asterisk/main/stasis.c

3193 lines
88 KiB
C
Raw Normal View History

/*
* Asterisk -- An open source telephony toolkit.
*
* Copyright (C) 2013, Digium, Inc.
*
* David M. Lee, II <dlee@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 Stasis Message Bus API.
*
* \author David M. Lee, II <dlee@digium.com>
*/
/*** MODULEINFO
<support_level>core</support_level>
***/
#include "asterisk.h"
#include "asterisk/astobj2.h"
Update Asterisk's CDRs for the new bridging framework This patch is the initial push to update Asterisk's CDR engine for the new bridging framework. This patch guts the existing CDR engine and builds the new on top of messages coming across Stasis. As changes in channel state and bridge state are detected, CDRs are built and dispatched accordingly. This fundamentally changes CDRs in a few ways. (1) CDRs are now *very* reflective of the actual state of channels and bridges. This means CDRs track well with what an actual channel is doing - which is useful in transfer scenarios (which were previously difficult to pin down). It does, however, mean that CDRs cannot be 'fooled'. Previous behavior in Asterisk allowed for CDR applications, channels, and other properties to be spoofed in parts of the code - this no longer works. (2) CDRs have defined behavior in multi-party scenarios. This behavior will not be what everyone wants, but it is a defined behavior and as such, it is predictable. (3) The CDR manipulation functions and applications have been overhauled. Major changes have been made to ResetCDR and ForkCDR in particular. Many of the options for these two applications no longer made any sense with the new framework and the (slightly) more immutable nature of CDRs. There are a plethora of other changes. For a full description of CDR behavior, see the CDR specification on the Asterisk wiki. (closes issue ASTERISK-21196) Review: https://reviewboard.asterisk.org/r/2486/ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@391947 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-06-17 03:00:38 +00:00
#include "asterisk/stasis_internal.h"
#include "asterisk/stasis.h"
#include "asterisk/taskprocessor.h"
main/stasis: Allow subscriptions to use a threadpool for message delivery Prior to this patch, all Stasis subscriptions would receive a dedicated thread for servicing published messages. In contrast, prior to r400178 (see review https://reviewboard.asterisk.org/r/2881/), the subscriptions shared a thread pool. It was discovered during some initial work on Stasis that, for a low subscription count with high message throughput, the threadpool was not as performant as simply having a dedicated thread per subscriber. For situations where a subscriber receives a substantial number of messages and is always present, the model of having a dedicated thread per subscriber makes sense. While we still have plenty of subscriptions that would follow this model, e.g., AMI, CDRs, CEL, etc., there are plenty that also fall into the following two categories: * Large number of subscriptions, specifically those tied to endpoints/peers. * Low number of messages. Some subscriptions exist specifically to coordinate a single message - the subscription is created, a message is published, the delivery is synchronized, and the subscription is destroyed. In both of the latter two cases, creating a dedicated thread is wasteful (and in the case of a large number of peers/endpoints, harmful). In those cases, having shared delivery threads is far more performant. This patch adds the ability of a subscriber to Stasis to choose whether or not their