasonably large performance improvement could be made in this area. This patch is the result of that project, which overhauled the media format architecture and its usage in Asterisk to improve performance. Generally, the new philosophy for handling formats is as follows: * The ast_format structure is reference counted. This removed a large amount of the memory allocations and copying that was done in prior versions. * In order to prevent race conditions while keeping things performant, the ast_format structure is immutable by convention and lock-free. Violate this tenet at your peril! * Because formats are reference counted, codecs are also reference counted. The Asterisk core generally provides built-in codecs and caches the ast_format structures created to represent them. Generally, to prevent inordinate amounts of module reference bumping, codecs and formats can be added at run-time but cannot be removed. * All compatibility with the bit field representation of codecs/formats has been moved to a compatibility API. The primary user of this representation is chan_iax2, which must continue to maintain its bit-field usage of formats for interoperability concerns. * When a format is negotiated with attributes, or when a format cannot be represented by one of the cached formats, a new format object is created or cloned from an existing format. That format may have the same codec underlying it, but is a different format than a version of the format with different attributes or without attributes. * While formats are reference counted objects, the reference count maintained on the format should be manipulated with care. Formats are generally cached and will persist for the lifetime of Asterisk and do not explicitly need to have their lifetime modified. An exception to this is when the user of a format does not know where the format came from *and* the user may outlive the provider of the format. This occurs, for example, when a format is read from a channel: the channel may have a format with attributes (hence, non-cached) and the user of the format may last longer than the channel (if the reference to the channel is released prior to the format's reference). For more information on this work, see the API design notes: https://wiki.asterisk.org/wiki/display/AST/Media+Format+Rewrite Finally, this work was the culmination of a large number of developer's efforts. Extra thanks goes to Corey Farrell, who took on a large amount of the work in the Asterisk core, chan_sip, and was an invaluable resource in peer reviews throughout this project. There were a substantial number of patches contributed during this work; the following issues/patch names simply reflect some of the work (and will cause the release scripts to give attribution to the individuals who work on them). Reviews: https://reviewboard.asterisk.org/r/3814 https://reviewboard.asterisk.org/r/3808 https://reviewboard.asterisk.org/r/3805 https://reviewboard.asterisk.org/r/3803 https://reviewboard.asterisk.org/r/3801 https://reviewboard.asterisk.org/r/3798 https://reviewboard.asterisk.org/r/3800 https://reviewboard.asterisk.org/r/3794 https://reviewboard.asterisk.org/r/3793 https://reviewboard.asterisk.org/r/3792 https://reviewboard.asterisk.org/r/3791 https://reviewboard.asterisk.org/r/3790 https://reviewboard.asterisk.org/r/3789 https://reviewboard.asterisk.org/r/3788 https://reviewboard.asterisk.org/r/3787 https://reviewboard.asterisk.org/r/3786 https://reviewboard.asterisk.org/r/3784 https://reviewboard.asterisk.org/r/3783 https://reviewboard.asterisk.org/r/3778 https://reviewboard.asterisk.org/r/3774 https://reviewboard.asterisk.org/r/3775 https://reviewboard.asterisk.org/r/3772 https://reviewboard.asterisk.org/r/3761 https://reviewboard.asterisk.org/r/3754 https://reviewboard.asterisk.org/r/3753 https://reviewboard.asterisk.org/r/3751 https://reviewboard.asterisk.org/r/3750 https://reviewboard.asterisk.org/r/3748 https://reviewboard.asterisk.org/r/3747 https://reviewboard.asterisk.org/r/3746 https://reviewboard.asterisk.org/r/3742 https://reviewboard.asterisk.org/r/3740 https://reviewboard.asterisk.org/r/3739 https://reviewboard.asterisk.org/r/3738 https://reviewboard.asterisk.org/r/3737 https://reviewboard.asterisk.org/r/3736 https://reviewboard.asterisk.org/r/3734 https://reviewboard.asterisk.org/r/3722 https://reviewboard.asterisk.org/r/3713 https://reviewboard.asterisk.org/r/3703 https://reviewboard.asterisk.org/r/3689 https://reviewboard.asterisk.org/r/3687 https://reviewboard.asterisk.org/r/3674 https://reviewboard.asterisk.org/r/3671 https://reviewboard.asterisk.org/r/3667 https://reviewboard.asterisk.org/r/3665 https://reviewboard.asterisk.org/r/3625 https://reviewboard.asterisk.org/r/3602 https://reviewboard.asterisk.org/r/3519 https://reviewboard.asterisk.org/r/3518 https://reviewboard.asterisk.org/r/3516 https://reviewboard.asterisk.org/r/3515 https://reviewboard.asterisk.org/r/3512 https://reviewboard.asterisk.org/r/3506 https://reviewboard.asterisk.org/r/3413 https://reviewboard.asterisk.org/r/3410 https://reviewboard.asterisk.org/r/3387 https://reviewboard.asterisk.org/r/3388 https://reviewboard.asterisk.org/r/3389 https://reviewboard.asterisk.org/r/3390 https://reviewboard.asterisk.org/r/3321 https://reviewboard.asterisk.org/r/3320 https://reviewboard.asterisk.org/r/3319 https://reviewboard.asterisk.org/r/3318 https://reviewboard.asterisk.org/r/3266 https://reviewboard.asterisk.org/r/3265 https://reviewboard.asterisk.org/r/3234 https://reviewboard.asterisk.org/r/3178 ASTERISK-23114 #close Reported by: mjordan media_formats_translation_core.diff uploaded by kharwell (License 6464) rb3506.diff uploaded by mjordan (License 6283) media_format_app_file.diff uploaded by kharwell (License 6464) misc-2.diff uploaded by file (License 5000) chan_mild-3.diff uploaded by file (License 5000) chan_obscure.diff uploaded by file (License 5000) jingle.diff uploaded by file (License 5000) funcs.diff uploaded by file (License 5000) formats.diff uploaded by file (License 5000) core.diff uploaded by file (License 5000) bridges.diff uploaded by file (License 5000) mf-codecs-2.diff uploaded by file (License 5000) mf-app_fax.diff uploaded by file (License 5000) mf-apps-3.diff uploaded by file (License 5000) media-formats-3.diff uploaded by file (License 5000) ASTERISK-23715 rb3713.patch uploaded by coreyfarrell (License 5909) rb3689.patch uploaded by mjordan (License 6283) ASTERISK-23957 rb3722.patch uploaded by mjordan (License 6283) mf-attributes-3.diff uploaded by file (License 5000) ASTERISK-23958 Tested by: jrose rb3822.patch uploaded by coreyfarrell (License 5909) rb3800.patch uploaded by jrose (License 6182) chan_sip.diff uploaded by mjordan (License 6283) rb3747.patch uploaded by jrose (License 6182) ASTERISK-23959 #close Tested by: sgriepentrog, mjordan, coreyfarrell sip_cleanup.diff uploaded by opticron (License 6273) chan_sip_caps.diff uploaded by mjordan (License 6283) rb3751.patch uploaded by coreyfarrell (License 5909) chan_sip-3.diff uploaded by file (License 5000) ASTERISK-23960 #close Tested by: opticron direct_media.diff uploaded by opticron (License 6273) pjsip-direct-media.diff uploaded by file (License 5000) format_cap_remove.diff uploaded by opticron (License 6273) media_format_fixes.diff uploaded by opticron (License 6273) chan_pjsip-2.diff uploaded by file (License 5000) ASTERISK-23966 #close Tested by: rmudgett rb3803.patch uploaded by rmudgetti (License 5621) chan_dahdi.diff uploaded by file (License 5000) ASTERISK-24064 #close Tested by: coreyfarrell, mjordan, opticron, file, rmudgett, sgriepentrog, jrose rb3814.patch uploaded by rmudgett (License 5621) moh_cleanup.diff uploaded by opticron (License 6273) bridge_leak.diff uploaded by opticron (License 6273) translate.diff uploaded by file (License 5000) rb3795.patch uploaded by rmudgett (License 5621) tls_fix.diff uploaded by mjordan (License 6283) fax-mf-fix-2.diff uploaded by file (License 5000) rtp_transfer_stuff uploaded by mjordan (License 6283) rb3787.patch uploaded by rmudgett (License 5621) media-formats-explicit-translate-format-3.diff uploaded by file (License 5000) format_cache_case_fix.diff uploaded by opticron (License 6273) rb3774.patch uploaded by rmudgett (License 5621) rb3775.patch uploaded by rmudgett (License 5621) rtp_engine_fix.diff uploaded by opticron (License 6273) rtp_crash_fix.diff uploaded by opticron (License 6273) rb3753.patch uploaded by mjordan (License 6283) rb3750.patch uploaded by mjordan (License 6283) rb3748.patch uploaded by rmudgett (License 5621) media_format_fixes.diff uploaded by opticron (License 6273) rb3740.patch uploaded by mjordan (License 6283) rb3739.patch uploaded by mjordan (License 6283) rb3734.patch uploaded by mjordan (License 6283) rb3689.patch uploaded by mjordan (License 6283) rb3674.patch uploaded by coreyfarrell (License 5909) rb3671.patch uploaded by coreyfarrell (License 5909) rb3667.patch uploaded by coreyfarrell (License 5909) rb3665.patch uploaded by mjordan (License 6283) rb3625.patch uploaded by coreyfarrell (License 5909) rb3602.patch uploaded by coreyfarrell (License 5909) format_compatibility-2.diff uploaded by file (License 5000) core.diff uploaded by file (License 5000) git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@419044 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-07-20 22:06:33 +00:00
ao2_replace(rtp->lasttxformat, format);
if (rtp->smoother) {
ast_smoother_free(rtp->smoother);
rtp->smoother = NULL;
}
}
/* If no smoother is present see if we have to set one up */
media formats: re-architect handling of media for performance improvements In the old times media formats were represented using a bit field. This was fast but had a few limitations. 1. Asterisk was limited in how many formats it could handle. 2. Formats, being a bit field, could not include any attribute information. A format was strictly its type, e.g., "this is ulaw". This was changed in Asterisk 10 (see https://wiki.asterisk.org/wiki/display/AST/Media+Architecture+Proposal for notes on that work) which led to the creation of the ast_format structure. This structure allowed Asterisk to handle attributes and bundle information with a format. Additionally, ast_format_cap was created to act as a container for multiple formats that, together, formed the capability of some entity. Another mechanism was added to allow logic to be registered which performed format attribute negotiation. Everywhere throughout the codebase Asterisk was changed to use this strategy. Unfortunately, in software, there is no free lunch. These new capabilities came at a cost. Performance analysis and profiling showed that we spend an inordinate amount of time comparing, copying, and generally manipulating formats and their related structures. Basic prototyping has shown that a reasonably large performance improvement could be made in this area. This patch is the result of that project, which overhauled the media format architecture and its usage in Asterisk to improve performance. Generally, the new philosophy for handling formats is as follows: * The ast_format structure is reference counted. This removed a large amount of the memory allocations and copying that was done in prior versions. * In order to prevent race conditions while keeping things performant, the ast_format structure is immutable by convention and lock-free. Violate this tenet at your peril! * Because formats are reference counted, codecs are also reference counted. The Asterisk core generally provides built-in codecs and caches the ast_format structures created to represent them. Generally, to prevent inordinate amounts of module reference bumping, codecs and formats can be added at run-time but cannot be removed. * All compatibility with the bit field representation of codecs/formats has been moved to a compatibility API. The primary user of this representation is chan_iax2, which must continue to maintain its bit-field usage of formats for interoperability concerns. * When a format is negotiated with attributes, or when a format cannot be represented by one of the cached formats, a new format object is created or cloned from an existing format. That format may have the same codec underlying it, but is a different format than a version of the format with different attributes or without attributes. * While formats are reference counted objects, the reference count maintained on the format should be manipulated with care. Formats are generally cached and will persist for the lifetime of Asterisk and do not explicitly need to have their lifetime modified. An exception to this is when the user of a format does not know where the format came from *and* the user may outlive the provider of the format. This occurs, for example, when a format is read from a channel: the channel may have a format with attributes (hence, non-cached) and the user of the format may last longer than the channel (if the reference to the channel is released prior to the format's reference). For more information on this work, see the API design notes: https://wiki.asterisk.org/wiki/display/AST/Media+Format+Rewrite Finally, this work was the culmination of a large number of developer's efforts. Extra thanks goes to Corey Farrell, who took on a large amount of the work in the Asterisk core, chan_sip, and was an invaluable resource in peer reviews throughout this project. There were a substantial number of patches contributed during this work; the following issues/patch names simply reflect some of the work (and will cause the release scripts to give attribution to the individuals who work on them). Reviews: https://reviewboard.asterisk.org/r/3814 https://reviewboard.asterisk.org/r/3808 https://reviewboard.asterisk.org/r/3805 https://reviewboard.asterisk.org/r/3803 https://reviewboard.asterisk.org/r/3801 https://reviewboard.asterisk.org/r/3798 https://reviewboard.asterisk.org/r/3800 https://reviewboard.asterisk.org/r/3794 https://reviewboard.asterisk.org/r/3793 https://reviewboard.asterisk.org/r/3792 https://reviewboard.asterisk.org/r/3791 https://reviewboard.asterisk.org/r/3790 https://reviewboard.asterisk.org/r/3789 https://reviewboard.asterisk.org/r/3788 https://reviewboard.asterisk.org/r/3787 https://reviewboard.asterisk.org/r/3786 https://reviewboard.asterisk.org/r/3784 https://reviewboard.asterisk.org/r/3783 https://reviewboard.asterisk.org/r/3778 https://reviewboard.asterisk.org/r/3774 https://reviewboard.asterisk.org/r/3775 https://reviewboard.asterisk.org/r/3772 https://reviewboard.asterisk.org/r/3761 https://reviewboard.asterisk.org/r/3754 https://reviewboard.asterisk.org/r/3753 https://reviewboard.asterisk.org/r/3751 https://reviewboard.asterisk.org/r/3750 https://reviewboard.asterisk.org/r/3748 https://reviewboard.asterisk.org/r/3747 https://reviewboard.asterisk.org/r/3746 https://reviewboard.asterisk.org/r/3742 https://reviewboard.asterisk.org/r/3740 https://reviewboard.asterisk.org/r/3739 https://reviewboard.asterisk.org/r/3738 https://reviewboard.asterisk.org/r/3737 https://reviewboard.asterisk.org/r/3736 https://reviewboard.asterisk.org/r/3734 https://reviewboard.asterisk.org/r/3722 https://reviewboard.asterisk.org/r/3713 https://reviewboard.asterisk.org/r/3703 https://reviewboard.asterisk.org/r/3689 https://reviewboard.asterisk.org/r/3687 https://reviewboard.asterisk.org/r/3674 https://reviewboard.asterisk.org/r/3671 https://reviewboard.asterisk.org/r/3667 https://reviewboard.asterisk.org/r/3665 https://reviewboard.asterisk.org/r/3625 https://reviewboard.asterisk.org/r/3602 https://reviewboard.asterisk.org/r/3519 https://reviewboard.asterisk.org/r/3518 https://reviewboard.asterisk.org/r/3516 https://reviewboard.asterisk.org/r/3515 https://reviewboard.asterisk.org/r/3512 https://reviewboard.asterisk.org/r/3506 https://reviewboard.asterisk.org/r/3413 https://reviewboard.asterisk.org/r/3410 https://reviewboard.asterisk.org/r/3387 https://reviewboard.asterisk.org/r/3388 https://reviewboard.asterisk.org/r/3389 https://reviewboard.asterisk.org/r/3390 https://reviewboard.asterisk.org/r/3321 https://reviewboard.asterisk.org/r/3320 https://reviewboard.asterisk.org/r/3319 https://reviewboard.asterisk.org/r/3318 https://reviewboard.asterisk.org/r/3266 https://reviewboard.asterisk.org/r/3265 https://reviewboard.asterisk.org/r/3234 https://reviewboard.asterisk.org/r/3178 ASTERISK-23114 #close Reported by: mjordan media_formats_translation_core.diff uploaded by kharwell (License 6464) rb3506.diff uploaded by mjordan (License 6283) media_format_app_file.diff uploaded by kharwell (License 6464) misc-2.diff uploaded by file (License 5000) chan_mild-3.diff uploaded by file (License 5000) chan_obscure.diff uploaded by file (License 5000) jingle.diff uploaded by file (License 5000) funcs.diff uploaded by file (License 5000) formats.diff uploaded by file (License 5000) core.diff uploaded by file (License 5000) bridges.diff uploaded by file (License 5000) mf-codecs-2.diff uploaded by file (License 5000) mf-app_fax.diff uploaded by file (License 5000) mf-apps-3.diff uploaded by file (License 5000) media-formats-3.diff uploaded by file (License 5000) ASTERISK-23715 rb3713.patch uploaded by coreyfarrell (License 5909) rb3689.patch uploaded by mjordan (License 6283) ASTERISK-23957 rb3722.patch uploaded by mjordan (License 6283) mf-attributes-3.diff uploaded by file (License 5000) ASTERISK-23958 Tested by: jrose rb3822.patch uploaded by coreyfarrell (License 5909) rb3800.patch uploaded by jrose (License 6182) chan_sip.diff uploaded by mjordan (License 6283) rb3747.patch uploaded by jrose (License 6182) ASTERISK-23959 #close Tested by: sgriepentrog, mjordan, coreyfarrell sip_cleanup.diff uploaded by opticron (License 6273) chan_sip_caps.diff uploaded by mjordan (License 6283) rb3751.patch uploaded by coreyfarrell (License 5909) chan_sip-3.diff uploaded by file (License 5000) ASTERISK-23960 #close Tested by: opticron direct_media.diff uploaded by opticron (License 6273) pjsip-direct-media.diff uploaded by file (License 5000) format_cap_remove.diff uploaded by opticron (License 6273) media_format_fixes.diff uploaded by opticron (License 6273) chan_pjsip-2.diff uploaded by file (License 5000) ASTERISK-23966 #close Tested by: rmudgett rb3803.patch uploaded by rmudgetti (License 5621) chan_dahdi.diff uploaded by file (License 5000) ASTERISK-24064 #close Tested by: coreyfarrell, mjordan, opticron, file, rmudgett, sgriepentrog, jrose rb3814.patch uploaded by rmudgett (License 5621) moh_cleanup.diff uploaded by opticron (License 6273) bridge_leak.diff uploaded by opticron (License 6273) translate.diff uploaded by file (License 5000) rb3795.patch uploaded by rmudgett (License 5621) tls_fix.diff uploaded by mjordan (License 6283) fax-mf-fix-2.diff uploaded by file (License 5000) rtp_transfer_stuff uploaded by mjordan (License 6283) rb3787.patch uploaded by rmudgett (License 5621) media-formats-explicit-translate-format-3.diff uploaded by file (License 5000) format_cache_case_fix.diff uploaded by opticron (License 6273) rb3774.patch uploaded by rmudgett (License 5621) rb3775.patch uploaded by rmudgett (License 5621) rtp_engine_fix.diff uploaded by opticron (License 6273) rtp_crash_fix.diff uploaded by opticron (License 6273) rb3753.patch uploaded by mjordan (License 6283) rb3750.patch uploaded by mjordan (License 6283) rb3748.patch uploaded by rmudgett (License 5621) media_format_fixes.diff uploaded by opticron (License 6273) rb3740.patch uploaded by mjordan (License 6283) rb3739.patch uploaded by mjordan (License 6283) rb3734.patch uploaded by mjordan (License 6283) rb3689.patch uploaded by mjordan (License 6283) rb3674.patch uploaded by coreyfarrell (License 5909) rb3671.patch uploaded by coreyfarrell (License 5909) rb3667.patch uploaded by coreyfarrell (License 5909) rb3665.patch uploaded by mjordan (License 6283) rb3625.patch uploaded by coreyfarrell (License 5909) rb3602.patch uploaded by coreyfarrell (License 5909) format_compatibility-2.diff uploaded by file (License 5000) core.diff uploaded by file (License 5000) git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@419044 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-07-20 22:06:33 +00:00
if (!rtp->smoother && ast_format_can_be_smoothed(format)) {
unsigned int smoother_flags = ast_format_get_smoother_flags(format);
media formats: re-architect handling of media for performance improvements In the old times media formats were represented using a bit field. This was fast but had a few limitations. 1. Asterisk was limited in how many formats it could handle. 2. Formats, being a bit field, could not include any attribute information. A format was strictly its type, e.g., "this is ulaw". This was changed in Asterisk 10 (see https://wiki.asterisk.org/wiki/display/AST/Media+Architecture+Proposal for notes on that work) which led to the creation of the ast_format structure. This structure allowed Asterisk to handle attributes and bundle information with a format. Additionally, ast_format_cap was created to act as a container for multiple formats that, together, formed the capability of some entity. Another mechanism was added to allow logic to be registered which performed format attribute negotiation. Everywhere throughout the codebase Asterisk was changed to use this strategy. Unfortunately, in software, there is no free lunch. These new capabilities came at a cost. Performance analysis and profiling showed that we spend an inordinate amount of time comparing, copying, and generally manipulating formats and their related structures. Basic prototyping has shown that a reasonably large performance improvement could be made in this area. This patch is the result of that project, which overhauled the media format architecture and its usage in Asterisk to improve performance. Generally, the new philosophy for handling formats is as follows: * The ast_format structure is reference counted. This removed a large amount of the memory allocations and copying that was done in prior versions. * In order to prevent race conditions while keeping things performant, the ast_format structure is immutable by convention and lock-free. Violate this tenet at your peril! * Because formats are reference counted, codecs are also reference counted. The Asterisk core generally provides built-in codecs and caches the ast_format structures created to represent them. Generally, to prevent inordinate amounts of module reference bumping, codecs and formats can be added at run-time but cannot be removed. * All compatibility with the bit field representation of codecs/formats has been moved to a compatibility API. The primary user of this representation is chan_iax2, which must continue to maintain its bit-field usage of formats for interoperability concerns. * When a format is negotiated with attributes, or when a format cannot be represented by one of the cached formats, a new format object is created or cloned from an existing format. That format may have the same codec underlying it, but is a different format than a version of the format with different attributes or without attributes. * While formats are reference counted objects, the reference count maintained on the format should be manipulated with care. Formats are generally cached and will persist for the lifetime of Asterisk and do not explicitly need to have their lifetime modified. An exception to this is when the user of a format does not know where the format came from *and* the user may outlive the provider of the format. This occurs, for example, when a format is read from a channel: the channel may have a format with attributes (hence, non-cached) and the user of the format may last longer than the channel (if the reference to the channel is released prior to the format's reference). For more information on this work, see the API design notes: https://wiki.asterisk.org/wiki/display/AST/Media+Format+Rewrite Finally, this work was the culmination of a large number of developer's efforts. Extra thanks goes to Corey Farrell, who took on a large amount of the work in the Asterisk core, chan_sip, and was an invaluable resource in peer reviews throughout this project. There were a substantial number of patches contributed during this work; the following issues/patch names simply reflect some of the work (and will cause the release scripts to give attribution to the individuals who work on them). Reviews: https://reviewboard.asterisk.org/r/3814 https://reviewboard.asterisk.org/r/3808 https://reviewboard.asterisk.org/r/3805 https://reviewboard.asterisk.org/r/3803 https://reviewboard.asterisk.org/r/3801 https://reviewboard.asterisk.org/r/3798 https://reviewboard.asterisk.org/r/3800 https://reviewboard.asterisk.org/r/3794 https://reviewboard.asterisk.org/r/3793 https://reviewboard.asterisk.org/r/3792 https://reviewboard.asterisk.org/r/3791 https://reviewboard.asterisk.org/r/3790 https://reviewboard.asterisk.org/r/3789 https://reviewboard.asterisk.org/r/3788 https://reviewboard.asterisk.org/r/3787 https://reviewboard.asterisk.org/r/3786 https://reviewboard.asterisk.org/r/3784 https://reviewboard.asterisk.org/r/3783 https://reviewboard.asterisk.org/r/3778 https://reviewboard.asterisk.org/r/3774 https://reviewboard.asterisk.org/r/3775 https://reviewboard.asterisk.org/r/3772 https://reviewboard.asterisk.org/r/3761 https://reviewboard.asterisk.org/r/3754 https://reviewboard.asterisk.org/r/3753 https://reviewboard.asterisk.org/r/3751 https://reviewboard.asterisk.org/r/3750 https://reviewboard.asterisk.org/r/3748 https://reviewboard.asterisk.org/r/3747 https://reviewboard.asterisk.org/r/3746 https://reviewboard.asterisk.org/r/3742 https://reviewboard.asterisk.org/r/3740 https://reviewboard.asterisk.org/r/3739 https://reviewboard.asterisk.org/r/3738 https://reviewboard.asterisk.org/r/3737 https://reviewboard.asterisk.org/r/3736 https://reviewboard.asterisk.org/r/3734 https://reviewboard.asterisk.org/r/3722 https://reviewboard.asterisk.org/r/3713 https://reviewboard.asterisk.org/r/3703 https://reviewboard.asterisk.org/r/3689 https://reviewboard.asterisk.org/r/3687 https://reviewboard.asterisk.org/r/3674 https://reviewboard.asterisk.org/r/3671 https://reviewboard.asterisk.org/r/3667 https://reviewboard.asterisk.org/r/3665 https://reviewboard.asterisk.org/r/3625 https://reviewboard.asterisk.org/r/3602 https://reviewboard.asterisk.org/r/3519 https://reviewboard.asterisk.org/r/3518 https://reviewboard.asterisk.org/r/3516 https://reviewboard.asterisk.org/r/3515 https://reviewboard.asterisk.org/r/3512 https://reviewboard.asterisk.org/r/3506 https://reviewboard.asterisk.org/r/3413 https://reviewboard.asterisk.org/r/3410 https://reviewboard.asterisk.org/r/3387 https://reviewboard.asterisk.org/r/3388 https://reviewboard.asterisk.org/r/3389 https://reviewboard.asterisk.org/r/3390 https://reviewboard.asterisk.org/r/3321 https://reviewboard.asterisk.org/r/3320 https://reviewboard.asterisk.org/r/3319 https://reviewboard.asterisk.org/r/3318 https://reviewboard.asterisk.org/r/3266 https://reviewboard.asterisk.org/r/3265 https://reviewboard.asterisk.org/r/3234 https://reviewboard.asterisk.org/r/3178 ASTERISK-23114 #close Reported by: mjordan media_formats_translation_core.diff uploaded by kharwell (License 6464) rb3506.diff uploaded by mjordan (License 6283) media_format_app_file.diff uploaded by kharwell (License 6464) misc-2.diff uploaded by file (License 5000) chan_mild-3.diff uploaded by file (License 5000) chan_obscure.diff uploaded by file (License 5000) jingle.diff uploaded by file (License 5000) funcs.diff uploaded by file (License 5000) formats.diff uploaded by file (License 5000) core.diff uploaded by file (License 5000) bridges.diff uploaded by file (License 5000) mf-codecs-2.diff uploaded by file (License 5000) mf-app_fax.diff uploaded by file (License 5000) mf-apps-3.diff uploaded by file (License 5000) media-formats-3.diff uploaded by file (License 5000) ASTERISK-23715 rb3713.patch uploaded by coreyfarrell (License 5909) rb3689.patch uploaded by mjordan (License 6283) ASTERISK-23957 rb3722.patch uploaded by mjordan (License 6283) mf-attributes-3.diff uploaded by file (License 5000) ASTERISK-23958 Tested by: jrose rb3822.patch uploaded by coreyfarrell (License 5909) rb3800.patch uploaded by jrose (License 6182) chan_sip.diff uploaded by mjordan (License 6283) rb3747.patch uploaded by jrose (License 6182) ASTERISK-23959 #close Tested by: sgriepentrog, mjordan, coreyfarrell sip_cleanup.diff uploaded by opticron (License 6273) chan_sip_caps.diff uploaded by mjordan (License 6283) rb3751.patch uploaded by coreyfarrell (License 5909) chan_sip-3.diff uploaded by file (License 5000) ASTERISK-23960 #close Tested by: opticron direct_media.diff uploaded by opticron (License 6273) pjsip-direct-media.diff uploaded by file (License 5000) format_cap_remove.diff uploaded by opticron (License 6273) media_format_fixes.diff uploaded by opticron (License 6273) chan_pjsip-2.diff uploaded by file (License 5000) ASTERISK-23966 #close Tested by: rmudgett rb3803.patch uploaded by rmudgetti (License 5621) chan_dahdi.diff uploaded by file (License 5000) ASTERISK-24064 #close Tested by: coreyfarrell, mjordan, opticron, file, rmudgett, sgriepentrog, jrose rb3814.patch uploaded by rmudgett (License 5621) moh_cleanup.diff uploaded by opticron (License 6273) bridge_leak.diff uploaded by opticron (License 6273) translate.diff uploaded by file (License 5000) rb3795.patch uploaded by rmudgett (License 5621) tls_fix.diff uploaded by mjordan (License 6283) fax-mf-fix-2.diff uploaded by file (License 5000) rtp_transfer_stuff uploaded by mjordan (License 6283) rb3787.patch uploaded by rmudgett (License 5621) media-formats-explicit-translate-format-3.diff uploaded by file (License 5000) format_cache_case_fix.diff uploaded by opticron (License 6273) rb3774.patch uploaded by rmudgett (License 5621) rb3775.patch uploaded by rmudgett (License 5621) rtp_engine_fix.diff uploaded by opticron (License 6273) rtp_crash_fix.diff uploaded by opticron (License 6273) rb3753.patch uploaded by mjordan (License 6283) rb3750.patch uploaded by mjordan (License 6283) rb3748.patch uploaded by rmudgett (License 5621) media_format_fixes.diff uploaded by opticron (License 6273) rb3740.patch uploaded by mjordan (License 6283) rb3739.patch uploaded by mjordan (License 6283) rb3734.patch uploaded by mjordan (License 6283) rb3689.patch uploaded by mjordan (License 6283) rb3674.patch uploaded by coreyfarrell (License 5909) rb3671.patch uploaded by coreyfarrell (License 5909) rb3667.patch uploaded by coreyfarrell (License 5909) rb3665.patch uploaded by mjordan (License 6283) rb3625.patch uploaded by coreyfarrell (License 5909) rb3602.patch uploaded by coreyfarrell (License 5909) format_compatibility-2.diff uploaded by file (License 5000) core.diff uploaded by file (License 5000) git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@419044 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-07-20 22:06:33 +00:00
unsigned int framing_ms = ast_rtp_codecs_get_framing(ast_rtp_instance_get_codecs(instance));
if (!framing_ms && (smoother_flags & AST_SMOOTHER_FLAG_FORCED)) {
framing_ms = ast_format_get_default_ms(format);
}
media formats: re-architect handling of media for performance improvements In the old times media formats were represented using a bit field. This was fast but had a few limitations. 1. Asterisk was limited in how many formats it could handle. 2. Formats, being a bit field, could not include any attribute information. A format was strictly its type, e.g., "this is ulaw". This was changed in Asterisk 10 (see https://wiki.asterisk.org/wiki/display/AST/Media+Architecture+Proposal for notes on that work) which led to the creation of the ast_format structure. This structure allowed Asterisk to handle attributes and bundle information with a format. Additionally, ast_format_cap was created to act as a container for multiple formats that, together, formed the capability of some entity. Another mechanism was added to allow logic to be registered which performed format attribute negotiation. Everywhere throughout the codebase Asterisk was changed to use this strategy. Unfortunately, in software, there is no free lunch. These new capabilities came at a cost. Performance analysis and profiling showed that we spend an inordinate amount of time comparing, copying, and generally manipulating formats and their related structures. Basic prototyping has shown that a reasonably large performance improvement could be made in this area. This patch is the result of that project, which overhauled the media format architecture and its usage in Asterisk to improve performance. Generally, the new philosophy for handling formats is as follows: * The ast_format structure is reference counted. This removed a large amount of the memory allocations and copying that was done in prior versions. * In order to prevent race conditions while keeping things performant, the ast_format structure is immutable by convention and lock-free. Violate this tenet at your peril! * Because formats are reference counted, codecs are also reference counted. The Asterisk core generally provides built-in codecs and caches the ast_format structures created to represent them. Generally, to prevent inordinate amounts of module reference bumping, codecs and formats can be added at run-time but cannot be removed. * All compatibility with the bit field representation of codecs/formats has been moved to a compatibility API. The primary user of this representation is chan_iax2, which must continue to maintain its bit-field usage of formats for interoperability concerns. * When a format is negotiated with attributes, or when a format cannot be represented by one of the cached formats, a new format object is created or cloned from an existing format. That format may have the same codec underlying it, but is a different format than a version of the format with different attributes or without attributes. * While formats are reference counted objects, the reference count maintained on the format should be manipulated with care. Formats are generally cached and will persist for the lifetime of Asterisk and do not explicitly need to have their lifetime modified. An exception to this is when the user of a format does not know where the format came from *and* the user may outlive the provider of the format. This occurs, for example, when a format is read from a channel: the channel may have a format with attributes (hence, non-cached) and the user of the format may last longer than the channel (if the reference to the channel is released prior to the format's reference). For more information on this work, see the API design notes: https://wiki.asterisk.org/wiki/display/AST/Media+Format+Rewrite Finally, this work was the culmination of a large number of developer's efforts. Extra thanks goes to Corey Farrell, who took on a large amount of the work in the Asterisk core, chan_sip, and was an invaluable resource in peer reviews throughout this project. There were a substantial number of patches contributed during this work; the following issues/patch names simply reflect some of the work (and will cause the release scripts to give attribution to the individuals who work on them). Reviews: https://reviewboard.asterisk.org/r/3814 https://reviewboard.asterisk.org/r/3808 https://reviewboard.asterisk.org/r/3805 https://reviewboard.asterisk.org/r/3803 https://reviewboard.asterisk.org/r/3801 https://reviewboard.asterisk.org/r/3798 https://reviewboard.asterisk.org/r/3800 https://reviewboard.asterisk.org/r/3794 https://reviewboard.asterisk.org/r/3793 https://reviewboard.asterisk.org/r/3792 https://reviewboard.asterisk.org/r/3791 https://reviewboard.asterisk.org/r/3790 https://reviewboard.asterisk.org/r/3789 https://reviewboard.asterisk.org/r/3788 https://reviewboard.asterisk.org/r/3787 https://reviewboard.asterisk.org/r/3786 https://reviewboard.asterisk.org/r/3784 https://reviewboard.asterisk.org/r/3783 https://reviewboard.asterisk.org/r/3778 https://reviewboard.asterisk.org/r/3774 https://reviewboard.asterisk.org/r/3775 https://reviewboard.asterisk.org/r/3772 https://reviewboard.asterisk.org/r/3761 https://reviewboard.asterisk.org/r/3754 https://reviewboard.asterisk.org/r/3753 https://reviewboard.asterisk.org/r/3751 https://reviewboard.asterisk.org/r/3750 https://reviewboard.asterisk.org/r/3748 https://reviewboard.asterisk.org/r/3747 https://reviewboard.asterisk.org/r/3746 https://reviewboard.asterisk.org/r/3742 https://reviewboard.asterisk.org/r/3740 https://reviewboard.asterisk.org/r/3739 https://reviewboard.asterisk.org/r/3738 https://reviewboard.asterisk.org/r/3737 https://reviewboard.asterisk.org/r/3736 https://reviewboard.asterisk.org/r/3734 https://reviewboard.asterisk.org/r/3722 https://reviewboard.asterisk.org/r/3713 https://reviewboard.asterisk.org/r/3703 https://reviewboard.asterisk.org/r/3689 https://reviewboard.asterisk.org/r/3687 https://reviewboard.asterisk.org/r/3674 https://reviewboard.asterisk.org/r/3671 https://reviewboard.asterisk.org/r/3667 https://reviewboard.asterisk.org/r/3665 https://reviewboard.asterisk.org/r/3625 https://reviewboard.asterisk.org/r/3602 https://reviewboard.asterisk.org/r/3519 https://reviewboard.asterisk.org/r/3518 https://reviewboard.asterisk.org/r/3516 https://reviewboard.asterisk.org/r/3515 https://reviewboard.asterisk.org/r/3512 https://reviewboard.asterisk.org/r/3506 https://reviewboard.asterisk.org/r/3413 https://reviewboard.asterisk.org/r/3410 https://reviewboard.asterisk.org/r/3387 https://reviewboard.asterisk.org/r/3388 https://reviewboard.asterisk.org/r/3389 https://reviewboard.asterisk.org/r/3390 https://reviewboard.asterisk.org/r/3321 https://reviewboard.asterisk.org/r/3320 https://reviewboard.asterisk.org/r/3319 https://reviewboard.asterisk.org/r/3318 https://reviewboard.asterisk.org/r/3266 https://reviewboard.asterisk.org/r/3265 https://reviewboard.asterisk.org/r/3234 https://reviewboard.asterisk.org/r/3178 ASTERISK-23114 #close Reported by: mjordan media_formats_translation_core.diff uploaded by kharwell (License 6464) rb3506.diff uploaded by mjordan (License 6283) media_format_app_file.diff uploaded by kharwell (License 6464) misc-2.diff uploaded by file (License 5000) chan_mild-3.diff uploaded by file (License 5000) chan_obscure.diff uploaded by file (License 5000) jingle.diff uploaded by file (License 5000) funcs.diff uploaded by file (License 5000) formats.diff uploaded by file (License 5000) core.diff uploaded by file (License 5000) bridges.diff uploaded by file (License 5000) mf-codecs-2.diff uploaded by file (License 5000) mf-app_fax.diff uploaded by file (License 5000) mf-apps-3.diff uploaded by file (License 5000) media-formats-3.diff uploaded by file (License 5000) ASTERISK-23715 rb3713.patch uploaded by coreyfarrell (License 5909) rb3689.patch uploaded by mjordan (License 6283) ASTERISK-23957 rb3722.patch uploaded by mjordan (License 6283) mf-attributes-3.diff uploaded by file (License 5000) ASTERISK-23958 Tested by: jrose rb3822.patch uploaded by coreyfarrell (License 5909) rb3800.patch uploaded by jrose (License 6182) chan_sip.diff uploaded by mjordan (License 6283) rb3747.patch uploaded by jrose (License 6182) ASTERISK-23959 #close Tested by: sgriepentrog, mjordan, coreyfarrell sip_cleanup.diff uploaded by opticron (License 6273) chan_sip_caps.diff uploaded by mjordan (License 6283) rb3751.patch uploaded by coreyfarrell (License 5909) chan_sip-3.diff uploaded by file (License 5000) ASTERISK-23960 #close Tested by: opticron direct_media.diff uploaded by opticron (License 6273) pjsip-direct-media.diff uploaded by file (License 5000) format_cap_remove.diff uploaded by opticron (License 6273) media_format_fixes.diff uploaded by opticron (License 6273) chan_pjsip-2.diff uploaded by file (License 5000) ASTERISK-23966 #close Tested by: rmudgett rb3803.patch uploaded by rmudgetti (License 5621) chan_dahdi.diff uploaded by file (License 5000) ASTERISK-24064 #close Tested by: coreyfarrell, mjordan, opticron, file, rmudgett, sgriepentrog, jrose rb3814.patch uploaded by rmudgett (License 5621) moh_cleanup.diff uploaded by opticron (License 6273) bridge_leak.diff uploaded by opticron (License 6273) translate.diff uploaded by file (License 5000) rb3795.patch uploaded by rmudgett (License 5621) tls_fix.diff uploaded by mjordan (License 6283) fax-mf-fix-2.diff uploaded by file (License 5000) rtp_transfer_stuff uploaded by mjordan (License 6283) rb3787.patch uploaded by rmudgett (License 5621) media-formats-explicit-translate-format-3.diff uploaded by file (License 5000) format_cache_case_fix.diff uploaded by opticron (License 6273) rb3774.patch uploaded by rmudgett (License 5621) rb3775.patch uploaded by rmudgett (License 5621) rtp_engine_fix.diff uploaded by opticron (License 6273) rtp_crash_fix.diff uploaded by opticron (License 6273) rb3753.patch uploaded by mjordan (License 6283) rb3750.patch uploaded by mjordan (License 6283) rb3748.patch uploaded by rmudgett (License 5621) media_format_fixes.diff uploaded by opticron (License 6273) rb3740.patch uploaded by mjordan (License 6283) rb3739.patch uploaded by mjordan (License 6283) rb3734.patch uploaded by mjordan (License 6283) rb3689.patch uploaded by mjordan (License 6283) rb3674.patch uploaded by coreyfarrell (License 5909) rb3671.patch uploaded by coreyfarrell (License 5909) rb3667.patch uploaded by coreyfarrell (License 5909) rb3665.patch uploaded by mjordan (License 6283) rb3625.patch uploaded by coreyfarrell (License 5909) rb3602.patch uploaded by coreyfarrell (License 5909) format_compatibility-2.diff uploaded by file (License 5000) core.diff uploaded by file (License 5000) git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@419044 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-07-20 22:06:33 +00:00
if (framing_ms) {
rtp->smoother = ast_smoother_new((framing_ms * ast_format_get_minimum_bytes(format)) / ast_format_get_minimum_ms(format));
if (!rtp->smoother) {
ast_log(LOG_WARNING, "Unable to create smoother: format %s ms: %u len: %u\n",
ast_format_get_name(format), framing_ms, ast_format_get_minimum_bytes(format));
return -1;
}
ast_smoother_set_flags(rtp->smoother, smoother_flags);
}
}
/* Feed audio frames into the actual function that will create a frame and send it */
if (rtp->smoother) {
struct ast_frame *f;
if (ast_smoother_test_flag(rtp->smoother, AST_SMOOTHER_FLAG_BE)) {
ast_smoother_feed_be(rtp->smoother, frame);
} else {
ast_smoother_feed(rtp->smoother, frame);
}
while ((f = ast_smoother_read(rtp->smoother)) && (f->data.ptr)) {
rtp_raw_write(instance, f, codec);
}
} else {
int hdrlen = 12;
struct ast_frame *f = NULL;
if (frame->offset < hdrlen) {
f = ast_frdup(frame);
} else {
f = frame;
}
if (f->data.ptr) {
rtp_raw_write(instance, f, codec);
}
if (f != frame) {
ast_frfree(f);
}
}
return 0;
}
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
static void calc_rxstamp_and_jitter(struct timeval *tv,
struct ast_rtp *rtp, unsigned int rx_rtp_ts,
int mark)
{
int rate = ast_rtp_get_rate(rtp->f.subclass.format);
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
double jitter = 0.0;
double prev_jitter = 0.0;
struct timeval now;
struct timeval tmp;
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
double rxnow;
double arrival_sec;
unsigned int arrival;
int transit;
int d;
gettimeofday(&now,NULL);
if (rtp->rxcount == 1 || mark) {
rtp->rxstart = ast_tv2double(&now);
rtp->remote_seed_rx_rtp_ts = rx_rtp_ts;
/*
* "tv" is placed in the received frame's
* "delivered" field and when this frame is
* sent out again on the other side, it's
* used to calculate the timestamp on the
* outgoing RTP packets.
*
* NOTE: We need to do integer math here
* because double math rounding issues can
* generate incorrect timestamps.
*/
rtp->rxcore = now;
tmp = ast_samp2tv(rx_rtp_ts, rate);
rtp->rxcore = ast_tvsub(rtp->rxcore, tmp);
rtp->rxcore.tv_usec -= rtp->rxcore.tv_usec % 100;
*tv = ast_tvadd(rtp->rxcore, tmp);
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
ast_debug_rtcp(3, "%s: "
"Seed ts: %u current time: %f\n",
ast_rtp_instance_get_channel_id(rtp->owner)
, rx_rtp_ts
, rtp->rxstart
);
return;
}
tmp = ast_samp2tv(rx_rtp_ts, rate);
/* See the comment about "tv" above. Even if
* we don't use this received packet for jitter
* calculations, we still need to set tv so the
* timestamp will be correct when this packet is
* sent out again.
*/
*tv = ast_tvadd(rtp->rxcore, tmp);
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
/*
* The first few packets are generally unstable so let's
* not use them in the calculations.
*/
if (rtp->rxcount < RTP_IGNORE_FIRST_PACKETS_COUNT) {
ast_debug_rtcp(3, "%s: Packet %d < %d. Ignoring\n",
ast_rtp_instance_get_channel_id(rtp->owner)
, rtp->rxcount
, RTP_IGNORE_FIRST_PACKETS_COUNT
);
return;
}
/*
* First good packet. Capture the start time and timestamp
* but don't actually use this packet for calculation.
*/
if (rtp->rxcount == RTP_IGNORE_FIRST_PACKETS_COUNT) {
rtp->rxstart_stable = ast_tv2double(&now);
rtp->remote_seed_rx_rtp_ts_stable = rx_rtp_ts;
rtp->last_transit_time_samples = -rx_rtp_ts;
ast_debug_rtcp(3, "%s: "
"pkt: %5u Stable Seed ts: %u current time: %f\n",
ast_rtp_instance_get_channel_id(rtp->owner)
, rtp->rxcount
, rx_rtp_ts
, rtp->rxstart_stable
);
return;
}
/*
* If the current packet isn't in sequence, don't
* use it in any calculations as remote_current_rx_rtp_ts
* is not going to be correct.
*/
if (rtp->lastrxseqno != rtp->prevrxseqno + 1) {
ast_debug_rtcp(3, "%s: Current packet seq %d != last packet seq %d + 1. Ignoring\n",
ast_rtp_instance_get_channel_id(rtp->owner)
, rtp->lastrxseqno
, rtp->prevrxseqno
);
return;
}
/*
* The following calculations are taken from
* https://www.rfc-editor.org/rfc/rfc3550#appendix-A.8
*
* The received rtp timestamp is the random "seed"
* timestamp chosen by the sender when they sent the
* first packet, plus the number of samples since then.
*
* To get our arrival time in the same units, we
* calculate the time difference in seconds between
* when we received the first packet and when we
* received this packet and convert that to samples.
*/
rxnow = ast_tv2double(&now);
arrival_sec = rxnow - rtp->rxstart_stable;
arrival = ast_sec2samp(arrival_sec, rate);
/*
* Now we can use the exact formula in
* https://www.rfc-editor.org/rfc/rfc3550#appendix-A.8 :
*
* int transit = arrival - r->ts;
* int d = transit - s->transit;
* s->transit = transit;
* if (d < 0) d = -d;
* s->jitter += (1./16.) * ((double)d - s->jitter);
*
* Our rx_rtp_ts is their r->ts.
* Our rtp->last_transit_time_samples is their s->transit.
* Our rtp->rxjitter is their s->jitter.
*/
transit = arrival - rx_rtp_ts;
d = transit - rtp->last_transit_time_samples;
if (d < 0) {
d = -d;
}
prev_jitter = rtp->rxjitter_samples;
jitter = (1.0/16.0) * (((double)d) - prev_jitter);
rtp->rxjitter_samples = prev_jitter + jitter;
/*
* We need to hang on to jitter in both samples and seconds.
*/
rtp->rxjitter = ast_samp2sec(rtp->rxjitter_samples, rate);
ast_debug_rtcp(3, "%s: pkt: %5u "
"Arrival sec: %7.3f Arrival ts: %10u RX ts: %10u "
"Transit samp: %6d Last transit samp: %6d d: %4d "
"Curr jitter: %7.0f(%7.3f) Prev Jitter: %7.0f(%7.3f) New Jitter: %7.0f(%7.3f)\n",
ast_rtp_instance_get_channel_id(rtp->owner)
, rtp->rxcount
, arrival_sec
, arrival
, rx_rtp_ts
, transit
, rtp->last_transit_time_samples
, d
, jitter
, ast_samp2sec(jitter, rate)
, prev_jitter
, ast_samp2sec(prev_jitter, rate)
, rtp->rxjitter_samples
, rtp->rxjitter
);
rtp->last_transit_time_samples = transit;
/*
* Update all the stats.
*/
if (rtp->rtcp) {
if (rtp->rxjitter > rtp->rtcp->maxrxjitter)
rtp->rtcp->maxrxjitter = rtp->rxjitter;
if (rtp->rtcp->rxjitter_count == 1)
rtp->rtcp->minrxjitter = rtp->rxjitter;
if (rtp->rtcp && rtp->rxjitter < rtp->rtcp->minrxjitter)
rtp->rtcp->minrxjitter = rtp->rxjitter;
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
calc_mean_and_standard_deviation(rtp->rxjitter,
&rtp->rtcp->normdev_rxjitter, &rtp->rtcp->stdev_rxjitter,
&rtp->rtcp->rxjitter_count);
}
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
return;
}
Recorded merge of revisions 254452 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r254452 | mmichelson | 2010-03-25 10:59:56 -0500 (Thu, 25 Mar 2010) | 44 lines Several fixes regarding RFC2833 DTMF detection. Here is a copy and paste of the details from my request on reviewboard that dealt with these changes: Fix 1. The first change in place is to fix Mantis issue 15811, which deals with a situation where Asterisk will incorrectly interpret out of order RFC2833 frames as duplicate DTMF digits. For instance, we would receive a sequence like: seqno 1: DTMF 1 seqno 2: DTMF 1 seqno 3: DTMF 1 seqno 4: DTMF 1 seqno 6: DTMF 1 (end) seqno 5: DTMF 1 seqno 7: DTMF 1 (end) seqno 8: DTMF 1 (end) Prior to this patch when we received the frame with seqno 5, we would interpret this as a new DTMF 1. With this patch, we will check the seqno of the incoming digit and not process the frame if the seqno is lower than the last recorded seqno. Note that we do not record the seqno of the dropped DTMF frame for future processing. While the above situation is what was designed to be fixed, the patch is written in such a way that the following would also be fixed too: seqno 9: DTMF 1 seqno 10: DTMF 1 (end) seqno 11: DTMF 1 (end) seqno 13: DTMF 2 seqno 12: DTMF 1 (end) seqno 14: DTMF 2 seqno 15: DTMF 2 (end) seqno 16: DTMF 2 (end) seqno 17: DTMF 2 (end) In this second situation, the beginning of the DTMF 2 arrives before the final end frame of the DTMF 1. With the patch, seqno 12 is no processed and thus we properly interpret the DTMF. Fix 2. The second change in place is to fix an issue like the following: seqno 1: DTMF 1 seqno 2: DTMF 1 seqno 3: DTMF 1 (end) *packet lost* seqno 4: DTMF 1 (end) *packet lost* seqno 5: DTMF 1 (end) *packet lost* seqno 6: DTMF 2 When we receive seqno 6, we had code in place that was supposed to properly end the previously unended DTMF 1. The problem was that the code was essentially a no-op. The code would set up an end frame for the DTMF 1 but would immediately overwrite the frame with the begin for DTMF 2. I changed process_dtmf_rfc2833() so that instead of returning a single frame, it is given as an output parameter a list of frames. Each frame that needs to be returned is appended to this list. Fix 3. The final change is a minor one where an AST_CONTROL_SRCCHANGE frame could get lost. If we process a cisco DTMF or an RFC 3389 frame and no frame was returned, then we would return &ast_null_frame. The problem is that earlier in the function, we may have generated an AST_CONTROL_SRCCHANGE frame and put it in the list of frames we wish to return. This frame would be lost in such a case. The patch fixes this problem ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@254454 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2010-03-25 16:04:48 +00:00
static struct ast_frame *create_dtmf_frame(struct ast_rtp_instance *instance, enum ast_frame_type type, int compensate)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
struct ast_sockaddr remote_address = { {0,} };
ast_rtp_instance_get_remote_address(instance, &remote_address);
if (((compensate && type == AST_FRAME_DTMF_END) || (type == AST_FRAME_DTMF_BEGIN)) && ast_tvcmp(ast_tvnow(), rtp->dtmfmute) < 0) {
ast_debug_rtp(1, "(%p) RTP ignore potential DTMF echo from '%s'\n",
instance, ast_sockaddr_stringify(&remote_address));
rtp->resp = 0;
rtp->dtmfsamples = 0;
return &ast_null_frame;
} else if (type == AST_FRAME_DTMF_BEGIN && rtp->resp == 'X') {
ast_debug_rtp(1, "(%p) RTP ignore flash begin from '%s'\n",
instance, ast_sockaddr_stringify(&remote_address));
rtp->resp = 0;
rtp->dtmfsamples = 0;
return &ast_null_frame;
}
if (rtp->resp == 'X') {
ast_debug_rtp(1, "(%p) RTP creating flash Frame at %s\n",
instance, ast_sockaddr_stringify(&remote_address));
rtp->f.frametype = AST_FRAME_CONTROL;
rtp->f.subclass.integer = AST_CONTROL_FLASH;
} else {
ast_debug_rtp(1, "(%p) RTP creating %s DTMF Frame: %d (%c), at %s\n",
instance, type == AST_FRAME_DTMF_END ? "END" : "BEGIN",
rtp->resp, rtp->resp,
ast_sockaddr_stringify(&remote_address));
rtp->f.frametype = type;
rtp->f.subclass.integer = rtp->resp;
}
rtp->f.datalen = 0;
rtp->f.samples = 0;
rtp->f.mallocd = 0;
rtp->f.src = "RTP";
Recorded merge of revisions 254452 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r254452 | mmichelson | 2010-03-25 10:59:56 -0500 (Thu, 25 Mar 2010) | 44 lines Several fixes regarding RFC2833 DTMF detection. Here is a copy and paste of the details from my request on reviewboard that dealt with these changes: Fix 1. The first change in place is to fix Mantis issue 15811, which deals with a situation where Asterisk will incorrectly interpret out of order RFC2833 frames as duplicate DTMF digits. For instance, we would receive a sequence like: seqno 1: DTMF 1 seqno 2: DTMF 1 seqno 3: DTMF 1 seqno 4: DTMF 1 seqno 6: DTMF 1 (end) seqno 5: DTMF 1 seqno 7: DTMF 1 (end) seqno 8: DTMF 1 (end) Prior to this patch when we received the frame with seqno 5, we would interpret this as a new DTMF 1. With this patch, we will check the seqno of the incoming digit and not process the frame if the seqno is lower than the last recorded seqno. Note that we do not record the seqno of the dropped DTMF frame for future processing. While the above situation is what was designed to be fixed, the patch is written in such a way that the following would also be fixed too: seqno 9: DTMF 1 seqno 10: DTMF 1 (end) seqno 11: DTMF 1 (end) seqno 13: DTMF 2 seqno 12: DTMF 1 (end) seqno 14: DTMF 2 seqno 15: DTMF 2 (end) seqno 16: DTMF 2 (end) seqno 17: DTMF 2 (end) In this second situation, the beginning of the DTMF 2 arrives before the final end frame of the DTMF 1. With the patch, seqno 12 is no processed and thus we properly interpret the DTMF. Fix 2. The second change in place is to fix an issue like the following: seqno 1: DTMF 1 seqno 2: DTMF 1 seqno 3: DTMF 1 (end) *packet lost* seqno 4: DTMF 1 (end) *packet lost* seqno 5: DTMF 1 (end) *packet lost* seqno 6: DTMF 2 When we receive seqno 6, we had code in place that was supposed to properly end the previously unended DTMF 1. The problem was that the code was essentially a no-op. The code would set up an end frame for the DTMF 1 but would immediately overwrite the frame with the begin for DTMF 2. I changed process_dtmf_rfc2833() so that instead of returning a single frame, it is given as an output parameter a list of frames. Each frame that needs to be returned is appended to this list. Fix 3. The final change is a minor one where an AST_CONTROL_SRCCHANGE frame could get lost. If we process a cisco DTMF or an RFC 3389 frame and no frame was returned, then we would return &ast_null_frame. The problem is that earlier in the function, we may have generated an AST_CONTROL_SRCCHANGE frame and put it in the list of frames we wish to return. This frame would be lost in such a case. The patch fixes this problem ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@254454 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2010-03-25 16:04:48 +00:00
AST_LIST_NEXT(&rtp->f, frame_list) = NULL;
return &rtp->f;
}
2018-05-10 18:11:06 +00:00
static void process_dtmf_rfc2833(struct ast_rtp_instance *instance, unsigned char *data, int len, unsigned int seqno, unsigned int timestamp, int payloadtype, int mark, struct frame_list *frames)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
struct ast_sockaddr remote_address = { {0,} };
unsigned int event, event_end, samples;
char resp = 0;
struct ast_frame *f = NULL;
ast_rtp_instance_get_remote_address(instance, &remote_address);
/* Figure out event, event end, and samples */
event = ntohl(*((unsigned int *)(data)));
event >>= 24;
event_end = ntohl(*((unsigned int *)(data)));
event_end <<= 8;
event_end >>= 24;
samples = ntohl(*((unsigned int *)(data)));
samples &= 0xFFFF;
if (rtp_debug_test_addr(&remote_address)) {
ast_verbose("Got RTP RFC2833 from %s (type %-2.2d, seq %-6.6u, ts %-6.6u, len %-6.6d, mark %d, event %08x, end %d, duration %-5.5u) \n",
ast_sockaddr_stringify(&remote_address),
payloadtype, seqno, timestamp, len, (mark?1:0), event, ((event_end & 0x80)?1:0), samples);
}
/* Print out debug if turned on */
if (ast_debug_rtp_packet_is_allowed)
ast_debug(0, "- RTP 2833 Event: %08x (len = %d)\n", event, len);
/* Figure out what digit was pressed */
if (event < 10) {
resp = '0' + event;
} else if (event < 11) {
resp = '*';
} else if (event < 12) {
resp = '#';
} else if (event < 16) {
resp = 'A' + (event - 12);
} else if (event < 17) { /* Event 16: Hook flash */
resp = 'X';
} else {
/* Not a supported event */
ast_debug_rtp(1, "(%p) RTP ignoring RTP 2833 Event: %08x. Not a DTMF Digit.\n", instance, event);
Recorded merge of revisions 254452 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r254452 | mmichelson | 2010-03-25 10:59:56 -0500 (Thu, 25 Mar 2010) | 44 lines Several fixes regarding RFC2833 DTMF detection. Here is a copy and paste of the details from my request on reviewboard that dealt with these changes: Fix 1. The first change in place is to fix Mantis issue 15811, which deals with a situation where Asterisk will incorrectly interpret out of order RFC2833 frames as duplicate DTMF digits. For instance, we would receive a sequence like: seqno 1: DTMF 1 seqno 2: DTMF 1 seqno 3: DTMF 1 seqno 4: DTMF 1 seqno 6: DTMF 1 (end) seqno 5: DTMF 1 seqno 7: DTMF 1 (end) seqno 8: DTMF 1 (end) Prior to this patch when we received the frame with seqno 5, we would interpret this as a new DTMF 1. With this patch, we will check the seqno of the incoming digit and not process the frame if the seqno is lower than the last recorded seqno. Note that we do not record the seqno of the dropped DTMF frame for future processing. While the above situation is what was designed to be fixed, the patch is written in such a way that the following would also be fixed too: seqno 9: DTMF 1 seqno 10: DTMF 1 (end) seqno 11: DTMF 1 (end) seqno 13: DTMF 2 seqno 12: DTMF 1 (end) seqno 14: DTMF 2 seqno 15: DTMF 2 (end) seqno 16: DTMF 2 (end) seqno 17: DTMF 2 (end) In this second situation, the beginning of the DTMF 2 arrives before the final end frame of the DTMF 1. With the patch, seqno 12 is no processed and thus we properly interpret the DTMF. Fix 2. The second change in place is to fix an issue like the following: seqno 1: DTMF 1 seqno 2: DTMF 1 seqno 3: DTMF 1 (end) *packet lost* seqno 4: DTMF 1 (end) *packet lost* seqno 5: DTMF 1 (end) *packet lost* seqno 6: DTMF 2 When we receive seqno 6, we had code in place that was supposed to properly end the previously unended DTMF 1. The problem was that the code was essentially a no-op. The code would set up an end frame for the DTMF 1 but would immediately overwrite the frame with the begin for DTMF 2. I changed process_dtmf_rfc2833() so that instead of returning a single frame, it is given as an output parameter a list of frames. Each frame that needs to be returned is appended to this list. Fix 3. The final change is a minor one where an AST_CONTROL_SRCCHANGE frame could get lost. If we process a cisco DTMF or an RFC 3389 frame and no frame was returned, then we would return &ast_null_frame. The problem is that earlier in the function, we may have generated an AST_CONTROL_SRCCHANGE frame and put it in the list of frames we wish to return. This frame would be lost in such a case. The patch fixes this problem ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@254454 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2010-03-25 16:04:48 +00:00
return;
}
if (ast_rtp_instance_get_prop(instance, AST_RTP_PROPERTY_DTMF_COMPENSATE)) {
if (!rtp->last_end_timestamp.is_set || rtp->last_end_timestamp.ts != timestamp || (rtp->resp && rtp->resp != resp)) {
rtp->resp = resp;
rtp->dtmf_timeout = 0;
Recorded merge of revisions 254452 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r254452 | mmichelson | 2010-03-25 10:59:56 -0500 (Thu, 25 Mar 2010) | 44 lines Several fixes regarding RFC2833 DTMF detection. Here is a copy and paste of the details from my request on reviewboard that dealt with these changes: Fix 1. The first change in place is to fix Mantis issue 15811, which deals with a situation where Asterisk will incorrectly interpret out of order RFC2833 frames as duplicate DTMF digits. For instance, we would receive a sequence like: seqno 1: DTMF 1 seqno 2: DTMF 1 seqno 3: DTMF 1 seqno 4: DTMF 1 seqno 6: DTMF 1 (end) seqno 5: DTMF 1 seqno 7: DTMF 1 (end) seqno 8: DTMF 1 (end) Prior to this patch when we received the frame with seqno 5, we would interpret this as a new DTMF 1. With this patch, we will check the seqno of the incoming digit and not process the frame if the seqno is lower than the last recorded seqno. Note that we do not record the seqno of the dropped DTMF frame for future processing. While the above situation is what was designed to be fixed, the patch is written in such a way that the following would also be fixed too: seqno 9: DTMF 1 seqno 10: DTMF 1 (end) seqno 11: DTMF 1 (end) seqno 13: DTMF 2 seqno 12: DTMF 1 (end) seqno 14: DTMF 2 seqno 15: DTMF 2 (end) seqno 16: DTMF 2 (end) seqno 17: DTMF 2 (end) In this second situation, the beginning of the DTMF 2 arrives before the final end frame of the DTMF 1. With the patch, seqno 12 is no processed and thus we properly interpret the DTMF. Fix 2. The second change in place is to fix an issue like the following: seqno 1: DTMF 1 seqno 2: DTMF 1 seqno 3: DTMF 1 (end) *packet lost* seqno 4: DTMF 1 (end) *packet lost* seqno 5: DTMF 1 (end) *packet lost* seqno 6: DTMF 2 When we receive seqno 6, we had code in place that was supposed to properly end the previously unended DTMF 1. The problem was that the code was essentially a no-op. The code would set up an end frame for the DTMF 1 but would immediately overwrite the frame with the begin for DTMF 2. I changed process_dtmf_rfc2833() so that instead of returning a single frame, it is given as an output parameter a list of frames. Each frame that needs to be returned is appended to this list. Fix 3. The final change is a minor one where an AST_CONTROL_SRCCHANGE frame could get lost. If we process a cisco DTMF or an RFC 3389 frame and no frame was returned, then we would return &ast_null_frame. The problem is that earlier in the function, we may have generated an AST_CONTROL_SRCCHANGE frame and put it in the list of frames we wish to return. This frame would be lost in such a case. The patch fixes this problem ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@254454 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2010-03-25 16:04:48 +00:00
f = ast_frdup(create_dtmf_frame(instance, AST_FRAME_DTMF_END, ast_rtp_instance_get_prop(instance, AST_RTP_PROPERTY_DTMF_COMPENSATE)));
f->len = 0;
rtp->last_end_timestamp.ts = timestamp;
rtp->last_end_timestamp.is_set = 1;
Recorded merge of revisions 254452 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r254452 | mmichelson | 2010-03-25 10:59:56 -0500 (Thu, 25 Mar 2010) | 44 lines Several fixes regarding RFC2833 DTMF detection. Here is a copy and paste of the details from my request on reviewboard that dealt with these changes: Fix 1. The first change in place is to fix Mantis issue 15811, which deals with a situation where Asterisk will incorrectly interpret out of order RFC2833 frames as duplicate DTMF digits. For instance, we would receive a sequence like: seqno 1: DTMF 1 seqno 2: DTMF 1 seqno 3: DTMF 1 seqno 4: DTMF 1 seqno 6: DTMF 1 (end) seqno 5: DTMF 1 seqno 7: DTMF 1 (end) seqno 8: DTMF 1 (end) Prior to this patch when we received the frame with seqno 5, we would interpret this as a new DTMF 1. With this patch, we will check the seqno of the incoming digit and not process the frame if the seqno is lower than the last recorded seqno. Note that we do not record the seqno of the dropped DTMF frame for future processing. While the above situation is what was designed to be fixed, the patch is written in such a way that the following would also be fixed too: seqno 9: DTMF 1 seqno 10: DTMF 1 (end) seqno 11: DTMF 1 (end) seqno 13: DTMF 2 seqno 12: DTMF 1 (end) seqno 14: DTMF 2 seqno 15: DTMF 2 (end) seqno 16: DTMF 2 (end) seqno 17: DTMF 2 (end) In this second situation, the beginning of the DTMF 2 arrives before the final end frame of the DTMF 1. With the patch, seqno 12 is no processed and thus we properly interpret the DTMF. Fix 2. The second change in place is to fix an issue like the following: seqno 1: DTMF 1 seqno 2: DTMF 1 seqno 3: DTMF 1 (end) *packet lost* seqno 4: DTMF 1 (end) *packet lost* seqno 5: DTMF 1 (end) *packet lost* seqno 6: DTMF 2 When we receive seqno 6, we had code in place that was supposed to properly end the previously unended DTMF 1. The problem was that the code was essentially a no-op. The code would set up an end frame for the DTMF 1 but would immediately overwrite the frame with the begin for DTMF 2. I changed process_dtmf_rfc2833() so that instead of returning a single frame, it is given as an output parameter a list of frames. Each frame that needs to be returned is appended to this list. Fix 3. The final change is a minor one where an AST_CONTROL_SRCCHANGE frame could get lost. If we process a cisco DTMF or an RFC 3389 frame and no frame was returned, then we would return &ast_null_frame. The problem is that earlier in the function, we may have generated an AST_CONTROL_SRCCHANGE frame and put it in the list of frames we wish to return. This frame would be lost in such a case. The patch fixes this problem ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@254454 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2010-03-25 16:04:48 +00:00
AST_LIST_INSERT_TAIL(frames, f, frame_list);
}
} else {
/* The duration parameter measures the complete
duration of the event (from the beginning) - RFC2833.
Account for the fact that duration is only 16 bits long
(about 8 seconds at 8000 Hz) and can wrap is digit
is hold for too long. */
unsigned int new_duration = rtp->dtmf_duration;
unsigned int last_duration = new_duration & 0xFFFF;
if (last_duration > 64000 && samples < last_duration) {
new_duration += 0xFFFF + 1;
}
new_duration = (new_duration & ~0xFFFF) | samples;
if (event_end & 0x80) {
/* End event */
if (rtp->last_seqno != seqno && (!rtp->last_end_timestamp.is_set || timestamp > rtp->last_end_timestamp.ts)) {
rtp->last_end_timestamp.ts = timestamp;
rtp->last_end_timestamp.is_set = 1;
rtp->dtmf_duration = new_duration;
Handle extremely out of order RFC 2833 DTMF The current implementation of RFC 2833 DTMF handling in res_rtp_asterisk will, if a packet arrives out of order, drop the packet. This is to prevent duplicate ton generation in the Asterisk core. Since the RTP layer does not buffer data itself, this is the only option the RTP layer currently has for handling packets that arrive out of order. For the most part, this doesn't matter. For a particular digit, so long as a BEGIN packet arrives before the first END packet, the digit will be produced. If subsequent BEGIN packets arrive interleaved with the ENDs, they will be dropped; likewise, if the BEGIN or END packets themselves are out of order, those packets are dropped but sufficient information is conveyed to the Asterisk core to produce the appropriate digit. For certain sequences of DTMF packets - most notably when, for a particular digit, an END packet arrives before any BEGIN packet for that digit - this is a real problem. When an END arrives before any BEGINs, the END packet is dropped - but at the same time, it causes subsequent BEGIN packets for that digit to be ignored. When the next in order END packet arrives, it too is dropped - Asterisk believes that there was no initial BEGIN. The solution this patch provides is to trust the END packet to convey the information needed for the Asterisk core to produce the DTMF digit. If we receive an END packet, and it: * Has a timestamp greater then the last timestamp received from an END packet * Does not have the same sequence number as the last received sequence number (and is thus not an END packet retransmission) Then we send the END frame up to the Asterisk core. It contains enough DTMF information for Asterisk to produce the digit. On the other hand, if we receive a BEGIN or continuation packet that occurs with a timestamp equal to or less then the last END timestamp, then we've received something out of order - but we already have received enough information to produce the digit. These packets are dropped. Much thanks goes to Olle Johansson (oej) for providing the idea for this solution. Review: https://reviewboard.asterisk.org/r/2033/ (closes issue ASTERISK-18404) Reported by: Stephane Chazelas Tested by: Matt Jordan ........ Merged revisions 370252 from http://svn.asterisk.org/svn/asterisk/branches/1.8 ........ Merged revisions 370271 from http://svn.asterisk.org/svn/asterisk/branches/10 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@370272 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2012-07-19 21:45:20 +00:00
rtp->resp = resp;
Recorded merge of revisions 254452 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r254452 | mmichelson | 2010-03-25 10:59:56 -0500 (Thu, 25 Mar 2010) | 44 lines Several fixes regarding RFC2833 DTMF detection. Here is a copy and paste of the details from my request on reviewboard that dealt with these changes: Fix 1. The first change in place is to fix Mantis issue 15811, which deals with a situation where Asterisk will incorrectly interpret out of order RFC2833 frames as duplicate DTMF digits. For instance, we would receive a sequence like: seqno 1: DTMF 1 seqno 2: DTMF 1 seqno 3: DTMF 1 seqno 4: DTMF 1 seqno 6: DTMF 1 (end) seqno 5: DTMF 1 seqno 7: DTMF 1 (end) seqno 8: DTMF 1 (end) Prior to this patch when we received the frame with seqno 5, we would interpret this as a new DTMF 1. With this patch, we will check the seqno of the incoming digit and not process the frame if the seqno is lower than the last recorded seqno. Note that we do not record the seqno of the dropped DTMF frame for future processing. While the above situation is what was designed to be fixed, the patch is written in such a way that the following would also be fixed too: seqno 9: DTMF 1 seqno 10: DTMF 1 (end) seqno 11: DTMF 1 (end) seqno 13: DTMF 2 seqno 12: DTMF 1 (end) seqno 14: DTMF 2 seqno 15: DTMF 2 (end) seqno 16: DTMF 2 (end) seqno 17: DTMF 2 (end) In this second situation, the beginning of the DTMF 2 arrives before the final end frame of the DTMF 1. With the patch, seqno 12 is no processed and thus we properly interpret the DTMF. Fix 2. The second change in place is to fix an issue like the following: seqno 1: DTMF 1 seqno 2: DTMF 1 seqno 3: DTMF 1 (end) *packet lost* seqno 4: DTMF 1 (end) *packet lost* seqno 5: DTMF 1 (end) *packet lost* seqno 6: DTMF 2 When we receive seqno 6, we had code in place that was supposed to properly end the previously unended DTMF 1. The problem was that the code was essentially a no-op. The code would set up an end frame for the DTMF 1 but would immediately overwrite the frame with the begin for DTMF 2. I changed process_dtmf_rfc2833() so that instead of returning a single frame, it is given as an output parameter a list of frames. Each frame that needs to be returned is appended to this list. Fix 3. The final change is a minor one where an AST_CONTROL_SRCCHANGE frame could get lost. If we process a cisco DTMF or an RFC 3389 frame and no frame was returned, then we would return &ast_null_frame. The problem is that earlier in the function, we may have generated an AST_CONTROL_SRCCHANGE frame and put it in the list of frames we wish to return. This frame would be lost in such a case. The patch fixes this problem ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@254454 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2010-03-25 16:04:48 +00:00
f = ast_frdup(create_dtmf_frame(instance, AST_FRAME_DTMF_END, 0));
f->len = ast_tvdiff_ms(ast_samp2tv(rtp->dtmf_duration, ast_rtp_get_rate(f->subclass.format)), ast_tv(0, 0));
rtp->resp = 0;
rtp->dtmf_duration = rtp->dtmf_timeout = 0;
Recorded merge of revisions 254452 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r254452 | mmichelson | 2010-03-25 10:59:56 -0500 (Thu, 25 Mar 2010) | 44 lines Several fixes regarding RFC2833 DTMF detection. Here is a copy and paste of the details from my request on reviewboard that dealt with these changes: Fix 1. The first change in place is to fix Mantis issue 15811, which deals with a situation where Asterisk will incorrectly interpret out of order RFC2833 frames as duplicate DTMF digits. For instance, we would receive a sequence like: seqno 1: DTMF 1 seqno 2: DTMF 1 seqno 3: DTMF 1 seqno 4: DTMF 1 seqno 6: DTMF 1 (end) seqno 5: DTMF 1 seqno 7: DTMF 1 (end) seqno 8: DTMF 1 (end) Prior to this patch when we received the frame with seqno 5, we would interpret this as a new DTMF 1. With this patch, we will check the seqno of the incoming digit and not process the frame if the seqno is lower than the last recorded seqno. Note that we do not record the seqno of the dropped DTMF frame for future processing. While the above situation is what was designed to be fixed, the patch is written in such a way that the following would also be fixed too: seqno 9: DTMF 1 seqno 10: DTMF 1 (end) seqno 11: DTMF 1 (end) seqno 13: DTMF 2 seqno 12: DTMF 1 (end) seqno 14: DTMF 2 seqno 15: DTMF 2 (end) seqno 16: DTMF 2 (end) seqno 17: DTMF 2 (end) In this second situation, the beginning of the DTMF 2 arrives before the final end frame of the DTMF 1. With the patch, seqno 12 is no processed and thus we properly interpret the DTMF. Fix 2. The second change in place is to fix an issue like the following: seqno 1: DTMF 1 seqno 2: DTMF 1 seqno 3: DTMF 1 (end) *packet lost* seqno 4: DTMF 1 (end) *packet lost* seqno 5: DTMF 1 (end) *packet lost* seqno 6: DTMF 2 When we receive seqno 6, we had code in place that was supposed to properly end the previously unended DTMF 1. The problem was that the code was essentially a no-op. The code would set up an end frame for the DTMF 1 but would immediately overwrite the frame with the begin for DTMF 2. I changed process_dtmf_rfc2833() so that instead of returning a single frame, it is given as an output parameter a list of frames. Each frame that needs to be returned is appended to this list. Fix 3. The final change is a minor one where an AST_CONTROL_SRCCHANGE frame could get lost. If we process a cisco DTMF or an RFC 3389 frame and no frame was returned, then we would return &ast_null_frame. The problem is that earlier in the function, we may have generated an AST_CONTROL_SRCCHANGE frame and put it in the list of frames we wish to return. This frame would be lost in such a case. The patch fixes this problem ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@254454 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2010-03-25 16:04:48 +00:00
AST_LIST_INSERT_TAIL(frames, f, frame_list);
} else if (ast_debug_rtp_packet_is_allowed) {
ast_debug_rtp(1, "(%p) RTP dropping duplicate or out of order DTMF END frame (seqno: %u, ts %u, digit %c)\n",
instance, seqno, timestamp, resp);
}
} else {
/* Begin/continuation */
Handle extremely out of order RFC 2833 DTMF The current implementation of RFC 2833 DTMF handling in res_rtp_asterisk will, if a packet arrives out of order, drop the packet. This is to prevent duplicate ton generation in the Asterisk core. Since the RTP layer does not buffer data itself, this is the only option the RTP layer currently has for handling packets that arrive out of order. For the most part, this doesn't matter. For a particular digit, so long as a BEGIN packet arrives before the first END packet, the digit will be produced. If subsequent BEGIN packets arrive interleaved with the ENDs, they will be dropped; likewise, if the BEGIN or END packets themselves are out of order, those packets are dropped but sufficient information is conveyed to the Asterisk core to produce the appropriate digit. For certain sequences of DTMF packets - most notably when, for a particular digit, an END packet arrives before any BEGIN packet for that digit - this is a real problem. When an END arrives before any BEGINs, the END packet is dropped - but at the same time, it causes subsequent BEGIN packets for that digit to be ignored. When the next in order END packet arrives, it too is dropped - Asterisk believes that there was no initial BEGIN. The solution this patch provides is to trust the END packet to convey the information needed for the Asterisk core to produce the DTMF digit. If we receive an END packet, and it: * Has a timestamp greater then the last timestamp received from an END packet * Does not have the same sequence number as the last received sequence number (and is thus not an END packet retransmission) Then we send the END frame up to the Asterisk core. It contains enough DTMF information for Asterisk to produce the digit. On the other hand, if we receive a BEGIN or continuation packet that occurs with a timestamp equal to or less then the last END timestamp, then we've received something out of order - but we already have received enough information to produce the digit. These packets are dropped. Much thanks goes to Olle Johansson (oej) for providing the idea for this solution. Review: https://reviewboard.asterisk.org/r/2033/ (closes issue ASTERISK-18404) Reported by: Stephane Chazelas Tested by: Matt Jordan ........ Merged revisions 370252 from http://svn.asterisk.org/svn/asterisk/branches/1.8 ........ Merged revisions 370271 from http://svn.asterisk.org/svn/asterisk/branches/10 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@370272 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2012-07-19 21:45:20 +00:00
/* The second portion of the seqno check is to not mistakenly
* stop accepting DTMF if the seqno rolls over beyond
* 65535.
*/
if ((rtp->last_seqno > seqno && rtp->last_seqno - seqno < 50)
|| (rtp->last_end_timestamp.is_set
&& timestamp <= rtp->last_end_timestamp.ts)) {
Handle extremely out of order RFC 2833 DTMF The current implementation of RFC 2833 DTMF handling in res_rtp_asterisk will, if a packet arrives out of order, drop the packet. This is to prevent duplicate ton generation in the Asterisk core. Since the RTP layer does not buffer data itself, this is the only option the RTP layer currently has for handling packets that arrive out of order. For the most part, this doesn't matter. For a particular digit, so long as a BEGIN packet arrives before the first END packet, the digit will be produced. If subsequent BEGIN packets arrive interleaved with the ENDs, they will be dropped; likewise, if the BEGIN or END packets themselves are out of order, those packets are dropped but sufficient information is conveyed to the Asterisk core to produce the appropriate digit. For certain sequences of DTMF packets - most notably when, for a particular digit, an END packet arrives before any BEGIN packet for that digit - this is a real problem. When an END arrives before any BEGINs, the END packet is dropped - but at the same time, it causes subsequent BEGIN packets for that digit to be ignored. When the next in order END packet arrives, it too is dropped - Asterisk believes that there was no initial BEGIN. The solution this patch provides is to trust the END packet to convey the information needed for the Asterisk core to produce the DTMF digit. If we receive an END packet, and it: * Has a timestamp greater then the last timestamp received from an END packet * Does not have the same sequence number as the last received sequence number (and is thus not an END packet retransmission) Then we send the END frame up to the Asterisk core. It contains enough DTMF information for Asterisk to produce the digit. On the other hand, if we receive a BEGIN or continuation packet that occurs with a timestamp equal to or less then the last END timestamp, then we've received something out of order - but we already have received enough information to produce the digit. These packets are dropped. Much thanks goes to Olle Johansson (oej) for providing the idea for this solution. Review: https://reviewboard.asterisk.org/r/2033/ (closes issue ASTERISK-18404) Reported by: Stephane Chazelas Tested by: Matt Jordan ........ Merged revisions 370252 from http://svn.asterisk.org/svn/asterisk/branches/1.8 ........ Merged revisions 370271 from http://svn.asterisk.org/svn/asterisk/branches/10 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@370272 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2012-07-19 21:45:20 +00:00
/* Out of order frame. Processing this can cause us to
* improperly duplicate incoming DTMF, so just drop
* this.
*/
if (ast_debug_rtp_packet_is_allowed) {
ast_debug(0, "Dropping out of order DTMF frame (seqno %u, ts %u, digit %c)\n",
Handle extremely out of order RFC 2833 DTMF The current implementation of RFC 2833 DTMF handling in res_rtp_asterisk will, if a packet arrives out of order, drop the packet. This is to prevent duplicate ton generation in the Asterisk core. Since the RTP layer does not buffer data itself, this is the only option the RTP layer currently has for handling packets that arrive out of order. For the most part, this doesn't matter. For a particular digit, so long as a BEGIN packet arrives before the first END packet, the digit will be produced. If subsequent BEGIN packets arrive interleaved with the ENDs, they will be dropped; likewise, if the BEGIN or END packets themselves are out of order, those packets are dropped but sufficient information is conveyed to the Asterisk core to produce the appropriate digit. For certain sequences of DTMF packets - most notably when, for a particular digit, an END packet arrives before any BEGIN packet for that digit - this is a real problem. When an END arrives before any BEGINs, the END packet is dropped - but at the same time, it causes subsequent BEGIN packets for that digit to be ignored. When the next in order END packet arrives, it too is dropped - Asterisk believes that there was no initial BEGIN. The solution this patch provides is to trust the END packet to convey the information needed for the Asterisk core to produce the DTMF digit. If we receive an END packet, and it: * Has a timestamp greater then the last timestamp received from an END packet * Does not have the same sequence number as the last received sequence number (and is thus not an END packet retransmission) Then we send the END frame up to the Asterisk core. It contains enough DTMF information for Asterisk to produce the digit. On the other hand, if we receive a BEGIN or continuation packet that occurs with a timestamp equal to or less then the last END timestamp, then we've received something out of order - but we already have received enough information to produce the digit. These packets are dropped. Much thanks goes to Olle Johansson (oej) for providing the idea for this solution. Review: https://reviewboard.asterisk.org/r/2033/ (closes issue ASTERISK-18404) Reported by: Stephane Chazelas Tested by: Matt Jordan ........ Merged revisions 370252 from http://svn.asterisk.org/svn/asterisk/branches/1.8 ........ Merged revisions 370271 from http://svn.asterisk.org/svn/asterisk/branches/10 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@370272 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2012-07-19 21:45:20 +00:00
seqno, timestamp, resp);
}
return;
}
if (rtp->resp && rtp->resp != resp) {
/* Another digit already began. End it */
Recorded merge of revisions 254452 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r254452 | mmichelson | 2010-03-25 10:59:56 -0500 (Thu, 25 Mar 2010) | 44 lines Several fixes regarding RFC2833 DTMF detection. Here is a copy and paste of the details from my request on reviewboard that dealt with these changes: Fix 1. The first change in place is to fix Mantis issue 15811, which deals with a situation where Asterisk will incorrectly interpret out of order RFC2833 frames as duplicate DTMF digits. For instance, we would receive a sequence like: seqno 1: DTMF 1 seqno 2: DTMF 1 seqno 3: DTMF 1 seqno 4: DTMF 1 seqno 6: DTMF 1 (end) seqno 5: DTMF 1 seqno 7: DTMF 1 (end) seqno 8: DTMF 1 (end) Prior to this patch when we received the frame with seqno 5, we would interpret this as a new DTMF 1. With this patch, we will check the seqno of the incoming digit and not process the frame if the seqno is lower than the last recorded seqno. Note that we do not record the seqno of the dropped DTMF frame for future processing. While the above situation is what was designed to be fixed, the patch is written in such a way that the following would also be fixed too: seqno 9: DTMF 1 seqno 10: DTMF 1 (end) seqno 11: DTMF 1 (end) seqno 13: DTMF 2 seqno 12: DTMF 1 (end) seqno 14: DTMF 2 seqno 15: DTMF 2 (end) seqno 16: DTMF 2 (end) seqno 17: DTMF 2 (end) In this second situation, the beginning of the DTMF 2 arrives before the final end frame of the DTMF 1. With the patch, seqno 12 is no processed and thus we properly interpret the DTMF. Fix 2. The second change in place is to fix an issue like the following: seqno 1: DTMF 1 seqno 2: DTMF 1 seqno 3: DTMF 1 (end) *packet lost* seqno 4: DTMF 1 (end) *packet lost* seqno 5: DTMF 1 (end) *packet lost* seqno 6: DTMF 2 When we receive seqno 6, we had code in place that was supposed to properly end the previously unended DTMF 1. The problem was that the code was essentially a no-op. The code would set up an end frame for the DTMF 1 but would immediately overwrite the frame with the begin for DTMF 2. I changed process_dtmf_rfc2833() so that instead of returning a single frame, it is given as an output parameter a list of frames. Each frame that needs to be returned is appended to this list. Fix 3. The final change is a minor one where an AST_CONTROL_SRCCHANGE frame could get lost. If we process a cisco DTMF or an RFC 3389 frame and no frame was returned, then we would return &ast_null_frame. The problem is that earlier in the function, we may have generated an AST_CONTROL_SRCCHANGE frame and put it in the list of frames we wish to return. This frame would be lost in such a case. The patch fixes this problem ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@254454 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2010-03-25 16:04:48 +00:00
f = ast_frdup(create_dtmf_frame(instance, AST_FRAME_DTMF_END, 0));
f->len = ast_tvdiff_ms(ast_samp2tv(rtp->dtmf_duration, ast_rtp_get_rate(f->subclass.format)), ast_tv(0, 0));
rtp->resp = 0;
rtp->dtmf_duration = rtp->dtmf_timeout = 0;
Recorded merge of revisions 254452 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r254452 | mmichelson | 2010-03-25 10:59:56 -0500 (Thu, 25 Mar 2010) | 44 lines Several fixes regarding RFC2833 DTMF detection. Here is a copy and paste of the details from my request on reviewboard that dealt with these changes: Fix 1. The first change in place is to fix Mantis issue 15811, which deals with a situation where Asterisk will incorrectly interpret out of order RFC2833 frames as duplicate DTMF digits. For instance, we would receive a sequence like: seqno 1: DTMF 1 seqno 2: DTMF 1 seqno 3: DTMF 1 seqno 4: DTMF 1 seqno 6: DTMF 1 (end) seqno 5: DTMF 1 seqno 7: DTMF 1 (end) seqno 8: DTMF 1 (end) Prior to this patch when we received the frame with seqno 5, we would interpret this as a new DTMF 1. With this patch, we will check the seqno of the incoming digit and not process the frame if the seqno is lower than the last recorded seqno. Note that we do not record the seqno of the dropped DTMF frame for future processing. While the above situation is what was designed to be fixed, the patch is written in such a way that the following would also be fixed too: seqno 9: DTMF 1 seqno 10: DTMF 1 (end) seqno 11: DTMF 1 (end) seqno 13: DTMF 2 seqno 12: DTMF 1 (end) seqno 14: DTMF 2 seqno 15: DTMF 2 (end) seqno 16: DTMF 2 (end) seqno 17: DTMF 2 (end) In this second situation, the beginning of the DTMF 2 arrives before the final end frame of the DTMF 1. With the patch, seqno 12 is no processed and thus we properly interpret the DTMF. Fix 2. The second change in place is to fix an issue like the following: seqno 1: DTMF 1 seqno 2: DTMF 1 seqno 3: DTMF 1 (end) *packet lost* seqno 4: DTMF 1 (end) *packet lost* seqno 5: DTMF 1 (end) *packet lost* seqno 6: DTMF 2 When we receive seqno 6, we had code in place that was supposed to properly end the previously unended DTMF 1. The problem was that the code was essentially a no-op. The code would set up an end frame for the DTMF 1 but would immediately overwrite the frame with the begin for DTMF 2. I changed process_dtmf_rfc2833() so that instead of returning a single frame, it is given as an output parameter a list of frames. Each frame that needs to be returned is appended to this list. Fix 3. The final change is a minor one where an AST_CONTROL_SRCCHANGE frame could get lost. If we process a cisco DTMF or an RFC 3389 frame and no frame was returned, then we would return &ast_null_frame. The problem is that earlier in the function, we may have generated an AST_CONTROL_SRCCHANGE frame and put it in the list of frames we wish to return. This frame would be lost in such a case. The patch fixes this problem ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@254454 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2010-03-25 16:04:48 +00:00
AST_LIST_INSERT_TAIL(frames, f, frame_list);
}
if (rtp->resp) {
/* Digit continues */
rtp->dtmf_duration = new_duration;
} else {
/* New digit began */
rtp->resp = resp;
Recorded merge of revisions 254452 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r254452 | mmichelson | 2010-03-25 10:59:56 -0500 (Thu, 25 Mar 2010) | 44 lines Several fixes regarding RFC2833 DTMF detection. Here is a copy and paste of the details from my request on reviewboard that dealt with these changes: Fix 1. The first change in place is to fix Mantis issue 15811, which deals with a situation where Asterisk will incorrectly interpret out of order RFC2833 frames as duplicate DTMF digits. For instance, we would receive a sequence like: seqno 1: DTMF 1 seqno 2: DTMF 1 seqno 3: DTMF 1 seqno 4: DTMF 1 seqno 6: DTMF 1 (end) seqno 5: DTMF 1 seqno 7: DTMF 1 (end) seqno 8: DTMF 1 (end) Prior to this patch when we received the frame with seqno 5, we would interpret this as a new DTMF 1. With this patch, we will check the seqno of the incoming digit and not process the frame if the seqno is lower than the last recorded seqno. Note that we do not record the seqno of the dropped DTMF frame for future processing. While the above situation is what was designed to be fixed, the patch is written in such a way that the following would also be fixed too: seqno 9: DTMF 1 seqno 10: DTMF 1 (end) seqno 11: DTMF 1 (end) seqno 13: DTMF 2 seqno 12: DTMF 1 (end) seqno 14: DTMF 2 seqno 15: DTMF 2 (end) seqno 16: DTMF 2 (end) seqno 17: DTMF 2 (end) In this second situation, the beginning of the DTMF 2 arrives before the final end frame of the DTMF 1. With the patch, seqno 12 is no processed and thus we properly interpret the DTMF. Fix 2. The second change in place is to fix an issue like the following: seqno 1: DTMF 1 seqno 2: DTMF 1 seqno 3: DTMF 1 (end) *packet lost* seqno 4: DTMF 1 (end) *packet lost* seqno 5: DTMF 1 (end) *packet lost* seqno 6: DTMF 2 When we receive seqno 6, we had code in place that was supposed to properly end the previously unended DTMF 1. The problem was that the code was essentially a no-op. The code would set up an end frame for the DTMF 1 but would immediately overwrite the frame with the begin for DTMF 2. I changed process_dtmf_rfc2833() so that instead of returning a single frame, it is given as an output parameter a list of frames. Each frame that needs to be returned is appended to this list. Fix 3. The final change is a minor one where an AST_CONTROL_SRCCHANGE frame could get lost. If we process a cisco DTMF or an RFC 3389 frame and no frame was returned, then we would return &ast_null_frame. The problem is that earlier in the function, we may have generated an AST_CONTROL_SRCCHANGE frame and put it in the list of frames we wish to return. This frame would be lost in such a case. The patch fixes this problem ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@254454 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2010-03-25 16:04:48 +00:00
f = ast_frdup(create_dtmf_frame(instance, AST_FRAME_DTMF_BEGIN, 0));
rtp->dtmf_duration = samples;
Recorded merge of revisions 254452 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r254452 | mmichelson | 2010-03-25 10:59:56 -0500 (Thu, 25 Mar 2010) | 44 lines Several fixes regarding RFC2833 DTMF detection. Here is a copy and paste of the details from my request on reviewboard that dealt with these changes: Fix 1. The first change in place is to fix Mantis issue 15811, which deals with a situation where Asterisk will incorrectly interpret out of order RFC2833 frames as duplicate DTMF digits. For instance, we would receive a sequence like: seqno 1: DTMF 1 seqno 2: DTMF 1 seqno 3: DTMF 1 seqno 4: DTMF 1 seqno 6: DTMF 1 (end) seqno 5: DTMF 1 seqno 7: DTMF 1 (end) seqno 8: DTMF 1 (end) Prior to this patch when we received the frame with seqno 5, we would interpret this as a new DTMF 1. With this patch, we will check the seqno of the incoming digit and not process the frame if the seqno is lower than the last recorded seqno. Note that we do not record the seqno of the dropped DTMF frame for future processing. While the above situation is what was designed to be fixed, the patch is written in such a way that the following would also be fixed too: seqno 9: DTMF 1 seqno 10: DTMF 1 (end) seqno 11: DTMF 1 (end) seqno 13: DTMF 2 seqno 12: DTMF 1 (end) seqno 14: DTMF 2 seqno 15: DTMF 2 (end) seqno 16: DTMF 2 (end) seqno 17: DTMF 2 (end) In this second situation, the beginning of the DTMF 2 arrives before the final end frame of the DTMF 1. With the patch, seqno 12 is no processed and thus we properly interpret the DTMF. Fix 2. The second change in place is to fix an issue like the following: seqno 1: DTMF 1 seqno 2: DTMF 1 seqno 3: DTMF 1 (end) *packet lost* seqno 4: DTMF 1 (end) *packet lost* seqno 5: DTMF 1 (end) *packet lost* seqno 6: DTMF 2 When we receive seqno 6, we had code in place that was supposed to properly end the previously unended DTMF 1. The problem was that the code was essentially a no-op. The code would set up an end frame for the DTMF 1 but would immediately overwrite the frame with the begin for DTMF 2. I changed process_dtmf_rfc2833() so that instead of returning a single frame, it is given as an output parameter a list of frames. Each frame that needs to be returned is appended to this list. Fix 3. The final change is a minor one where an AST_CONTROL_SRCCHANGE frame could get lost. If we process a cisco DTMF or an RFC 3389 frame and no frame was returned, then we would return &ast_null_frame. The problem is that earlier in the function, we may have generated an AST_CONTROL_SRCCHANGE frame and put it in the list of frames we wish to return. This frame would be lost in such a case. The patch fixes this problem ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@254454 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2010-03-25 16:04:48 +00:00
AST_LIST_INSERT_TAIL(frames, f, frame_list);
}
rtp->dtmf_timeout = timestamp + rtp->dtmf_duration + dtmftimeout;
}
Handle extremely out of order RFC 2833 DTMF The current implementation of RFC 2833 DTMF handling in res_rtp_asterisk will, if a packet arrives out of order, drop the packet. This is to prevent duplicate ton generation in the Asterisk core. Since the RTP layer does not buffer data itself, this is the only option the RTP layer currently has for handling packets that arrive out of order. For the most part, this doesn't matter. For a particular digit, so long as a BEGIN packet arrives before the first END packet, the digit will be produced. If subsequent BEGIN packets arrive interleaved with the ENDs, they will be dropped; likewise, if the BEGIN or END packets themselves are out of order, those packets are dropped but sufficient information is conveyed to the Asterisk core to produce the appropriate digit. For certain sequences of DTMF packets - most notably when, for a particular digit, an END packet arrives before any BEGIN packet for that digit - this is a real problem. When an END arrives before any BEGINs, the END packet is dropped - but at the same time, it causes subsequent BEGIN packets for that digit to be ignored. When the next in order END packet arrives, it too is dropped - Asterisk believes that there was no initial BEGIN. The solution this patch provides is to trust the END packet to convey the information needed for the Asterisk core to produce the DTMF digit. If we receive an END packet, and it: * Has a timestamp greater then the last timestamp received from an END packet * Does not have the same sequence number as the last received sequence number (and is thus not an END packet retransmission) Then we send the END frame up to the Asterisk core. It contains enough DTMF information for Asterisk to produce the digit. On the other hand, if we receive a BEGIN or continuation packet that occurs with a timestamp equal to or less then the last END timestamp, then we've received something out of order - but we already have received enough information to produce the digit. These packets are dropped. Much thanks goes to Olle Johansson (oej) for providing the idea for this solution. Review: https://reviewboard.asterisk.org/r/2033/ (closes issue ASTERISK-18404) Reported by: Stephane Chazelas Tested by: Matt Jordan ........ Merged revisions 370252 from http://svn.asterisk.org/svn/asterisk/branches/1.8 ........ Merged revisions 370271 from http://svn.asterisk.org/svn/asterisk/branches/10 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@370272 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2012-07-19 21:45:20 +00:00
rtp->last_seqno = seqno;
}
rtp->dtmfsamples = samples;
Recorded merge of revisions 254452 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r254452 | mmichelson | 2010-03-25 10:59:56 -0500 (Thu, 25 Mar 2010) | 44 lines Several fixes regarding RFC2833 DTMF detection. Here is a copy and paste of the details from my request on reviewboard that dealt with these changes: Fix 1. The first change in place is to fix Mantis issue 15811, which deals with a situation where Asterisk will incorrectly interpret out of order RFC2833 frames as duplicate DTMF digits. For instance, we would receive a sequence like: seqno 1: DTMF 1 seqno 2: DTMF 1 seqno 3: DTMF 1 seqno 4: DTMF 1 seqno 6: DTMF 1 (end) seqno 5: DTMF 1 seqno 7: DTMF 1 (end) seqno 8: DTMF 1 (end) Prior to this patch when we received the frame with seqno 5, we would interpret this as a new DTMF 1. With this patch, we will check the seqno of the incoming digit and not process the frame if the seqno is lower than the last recorded seqno. Note that we do not record the seqno of the dropped DTMF frame for future processing. While the above situation is what was designed to be fixed, the patch is written in such a way that the following would also be fixed too: seqno 9: DTMF 1 seqno 10: DTMF 1 (end) seqno 11: DTMF 1 (end) seqno 13: DTMF 2 seqno 12: DTMF 1 (end) seqno 14: DTMF 2 seqno 15: DTMF 2 (end) seqno 16: DTMF 2 (end) seqno 17: DTMF 2 (end) In this second situation, the beginning of the DTMF 2 arrives before the final end frame of the DTMF 1. With the patch, seqno 12 is no processed and thus we properly interpret the DTMF. Fix 2. The second change in place is to fix an issue like the following: seqno 1: DTMF 1 seqno 2: DTMF 1 seqno 3: DTMF 1 (end) *packet lost* seqno 4: DTMF 1 (end) *packet lost* seqno 5: DTMF 1 (end) *packet lost* seqno 6: DTMF 2 When we receive seqno 6, we had code in place that was supposed to properly end the previously unended DTMF 1. The problem was that the code was essentially a no-op. The code would set up an end frame for the DTMF 1 but would immediately overwrite the frame with the begin for DTMF 2. I changed process_dtmf_rfc2833() so that instead of returning a single frame, it is given as an output parameter a list of frames. Each frame that needs to be returned is appended to this list. Fix 3. The final change is a minor one where an AST_CONTROL_SRCCHANGE frame could get lost. If we process a cisco DTMF or an RFC 3389 frame and no frame was returned, then we would return &ast_null_frame. The problem is that earlier in the function, we may have generated an AST_CONTROL_SRCCHANGE frame and put it in the list of frames we wish to return. This frame would be lost in such a case. The patch fixes this problem ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@254454 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2010-03-25 16:04:48 +00:00
return;
}
2018-05-10 18:11:06 +00:00
static struct ast_frame *process_dtmf_cisco(struct ast_rtp_instance *instance, unsigned char *data, int len, unsigned int seqno, unsigned int timestamp, int payloadtype, int mark)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
unsigned int event, flags, power;
char resp = 0;
unsigned char seq;
struct ast_frame *f = NULL;
if (len < 4) {
return NULL;
}
/* The format of Cisco RTP DTMF packet looks like next:
+0 - sequence number of DTMF RTP packet (begins from 1,
wrapped to 0)
+1 - set of flags
+1 (bit 0) - flaps by different DTMF digits delimited by audio
or repeated digit without audio???
+2 (+4,+6,...) - power level? (rises from 0 to 32 at begin of tone
then falls to 0 at its end)
+3 (+5,+7,...) - detected DTMF digit (0..9,*,#,A-D,...)
Repeated DTMF information (bytes 4/5, 6/7) is history shifted right
by each new packet and thus provides some redundancy.
Sample of Cisco RTP DTMF packet is (all data in hex):
19 07 00 02 12 02 20 02
showing end of DTMF digit '2'.
The packets
27 07 00 02 0A 02 20 02
28 06 20 02 00 02 0A 02
shows begin of new digit '2' with very short pause (20 ms) after
previous digit '2'. Bit +1.0 flips at begin of new digit.
Cisco RTP DTMF packets comes as replacement of audio RTP packets
so its uses the same sequencing and timestamping rules as replaced
audio packets. Repeat interval of DTMF packets is 20 ms and not rely
on audio framing parameters. Marker bit isn't used within stream of
DTMFs nor audio stream coming immediately after DTMF stream. Timestamps
are not sequential at borders between DTMF and audio streams,
*/
seq = data[0];
flags = data[1];
power = data[2];
event = data[3] & 0x1f;
if (ast_debug_rtp_packet_is_allowed)
ast_debug(0, "Cisco DTMF Digit: %02x (len=%d, seq=%d, flags=%02x, power=%u, history count=%d)\n", event, len, seq, flags, power, (len - 4) / 2);
if (event < 10) {
resp = '0' + event;
} else if (event < 11) {
resp = '*';
} else if (event < 12) {
resp = '#';
} else if (event < 16) {
resp = 'A' + (event - 12);
} else if (event < 17) {
resp = 'X';
}
if ((!rtp->resp && power) || (rtp->resp && (rtp->resp != resp))) {
rtp->resp = resp;
/* Why we should care on DTMF compensation at reception? */
if (ast_rtp_instance_get_prop(instance, AST_RTP_PROPERTY_DTMF_COMPENSATE)) {
Recorded merge of revisions 254452 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r254452 | mmichelson | 2010-03-25 10:59:56 -0500 (Thu, 25 Mar 2010) | 44 lines Several fixes regarding RFC2833 DTMF detection. Here is a copy and paste of the details from my request on reviewboard that dealt with these changes: Fix 1. The first change in place is to fix Mantis issue 15811, which deals with a situation where Asterisk will incorrectly interpret out of order RFC2833 frames as duplicate DTMF digits. For instance, we would receive a sequence like: seqno 1: DTMF 1 seqno 2: DTMF 1 seqno 3: DTMF 1 seqno 4: DTMF 1 seqno 6: DTMF 1 (end) seqno 5: DTMF 1 seqno 7: DTMF 1 (end) seqno 8: DTMF 1 (end) Prior to this patch when we received the frame with seqno 5, we would interpret this as a new DTMF 1. With this patch, we will check the seqno of the incoming digit and not process the frame if the seqno is lower than the last recorded seqno. Note that we do not record the seqno of the dropped DTMF frame for future processing. While the above situation is what was designed to be fixed, the patch is written in such a way that the following would also be fixed too: seqno 9: DTMF 1 seqno 10: DTMF 1 (end) seqno 11: DTMF 1 (end) seqno 13: DTMF 2 seqno 12: DTMF 1 (end) seqno 14: DTMF 2 seqno 15: DTMF 2 (end) seqno 16: DTMF 2 (end) seqno 17: DTMF 2 (end) In this second situation, the beginning of the DTMF 2 arrives before the final end frame of the DTMF 1. With the patch, seqno 12 is no processed and thus we properly interpret the DTMF. Fix 2. The second change in place is to fix an issue like the following: seqno 1: DTMF 1 seqno 2: DTMF 1 seqno 3: DTMF 1 (end) *packet lost* seqno 4: DTMF 1 (end) *packet lost* seqno 5: DTMF 1 (end) *packet lost* seqno 6: DTMF 2 When we receive seqno 6, we had code in place that was supposed to properly end the previously unended DTMF 1. The problem was that the code was essentially a no-op. The code would set up an end frame for the DTMF 1 but would immediately overwrite the frame with the begin for DTMF 2. I changed process_dtmf_rfc2833() so that instead of returning a single frame, it is given as an output parameter a list of frames. Each frame that needs to be returned is appended to this list. Fix 3. The final change is a minor one where an AST_CONTROL_SRCCHANGE frame could get lost. If we process a cisco DTMF or an RFC 3389 frame and no frame was returned, then we would return &ast_null_frame. The problem is that earlier in the function, we may have generated an AST_CONTROL_SRCCHANGE frame and put it in the list of frames we wish to return. This frame would be lost in such a case. The patch fixes this problem ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@254454 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2010-03-25 16:04:48 +00:00
f = create_dtmf_frame(instance, AST_FRAME_DTMF_BEGIN, 0);
rtp->dtmfsamples = 0;
}
} else if ((rtp->resp == resp) && !power) {
Recorded merge of revisions 254452 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r254452 | mmichelson | 2010-03-25 10:59:56 -0500 (Thu, 25 Mar 2010) | 44 lines Several fixes regarding RFC2833 DTMF detection. Here is a copy and paste of the details from my request on reviewboard that dealt with these changes: Fix 1. The first change in place is to fix Mantis issue 15811, which deals with a situation where Asterisk will incorrectly interpret out of order RFC2833 frames as duplicate DTMF digits. For instance, we would receive a sequence like: seqno 1: DTMF 1 seqno 2: DTMF 1 seqno 3: DTMF 1 seqno 4: DTMF 1 seqno 6: DTMF 1 (end) seqno 5: DTMF 1 seqno 7: DTMF 1 (end) seqno 8: DTMF 1 (end) Prior to this patch when we received the frame with seqno 5, we would interpret this as a new DTMF 1. With this patch, we will check the seqno of the incoming digit and not process the frame if the seqno is lower than the last recorded seqno. Note that we do not record the seqno of the dropped DTMF frame for future processing. While the above situation is what was designed to be fixed, the patch is written in such a way that the following would also be fixed too: seqno 9: DTMF 1 seqno 10: DTMF 1 (end) seqno 11: DTMF 1 (end) seqno 13: DTMF 2 seqno 12: DTMF 1 (end) seqno 14: DTMF 2 seqno 15: DTMF 2 (end) seqno 16: DTMF 2 (end) seqno 17: DTMF 2 (end) In this second situation, the beginning of the DTMF 2 arrives before the final end frame of the DTMF 1. With the patch, seqno 12 is no processed and thus we properly interpret the DTMF. Fix 2. The second change in place is to fix an issue like the following: seqno 1: DTMF 1 seqno 2: DTMF 1 seqno 3: DTMF 1 (end) *packet lost* seqno 4: DTMF 1 (end) *packet lost* seqno 5: DTMF 1 (end) *packet lost* seqno 6: DTMF 2 When we receive seqno 6, we had code in place that was supposed to properly end the previously unended DTMF 1. The problem was that the code was essentially a no-op. The code would set up an end frame for the DTMF 1 but would immediately overwrite the frame with the begin for DTMF 2. I changed process_dtmf_rfc2833() so that instead of returning a single frame, it is given as an output parameter a list of frames. Each frame that needs to be returned is appended to this list. Fix 3. The final change is a minor one where an AST_CONTROL_SRCCHANGE frame could get lost. If we process a cisco DTMF or an RFC 3389 frame and no frame was returned, then we would return &ast_null_frame. The problem is that earlier in the function, we may have generated an AST_CONTROL_SRCCHANGE frame and put it in the list of frames we wish to return. This frame would be lost in such a case. The patch fixes this problem ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@254454 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2010-03-25 16:04:48 +00:00
f = create_dtmf_frame(instance, AST_FRAME_DTMF_END, ast_rtp_instance_get_prop(instance, AST_RTP_PROPERTY_DTMF_COMPENSATE));
f->samples = rtp->dtmfsamples * (ast_rtp_get_rate(rtp->lastrxformat) / 1000);
rtp->resp = 0;
} else if (rtp->resp == resp) {
rtp->dtmfsamples += 20 * (ast_rtp_get_rate(rtp->lastrxformat) / 1000);
}
rtp->dtmf_timeout = 0;
return f;
}
2018-05-10 18:11:06 +00:00
static struct ast_frame *process_cn_rfc3389(struct ast_rtp_instance *instance, unsigned char *data, int len, unsigned int seqno, unsigned int timestamp, int payloadtype, int mark)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
/* Convert comfort noise into audio with various codecs. Unfortunately this doesn't
totally help us out because we don't have an engine to keep it going and we are not
guaranteed to have it every 20ms or anything */
if (ast_debug_rtp_packet_is_allowed) {
media formats: re-architect handling of media for performance improvements In the old times media formats were represented using a bit field. This was fast but had a few limitations. 1. Asterisk was limited in how many formats it could handle. 2. Formats, being a bit field, could not include any attribute information. A format was strictly its type, e.g., "this is ulaw". This was changed in Asterisk 10 (see https://wiki.asterisk.org/wiki/display/AST/Media+Architecture+Proposal for notes on that work) which led to the creation of the ast_format structure. This structure allowed Asterisk to handle attributes and bundle information with a format. Additionally, ast_format_cap was created to act as a container for multiple formats that, together, formed the capability of some entity. Another mechanism was added to allow logic to be registered which performed format attribute negotiation. Everywhere throughout the codebase Asterisk was changed to use this strategy. Unfortunately, in software, there is no free lunch. These new capabilities came at a cost. Performance analysis and profiling showed that we spend an inordinate amount of time comparing, copying, and generally manipulating formats and their related structures. Basic prototyping has shown that a reasonably large performance improvement could be made in this area. This patch is the result of that project, which overhauled the media format architecture and its usage in Asterisk to improve performance. Generally, the new philosophy for handling formats is as follows: * The ast_format structure is reference counted. This removed a large amount of the memory allocations and copying that was done in prior versions. * In order to prevent race conditions while keeping things performant, the ast_format structure is immutable by convention and lock-free. Violate this tenet at your peril! * Because formats are reference counted, codecs are also reference counted. The Asterisk core generally provides built-in codecs and caches the ast_format structures created to represent them. Generally, to prevent inordinate amounts of module reference bumping, codecs and formats can be added at run-time but cannot be removed. * All compatibility with the bit field representation of codecs/formats has been moved to a compatibility API. The primary user of this representation is chan_iax2, which must continue to maintain its bit-field usage of formats for interoperability concerns. * When a format is negotiated with attributes, or when a format cannot be represented by one of the cached formats, a new format object is created or cloned from an existing format. That format may have the same codec underlying it, but is a different format than a version of the format with different attributes or without attributes. * While formats are reference counted objects, the reference count maintained on the format should be manipulated with care. Formats are generally cached and will persist for the lifetime of Asterisk and do not explicitly need to have their lifetime modified. An exception to this is when the user of a format does not know where the format came from *and* the user may outlive the provider of the format. This occurs, for example, when a format is read from a channel: the channel may have a format with attributes (hence, non-cached) and the user of the format may last longer than the channel (if the reference to the channel is released prior to the format's reference). For more information on this work, see the API design notes: https://wiki.asterisk.org/wiki/display/AST/Media+Format+Rewrite Finally, this work was the culmination of a large number of developer's efforts. Extra thanks goes to Corey Farrell, who took on a large amount of the work in the Asterisk core, chan_sip, and was an invaluable resource in peer reviews throughout this project. There were a substantial number of patches contributed during this work; the following issues/patch names simply reflect some of the work (and will cause the release scripts to give attribution to the individuals who work on them). Reviews: https://reviewboard.asterisk.org/r/3814 https://reviewboard.asterisk.org/r/3808 https://reviewboard.asterisk.org/r/3805 https://reviewboard.asterisk.org/r/3803 https://reviewboard.asterisk.org/r/3801 https://reviewboard.asterisk.org/r/3798 https://reviewboard.asterisk.org/r/3800 https://reviewboard.asterisk.org/r/3794 https://reviewboard.asterisk.org/r/3793 https://reviewboard.asterisk.org/r/3792 https://reviewboard.asterisk.org/r/3791 https://reviewboard.asterisk.org/r/3790 https://reviewboard.asterisk.org/r/3789 https://reviewboard.asterisk.org/r/3788 https://reviewboard.asterisk.org/r/3787 https://reviewboard.asterisk.org/r/3786 https://reviewboard.asterisk.org/r/3784 https://reviewboard.asterisk.org/r/3783 https://reviewboard.asterisk.org/r/3778 https://reviewboard.asterisk.org/r/3774 https://reviewboard.asterisk.org/r/3775 https://reviewboard.asterisk.org/r/3772 https://reviewboard.asterisk.org/r/3761 https://reviewboard.asterisk.org/r/3754 https://reviewboard.asterisk.org/r/3753 https://reviewboard.asterisk.org/r/3751 https://reviewboard.asterisk.org/r/3750 https://reviewboard.asterisk.org/r/3748 https://reviewboard.asterisk.org/r/3747 https://reviewboard.asterisk.org/r/3746 https://reviewboard.asterisk.org/r/3742 https://reviewboard.asterisk.org/r/3740 https://reviewboard.asterisk.org/r/3739 https://reviewboard.asterisk.org/r/3738 https://reviewboard.asterisk.org/r/3737 https://reviewboard.asterisk.org/r/3736 https://reviewboard.asterisk.org/r/3734 https://reviewboard.asterisk.org/r/3722 https://reviewboard.asterisk.org/r/3713 https://reviewboard.asterisk.org/r/3703 https://reviewboard.asterisk.org/r/3689 https://reviewboard.asterisk.org/r/3687 https://reviewboard.asterisk.org/r/3674 https://reviewboard.asterisk.org/r/3671 https://reviewboard.asterisk.org/r/3667 https://reviewboard.asterisk.org/r/3665 https://reviewboard.asterisk.org/r/3625 https://reviewboard.asterisk.org/r/3602 https://reviewboard.asterisk.org/r/3519 https://reviewboard.asterisk.org/r/3518 https://reviewboard.asterisk.org/r/3516 https://reviewboard.asterisk.org/r/3515 https://reviewboard.asterisk.org/r/3512 https://reviewboard.asterisk.org/r/3506 https://reviewboard.asterisk.org/r/3413 https://reviewboard.asterisk.org/r/3410 https://reviewboard.asterisk.org/r/3387 https://reviewboard.asterisk.org/r/3388 https://reviewboard.asterisk.org/r/3389 https://reviewboard.asterisk.org/r/3390 https://reviewboard.asterisk.org/r/3321 https://reviewboard.asterisk.org/r/3320 https://reviewboard.asterisk.org/r/3319 https://reviewboard.asterisk.org/r/3318 https://reviewboard.asterisk.org/r/3266 https://reviewboard.asterisk.org/r/3265 https://reviewboard.asterisk.org/r/3234 https://reviewboard.asterisk.org/r/3178 ASTERISK-23114 #close Reported by: mjordan media_formats_translation_core.diff uploaded by kharwell (License 6464) rb3506.diff uploaded by mjordan (License 6283) media_format_app_file.diff uploaded by kharwell (License 6464) misc-2.diff uploaded by file (License 5000) chan_mild-3.diff uploaded by file (License 5000) chan_obscure.diff uploaded by file (License 5000) jingle.diff uploaded by file (License 5000) funcs.diff uploaded by file (License 5000) formats.diff uploaded by file (License 5000) core.diff uploaded by file (License 5000) bridges.diff uploaded by file (License 5000) mf-codecs-2.diff uploaded by file (License 5000) mf-app_fax.diff uploaded by file (License 5000) mf-apps-3.diff uploaded by file (License 5000) media-formats-3.diff uploaded by file (License 5000) ASTERISK-23715 rb3713.patch uploaded by coreyfarrell (License 5909) rb3689.patch uploaded by mjordan (License 6283) ASTERISK-23957 rb3722.patch uploaded by mjordan (License 6283) mf-attributes-3.diff uploaded by file (License 5000) ASTERISK-23958 Tested by: jrose rb3822.patch uploaded by coreyfarrell (License 5909) rb3800.patch uploaded by jrose (License 6182) chan_sip.diff uploaded by mjordan (License 6283) rb3747.patch uploaded by jrose (License 6182) ASTERISK-23959 #close Tested by: sgriepentrog, mjordan, coreyfarrell sip_cleanup.diff uploaded by opticron (License 6273) chan_sip_caps.diff uploaded by mjordan (License 6283) rb3751.patch uploaded by coreyfarrell (License 5909) chan_sip-3.diff uploaded by file (License 5000) ASTERISK-23960 #close Tested by: opticron direct_media.diff uploaded by opticron (License 6273) pjsip-direct-media.diff uploaded by file (License 5000) format_cap_remove.diff uploaded by opticron (License 6273) media_format_fixes.diff uploaded by opticron (License 6273) chan_pjsip-2.diff uploaded by file (License 5000) ASTERISK-23966 #close Tested by: rmudgett rb3803.patch uploaded by rmudgetti (License 5621) chan_dahdi.diff uploaded by file (License 5000) ASTERISK-24064 #close Tested by: coreyfarrell, mjordan, opticron, file, rmudgett, sgriepentrog, jrose rb3814.patch uploaded by rmudgett (License 5621) moh_cleanup.diff uploaded by opticron (License 6273) bridge_leak.diff uploaded by opticron (License 6273) translate.diff uploaded by file (License 5000) rb3795.patch uploaded by rmudgett (License 5621) tls_fix.diff uploaded by mjordan (License 6283) fax-mf-fix-2.diff uploaded by file (License 5000) rtp_transfer_stuff uploaded by mjordan (License 6283) rb3787.patch uploaded by rmudgett (License 5621) media-formats-explicit-translate-format-3.diff uploaded by file (License 5000) format_cache_case_fix.diff uploaded by opticron (License 6273) rb3774.patch uploaded by rmudgett (License 5621) rb3775.patch uploaded by rmudgett (License 5621) rtp_engine_fix.diff uploaded by opticron (License 6273) rtp_crash_fix.diff uploaded by opticron (License 6273) rb3753.patch uploaded by mjordan (License 6283) rb3750.patch uploaded by mjordan (License 6283) rb3748.patch uploaded by rmudgett (License 5621) media_format_fixes.diff uploaded by opticron (License 6273) rb3740.patch uploaded by mjordan (License 6283) rb3739.patch uploaded by mjordan (License 6283) rb3734.patch uploaded by mjordan (License 6283) rb3689.patch uploaded by mjordan (License 6283) rb3674.patch uploaded by coreyfarrell (License 5909) rb3671.patch uploaded by coreyfarrell (License 5909) rb3667.patch uploaded by coreyfarrell (License 5909) rb3665.patch uploaded by mjordan (License 6283) rb3625.patch uploaded by coreyfarrell (License 5909) rb3602.patch uploaded by coreyfarrell (License 5909) format_compatibility-2.diff uploaded by file (License 5000) core.diff uploaded by file (License 5000) git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@419044 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-07-20 22:06:33 +00:00
ast_debug(0, "- RTP 3389 Comfort noise event: Format %s (len = %d)\n",
ast_format_get_name(rtp->lastrxformat), len);
}
if (!ast_test_flag(rtp, FLAG_3389_WARNING)) {
struct ast_sockaddr remote_address = { {0,} };
ast_rtp_instance_get_remote_address(instance, &remote_address);
ast_log(LOG_NOTICE, "Comfort noise support incomplete in Asterisk (RFC 3389). Please turn off on client if possible. Client address: %s\n",
ast_sockaddr_stringify(&remote_address));
ast_set_flag(rtp, FLAG_3389_WARNING);
}
/* Must have at least one byte */
if (!len) {
return NULL;
}
if (len < 24) {
rtp->f.data.ptr = rtp->rawdata + AST_FRIENDLY_OFFSET;
rtp->f.datalen = len - 1;
rtp->f.offset = AST_FRIENDLY_OFFSET;
memcpy(rtp->f.data.ptr, data + 1, len - 1);
} else {
rtp->f.data.ptr = NULL;
rtp->f.offset = 0;
rtp->f.datalen = 0;
}
rtp->f.frametype = AST_FRAME_CNG;
rtp->f.subclass.integer = data[0] & 0x7f;
rtp->f.samples = 0;
rtp->f.delivery.tv_usec = rtp->f.delivery.tv_sec = 0;
return &rtp->f;
}
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
static int update_rtt_stats(struct ast_rtp *rtp, unsigned int lsr, unsigned int dlsr)
{
struct timeval now;
struct timeval rtt_tv;
unsigned int msw;
unsigned int lsw;
unsigned int rtt_msw;
unsigned int rtt_lsw;
unsigned int lsr_a;
unsigned int rtt;
gettimeofday(&now, NULL);
timeval2ntp(now, &msw, &lsw);
lsr_a = ((msw & 0x0000ffff) << 16) | ((lsw & 0xffff0000) >> 16);
rtt = lsr_a - lsr - dlsr;
rtt_msw = (rtt & 0xffff0000) >> 16;
rtt_lsw = (rtt & 0x0000ffff);
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
rtt_tv.tv_sec = rtt_msw;
/*
* Convert 16.16 fixed point rtt_lsw to usec without
* overflow.
*
* = rtt_lsw * 10^6 / 2^16
* = rtt_lsw * (2^6 * 5^6) / 2^16
* = rtt_lsw * 5^6 / 2^10
*
* The rtt_lsw value is in 16.16 fixed point format and 5^6
* requires 14 bits to represent. We have enough space to
* directly do the conversion because there is no integer
* component in rtt_lsw.
*/
rtt_tv.tv_usec = (rtt_lsw * 15625) >> 10;
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
rtp->rtcp->rtt = (double)rtt_tv.tv_sec + ((double)rtt_tv.tv_usec / 1000000);
if (lsr_a - dlsr < lsr) {
return 1;
}
rtp->rtcp->accumulated_transit += rtp->rtcp->rtt;
if (rtp->rtcp->rtt_count == 0 || rtp->rtcp->minrtt > rtp->rtcp->rtt) {
rtp->rtcp->minrtt = rtp->rtcp->rtt;
}
if (rtp->rtcp->maxrtt < rtp->rtcp->rtt) {
rtp->rtcp->maxrtt = rtp->rtcp->rtt;
}
calc_mean_and_standard_deviation(rtp->rtcp->rtt, &rtp->rtcp->normdevrtt,
&rtp->rtcp->stdevrtt, &rtp->rtcp->rtt_count);
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
return 0;
}
/*!
* \internal
* \brief Update RTCP interarrival jitter stats
*/
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
static void update_jitter_stats(struct ast_rtp *rtp, unsigned int ia_jitter)
{
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
int rate = ast_rtp_get_rate(rtp->f.subclass.format);
rtp->rtcp->reported_jitter = ast_samp2sec(ia_jitter, rate);
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
if (rtp->rtcp->reported_jitter_count == 0) {
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
rtp->rtcp->reported_minjitter = rtp->rtcp->reported_jitter;
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
}
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
if (rtp->rtcp->reported_jitter < rtp->rtcp->reported_minjitter) {
rtp->rtcp->reported_minjitter = rtp->rtcp->reported_jitter;
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
}
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
if (rtp->rtcp->reported_jitter > rtp->rtcp->reported_maxjitter) {
rtp->rtcp->reported_maxjitter = rtp->rtcp->reported_jitter;
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
}
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
calc_mean_and_standard_deviation(rtp->rtcp->reported_jitter,
&rtp->rtcp->reported_normdev_jitter, &rtp->rtcp->reported_stdev_jitter,
&rtp->rtcp->reported_jitter_count);
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
}
/*!
* \internal
* \brief Update RTCP lost packet stats
*/
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
static void update_lost_stats(struct ast_rtp *rtp, unsigned int lost_packets)
{
double reported_lost;
rtp->rtcp->reported_lost = lost_packets;
reported_lost = (double)rtp->rtcp->reported_lost;
if (rtp->rtcp->reported_lost_count == 0) {
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
rtp->rtcp->reported_minlost = reported_lost;
}
if (reported_lost < rtp->rtcp->reported_minlost) {
rtp->rtcp->reported_minlost = reported_lost;
}
if (reported_lost > rtp->rtcp->reported_maxlost) {
rtp->rtcp->reported_maxlost = reported_lost;
}
calc_mean_and_standard_deviation(reported_lost, &rtp->rtcp->reported_normdev_lost,
&rtp->rtcp->reported_stdev_lost, &rtp->rtcp->reported_lost_count);
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
}
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
#define RESCALE(in, inmin, inmax, outmin, outmax) ((((in - inmin)/(inmax-inmin))*(outmax-outmin))+outmin)
/*!
* \brief Calculate a "media experience score" based on given data
*
* Technically, a mean opinion score (MOS) cannot be calculated without the involvement
* of human eyes (video) and ears (audio). Thus instead we'll approximate an opinion
* using the given parameters, and call it a media experience score.
*
* The tallied score is based upon recommendations and formulas from ITU-T G.107,
* ITU-T G.109, ITU-T G.113, and other various internet sources.
*
* \param instance RTP instance
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
* \param normdevrtt The average round trip time
* \param normdev_rxjitter The smoothed jitter
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
* \param stdev_rxjitter The jitter standard deviation value
* \param normdev_rxlost The average number of packets lost since last check
*
* \return A media experience score.
*
* \note The calculations in this function could probably be simplified
* but calculating a MOS using the information available publicly,
* then re-scaling it to 0.0 -> 100.0 makes the process clearer and
* easier to troubleshoot or change.
*/
static double calc_media_experience_score(struct ast_rtp_instance *instance,
double normdevrtt, double normdev_rxjitter, double stdev_rxjitter,
double normdev_rxlost)
{
double r_value;
double pseudo_mos;
double mes = 0;
/*
* While the media itself might be okay, a significant enough delay could make
* for an unpleasant user experience.
*
* Calculate the effective latency by using the given round trip time, and adding
* jitter scaled according to its standard deviation. The scaling is done in order
* to increase jitter's weight since a higher deviation can result in poorer overall
* quality.
*/
double effective_latency = (normdevrtt * 1000)
+ ((normdev_rxjitter * 2) * (stdev_rxjitter / 3))
+ 10;
/*
* Using the defaults for the standard transmission rating factor ("R" value)
* one arrives at 93.2 (see ITU-T G.107 for more details), so we'll use that
* as the starting value and subtract deficiencies that could affect quality.
*
* Calculate the impact of the effective latency. Influence increases with
* values over 160 as the significant "lag" can degrade user experience.
*/
if (effective_latency < 160) {
r_value = 93.2 - (effective_latency / 40);
} else {
r_value = 93.2 - (effective_latency - 120) / 10;
}
/* Next evaluate the impact of lost packets */
r_value = r_value - (normdev_rxlost * 2.0);
/*
* Finally convert the "R" value into a opinion/quality score between 1 (really anything
* below 3 should be considered poor) and 4.5 (the highest achievable for VOIP).
*/
if (r_value < 0) {
pseudo_mos = 1.0;
} else if (r_value > 100) {
pseudo_mos = 4.5;
} else {
pseudo_mos = 1 + (0.035 * r_value) + (r_value * (r_value - 60) * (100 - r_value) * 0.0000007);
}
/*
* We're going to rescale the 0.0->5.0 pseudo_mos to the 0.0->100.0 MES.
* For those ranges, we could actually just multiply the pseudo_mos
* by 20 but we may want to change the scale later.
*/
mes = RESCALE(pseudo_mos, 0.0, 5.0, 0.0, 100.0);
return mes;
}
/*!
* \internal
* \brief Update MES stats based on info received in an SR or RR.
* This is RTP we sent and they received.
*/
static void update_reported_mes_stats(struct ast_rtp *rtp)
{
double mes = calc_media_experience_score(rtp->owner,
rtp->rtcp->normdevrtt,
rtp->rtcp->reported_jitter,
rtp->rtcp->reported_stdev_jitter,
rtp->rtcp->reported_normdev_lost);
rtp->rtcp->reported_mes = mes;
if (rtp->rtcp->reported_mes_count == 0) {
rtp->rtcp->reported_minmes = mes;
}
if (mes < rtp->rtcp->reported_minmes) {
rtp->rtcp->reported_minmes = mes;
}
if (mes > rtp->rtcp->reported_maxmes) {
rtp->rtcp->reported_maxmes = mes;
}
calc_mean_and_standard_deviation(mes, &rtp->rtcp->reported_normdev_mes,
&rtp->rtcp->reported_stdev_mes, &rtp->rtcp->reported_mes_count);
ast_debug_rtcp(2, "%s: rtt: %.9f j: %.9f sjh: %.9f lost: %.9f mes: %4.1f\n",
ast_rtp_instance_get_channel_id(rtp->owner),
rtp->rtcp->normdevrtt,
rtp->rtcp->reported_jitter,
rtp->rtcp->reported_stdev_jitter,
rtp->rtcp->reported_normdev_lost, mes);
}
/*!
* \internal
* \brief Update MES stats based on info we will send in an SR or RR.
* This is RTP they sent and we received.
*/
static void update_local_mes_stats(struct ast_rtp *rtp)
{
rtp->rxmes = calc_media_experience_score(rtp->owner,
rtp->rtcp->normdevrtt,
rtp->rxjitter,
rtp->rtcp->stdev_rxjitter,
rtp->rtcp->normdev_rxlost);
if (rtp->rtcp->rxmes_count == 0) {
rtp->rtcp->minrxmes = rtp->rxmes;
}
if (rtp->rxmes < rtp->rtcp->minrxmes) {
rtp->rtcp->minrxmes = rtp->rxmes;
}
if (rtp->rxmes > rtp->rtcp->maxrxmes) {
rtp->rtcp->maxrxmes = rtp->rxmes;
}
calc_mean_and_standard_deviation(rtp->rxmes, &rtp->rtcp->normdev_rxmes,
&rtp->rtcp->stdev_rxmes, &rtp->rtcp->rxmes_count);
ast_debug_rtcp(2, " %s: rtt: %.9f j: %.9f sjh: %.9f lost: %.9f mes: %4.1f\n",
ast_rtp_instance_get_channel_id(rtp->owner),
rtp->rtcp->normdevrtt,
rtp->rxjitter,
rtp->rtcp->stdev_rxjitter,
rtp->rtcp->normdev_rxlost, rtp->rxmes);
}
/*! \pre instance is locked */
static struct ast_rtp_instance *__rtp_find_instance_by_ssrc(struct ast_rtp_instance *instance,
struct ast_rtp *rtp, unsigned int ssrc, int source)
{
int index;
if (!AST_VECTOR_SIZE(&rtp->ssrc_mapping)) {
/* This instance is not bundled */
return instance;
}
/* Find the bundled child instance */
for (index = 0; index < AST_VECTOR_SIZE(&rtp->ssrc_mapping); ++index) {
struct rtp_ssrc_mapping *mapping = AST_VECTOR_GET_ADDR(&rtp->ssrc_mapping, index);
unsigned int mapping_ssrc = source ? ast_rtp_get_ssrc(mapping->instance) : mapping->ssrc;
if (mapping->ssrc_valid && mapping_ssrc == ssrc) {
return mapping->instance;
}
}
/* Does the SSRC match the bundled parent? */
if (rtp->themssrc_valid && rtp->themssrc == ssrc) {
return instance;
}
return NULL;
}
/*! \pre instance is locked */
static struct ast_rtp_instance *rtp_find_instance_by_packet_source_ssrc(struct ast_rtp_instance *instance,
struct ast_rtp *rtp, unsigned int ssrc)
{
return __rtp_find_instance_by_ssrc(instance, rtp, ssrc, 0);
}
/*! \pre instance is locked */
static struct ast_rtp_instance *rtp_find_instance_by_media_source_ssrc(struct ast_rtp_instance *instance,
struct ast_rtp *rtp, unsigned int ssrc)
{
return __rtp_find_instance_by_ssrc(instance, rtp, ssrc, 1);
}
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
static const char *rtcp_payload_type2str(unsigned int pt)
{
const char *str;
switch (pt) {
case RTCP_PT_SR:
str = "Sender Report";
break;
case RTCP_PT_RR:
str = "Receiver Report";
break;
case RTCP_PT_FUR:
/* Full INTRA-frame Request / Fast Update Request */
str = "H.261 FUR";
break;
case RTCP_PT_PSFB:
/* Payload Specific Feed Back */
str = "PSFB";
break;
case RTCP_PT_SDES:
str = "Source Description";
break;
case RTCP_PT_BYE:
str = "BYE";
break;
default:
str = "Unknown";
break;
}
return str;
}
static const char *rtcp_payload_subtype2str(unsigned int pt, unsigned int subtype)
{
switch (pt) {
case AST_RTP_RTCP_RTPFB:
if (subtype == AST_RTP_RTCP_FMT_NACK) {
return "NACK";
}
break;
case RTCP_PT_PSFB:
if (subtype == AST_RTP_RTCP_FMT_REMB) {
return "REMB";
}
break;
default:
break;
}
return NULL;
}
/*! \pre instance is locked */
static int ast_rtp_rtcp_handle_nack(struct ast_rtp_instance *instance, unsigned int *nackdata, unsigned int position,
unsigned int length)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
int res = 0;
int blp_index;
int packet_index;
int ice;
struct ast_rtp_rtcp_nack_payload *payload;
unsigned int current_word;
unsigned int pid; /* Packet ID which refers to seqno of lost packet */
unsigned int blp; /* Bitmask of following lost packets */
struct ast_sockaddr remote_address = { {0,} };
int abs_send_time_id;
unsigned int now_msw = 0;
unsigned int now_lsw = 0;
unsigned int packets_not_found = 0;
if (!rtp->send_buffer) {
ast_debug_rtcp(1, "(%p) RTCP tried to handle NACK request, "
"but we don't have a RTP packet storage!\n", instance);
return res;
}
abs_send_time_id = ast_rtp_instance_extmap_get_id(instance, AST_RTP_EXTENSION_ABS_SEND_TIME);
if (abs_send_time_id != -1) {
timeval2ntp(ast_tvnow(), &now_msw, &now_lsw);
}
ast_rtp_instance_get_remote_address(instance, &remote_address);
/*
* We use index 3 because with feedback messages, the FCI (Feedback Control Information)
* does not begin until after the version, packet SSRC, and media SSRC words.
*/
for (packet_index = 3; packet_index < length; packet_index++) {
current_word = ntohl(nackdata[position + packet_index]);
pid = current_word >> 16;
/* We know the remote end is missing this packet. Go ahead and send it if we still have it. */
payload = (struct ast_rtp_rtcp_nack_payload *)ast_data_buffer_get(rtp->send_buffer, pid);
if (payload) {
if (abs_send_time_id != -1) {
/* On retransmission we need to update the timestamp within the packet, as it
* is supposed to contain when the packet was actually sent.
*/
put_unaligned_time24(payload->buf + 17, now_msw, now_lsw);
}
res += rtp_sendto(instance, payload->buf, payload->size, 0, &remote_address, &ice);
} else {
ast_debug_rtcp(1, "(%p) RTCP received NACK request for RTP packet with seqno %d, "
"but we don't have it\n", instance, pid);
packets_not_found++;
}
/*
* The bitmask. Denoting the least significant bit as 1 and its most significant bit
* as 16, then bit i of the bitmask is set to 1 if the receiver has not received RTP
* packet (pid+i)(modulo 2^16). Otherwise, it is set to 0. We cannot assume bits set
* to 0 after a bit set to 1 have actually been received.
*/
blp = current_word & 0xffff;
blp_index = 1;
while (blp) {
if (blp & 1) {
/* Packet (pid + i)(modulo 2^16) is missing too. */
unsigned int seqno = (pid + blp_index) % 65536;
payload = (struct ast_rtp_rtcp_nack_payload *)ast_data_buffer_get(rtp->send_buffer, seqno);
if (payload) {
if (abs_send_time_id != -1) {
put_unaligned_time24(payload->buf + 17, now_msw, now_lsw);
}
res += rtp_sendto(instance, payload->buf, payload->size, 0, &remote_address, &ice);
} else {
ast_debug_rtcp(1, "(%p) RTCP remote end also requested RTP packet with seqno %d, "
"but we don't have it\n", instance, seqno);
packets_not_found++;
}
}
blp >>= 1;
blp_index++;
}
}
if (packets_not_found) {
/* Grow the send buffer based on how many packets were not found in the buffer, but
* enforce a maximum.
*/
ast_data_buffer_resize(rtp->send_buffer, MIN(MAXIMUM_RTP_SEND_BUFFER_SIZE,
ast_data_buffer_max(rtp->send_buffer) + packets_not_found));
ast_debug_rtcp(2, "(%p) RTCP send buffer on RTP instance is now at maximum of %zu\n",
instance, ast_data_buffer_max(rtp->send_buffer));
}
return res;
}
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
/*
* Unshifted RTCP header bit field masks
*/
#define RTCP_LENGTH_MASK 0xFFFF
#define RTCP_PAYLOAD_TYPE_MASK 0xFF
#define RTCP_REPORT_COUNT_MASK 0x1F
#define RTCP_PADDING_MASK 0x01
#define RTCP_VERSION_MASK 0x03
/*
* RTCP header bit field shift offsets
*/
#define RTCP_LENGTH_SHIFT 0
#define RTCP_PAYLOAD_TYPE_SHIFT 16
#define RTCP_REPORT_COUNT_SHIFT 24
#define RTCP_PADDING_SHIFT 29
#define RTCP_VERSION_SHIFT 30
#define RTCP_VERSION 2U
#define RTCP_VERSION_SHIFTED (RTCP_VERSION << RTCP_VERSION_SHIFT)
#define RTCP_VERSION_MASK_SHIFTED (RTCP_VERSION_MASK << RTCP_VERSION_SHIFT)
/*
* RTCP first packet record validity header mask and value.
*
* RFC3550 intentionally defines the encoding of RTCP_PT_SR and RTCP_PT_RR
* such that they differ in the least significant bit. Either of these two
* payload types MUST be the first RTCP packet record in a compound packet.
*
* RFC3550 checks the padding bit in the algorithm they use to check the
* RTCP packet for validity. However, we aren't masking the padding bit
* to check since we don't know if it is a compound RTCP packet or not.
*/
#define RTCP_VALID_MASK (RTCP_VERSION_MASK_SHIFTED | (((RTCP_PAYLOAD_TYPE_MASK & ~0x1)) << RTCP_PAYLOAD_TYPE_SHIFT))
#define RTCP_VALID_VALUE (RTCP_VERSION_SHIFTED | (RTCP_PT_SR << RTCP_PAYLOAD_TYPE_SHIFT))
#define RTCP_SR_BLOCK_WORD_LENGTH 5
#define RTCP_RR_BLOCK_WORD_LENGTH 6
#define RTCP_HEADER_SSRC_LENGTH 2
#define RTCP_FB_REMB_BLOCK_WORD_LENGTH 4
#define RTCP_FB_NACK_BLOCK_WORD_LENGTH 2
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
2018-05-10 18:11:06 +00:00
static struct ast_frame *ast_rtcp_interpret(struct ast_rtp_instance *instance, struct ast_srtp *srtp,
const unsigned char *rtcpdata, size_t size, struct ast_sockaddr *addr)
{
struct ast_rtp_instance *transport = instance;
struct ast_rtp *transport_rtp = ast_rtp_instance_get_data(instance);
2018-05-10 18:11:06 +00:00
int len = size;
unsigned int *rtcpheader = (unsigned int *)(rtcpdata);
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
unsigned int packetwords;
unsigned int position;
unsigned int first_word;
/*! True if we have seen an acceptable SSRC to learn the remote RTCP address */
unsigned int ssrc_seen;
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
struct ast_rtp_rtcp_report_block *report_block;
struct ast_frame *f = &ast_null_frame;
#ifdef TEST_FRAMEWORK
struct ast_rtp_engine_test *test_engine;
#endif
2018-05-10 18:11:06 +00:00
/* If this is encrypted then decrypt the payload */
if ((*rtcpheader & 0xC0) && res_srtp && srtp && res_srtp->unprotect(
srtp, rtcpheader, &len, 1 | (srtp_replay_protection << 1)) < 0) {
2018-05-10 18:11:06 +00:00
return &ast_null_frame;
}
packetwords = len / 4;
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
ast_debug_rtcp(2, "(%s) RTCP got report of %d bytes from %s\n",
ast_rtp_instance_get_channel_id(instance),
len, ast_sockaddr_stringify(addr));
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
/*
* Validate the RTCP packet according to an adapted and slightly
* modified RFC3550 validation algorithm.
*/
if (packetwords < RTCP_HEADER_SSRC_LENGTH) {
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
ast_debug_rtcp(2, "(%s) RTCP %p -- from %s: Frame size (%u words) is too short\n",
ast_rtp_instance_get_channel_id(instance),
transport_rtp, ast_sockaddr_stringify(addr), packetwords);
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
return &ast_null_frame;
}
position = 0;
first_word = ntohl(rtcpheader[position]);
if ((first_word & RTCP_VALID_MASK) != RTCP_VALID_VALUE) {
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
ast_debug_rtcp(2, "(%s) RTCP %p -- from %s: Failed first packet validity check\n",
ast_rtp_instance_get_channel_id(instance),
transport_rtp, ast_sockaddr_stringify(addr));
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
return &ast_null_frame;
}
do {
position += ((first_word >> RTCP_LENGTH_SHIFT) & RTCP_LENGTH_MASK) + 1;
if (packetwords <= position) {
break;
}
first_word = ntohl(rtcpheader[position]);
} while ((first_word & RTCP_VERSION_MASK_SHIFTED) == RTCP_VERSION_SHIFTED);
if (position != packetwords) {
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
ast_debug_rtcp(2, "(%s) RTCP %p -- from %s: Failed packet version or length check\n",
ast_rtp_instance_get_channel_id(instance),
transport_rtp, ast_sockaddr_stringify(addr));
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
return &ast_null_frame;
}
/*
* Note: RFC3605 points out that true NAT (vs NAPT) can cause RTCP
* to have a different IP address and port than RTP. Otherwise, when
* strictrtp is enabled we could reject RTCP packets not coming from
* the learned RTP IP address if it is available.
*/
/*
* strictrtp safety needs SSRC to match before we use the
* sender's address for symmetrical RTP to send our RTCP
* reports.
*
* If strictrtp is not enabled then claim to have already seen
* a matching SSRC so we'll accept this packet's address for
* symmetrical RTP.
*/
ssrc_seen = transport_rtp->strict_rtp_state == STRICT_RTP_OPEN;
position = 0;
while (position < packetwords) {
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
unsigned int i;
unsigned int pt;
unsigned int rc;
unsigned int ssrc;
/*! True if the ssrc value we have is valid and not garbage because it doesn't exist. */
unsigned int ssrc_valid;
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
unsigned int length;
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
unsigned int min_length;
/*! Always use packet source SSRC to find the rtp instance unless explicitly told not to. */
unsigned int use_packet_source = 1;
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
struct ast_json *message_blob;
RAII_VAR(struct ast_rtp_rtcp_report *, rtcp_report, NULL, ao2_cleanup);
struct ast_rtp_instance *child;
struct ast_rtp *rtp;
struct ast_rtp_rtcp_feedback *feedback;
i = position;
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
first_word = ntohl(rtcpheader[i]);
pt = (first_word >> RTCP_PAYLOAD_TYPE_SHIFT) & RTCP_PAYLOAD_TYPE_MASK;
rc = (first_word >> RTCP_REPORT_COUNT_SHIFT) & RTCP_REPORT_COUNT_MASK;
/* RFC3550 says 'length' is the number of words in the packet - 1 */
length = ((first_word >> RTCP_LENGTH_SHIFT) & RTCP_LENGTH_MASK) + 1;
/* Check expected RTCP packet record length */
min_length = RTCP_HEADER_SSRC_LENGTH;
switch (pt) {
case RTCP_PT_SR:
min_length += RTCP_SR_BLOCK_WORD_LENGTH;
/* fall through */
case RTCP_PT_RR:
min_length += (rc * RTCP_RR_BLOCK_WORD_LENGTH);
use_packet_source = 0;
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
break;
case RTCP_PT_FUR:
break;
case AST_RTP_RTCP_RTPFB:
switch (rc) {
case AST_RTP_RTCP_FMT_NACK:
min_length += RTCP_FB_NACK_BLOCK_WORD_LENGTH;
break;
default:
break;
}
use_packet_source = 0;
break;
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
case RTCP_PT_PSFB:
switch (rc) {
case AST_RTP_RTCP_FMT_REMB:
min_length += RTCP_FB_REMB_BLOCK_WORD_LENGTH;
break;
default:
break;
}
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
break;
case RTCP_PT_SDES:
case RTCP_PT_BYE:
/*
* There may not be a SSRC/CSRC present. The packet is
* useless but still valid if it isn't present.
*
* We don't know what min_length should be so disable the check
*/
min_length = length;
break;
default:
ast_debug_rtcp(1, "(%p) RTCP %p -- from %s: %u(%s) skipping record\n",
instance, transport_rtp, ast_sockaddr_stringify(addr), pt, rtcp_payload_type2str(pt));
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
if (rtcp_debug_test_addr(addr)) {
ast_verbose("\n");
ast_verbose("RTCP from %s: %u(%s) skipping record\n",
ast_sockaddr_stringify(addr), pt, rtcp_payload_type2str(pt));
}
position += length;
continue;
}
if (length < min_length) {
ast_debug_rtcp(1, "(%p) RTCP %p -- from %s: %u(%s) length field less than expected minimum. Min:%u Got:%u\n",
instance, transport_rtp, ast_sockaddr_stringify(addr), pt, rtcp_payload_type2str(pt),
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
min_length - 1, length - 1);
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
return &ast_null_frame;
}
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
/* Get the RTCP record SSRC if defined for the record */
ssrc_valid = 1;
switch (pt) {
case RTCP_PT_SR:
case RTCP_PT_RR:
rtcp_report = ast_rtp_rtcp_report_alloc(rc);
if (!rtcp_report) {
return &ast_null_frame;
}
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
rtcp_report->reception_report_count = rc;
ssrc = ntohl(rtcpheader[i + 2]);
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
rtcp_report->ssrc = ssrc;
break;
case RTCP_PT_FUR:
case RTCP_PT_PSFB:
ssrc = ntohl(rtcpheader[i + 1]);
break;
case AST_RTP_RTCP_RTPFB:
ssrc = ntohl(rtcpheader[i + 2]);
break;
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
case RTCP_PT_SDES:
case RTCP_PT_BYE:
default:
ssrc = 0;
ssrc_valid = 0;
break;
}
if (rtcp_debug_test_addr(addr)) {
const char *subtype = rtcp_payload_subtype2str(pt, rc);
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
ast_verbose("\n");
ast_verbose("RTCP from %s\n", ast_sockaddr_stringify(addr));
ast_verbose("PT: %u (%s)\n", pt, rtcp_payload_type2str(pt));
if (subtype) {
ast_verbose("Packet Subtype: %u (%s)\n", rc, subtype);
} else {
ast_verbose("Reception reports: %u\n", rc);
}
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
ast_verbose("SSRC of sender: %u\n", ssrc);
}
/* Determine the appropriate instance for this */
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
if (ssrc_valid) {
/*
* Depending on the payload type, either the packet source or media source
* SSRC is used.
*/
if (use_packet_source) {
child = rtp_find_instance_by_packet_source_ssrc(transport, transport_rtp, ssrc);
} else {
child = rtp_find_instance_by_media_source_ssrc(transport, transport_rtp, ssrc);
}
if (child && child != transport) {
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
/*
* It is safe to hold the child lock while holding the parent lock.
* We guarantee that the locking order is always parent->child or
* that the child lock is not held when acquiring the parent lock.
*/
ao2_lock(child);
instance = child;
rtp = ast_rtp_instance_get_data(instance);
} else {
/* The child is the parent! We don't need to unlock it. */
child = NULL;
rtp = transport_rtp;
}
} else {
child = NULL;
rtp = transport_rtp;
}
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
if (ssrc_valid && rtp->themssrc_valid) {
/*
* If the SSRC is 1, we still need to handle RTCP since this could be a
* special case. For example, if we have a unidirectional video stream, the
* SSRC may be set to 1 by the browser (in the case of chromium), and requests
* will still need to be processed so that video can flow as expected. This
* should only be done for PLI and FUR, since there is not a way to get the
* appropriate rtp instance when the SSRC is 1.
*/
int exception = (ssrc == 1 && !((pt == RTCP_PT_PSFB && rc == AST_RTP_RTCP_FMT_PLI) || pt == RTCP_PT_FUR));
if ((ssrc != rtp->themssrc && use_packet_source && ssrc != 1)
|| exception) {
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
/*
* Skip over this RTCP record as it does not contain the
* correct SSRC. We should not act upon RTCP records
* for a different stream.
*/
position += length;
ast_debug_rtcp(1, "(%p) RTCP %p -- from %s: Skipping record, received SSRC '%u' != expected '%u'\n",
instance, rtp, ast_sockaddr_stringify(addr), ssrc, rtp->themssrc);
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
if (child) {
ao2_unlock(child);
}
continue;
}
ssrc_seen = 1;
}
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
if (ssrc_seen && ast_rtp_instance_get_prop(instance, AST_RTP_PROPERTY_NAT)) {
/* Send to whoever sent to us */
if (ast_sockaddr_cmp(&rtp->rtcp->them, addr)) {
ast_sockaddr_copy(&rtp->rtcp->them, addr);
if (ast_debug_rtp_packet_is_allowed) {
ast_debug(0, "(%p) RTCP NAT: Got RTCP from other end. Now sending to address %s\n",
instance, ast_sockaddr_stringify(addr));
}
}
}
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
i += RTCP_HEADER_SSRC_LENGTH; /* Advance past header and ssrc */
switch (pt) {
case RTCP_PT_SR:
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
gettimeofday(&rtp->rtcp->rxlsr, NULL);
rtp->rtcp->themrxlsr = ((ntohl(rtcpheader[i]) & 0x0000ffff) << 16) | ((ntohl(rtcpheader[i + 1]) & 0xffff0000) >> 16);
rtp->rtcp->spc = ntohl(rtcpheader[i + 3]);
rtp->rtcp->soc = ntohl(rtcpheader[i + 4]);
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
rtcp_report->type = RTCP_PT_SR;
rtcp_report->sender_information.packet_count = rtp->rtcp->spc;
rtcp_report->sender_information.octet_count = rtp->rtcp->soc;
ntp2timeval((unsigned int)ntohl(rtcpheader[i]),
(unsigned int)ntohl(rtcpheader[i + 1]),
&rtcp_report->sender_information.ntp_timestamp);
rtcp_report->sender_information.rtp_timestamp = ntohl(rtcpheader[i + 2]);
if (rtcp_debug_test_addr(addr)) {
ast_verbose("NTP timestamp: %u.%06u\n",
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
(unsigned int)rtcp_report->sender_information.ntp_timestamp.tv_sec,
(unsigned int)rtcp_report->sender_information.ntp_timestamp.tv_usec);
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
ast_verbose("RTP timestamp: %u\n", rtcp_report->sender_information.rtp_timestamp);
ast_verbose("SPC: %u\tSOC: %u\n",
rtcp_report->sender_information.packet_count,
rtcp_report->sender_information.octet_count);
}
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
i += RTCP_SR_BLOCK_WORD_LENGTH;
/* Intentional fall through */
case RTCP_PT_RR:
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
if (rtcp_report->type != RTCP_PT_SR) {
rtcp_report->type = RTCP_PT_RR;
}
res_hep_rtcp: Add module that sends RTCP information to a Homer Server This patch adds a new module to Asterisk, res_hep_rtcp. The module subscribes to the RTCP topics in Stasis and receives RTCP information back from the message bus. It encodes into HEPv3 packets and sends the information to the res_hep module for transmission. Using this, someone with a Homer server can get live call quality monitoring for all RTP-based channels in their Asterisk 12+ systems. In addition, there were a few bugs in the RTP engine, res_rtp_asterisk, and chan_pjsip that were uncovered by the tests written for the Asterisk Test Suite. This patch fixes the following: 1) chan_pjsip failed to set its channel unique ids on its RTP instance on outbound calls. It now does this in the appropriate location, in the serialized call callback. 2) The rtp_engine was overflowing some values when packed into JSON. Specifically, some longs and unsigned ints can't be be packed into integer values, for obvious reasons. Since libjansson only supports integers, floats, strings, booleans, and objects, we print these values into strings. 3) res_rtp_asterisk had a few problems: (a) it would emit a source IP address of 0.0.0.0 if bound to that IP address. We now use ast_find_ourip to get a better IP address, and properly marshal the result into an ast_strdupa'd string. (b) Reports can be generated with no report bodies. In particular, this occurs when a sender is transmitting information to a receiver (who will send no RTP back to the sender). As such, the sender has no report body for what it received. We now properly handle this case, and the sender will emit SR reports with no body. Likewise, if we receive an RTCP packet with no report body, we will still generate the appropriate events. ASTERISK-24119 #close ........ Merged revisions 419823 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@419825 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-07-31 11:57:51 +00:00
if (rc > 0) {
/* Don't handle multiple reception reports (rc > 1) yet */
report_block = ast_calloc(1, sizeof(*report_block));
if (!report_block) {
if (child) {
ao2_unlock(child);
}
res_hep_rtcp: Add module that sends RTCP information to a Homer Server This patch adds a new module to Asterisk, res_hep_rtcp. The module subscribes to the RTCP topics in Stasis and receives RTCP information back from the message bus. It encodes into HEPv3 packets and sends the information to the res_hep module for transmission. Using this, someone with a Homer server can get live call quality monitoring for all RTP-based channels in their Asterisk 12+ systems. In addition, there were a few bugs in the RTP engine, res_rtp_asterisk, and chan_pjsip that were uncovered by the tests written for the Asterisk Test Suite. This patch fixes the following: 1) chan_pjsip failed to set its channel unique ids on its RTP instance on outbound calls. It now does this in the appropriate location, in the serialized call callback. 2) The rtp_engine was overflowing some values when packed into JSON. Specifically, some longs and unsigned ints can't be be packed into integer values, for obvious reasons. Since libjansson only supports integers, floats, strings, booleans, and objects, we print these values into strings. 3) res_rtp_asterisk had a few problems: (a) it would emit a source IP address of 0.0.0.0 if bound to that IP address. We now use ast_find_ourip to get a better IP address, and properly marshal the result into an ast_strdupa'd string. (b) Reports can be generated with no report bodies. In particular, this occurs when a sender is transmitting information to a receiver (who will send no RTP back to the sender). As such, the sender has no report body for what it received. We now properly handle this case, and the sender will emit SR reports with no body. Likewise, if we receive an RTCP packet with no report body, we will still generate the appropriate events. ASTERISK-24119 #close ........ Merged revisions 419823 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@419825 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-07-31 11:57:51 +00:00
return &ast_null_frame;
}
rtcp_report->report_block[0] = report_block;
res_hep_rtcp: Add module that sends RTCP information to a Homer Server This patch adds a new module to Asterisk, res_hep_rtcp. The module subscribes to the RTCP topics in Stasis and receives RTCP information back from the message bus. It encodes into HEPv3 packets and sends the information to the res_hep module for transmission. Using this, someone with a Homer server can get live call quality monitoring for all RTP-based channels in their Asterisk 12+ systems. In addition, there were a few bugs in the RTP engine, res_rtp_asterisk, and chan_pjsip that were uncovered by the tests written for the Asterisk Test Suite. This patch fixes the following: 1) chan_pjsip failed to set its channel unique ids on its RTP instance on outbound calls. It now does this in the appropriate location, in the serialized call callback. 2) The rtp_engine was overflowing some values when packed into JSON. Specifically, some longs and unsigned ints can't be be packed into integer values, for obvious reasons. Since libjansson only supports integers, floats, strings, booleans, and objects, we print these values into strings. 3) res_rtp_asterisk had a few problems: (a) it would emit a source IP address of 0.0.0.0 if bound to that IP address. We now use ast_find_ourip to get a better IP address, and properly marshal the result into an ast_strdupa'd string. (b) Reports can be generated with no report bodies. In particular, this occurs when a sender is transmitting information to a receiver (who will send no RTP back to the sender). As such, the sender has no report body for what it received. We now properly handle this case, and the sender will emit SR reports with no body. Likewise, if we receive an RTCP packet with no report body, we will still generate the appropriate events. ASTERISK-24119 #close ........ Merged revisions 419823 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@419825 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-07-31 11:57:51 +00:00
report_block->source_ssrc = ntohl(rtcpheader[i]);
report_block->lost_count.packets = ntohl(rtcpheader[i + 1]) & 0x00ffffff;
report_block->lost_count.fraction = ((ntohl(rtcpheader[i + 1]) & 0xff000000) >> 24);
report_block->highest_seq_no = ntohl(rtcpheader[i + 2]);
report_block->ia_jitter = ntohl(rtcpheader[i + 3]);
report_block->lsr = ntohl(rtcpheader[i + 4]);
report_block->dlsr = ntohl(rtcpheader[i + 5]);
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
if (report_block->lsr) {
int skewed = update_rtt_stats(rtp, report_block->lsr, report_block->dlsr);
if (skewed && rtcp_debug_test_addr(addr)) {
struct timeval now;
unsigned int lsr_now, lsw, msw;
gettimeofday(&now, NULL);
timeval2ntp(now, &msw, &lsw);
lsr_now = (((msw & 0xffff) << 16) | ((lsw & 0xffff0000) >> 16));
ast_verbose("Internal RTCP NTP clock skew detected: "
"lsr=%u, now=%u, dlsr=%u (%u:%03ums), "
res_hep_rtcp: Add module that sends RTCP information to a Homer Server This patch adds a new module to Asterisk, res_hep_rtcp. The module subscribes to the RTCP topics in Stasis and receives RTCP information back from the message bus. It encodes into HEPv3 packets and sends the information to the res_hep module for transmission. Using this, someone with a Homer server can get live call quality monitoring for all RTP-based channels in their Asterisk 12+ systems. In addition, there were a few bugs in the RTP engine, res_rtp_asterisk, and chan_pjsip that were uncovered by the tests written for the Asterisk Test Suite. This patch fixes the following: 1) chan_pjsip failed to set its channel unique ids on its RTP instance on outbound calls. It now does this in the appropriate location, in the serialized call callback. 2) The rtp_engine was overflowing some values when packed into JSON. Specifically, some longs and unsigned ints can't be be packed into integer values, for obvious reasons. Since libjansson only supports integers, floats, strings, booleans, and objects, we print these values into strings. 3) res_rtp_asterisk had a few problems: (a) it would emit a source IP address of 0.0.0.0 if bound to that IP address. We now use ast_find_ourip to get a better IP address, and properly marshal the result into an ast_strdupa'd string. (b) Reports can be generated with no report bodies. In particular, this occurs when a sender is transmitting information to a receiver (who will send no RTP back to the sender). As such, the sender has no report body for what it received. We now properly handle this case, and the sender will emit SR reports with no body. Likewise, if we receive an RTCP packet with no report body, we will still generate the appropriate events. ASTERISK-24119 #close ........ Merged revisions 419823 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@419825 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-07-31 11:57:51 +00:00
"diff=%u\n",
report_block->lsr, lsr_now, report_block->dlsr, report_block->dlsr / 65536,
(report_block->dlsr % 65536) * 1000 / 65536,
report_block->dlsr - (lsr_now - report_block->lsr));
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
}
res_hep_rtcp: Add module that sends RTCP information to a Homer Server This patch adds a new module to Asterisk, res_hep_rtcp. The module subscribes to the RTCP topics in Stasis and receives RTCP information back from the message bus. It encodes into HEPv3 packets and sends the information to the res_hep module for transmission. Using this, someone with a Homer server can get live call quality monitoring for all RTP-based channels in their Asterisk 12+ systems. In addition, there were a few bugs in the RTP engine, res_rtp_asterisk, and chan_pjsip that were uncovered by the tests written for the Asterisk Test Suite. This patch fixes the following: 1) chan_pjsip failed to set its channel unique ids on its RTP instance on outbound calls. It now does this in the appropriate location, in the serialized call callback. 2) The rtp_engine was overflowing some values when packed into JSON. Specifically, some longs and unsigned ints can't be be packed into integer values, for obvious reasons. Since libjansson only supports integers, floats, strings, booleans, and objects, we print these values into strings. 3) res_rtp_asterisk had a few problems: (a) it would emit a source IP address of 0.0.0.0 if bound to that IP address. We now use ast_find_ourip to get a better IP address, and properly marshal the result into an ast_strdupa'd string. (b) Reports can be generated with no report bodies. In particular, this occurs when a sender is transmitting information to a receiver (who will send no RTP back to the sender). As such, the sender has no report body for what it received. We now properly handle this case, and the sender will emit SR reports with no body. Likewise, if we receive an RTCP packet with no report body, we will still generate the appropriate events. ASTERISK-24119 #close ........ Merged revisions 419823 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@419825 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-07-31 11:57:51 +00:00
}
update_jitter_stats(rtp, report_block->ia_jitter);
update_lost_stats(rtp, report_block->lost_count.packets);
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
/*
* update_reported_mes_stats must be called AFTER
* update_rtt_stats, update_jitter_stats and
* update_lost_stats.
*/
update_reported_mes_stats(rtp);
res_hep_rtcp: Add module that sends RTCP information to a Homer Server This patch adds a new module to Asterisk, res_hep_rtcp. The module subscribes to the RTCP topics in Stasis and receives RTCP information back from the message bus. It encodes into HEPv3 packets and sends the information to the res_hep module for transmission. Using this, someone with a Homer server can get live call quality monitoring for all RTP-based channels in their Asterisk 12+ systems. In addition, there were a few bugs in the RTP engine, res_rtp_asterisk, and chan_pjsip that were uncovered by the tests written for the Asterisk Test Suite. This patch fixes the following: 1) chan_pjsip failed to set its channel unique ids on its RTP instance on outbound calls. It now does this in the appropriate location, in the serialized call callback. 2) The rtp_engine was overflowing some values when packed into JSON. Specifically, some longs and unsigned ints can't be be packed into integer values, for obvious reasons. Since libjansson only supports integers, floats, strings, booleans, and objects, we print these values into strings. 3) res_rtp_asterisk had a few problems: (a) it would emit a source IP address of 0.0.0.0 if bound to that IP address. We now use ast_find_ourip to get a better IP address, and properly marshal the result into an ast_strdupa'd string. (b) Reports can be generated with no report bodies. In particular, this occurs when a sender is transmitting information to a receiver (who will send no RTP back to the sender). As such, the sender has no report body for what it received. We now properly handle this case, and the sender will emit SR reports with no body. Likewise, if we receive an RTCP packet with no report body, we will still generate the appropriate events. ASTERISK-24119 #close ........ Merged revisions 419823 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@419825 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-07-31 11:57:51 +00:00
if (rtcp_debug_test_addr(addr)) {
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
int rate = ast_rtp_get_rate(rtp->f.subclass.format);
res_hep_rtcp: Add module that sends RTCP information to a Homer Server This patch adds a new module to Asterisk, res_hep_rtcp. The module subscribes to the RTCP topics in Stasis and receives RTCP information back from the message bus. It encodes into HEPv3 packets and sends the information to the res_hep module for transmission. Using this, someone with a Homer server can get live call quality monitoring for all RTP-based channels in their Asterisk 12+ systems. In addition, there were a few bugs in the RTP engine, res_rtp_asterisk, and chan_pjsip that were uncovered by the tests written for the Asterisk Test Suite. This patch fixes the following: 1) chan_pjsip failed to set its channel unique ids on its RTP instance on outbound calls. It now does this in the appropriate location, in the serialized call callback. 2) The rtp_engine was overflowing some values when packed into JSON. Specifically, some longs and unsigned ints can't be be packed into integer values, for obvious reasons. Since libjansson only supports integers, floats, strings, booleans, and objects, we print these values into strings. 3) res_rtp_asterisk had a few problems: (a) it would emit a source IP address of 0.0.0.0 if bound to that IP address. We now use ast_find_ourip to get a better IP address, and properly marshal the result into an ast_strdupa'd string. (b) Reports can be generated with no report bodies. In particular, this occurs when a sender is transmitting information to a receiver (who will send no RTP back to the sender). As such, the sender has no report body for what it received. We now properly handle this case, and the sender will emit SR reports with no body. Likewise, if we receive an RTCP packet with no report body, we will still generate the appropriate events. ASTERISK-24119 #close ........ Merged revisions 419823 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@419825 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-07-31 11:57:51 +00:00
ast_verbose(" Fraction lost: %d\n", report_block->lost_count.fraction);
ast_verbose(" Packets lost so far: %u\n", report_block->lost_count.packets);
ast_verbose(" Highest sequence number: %u\n", report_block->highest_seq_no & 0x0000ffff);
ast_verbose(" Sequence number cycles: %u\n", report_block->highest_seq_no >> 16);
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
ast_verbose(" Interarrival jitter (samp): %u\n", report_block->ia_jitter);
ast_verbose(" Interarrival jitter (secs): %.6f\n", ast_samp2sec(report_block->ia_jitter, rate));
res_hep_rtcp: Add module that sends RTCP information to a Homer Server This patch adds a new module to Asterisk, res_hep_rtcp. The module subscribes to the RTCP topics in Stasis and receives RTCP information back from the message bus. It encodes into HEPv3 packets and sends the information to the res_hep module for transmission. Using this, someone with a Homer server can get live call quality monitoring for all RTP-based channels in their Asterisk 12+ systems. In addition, there were a few bugs in the RTP engine, res_rtp_asterisk, and chan_pjsip that were uncovered by the tests written for the Asterisk Test Suite. This patch fixes the following: 1) chan_pjsip failed to set its channel unique ids on its RTP instance on outbound calls. It now does this in the appropriate location, in the serialized call callback. 2) The rtp_engine was overflowing some values when packed into JSON. Specifically, some longs and unsigned ints can't be be packed into integer values, for obvious reasons. Since libjansson only supports integers, floats, strings, booleans, and objects, we print these values into strings. 3) res_rtp_asterisk had a few problems: (a) it would emit a source IP address of 0.0.0.0 if bound to that IP address. We now use ast_find_ourip to get a better IP address, and properly marshal the result into an ast_strdupa'd string. (b) Reports can be generated with no report bodies. In particular, this occurs when a sender is transmitting information to a receiver (who will send no RTP back to the sender). As such, the sender has no report body for what it received. We now properly handle this case, and the sender will emit SR reports with no body. Likewise, if we receive an RTCP packet with no report body, we will still generate the appropriate events. ASTERISK-24119 #close ........ Merged revisions 419823 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@419825 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-07-31 11:57:51 +00:00
ast_verbose(" Last SR(our NTP): %lu.%010lu\n",(unsigned long)(report_block->lsr) >> 16,((unsigned long)(report_block->lsr) << 16) * 4096);
ast_verbose(" DLSR: %4.4f (sec)\n",(double)report_block->dlsr / 65536.0);
ast_verbose(" RTT: %4.4f(sec)\n", rtp->rtcp->rtt);
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
ast_verbose(" MES: %4.1f\n", rtp->rtcp->reported_mes);
res_hep_rtcp: Add module that sends RTCP information to a Homer Server This patch adds a new module to Asterisk, res_hep_rtcp. The module subscribes to the RTCP topics in Stasis and receives RTCP information back from the message bus. It encodes into HEPv3 packets and sends the information to the res_hep module for transmission. Using this, someone with a Homer server can get live call quality monitoring for all RTP-based channels in their Asterisk 12+ systems. In addition, there were a few bugs in the RTP engine, res_rtp_asterisk, and chan_pjsip that were uncovered by the tests written for the Asterisk Test Suite. This patch fixes the following: 1) chan_pjsip failed to set its channel unique ids on its RTP instance on outbound calls. It now does this in the appropriate location, in the serialized call callback. 2) The rtp_engine was overflowing some values when packed into JSON. Specifically, some longs and unsigned ints can't be be packed into integer values, for obvious reasons. Since libjansson only supports integers, floats, strings, booleans, and objects, we print these values into strings. 3) res_rtp_asterisk had a few problems: (a) it would emit a source IP address of 0.0.0.0 if bound to that IP address. We now use ast_find_ourip to get a better IP address, and properly marshal the result into an ast_strdupa'd string. (b) Reports can be generated with no report bodies. In particular, this occurs when a sender is transmitting information to a receiver (who will send no RTP back to the sender). As such, the sender has no report body for what it received. We now properly handle this case, and the sender will emit SR reports with no body. Likewise, if we receive an RTCP packet with no report body, we will still generate the appropriate events. ASTERISK-24119 #close ........ Merged revisions 419823 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@419825 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-07-31 11:57:51 +00:00
}
}
res_hep_rtcp: Add module that sends RTCP information to a Homer Server This patch adds a new module to Asterisk, res_hep_rtcp. The module subscribes to the RTCP topics in Stasis and receives RTCP information back from the message bus. It encodes into HEPv3 packets and sends the information to the res_hep module for transmission. Using this, someone with a Homer server can get live call quality monitoring for all RTP-based channels in their Asterisk 12+ systems. In addition, there were a few bugs in the RTP engine, res_rtp_asterisk, and chan_pjsip that were uncovered by the tests written for the Asterisk Test Suite. This patch fixes the following: 1) chan_pjsip failed to set its channel unique ids on its RTP instance on outbound calls. It now does this in the appropriate location, in the serialized call callback. 2) The rtp_engine was overflowing some values when packed into JSON. Specifically, some longs and unsigned ints can't be be packed into integer values, for obvious reasons. Since libjansson only supports integers, floats, strings, booleans, and objects, we print these values into strings. 3) res_rtp_asterisk had a few problems: (a) it would emit a source IP address of 0.0.0.0 if bound to that IP address. We now use ast_find_ourip to get a better IP address, and properly marshal the result into an ast_strdupa'd string. (b) Reports can be generated with no report bodies. In particular, this occurs when a sender is transmitting information to a receiver (who will send no RTP back to the sender). As such, the sender has no report body for what it received. We now properly handle this case, and the sender will emit SR reports with no body. Likewise, if we receive an RTCP packet with no report body, we will still generate the appropriate events. ASTERISK-24119 #close ........ Merged revisions 419823 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@419825 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-07-31 11:57:51 +00:00
/* If and when we handle more than one report block, this should occur outside
* this loop.
*/
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
message_blob = ast_json_pack("{s: s, s: s, s: f, s: f}",
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
"from", ast_sockaddr_stringify(addr),
"to", transport_rtp->rtcp->local_addr_str,
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
"rtt", rtp->rtcp->rtt,
"mes", rtp->rtcp->reported_mes);
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
ast_rtp_publish_rtcp_message(instance, ast_rtp_rtcp_received_type(),
rtcp_report,
message_blob);
ast_json_unref(message_blob);
/* Return an AST_FRAME_RTCP frame with the ast_rtp_rtcp_report
* object as a its data */
transport_rtp->f.frametype = AST_FRAME_RTCP;
transport_rtp->f.subclass.integer = pt;
transport_rtp->f.data.ptr = rtp->rtcp->frame_buf + AST_FRIENDLY_OFFSET;
memcpy(transport_rtp->f.data.ptr, rtcp_report, sizeof(struct ast_rtp_rtcp_report));
transport_rtp->f.datalen = sizeof(struct ast_rtp_rtcp_report);
if (rc > 0) {
/* There's always a single report block stored, here */
struct ast_rtp_rtcp_report *rtcp_report2;
report_block = transport_rtp->f.data.ptr + transport_rtp->f.datalen + sizeof(struct ast_rtp_rtcp_report_block *);
memcpy(report_block, rtcp_report->report_block[0], sizeof(struct ast_rtp_rtcp_report_block));
rtcp_report2 = (struct ast_rtp_rtcp_report *)transport_rtp->f.data.ptr;
rtcp_report2->report_block[0] = report_block;
transport_rtp->f.datalen += sizeof(struct ast_rtp_rtcp_report_block);
}
transport_rtp->f.offset = AST_FRIENDLY_OFFSET;
transport_rtp->f.samples = 0;
transport_rtp->f.mallocd = 0;
transport_rtp->f.delivery.tv_sec = 0;
transport_rtp->f.delivery.tv_usec = 0;
transport_rtp->f.src = "RTP";
transport_rtp->f.stream_num = rtp->stream_num;
f = &transport_rtp->f;
break;
case AST_RTP_RTCP_RTPFB:
switch (rc) {
case AST_RTP_RTCP_FMT_NACK:
/* If retransmissions are not enabled ignore this message */
if (!rtp->send_buffer) {
break;
}
if (rtcp_debug_test_addr(addr)) {
ast_verbose("Received generic RTCP NACK message\n");
}
ast_rtp_rtcp_handle_nack(instance, rtcpheader, position, length);
break;
default:
break;
}
break;
case RTCP_PT_FUR:
/* Handle RTCP FUR as FIR by setting the format to 4 */
rc = AST_RTP_RTCP_FMT_FIR;
case RTCP_PT_PSFB:
switch (rc) {
case AST_RTP_RTCP_FMT_PLI:
case AST_RTP_RTCP_FMT_FIR:
if (rtcp_debug_test_addr(addr)) {
ast_verbose("Received an RTCP Fast Update Request\n");
}
transport_rtp->f.frametype = AST_FRAME_CONTROL;
transport_rtp->f.subclass.integer = AST_CONTROL_VIDUPDATE;
transport_rtp->f.datalen = 0;
transport_rtp->f.samples = 0;
transport_rtp->f.mallocd = 0;
transport_rtp->f.src = "RTP";
f = &transport_rtp->f;
break;
case AST_RTP_RTCP_FMT_REMB:
/* If REMB support is not enabled ignore this message */
if (!ast_rtp_instance_get_prop(instance, AST_RTP_PROPERTY_REMB)) {
break;
}
if (rtcp_debug_test_addr(addr)) {
ast_verbose("Received REMB report\n");
}
transport_rtp->f.frametype = AST_FRAME_RTCP;
transport_rtp->f.subclass.integer = pt;
transport_rtp->f.stream_num = rtp->stream_num;
transport_rtp->f.data.ptr = rtp->rtcp->frame_buf + AST_FRIENDLY_OFFSET;
feedback = transport_rtp->f.data.ptr;
feedback->fmt = rc;
/* We don't actually care about the SSRC information in the feedback message */
first_word = ntohl(rtcpheader[i + 2]);
feedback->remb.br_exp = (first_word >> 18) & ((1 << 6) - 1);
feedback->remb.br_mantissa = first_word & ((1 << 18) - 1);
transport_rtp->f.datalen = sizeof(struct ast_rtp_rtcp_feedback);
transport_rtp->f.offset = AST_FRIENDLY_OFFSET;
transport_rtp->f.samples = 0;
transport_rtp->f.mallocd = 0;
transport_rtp->f.delivery.tv_sec = 0;
transport_rtp->f.delivery.tv_usec = 0;
transport_rtp->f.src = "RTP";
f = &transport_rtp->f;
break;
default:
break;
}
break;
case RTCP_PT_SDES:
if (rtcp_debug_test_addr(addr)) {
ast_verbose("Received an SDES from %s\n",
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
ast_sockaddr_stringify(addr));
}
#ifdef TEST_FRAMEWORK
if ((test_engine = ast_rtp_instance_get_test(instance))) {
test_engine->sdes_received = 1;
}
#endif
break;
case RTCP_PT_BYE:
if (rtcp_debug_test_addr(addr)) {
ast_verbose("Received a BYE from %s\n",
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
ast_sockaddr_stringify(addr));
}
break;
default:
break;
}
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
position += length;
rtp->rtcp->rtcp_info = 1;
if (child) {
ao2_unlock(child);
}
}
return f;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static struct ast_frame *ast_rtcp_read(struct ast_rtp_instance *instance)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
2018-05-10 18:11:06 +00:00
struct ast_srtp *srtp = ast_rtp_instance_get_srtp(instance, 1);
struct ast_sockaddr addr;
unsigned char rtcpdata[8192 + AST_FRIENDLY_OFFSET];
unsigned char *read_area = rtcpdata + AST_FRIENDLY_OFFSET;
size_t read_area_size = sizeof(rtcpdata) - AST_FRIENDLY_OFFSET;
int res;
/* Read in RTCP data from the socket */
if ((res = rtcp_recvfrom(instance, read_area, read_area_size,
0, &addr)) < 0) {
if (res == RTP_DTLS_ESTABLISHED) {
rtp->f.frametype = AST_FRAME_CONTROL;
rtp->f.subclass.integer = AST_CONTROL_SRCCHANGE;
return &rtp->f;
}
ast_assert(errno != EBADF);
if (errno != EAGAIN) {
ast_log(LOG_WARNING, "RTCP Read error: %s. Hanging up.\n",
(errno) ? strerror(errno) : "Unspecified");
return NULL;
}
return &ast_null_frame;
}
/* If this was handled by the ICE session don't do anything further */
if (!res) {
return &ast_null_frame;
}
if (!*read_area) {
struct sockaddr_in addr_tmp;
struct ast_sockaddr addr_v4;
if (ast_sockaddr_is_ipv4(&addr)) {
ast_sockaddr_to_sin(&addr, &addr_tmp);
} else if (ast_sockaddr_ipv4_mapped(&addr, &addr_v4)) {
ast_debug_stun(2, "(%p) STUN using IPv6 mapped address %s\n",
instance, ast_sockaddr_stringify(&addr));
ast_sockaddr_to_sin(&addr_v4, &addr_tmp);
} else {
ast_debug_stun(2, "(%p) STUN cannot do for non IPv4 address %s\n",
instance, ast_sockaddr_stringify(&addr));
return &ast_null_frame;
}
if ((ast_stun_handle_packet(rtp->rtcp->s, &addr_tmp, read_area, res, NULL, NULL) == AST_STUN_ACCEPT)) {
ast_sockaddr_from_sin(&addr, &addr_tmp);
ast_sockaddr_copy(&rtp->rtcp->them, &addr);
}
return &ast_null_frame;
}
2018-05-10 18:11:06 +00:00
return ast_rtcp_interpret(instance, srtp, read_area, res, &addr);
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static int bridge_p2p_rtp_write(struct ast_rtp_instance *instance,
struct ast_rtp_instance *instance1, unsigned int *rtpheader, int len, int hdrlen)
{
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
struct ast_rtp *bridged;
int res = 0, payload = 0, bridged_payload = 0, mark;
media formats: re-architect handling of media for performance improvements In the old times media formats were represented using a bit field. This was fast but had a few limitations. 1. Asterisk was limited in how many formats it could handle. 2. Formats, being a bit field, could not include any attribute information. A format was strictly its type, e.g., "this is ulaw". This was changed in Asterisk 10 (see https://wiki.asterisk.org/wiki/display/AST/Media+Architecture+Proposal for notes on that work) which led to the creation of the ast_format structure. This structure allowed Asterisk to handle attributes and bundle information with a format. Additionally, ast_format_cap was created to act as a container for multiple formats that, together, formed the capability of some entity. Another mechanism was added to allow logic to be registered which performed format attribute negotiation. Everywhere throughout the codebase Asterisk was changed to use this strategy. Unfortunately, in software, there is no free lunch. These new capabilities came at a cost. Performance analysis and profiling showed that we spend an inordinate amount of time comparing, copying, and generally manipulating formats and their related structures. Basic prototyping has shown that a reasonably large performance improvement could be made in this area. This patch is the result of that project, which overhauled the media format architecture and its usage in Asterisk to improve performance. Generally, the new philosophy for handling formats is as follows: * The ast_format structure is reference counted. This removed a large amount of the memory allocations and copying that was done in prior versions. * In order to prevent race conditions while keeping things performant, the ast_format structure is immutable by convention and lock-free. Violate this tenet at your peril! * Because formats are reference counted, codecs are also reference counted. The Asterisk core generally provides built-in codecs and caches the ast_format structures created to represent them. Generally, to prevent inordinate amounts of module reference bumping, codecs and formats can be added at run-time but cannot be removed. * All compatibility with the bit field representation of codecs/formats has been moved to a compatibility API. The primary user of this representation is chan_iax2, which must continue to maintain its bit-field usage of formats for interoperability concerns. * When a format is negotiated with attributes, or when a format cannot be represented by one of the cached formats, a new format object is created or cloned from an existing format. That format may have the same codec underlying it, but is a different format than a version of the format with different attributes or without attributes. * While formats are reference counted objects, the reference count maintained on the format should be manipulated with care. Formats are generally cached and will persist for the lifetime of Asterisk and do not explicitly need to have their lifetime modified. An exception to this is when the user of a format does not know where the format came from *and* the user may outlive the provider of the format. This occurs, for example, when a format is read from a channel: the channel may have a format with attributes (hence, non-cached) and the user of the format may last longer than the channel (if the reference to the channel is released prior to the format's reference). For more information on this work, see the API design notes: https://wiki.asterisk.org/wiki/display/AST/Media+Format+Rewrite Finally, this work was the culmination of a large number of developer's efforts. Extra thanks goes to Corey Farrell, who took on a large amount of the work in the Asterisk core, chan_sip, and was an invaluable resource in peer reviews throughout this project. There were a substantial number of patches contributed during this work; the following issues/patch names simply reflect some of the work (and will cause the release scripts to give attribution to the individuals who work on them). Reviews: https://reviewboard.asterisk.org/r/3814 https://reviewboard.asterisk.org/r/3808 https://reviewboard.asterisk.org/r/3805 https://reviewboard.asterisk.org/r/3803 https://reviewboard.asterisk.org/r/3801 https://reviewboard.asterisk.org/r/3798 https://reviewboard.asterisk.org/r/3800 https://reviewboard.asterisk.org/r/3794 https://reviewboard.asterisk.org/r/3793 https://reviewboard.asterisk.org/r/3792 https://reviewboard.asterisk.org/r/3791 https://reviewboard.asterisk.org/r/3790 https://reviewboard.asterisk.org/r/3789 https://reviewboard.asterisk.org/r/3788 https://reviewboard.asterisk.org/r/3787 https://reviewboard.asterisk.org/r/3786 https://reviewboard.asterisk.org/r/3784 https://reviewboard.asterisk.org/r/3783 https://reviewboard.asterisk.org/r/3778 https://reviewboard.asterisk.org/r/3774 https://reviewboard.asterisk.org/r/3775 https://reviewboard.asterisk.org/r/3772 https://reviewboard.asterisk.org/r/3761 https://reviewboard.asterisk.org/r/3754 https://reviewboard.asterisk.org/r/3753 https://reviewboard.asterisk.org/r/3751 https://reviewboard.asterisk.org/r/3750 https://reviewboard.asterisk.org/r/3748 https://reviewboard.asterisk.org/r/3747 https://reviewboard.asterisk.org/r/3746 https://reviewboard.asterisk.org/r/3742 https://reviewboard.asterisk.org/r/3740 https://reviewboard.asterisk.org/r/3739 https://reviewboard.asterisk.org/r/3738 https://reviewboard.asterisk.org/r/3737 https://reviewboard.asterisk.org/r/3736 https://reviewboard.asterisk.org/r/3734 https://reviewboard.asterisk.org/r/3722 https://reviewboard.asterisk.org/r/3713 https://reviewboard.asterisk.org/r/3703 https://reviewboard.asterisk.org/r/3689 https://reviewboard.asterisk.org/r/3687 https://reviewboard.asterisk.org/r/3674 https://reviewboard.asterisk.org/r/3671 https://reviewboard.asterisk.org/r/3667 https://reviewboard.asterisk.org/r/3665 https://reviewboard.asterisk.org/r/3625 https://reviewboard.asterisk.org/r/3602 https://reviewboard.asterisk.org/r/3519 https://reviewboard.asterisk.org/r/3518 https://reviewboard.asterisk.org/r/3516 https://reviewboard.asterisk.org/r/3515 https://reviewboard.asterisk.org/r/3512 https://reviewboard.asterisk.org/r/3506 https://reviewboard.asterisk.org/r/3413 https://reviewboard.asterisk.org/r/3410 https://reviewboard.asterisk.org/r/3387 https://reviewboard.asterisk.org/r/3388 https://reviewboard.asterisk.org/r/3389 https://reviewboard.asterisk.org/r/3390 https://reviewboard.asterisk.org/r/3321 https://reviewboard.asterisk.org/r/3320 https://reviewboard.asterisk.org/r/3319 https://reviewboard.asterisk.org/r/3318 https://reviewboard.asterisk.org/r/3266 https://reviewboard.asterisk.org/r/3265 https://reviewboard.asterisk.org/r/3234 https://reviewboard.asterisk.org/r/3178 ASTERISK-23114 #close Reported by: mjordan media_formats_translation_core.diff uploaded by kharwell (License 6464) rb3506.diff uploaded by mjordan (License 6283) media_format_app_file.diff uploaded by kharwell (License 6464) misc-2.diff uploaded by file (License 5000) chan_mild-3.diff uploaded by file (License 5000) chan_obscure.diff uploaded by file (License 5000) jingle.diff uploaded by file (License 5000) funcs.diff uploaded by file (License 5000) formats.diff uploaded by file (License 5000) core.diff uploaded by file (License 5000) bridges.diff uploaded by file (License 5000) mf-codecs-2.diff uploaded by file (License 5000) mf-app_fax.diff uploaded by file (License 5000) mf-apps-3.diff uploaded by file (License 5000) media-formats-3.diff uploaded by file (License 5000) ASTERISK-23715 rb3713.patch uploaded by coreyfarrell (License 5909) rb3689.patch uploaded by mjordan (License 6283) ASTERISK-23957 rb3722.patch uploaded by mjordan (License 6283) mf-attributes-3.diff uploaded by file (License 5000) ASTERISK-23958 Tested by: jrose rb3822.patch uploaded by coreyfarrell (License 5909) rb3800.patch uploaded by jrose (License 6182) chan_sip.diff uploaded by mjordan (License 6283) rb3747.patch uploaded by jrose (License 6182) ASTERISK-23959 #close Tested by: sgriepentrog, mjordan, coreyfarrell sip_cleanup.diff uploaded by opticron (License 6273) chan_sip_caps.diff uploaded by mjordan (License 6283) rb3751.patch uploaded by coreyfarrell (License 5909) chan_sip-3.diff uploaded by file (License 5000) ASTERISK-23960 #close Tested by: opticron direct_media.diff uploaded by opticron (License 6273) pjsip-direct-media.diff uploaded by file (License 5000) format_cap_remove.diff uploaded by opticron (License 6273) media_format_fixes.diff uploaded by opticron (License 6273) chan_pjsip-2.diff uploaded by file (License 5000) ASTERISK-23966 #close Tested by: rmudgett rb3803.patch uploaded by rmudgetti (License 5621) chan_dahdi.diff uploaded by file (License 5000) ASTERISK-24064 #close Tested by: coreyfarrell, mjordan, opticron, file, rmudgett, sgriepentrog, jrose rb3814.patch uploaded by rmudgett (License 5621) moh_cleanup.diff uploaded by opticron (License 6273) bridge_leak.diff uploaded by opticron (License 6273) translate.diff uploaded by file (License 5000) rb3795.patch uploaded by rmudgett (License 5621) tls_fix.diff uploaded by mjordan (License 6283) fax-mf-fix-2.diff uploaded by file (License 5000) rtp_transfer_stuff uploaded by mjordan (License 6283) rb3787.patch uploaded by rmudgett (License 5621) media-formats-explicit-translate-format-3.diff uploaded by file (License 5000) format_cache_case_fix.diff uploaded by opticron (License 6273) rb3774.patch uploaded by rmudgett (License 5621) rb3775.patch uploaded by rmudgett (License 5621) rtp_engine_fix.diff uploaded by opticron (License 6273) rtp_crash_fix.diff uploaded by opticron (License 6273) rb3753.patch uploaded by mjordan (License 6283) rb3750.patch uploaded by mjordan (License 6283) rb3748.patch uploaded by rmudgett (License 5621) media_format_fixes.diff uploaded by opticron (License 6273) rb3740.patch uploaded by mjordan (License 6283) rb3739.patch uploaded by mjordan (License 6283) rb3734.patch uploaded by mjordan (License 6283) rb3689.patch uploaded by mjordan (License 6283) rb3674.patch uploaded by coreyfarrell (License 5909) rb3671.patch uploaded by coreyfarrell (License 5909) rb3667.patch uploaded by coreyfarrell (License 5909) rb3665.patch uploaded by mjordan (License 6283) rb3625.patch uploaded by coreyfarrell (License 5909) rb3602.patch uploaded by coreyfarrell (License 5909) format_compatibility-2.diff uploaded by file (License 5000) core.diff uploaded by file (License 5000) git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@419044 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-07-20 22:06:33 +00:00
RAII_VAR(struct ast_rtp_payload_type *, payload_type, NULL, ao2_cleanup);
int reconstruct = ntohl(rtpheader[0]);
struct ast_sockaddr remote_address = { {0,} };
int ice;
unsigned int timestamp = ntohl(rtpheader[1]);
/* Get fields from packet */
payload = (reconstruct & 0x7f0000) >> 16;
mark = (reconstruct & 0x800000) >> 23;
/* Check what the payload value should be */
media formats: re-architect handling of media for performance improvements In the old times media formats were represented using a bit field. This was fast but had a few limitations. 1. Asterisk was limited in how many formats it could handle. 2. Formats, being a bit field, could not include any attribute information. A format was strictly its type, e.g., "this is ulaw". This was changed in Asterisk 10 (see https://wiki.asterisk.org/wiki/display/AST/Media+Architecture+Proposal for notes on that work) which led to the creation of the ast_format structure. This structure allowed Asterisk to handle attributes and bundle information with a format. Additionally, ast_format_cap was created to act as a container for multiple formats that, together, formed the capability of some entity. Another mechanism was added to allow logic to be registered which performed format attribute negotiation. Everywhere throughout the codebase Asterisk was changed to use this strategy. Unfortunately, in software, there is no free lunch. These new capabilities came at a cost. Performance analysis and profiling showed that we spend an inordinate amount of time comparing, copying, and generally manipulating formats and their related structures. Basic prototyping has shown that a reasonably large performance improvement could be made in this area. This patch is the result of that project, which overhauled the media format architecture and its usage in Asterisk to improve performance. Generally, the new philosophy for handling formats is as follows: * The ast_format structure is reference counted. This removed a large amount of the memory allocations and copying that was done in prior versions. * In order to prevent race conditions while keeping things performant, the ast_format structure is immutable by convention and lock-free. Violate this tenet at your peril! * Because formats are reference counted, codecs are also reference counted. The Asterisk core generally provides built-in codecs and caches the ast_format structures created to represent them. Generally, to prevent inordinate amounts of module reference bumping, codecs and formats can be added at run-time but cannot be removed. * All compatibility with the bit field representation of codecs/formats has been moved to a compatibility API. The primary user of this representation is chan_iax2, which must continue to maintain its bit-field usage of formats for interoperability concerns. * When a format is negotiated with attributes, or when a format cannot be represented by one of the cached formats, a new format object is created or cloned from an existing format. That format may have the same codec underlying it, but is a different format than a version of the format with different attributes or without attributes. * While formats are reference counted objects, the reference count maintained on the format should be manipulated with care. Formats are generally cached and will persist for the lifetime of Asterisk and do not explicitly need to have their lifetime modified. An exception to this is when the user of a format does not know where the format came from *and* the user may outlive the provider of the format. This occurs, for example, when a format is read from a channel: the channel may have a format with attributes (hence, non-cached) and the user of the format may last longer than the channel (if the reference to the channel is released prior to the format's reference). For more information on this work, see the API design notes: https://wiki.asterisk.org/wiki/display/AST/Media+Format+Rewrite Finally, this work was the culmination of a large number of developer's efforts. Extra thanks goes to Corey Farrell, who took on a large amount of the work in the Asterisk core, chan_sip, and was an invaluable resource in peer reviews throughout this project. There were a substantial number of patches contributed during this work; the following issues/patch names simply reflect some of the work (and will cause the release scripts to give attribution to the individuals who work on them). Reviews: https://reviewboard.asterisk.org/r/3814 https://reviewboard.asterisk.org/r/3808 https://reviewboard.asterisk.org/r/3805 https://reviewboard.asterisk.org/r/3803 https://reviewboard.asterisk.org/r/3801 https://reviewboard.asterisk.org/r/3798 https://reviewboard.asterisk.org/r/3800 https://reviewboard.asterisk.org/r/3794 https://reviewboard.asterisk.org/r/3793 https://reviewboard.asterisk.org/r/3792 https://reviewboard.asterisk.org/r/3791 https://reviewboard.asterisk.org/r/3790 https://reviewboard.asterisk.org/r/3789 https://reviewboard.asterisk.org/r/3788 https://reviewboard.asterisk.org/r/3787 https://reviewboard.asterisk.org/r/3786 https://reviewboard.asterisk.org/r/3784 https://reviewboard.asterisk.org/r/3783 https://reviewboard.asterisk.org/r/3778 https://reviewboard.asterisk.org/r/3774 https://reviewboard.asterisk.org/r/3775 https://reviewboard.asterisk.org/r/3772 https://reviewboard.asterisk.org/r/3761 https://reviewboard.asterisk.org/r/3754 https://reviewboard.asterisk.org/r/3753 https://reviewboard.asterisk.org/r/3751 https://reviewboard.asterisk.org/r/3750 https://reviewboard.asterisk.org/r/3748 https://reviewboard.asterisk.org/r/3747 https://reviewboard.asterisk.org/r/3746 https://reviewboard.asterisk.org/r/3742 https://reviewboard.asterisk.org/r/3740 https://reviewboard.asterisk.org/r/3739 https://reviewboard.asterisk.org/r/3738 https://reviewboard.asterisk.org/r/3737 https://reviewboard.asterisk.org/r/3736 https://reviewboard.asterisk.org/r/3734 https://reviewboard.asterisk.org/r/3722 https://reviewboard.asterisk.org/r/3713 https://reviewboard.asterisk.org/r/3703 https://reviewboard.asterisk.org/r/3689 https://reviewboard.asterisk.org/r/3687 https://reviewboard.asterisk.org/r/3674 https://reviewboard.asterisk.org/r/3671 https://reviewboard.asterisk.org/r/3667 https://reviewboard.asterisk.org/r/3665 https://reviewboard.asterisk.org/r/3625 https://reviewboard.asterisk.org/r/3602 https://reviewboard.asterisk.org/r/3519 https://reviewboard.asterisk.org/r/3518 https://reviewboard.asterisk.org/r/3516 https://reviewboard.asterisk.org/r/3515 https://reviewboard.asterisk.org/r/3512 https://reviewboard.asterisk.org/r/3506 https://reviewboard.asterisk.org/r/3413 https://reviewboard.asterisk.org/r/3410 https://reviewboard.asterisk.org/r/3387 https://reviewboard.asterisk.org/r/3388 https://reviewboard.asterisk.org/r/3389 https://reviewboard.asterisk.org/r/3390 https://reviewboard.asterisk.org/r/3321 https://reviewboard.asterisk.org/r/3320 https://reviewboard.asterisk.org/r/3319 https://reviewboard.asterisk.org/r/3318 https://reviewboard.asterisk.org/r/3266 https://reviewboard.asterisk.org/r/3265 https://reviewboard.asterisk.org/r/3234 https://reviewboard.asterisk.org/r/3178 ASTERISK-23114 #close Reported by: mjordan media_formats_translation_core.diff uploaded by kharwell (License 6464) rb3506.diff uploaded by mjordan (License 6283) media_format_app_file.diff uploaded by kharwell (License 6464) misc-2.diff uploaded by file (License 5000) chan_mild-3.diff uploaded by file (License 5000) chan_obscure.diff uploaded by file (License 5000) jingle.diff uploaded by file (License 5000) funcs.diff uploaded by file (License 5000) formats.diff uploaded by file (License 5000) core.diff uploaded by file (License 5000) bridges.diff uploaded by file (License 5000) mf-codecs-2.diff uploaded by file (License 5000) mf-app_fax.diff uploaded by file (License 5000) mf-apps-3.diff uploaded by file (License 5000) media-formats-3.diff uploaded by file (License 5000) ASTERISK-23715 rb3713.patch uploaded by coreyfarrell (License 5909) rb3689.patch uploaded by mjordan (License 6283) ASTERISK-23957 rb3722.patch uploaded by mjordan (License 6283) mf-attributes-3.diff uploaded by file (License 5000) ASTERISK-23958 Tested by: jrose rb3822.patch uploaded by coreyfarrell (License 5909) rb3800.patch uploaded by jrose (License 6182) chan_sip.diff uploaded by mjordan (License 6283) rb3747.patch uploaded by jrose (License 6182) ASTERISK-23959 #close Tested by: sgriepentrog, mjordan, coreyfarrell sip_cleanup.diff uploaded by opticron (License 6273) chan_sip_caps.diff uploaded by mjordan (License 6283) rb3751.patch uploaded by coreyfarrell (License 5909) chan_sip-3.diff uploaded by file (License 5000) ASTERISK-23960 #close Tested by: opticron direct_media.diff uploaded by opticron (License 6273) pjsip-direct-media.diff uploaded by file (License 5000) format_cap_remove.diff uploaded by opticron (License 6273) media_format_fixes.diff uploaded by opticron (License 6273) chan_pjsip-2.diff uploaded by file (License 5000) ASTERISK-23966 #close Tested by: rmudgett rb3803.patch uploaded by rmudgetti (License 5621) chan_dahdi.diff uploaded by file (License 5000) ASTERISK-24064 #close Tested by: coreyfarrell, mjordan, opticron, file, rmudgett, sgriepentrog, jrose rb3814.patch uploaded by rmudgett (License 5621) moh_cleanup.diff uploaded by opticron (License 6273) bridge_leak.diff uploaded by opticron (License 6273) translate.diff uploaded by file (License 5000) rb3795.patch uploaded by rmudgett (License 5621) tls_fix.diff uploaded by mjordan (License 6283) fax-mf-fix-2.diff uploaded by file (License 5000) rtp_transfer_stuff uploaded by mjordan (License 6283) rb3787.patch uploaded by rmudgett (License 5621) media-formats-explicit-translate-format-3.diff uploaded by file (License 5000) format_cache_case_fix.diff uploaded by opticron (License 6273) rb3774.patch uploaded by rmudgett (License 5621) rb3775.patch uploaded by rmudgett (License 5621) rtp_engine_fix.diff uploaded by opticron (License 6273) rtp_crash_fix.diff uploaded by opticron (License 6273) rb3753.patch uploaded by mjordan (License 6283) rb3750.patch uploaded by mjordan (License 6283) rb3748.patch uploaded by rmudgett (License 5621) media_format_fixes.diff uploaded by opticron (License 6273) rb3740.patch uploaded by mjordan (License 6283) rb3739.patch uploaded by mjordan (License 6283) rb3734.patch uploaded by mjordan (License 6283) rb3689.patch uploaded by mjordan (License 6283) rb3674.patch uploaded by coreyfarrell (License 5909) rb3671.patch uploaded by coreyfarrell (License 5909) rb3667.patch uploaded by coreyfarrell (License 5909) rb3665.patch uploaded by mjordan (License 6283) rb3625.patch uploaded by coreyfarrell (License 5909) rb3602.patch uploaded by coreyfarrell (License 5909) format_compatibility-2.diff uploaded by file (License 5000) core.diff uploaded by file (License 5000) git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@419044 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-07-20 22:06:33 +00:00
payload_type = ast_rtp_codecs_get_payload(ast_rtp_instance_get_codecs(instance), payload);
if (!payload_type) {
return -1;
}
/* Otherwise adjust bridged payload to match */
rtp_engine.c: Initial split of payload types into rx and tx mappings. There are numerous problems with the current implementation of the RTP payload type mapping in Asterisk. It uses only one mapping structure to associate payload types to codecs. The single mapping is overkill if all of the payload type values are well known values. Dynamic payload type mappings do not work as well with the single mapping because RFC3264 allows each side of the link to negotiate different dynamic mappings for what they want to receive. Not only could you have the same codec mapped for sending and receiving on different payload types you could wind up with the same payload type mapped to different codecs for each direction. 1) An independent payload type mapping is needed for sending and receiving. 2) The receive mapping needs to keep track of previous mappings because of the slack to when negotiation happens and current packets in flight using the old mapping arrive. 3) The transmit mapping only needs to keep track of the current negotiated values since we are sending the packets and know when the switchover takes place. * Needed to create ast_rtp_codecs_payload_code_tx() and make some callers use the new function because ast_rtp_codecs_payload_code() was used for mappings in both directions. * Needed to create ast_rtp_codecs_payloads_xover() for cases where we need to pass preferred codec mappings to the peer channel for early media bridging or when we need to prefer the offered mapping that RFC3264 says we SHOULD use. * ast_rtp_codecs_payloads_xover() and ast_rtp_codecs_payload_code_tx() are the only new public functions created. All the others were only used for the tx or rx mapping direction so the function doxygen now reflects which direction the function operates. * chan_mgcp.c: Removed call to ast_rtp_codecs_payloads_clear() as doing that makes no sense when processing an incoming SDP. We would be wiping out any mappings that we set for the possible outgoing SDP we sent earlier. ASTERISK-25166 Reported by: Kevin Harwell ASTERISK-17410 Reported by: Boris Fox Change-Id: Iaf6c227bca68cb7c414cf2fd4108a8ac98bd45ac
2015-07-24 00:24:04 +00:00
bridged_payload = ast_rtp_codecs_payload_code_tx(ast_rtp_instance_get_codecs(instance1),
payload_type->asterisk_format, payload_type->format, payload_type->rtp_code);
/* If no codec could be matched between instance and instance1, then somehow things were made incompatible while we were still bridged. Bail. */
if (bridged_payload < 0) {
return -1;
}
/* If the payload coming in is not one of the negotiated ones then send it to the core, this will cause formats to change and the bridge to break */
if (ast_rtp_codecs_find_payload_code(ast_rtp_instance_get_codecs(instance1), bridged_payload) == -1) {
ast_debug_rtp(1, "(%p, %p) RTP unsupported payload type received\n", instance, instance1);
return -1;
}
/*
* Even if we are no longer in dtmf, we could still be receiving
* re-transmissions of the last dtmf end still. Feed those to the
* core so they can be filtered accordingly.
*/
if (rtp->last_end_timestamp.is_set && rtp->last_end_timestamp.ts == timestamp) {
ast_debug_rtp(1, "(%p, %p) RTP feeding packet with duplicate timestamp to core\n", instance, instance1);
return -1;
}
if (payload_type->asterisk_format) {
ao2_replace(rtp->lastrxformat, payload_type->format);
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*
* We have now determined that we need to send the RTP packet
* out the bridged instance to do local bridging so we must unlock
* the receiving instance to prevent deadlock with the bridged
* instance.
*
* Technically we should grab a ref to instance1 so it won't go
* away on us. However, we should be safe because the bridged
* instance won't change without both channels involved being
* locked and we currently have the channel lock for the receiving
* instance.
*/
ao2_unlock(instance);
ao2_lock(instance1);
/*
* Get the peer rtp pointer now to emphasize that using it
* must happen while instance1 is locked.
*/
bridged = ast_rtp_instance_get_data(instance1);
/* If bridged peer is in dtmf, feed all packets to core until it finishes to avoid infinite dtmf */
if (bridged->sending_digit) {
ast_debug_rtp(1, "(%p, %p) RTP Feeding packet to core until DTMF finishes\n", instance, instance1);
ao2_unlock(instance1);
ao2_lock(instance);
return -1;
}
if (payload_type->asterisk_format) {
/*
* If bridged peer has already received rtp, perform the asymmetric codec check
* if that feature has been activated
*/
if (!bridged->asymmetric_codec
&& bridged->lastrxformat != ast_format_none
&& ast_format_cmp(payload_type->format, bridged->lastrxformat) == AST_FORMAT_CMP_NOT_EQUAL) {
ast_debug_rtp(1, "(%p, %p) RTP asymmetric RTP codecs detected (TX: %s, RX: %s) sending frame to core\n",
instance, instance1, ast_format_get_name(payload_type->format),
ast_format_get_name(bridged->lastrxformat));
ao2_unlock(instance1);
ao2_lock(instance);
return -1;
}
ao2_replace(bridged->lasttxformat, payload_type->format);
}
ast_rtp_instance_get_remote_address(instance1, &remote_address);
if (ast_sockaddr_isnull(&remote_address)) {
ast_debug_rtp(5, "(%p, %p) RTP remote address is null, most likely RTP has been stopped\n",
instance, instance1);
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_unlock(instance1);
ao2_lock(instance);
return 0;
}
/* If the marker bit has been explicitly set turn it on */
if (ast_test_flag(bridged, FLAG_NEED_MARKER_BIT)) {
mark = 1;
ast_clear_flag(bridged, FLAG_NEED_MARKER_BIT);
}
/* Set the marker bit for the first local bridged packet which has the first bridged peer's SSRC. */
if (ast_test_flag(bridged, FLAG_REQ_LOCAL_BRIDGE_BIT)) {
mark = 1;
ast_clear_flag(bridged, FLAG_REQ_LOCAL_BRIDGE_BIT);
}
/* Reconstruct part of the packet */
reconstruct &= 0xFF80FFFF;
reconstruct |= (bridged_payload << 16);
reconstruct |= (mark << 23);
rtpheader[0] = htonl(reconstruct);
if (mark) {
/* make this rtp instance aware of the new ssrc it is sending */
bridged->ssrc = ntohl(rtpheader[2]);
}
/* Send the packet back out */
res = rtp_sendto(instance1, (void *)rtpheader, len, 0, &remote_address, &ice);
if (res < 0) {
if (!ast_rtp_instance_get_prop(instance1, AST_RTP_PROPERTY_NAT) || (ast_rtp_instance_get_prop(instance1, AST_RTP_PROPERTY_NAT) && (ast_test_flag(bridged, FLAG_NAT_ACTIVE) == FLAG_NAT_ACTIVE))) {
ast_log(LOG_WARNING,
"RTP Transmission error of packet to %s: %s\n",
ast_sockaddr_stringify(&remote_address),
strerror(errno));
} else if (((ast_test_flag(bridged, FLAG_NAT_ACTIVE) == FLAG_NAT_INACTIVE) || ast_debug_rtp_packet_is_allowed) && !ast_test_flag(bridged, FLAG_NAT_INACTIVE_NOWARN)) {
if (ast_debug_rtp_packet_is_allowed || DEBUG_ATLEAST(1)) {
ast_log(LOG_WARNING,
"RTP NAT: Can't write RTP to private "
"address %s, waiting for other end to "
"send audio...\n",
ast_sockaddr_stringify(&remote_address));
}
ast_set_flag(bridged, FLAG_NAT_INACTIVE_NOWARN);
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_unlock(instance1);
ao2_lock(instance);
return 0;
}
if (rtp_debug_test_addr(&remote_address)) {
ast_verbose("Sent RTP P2P packet to %s%s (type %-2.2d, len %-6.6d)\n",
ast_sockaddr_stringify(&remote_address),
ice ? " (via ICE)" : "",
bridged_payload, len - hdrlen);
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_unlock(instance1);
ao2_lock(instance);
return 0;
}
static void rtp_instance_unlock(struct ast_rtp_instance *instance)
{
if (instance) {
ao2_unlock(instance);
}
}
static int rtp_transport_wide_cc_packet_statistics_cmp(struct rtp_transport_wide_cc_packet_statistics a,
struct rtp_transport_wide_cc_packet_statistics b)
{
return a.seqno - b.seqno;
}
static void rtp_transport_wide_cc_feedback_status_vector_append(unsigned char *rtcpheader, int *packet_len, int *status_vector_chunk_bits,
uint16_t *status_vector_chunk, int status)
{
/* Appending this status will use up 2 bits */
*status_vector_chunk_bits -= 2;
/* We calculate which bits we want to update the status of. Since a status vector
* is 16 bits we take away 2 (for the header), and then we take away any that have
* already been used.
*/
*status_vector_chunk |= (status << (16 - 2 - (14 - *status_vector_chunk_bits)));
/* If there are still bits available we can return early */
if (*status_vector_chunk_bits) {
return;
}
/* Otherwise we have to place this chunk into the packet */
put_unaligned_uint16(rtcpheader + *packet_len, htons(*status_vector_chunk));
*status_vector_chunk_bits = 14;
/* The first bit being 1 indicates that this is a status vector chunk and the second
* bit being 1 indicates that we are using 2 bits to represent each status for a
* packet.
*/
*status_vector_chunk = (1 << 15) | (1 << 14);
*packet_len += 2;
}
static void rtp_transport_wide_cc_feedback_status_append(unsigned char *rtcpheader, int *packet_len, int *status_vector_chunk_bits,
uint16_t *status_vector_chunk, int *run_length_chunk_count, int *run_length_chunk_status, int status)
{
if (*run_length_chunk_status != status) {
while (*run_length_chunk_count > 0 && *run_length_chunk_count < 8) {
/* Realistically it only makes sense to use a run length chunk if there were 8 or more
* consecutive packets of the same type, otherwise we could end up making the packet larger
* if we have lots of small blocks of the same type. To help with this we backfill the status
* vector (since it always represents 7 packets). Best case we end up with only that single
* status vector and the rest are run length chunks.
*/
rtp_transport_wide_cc_feedback_status_vector_append(rtcpheader, packet_len, status_vector_chunk_bits,
status_vector_chunk, *run_length_chunk_status);
*run_length_chunk_count -= 1;
}
if (*run_length_chunk_count) {
/* There is a run length chunk which needs to be written out */
put_unaligned_uint16(rtcpheader + *packet_len, htons((0 << 15) | (*run_length_chunk_status << 13) | *run_length_chunk_count));
*packet_len += 2;
}
/* In all cases the run length chunk has to be reset */
*run_length_chunk_count = 0;
*run_length_chunk_status = -1;
if (*status_vector_chunk_bits == 14) {
/* We aren't in the middle of a status vector so we can try for a run length chunk */
*run_length_chunk_status = status;
*run_length_chunk_count = 1;
} else {
/* We're doing a status vector so populate it accordingly */
rtp_transport_wide_cc_feedback_status_vector_append(rtcpheader, packet_len, status_vector_chunk_bits,
status_vector_chunk, status);
}
} else {
/* This is easy, the run length chunk count can just get bumped up */
*run_length_chunk_count += 1;
}
}
static int rtp_transport_wide_cc_feedback_produce(const void *data)
{
struct ast_rtp_instance *instance = (struct ast_rtp_instance *) data;
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
unsigned char *rtcpheader;
char bdata[1024];
struct rtp_transport_wide_cc_packet_statistics *first_packet;
struct rtp_transport_wide_cc_packet_statistics *previous_packet;
int i;
int status_vector_chunk_bits = 14;
uint16_t status_vector_chunk = (1 << 15) | (1 << 14);
int run_length_chunk_count = 0;
int run_length_chunk_status = -1;
int packet_len = 20;
int delta_len = 0;
int packet_count = 0;
unsigned int received_msw;
unsigned int received_lsw;
struct ast_sockaddr remote_address = { { 0, } };
int res;
int ice;
unsigned int large_delta_count = 0;
unsigned int small_delta_count = 0;
unsigned int lost_count = 0;
if (!rtp || !rtp->rtcp || rtp->transport_wide_cc.schedid == -1) {
ao2_ref(instance, -1);
return 0;
}
ao2_lock(instance);
/* If no packets have been received then do nothing */
if (!AST_VECTOR_SIZE(&rtp->transport_wide_cc.packet_statistics)) {
ao2_unlock(instance);
return 1000;
}
rtcpheader = (unsigned char *)bdata;
/* The first packet in the vector acts as our base sequence number and reference time */
first_packet = AST_VECTOR_GET_ADDR(&rtp->transport_wide_cc.packet_statistics, 0);
previous_packet = first_packet;
/* We go through each packet that we have statistics for, adding it either to a status
* vector chunk or a run length chunk. The code tries to be as efficient as possible to
* reduce packet size and will favor run length chunks when it makes sense.
*/
for (i = 0; i < AST_VECTOR_SIZE(&rtp->transport_wide_cc.packet_statistics); ++i) {
struct rtp_transport_wide_cc_packet_statistics *statistics;
int lost = 0;
int res = 0;
statistics = AST_VECTOR_GET_ADDR(&rtp->transport_wide_cc.packet_statistics, i);
packet_count++;
if (first_packet != statistics) {
/* The vector stores statistics in a sorted fashion based on the sequence
* number. This ensures we can detect any packets that have been lost/not
* received by comparing the sequence numbers.
*/
lost = statistics->seqno - (previous_packet->seqno + 1);
lost_count += lost;
}
while (lost) {
/* We append a not received status until all the lost packets have been accounted for */
rtp_transport_wide_cc_feedback_status_append(rtcpheader, &packet_len, &status_vector_chunk_bits,
&status_vector_chunk, &run_length_chunk_count, &run_length_chunk_status, 0);
packet_count++;
/* If there is no more room left for storing packets stop now, we leave 20
* extra bits at the end just in case.
*/
if (packet_len + delta_len + 20 > sizeof(bdata)) {
res = -1;
break;
}
lost--;
}
/* If the lost packet appending bailed out because we have no more space, then exit here too */
if (res) {
break;
}
/* Per the spec the delta is in increments of 250 */
statistics->delta = ast_tvdiff_us(statistics->received, previous_packet->received) / 250;
/* Based on the delta determine the status of this packet */
if (statistics->delta < 0 || statistics->delta > 127) {
/* Large or negative delta */
rtp_transport_wide_cc_feedback_status_append(rtcpheader, &packet_len, &status_vector_chunk_bits,
&status_vector_chunk, &run_length_chunk_count, &run_length_chunk_status, 2);
delta_len += 2;
large_delta_count++;
} else {
/* Small delta */
rtp_transport_wide_cc_feedback_status_append(rtcpheader, &packet_len, &status_vector_chunk_bits,
&status_vector_chunk, &run_length_chunk_count, &run_length_chunk_status, 1);
delta_len += 1;
small_delta_count++;
}
previous_packet = statistics;
/* If there is no more room left in the packet stop handling of any subsequent packets */
if (packet_len + delta_len + 20 > sizeof(bdata)) {
break;
}
}
if (status_vector_chunk_bits != 14) {
/* If the status vector chunk has packets in it then place it in the RTCP packet */
put_unaligned_uint16(rtcpheader + packet_len, htons(status_vector_chunk));
packet_len += 2;
} else if (run_length_chunk_count) {
/* If there is a run length chunk in progress then place it in the RTCP packet */
put_unaligned_uint16(rtcpheader + packet_len, htons((0 << 15) | (run_length_chunk_status << 13) | run_length_chunk_count));
packet_len += 2;
}
/* We iterate again to build delta chunks */
for (i = 0; i < AST_VECTOR_SIZE(&rtp->transport_wide_cc.packet_statistics); ++i) {
struct rtp_transport_wide_cc_packet_statistics *statistics;
statistics = AST_VECTOR_GET_ADDR(&rtp->transport_wide_cc.packet_statistics, i);
if (statistics->delta < 0 || statistics->delta > 127) {
/* We need 2 bytes to store this delta */
put_unaligned_uint16(rtcpheader + packet_len, htons(statistics->delta));
packet_len += 2;
} else {
/* We can store this delta in 1 byte */
rtcpheader[packet_len] = statistics->delta;
packet_len += 1;
}
/* If this is the last packet handled by the run length chunk or status vector chunk code
* then we can go no further.
*/
if (statistics == previous_packet) {
break;
}
}
/* Zero pad the end of the packet */
while (packet_len % 4) {
rtcpheader[packet_len++] = 0;
}
/* Add the general RTCP header information */
put_unaligned_uint32(rtcpheader, htonl((2 << 30) | (AST_RTP_RTCP_FMT_TRANSPORT_WIDE_CC << 24)
| (AST_RTP_RTCP_RTPFB << 16) | ((packet_len / 4) - 1)));
put_unaligned_uint32(rtcpheader + 4, htonl(rtp->ssrc));
put_unaligned_uint32(rtcpheader + 8, htonl(rtp->themssrc));
/* Add the transport-cc specific header information */
put_unaligned_uint32(rtcpheader + 12, htonl((first_packet->seqno << 16) | packet_count));
timeval2ntp(first_packet->received, &received_msw, &received_lsw);
put_unaligned_time24(rtcpheader + 16, received_msw, received_lsw);
rtcpheader[19] = rtp->transport_wide_cc.feedback_count;
/* The packet is now fully constructed so send it out */
ast_sockaddr_copy(&remote_address, &rtp->rtcp->them);
ast_debug_rtcp(2, "(%p) RTCP sending transport-cc feedback packet of size '%d' on '%s' with packet count of %d (small = %d, large = %d, lost = %d)\n",
instance, packet_len, ast_rtp_instance_get_channel_id(instance), packet_count, small_delta_count, large_delta_count, lost_count);
res = rtcp_sendto(instance, (unsigned int *)rtcpheader, packet_len, 0, &remote_address, &ice);
if (res < 0) {
ast_log(LOG_ERROR, "RTCP transport-cc feedback error to %s due to %s\n",
ast_sockaddr_stringify(&remote_address), strerror(errno));
}
AST_VECTOR_RESET(&rtp->transport_wide_cc.packet_statistics, AST_VECTOR_ELEM_CLEANUP_NOOP);
rtp->transport_wide_cc.feedback_count++;
ao2_unlock(instance);
return 1000;
}
static void rtp_instance_parse_transport_wide_cc(struct ast_rtp_instance *instance, struct ast_rtp *rtp,
unsigned char *data, int len)
{
uint16_t *seqno = (uint16_t *)data;
struct rtp_transport_wide_cc_packet_statistics statistics;
struct ast_rtp_instance *transport = rtp->bundled ? rtp->bundled : instance;
struct ast_rtp *transport_rtp = ast_rtp_instance_get_data(transport);
/* If the sequence number has cycled over then record it as such */
if (((int)transport_rtp->transport_wide_cc.last_seqno - (int)ntohs(*seqno)) > 100) {
transport_rtp->transport_wide_cc.cycles += RTP_SEQ_MOD;
}
/* Populate the statistics information for this packet */
statistics.seqno = transport_rtp->transport_wide_cc.cycles + ntohs(*seqno);
statistics.received = ast_tvnow();
/* We allow at a maximum 1000 packet statistics in play at a time, if we hit the
* limit we give up and start fresh.
*/
if (AST_VECTOR_SIZE(&transport_rtp->transport_wide_cc.packet_statistics) > 1000) {
AST_VECTOR_RESET(&rtp->transport_wide_cc.packet_statistics, AST_VECTOR_ELEM_CLEANUP_NOOP);
}
if (!AST_VECTOR_SIZE(&transport_rtp->transport_wide_cc.packet_statistics) ||
statistics.seqno > transport_rtp->transport_wide_cc.last_extended_seqno) {
/* This is the expected path */
if (AST_VECTOR_APPEND(&transport_rtp->transport_wide_cc.packet_statistics, statistics)) {
return;
}
transport_rtp->transport_wide_cc.last_extended_seqno = statistics.seqno;
transport_rtp->transport_wide_cc.last_seqno = ntohs(*seqno);
} else {
/* This packet was out of order, so reorder it within the vector accordingly */
if (AST_VECTOR_ADD_SORTED(&transport_rtp->transport_wide_cc.packet_statistics, statistics,
rtp_transport_wide_cc_packet_statistics_cmp)) {
return;
}
}
/* If we have not yet scheduled the periodic sending of feedback for this transport then do so */
if (transport_rtp->transport_wide_cc.schedid < 0 && transport_rtp->rtcp) {
ast_debug_rtcp(1, "(%p) RTCP starting transport-cc feedback transmission on RTP instance '%p'\n", instance, transport);
ao2_ref(transport, +1);
transport_rtp->transport_wide_cc.schedid = ast_sched_add(rtp->sched, 1000,
rtp_transport_wide_cc_feedback_produce, transport);
if (transport_rtp->transport_wide_cc.schedid < 0) {
ao2_ref(transport, -1);
ast_log(LOG_WARNING, "Scheduling RTCP transport-cc feedback transmission failed on RTP instance '%p'\n",
transport);
}
}
}
static void rtp_instance_parse_extmap_extensions(struct ast_rtp_instance *instance, struct ast_rtp *rtp,
unsigned char *extension, int len)
{
int transport_wide_cc_id = ast_rtp_instance_extmap_get_id(instance, AST_RTP_EXTENSION_TRANSPORT_WIDE_CC);
int pos = 0;
/* We currently only care about the transport-cc extension, so if that's not negotiated then do nothing */
if (transport_wide_cc_id == -1) {
return;
}
/* Only while we do not exceed available extension data do we continue */
while (pos < len) {
int id = extension[pos] >> 4;
int extension_len = (extension[pos] & 0xF) + 1;
/* We've handled the first byte as it contains the extension id and length, so always
* skip ahead now
*/
pos += 1;
if (id == 0) {
/* From the RFC:
* In both forms, padding bytes have the value of 0 (zero). They may be
* placed between extension elements, if desired for alignment, or after
* the last extension element, if needed for padding. A padding byte
* does not supply the ID of an element, nor the length field. When a
* padding byte is found, it is ignored and the parser moves on to
* interpreting the next byte.
*/
continue;
} else if (id == 15) {
/* From the RFC:
* The local identifier value 15 is reserved for future extension and
* MUST NOT be used as an identifier. If the ID value 15 is
* encountered, its length field should be ignored, processing of the
* entire extension should terminate at that point, and only the
* extension elements present prior to the element with ID 15
* considered.
*/
break;
} else if ((pos + extension_len) > len) {
/* The extension is corrupted and is stating that it contains more data than is
* available in the extensions data.
*/
break;
}
/* If this is transport-cc then we need to parse it further */
if (id == transport_wide_cc_id) {
rtp_instance_parse_transport_wide_cc(instance, rtp, extension + pos, extension_len);
}
/* Skip ahead to the next extension */
pos += extension_len;
}
}
2018-05-10 18:11:06 +00:00
static struct ast_frame *ast_rtp_interpret(struct ast_rtp_instance *instance, struct ast_srtp *srtp,
const struct ast_sockaddr *remote_address, unsigned char *read_area, int length, int prev_seqno,
unsigned int bundled)
{
2018-05-10 18:11:06 +00:00
unsigned int *rtpheader = (unsigned int*)(read_area);
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
struct ast_rtp_instance *instance1;
int res = length, hdrlen = 12, ssrc, seqno, payloadtype, padding, mark, ext, cc;
unsigned int timestamp;
media formats: re-architect handling of media for performance improvements In the old times media formats were represented using a bit field. This was fast but had a few limitations. 1. Asterisk was limited in how many formats it could handle. 2. Formats, being a bit field, could not include any attribute information. A format was strictly its type, e.g., "this is ulaw". This was changed in Asterisk 10 (see https://wiki.asterisk.org/wiki/display/AST/Media+Architecture+Proposal for notes on that work) which led to the creation of the ast_format structure. This structure allowed Asterisk to handle attributes and bundle information with a format. Additionally, ast_format_cap was created to act as a container for multiple formats that, together, formed the capability of some entity. Another mechanism was added to allow logic to be registered which performed format attribute negotiation. Everywhere throughout the codebase Asterisk was changed to use this strategy. Unfortunately, in software, there is no free lunch. These new capabilities came at a cost. Performance analysis and profiling showed that we spend an inordinate amount of time comparing, copying, and generally manipulating formats and their related structures. Basic prototyping has shown that a reasonably large performance improvement could be made in this area. This patch is the result of that project, which overhauled the media format architecture and its usage in Asterisk to improve performance. Generally, the new philosophy for handling formats is as follows: * The ast_format structure is reference counted. This removed a large amount of the memory allocations and copying that was done in prior versions. * In order to prevent race conditions while keeping things performant, the ast_format structure is immutable by convention and lock-free. Violate this tenet at your peril! * Because formats are reference counted, codecs are also reference counted. The Asterisk core generally provides built-in codecs and caches the ast_format structures created to represent them. Generally, to prevent inordinate amounts of module reference bumping, codecs and formats can be added at run-time but cannot be removed. * All compatibility with the bit field representation of codecs/formats has been moved to a compatibility API. The primary user of this representation is chan_iax2, which must continue to maintain its bit-field usage of formats for interoperability concerns. * When a format is negotiated with attributes, or when a format cannot be represented by one of the cached formats, a new format object is created or cloned from an existing format. That format may have the same codec underlying it, but is a different format than a version of the format with different attributes or without attributes. * While formats are reference counted objects, the reference count maintained on the format should be manipulated with care. Formats are generally cached and will persist for the lifetime of Asterisk and do not explicitly need to have their lifetime modified. An exception to this is when the user of a format does not know where the format came from *and* the user may outlive the provider of the format. This occurs, for example, when a format is read from a channel: the channel may have a format with attributes (hence, non-cached) and the user of the format may last longer than the channel (if the reference to the channel is released prior to the format's reference). For more information on this work, see the API design notes: https://wiki.asterisk.org/wiki/display/AST/Media+Format+Rewrite Finally, this work was the culmination of a large number of developer's efforts. Extra thanks goes to Corey Farrell, who took on a large amount of the work in the Asterisk core, chan_sip, and was an invaluable resource in peer reviews throughout this project. There were a substantial number of patches contributed during this work; the following issues/patch names simply reflect some of the work (and will cause the release scripts to give attribution to the individuals who work on them). Reviews: https://reviewboard.asterisk.org/r/3814 https://reviewboard.asterisk.org/r/3808 https://reviewboard.asterisk.org/r/3805 https://reviewboard.asterisk.org/r/3803 https://reviewboard.asterisk.org/r/3801 https://reviewboard.asterisk.org/r/3798 https://reviewboard.asterisk.org/r/3800 https://reviewboard.asterisk.org/r/3794 https://reviewboard.asterisk.org/r/3793 https://reviewboard.asterisk.org/r/3792 https://reviewboard.asterisk.org/r/3791 https://reviewboard.asterisk.org/r/3790 https://reviewboard.asterisk.org/r/3789 https://reviewboard.asterisk.org/r/3788 https://reviewboard.asterisk.org/r/3787 https://reviewboard.asterisk.org/r/3786 https://reviewboard.asterisk.org/r/3784 https://reviewboard.asterisk.org/r/3783 https://reviewboard.asterisk.org/r/3778 https://reviewboard.asterisk.org/r/3774 https://reviewboard.asterisk.org/r/3775 https://reviewboard.asterisk.org/r/3772 https://reviewboard.asterisk.org/r/3761 https://reviewboard.asterisk.org/r/3754 https://reviewboard.asterisk.org/r/3753 https://reviewboard.asterisk.org/r/3751 https://reviewboard.asterisk.org/r/3750 https://reviewboard.asterisk.org/r/3748 https://reviewboard.asterisk.org/r/3747 https://reviewboard.asterisk.org/r/3746 https://reviewboard.asterisk.org/r/3742 https://reviewboard.asterisk.org/r/3740 https://reviewboard.asterisk.org/r/3739 https://reviewboard.asterisk.org/r/3738 https://reviewboard.asterisk.org/r/3737 https://reviewboard.asterisk.org/r/3736 https://reviewboard.asterisk.org/r/3734 https://reviewboard.asterisk.org/r/3722 https://reviewboard.asterisk.org/r/3713 https://reviewboard.asterisk.org/r/3703 https://reviewboard.asterisk.org/r/3689 https://reviewboard.asterisk.org/r/3687 https://reviewboard.asterisk.org/r/3674 https://reviewboard.asterisk.org/r/3671 https://reviewboard.asterisk.org/r/3667 https://reviewboard.asterisk.org/r/3665 https://reviewboard.asterisk.org/r/3625 https://reviewboard.asterisk.org/r/3602 https://reviewboard.asterisk.org/r/3519 https://reviewboard.asterisk.org/r/3518 https://reviewboard.asterisk.org/r/3516 https://reviewboard.asterisk.org/r/3515 https://reviewboard.asterisk.org/r/3512 https://reviewboard.asterisk.org/r/3506 https://reviewboard.asterisk.org/r/3413 https://reviewboard.asterisk.org/r/3410 https://reviewboard.asterisk.org/r/3387 https://reviewboard.asterisk.org/r/3388 https://reviewboard.asterisk.org/r/3389 https://reviewboard.asterisk.org/r/3390 https://reviewboard.asterisk.org/r/3321 https://reviewboard.asterisk.org/r/3320 https://reviewboard.asterisk.org/r/3319 https://reviewboard.asterisk.org/r/3318 https://reviewboard.asterisk.org/r/3266 https://reviewboard.asterisk.org/r/3265 https://reviewboard.asterisk.org/r/3234 https://reviewboard.asterisk.org/r/3178 ASTERISK-23114 #close Reported by: mjordan media_formats_translation_core.diff uploaded by kharwell (License 6464) rb3506.diff uploaded by mjordan (License 6283) media_format_app_file.diff uploaded by kharwell (License 6464) misc-2.diff uploaded by file (License 5000) chan_mild-3.diff uploaded by file (License 5000) chan_obscure.diff uploaded by file (License 5000) jingle.diff uploaded by file (License 5000) funcs.diff uploaded by file (License 5000) formats.diff uploaded by file (License 5000) core.diff uploaded by file (License 5000) bridges.diff uploaded by file (License 5000) mf-codecs-2.diff uploaded by file (License 5000) mf-app_fax.diff uploaded by file (License 5000) mf-apps-3.diff uploaded by file (License 5000) media-formats-3.diff uploaded by file (License 5000) ASTERISK-23715 rb3713.patch uploaded by coreyfarrell (License 5909) rb3689.patch uploaded by mjordan (License 6283) ASTERISK-23957 rb3722.patch uploaded by mjordan (License 6283) mf-attributes-3.diff uploaded by file (License 5000) ASTERISK-23958 Tested by: jrose rb3822.patch uploaded by coreyfarrell (License 5909) rb3800.patch uploaded by jrose (License 6182) chan_sip.diff uploaded by mjordan (License 6283) rb3747.patch uploaded by jrose (License 6182) ASTERISK-23959 #close Tested by: sgriepentrog, mjordan, coreyfarrell sip_cleanup.diff uploaded by opticron (License 6273) chan_sip_caps.diff uploaded by mjordan (License 6283) rb3751.patch uploaded by coreyfarrell (License 5909) chan_sip-3.diff uploaded by file (License 5000) ASTERISK-23960 #close Tested by: opticron direct_media.diff uploaded by opticron (License 6273) pjsip-direct-media.diff uploaded by file (License 5000) format_cap_remove.diff uploaded by opticron (License 6273) media_format_fixes.diff uploaded by opticron (License 6273) chan_pjsip-2.diff uploaded by file (License 5000) ASTERISK-23966 #close Tested by: rmudgett rb3803.patch uploaded by rmudgetti (License 5621) chan_dahdi.diff uploaded by file (License 5000) ASTERISK-24064 #close Tested by: coreyfarrell, mjordan, opticron, file, rmudgett, sgriepentrog, jrose rb3814.patch uploaded by rmudgett (License 5621) moh_cleanup.diff uploaded by opticron (License 6273) bridge_leak.diff uploaded by opticron (License 6273) translate.diff uploaded by file (License 5000) rb3795.patch uploaded by rmudgett (License 5621) tls_fix.diff uploaded by mjordan (License 6283) fax-mf-fix-2.diff uploaded by file (License 5000) rtp_transfer_stuff uploaded by mjordan (License 6283) rb3787.patch uploaded by rmudgett (License 5621) media-formats-explicit-translate-format-3.diff uploaded by file (License 5000) format_cache_case_fix.diff uploaded by opticron (License 6273) rb3774.patch uploaded by rmudgett (License 5621) rb3775.patch uploaded by rmudgett (License 5621) rtp_engine_fix.diff uploaded by opticron (License 6273) rtp_crash_fix.diff uploaded by opticron (License 6273) rb3753.patch uploaded by mjordan (License 6283) rb3750.patch uploaded by mjordan (License 6283) rb3748.patch uploaded by rmudgett (License 5621) media_format_fixes.diff uploaded by opticron (License 6273) rb3740.patch uploaded by mjordan (License 6283) rb3739.patch uploaded by mjordan (License 6283) rb3734.patch uploaded by mjordan (License 6283) rb3689.patch uploaded by mjordan (License 6283) rb3674.patch uploaded by coreyfarrell (License 5909) rb3671.patch uploaded by coreyfarrell (License 5909) rb3667.patch uploaded by coreyfarrell (License 5909) rb3665.patch uploaded by mjordan (License 6283) rb3625.patch uploaded by coreyfarrell (License 5909) rb3602.patch uploaded by coreyfarrell (License 5909) format_compatibility-2.diff uploaded by file (License 5000) core.diff uploaded by file (License 5000) git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@419044 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-07-20 22:06:33 +00:00
RAII_VAR(struct ast_rtp_payload_type *, payload, NULL, ao2_cleanup);
Recorded merge of revisions 254452 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r254452 | mmichelson | 2010-03-25 10:59:56 -0500 (Thu, 25 Mar 2010) | 44 lines Several fixes regarding RFC2833 DTMF detection. Here is a copy and paste of the details from my request on reviewboard that dealt with these changes: Fix 1. The first change in place is to fix Mantis issue 15811, which deals with a situation where Asterisk will incorrectly interpret out of order RFC2833 frames as duplicate DTMF digits. For instance, we would receive a sequence like: seqno 1: DTMF 1 seqno 2: DTMF 1 seqno 3: DTMF 1 seqno 4: DTMF 1 seqno 6: DTMF 1 (end) seqno 5: DTMF 1 seqno 7: DTMF 1 (end) seqno 8: DTMF 1 (end) Prior to this patch when we received the frame with seqno 5, we would interpret this as a new DTMF 1. With this patch, we will check the seqno of the incoming digit and not process the frame if the seqno is lower than the last recorded seqno. Note that we do not record the seqno of the dropped DTMF frame for future processing. While the above situation is what was designed to be fixed, the patch is written in such a way that the following would also be fixed too: seqno 9: DTMF 1 seqno 10: DTMF 1 (end) seqno 11: DTMF 1 (end) seqno 13: DTMF 2 seqno 12: DTMF 1 (end) seqno 14: DTMF 2 seqno 15: DTMF 2 (end) seqno 16: DTMF 2 (end) seqno 17: DTMF 2 (end) In this second situation, the beginning of the DTMF 2 arrives before the final end frame of the DTMF 1. With the patch, seqno 12 is no processed and thus we properly interpret the DTMF. Fix 2. The second change in place is to fix an issue like the following: seqno 1: DTMF 1 seqno 2: DTMF 1 seqno 3: DTMF 1 (end) *packet lost* seqno 4: DTMF 1 (end) *packet lost* seqno 5: DTMF 1 (end) *packet lost* seqno 6: DTMF 2 When we receive seqno 6, we had code in place that was supposed to properly end the previously unended DTMF 1. The problem was that the code was essentially a no-op. The code would set up an end frame for the DTMF 1 but would immediately overwrite the frame with the begin for DTMF 2. I changed process_dtmf_rfc2833() so that instead of returning a single frame, it is given as an output parameter a list of frames. Each frame that needs to be returned is appended to this list. Fix 3. The final change is a minor one where an AST_CONTROL_SRCCHANGE frame could get lost. If we process a cisco DTMF or an RFC 3389 frame and no frame was returned, then we would return &ast_null_frame. The problem is that earlier in the function, we may have generated an AST_CONTROL_SRCCHANGE frame and put it in the list of frames we wish to return. This frame would be lost in such a case. The patch fixes this problem ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@254454 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2010-03-25 16:04:48 +00:00
struct frame_list frames;
2018-05-10 18:11:06 +00:00
/* If this payload is encrypted then decrypt it using the given SRTP instance */
if ((*read_area & 0xC0) && res_srtp && srtp && res_srtp->unprotect(
srtp, read_area, &res, 0 | (srtp_replay_protection << 1)) < 0) {
return &ast_null_frame;
}
2018-05-10 18:11:06 +00:00
/* If we are currently sending DTMF to the remote party send a continuation packet */
if (rtp->sending_digit) {
ast_rtp_dtmf_continuation(instance);
}
2018-05-10 18:11:06 +00:00
/* Pull out the various other fields we will need */
ssrc = ntohl(rtpheader[2]);
2018-05-10 18:11:06 +00:00
seqno = ntohl(rtpheader[0]);
payloadtype = (seqno & 0x7f0000) >> 16;
padding = seqno & (1 << 29);
mark = seqno & (1 << 23);
ext = seqno & (1 << 28);
cc = (seqno & 0xF000000) >> 24;
seqno &= 0xffff;
timestamp = ntohl(rtpheader[1]);
AST_LIST_HEAD_INIT_NOLOCK(&frames);
/* Remove any padding bytes that may be present */
if (padding) {
res -= read_area[res - 1];
}
2018-05-10 18:11:06 +00:00
/* Skip over any CSRC fields */
if (cc) {
hdrlen += cc * 4;
}
2018-05-10 18:11:06 +00:00
/* Look for any RTP extensions, currently we do not support any */
if (ext) {
int extensions_size = (ntohl(rtpheader[hdrlen/4]) & 0xffff) << 2;
unsigned int profile;
profile = (ntohl(rtpheader[3]) & 0xffff0000) >> 16;
if (profile == 0xbede) {
/* We skip over the first 4 bytes as they are just for the one byte extension header */
rtp_instance_parse_extmap_extensions(instance, rtp, read_area + hdrlen + 4, extensions_size);
} else if (DEBUG_ATLEAST(1)) {
2018-05-10 18:11:06 +00:00
if (profile == 0x505a) {
ast_log(LOG_DEBUG, "Found Zfone extension in RTP stream - zrtp - not supported.\n");
} else {
2018-05-10 18:11:06 +00:00
/* SDP negotiated RTP extensions can not currently be output in logging */
ast_log(LOG_DEBUG, "Found unknown RTP Extensions %x\n", profile);
}
}
hdrlen += extensions_size;
hdrlen += 4;
}
2018-05-10 18:11:06 +00:00
/* Make sure after we potentially mucked with the header length that it is once again valid */
if (res < hdrlen) {
ast_log(LOG_WARNING, "RTP Read too short (%d, expecting %d\n", res, hdrlen);
return AST_LIST_FIRST(&frames) ? AST_LIST_FIRST(&frames) : &ast_null_frame;
}
/* Only non-bundled instances can change/learn the remote's SSRC implicitly. */
if (!bundled) {
/* Force a marker bit and change SSRC if the SSRC changes */
if (rtp->themssrc_valid && rtp->themssrc != ssrc) {
struct ast_frame *f, srcupdate = {
AST_FRAME_CONTROL,
.subclass.integer = AST_CONTROL_SRCCHANGE,
};
if (!mark) {
if (ast_debug_rtp_packet_is_allowed) {
ast_debug(0, "(%p) RTP forcing Marker bit, because SSRC has changed\n", instance);
}
mark = 1;
}
f = ast_frisolate(&srcupdate);
AST_LIST_INSERT_TAIL(&frames, f, frame_list);
rtp->seedrxseqno = 0;
rtp->rxcount = 0;
rtp->rxoctetcount = 0;
rtp->cycles = 0;
prev_seqno = 0;
rtp->last_seqno = 0;
rtp->last_end_timestamp.ts = 0;
rtp->last_end_timestamp.is_set = 0;
if (rtp->rtcp) {
rtp->rtcp->expected_prior = 0;
rtp->rtcp->received_prior = 0;
}
}
rtp->themssrc = ssrc; /* Record their SSRC to put in future RR */
rtp->themssrc_valid = 1;
}
2018-05-10 18:11:06 +00:00
rtp->rxcount++;
rtp->rxoctetcount += (res - hdrlen);
if (rtp->rxcount == 1) {
rtp->seedrxseqno = seqno;
}
/* Do not schedule RR if RTCP isn't run */
if (rtp->rtcp && !ast_sockaddr_isnull(&rtp->rtcp->them) && rtp->rtcp->schedid < 0) {
/* Schedule transmission of Receiver Report */
ao2_ref(instance, +1);
rtp->rtcp->schedid = ast_sched_add(rtp->sched, ast_rtcp_calc_interval(rtp), ast_rtcp_write, instance);
if (rtp->rtcp->schedid < 0) {
ao2_ref(instance, -1);
ast_log(LOG_WARNING, "scheduling RTCP transmission failed.\n");
}
}
if ((int)prev_seqno - (int)seqno > 100) /* if so it would indicate that the sender cycled; allow for misordering */
2018-05-10 18:11:06 +00:00
rtp->cycles += RTP_SEQ_MOD;
/* If we are directly bridged to another instance send the audio directly out,
* but only after updating core information about the received traffic so that
* outgoing RTCP reflects it.
*/
instance1 = ast_rtp_instance_get_bridged(instance);
if (instance1
&& !bridge_p2p_rtp_write(instance, instance1, rtpheader, res, hdrlen)) {
struct timeval rxtime;
struct ast_frame *f;
/* Update statistics for jitter so they are correct in RTCP */
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
calc_rxstamp_and_jitter(&rxtime, rtp, timestamp, mark);
2018-05-10 18:11:06 +00:00
/* When doing P2P we don't need to raise any frames about SSRC change to the core */
while ((f = AST_LIST_REMOVE_HEAD(&frames, frame_list)) != NULL) {
ast_frfree(f);
}
return &ast_null_frame;
}
payload = ast_rtp_codecs_get_payload(ast_rtp_instance_get_codecs(instance), payloadtype);
if (!payload) {
/* Unknown payload type. */
return AST_LIST_FIRST(&frames) ? AST_LIST_FIRST(&frames) : &ast_null_frame;
}
/* If the payload is not actually an Asterisk one but a special one pass it off to the respective handler */
if (!payload->asterisk_format) {
struct ast_frame *f = NULL;
if (payload->rtp_code == AST_RTP_DTMF) {
/* process_dtmf_rfc2833 may need to return multiple frames. We do this
* by passing the pointer to the frame list to it so that the method
* can append frames to the list as needed.
*/
process_dtmf_rfc2833(instance, read_area + hdrlen, res - hdrlen, seqno, timestamp, payloadtype, mark, &frames);
} else if (payload->rtp_code == AST_RTP_CISCO_DTMF) {
f = process_dtmf_cisco(instance, read_area + hdrlen, res - hdrlen, seqno, timestamp, payloadtype, mark);
} else if (payload->rtp_code == AST_RTP_CN) {
f = process_cn_rfc3389(instance, read_area + hdrlen, res - hdrlen, seqno, timestamp, payloadtype, mark);
} else {
ast_log(LOG_NOTICE, "Unknown RTP codec %d received from '%s'\n",
payloadtype,
ast_sockaddr_stringify(remote_address));
}
if (f) {
AST_LIST_INSERT_TAIL(&frames, f, frame_list);
}
/* Even if no frame was returned by one of the above methods,
* we may have a frame to return in our frame list
*/
return AST_LIST_FIRST(&frames) ? AST_LIST_FIRST(&frames) : &ast_null_frame;
}
ao2_replace(rtp->lastrxformat, payload->format);
ao2_replace(rtp->f.subclass.format, payload->format);
switch (ast_format_get_type(rtp->f.subclass.format)) {
case AST_MEDIA_TYPE_AUDIO:
rtp->f.frametype = AST_FRAME_VOICE;
break;
case AST_MEDIA_TYPE_VIDEO:
rtp->f.frametype = AST_FRAME_VIDEO;
break;
case AST_MEDIA_TYPE_TEXT:
rtp->f.frametype = AST_FRAME_TEXT;
break;
case AST_MEDIA_TYPE_IMAGE:
/* Fall through */
default:
ast_log(LOG_WARNING, "Unknown or unsupported media type: %s\n",
ast_codec_media_type2str(ast_format_get_type(rtp->f.subclass.format)));
return &ast_null_frame;
}
if (rtp->dtmf_timeout && rtp->dtmf_timeout < timestamp) {
rtp->dtmf_timeout = 0;
if (rtp->resp) {
struct ast_frame *f;
f = create_dtmf_frame(instance, AST_FRAME_DTMF_END, 0);
f->len = ast_tvdiff_ms(ast_samp2tv(rtp->dtmf_duration, ast_rtp_get_rate(f->subclass.format)), ast_tv(0, 0));
2018-05-10 18:11:06 +00:00
rtp->resp = 0;
rtp->dtmf_timeout = rtp->dtmf_duration = 0;
AST_LIST_INSERT_TAIL(&frames, f, frame_list);
return AST_LIST_FIRST(&frames);
}
}
rtp->f.src = "RTP";
rtp->f.mallocd = 0;
rtp->f.datalen = res - hdrlen;
rtp->f.data.ptr = read_area + hdrlen;
rtp->f.offset = hdrlen + AST_FRIENDLY_OFFSET;
ast_set_flag(&rtp->f, AST_FRFLAG_HAS_SEQUENCE_NUMBER);
rtp->f.seqno = seqno;
rtp->f.stream_num = rtp->stream_num;
if ((ast_format_cmp(rtp->f.subclass.format, ast_format_t140) == AST_FORMAT_CMP_EQUAL)
&& ((int)seqno - (prev_seqno + 1) > 0)
&& ((int)seqno - (prev_seqno + 1) < 10)) {
unsigned char *data = rtp->f.data.ptr;
memmove(rtp->f.data.ptr+3, rtp->f.data.ptr, rtp->f.datalen);
rtp->f.datalen +=3;
*data++ = 0xEF;
*data++ = 0xBF;
*data = 0xBD;
}
if (ast_format_cmp(rtp->f.subclass.format, ast_format_t140_red) == AST_FORMAT_CMP_EQUAL) {
unsigned char *data = rtp->f.data.ptr;
unsigned char *header_end;
int num_generations;
int header_length;
int len;
int diff =(int)seqno - (prev_seqno+1); /* if diff = 0, no drop*/
int x;
ao2_replace(rtp->f.subclass.format, ast_format_t140);
header_end = memchr(data, ((*data) & 0x7f), rtp->f.datalen);
if (header_end == NULL) {
return AST_LIST_FIRST(&frames) ? AST_LIST_FIRST(&frames) : &ast_null_frame;
}
header_end++;
header_length = header_end - data;
num_generations = header_length / 4;
len = header_length;
if (!diff) {
for (x = 0; x < num_generations; x++)
len += data[x * 4 + 3];
if (!(rtp->f.datalen - len))
return AST_LIST_FIRST(&frames) ? AST_LIST_FIRST(&frames) : &ast_null_frame;
rtp->f.data.ptr += len;
rtp->f.datalen -= len;
} else if (diff > num_generations && diff < 10) {
len -= 3;
rtp->f.data.ptr += len;
rtp->f.datalen -= len;
data = rtp->f.data.ptr;
*data++ = 0xEF;
*data++ = 0xBF;
*data = 0xBD;
} else {
for ( x = 0; x < num_generations - diff; x++)
len += data[x * 4 + 3];
rtp->f.data.ptr += len;
rtp->f.datalen -= len;
}
}
if (ast_format_get_type(rtp->f.subclass.format) == AST_MEDIA_TYPE_AUDIO) {
rtp->f.samples = ast_codec_samples_count(&rtp->f);
if (ast_format_cache_is_slinear(rtp->f.subclass.format)) {
ast_frame_byteswap_be(&rtp->f);
}
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
calc_rxstamp_and_jitter(&rtp->f.delivery, rtp, timestamp, mark);
2018-05-10 18:11:06 +00:00
/* Add timing data to let ast_generic_bridge() put the frame into a jitterbuf */
ast_set_flag(&rtp->f, AST_FRFLAG_HAS_TIMING_INFO);
rtp->f.ts = timestamp / (ast_rtp_get_rate(rtp->f.subclass.format) / 1000);
2018-05-10 18:11:06 +00:00
rtp->f.len = rtp->f.samples / ((ast_format_get_sample_rate(rtp->f.subclass.format) / 1000));
} else if (ast_format_get_type(rtp->f.subclass.format) == AST_MEDIA_TYPE_VIDEO) {
/* Video -- samples is # of samples vs. 90000 */
if (!rtp->lastividtimestamp)
rtp->lastividtimestamp = timestamp;
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
calc_rxstamp_and_jitter(&rtp->f.delivery, rtp, timestamp, mark);
2018-05-10 18:11:06 +00:00
ast_set_flag(&rtp->f, AST_FRFLAG_HAS_TIMING_INFO);
rtp->f.ts = timestamp / (ast_rtp_get_rate(rtp->f.subclass.format) / 1000);
2018-05-10 18:11:06 +00:00
rtp->f.samples = timestamp - rtp->lastividtimestamp;
rtp->lastividtimestamp = timestamp;
rtp->f.delivery.tv_sec = 0;
rtp->f.delivery.tv_usec = 0;
/* Pass the RTP marker bit as bit */
rtp->f.subclass.frame_ending = mark ? 1 : 0;
} else if (ast_format_get_type(rtp->f.subclass.format) == AST_MEDIA_TYPE_TEXT) {
/* TEXT -- samples is # of samples vs. 1000 */
if (!rtp->lastitexttimestamp)
rtp->lastitexttimestamp = timestamp;
rtp->f.samples = timestamp - rtp->lastitexttimestamp;
rtp->lastitexttimestamp = timestamp;
rtp->f.delivery.tv_sec = 0;
rtp->f.delivery.tv_usec = 0;
} else {
ast_log(LOG_WARNING, "Unknown or unsupported media type: %s\n",
ast_codec_media_type2str(ast_format_get_type(rtp->f.subclass.format)));
return &ast_null_frame;
}
AST_LIST_INSERT_TAIL(&frames, &rtp->f, frame_list);
return AST_LIST_FIRST(&frames);
}
#ifdef AST_DEVMODE
struct rtp_drop_packets_data {
/* Whether or not to randomize the number of packets to drop. */
unsigned int use_random_num;
/* Whether or not to randomize the time interval between packets drops. */
unsigned int use_random_interval;
/* The total number of packets to drop. If 'use_random_num' is true then this
* value becomes the upper bound for a number of random packets to drop. */
unsigned int num_to_drop;
/* The current number of packets that have been dropped during an interval. */
unsigned int num_dropped;
/* The optional interval to use between packet drops. If 'use_random_interval'
* is true then this values becomes the upper bound for a random interval used. */
struct timeval interval;
/* The next time a packet drop should be triggered. */
struct timeval next;
/* An optional IP address from which to drop packets from. */
struct ast_sockaddr addr;
/* The optional port from which to drop packets from. */
unsigned int port;
};
static struct rtp_drop_packets_data drop_packets_data;
static void drop_packets_data_update(struct timeval tv)
{
/*
* num_dropped keeps up with the number of packets that have been dropped for a
* given interval. Once the specified number of packets have been dropped and
* the next time interval is ready to trigger then set this number to zero (drop
* the next 'n' packets up to 'num_to_drop'), or if 'use_random_num' is set to
* true then set to a random number between zero and 'num_to_drop'.
*/
drop_packets_data.num_dropped = drop_packets_data.use_random_num ?
ast_random() % drop_packets_data.num_to_drop : 0;
/*
* A specified number of packets can be dropped at a given interval (e.g every
* 30 seconds). If 'use_random_interval' is false simply add the interval to
* the given time to get the next trigger point. If set to true, then get a
* random time between the given time and up to the specified interval.
*/
if (drop_packets_data.use_random_interval) {
/* Calculate as a percentage of the specified drop packets interval */
struct timeval interval = ast_time_create_by_unit(ast_time_tv_to_usec(
&drop_packets_data.interval) * ((double)(ast_random() % 100 + 1) / 100),
TIME_UNIT_MICROSECOND);
drop_packets_data.next = ast_tvadd(tv, interval);
} else {
drop_packets_data.next = ast_tvadd(tv, drop_packets_data.interval);
}
}
static int should_drop_packets(struct ast_sockaddr *addr)
{
struct timeval tv;
if (!drop_packets_data.num_to_drop) {
return 0;
}
/*
* If an address has been specified then filter on it, and also the port if
* it too was included.
*/
if (!ast_sockaddr_isnull(&drop_packets_data.addr) &&
(drop_packets_data.port ?
ast_sockaddr_cmp(&drop_packets_data.addr, addr) :
ast_sockaddr_cmp_addr(&drop_packets_data.addr, addr)) != 0) {
/* Address and/or port does not match */
return 0;
}
/* Keep dropping packets until we've reached the total to drop */
if (drop_packets_data.num_dropped < drop_packets_data.num_to_drop) {
++drop_packets_data.num_dropped;
return 1;
}
/*
* Once the set number of packets has been dropped check to see if it's
* time to drop more.
*/
if (ast_tvzero(drop_packets_data.interval)) {
/* If no interval then drop specified number of packets and be done */
drop_packets_data.num_to_drop = 0;
return 0;
}
tv = ast_tvnow();
if (ast_tvcmp(tv, drop_packets_data.next) == -1) {
/* Still waiting for the next time interval to elapse */
return 0;
}
/*
* The next time interval has elapsed so update the tracking structure
* in order to start dropping more packets, and figure out when the next
* time interval is.
*/
drop_packets_data_update(tv);
return 1;
}
#endif
2018-05-10 18:11:06 +00:00
/*! \pre instance is locked */
static struct ast_frame *ast_rtp_read(struct ast_rtp_instance *instance, int rtcp)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
struct ast_srtp *srtp;
RAII_VAR(struct ast_rtp_instance *, child, NULL, rtp_instance_unlock);
struct ast_sockaddr addr;
int res, hdrlen = 12, version, payloadtype;
2018-05-10 18:11:06 +00:00
unsigned char *read_area = rtp->rawdata + AST_FRIENDLY_OFFSET;
size_t read_area_size = sizeof(rtp->rawdata) - AST_FRIENDLY_OFFSET;
unsigned int *rtpheader = (unsigned int*)(read_area), seqno, ssrc, timestamp, prev_seqno;
struct ast_sockaddr remote_address = { {0,} };
struct frame_list frames;
struct ast_frame *frame;
unsigned int bundled;
2018-05-10 18:11:06 +00:00
/* If this is actually RTCP let's hop on over and handle it */
if (rtcp) {
if (rtp->rtcp && rtp->rtcp->type == AST_RTP_INSTANCE_RTCP_STANDARD) {
return ast_rtcp_read(instance);
}
return &ast_null_frame;
}
/* Actually read in the data from the socket */
if ((res = rtp_recvfrom(instance, read_area, read_area_size, 0,
&addr)) < 0) {
if (res == RTP_DTLS_ESTABLISHED) {
rtp->f.frametype = AST_FRAME_CONTROL;
rtp->f.subclass.integer = AST_CONTROL_SRCCHANGE;
return &rtp->f;
}
ast_assert(errno != EBADF);
if (errno != EAGAIN) {
ast_log(LOG_WARNING, "RTP Read error: %s. Hanging up.\n",
(errno) ? strerror(errno) : "Unspecified");
return NULL;
}
return &ast_null_frame;
}
/* If this was handled by the ICE session don't do anything */
if (!res) {
return &ast_null_frame;
}
/* This could be a multiplexed RTCP packet. If so, be sure to interpret it correctly */
if (rtcp_mux(rtp, read_area)) {
return ast_rtcp_interpret(instance, ast_rtp_instance_get_srtp(instance, 1), read_area, res, &addr);
}
/* Make sure the data that was read in is actually enough to make up an RTP packet */
if (res < hdrlen) {
/* If this is a keepalive containing only nulls, don't bother with a warning */
int i;
for (i = 0; i < res; ++i) {
if (read_area[i] != '\0') {
ast_log(LOG_WARNING, "RTP Read too short\n");
return &ast_null_frame;
}
}
return &ast_null_frame;
}
/* Get fields and verify this is an RTP packet */
seqno = ntohl(rtpheader[0]);
ast_rtp_instance_get_remote_address(instance, &remote_address);
if (!(version = (seqno & 0xC0000000) >> 30)) {
struct sockaddr_in addr_tmp;
struct ast_sockaddr addr_v4;
if (ast_sockaddr_is_ipv4(&addr)) {
ast_sockaddr_to_sin(&addr, &addr_tmp);
} else if (ast_sockaddr_ipv4_mapped(&addr, &addr_v4)) {
ast_debug_stun(1, "(%p) STUN using IPv6 mapped address %s\n",
instance, ast_sockaddr_stringify(&addr));
ast_sockaddr_to_sin(&addr_v4, &addr_tmp);
} else {
ast_debug_stun(1, "(%p) STUN cannot do for non IPv4 address %s\n",
instance, ast_sockaddr_stringify(&addr));
return &ast_null_frame;
}
if ((ast_stun_handle_packet(rtp->s, &addr_tmp, read_area, res, NULL, NULL) == AST_STUN_ACCEPT) &&
ast_sockaddr_isnull(&remote_address)) {
ast_sockaddr_from_sin(&addr, &addr_tmp);
ast_rtp_instance_set_remote_address(instance, &addr);
}
return &ast_null_frame;
}
/* If the version is not what we expected by this point then just drop the packet */
if (version != 2) {
return &ast_null_frame;
}
/* We use the SSRC to determine what RTP instance this packet is actually for */
ssrc = ntohl(rtpheader[2]);
2018-05-10 18:11:06 +00:00
/* We use the SRTP data from the provided instance that it came in on, not the child */
srtp = ast_rtp_instance_get_srtp(instance, 0);
/* Determine the appropriate instance for this */
child = rtp_find_instance_by_packet_source_ssrc(instance, rtp, ssrc);
if (!child) {
/* Neither the bundled parent nor any child has this SSRC */
return &ast_null_frame;
}
if (child != instance) {
/* It is safe to hold the child lock while holding the parent lock, we guarantee that the locking order
* is always parent->child or that the child lock is not held when acquiring the parent lock.
*/
ao2_lock(child);
instance = child;
rtp = ast_rtp_instance_get_data(instance);
} else {
/* The child is the parent! We don't need to unlock it. */
child = NULL;
}
/* If strict RTP protection is enabled see if we need to learn the remote address or if we need to drop the packet */
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
switch (rtp->strict_rtp_state) {
case STRICT_RTP_LEARN:
/*
* Scenario setup:
* PartyA -- Ast1 -- Ast2 -- PartyB
*
* The learning timeout is necessary for Ast1 to handle the above
* setup where PartyA calls PartyB and Ast2 initiates direct media
* between Ast1 and PartyB. Ast1 may lock onto the Ast2 stream and
* never learn the PartyB stream when it starts. The timeout makes
* Ast1 stay in the learning state long enough to see and learn the
* RTP stream from PartyB.
*
* To mitigate against attack, the learning state cannot switch
* streams while there are competing streams. The competing streams
* interfere with each other's qualification. Once we accept a
* stream and reach the timeout, an attacker cannot interfere
* anymore.
*
* Here are a few scenarios and each one assumes that the streams
* are continuous:
*
* 1) We already have a known stream source address and the known
* stream wants to change to a new source address. An attacking
* stream will block learning the new stream source. After the
* timeout we re-lock onto the original stream source address which
* likely went away. The result is one way audio.
*
* 2) We already have a known stream source address and the known
* stream doesn't want to change source addresses. An attacking
* stream will not be able to replace the known stream. After the
* timeout we re-lock onto the known stream. The call is not
* affected.
*
* 3) We don't have a known stream source address. This presumably
* is the start of a call. Competing streams will result in staying
* in learning mode until a stream becomes the victor and we reach
* the timeout. We cannot exit learning if we have no known stream
* to lock onto. The result is one way audio until there is a victor.
*
* If we learn a stream source address before the timeout we will be
* in scenario 1) or 2) when a competing stream starts.
*/
if (!ast_sockaddr_isnull(&rtp->strict_rtp_address)
&& STRICT_RTP_LEARN_TIMEOUT < ast_tvdiff_ms(ast_tvnow(), rtp->rtp_source_learn.start)) {
ast_verb(4, "%p -- Strict RTP learning complete - Locking on source address %s\n",
rtp, ast_sockaddr_stringify(&rtp->strict_rtp_address));
ast_test_suite_event_notify("STRICT_RTP_LEARN", "Source: %s",
ast_sockaddr_stringify(&rtp->strict_rtp_address));
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
rtp->strict_rtp_state = STRICT_RTP_CLOSED;
} else {
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
struct ast_sockaddr target_address;
if (!ast_sockaddr_cmp(&rtp->strict_rtp_address, &addr)) {
/*
* We are open to learning a new address but have received
* traffic from the current address, accept it and reset
* the learning counts for a new source. When no more
* current source packets arrive a new source can take over
* once sufficient traffic is received.
*/
rtp_learning_seq_init(&rtp->rtp_source_learn, seqno);
break;
}
/*
* We give preferential treatment to the requested target address
* (negotiated SDP address) where we are to send our RTP. However,
* the other end has no obligation to send from that address even
* though it is practically a requirement when NAT is involved.
*/
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
ast_rtp_instance_get_requested_target_address(instance, &target_address);
if (!ast_sockaddr_cmp(&target_address, &addr)) {
/* Accept the negotiated target RTP stream as the source */
ast_verb(4, "%p -- Strict RTP switching to RTP target address %s as source\n",
rtp, ast_sockaddr_stringify(&addr));
ast_sockaddr_copy(&rtp->strict_rtp_address, &addr);
rtp_learning_seq_init(&rtp->rtp_source_learn, seqno);
break;
}
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
/*
* Trying to learn a new address. If we pass a probationary period
* with it, that means we've stopped getting RTP from the original
* source and we should switch to it.
*/
if (!ast_sockaddr_cmp(&rtp->rtp_source_learn.proposed_address, &addr)) {
if (rtp->rtp_source_learn.stream_type == AST_MEDIA_TYPE_UNKNOWN) {
struct ast_rtp_codecs *codecs;
codecs = ast_rtp_instance_get_codecs(instance);
rtp->rtp_source_learn.stream_type =
ast_rtp_codecs_get_stream_type(codecs);
ast_verb(4, "%p -- Strict RTP qualifying stream type: %s\n",
rtp, ast_codec_media_type2str(rtp->rtp_source_learn.stream_type));
}
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
if (!rtp_learning_rtp_seq_update(&rtp->rtp_source_learn, seqno)) {
/* Accept the new RTP stream */
ast_verb(4, "%p -- Strict RTP switching source address to %s\n",
rtp, ast_sockaddr_stringify(&addr));
ast_sockaddr_copy(&rtp->strict_rtp_address, &addr);
rtp_learning_seq_init(&rtp->rtp_source_learn, seqno);
break;
}
/* Not ready to accept the RTP stream candidate */
ast_debug_rtp(1, "(%p) RTP %p -- Received packet from %s, dropping due to strict RTP protection. Will switch to it in %d packets.\n",
instance, rtp, ast_sockaddr_stringify(&addr), rtp->rtp_source_learn.packets);
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
} else {
/*
* This is either an attacking stream or
* the start of the expected new stream.
*/
ast_sockaddr_copy(&rtp->rtp_source_learn.proposed_address, &addr);
rtp_learning_seq_init(&rtp->rtp_source_learn, seqno);
ast_debug_rtp(1, "(%p) RTP %p -- Received packet from %s, dropping due to strict RTP protection. Qualifying new stream.\n",
instance, rtp, ast_sockaddr_stringify(&addr));
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
}
return &ast_null_frame;
}
/* Fall through */
case STRICT_RTP_CLOSED:
/*
* We should not allow a stream address change if the SSRC matches
* once strictrtp learning is closed. Any kind of address change
* like this should have happened while we were in the learning
* state. We do not want to allow the possibility of an attacker
* interfering with the RTP stream after the learning period.
* An attacker could manage to get an RTCP packet redirected to
* them which can contain the SSRC value.
*/
if (!ast_sockaddr_cmp(&rtp->strict_rtp_address, &addr)) {
break;
}
ast_debug_rtp(1, "(%p) RTP %p -- Received packet from %s, dropping due to strict RTP protection.\n",
instance, rtp, ast_sockaddr_stringify(&addr));
#ifdef TEST_FRAMEWORK
{
static int strict_rtp_test_event = 1;
if (strict_rtp_test_event) {
ast_test_suite_event_notify("STRICT_RTP_CLOSED", "Source: %s",
ast_sockaddr_stringify(&addr));
strict_rtp_test_event = 0; /* Only run this event once to prevent possible spam */
}
}
#endif
return &ast_null_frame;
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
case STRICT_RTP_OPEN:
break;
}
/* If symmetric RTP is enabled see if the remote side is not what we expected and change where we are sending audio */
if (ast_rtp_instance_get_prop(instance, AST_RTP_PROPERTY_NAT)) {
if (ast_sockaddr_cmp(&remote_address, &addr)) {
/* do not update the originally given address, but only the remote */
ast_rtp_instance_set_incoming_source_address(instance, &addr);
ast_sockaddr_copy(&remote_address, &addr);
if (rtp->rtcp && rtp->rtcp->type == AST_RTP_INSTANCE_RTCP_STANDARD) {
ast_sockaddr_copy(&rtp->rtcp->them, &addr);
ast_sockaddr_set_port(&rtp->rtcp->them, ast_sockaddr_port(&addr) + 1);
}
ast_set_flag(rtp, FLAG_NAT_ACTIVE);
if (ast_debug_rtp_packet_is_allowed)
ast_debug(0, "(%p) RTP NAT: Got audio from other end. Now sending to address %s\n",
instance, ast_sockaddr_stringify(&remote_address));
}
}
/* Pull out the various other fields we will need */
payloadtype = (seqno & 0x7f0000) >> 16;
seqno &= 0xffff;
timestamp = ntohl(rtpheader[1]);
#ifdef AST_DEVMODE
if (should_drop_packets(&addr)) {
ast_debug(0, "(%p) RTP: drop received packet from %s (type %-2.2d, seq %-6.6u, ts %-6.6u, len %-6.6d)\n",
instance, ast_sockaddr_stringify(&addr), payloadtype, seqno, timestamp, res - hdrlen);
return &ast_null_frame;
}
#endif
2018-05-10 18:11:06 +00:00
if (rtp_debug_test_addr(&addr)) {
ast_verbose("Got RTP packet from %s (type %-2.2d, seq %-6.6u, ts %-6.6u, len %-6.6d)\n",
ast_sockaddr_stringify(&addr),
payloadtype, seqno, timestamp, res - hdrlen);
}
AST_LIST_HEAD_INIT_NOLOCK(&frames);
bundled = (child || AST_VECTOR_SIZE(&rtp->ssrc_mapping)) ? 1 : 0;
2018-05-10 18:11:06 +00:00
prev_seqno = rtp->lastrxseqno;
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
/* We need to save lastrxseqno for use by jitter before resetting it. */
rtp->prevrxseqno = rtp->lastrxseqno;
2018-05-10 18:11:06 +00:00
rtp->lastrxseqno = seqno;
2018-05-10 18:11:06 +00:00
if (!rtp->recv_buffer) {
/* If there is no receive buffer then we can pass back the frame directly */
frame = ast_rtp_interpret(instance, srtp, &addr, read_area, res, prev_seqno, bundled);
AST_LIST_INSERT_TAIL(&frames, frame, frame_list);
return AST_LIST_FIRST(&frames);
2018-05-10 18:11:06 +00:00
} else if (rtp->expectedrxseqno == -1 || seqno == rtp->expectedrxseqno) {
rtp->expectedrxseqno = seqno + 1;
/* We've cycled over, so go back to 0 */
if (rtp->expectedrxseqno == SEQNO_CYCLE_OVER) {
rtp->expectedrxseqno = 0;
}
2018-05-10 18:11:06 +00:00
/* If there are no buffered packets that will be placed after this frame then we can
* return it directly without duplicating it.
*/
if (!ast_data_buffer_count(rtp->recv_buffer)) {
frame = ast_rtp_interpret(instance, srtp, &addr, read_area, res, prev_seqno, bundled);
AST_LIST_INSERT_TAIL(&frames, frame, frame_list);
return AST_LIST_FIRST(&frames);
Merged revisions 336878 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/10 ................ r336878 | russell | 2011-09-19 20:03:55 -0500 (Mon, 19 Sep 2011) | 43 lines Merged revisions 336877 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.8 ........ r336877 | russell | 2011-09-19 19:56:20 -0500 (Mon, 19 Sep 2011) | 36 lines Fix crashes in ast_rtcp_write(). This patch addresses crashes related to RTCP handling. The backtraces just show a crash in ast_rtcp_write() where it appears that the RTP instance is no longer valid. There is a race condition with scheduled RTCP transmissions and the destruction of the RTP instance. This patch utilizes the fact that ast_rtp_instance is a reference counted object and ensures that it will not get destroyed while a reference is still around due to scheduled RTCP transmissions. RTCP transmissions are scheduled and executed from the chan_sip scheduler context. This scheduler context is processed in the SIP monitor thread. The destruction of an RTP instance occurs when the associated sip_pvt gets destroyed (which happens when the sip_pvt reference count reaches 0). However, the SIP monitor thread is not the only thread that can cause a sip_pvt to get destroyed. The sip_hangup function, executed from a channel thread, also decrements the reference count on a sip_pvt and could cause it to get destroyed. While this is being changed anyway, the patch also removes calling ast_sched_del() from within the RTCP scheduler callback. It's not helpful. Simply returning 0 prevents the callback from being rescheduled. (closes issue ASTERISK-18570) Related issues that look like they are the same problem: (issue ASTERISK-17560) (issue ASTERISK-15406) (issue ASTERISK-15257) (issue ASTERISK-13334) (issue ASTERISK-9977) (issue ASTERISK-9716) Review: https://reviewboard.asterisk.org/r/1444/ ........ ................ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@336879 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2011-09-20 01:11:18 +00:00
}
2018-05-10 18:11:06 +00:00
if (!AST_VECTOR_REMOVE_CMP_ORDERED(&rtp->missing_seqno, seqno, find_by_value,
AST_VECTOR_ELEM_CLEANUP_NOOP)) {
ast_debug_rtp(2, "(%p) RTP Packet with sequence number '%d' on instance is no longer missing\n",
instance, seqno);
2018-05-10 18:11:06 +00:00
}
2018-05-10 18:11:06 +00:00
/* If we don't have the next packet after this we can directly return the frame, as there is no
* chance it will be overwritten.
*/
if (!ast_data_buffer_get(rtp->recv_buffer, rtp->expectedrxseqno)) {
frame = ast_rtp_interpret(instance, srtp, &addr, read_area, res, prev_seqno, bundled);
AST_LIST_INSERT_TAIL(&frames, frame, frame_list);
return AST_LIST_FIRST(&frames);
2018-05-10 18:11:06 +00:00
}
2018-05-10 18:11:06 +00:00
/* Otherwise we need to dupe the frame so that the potential processing of frames placed after
* it do not overwrite the data. You may be thinking that we could just add the current packet
* to the head of the frames list and avoid having to duplicate it but this would result in out
* of order packet processing by libsrtp which we are trying to avoid.
*/
frame = ast_frdup(ast_rtp_interpret(instance, srtp, &addr, read_area, res, prev_seqno, bundled));
2018-05-10 18:11:06 +00:00
if (frame) {
AST_LIST_INSERT_TAIL(&frames, frame, frame_list);
prev_seqno = seqno;
2018-05-10 18:11:06 +00:00
}
2018-05-10 18:11:06 +00:00
/* Add any additional packets that we have buffered and that are available */
while (ast_data_buffer_count(rtp->recv_buffer)) {
struct ast_rtp_rtcp_nack_payload *payload;
2018-05-10 18:11:06 +00:00
payload = (struct ast_rtp_rtcp_nack_payload *)ast_data_buffer_remove(rtp->recv_buffer, rtp->expectedrxseqno);
if (!payload) {
break;
}
frame = ast_frdup(ast_rtp_interpret(instance, srtp, &addr, payload->buf, payload->size, prev_seqno, bundled));
2018-05-10 18:11:06 +00:00
ast_free(payload);
if (!frame) {
/* If this packet can't be interpreted due to being out of memory we return what we have and assume
2018-05-10 18:11:06 +00:00
* that we will determine it is a missing packet later and NACK for it.
*/
return AST_LIST_FIRST(&frames);
}
ast_debug_rtp(2, "(%p) RTP pulled buffered packet with sequence number '%d' to additionally return\n",
instance, frame->seqno);
2018-05-10 18:11:06 +00:00
AST_LIST_INSERT_TAIL(&frames, frame, frame_list);
prev_seqno = rtp->expectedrxseqno;
2018-05-10 18:11:06 +00:00
rtp->expectedrxseqno++;
if (rtp->expectedrxseqno == SEQNO_CYCLE_OVER) {
rtp->expectedrxseqno = 0;
}
}
2018-05-10 18:11:06 +00:00
return AST_LIST_FIRST(&frames);
} else if ((((seqno - rtp->expectedrxseqno) > 100) && timestamp > rtp->lastividtimestamp) ||
2018-05-10 18:11:06 +00:00
ast_data_buffer_count(rtp->recv_buffer) == ast_data_buffer_max(rtp->recv_buffer)) {
int inserted = 0;
2018-05-10 18:11:06 +00:00
/* We have a large number of outstanding buffered packets or we've jumped far ahead in time.
* To compensate we dump what we have in the buffer and place the current packet in a logical
* spot. In the case of video we also require a full frame to give the decoding side a fighting
* chance.
*/
2018-05-10 18:11:06 +00:00
if (rtp->rtp_source_learn.stream_type == AST_MEDIA_TYPE_VIDEO) {
ast_debug_rtp(2, "(%p) RTP source has wild gap or packet loss, sending FIR\n",
2018-05-10 18:11:06 +00:00
instance);
rtp_write_rtcp_fir(instance, rtp, &remote_address);
}
/* This works by going through the progression of the sequence number retrieving buffered packets
* or inserting the current received packet until we've run out of packets. This ensures that the
* packets are in the correct sequence number order.
*/
2018-05-10 18:11:06 +00:00
while (ast_data_buffer_count(rtp->recv_buffer)) {
struct ast_rtp_rtcp_nack_payload *payload;
/* If the packet we received is the one we are expecting at this point then add it in */
if (rtp->expectedrxseqno == seqno) {
frame = ast_frdup(ast_rtp_interpret(instance, srtp, &addr, read_area, res, prev_seqno, bundled));
if (frame) {
AST_LIST_INSERT_TAIL(&frames, frame, frame_list);
prev_seqno = seqno;
ast_debug_rtp(2, "(%p) RTP inserted just received packet with sequence number '%d' in correct order\n",
instance, seqno);
}
/* It is possible due to packet retransmission for this packet to also exist in the receive
* buffer so we explicitly remove it in case this occurs, otherwise the receive buffer will
* never be empty.
*/
payload = (struct ast_rtp_rtcp_nack_payload *)ast_data_buffer_remove(rtp->recv_buffer, seqno);
if (payload) {
ast_free(payload);
}
rtp->expectedrxseqno++;
if (rtp->expectedrxseqno == SEQNO_CYCLE_OVER) {
rtp->expectedrxseqno = 0;
}
inserted = 1;
2018-05-10 18:11:06 +00:00
continue;
}
payload = (struct ast_rtp_rtcp_nack_payload *)ast_data_buffer_remove(rtp->recv_buffer, rtp->expectedrxseqno);
if (payload) {
frame = ast_frdup(ast_rtp_interpret(instance, srtp, &addr, payload->buf, payload->size, prev_seqno, bundled));
if (frame) {
AST_LIST_INSERT_TAIL(&frames, frame, frame_list);
prev_seqno = rtp->expectedrxseqno;
ast_debug_rtp(2, "(%p) RTP emptying queue and returning packet with sequence number '%d'\n",
instance, frame->seqno);
2018-05-10 18:11:06 +00:00
}
ast_free(payload);
2018-05-10 18:11:06 +00:00
}
rtp->expectedrxseqno++;
if (rtp->expectedrxseqno == SEQNO_CYCLE_OVER) {
rtp->expectedrxseqno = 0;
2018-05-10 18:11:06 +00:00
}
}
2018-05-10 18:11:06 +00:00
if (!inserted) {
/* This current packet goes after them, and we assume that packets going forward will follow
* that new sequence number increment. It is okay for this to not be duplicated as it is guaranteed
* to be the last packet processed right now and it is also guaranteed that it will always return
* non-NULL.
*/
frame = ast_rtp_interpret(instance, srtp, &addr, read_area, res, prev_seqno, bundled);
2018-05-10 18:11:06 +00:00
AST_LIST_INSERT_TAIL(&frames, frame, frame_list);
rtp->expectedrxseqno = seqno + 1;
if (rtp->expectedrxseqno == SEQNO_CYCLE_OVER) {
rtp->expectedrxseqno = 0;
}
2018-05-10 18:11:06 +00:00
ast_debug_rtp(2, "(%p) RTP adding just received packet with sequence number '%d' to end of dumped queue\n",
instance, seqno);
Recorded merge of revisions 254452 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r254452 | mmichelson | 2010-03-25 10:59:56 -0500 (Thu, 25 Mar 2010) | 44 lines Several fixes regarding RFC2833 DTMF detection. Here is a copy and paste of the details from my request on reviewboard that dealt with these changes: Fix 1. The first change in place is to fix Mantis issue 15811, which deals with a situation where Asterisk will incorrectly interpret out of order RFC2833 frames as duplicate DTMF digits. For instance, we would receive a sequence like: seqno 1: DTMF 1 seqno 2: DTMF 1 seqno 3: DTMF 1 seqno 4: DTMF 1 seqno 6: DTMF 1 (end) seqno 5: DTMF 1 seqno 7: DTMF 1 (end) seqno 8: DTMF 1 (end) Prior to this patch when we received the frame with seqno 5, we would interpret this as a new DTMF 1. With this patch, we will check the seqno of the incoming digit and not process the frame if the seqno is lower than the last recorded seqno. Note that we do not record the seqno of the dropped DTMF frame for future processing. While the above situation is what was designed to be fixed, the patch is written in such a way that the following would also be fixed too: seqno 9: DTMF 1 seqno 10: DTMF 1 (end) seqno 11: DTMF 1 (end) seqno 13: DTMF 2 seqno 12: DTMF 1 (end) seqno 14: DTMF 2 seqno 15: DTMF 2 (end) seqno 16: DTMF 2 (end) seqno 17: DTMF 2 (end) In this second situation, the beginning of the DTMF 2 arrives before the final end frame of the DTMF 1. With the patch, seqno 12 is no processed and thus we properly interpret the DTMF. Fix 2. The second change in place is to fix an issue like the following: seqno 1: DTMF 1 seqno 2: DTMF 1 seqno 3: DTMF 1 (end) *packet lost* seqno 4: DTMF 1 (end) *packet lost* seqno 5: DTMF 1 (end) *packet lost* seqno 6: DTMF 2 When we receive seqno 6, we had code in place that was supposed to properly end the previously unended DTMF 1. The problem was that the code was essentially a no-op. The code would set up an end frame for the DTMF 1 but would immediately overwrite the frame with the begin for DTMF 2. I changed process_dtmf_rfc2833() so that instead of returning a single frame, it is given as an output parameter a list of frames. Each frame that needs to be returned is appended to this list. Fix 3. The final change is a minor one where an AST_CONTROL_SRCCHANGE frame could get lost. If we process a cisco DTMF or an RFC 3389 frame and no frame was returned, then we would return &ast_null_frame. The problem is that earlier in the function, we may have generated an AST_CONTROL_SRCCHANGE frame and put it in the list of frames we wish to return. This frame would be lost in such a case. The patch fixes this problem ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@254454 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2010-03-25 16:04:48 +00:00
}
2018-05-10 18:11:06 +00:00
/* When we flush increase our chance for next time by growing the receive buffer when possible
* by how many packets we missed, to give ourselves a bit more breathing room.
*/
ast_data_buffer_resize(rtp->recv_buffer, MIN(MAXIMUM_RTP_RECV_BUFFER_SIZE,
ast_data_buffer_max(rtp->recv_buffer) + AST_VECTOR_SIZE(&rtp->missing_seqno)));
ast_debug_rtp(2, "(%p) RTP receive buffer is now at maximum of %zu\n", instance, ast_data_buffer_max(rtp->recv_buffer));
2018-05-10 18:11:06 +00:00
/* As there is such a large gap we don't want to flood the order side with missing packets, so we
* give up and start anew.
Recorded merge of revisions 254452 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.4 ........ r254452 | mmichelson | 2010-03-25 10:59:56 -0500 (Thu, 25 Mar 2010) | 44 lines Several fixes regarding RFC2833 DTMF detection. Here is a copy and paste of the details from my request on reviewboard that dealt with these changes: Fix 1. The first change in place is to fix Mantis issue 15811, which deals with a situation where Asterisk will incorrectly interpret out of order RFC2833 frames as duplicate DTMF digits. For instance, we would receive a sequence like: seqno 1: DTMF 1 seqno 2: DTMF 1 seqno 3: DTMF 1 seqno 4: DTMF 1 seqno 6: DTMF 1 (end) seqno 5: DTMF 1 seqno 7: DTMF 1 (end) seqno 8: DTMF 1 (end) Prior to this patch when we received the frame with seqno 5, we would interpret this as a new DTMF 1. With this patch, we will check the seqno of the incoming digit and not process the frame if the seqno is lower than the last recorded seqno. Note that we do not record the seqno of the dropped DTMF frame for future processing. While the above situation is what was designed to be fixed, the patch is written in such a way that the following would also be fixed too: seqno 9: DTMF 1 seqno 10: DTMF 1 (end) seqno 11: DTMF 1 (end) seqno 13: DTMF 2 seqno 12: DTMF 1 (end) seqno 14: DTMF 2 seqno 15: DTMF 2 (end) seqno 16: DTMF 2 (end) seqno 17: DTMF 2 (end) In this second situation, the beginning of the DTMF 2 arrives before the final end frame of the DTMF 1. With the patch, seqno 12 is no processed and thus we properly interpret the DTMF. Fix 2. The second change in place is to fix an issue like the following: seqno 1: DTMF 1 seqno 2: DTMF 1 seqno 3: DTMF 1 (end) *packet lost* seqno 4: DTMF 1 (end) *packet lost* seqno 5: DTMF 1 (end) *packet lost* seqno 6: DTMF 2 When we receive seqno 6, we had code in place that was supposed to properly end the previously unended DTMF 1. The problem was that the code was essentially a no-op. The code would set up an end frame for the DTMF 1 but would immediately overwrite the frame with the begin for DTMF 2. I changed process_dtmf_rfc2833() so that instead of returning a single frame, it is given as an output parameter a list of frames. Each frame that needs to be returned is appended to this list. Fix 3. The final change is a minor one where an AST_CONTROL_SRCCHANGE frame could get lost. If we process a cisco DTMF or an RFC 3389 frame and no frame was returned, then we would return &ast_null_frame. The problem is that earlier in the function, we may have generated an AST_CONTROL_SRCCHANGE frame and put it in the list of frames we wish to return. This frame would be lost in such a case. The patch fixes this problem ........ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@254454 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2010-03-25 16:04:48 +00:00
*/
2018-05-10 18:11:06 +00:00
AST_VECTOR_RESET(&rtp->missing_seqno, AST_VECTOR_ELEM_CLEANUP_NOOP);
2018-05-10 18:11:06 +00:00
return AST_LIST_FIRST(&frames);
}
/* We're finished with the frames list */
ast_frame_free(AST_LIST_FIRST(&frames), 0);
/* Determine if the received packet is from the last OLD_PACKET_COUNT (1000 by default) packets or not.
* For the case where the received sequence number exceeds that of the expected sequence number we calculate
* the past sequence number that would be 1000 sequence numbers ago. If the received sequence number
* exceeds or meets that then it is within OLD_PACKET_COUNT packets ago. For example if the expected
* sequence number is 100 and we receive 65530, then it would be considered old. This is because
* 65535 - 1000 + 100 = 64635 which gives us the sequence number at which we would consider the packets
* old. Since 65530 is above that, it would be considered old.
* For the case where the received sequence number is less than the expected sequence number we can do
* a simple subtraction to see if it is 1000 packets ago or not.
*/
if ((seqno < rtp->expectedrxseqno && ((rtp->expectedrxseqno - seqno) <= OLD_PACKET_COUNT)) ||
(seqno > rtp->expectedrxseqno && (seqno >= (65535 - OLD_PACKET_COUNT + rtp->expectedrxseqno)))) {
2018-05-10 18:11:06 +00:00
/* If this is a packet from the past then we have received a duplicate packet, so just drop it */
ast_debug_rtp(2, "(%p) RTP received an old packet with sequence number '%d', dropping it\n",
instance, seqno);
media formats: re-architect handling of media for performance improvements In the old times media formats were represented using a bit field. This was fast but had a few limitations. 1. Asterisk was limited in how many formats it could handle. 2. Formats, being a bit field, could not include any attribute information. A format was strictly its type, e.g., "this is ulaw". This was changed in Asterisk 10 (see https://wiki.asterisk.org/wiki/display/AST/Media+Architecture+Proposal for notes on that work) which led to the creation of the ast_format structure. This structure allowed Asterisk to handle attributes and bundle information with a format. Additionally, ast_format_cap was created to act as a container for multiple formats that, together, formed the capability of some entity. Another mechanism was added to allow logic to be registered which performed format attribute negotiation. Everywhere throughout the codebase Asterisk was changed to use this strategy. Unfortunately, in software, there is no free lunch. These new capabilities came at a cost. Performance analysis and profiling showed that we spend an inordinate amount of time comparing, copying, and generally manipulating formats and their related structures. Basic prototyping has shown that a reasonably large performance improvement could be made in this area. This patch is the result of that project, which overhauled the media format architecture and its usage in Asterisk to improve performance. Generally, the new philosophy for handling formats is as follows: * The ast_format structure is reference counted. This removed a large amount of the memory allocations and copying that was done in prior versions. * In order to prevent race conditions while keeping things performant, the ast_format structure is immutable by convention and lock-free. Violate this tenet at your peril! * Because formats are reference counted, codecs are also reference counted. The Asterisk core generally provides built-in codecs and caches the ast_format structures created to represent them. Generally, to prevent inordinate amounts of module reference bumping, codecs and formats can be added at run-time but cannot be removed. * All compatibility with the bit field representation of codecs/formats has been moved to a compatibility API. The primary user of this representation is chan_iax2, which must continue to maintain its bit-field usage of formats for interoperability concerns. * When a format is negotiated with attributes, or when a format cannot be represented by one of the cached formats, a new format object is created or cloned from an existing format. That format may have the same codec underlying it, but is a different format than a version of the format with different attributes or without attributes. * While formats are reference counted objects, the reference count maintained on the format should be manipulated with care. Formats are generally cached and will persist for the lifetime of Asterisk and do not explicitly need to have their lifetime modified. An exception to this is when the user of a format does not know where the format came from *and* the user may outlive the provider of the format. This occurs, for example, when a format is read from a channel: the channel may have a format with attributes (hence, non-cached) and the user of the format may last longer than the channel (if the reference to the channel is released prior to the format's reference). For more information on this work, see the API design notes: https://wiki.asterisk.org/wiki/display/AST/Media+Format+Rewrite Finally, this work was the culmination of a large number of developer's efforts. Extra thanks goes to Corey Farrell, who took on a large amount of the work in the Asterisk core, chan_sip, and was an invaluable resource in peer reviews throughout this project. There were a substantial number of patches contributed during this work; the following issues/patch names simply reflect some of the work (and will cause the release scripts to give attribution to the individuals who work on them). Reviews: https://reviewboard.asterisk.org/r/3814 https://reviewboard.asterisk.org/r/3808 https://reviewboard.asterisk.org/r/3805 https://reviewboard.asterisk.org/r/3803 https://reviewboard.asterisk.org/r/3801 https://reviewboard.asterisk.org/r/3798 https://reviewboard.asterisk.org/r/3800 https://reviewboard.asterisk.org/r/3794 https://reviewboard.asterisk.org/r/3793 https://reviewboard.asterisk.org/r/3792 https://reviewboard.asterisk.org/r/3791 https://reviewboard.asterisk.org/r/3790 https://reviewboard.asterisk.org/r/3789 https://reviewboard.asterisk.org/r/3788 https://reviewboard.asterisk.org/r/3787 https://reviewboard.asterisk.org/r/3786 https://reviewboard.asterisk.org/r/3784 https://reviewboard.asterisk.org/r/3783 https://reviewboard.asterisk.org/r/3778 https://reviewboard.asterisk.org/r/3774 https://reviewboard.asterisk.org/r/3775 https://reviewboard.asterisk.org/r/3772 https://reviewboard.asterisk.org/r/3761 https://reviewboard.asterisk.org/r/3754 https://reviewboard.asterisk.org/r/3753 https://reviewboard.asterisk.org/r/3751 https://reviewboard.asterisk.org/r/3750 https://reviewboard.asterisk.org/r/3748 https://reviewboard.asterisk.org/r/3747 https://reviewboard.asterisk.org/r/3746 https://reviewboard.asterisk.org/r/3742 https://reviewboard.asterisk.org/r/3740 https://reviewboard.asterisk.org/r/3739 https://reviewboard.asterisk.org/r/3738 https://reviewboard.asterisk.org/r/3737 https://reviewboard.asterisk.org/r/3736 https://reviewboard.asterisk.org/r/3734 https://reviewboard.asterisk.org/r/3722 https://reviewboard.asterisk.org/r/3713 https://reviewboard.asterisk.org/r/3703 https://reviewboard.asterisk.org/r/3689 https://reviewboard.asterisk.org/r/3687 https://reviewboard.asterisk.org/r/3674 https://reviewboard.asterisk.org/r/3671 https://reviewboard.asterisk.org/r/3667 https://reviewboard.asterisk.org/r/3665 https://reviewboard.asterisk.org/r/3625 https://reviewboard.asterisk.org/r/3602 https://reviewboard.asterisk.org/r/3519 https://reviewboard.asterisk.org/r/3518 https://reviewboard.asterisk.org/r/3516 https://reviewboard.asterisk.org/r/3515 https://reviewboard.asterisk.org/r/3512 https://reviewboard.asterisk.org/r/3506 https://reviewboard.asterisk.org/r/3413 https://reviewboard.asterisk.org/r/3410 https://reviewboard.asterisk.org/r/3387 https://reviewboard.asterisk.org/r/3388 https://reviewboard.asterisk.org/r/3389 https://reviewboard.asterisk.org/r/3390 https://reviewboard.asterisk.org/r/3321 https://reviewboard.asterisk.org/r/3320 https://reviewboard.asterisk.org/r/3319 https://reviewboard.asterisk.org/r/3318 https://reviewboard.asterisk.org/r/3266 https://reviewboard.asterisk.org/r/3265 https://reviewboard.asterisk.org/r/3234 https://reviewboard.asterisk.org/r/3178 ASTERISK-23114 #close Reported by: mjordan media_formats_translation_core.diff uploaded by kharwell (License 6464) rb3506.diff uploaded by mjordan (License 6283) media_format_app_file.diff uploaded by kharwell (License 6464) misc-2.diff uploaded by file (License 5000) chan_mild-3.diff uploaded by file (License 5000) chan_obscure.diff uploaded by file (License 5000) jingle.diff uploaded by file (License 5000) funcs.diff uploaded by file (License 5000) formats.diff uploaded by file (License 5000) core.diff uploaded by file (License 5000) bridges.diff uploaded by file (License 5000) mf-codecs-2.diff uploaded by file (License 5000) mf-app_fax.diff uploaded by file (License 5000) mf-apps-3.diff uploaded by file (License 5000) media-formats-3.diff uploaded by file (License 5000) ASTERISK-23715 rb3713.patch uploaded by coreyfarrell (License 5909) rb3689.patch uploaded by mjordan (License 6283) ASTERISK-23957 rb3722.patch uploaded by mjordan (License 6283) mf-attributes-3.diff uploaded by file (License 5000) ASTERISK-23958 Tested by: jrose rb3822.patch uploaded by coreyfarrell (License 5909) rb3800.patch uploaded by jrose (License 6182) chan_sip.diff uploaded by mjordan (License 6283) rb3747.patch uploaded by jrose (License 6182) ASTERISK-23959 #close Tested by: sgriepentrog, mjordan, coreyfarrell sip_cleanup.diff uploaded by opticron (License 6273) chan_sip_caps.diff uploaded by mjordan (License 6283) rb3751.patch uploaded by coreyfarrell (License 5909) chan_sip-3.diff uploaded by file (License 5000) ASTERISK-23960 #close Tested by: opticron direct_media.diff uploaded by opticron (License 6273) pjsip-direct-media.diff uploaded by file (License 5000) format_cap_remove.diff uploaded by opticron (License 6273) media_format_fixes.diff uploaded by opticron (License 6273) chan_pjsip-2.diff uploaded by file (License 5000) ASTERISK-23966 #close Tested by: rmudgett rb3803.patch uploaded by rmudgetti (License 5621) chan_dahdi.diff uploaded by file (License 5000) ASTERISK-24064 #close Tested by: coreyfarrell, mjordan, opticron, file, rmudgett, sgriepentrog, jrose rb3814.patch uploaded by rmudgett (License 5621) moh_cleanup.diff uploaded by opticron (License 6273) bridge_leak.diff uploaded by opticron (License 6273) translate.diff uploaded by file (License 5000) rb3795.patch uploaded by rmudgett (License 5621) tls_fix.diff uploaded by mjordan (License 6283) fax-mf-fix-2.diff uploaded by file (License 5000) rtp_transfer_stuff uploaded by mjordan (License 6283) rb3787.patch uploaded by rmudgett (License 5621) media-formats-explicit-translate-format-3.diff uploaded by file (License 5000) format_cache_case_fix.diff uploaded by opticron (License 6273) rb3774.patch uploaded by rmudgett (License 5621) rb3775.patch uploaded by rmudgett (License 5621) rtp_engine_fix.diff uploaded by opticron (License 6273) rtp_crash_fix.diff uploaded by opticron (License 6273) rb3753.patch uploaded by mjordan (License 6283) rb3750.patch uploaded by mjordan (License 6283) rb3748.patch uploaded by rmudgett (License 5621) media_format_fixes.diff uploaded by opticron (License 6273) rb3740.patch uploaded by mjordan (License 6283) rb3739.patch uploaded by mjordan (License 6283) rb3734.patch uploaded by mjordan (License 6283) rb3689.patch uploaded by mjordan (License 6283) rb3674.patch uploaded by coreyfarrell (License 5909) rb3671.patch uploaded by coreyfarrell (License 5909) rb3667.patch uploaded by coreyfarrell (License 5909) rb3665.patch uploaded by mjordan (License 6283) rb3625.patch uploaded by coreyfarrell (License 5909) rb3602.patch uploaded by coreyfarrell (License 5909) format_compatibility-2.diff uploaded by file (License 5000) core.diff uploaded by file (License 5000) git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@419044 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-07-20 22:06:33 +00:00
return &ast_null_frame;
2018-05-10 18:11:06 +00:00
} else if (ast_data_buffer_get(rtp->recv_buffer, seqno)) {
/* If this is a packet we already have buffered then it is a duplicate, so just drop it */
ast_debug_rtp(2, "(%p) RTP received a duplicate transmission of packet with sequence number '%d', dropping it\n",
instance, seqno);
2018-05-10 18:11:06 +00:00
return &ast_null_frame;
} else {
/* This is an out of order packet from the future */
struct ast_rtp_rtcp_nack_payload *payload;
int missing_seqno;
int remove_failed;
unsigned int missing_seqnos_added = 0;
2018-05-10 18:11:06 +00:00
ast_debug_rtp(2, "(%p) RTP received an out of order packet with sequence number '%d' while expecting '%d' from the future\n",
instance, seqno, rtp->expectedrxseqno);
2018-05-10 18:11:06 +00:00
payload = ast_malloc(sizeof(*payload) + res);
if (!payload) {
/* If the payload can't be allocated then we can't defer this packet right now.
* Instead of dumping what we have we pretend we lost this packet. It will then
* get NACKed later or the existing buffer will be returned entirely. Well, we may
* try since we're seemingly out of memory. It's a bad situation all around and
* packets are likely to get lost anyway.
*/
return &ast_null_frame;
}
2018-05-10 18:11:06 +00:00
payload->size = res;
memcpy(payload->buf, rtpheader, res);
if (ast_data_buffer_put(rtp->recv_buffer, seqno, payload) == -1) {
ast_free(payload);
}
/* If this sequence number is removed that means we had a gap and this packet has filled it in
* some. Since it was part of the gap we will have already added any other missing sequence numbers
* before it (and possibly after it) to the vector so we don't need to do that again. Note that
* remove_failed will be set to -1 if the sequence number isn't removed, and 0 if it is.
*/
remove_failed = AST_VECTOR_REMOVE_CMP_ORDERED(&rtp->missing_seqno, seqno, find_by_value,
2018-05-10 18:11:06 +00:00
AST_VECTOR_ELEM_CLEANUP_NOOP);
if (!remove_failed) {
ast_debug_rtp(2, "(%p) RTP packet with sequence number '%d' is no longer missing\n",
instance, seqno);
}
/* The missing sequence number code works by taking the sequence number of the
* packet we've just received and going backwards until we hit the sequence number
* of the last packet we've received. While doing so we check to make sure that the
* sequence number is not already missing and that it is not already buffered.
*/
missing_seqno = seqno;
while (remove_failed) {
missing_seqno -= 1;
/* If we've cycled backwards then start back at the top */
if (missing_seqno < 0) {
missing_seqno = 65535;
}
/* We've gone backwards enough such that we've hit the previous sequence number */
if (missing_seqno == prev_seqno) {
break;
}
2018-05-10 18:11:06 +00:00
/* We don't want missing sequence number duplicates. If, for some reason,
* packets are really out of order, we could end up in this scenario:
*
* We are expecting sequence number 100
* We receive sequence number 105
* Sequence numbers 100 through 104 get added to the vector
* We receive sequence number 101 (this section is skipped)
* We receive sequence number 103
* Sequence number 102 is added to the vector
*
* This will prevent the duplicate from being added.
*/
if (AST_VECTOR_GET_CMP(&rtp->missing_seqno, missing_seqno,
2018-05-10 18:11:06 +00:00
find_by_value)) {
continue;
}
/* If this packet has been buffered already then don't count it amongst the
* missing.
*/
if (ast_data_buffer_get(rtp->recv_buffer, missing_seqno)) {
2018-05-10 18:11:06 +00:00
continue;
}
ast_debug_rtp(2, "(%p) RTP added missing sequence number '%d'\n",
instance, missing_seqno);
AST_VECTOR_ADD_SORTED(&rtp->missing_seqno, missing_seqno,
2018-05-10 18:11:06 +00:00
compare_by_value);
missing_seqnos_added++;
2018-05-10 18:11:06 +00:00
}
/* When we add a large number of missing sequence numbers we assume there was a substantial
* gap in reception so we trigger an immediate NACK. When our data buffer is 1/4 full we
* assume that the packets aren't just out of order but have actually been lost. At 1/2
* full we get more aggressive and ask for retransmission when we get a new packet.
* To get them back we construct and send a NACK causing the sender to retransmit them.
2018-05-10 18:11:06 +00:00
*/
if (missing_seqnos_added >= MISSING_SEQNOS_ADDED_TRIGGER ||
ast_data_buffer_count(rtp->recv_buffer) == ast_data_buffer_max(rtp->recv_buffer) / 4 ||
ast_data_buffer_count(rtp->recv_buffer) >= ast_data_buffer_max(rtp->recv_buffer) / 2) {
2018-05-10 18:11:06 +00:00
int packet_len = 0;
int res = 0;
int ice;
int sr;
size_t data_size = AST_UUID_STR_LEN + 128 + (AST_VECTOR_SIZE(&rtp->missing_seqno) * 4);
2018-05-10 18:11:06 +00:00
RAII_VAR(unsigned char *, rtcpheader, NULL, ast_free_ptr);
RAII_VAR(struct ast_rtp_rtcp_report *, rtcp_report,
ast_rtp_rtcp_report_alloc(rtp->themssrc_valid ? 1 : 0),
ao2_cleanup);
/* Sufficient space for RTCP headers and report, SDES with CNAME, NACK header,
* and worst case 4 bytes per missing sequence number.
*/
2018-05-10 18:11:06 +00:00
rtcpheader = ast_malloc(sizeof(*rtcpheader) + data_size);
if (!rtcpheader) {
ast_debug_rtcp(1, "(%p) RTCP failed to allocate memory for NACK\n", instance);
2018-05-10 18:11:06 +00:00
return &ast_null_frame;
}
2018-05-10 18:11:06 +00:00
memset(rtcpheader, 0, data_size);
res = ast_rtcp_generate_compound_prefix(instance, rtcpheader, rtcp_report, &sr);
2018-05-10 18:11:06 +00:00
if (res == 0 || res == 1) {
return &ast_null_frame;
}
2018-05-10 18:11:06 +00:00
packet_len += res;
2018-05-10 18:11:06 +00:00
res = ast_rtcp_generate_nack(instance, rtcpheader + packet_len);
2018-05-10 18:11:06 +00:00
if (res == 0) {
ast_debug_rtcp(1, "(%p) RTCP failed to construct NACK, stopping here\n", instance);
2018-05-10 18:11:06 +00:00
return &ast_null_frame;
}
2018-05-10 18:11:06 +00:00
packet_len += res;
2018-05-10 18:11:06 +00:00
res = rtcp_sendto(instance, rtcpheader, packet_len, 0, &remote_address, &ice);
if (res < 0) {
ast_debug_rtcp(1, "(%p) RTCP failed to send NACK request out\n", instance);
2018-05-10 18:11:06 +00:00
} else {
ast_debug_rtcp(2, "(%p) RTCP sending a NACK request to get missing packets\n", instance);
2018-05-10 18:11:06 +00:00
/* Update RTCP SR/RR statistics */
ast_rtcp_calculate_sr_rr_statistics(instance, rtcp_report, remote_address, ice, sr);
}
}
}
2018-05-10 18:11:06 +00:00
return &ast_null_frame;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static void ast_rtp_prop_set(struct ast_rtp_instance *instance, enum ast_rtp_property property, int value)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
if (property == AST_RTP_PROPERTY_RTCP) {
if (value) {
struct ast_sockaddr local_addr;
if (rtp->rtcp && rtp->rtcp->type == value) {
ast_debug_rtcp(1, "(%p) RTCP ignoring duplicate property\n", instance);
return;
}
if (!rtp->rtcp) {
rtp->rtcp = ast_calloc(1, sizeof(*rtp->rtcp));
if (!rtp->rtcp) {
return;
}
rtp->rtcp->s = -1;
#if defined(HAVE_OPENSSL) && (OPENSSL_VERSION_NUMBER >= 0x10001000L) && !defined(OPENSSL_NO_SRTP)
rtp->rtcp->dtls.timeout_timer = -1;
#endif
rtp->rtcp->schedid = -1;
}
rtp->rtcp->type = value;
/* Grab the IP address and port we are going to use */
ast_rtp_instance_get_local_address(instance, &rtp->rtcp->us);
if (value == AST_RTP_INSTANCE_RTCP_STANDARD) {
ast_sockaddr_set_port(&rtp->rtcp->us,
ast_sockaddr_port(&rtp->rtcp->us) + 1);
}
ast_sockaddr_copy(&local_addr, &rtp->rtcp->us);
if (!ast_find_ourip(&local_addr, &rtp->rtcp->us, 0)) {
ast_sockaddr_set_port(&local_addr, ast_sockaddr_port(&rtp->rtcp->us));
} else {
/* Failed to get local address reset to use default. */
ast_sockaddr_copy(&local_addr, &rtp->rtcp->us);
}
ast_free(rtp->rtcp->local_addr_str);
rtp->rtcp->local_addr_str = ast_strdup(ast_sockaddr_stringify(&local_addr));
if (!rtp->rtcp->local_addr_str) {
ast_free(rtp->rtcp);
rtp->rtcp = NULL;
return;
}
if (value == AST_RTP_INSTANCE_RTCP_STANDARD) {
/* We're either setting up RTCP from scratch or
* switching from MUX. Either way, we won't have
* a socket set up, and we need to set it up
*/
if ((rtp->rtcp->s =
create_new_socket("RTCP",
ast_sockaddr_is_ipv4(&rtp->rtcp->us) ?
AF_INET :
ast_sockaddr_is_ipv6(&rtp->rtcp->us) ?
AF_INET6 : -1)) < 0) {
ast_debug_rtcp(1, "(%p) RTCP failed to create a new socket\n", instance);
ast_free(rtp->rtcp->local_addr_str);
ast_free(rtp->rtcp);
rtp->rtcp = NULL;
return;
}
/* Try to actually bind to the IP address and port we are going to use for RTCP, if this fails we have to bail out */
if (ast_bind(rtp->rtcp->s, &rtp->rtcp->us)) {
ast_debug_rtcp(1, "(%p) RTCP failed to setup RTP instance\n", instance);
close(rtp->rtcp->s);
ast_free(rtp->rtcp->local_addr_str);
ast_free(rtp->rtcp);
rtp->rtcp = NULL;
return;
}
#ifdef HAVE_PJPROJECT
if (rtp->ice) {
rtp_add_candidates_to_ice(instance, rtp, &rtp->rtcp->us, ast_sockaddr_port(&rtp->rtcp->us), AST_RTP_ICE_COMPONENT_RTCP, TRANSPORT_SOCKET_RTCP);
}
#endif
#if defined(HAVE_OPENSSL) && (OPENSSL_VERSION_NUMBER >= 0x10001000L) && !defined(OPENSSL_NO_SRTP)
dtls_setup_rtcp(instance);
#endif
} else {
struct ast_sockaddr addr;
/* RTCPMUX uses the same socket as RTP. If we were previously using standard RTCP
* then close the socket we previously created.
*
* It may seem as though there is a possible race condition here where we might try
* to close the RTCP socket while it is being used to send data. However, this is not
* a problem in practice since setting and adjusting of RTCP properties happens prior
* to activating RTP. It is not until RTP is activated that timers start for RTCP
* transmission
*/
if (rtp->rtcp->s > -1 && rtp->rtcp->s != rtp->s) {
close(rtp->rtcp->s);
}
rtp->rtcp->s = rtp->s;
ast_rtp_instance_get_remote_address(instance, &addr);
ast_sockaddr_copy(&rtp->rtcp->them, &addr);
#if defined(HAVE_OPENSSL) && (OPENSSL_VERSION_NUMBER >= 0x10001000L) && !defined(OPENSSL_NO_SRTP)
if (rtp->rtcp->dtls.ssl && rtp->rtcp->dtls.ssl != rtp->dtls.ssl) {
SSL_free(rtp->rtcp->dtls.ssl);
}
rtp->rtcp->dtls.ssl = rtp->dtls.ssl;
#endif
}
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
ast_debug_rtcp(1, "(%s) RTCP setup on RTP instance\n",
ast_rtp_instance_get_channel_id(instance));
} else {
if (rtp->rtcp) {
if (rtp->rtcp->schedid > -1) {
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_unlock(instance);
if (!ast_sched_del(rtp->sched, rtp->rtcp->schedid)) {
/* Successfully cancelled scheduler entry. */
ao2_ref(instance, -1);
} else {
/* Unable to cancel scheduler entry */
ast_debug_rtcp(1, "(%p) RTCP failed to tear down RTCP\n", instance);
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_lock(instance);
return;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_lock(instance);
rtp->rtcp->schedid = -1;
}
if (rtp->transport_wide_cc.schedid > -1) {
ao2_unlock(instance);
if (!ast_sched_del(rtp->sched, rtp->transport_wide_cc.schedid)) {
ao2_ref(instance, -1);
} else {
ast_debug_rtcp(1, "(%p) RTCP failed to tear down transport-cc feedback\n", instance);
ao2_lock(instance);
return;
}
ao2_lock(instance);
rtp->transport_wide_cc.schedid = -1;
}
if (rtp->rtcp->s > -1 && rtp->rtcp->s != rtp->s) {
close(rtp->rtcp->s);
}
#if defined(HAVE_OPENSSL) && (OPENSSL_VERSION_NUMBER >= 0x10001000L) && !defined(OPENSSL_NO_SRTP)
ao2_unlock(instance);
dtls_srtp_stop_timeout_timer(instance, rtp, 1);
ao2_lock(instance);
if (rtp->rtcp->dtls.ssl && rtp->rtcp->dtls.ssl != rtp->dtls.ssl) {
SSL_free(rtp->rtcp->dtls.ssl);
}
#endif
ast_free(rtp->rtcp->local_addr_str);
ast_free(rtp->rtcp);
rtp->rtcp = NULL;
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
ast_debug_rtcp(1, "(%s) RTCP torn down on RTP instance\n",
ast_rtp_instance_get_channel_id(instance));
}
}
} else if (property == AST_RTP_PROPERTY_ASYMMETRIC_CODEC) {
rtp->asymmetric_codec = value;
} else if (property == AST_RTP_PROPERTY_RETRANS_SEND) {
if (value) {
if (!rtp->send_buffer) {
rtp->send_buffer = ast_data_buffer_alloc(ast_free_ptr, DEFAULT_RTP_SEND_BUFFER_SIZE);
}
} else {
if (rtp->send_buffer) {
ast_data_buffer_free(rtp->send_buffer);
rtp->send_buffer = NULL;
}
}
2018-05-10 18:11:06 +00:00
} else if (property == AST_RTP_PROPERTY_RETRANS_RECV) {
if (value) {
if (!rtp->recv_buffer) {
rtp->recv_buffer = ast_data_buffer_alloc(ast_free_ptr, DEFAULT_RTP_RECV_BUFFER_SIZE);
AST_VECTOR_INIT(&rtp->missing_seqno, 0);
}
} else {
if (rtp->recv_buffer) {
ast_data_buffer_free(rtp->recv_buffer);
rtp->recv_buffer = NULL;
AST_VECTOR_FREE(&rtp->missing_seqno);
}
}
}
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static int ast_rtp_fd(struct ast_rtp_instance *instance, int rtcp)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
return rtcp ? (rtp->rtcp ? rtp->rtcp->s : -1) : rtp->s;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static void ast_rtp_remote_address_set(struct ast_rtp_instance *instance, struct ast_sockaddr *addr)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
struct ast_sockaddr local;
int index;
ast_rtp_instance_get_local_address(instance, &local);
if (!ast_sockaddr_isnull(addr)) {
/* Update the local RTP address with what is being used */
if (ast_ouraddrfor(addr, &local)) {
/* Failed to update our address so reuse old local address */
ast_rtp_instance_get_local_address(instance, &local);
} else {
ast_rtp_instance_set_local_address(instance, &local);
}
}
if (rtp->rtcp && !ast_sockaddr_isnull(addr)) {
ast_debug_rtcp(1, "(%p) RTCP setting address on RTP instance\n", instance);
ast_sockaddr_copy(&rtp->rtcp->them, addr);
if (rtp->rtcp->type == AST_RTP_INSTANCE_RTCP_STANDARD) {
ast_sockaddr_set_port(&rtp->rtcp->them, ast_sockaddr_port(addr) + 1);
/* Update the local RTCP address with what is being used */
ast_sockaddr_set_port(&local, ast_sockaddr_port(&local) + 1);
}
ast_sockaddr_copy(&rtp->rtcp->us, &local);
ast_free(rtp->rtcp->local_addr_str);
rtp->rtcp->local_addr_str = ast_strdup(ast_sockaddr_stringify(&local));
}
/* Update any bundled RTP instances */
for (index = 0; index < AST_VECTOR_SIZE(&rtp->ssrc_mapping); ++index) {
struct rtp_ssrc_mapping *mapping = AST_VECTOR_GET_ADDR(&rtp->ssrc_mapping, index);
ast_rtp_instance_set_remote_address(mapping->instance, addr);
}
/* Need to reset the DTMF last sequence number and the timestamp of the last END packet */
rtp->last_seqno = 0;
rtp->last_end_timestamp.ts = 0;
rtp->last_end_timestamp.is_set = 0;
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
if (strictrtp && rtp->strict_rtp_state != STRICT_RTP_OPEN
&& !ast_sockaddr_isnull(addr) && ast_sockaddr_cmp(addr, &rtp->strict_rtp_address)) {
/* We only need to learn a new strict source address if we've been told the source is
* changing to something different.
*/
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
ast_verb(4, "%p -- Strict RTP learning after remote address set to: %s\n",
rtp, ast_sockaddr_stringify(addr));
rtp_learning_start(rtp);
}
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*!
* \brief Write t140 redundancy frame
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
*
* \param data primary data to be buffered
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
*
* Scheduler callback
*/
static int red_write(const void *data)
{
struct ast_rtp_instance *instance = (struct ast_rtp_instance*) data;
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_lock(instance);
if (rtp->red->t140.datalen > 0) {
ast_rtp_write(instance, &rtp->red->t140);
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_unlock(instance);
return 1;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static int rtp_red_init(struct ast_rtp_instance *instance, int buffer_time, int *payloads, int generations)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
int x;
rtp->red = ast_calloc(1, sizeof(*rtp->red));
if (!rtp->red) {
return -1;
}
rtp->red->t140.frametype = AST_FRAME_TEXT;
rtp->red->t140.subclass.format = ast_format_t140_red;
rtp->red->t140.data.ptr = &rtp->red->buf_data;
rtp->red->t140red = rtp->red->t140;
rtp->red->t140red.data.ptr = &rtp->red->t140red_data;
rtp->red->ti = buffer_time;
rtp->red->num_gen = generations;
rtp->red->hdrlen = generations * 4 + 1;
for (x = 0; x < generations; x++) {
rtp->red->pt[x] = payloads[x];
rtp->red->pt[x] |= 1 << 7; /* mark redundant generations pt */
rtp->red->t140red_data[x*4] = rtp->red->pt[x];
}
rtp->red->t140red_data[x*4] = rtp->red->pt[x] = payloads[x]; /* primary pt */
rtp->red->schedid = ast_sched_add(rtp->sched, generations, red_write, instance);
return 0;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static int rtp_red_buffer(struct ast_rtp_instance *instance, struct ast_frame *frame)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
struct rtp_red *red = rtp->red;
if (!red) {
return 0;
}
if (frame->datalen > 0) {
if (red->t140.datalen > 0) {
const unsigned char *primary = red->buf_data;
/* There is something already in the T.140 buffer */
if (primary[0] == 0x08 || primary[0] == 0x0a || primary[0] == 0x0d) {
/* Flush the previous T.140 packet if it is a command */
ast_rtp_write(instance, &rtp->red->t140);
} else {
primary = frame->data.ptr;
if (primary[0] == 0x08 || primary[0] == 0x0a || primary[0] == 0x0d) {
/* Flush the previous T.140 packet if we are buffering a command now */
ast_rtp_write(instance, &rtp->red->t140);
}
}
}
memcpy(&red->buf_data[red->t140.datalen], frame->data.ptr, frame->datalen);
red->t140.datalen += frame->datalen;
red->t140.ts = frame->ts;
}
return 0;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre Neither instance0 nor instance1 are locked */
static int ast_rtp_local_bridge(struct ast_rtp_instance *instance0, struct ast_rtp_instance *instance1)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance0);
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_lock(instance0);
ast_set_flag(rtp, FLAG_NEED_MARKER_BIT | FLAG_REQ_LOCAL_BRIDGE_BIT);
if (rtp->smoother) {
ast_smoother_free(rtp->smoother);
rtp->smoother = NULL;
}
/* We must use a new SSRC when local bridge ends */
if (!instance1) {
rtp->ssrc = rtp->ssrc_orig;
rtp->ssrc_orig = 0;
rtp->ssrc_saved = 0;
} else if (!rtp->ssrc_saved) {
/* In case ast_rtp_local_bridge is called multiple times, only save the ssrc from before local bridge began */
rtp->ssrc_orig = rtp->ssrc;
rtp->ssrc_saved = 1;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_unlock(instance0);
return 0;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static int ast_rtp_get_stat(struct ast_rtp_instance *instance, struct ast_rtp_instance_stats *stats, enum ast_rtp_instance_stat stat)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
if (!rtp->rtcp) {
return -1;
}
AST_RTP_STAT_SET(AST_RTP_INSTANCE_STAT_TXCOUNT, -1, stats->txcount, rtp->txcount);
AST_RTP_STAT_SET(AST_RTP_INSTANCE_STAT_RXCOUNT, -1, stats->rxcount, rtp->rxcount);
AST_RTP_STAT_SET(AST_RTP_INSTANCE_STAT_TXOCTETCOUNT, -1, stats->txoctetcount, rtp->txoctetcount);
AST_RTP_STAT_SET(AST_RTP_INSTANCE_STAT_RXOCTETCOUNT, -1, stats->rxoctetcount, rtp->rxoctetcount);
AST_RTP_STAT_SET(AST_RTP_INSTANCE_STAT_TXPLOSS, AST_RTP_INSTANCE_STAT_COMBINED_LOSS, stats->txploss, rtp->rtcp->reported_lost);
AST_RTP_STAT_SET(AST_RTP_INSTANCE_STAT_RXPLOSS, AST_RTP_INSTANCE_STAT_COMBINED_LOSS, stats->rxploss, rtp->rtcp->expected_prior - rtp->rtcp->received_prior);
AST_RTP_STAT_SET(AST_RTP_INSTANCE_STAT_REMOTE_MAXRXPLOSS, AST_RTP_INSTANCE_STAT_COMBINED_LOSS, stats->remote_maxrxploss, rtp->rtcp->reported_maxlost);
AST_RTP_STAT_SET(AST_RTP_INSTANCE_STAT_REMOTE_MINRXPLOSS, AST_RTP_INSTANCE_STAT_COMBINED_LOSS, stats->remote_minrxploss, rtp->rtcp->reported_minlost);
AST_RTP_STAT_SET(AST_RTP_INSTANCE_STAT_REMOTE_NORMDEVRXPLOSS, AST_RTP_INSTANCE_STAT_COMBINED_LOSS, stats->remote_normdevrxploss, rtp->rtcp->reported_normdev_lost);
AST_RTP_STAT_SET(AST_RTP_INSTANCE_STAT_REMOTE_STDEVRXPLOSS, AST_RTP_INSTANCE_STAT_COMBINED_LOSS, stats->remote_stdevrxploss, rtp->rtcp->reported_stdev_lost);
AST_RTP_STAT_SET(AST_RTP_INSTANCE_STAT_LOCAL_MAXRXPLOSS, AST_RTP_INSTANCE_STAT_COMBINED_LOSS, stats->local_maxrxploss, rtp->rtcp->maxrxlost);
AST_RTP_STAT_SET(AST_RTP_INSTANCE_STAT_LOCAL_MINRXPLOSS, AST_RTP_INSTANCE_STAT_COMBINED_LOSS, stats->local_minrxploss, rtp->rtcp->minrxlost);
AST_RTP_STAT_SET(AST_RTP_INSTANCE_STAT_LOCAL_NORMDEVRXPLOSS, AST_RTP_INSTANCE_STAT_COMBINED_LOSS, stats->local_normdevrxploss, rtp->rtcp->normdev_rxlost);
AST_RTP_STAT_SET(AST_RTP_INSTANCE_STAT_LOCAL_STDEVRXPLOSS, AST_RTP_INSTANCE_STAT_COMBINED_LOSS, stats->local_stdevrxploss, rtp->rtcp->stdev_rxlost);
AST_RTP_STAT_TERMINATOR(AST_RTP_INSTANCE_STAT_COMBINED_LOSS);
AST_RTP_STAT_SET(AST_RTP_INSTANCE_STAT_TXJITTER, AST_RTP_INSTANCE_STAT_COMBINED_JITTER, stats->txjitter, rtp->rxjitter);
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
AST_RTP_STAT_SET(AST_RTP_INSTANCE_STAT_RXJITTER, AST_RTP_INSTANCE_STAT_COMBINED_JITTER, stats->rxjitter, rtp->rtcp->reported_jitter);
AST_RTP_STAT_SET(AST_RTP_INSTANCE_STAT_REMOTE_MAXJITTER, AST_RTP_INSTANCE_STAT_COMBINED_JITTER, stats->remote_maxjitter, rtp->rtcp->reported_maxjitter);
AST_RTP_STAT_SET(AST_RTP_INSTANCE_STAT_REMOTE_MINJITTER, AST_RTP_INSTANCE_STAT_COMBINED_JITTER, stats->remote_minjitter, rtp->rtcp->reported_minjitter);
AST_RTP_STAT_SET(AST_RTP_INSTANCE_STAT_REMOTE_NORMDEVJITTER, AST_RTP_INSTANCE_STAT_COMBINED_JITTER, stats->remote_normdevjitter, rtp->rtcp->reported_normdev_jitter);
AST_RTP_STAT_SET(AST_RTP_INSTANCE_STAT_REMOTE_STDEVJITTER, AST_RTP_INSTANCE_STAT_COMBINED_JITTER, stats->remote_stdevjitter, rtp->rtcp->reported_stdev_jitter);
AST_RTP_STAT_SET(AST_RTP_INSTANCE_STAT_LOCAL_MAXJITTER, AST_RTP_INSTANCE_STAT_COMBINED_JITTER, stats->local_maxjitter, rtp->rtcp->maxrxjitter);
AST_RTP_STAT_SET(AST_RTP_INSTANCE_STAT_LOCAL_MINJITTER, AST_RTP_INSTANCE_STAT_COMBINED_JITTER, stats->local_minjitter, rtp->rtcp->minrxjitter);
AST_RTP_STAT_SET(AST_RTP_INSTANCE_STAT_LOCAL_NORMDEVJITTER, AST_RTP_INSTANCE_STAT_COMBINED_JITTER, stats->local_normdevjitter, rtp->rtcp->normdev_rxjitter);
AST_RTP_STAT_SET(AST_RTP_INSTANCE_STAT_LOCAL_STDEVJITTER, AST_RTP_INSTANCE_STAT_COMBINED_JITTER, stats->local_stdevjitter, rtp->rtcp->stdev_rxjitter);
AST_RTP_STAT_TERMINATOR(AST_RTP_INSTANCE_STAT_COMBINED_JITTER);
AST_RTP_STAT_SET(AST_RTP_INSTANCE_STAT_RTT, AST_RTP_INSTANCE_STAT_COMBINED_RTT, stats->rtt, rtp->rtcp->rtt);
AST_RTP_STAT_SET(AST_RTP_INSTANCE_STAT_MAX_RTT, AST_RTP_INSTANCE_STAT_COMBINED_RTT, stats->maxrtt, rtp->rtcp->maxrtt);
AST_RTP_STAT_SET(AST_RTP_INSTANCE_STAT_MIN_RTT, AST_RTP_INSTANCE_STAT_COMBINED_RTT, stats->minrtt, rtp->rtcp->minrtt);
AST_RTP_STAT_SET(AST_RTP_INSTANCE_STAT_NORMDEVRTT, AST_RTP_INSTANCE_STAT_COMBINED_RTT, stats->normdevrtt, rtp->rtcp->normdevrtt);
AST_RTP_STAT_SET(AST_RTP_INSTANCE_STAT_STDEVRTT, AST_RTP_INSTANCE_STAT_COMBINED_RTT, stats->stdevrtt, rtp->rtcp->stdevrtt);
AST_RTP_STAT_TERMINATOR(AST_RTP_INSTANCE_STAT_COMBINED_RTT);
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
AST_RTP_STAT_SET(AST_RTP_INSTANCE_STAT_TXMES, AST_RTP_INSTANCE_STAT_COMBINED_MES, stats->txmes, rtp->rxmes);
AST_RTP_STAT_SET(AST_RTP_INSTANCE_STAT_RXMES, AST_RTP_INSTANCE_STAT_COMBINED_MES, stats->rxmes, rtp->rtcp->reported_mes);
AST_RTP_STAT_SET(AST_RTP_INSTANCE_STAT_REMOTE_MAXMES, AST_RTP_INSTANCE_STAT_COMBINED_MES, stats->remote_maxmes, rtp->rtcp->reported_maxmes);
AST_RTP_STAT_SET(AST_RTP_INSTANCE_STAT_REMOTE_MINMES, AST_RTP_INSTANCE_STAT_COMBINED_MES, stats->remote_minmes, rtp->rtcp->reported_minmes);
AST_RTP_STAT_SET(AST_RTP_INSTANCE_STAT_REMOTE_NORMDEVMES, AST_RTP_INSTANCE_STAT_COMBINED_MES, stats->remote_normdevmes, rtp->rtcp->reported_normdev_mes);
AST_RTP_STAT_SET(AST_RTP_INSTANCE_STAT_REMOTE_STDEVMES, AST_RTP_INSTANCE_STAT_COMBINED_MES, stats->remote_stdevmes, rtp->rtcp->reported_stdev_mes);
AST_RTP_STAT_SET(AST_RTP_INSTANCE_STAT_LOCAL_MAXMES, AST_RTP_INSTANCE_STAT_COMBINED_MES, stats->local_maxmes, rtp->rtcp->maxrxmes);
AST_RTP_STAT_SET(AST_RTP_INSTANCE_STAT_LOCAL_MINMES, AST_RTP_INSTANCE_STAT_COMBINED_MES, stats->local_minmes, rtp->rtcp->minrxmes);
AST_RTP_STAT_SET(AST_RTP_INSTANCE_STAT_LOCAL_NORMDEVMES, AST_RTP_INSTANCE_STAT_COMBINED_MES, stats->local_normdevmes, rtp->rtcp->normdev_rxmes);
AST_RTP_STAT_SET(AST_RTP_INSTANCE_STAT_LOCAL_STDEVMES, AST_RTP_INSTANCE_STAT_COMBINED_MES, stats->local_stdevmes, rtp->rtcp->stdev_rxjitter);
AST_RTP_STAT_TERMINATOR(AST_RTP_INSTANCE_STAT_COMBINED_MES);
AST_RTP_STAT_SET(AST_RTP_INSTANCE_STAT_LOCAL_SSRC, -1, stats->local_ssrc, rtp->ssrc);
AST_RTP_STAT_SET(AST_RTP_INSTANCE_STAT_REMOTE_SSRC, -1, stats->remote_ssrc, rtp->themssrc);
Refactor RTCP events over to Stasis; associate with channels This patch does the following: * It merges Jaco Kroon's patch from ASTERISK-20754, which provides channel information in the RTCP events. Because Stasis provides a cache, Jaco's patch was modified to pass the channel uniqueid to the RTP layer as opposed to a pointer to the channel. This has the following benefits: (1) It keeps the RTP engine 'clean' of references back to channels (2) It prevents circular dependencies and other potential ref counting issues * The RTP engine now allows any RTP implementation to raise RTCP messages. Potentially, other implementations (such as res_rtp_multicast) could also raise RTCP information. The engine provides structs to represent RTCP headers and RTCP SR/RR reports. * Some general refactoring in res_rtp_asterisk was done to try and tame the RTCP code. It isn't perfect - that's *way* beyond the scope of this work - but it does feel marginally better. * A few random bugs were fixed in the RTCP statistics. (Example: performing an assignment of a = a is probably not correct) * We now raise RTCP events for each SR/RR sent/received. Previously we wouldn't raise an event when we sent a RR report. Note that this work will be of use to others who want to monitor call quality or build modules that report call quality statistics. Since the events are now moving across the Stasis message bus, this is far easier to accomplish. It is also a first step (though by no means the last step) towards getting Olle's pinefrog work incorporated. Again: note that the patch by Jaco Kroon was modified slightly for this work; however, he did all of the hard work in finding the right places to set the channel in the RTP engine across the channel drivers. Much thanks goes to Jaco for his hard work here. Review: https://reviewboard.asterisk.org/r/2603/ (closes issue ASTERISK-20574) Reported by: Jaco Kroon patches: asterisk-rtcp-channel.patch uploaded by jkroon (License 5671) (closes issue ASTERISK-21471) Reported by: Matt Jordan git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@393740 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-07-05 17:33:33 +00:00
AST_RTP_STAT_STRCPY(AST_RTP_INSTANCE_STAT_CHANNEL_UNIQUEID, -1, stats->channel_uniqueid, ast_rtp_instance_get_channel_id(instance));
return 0;
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre Neither instance0 nor instance1 are locked */
static int ast_rtp_dtmf_compatible(struct ast_channel *chan0, struct ast_rtp_instance *instance0, struct ast_channel *chan1, struct ast_rtp_instance *instance1)
{
/* If both sides are not using the same method of DTMF transmission
* (ie: one is RFC2833, other is INFO... then we can not do direct media.
* --------------------------------------------------
* | DTMF Mode | HAS_DTMF | Accepts Begin Frames |
* |-----------|------------|-----------------------|
* | Inband | False | True |
* | RFC2833 | True | True |
* | SIP INFO | False | False |
* --------------------------------------------------
*/
return (((ast_rtp_instance_get_prop(instance0, AST_RTP_PROPERTY_DTMF) != ast_rtp_instance_get_prop(instance1, AST_RTP_PROPERTY_DTMF)) ||
(!ast_channel_tech(chan0)->send_digit_begin != !ast_channel_tech(chan1)->send_digit_begin)) ? 0 : 1);
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is NOT locked */
static void ast_rtp_stun_request(struct ast_rtp_instance *instance, struct ast_sockaddr *suggestion, const char *username)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
struct sockaddr_in suggestion_tmp;
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*
* The instance should not be locked because we can block
* waiting for a STUN respone.
*/
ast_sockaddr_to_sin(suggestion, &suggestion_tmp);
ast_stun_request(rtp->s, &suggestion_tmp, username, NULL);
ast_sockaddr_from_sin(suggestion, &suggestion_tmp);
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static void ast_rtp_stop(struct ast_rtp_instance *instance)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
struct ast_sockaddr addr = { {0,} };
#if defined(HAVE_OPENSSL) && (OPENSSL_VERSION_NUMBER >= 0x10001000L) && !defined(OPENSSL_NO_SRTP)
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_unlock(instance);
AST_SCHED_DEL_UNREF(rtp->sched, rtp->rekeyid, ao2_ref(instance, -1));
dtls_srtp_stop_timeout_timer(instance, rtp, 0);
if (rtp->rtcp) {
dtls_srtp_stop_timeout_timer(instance, rtp, 1);
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_lock(instance);
#endif
res_rtp_asterisk: Asterisk Media Experience Score (MES) ----------------- This commit reinstates MES with some casting fixes to the functions in time.h that convert between doubles and timeval structures. The casting issues were causing incorrect timestamps to be calculated which caused transcoding from/to G722 to produce bad or no audio. ASTERISK-30391 ----------------- This module has been updated to provide additional quality statistics in the form of an Asterisk Media Experience Score. The score is avilable using the same mechanisms you'd use to retrieve jitter, loss, and rtt statistics. For more information about the score and how to retrieve it, see https://wiki.asterisk.org/wiki/display/AST/Media+Experience+Score * Updated chan_pjsip to set quality channel variables when a call ends. * Updated channels/pjsip/dialplan_functions.c to add the ability to retrieve the MES along with the existing rtcp stats when using the CHANNEL dialplan function. * Added the ast_debug_rtp_is_allowed and ast_debug_rtcp_is_allowed checks for debugging purposes. * Added several function to time.h for manipulating time-in-samples and times represented as double seconds. * Updated rtp_engine.c to pass through the MES when stats are requested. Also debug output that dumps the stats when an rtp instance is destroyed. * Updated res_rtp_asterisk.c to implement the calculation of the MES. In the process, also had to update the calculation of jitter. Many debugging statements were also changed to be more informative. * Added a unit test for internal testing. The test should not be run during normal operation and is disabled by default. Change-Id: I4fce265965e68c3fdfeca55e614371ee69c65038
2022-10-28 10:57:56 +00:00
ast_debug_rtp(1, "(%s) RTP Stop\n",
ast_rtp_instance_get_channel_id(instance));
if (rtp->rtcp && rtp->rtcp->schedid > -1) {
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_unlock(instance);
Merged revisions 336878 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/10 ................ r336878 | russell | 2011-09-19 20:03:55 -0500 (Mon, 19 Sep 2011) | 43 lines Merged revisions 336877 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.8 ........ r336877 | russell | 2011-09-19 19:56:20 -0500 (Mon, 19 Sep 2011) | 36 lines Fix crashes in ast_rtcp_write(). This patch addresses crashes related to RTCP handling. The backtraces just show a crash in ast_rtcp_write() where it appears that the RTP instance is no longer valid. There is a race condition with scheduled RTCP transmissions and the destruction of the RTP instance. This patch utilizes the fact that ast_rtp_instance is a reference counted object and ensures that it will not get destroyed while a reference is still around due to scheduled RTCP transmissions. RTCP transmissions are scheduled and executed from the chan_sip scheduler context. This scheduler context is processed in the SIP monitor thread. The destruction of an RTP instance occurs when the associated sip_pvt gets destroyed (which happens when the sip_pvt reference count reaches 0). However, the SIP monitor thread is not the only thread that can cause a sip_pvt to get destroyed. The sip_hangup function, executed from a channel thread, also decrements the reference count on a sip_pvt and could cause it to get destroyed. While this is being changed anyway, the patch also removes calling ast_sched_del() from within the RTCP scheduler callback. It's not helpful. Simply returning 0 prevents the callback from being rescheduled. (closes issue ASTERISK-18570) Related issues that look like they are the same problem: (issue ASTERISK-17560) (issue ASTERISK-15406) (issue ASTERISK-15257) (issue ASTERISK-13334) (issue ASTERISK-9977) (issue ASTERISK-9716) Review: https://reviewboard.asterisk.org/r/1444/ ........ ................ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@336879 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2011-09-20 01:11:18 +00:00
if (!ast_sched_del(rtp->sched, rtp->rtcp->schedid)) {
/* successfully cancelled scheduler entry. */
ao2_ref(instance, -1);
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_lock(instance);
Merged revisions 336878 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/10 ................ r336878 | russell | 2011-09-19 20:03:55 -0500 (Mon, 19 Sep 2011) | 43 lines Merged revisions 336877 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.8 ........ r336877 | russell | 2011-09-19 19:56:20 -0500 (Mon, 19 Sep 2011) | 36 lines Fix crashes in ast_rtcp_write(). This patch addresses crashes related to RTCP handling. The backtraces just show a crash in ast_rtcp_write() where it appears that the RTP instance is no longer valid. There is a race condition with scheduled RTCP transmissions and the destruction of the RTP instance. This patch utilizes the fact that ast_rtp_instance is a reference counted object and ensures that it will not get destroyed while a reference is still around due to scheduled RTCP transmissions. RTCP transmissions are scheduled and executed from the chan_sip scheduler context. This scheduler context is processed in the SIP monitor thread. The destruction of an RTP instance occurs when the associated sip_pvt gets destroyed (which happens when the sip_pvt reference count reaches 0). However, the SIP monitor thread is not the only thread that can cause a sip_pvt to get destroyed. The sip_hangup function, executed from a channel thread, also decrements the reference count on a sip_pvt and could cause it to get destroyed. While this is being changed anyway, the patch also removes calling ast_sched_del() from within the RTCP scheduler callback. It's not helpful. Simply returning 0 prevents the callback from being rescheduled. (closes issue ASTERISK-18570) Related issues that look like they are the same problem: (issue ASTERISK-17560) (issue ASTERISK-15406) (issue ASTERISK-15257) (issue ASTERISK-13334) (issue ASTERISK-9977) (issue ASTERISK-9716) Review: https://reviewboard.asterisk.org/r/1444/ ........ ................ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@336879 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2011-09-20 01:11:18 +00:00
rtp->rtcp->schedid = -1;
}
Merged revisions 336878 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/10 ................ r336878 | russell | 2011-09-19 20:03:55 -0500 (Mon, 19 Sep 2011) | 43 lines Merged revisions 336877 via svnmerge from https://origsvn.digium.com/svn/asterisk/branches/1.8 ........ r336877 | russell | 2011-09-19 19:56:20 -0500 (Mon, 19 Sep 2011) | 36 lines Fix crashes in ast_rtcp_write(). This patch addresses crashes related to RTCP handling. The backtraces just show a crash in ast_rtcp_write() where it appears that the RTP instance is no longer valid. There is a race condition with scheduled RTCP transmissions and the destruction of the RTP instance. This patch utilizes the fact that ast_rtp_instance is a reference counted object and ensures that it will not get destroyed while a reference is still around due to scheduled RTCP transmissions. RTCP transmissions are scheduled and executed from the chan_sip scheduler context. This scheduler context is processed in the SIP monitor thread. The destruction of an RTP instance occurs when the associated sip_pvt gets destroyed (which happens when the sip_pvt reference count reaches 0). However, the SIP monitor thread is not the only thread that can cause a sip_pvt to get destroyed. The sip_hangup function, executed from a channel thread, also decrements the reference count on a sip_pvt and could cause it to get destroyed. While this is being changed anyway, the patch also removes calling ast_sched_del() from within the RTCP scheduler callback. It's not helpful. Simply returning 0 prevents the callback from being rescheduled. (closes issue ASTERISK-18570) Related issues that look like they are the same problem: (issue ASTERISK-17560) (issue ASTERISK-15406) (issue ASTERISK-15257) (issue ASTERISK-13334) (issue ASTERISK-9977) (issue ASTERISK-9716) Review: https://reviewboard.asterisk.org/r/1444/ ........ ................ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@336879 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2011-09-20 01:11:18 +00:00
if (rtp->transport_wide_cc.schedid > -1) {
ao2_unlock(instance);
if (!ast_sched_del(rtp->sched, rtp->transport_wide_cc.schedid)) {
ao2_ref(instance, -1);
}
ao2_lock(instance);
rtp->transport_wide_cc.schedid = -1;
}
if (rtp->red) {
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_unlock(instance);
AST_SCHED_DEL(rtp->sched, rtp->red->schedid);
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
ao2_lock(instance);
ast_free(rtp->red);
rtp->red = NULL;
}
ast_rtp_instance_set_remote_address(instance, &addr);
ast_set_flag(rtp, FLAG_NEED_MARKER_BIT);
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*! \pre instance is locked */
static int ast_rtp_qos_set(struct ast_rtp_instance *instance, int tos, int cos, const char *desc)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
return ast_set_qos(rtp->s, tos, cos, desc);
}
rtp_engine/res_rtp_asterisk: Fix RTP struct reentrancy crashes. The struct ast_rtp_instance has historically been indirectly protected from reentrancy issues by the channel lock because early channel drivers held the lock for really long times. Holding the channel lock for such a long time has caused many deadlock problems in the past. Along comes chan_pjsip/res_pjsip which doesn't necessarily hold the channel lock because sometimes there may not be an associated channel created yet or the channel pointer isn't available. In the case of ASTERISK-26835 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up changing the rtp->rtcp->local_addr_str string and interfering with each other. The classic reentrancy problem resulted in a crash. In the case of ASTERISK-26853 a pjsip serializer thread was processing a message's SDP body while another thread was reading a RTP packet from the socket. Both threads wound up processing ICE candidates in PJPROJECT and interfering with each other. The classic reentrancy problem resulted in a crash. * rtp_engine.c: Make the ast_rtp_instance_xxx() calls lock the RTP instance struct. * rtp_engine.c: Make ICE and DTLS wrapper functions to lock the RTP instance struct for the API call. * res_rtp_asterisk.c: Lock the RTP instance to prevent a reentrancy problem with rtp->rtcp->local_addr_str in the scheduler thread running ast_rtcp_write(). * res_rtp_asterisk.c: Avoid deadlock when local RTP bridging in bridge_p2p_rtp_write() because there are two RTP instance structs involved. * res_rtp_asterisk.c: Avoid deadlock when trying to stop scheduler callbacks. We cannot hold the instance lock when trying to stop a scheduler callback. * res_rtp_asterisk.c: Remove the lock in struct dtls_details and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads to prevent a race condition between starting and stopping a timeout timer. The race condition is no longer present between dtls_perform_handshake() and __rtp_recvfrom() because the instance lock prevents these functions from overlapping each other with regards to the timeout timer. * res_rtp_asterisk.c: Remove the lock in struct ast_rtp and use the struct ast_rtp_instance ao2 object lock instead. The lock was used to synchronize two threads using a condition signal to know when TURN negotiations complete. * res_rtp_asterisk.c: Avoid deadlock when trying to stop the TURN ioqueue_worker_thread(). We cannot hold the instance lock when trying to create or shut down the worker thread without a risk of deadlock. This patch exposed a race condition between a PJSIP serializer thread setting up an ICE session in ice_create() and another thread reading RTP packets. * res_rtp_asterisk.c:ice_create(): Set the new rtp->ice pointer after we have re-locked the RTP instance to prevent the other thread from trying to process ICE packets on an incomplete ICE session setup. A similar race condition is between a PJSIP serializer thread resetting up an ICE session in ice_create() and the timer_worker_thread() processing the completion of the previous ICE session. * res_rtp_asterisk.c:ast_rtp_on_ice_complete(): Protect against an uninitialized/null remote_address after calling update_address_with_ice_candidate(). * res_rtp_asterisk.c: Eliminate the chance of ice_reset_session() destroying and setting the rtp->ice pointer to NULL while other threads are using it by adding an ao2 wrapper around the PJPROJECT ice pointer. Now when we have to unlock the RTP instance object to call a PJPROJECT ICE function we will hold a ref to the wrapper. Also added some rtp->ice NULL checks after we relock the RTP instance and have to do something with the ICE structure. ASTERISK-26835 #close ASTERISK-26853 #close Change-Id: I780b39ec935dcefcce880d50c1a7261744f1d1b4
2017-03-22 21:05:49 +00:00
/*!
* \brief generate comfort noice (CNG)
*
* \pre instance is locked
*/
static int ast_rtp_sendcng(struct ast_rtp_instance *instance, int level)
{
unsigned int *rtpheader;
int hdrlen = 12;
int res, payload = 0;
char data[256];
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
struct ast_sockaddr remote_address = { {0,} };
int ice;
ast_rtp_instance_get_remote_address(instance, &remote_address);
if (ast_sockaddr_isnull(&remote_address)) {
return -1;
}
rtp_engine.c: Initial split of payload types into rx and tx mappings. There are numerous problems with the current implementation of the RTP payload type mapping in Asterisk. It uses only one mapping structure to associate payload types to codecs. The single mapping is overkill if all of the payload type values are well known values. Dynamic payload type mappings do not work as well with the single mapping because RFC3264 allows each side of the link to negotiate different dynamic mappings for what they want to receive. Not only could you have the same codec mapped for sending and receiving on different payload types you could wind up with the same payload type mapped to different codecs for each direction. 1) An independent payload type mapping is needed for sending and receiving. 2) The receive mapping needs to keep track of previous mappings because of the slack to when negotiation happens and current packets in flight using the old mapping arrive. 3) The transmit mapping only needs to keep track of the current negotiated values since we are sending the packets and know when the switchover takes place. * Needed to create ast_rtp_codecs_payload_code_tx() and make some callers use the new function because ast_rtp_codecs_payload_code() was used for mappings in both directions. * Needed to create ast_rtp_codecs_payloads_xover() for cases where we need to pass preferred codec mappings to the peer channel for early media bridging or when we need to prefer the offered mapping that RFC3264 says we SHOULD use. * ast_rtp_codecs_payloads_xover() and ast_rtp_codecs_payload_code_tx() are the only new public functions created. All the others were only used for the tx or rx mapping direction so the function doxygen now reflects which direction the function operates. * chan_mgcp.c: Removed call to ast_rtp_codecs_payloads_clear() as doing that makes no sense when processing an incoming SDP. We would be wiping out any mappings that we set for the possible outgoing SDP we sent earlier. ASTERISK-25166 Reported by: Kevin Harwell ASTERISK-17410 Reported by: Boris Fox Change-Id: Iaf6c227bca68cb7c414cf2fd4108a8ac98bd45ac
2015-07-24 00:24:04 +00:00
payload = ast_rtp_codecs_payload_code_tx(ast_rtp_instance_get_codecs(instance), 0, NULL, AST_RTP_CN);
level = 127 - (level & 0x7f);
rtp->dtmfmute = ast_tvadd(ast_tvnow(), ast_tv(0, 500000));
/* Get a pointer to the header */
rtpheader = (unsigned int *)data;
rtpheader[0] = htonl((2 << 30) | (payload << 16) | (rtp->seqno));
rtpheader[1] = htonl(rtp->lastts);
rtpheader[2] = htonl(rtp->ssrc);
data[12] = level;
res = rtp_sendto(instance, (void *) rtpheader, hdrlen + 1, 0, &remote_address, &ice);
if (res < 0) {
ast_log(LOG_ERROR, "RTP Comfort Noise Transmission error to %s: %s\n", ast_sockaddr_stringify(&remote_address), strerror(errno));
return res;
}
if (rtp_debug_test_addr(&remote_address)) {
ast_verbose("Sent Comfort Noise RTP packet to %s%s (type %-2.2d, seq %-6.6d, ts %-6.6u, len %-6.6d)\n",
ast_sockaddr_stringify(&remote_address),
ice ? " (via ICE)" : "",
AST_RTP_CN, rtp->seqno, rtp->lastdigitts, res - hdrlen);
}
rtp->seqno++;
return res;
}
/*! \pre instance is locked */
static unsigned int ast_rtp_get_ssrc(struct ast_rtp_instance *instance)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
return rtp->ssrc;
}
/*! \pre instance is locked */
static const char *ast_rtp_get_cname(struct ast_rtp_instance *instance)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
return rtp->cname;
}
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
/*! \pre instance is locked */
static void ast_rtp_set_remote_ssrc(struct ast_rtp_instance *instance, unsigned int ssrc)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
if (rtp->themssrc_valid && rtp->themssrc == ssrc) {
return;
}
rtp->themssrc = ssrc;
rtp->themssrc_valid = 1;
/* If this is bundled we need to update the SSRC mapping */
if (rtp->bundled) {
struct ast_rtp *bundled_rtp;
int index;
ao2_unlock(instance);
/* The child lock can't be held while accessing the parent */
ao2_lock(rtp->bundled);
bundled_rtp = ast_rtp_instance_get_data(rtp->bundled);
for (index = 0; index < AST_VECTOR_SIZE(&bundled_rtp->ssrc_mapping); ++index) {
struct rtp_ssrc_mapping *mapping = AST_VECTOR_GET_ADDR(&bundled_rtp->ssrc_mapping, index);
if (mapping->instance == instance) {
mapping->ssrc = ssrc;
mapping->ssrc_valid = 1;
break;
}
}
ao2_unlock(rtp->bundled);
ao2_lock(instance);
}
}
static void ast_rtp_set_stream_num(struct ast_rtp_instance *instance, int stream_num)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
rtp->stream_num = stream_num;
}
static int ast_rtp_extension_enable(struct ast_rtp_instance *instance, enum ast_rtp_extension extension)
{
switch (extension) {
case AST_RTP_EXTENSION_ABS_SEND_TIME:
case AST_RTP_EXTENSION_TRANSPORT_WIDE_CC:
return 1;
default:
return 0;
}
}
AST-2017-008: Improve RTP and RTCP packet processing. Validate RTCP packets before processing them. * Validate that the received packet is of a minimum length and apply the RFC3550 RTCP packet validation checks. * Fixed potentially reading garbage beyond the received RTCP record data. * Fixed rtp->themssrc only being set once when the remote could change the SSRC. We would effectively stop handling the RTCP statistic records. * Fixed rtp->themssrc to not treat a zero value as special by adding rtp->themssrc_valid to indicate if rtp->themssrc is available. ASTERISK-27274 Make strict RTP learning more flexible. Direct media can cause strict RTP to attempt to learn a remote address again before it has had a chance to learn the remote address the first time. Because of the rapid relearn requests, strict RTP could latch onto the first remote address and fail to latch onto the direct media remote address. As a result, you have one way audio until the call is placed on and off hold. The new algorithm learns remote addresses for a set time (1.5 seconds) before locking the remote address. In addition, we must see a configured number of remote packets from the same address in a row before switching. * Fixed strict RTP learning from always accepting the first new address packet as the new stream. * Fixed strict RTP to initialize the expected sequence number with the last received sequence number instead of the last transmitted sequence number. * Fixed the predicted next sequence number calculation in rtp_learning_rtp_seq_update() to handle overflow. ASTERISK-27252 Change-Id: Ia2d3aa6e0f22906c25971e74f10027d96525f31c
2017-08-25 22:01:57 +00:00
/*! \pre child is locked */
static int ast_rtp_bundle(struct ast_rtp_instance *child, struct ast_rtp_instance *parent)
{
struct ast_rtp *child_rtp = ast_rtp_instance_get_data(child);
struct ast_rtp *parent_rtp;
struct rtp_ssrc_mapping mapping;
struct ast_sockaddr them = { { 0, } };
if (child_rtp->bundled == parent) {
return 0;
}
/* If this instance was already bundled then remove the SSRC mapping */
if (child_rtp->bundled) {
struct ast_rtp *bundled_rtp;
ao2_unlock(child);
/* The child lock can't be held while accessing the parent */
ao2_lock(child_rtp->bundled);
bundled_rtp = ast_rtp_instance_get_data(child_rtp->bundled);
AST_VECTOR_REMOVE_CMP_UNORDERED(&bundled_rtp->ssrc_mapping, child, SSRC_MAPPING_ELEM_CMP, AST_VECTOR_ELEM_CLEANUP_NOOP);
ao2_unlock(child_rtp->bundled);
ao2_lock(child);
ao2_ref(child_rtp->bundled, -1);
child_rtp->bundled = NULL;
}
if (!parent) {
/* We transitioned away from bundle so we need our own transport resources once again */
rtp_allocate_transport(child, child_rtp);
return 0;
}
parent_rtp = ast_rtp_instance_get_data(parent);
/* We no longer need any transport related resources as we will use our parent RTP instance instead */
rtp_deallocate_transport(child, child_rtp);
/* Children maintain a reference to the parent to guarantee that the transport doesn't go away on them */
child_rtp->bundled = ao2_bump(parent);
mapping.ssrc = child_rtp->themssrc;
mapping.ssrc_valid = child_rtp->themssrc_valid;
mapping.instance = child;
ao2_unlock(child);
ao2_lock(parent);
AST_VECTOR_APPEND(&parent_rtp->ssrc_mapping, mapping);
#if defined(HAVE_OPENSSL) && (OPENSSL_VERSION_NUMBER >= 0x10001000L) && !defined(OPENSSL_NO_SRTP)
/* If DTLS-SRTP is already in use then add the local SSRC to it, otherwise it will get added once DTLS
* negotiation has been completed.
*/
if (parent_rtp->dtls.connection == AST_RTP_DTLS_CONNECTION_EXISTING) {
dtls_srtp_add_local_ssrc(parent_rtp, parent, 0, child_rtp->ssrc, 0);
}
#endif
/* Bundle requires that RTCP-MUX be in use so only the main remote address needs to match */
ast_rtp_instance_get_remote_address(parent, &them);
ao2_unlock(parent);
ao2_lock(child);
ast_rtp_instance_set_remote_address(child, &them);
return 0;
}
#ifdef HAVE_PJPROJECT
static void stunaddr_resolve_callback(const struct ast_dns_query *query)
{
const int lowest_ttl = ast_dns_result_get_lowest_ttl(ast_dns_query_get_result(query));
const char *stunaddr_name = ast_dns_query_get_name(query);
const char *stunaddr_resolved_str;
if (!store_stunaddr_resolved(query)) {
ast_log(LOG_WARNING, "Failed to resolve stunaddr '%s'. Cancelling recurring resolution.\n", stunaddr_name);
return;
}
if (DEBUG_ATLEAST(2)) {
ast_rwlock_rdlock(&stunaddr_lock);
stunaddr_resolved_str = ast_inet_ntoa(stunaddr.sin_addr);
ast_rwlock_unlock(&stunaddr_lock);
ast_debug_stun(2, "Resolved stunaddr '%s' to '%s'. Lowest TTL = %d.\n",
stunaddr_name,
stunaddr_resolved_str,
lowest_ttl);
}
if (!lowest_ttl) {
ast_log(LOG_WARNING, "Resolution for stunaddr '%s' returned TTL = 0. Recurring resolution was cancelled.\n", ast_dns_query_get_name(query));
}
}
static int store_stunaddr_resolved(const struct ast_dns_query *query)
{
const struct ast_dns_result *result = ast_dns_query_get_result(query);
const struct ast_dns_record *record;
for (record = ast_dns_result_get_records(result); record; record = ast_dns_record_get_next(record)) {
const size_t data_size = ast_dns_record_get_data_size(record);
const unsigned char *data = (unsigned char *)ast_dns_record_get_data(record);
const int rr_type = ast_dns_record_get_rr_type(record);
if (rr_type == ns_t_a && data_size == 4) {
ast_rwlock_wrlock(&stunaddr_lock);
memcpy(&stunaddr.sin_addr, data, data_size);
stunaddr.sin_family = AF_INET;
ast_rwlock_unlock(&stunaddr_lock);
return 1;
} else {
ast_debug_stun(3, "Unrecognized rr_type '%u' or data_size '%zu' from DNS query for stunaddr '%s'\n",
rr_type, data_size, ast_dns_query_get_name(query));
continue;
}
}
return 0;
}
static void clean_stunaddr(void) {
if (stunaddr_resolver) {
if (ast_dns_resolve_recurring_cancel(stunaddr_resolver)) {
ast_log(LOG_ERROR, "Failed to cancel recurring DNS resolution of previous stunaddr.\n");
}
ao2_ref(stunaddr_resolver, -1);
stunaddr_resolver = NULL;
}
ast_rwlock_wrlock(&stunaddr_lock);
memset(&stunaddr, 0, sizeof(stunaddr));
ast_rwlock_unlock(&stunaddr_lock);
}
#endif
#if defined(HAVE_OPENSSL) && (OPENSSL_VERSION_NUMBER >= 0x10001000L) && !defined(OPENSSL_NO_SRTP)
/*! \pre instance is locked */
static int ast_rtp_activate(struct ast_rtp_instance *instance)
{
struct ast_rtp *rtp = ast_rtp_instance_get_data(instance);
/* If ICE negotiation is enabled the DTLS Handshake will be performed upon completion of it */
#ifdef HAVE_PJPROJECT
if (rtp->ice) {
return 0;
}
#endif
ast_debug_dtls(3, "(%p) DTLS - ast_rtp_activate rtp=%p - setup and perform DTLS'\n", instance, rtp);
res_rtp_asterisk: bad audio (static) due to incomplete dtls/srtp setup There was a race condition between client initiated DTLS setup, and handling of server side ice completion that caused the underlying SSL object to get cleared during DTLS initialization. If this happened Asterisk would be left in a partial DTLS setup state. RTP packets were sent and received, but were not being encrypted and decrypted. This resulted in no audio, or static. Specifically, this occurred when '__rtp_recvfrom' was processing the handshake sequence from the client to the server, and then 'ast_rtp_on_ice_complete' gets called from another thread and clears the SSL object when calling the 'dtls_perform_setup' function. The timing had to be just right in the sense that from the external SSL library perspective SSL initialization completed (rtp recv), Asterisk clears/resets the SSL object (ice done), and then checks to see if SSL is intialized (rtp recv). Since it was cleared, Asterisk thinks it is not finished, thus not completing 'dtls_srtp_setup'. This patch removes calls to 'dtls_perform_setup', which clears the SSL object, in 'ast_rtp_on_ice_complete'. When ice completes, there is no reason to clear the underlying SSL object. If an ice candidate changes a full protocol level renegotiation occurs. Also, in the case of bundled ICE candidates are reused when a stream is added. So no real reason to have to clear, and reset in this instance. Also, this patch adds a bit of extra logging to aid in diagnosis of any future problems. ASTERISK-28742 #close Change-Id: I34c9e6bad5a39b087164646e2836e3e48fe6892f
2020-02-13 21:08:10 +00:00
dtls_perform_setup(&rtp->dtls);
dtls_perform_handshake(instance, &rtp->dtls, 0);
if (rtp->rtcp && rtp->rtcp->type == AST_RTP_INSTANCE_RTCP_STANDARD) {
dtls_perform_setup(&rtp->rtcp->dtls);
dtls_perform_handshake(instance, &rtp->rtcp->dtls, 1);
}
return 0;
}
#endif
static char *rtp_do_debug_ip(struct ast_cli_args *a)
{
char *arg = ast_strdupa(a->argv[4]);
char *debughost = NULL;
char *debugport = NULL;
if (!ast_sockaddr_parse(&rtpdebugaddr, arg, 0) || !ast_sockaddr_split_hostport(arg, &debughost, &debugport, 0)) {
ast_cli(a->fd, "Lookup failed for '%s'\n", arg);
return CLI_FAILURE;
}
rtpdebugport = (!ast_strlen_zero(debugport) && debugport[0] != '0');
ast_cli(a->fd, "RTP Packet Debugging Enabled for address: %s\n",
ast_sockaddr_stringify(&rtpdebugaddr));
ast_debug_category_set_sublevel(AST_LOG_CATEGORY_RTP_PACKET, AST_LOG_CATEGORY_ENABLED);
return CLI_SUCCESS;
}
static char *rtcp_do_debug_ip(struct ast_cli_args *a)
{
char *arg = ast_strdupa(a->argv[4]);
char *debughost = NULL;
char *debugport = NULL;
if (!ast_sockaddr_parse(&rtcpdebugaddr, arg, 0) || !ast_sockaddr_split_hostport(arg, &debughost, &debugport, 0)) {
ast_cli(a->fd, "Lookup failed for '%s'\n", arg);
return CLI_FAILURE;
}
rtcpdebugport = (!ast_strlen_zero(debugport) && debugport[0] != '0');
ast_cli(a->fd, "RTCP Packet Debugging Enabled for address: %s\n",
ast_sockaddr_stringify(&rtcpdebugaddr));
ast_debug_category_set_sublevel(AST_LOG_CATEGORY_RTCP_PACKET, AST_LOG_CATEGORY_ENABLED);
return CLI_SUCCESS;
}
static char *handle_cli_rtp_set_debug(struct ast_cli_entry *e, int cmd, struct ast_cli_args *a)
{
switch (cmd) {
case CLI_INIT:
e->command = "rtp set debug {on|off|ip}";
e->usage =
"Usage: rtp set debug {on|off|ip host[:port]}\n"
" Enable/Disable dumping of all RTP packets. If 'ip' is\n"
" specified, limit the dumped packets to those to and from\n"
" the specified 'host' with optional port.\n";
return NULL;
case CLI_GENERATE:
return NULL;
}
if (a->argc == e->args) { /* set on or off */
if (!strncasecmp(a->argv[e->args-1], "on", 2)) {
ast_debug_category_set_sublevel(AST_LOG_CATEGORY_RTP_PACKET, AST_LOG_CATEGORY_ENABLED);
memset(&rtpdebugaddr, 0, sizeof(rtpdebugaddr));
ast_cli(a->fd, "RTP Packet Debugging Enabled\n");
return CLI_SUCCESS;
} else if (!strncasecmp(a->argv[e->args-1], "off", 3)) {
ast_debug_category_set_sublevel(AST_LOG_CATEGORY_RTP_PACKET, AST_LOG_CATEGORY_DISABLED);
ast_cli(a->fd, "RTP Packet Debugging Disabled\n");
return CLI_SUCCESS;
}
} else if (a->argc == e->args +1) { /* ip */
return rtp_do_debug_ip(a);
}
return CLI_SHOWUSAGE; /* default, failure */
}
static char *handle_cli_rtp_settings(struct ast_cli_entry *e, int cmd, struct ast_cli_args *a)
{
#ifdef HAVE_PJPROJECT
struct sockaddr_in stunaddr_copy;
#endif
switch (cmd) {
case CLI_INIT:
e->command = "rtp show settings";
e->usage =
"Usage: rtp show settings\n"
" Display RTP configuration settings\n";
return NULL;
case CLI_GENERATE:
return NULL;
}
if (a->argc != 3) {
return CLI_SHOWUSAGE;
}
ast_cli(a->fd, "\n\nGeneral Settings:\n");
ast_cli(a->fd, "----------------\n");
ast_cli(a->fd, " Port start: %d\n", rtpstart);
ast_cli(a->fd, " Port end: %d\n", rtpend);
#ifdef SO_NO_CHECK
ast_cli(a->fd, " Checksums: %s\n", AST_CLI_YESNO(nochecksums == 0));
#endif
ast_cli(a->fd, " DTMF Timeout: %d\n", dtmftimeout);
ast_cli(a->fd, " Strict RTP: %s\n", AST_CLI_YESNO(strictrtp));
if (strictrtp) {
ast_cli(a->fd, " Probation: %d frames\n", learning_min_sequential);
}
ast_cli(a->fd, " Replay Protect: %s\n", AST_CLI_YESNO(srtp_replay_protection));
#ifdef HAVE_PJPROJECT
ast_cli(a->fd, " ICE support: %s\n", AST_CLI_YESNO(icesupport));
ast_rwlock_rdlock(&stunaddr_lock);
memcpy(&stunaddr_copy, &stunaddr, sizeof(stunaddr));
ast_rwlock_unlock(&stunaddr_lock);
ast_cli(a->fd, " STUN address: %s:%d\n", ast_inet_ntoa(stunaddr_copy.sin_addr), htons(stunaddr_copy.sin_port));
#endif
return CLI_SUCCESS;
}
static char *handle_cli_rtcp_set_debug(struct ast_cli_entry *e, int cmd, struct ast_cli_args *a)
{
switch (cmd) {
case CLI_INIT:
e->command = "rtcp set debug {on|off|ip}";
e->usage =
"Usage: rtcp set debug {on|off|ip host[:port]}\n"
" Enable/Disable dumping of all RTCP packets. If 'ip' is\n"
" specified, limit the dumped packets to those to and from\n"
" the specified 'host' with optional port.\n";
return NULL;
case CLI_GENERATE:
return NULL;
}
if (a->argc == e->args) { /* set on or off */
if (!strncasecmp(a->argv[e->args-1], "on", 2)) {
ast_debug_category_set_sublevel(AST_LOG_CATEGORY_RTCP_PACKET, AST_LOG_CATEGORY_ENABLED);
memset(&rtcpdebugaddr, 0, sizeof(rtcpdebugaddr));
ast_cli(a->fd, "RTCP Packet Debugging Enabled\n");
return CLI_SUCCESS;
} else if (!strncasecmp(a->argv[e->args-1], "off", 3)) {
ast_debug_category_set_sublevel(AST_LOG_CATEGORY_RTCP_PACKET, AST_LOG_CATEGORY_DISABLED);
ast_cli(a->fd, "RTCP Packet Debugging Disabled\n");
return CLI_SUCCESS;
}
} else if (a->argc == e->args +1) { /* ip */
return rtcp_do_debug_ip(a);
}
return CLI_SHOWUSAGE; /* default, failure */
}
static char *handle_cli_rtcp_set_stats(struct ast_cli_entry *e, int cmd, struct ast_cli_args *a)
{
switch (cmd) {
case CLI_INIT:
e->command = "rtcp set stats {on|off}";
e->usage =
"Usage: rtcp set stats {on|off}\n"
" Enable/Disable dumping of RTCP stats.\n";
return NULL;
case CLI_GENERATE:
return NULL;
}
if (a->argc != e->args)
return CLI_SHOWUSAGE;
if (!strncasecmp(a->argv[e->args-1], "on", 2))
rtcpstats = 1;
else if (!strncasecmp(a->argv[e->args-1], "off", 3))
rtcpstats = 0;
else
return CLI_SHOWUSAGE;
ast_cli(a->fd, "RTCP Stats %s\n", rtcpstats ? "Enabled" : "Disabled");
return CLI_SUCCESS;
}
#ifdef AST_DEVMODE
static unsigned int use_random(struct ast_cli_args *a, int pos, unsigned int index)
{
return pos >= index && !ast_strlen_zero(a->argv[index - 1]) &&
!strcasecmp(a->argv[index - 1], "random");
}
static char *handle_cli_rtp_drop_incoming_packets(struct ast_cli_entry *e, int cmd, struct ast_cli_args *a)
{
static const char * const completions_2[] = { "stop", "<N>", NULL };
static const char * const completions_3[] = { "random", "incoming packets", NULL };
static const char * const completions_5[] = { "on", "every", NULL };
static const char * const completions_units[] = { "random", "usec", "msec", "sec", "min", NULL };
unsigned int use_random_num = 0;
unsigned int use_random_interval = 0;
unsigned int num_to_drop = 0;
unsigned int interval = 0;
const char *interval_s = NULL;
const char *unit_s = NULL;
struct ast_sockaddr addr;
const char *addr_s = NULL;
switch (cmd) {
case CLI_INIT:
e->command = "rtp drop";
e->usage =
"Usage: rtp drop [stop|[<N> [random] incoming packets[ every <N> [random] {usec|msec|sec|min}][ on <ip[:port]>]]\n"
" Drop RTP incoming packets.\n";
return NULL;
case CLI_GENERATE:
use_random_num = use_random(a, a->pos, 4);
use_random_interval = use_random(a, a->pos, 8 + use_random_num) ||
use_random(a, a->pos, 10 + use_random_num);
switch (a->pos - use_random_num - use_random_interval) {
case 2:
return ast_cli_complete(a->word, completions_2, a->n);
case 3:
return ast_cli_complete(a->word, completions_3 + use_random_num, a->n);
case 5:
return ast_cli_complete(a->word, completions_5, a->n);
case 7:
if (!strcasecmp(a->argv[a->pos - 2], "on")) {
ast_cli_completion_add(ast_strdup("every"));
break;
}
/* Fall through */
case 9:
if (!strcasecmp(a->argv[a->pos - 2 - use_random_interval], "every")) {
return ast_cli_complete(a->word, completions_units + use_random_interval, a->n);
}
break;
case 8:
if (!strcasecmp(a->argv[a->pos - 3 - use_random_interval], "every")) {
ast_cli_completion_add(ast_strdup("on"));
}
break;
}
return NULL;
}
if (a->argc < 3) {
return CLI_SHOWUSAGE;
}
use_random_num = use_random(a, a->argc, 4);
use_random_interval = use_random(a, a->argc, 8 + use_random_num) ||
use_random(a, a->argc, 10 + use_random_num);
if (!strcasecmp(a->argv[2], "stop")) {
/* rtp drop stop */
} else if (a->argc < 5) {
return CLI_SHOWUSAGE;
} else if (ast_str_to_uint(a->argv[2], &num_to_drop)) {
ast_cli(a->fd, "%s is not a valid number of packets to drop\n", a->argv[2]);
return CLI_FAILURE;
} else if (a->argc - use_random_num == 5) {
/* rtp drop <N> [random] incoming packets */
} else if (a->argc - use_random_num >= 7 && !strcasecmp(a->argv[5 + use_random_num], "on")) {
/* rtp drop <N> [random] incoming packets on <ip[:port]> */
addr_s = a->argv[6 + use_random_num];
if (a->argc - use_random_num - use_random_interval == 10 &&
!strcasecmp(a->argv[7 + use_random_num], "every")) {
/* rtp drop <N> [random] incoming packets on <ip[:port]> every <N> [random] {usec|msec|sec|min} */
interval_s = a->argv[8 + use_random_num];
unit_s = a->argv[9 + use_random_num + use_random_interval];
}
} else if (a->argc - use_random_num >= 8 && !strcasecmp(a->argv[5 + use_random_num], "every")) {
/* rtp drop <N> [random] incoming packets every <N> [random] {usec|msec|sec|min} */
interval_s = a->argv[6 + use_random_num];
unit_s = a->argv[7 + use_random_num + use_random_interval];
if (a->argc == 10 + use_random_num + use_random_interval &&
!strcasecmp(a->argv[8 + use_random_num + use_random_interval], "on")) {
/* rtp drop <N> [random] incoming packets every <N> [random] {usec|msec|sec|min} on <ip[:port]> */
addr_s = a->argv[9 + use_random_num + use_random_interval];
}
} else {
return CLI_SHOWUSAGE;
}
if (a->argc - use_random_num >= 8 && !interval_s && !addr_s) {
return CLI_SHOWUSAGE;
}
if (interval_s && ast_str_to_uint(interval_s, &interval)) {
ast_cli(a->fd, "%s is not a valid interval number\n", interval_s);
return CLI_FAILURE;
}
memset(&addr, 0, sizeof(addr));
if (addr_s && !ast_sockaddr_parse(&addr, addr_s, 0)) {
ast_cli(a->fd, "%s is not a valid hostname[:port]\n", addr_s);
return CLI_FAILURE;
}
drop_packets_data.use_random_num = use_random_num;
drop_packets_data.use_random_interval = use_random_interval;
drop_packets_data.num_to_drop = num_to_drop;
drop_packets_data.interval = ast_time_create_by_unit_str(interval, unit_s);
ast_sockaddr_copy(&drop_packets_data.addr, &addr);
drop_packets_data.port = ast_sockaddr_port(&addr);
drop_packets_data_update(ast_tvnow());
return CLI_SUCCESS;
}
#endif
static struct ast_cli_entry cli_rtp[] = {
AST_CLI_DEFINE(handle_cli_rtp_set_debug, "Enable/Disable RTP debugging"),
AST_CLI_DEFINE(handle_cli_rtp_settings, "Display RTP settings"),
AST_CLI_DEFINE(handle_cli_rtcp_set_debug, "Enable/Disable RTCP debugging"),
AST_CLI_DEFINE(handle_cli_rtcp_set_stats, "Enable/Disable RTCP stats"),
#ifdef AST_DEVMODE
AST_CLI_DEFINE(handle_cli_rtp_drop_incoming_packets, "Drop RTP incoming packets"),
#endif
};
static int rtp_reload(int reload, int by_external_config)
{
struct ast_config *cfg;
const char *s;
struct ast_flags config_flags = { (reload && !by_external_config) ? CONFIG_FLAG_FILEUNCHANGED : 0 };
#ifdef HAVE_PJPROJECT
struct ast_variable *var;
struct ast_ice_host_candidate *candidate;
int acl_subscription_flag = 0;
#endif
cfg = ast_config_load2("rtp.conf", "rtp", config_flags);
if (!cfg || cfg == CONFIG_STATUS_FILEUNCHANGED || cfg == CONFIG_STATUS_FILEINVALID) {
return 0;
}
#ifdef SO_NO_CHECK
nochecksums = 0;
#endif
rtpstart = DEFAULT_RTP_START;
rtpend = DEFAULT_RTP_END;
rtcpinterval = RTCP_DEFAULT_INTERVALMS;
dtmftimeout = DEFAULT_DTMF_TIMEOUT;
strictrtp = DEFAULT_STRICT_RTP;
learning_min_sequential = DEFAULT_LEARNING_MIN_SEQUENTIAL;
learning_min_duration = DEFAULT_LEARNING_MIN_DURATION;
srtp_replay_protection = DEFAULT_SRTP_REPLAY_PROTECTION;
/** This resource is not "reloaded" so much as unloaded and loaded again.
* In the case of the TURN related variables, the memory referenced by a
* previously loaded instance *should* have been released when the
* corresponding pool was destroyed. If at some point in the future this
* resource were to support ACTUAL live reconfiguration and did NOT release
* the pool this will cause a small memory leak.
*/
#ifdef HAVE_PJPROJECT
icesupport = DEFAULT_ICESUPPORT;
stun_software_attribute = DEFAULT_STUN_SOFTWARE_ATTRIBUTE;
turnport = DEFAULT_TURN_PORT;
clean_stunaddr();
turnaddr = pj_str(NULL);
turnusername = pj_str(NULL);
turnpassword = pj_str(NULL);
host_candidate_overrides_clear();
#endif
#if defined(HAVE_OPENSSL) && (OPENSSL_VERSION_NUMBER >= 0x10001000L) && !defined(OPENSSL_NO_SRTP)
dtls_mtu = DEFAULT_DTLS_MTU;
#endif
if ((s = ast_variable_retrieve(cfg, "general", "rtpstart"))) {
rtpstart = atoi(s);
if (rtpstart < MINIMUM_RTP_PORT)
rtpstart = MINIMUM_RTP_PORT;
if (rtpstart > MAXIMUM_RTP_PORT)
rtpstart = MAXIMUM_RTP_PORT;
}
if ((s = ast_variable_retrieve(cfg, "general", "rtpend"))) {
rtpend = atoi(s);
if (rtpend < MINIMUM_RTP_PORT)
rtpend = MINIMUM_RTP_PORT;
if (rtpend > MAXIMUM_RTP_PORT)
rtpend = MAXIMUM_RTP_PORT;
}
if ((s = ast_variable_retrieve(cfg, "general", "rtcpinterval"))) {
rtcpinterval = atoi(s);
if (rtcpinterval == 0)
rtcpinterval = 0; /* Just so we're clear... it's zero */
if (rtcpinterval < RTCP_MIN_INTERVALMS)
rtcpinterval = RTCP_MIN_INTERVALMS; /* This catches negative numbers too */
if (rtcpinterval > RTCP_MAX_INTERVALMS)
rtcpinterval = RTCP_MAX_INTERVALMS;
}
if ((s = ast_variable_retrieve(cfg, "general", "rtpchecksums"))) {
#ifdef SO_NO_CHECK
nochecksums = ast_false(s) ? 1 : 0;
#else
if (ast_false(s))
ast_log(LOG_WARNING, "Disabling RTP checksums is not supported on this operating system!\n");
#endif
}
if ((s = ast_variable_retrieve(cfg, "general", "dtmftimeout"))) {
dtmftimeout = atoi(s);
if ((dtmftimeout < 0) || (dtmftimeout > 64000)) {
ast_log(LOG_WARNING, "DTMF timeout of '%d' outside range, using default of '%d' instead\n",
dtmftimeout, DEFAULT_DTMF_TIMEOUT);
dtmftimeout = DEFAULT_DTMF_TIMEOUT;
};
}
if ((s = ast_variable_retrieve(cfg, "general", "strictrtp"))) {
if (ast_true(s)) {
strictrtp = STRICT_RTP_YES;
} else if (!strcasecmp(s, "seqno")) {
strictrtp = STRICT_RTP_SEQNO;
} else {
strictrtp = STRICT_RTP_NO;
}
}
if ((s = ast_variable_retrieve(cfg, "general", "probation"))) {
if ((sscanf(s, "%d", &learning_min_sequential) != 1) || learning_min_sequential <= 1) {
ast_log(LOG_WARNING, "Value for 'probation' could not be read, using default of '%d' instead\n",
DEFAULT_LEARNING_MIN_SEQUENTIAL);
learning_min_sequential = DEFAULT_LEARNING_MIN_SEQUENTIAL;
}
learning_min_duration = CALC_LEARNING_MIN_DURATION(learning_min_sequential);
}
if ((s = ast_variable_retrieve(cfg, "general", "srtpreplayprotection"))) {
srtp_replay_protection = ast_true(s);
}
#ifdef HAVE_PJPROJECT
if ((s = ast_variable_retrieve(cfg, "general", "icesupport"))) {
icesupport = ast_true(s);
}
if ((s = ast_variable_retrieve(cfg, "general", "stun_software_attribute"))) {
stun_software_attribute = ast_true(s);
}
if ((s = ast_variable_retrieve(cfg, "general", "stunaddr"))) {
char *hostport, *host, *port;
unsigned int port_parsed = STANDARD_STUN_PORT;
struct ast_sockaddr stunaddr_parsed;
hostport = ast_strdupa(s);
if (!ast_parse_arg(hostport, PARSE_ADDR, &stunaddr_parsed)) {
ast_debug_stun(3, "stunaddr = '%s' does not need name resolution\n",
ast_sockaddr_stringify_host(&stunaddr_parsed));
if (!ast_sockaddr_port(&stunaddr_parsed)) {
ast_sockaddr_set_port(&stunaddr_parsed, STANDARD_STUN_PORT);
}
ast_rwlock_wrlock(&stunaddr_lock);
ast_sockaddr_to_sin(&stunaddr_parsed, &stunaddr);
ast_rwlock_unlock(&stunaddr_lock);
} else if (ast_sockaddr_split_hostport(hostport, &host, &port, 0)) {
if (port) {
ast_parse_arg(port, PARSE_UINT32|PARSE_IN_RANGE, &port_parsed, 1, 65535);
}
stunaddr.sin_port = htons(port_parsed);
stunaddr_resolver = ast_dns_resolve_recurring(host, T_A, C_IN,
&stunaddr_resolve_callback, NULL);
if (!stunaddr_resolver) {
ast_log(LOG_ERROR, "Failed to setup recurring DNS resolution of stunaddr '%s'",
host);
}
} else {
ast_log(LOG_ERROR, "Failed to parse stunaddr '%s'", hostport);
}
}
if ((s = ast_variable_retrieve(cfg, "general", "turnaddr"))) {
struct sockaddr_in addr;
addr.sin_port = htons(DEFAULT_TURN_PORT);
if (ast_parse_arg(s, PARSE_INADDR, &addr)) {
ast_log(LOG_WARNING, "Invalid TURN server address: %s\n", s);
} else {
pj_strdup2_with_null(pool, &turnaddr, ast_inet_ntoa(addr.sin_addr));
/* ntohs() is not a bug here. The port number is used in host byte order with
* a pjnat API. */
turnport = ntohs(addr.sin_port);
}
}
if ((s = ast_variable_retrieve(cfg, "general", "turnusername"))) {
pj_strdup2_with_null(pool, &turnusername, s);
}
if ((s = ast_variable_retrieve(cfg, "general", "turnpassword"))) {
pj_strdup2_with_null(pool, &turnpassword, s);
}
AST_RWLIST_WRLOCK(&host_candidates);
for (var = ast_variable_browse(cfg, "ice_host_candidates"); var; var = var->next) {
struct ast_sockaddr local_addr, advertised_addr;
unsigned int include_local_address = 0;
char *sep;
ast_sockaddr_setnull(&local_addr);
ast_sockaddr_setnull(&advertised_addr);
if (ast_parse_arg(var->name, PARSE_ADDR | PARSE_PORT_IGNORE, &local_addr)) {
ast_log(LOG_WARNING, "Invalid local ICE host address: %s\n", var->name);
continue;
}
sep = strchr(var->value,',');
if (sep) {
*sep = '\0';
sep++;
sep = ast_skip_blanks(sep);
include_local_address = strcmp(sep, "include_local_address") == 0;
}
if (ast_parse_arg(var->value, PARSE_ADDR | PARSE_PORT_IGNORE, &advertised_addr)) {
ast_log(LOG_WARNING, "Invalid advertised ICE host address: %s\n", var->value);
continue;
}
if (!(candidate = ast_calloc(1, sizeof(*candidate)))) {
ast_log(LOG_ERROR, "Failed to allocate ICE host candidate mapping.\n");
break;
}
candidate->include_local = include_local_address;
ast_sockaddr_copy(&candidate->local, &local_addr);
ast_sockaddr_copy(&candidate->advertised, &advertised_addr);
AST_RWLIST_INSERT_TAIL(&host_candidates, candidate, next);
}
AST_RWLIST_UNLOCK(&host_candidates);
ast_rwlock_wrlock(&ice_acl_lock);
ast_rwlock_wrlock(&stun_acl_lock);
ice_acl = ast_free_acl_list(ice_acl);
stun_acl = ast_free_acl_list(stun_acl);
for (var = ast_variable_browse(cfg, "general"); var; var = var->next) {
const char* sense = NULL;
struct ast_acl_list **acl = NULL;
if (strncasecmp(var->name, "ice_", 4) == 0) {
sense = var->name + 4;
acl = &ice_acl;
} else if (strncasecmp(var->name, "stun_", 5) == 0) {
sense = var->name + 5;
acl = &stun_acl;
} else {
continue;
}
if (strcasecmp(sense, "blacklist") == 0) {
sense = "deny";
}
if (strcasecmp(sense, "acl") && strcasecmp(sense, "permit") && strcasecmp(sense, "deny")) {
continue;
}
ast_append_acl(sense, var->value, acl, NULL, &acl_subscription_flag);
}
ast_rwlock_unlock(&ice_acl_lock);
ast_rwlock_unlock(&stun_acl_lock);
if (acl_subscription_flag && !acl_change_sub) {
acl_change_sub = stasis_subscribe(ast_security_topic(), acl_change_stasis_cb, NULL);
stasis_subscription_accept_message_type(acl_change_sub, ast_named_acl_change_type());
stasis_subscription_set_filter(acl_change_sub, STASIS_SUBSCRIPTION_FILTER_SELECTIVE);
} else if (!acl_subscription_flag && acl_change_sub) {
acl_change_sub = stasis_unsubscribe_and_join(acl_change_sub);
}
#endif
#if defined(HAVE_OPENSSL) && (OPENSSL_VERSION_NUMBER >= 0x10001000L) && !defined(OPENSSL_NO_SRTP)
if ((s = ast_variable_retrieve(cfg, "general", "dtls_mtu"))) {
if ((sscanf(s, "%d", &dtls_mtu) != 1) || dtls_mtu < 256) {
ast_log(LOG_WARNING, "Value for 'dtls_mtu' could not be read, using default of '%d' instead\n",
DEFAULT_DTLS_MTU);
dtls_mtu = DEFAULT_DTLS_MTU;
}
}
#endif
ast_config_destroy(cfg);
/* Choosing an odd start port casues issues (like a potential infinite loop) and as odd parts are not
chosen anyway, we are going to round up and issue a warning */
if (rtpstart & 1) {
rtpstart++;
ast_log(LOG_WARNING, "Odd start value for RTP port in rtp.conf, rounding up to %d\n", rtpstart);
}
if (rtpstart >= rtpend) {
ast_log(LOG_WARNING, "Unreasonable values for RTP start/end port in rtp.conf\n");
rtpstart = DEFAULT_RTP_START;
rtpend = DEFAULT_RTP_END;
}
ast_verb(2, "RTP Allocating from port range %d -> %d\n", rtpstart, rtpend);
return 0;
}
static int reload_module(void)
{
rtp_reload(1, 0);
return 0;
}
#ifdef HAVE_PJPROJECT
static void rtp_terminate_pjproject(void)
{
pj_thread_register_check();
if (timer_thread) {
timer_terminate = 1;
pj_thread_join(timer_thread);
pj_thread_destroy(timer_thread);
}
ast_pjproject_caching_pool_destroy(&cachingpool);
pj_shutdown();
}
static void acl_change_stasis_cb(void *data, struct stasis_subscription *sub, struct stasis_message *message)
{
if (stasis_message_type(message) != ast_named_acl_change_type()) {
return;
}
/* There is no simple way to just reload the ACLs, so just execute a forced reload. */
rtp_reload(1, 1);
}
#endif
static int load_module(void)
{
#ifdef HAVE_PJPROJECT
pj_lock_t *lock;
ast_sockaddr_parse(&lo6, "::1", PARSE_PORT_IGNORE);
AST_PJPROJECT_INIT_LOG_LEVEL();
if (pj_init() != PJ_SUCCESS) {
return AST_MODULE_LOAD_DECLINE;
}
if (pjlib_util_init() != PJ_SUCCESS) {
rtp_terminate_pjproject();
return AST_MODULE_LOAD_DECLINE;
}
if (pjnath_init() != PJ_SUCCESS) {
rtp_terminate_pjproject();
return AST_MODULE_LOAD_DECLINE;
}
ast_pjproject_caching_pool_init(&cachingpool, &pj_pool_factory_default_policy, 0);
pool = pj_pool_create(&cachingpool.factory, "timer", 512, 512, NULL);
if (pj_timer_heap_create(pool, 100, &timer_heap) != PJ_SUCCESS) {
rtp_terminate_pjproject();
return AST_MODULE_LOAD_DECLINE;
}
if (pj_lock_create_recursive_mutex(pool, "rtp%p", &lock) != PJ_SUCCESS) {
rtp_terminate_pjproject();
return AST_MODULE_LOAD_DECLINE;
}
pj_timer_heap_set_lock(timer_heap, lock, PJ_TRUE);
if (pj_thread_create(pool, "timer", &timer_worker_thread, NULL, 0, 0, &timer_thread) != PJ_SUCCESS) {
rtp_terminate_pjproject();
return AST_MODULE_LOAD_DECLINE;
}
#endif
#if defined(HAVE_OPENSSL) && (OPENSSL_VERSION_NUMBER >= 0x10001000L) && !defined(OPENSSL_NO_SRTP) && defined(HAVE_OPENSSL_BIO_METHOD)
dtls_bio_methods = BIO_meth_new(BIO_TYPE_BIO, "rtp write");
if (!dtls_bio_methods) {
#ifdef HAVE_PJPROJECT
rtp_terminate_pjproject();
#endif
return AST_MODULE_LOAD_DECLINE;
}
BIO_meth_set_write(dtls_bio_methods, dtls_bio_write);
BIO_meth_set_ctrl(dtls_bio_methods, dtls_bio_ctrl);
BIO_meth_set_create(dtls_bio_methods, dtls_bio_new);
BIO_meth_set_destroy(dtls_bio_methods, dtls_bio_free);
#endif
if (ast_rtp_engine_register(&asterisk_rtp_engine)) {
#if defined(HAVE_OPENSSL) && (OPENSSL_VERSION_NUMBER >= 0x10001000L) && !defined(OPENSSL_NO_SRTP) && defined(HAVE_OPENSSL_BIO_METHOD)
BIO_meth_free(dtls_bio_methods);
#endif
#ifdef HAVE_PJPROJECT
rtp_terminate_pjproject();
#endif
return AST_MODULE_LOAD_DECLINE;
}
if (ast_cli_register_multiple(cli_rtp, ARRAY_LEN(cli_rtp))) {
#if defined(HAVE_OPENSSL) && (OPENSSL_VERSION_NUMBER >= 0x10001000L) && !defined(OPENSSL_NO_SRTP) && defined(HAVE_OPENSSL_BIO_METHOD)
BIO_meth_free(dtls_bio_methods);
#endif
#ifdef HAVE_PJPROJECT
ast_rtp_engine_unregister(&asterisk_rtp_engine);
rtp_terminate_pjproject();
#endif
return AST_MODULE_LOAD_DECLINE;
}
rtp_reload(0, 0);
return AST_MODULE_LOAD_SUCCESS;
}
static int unload_module(void)
{
ast_rtp_engine_unregister(&asterisk_rtp_engine);
ast_cli_unregister_multiple(cli_rtp, ARRAY_LEN(cli_rtp));
#if defined(HAVE_OPENSSL) && (OPENSSL_VERSION_NUMBER >= 0x10001000L) && !defined(OPENSSL_NO_SRTP) && defined(HAVE_OPENSSL_BIO_METHOD)
if (dtls_bio_methods) {
BIO_meth_free(dtls_bio_methods);
}
#endif
#ifdef HAVE_PJPROJECT
host_candidate_overrides_clear();
pj_thread_register_check();
rtp_terminate_pjproject();
acl_change_sub = stasis_unsubscribe_and_join(acl_change_sub);
rtp_unload_acl(&ice_acl_lock, &ice_acl);
rtp_unload_acl(&stun_acl_lock, &stun_acl);
clean_stunaddr();
#endif
return 0;
}
AST_MODULE_INFO(ASTERISK_GPL_KEY, AST_MODFLAG_LOAD_ORDER, "Asterisk RTP Stack",
.support_level = AST_MODULE_SUPPORT_CORE,
.load = load_module,
.unload = unload_module,
.reload = reload_module,
.load_pri = AST_MODPRI_CHANNEL_DEPEND,
#ifdef HAVE_PJPROJECT
.requires = "res_pjproject",
#endif
);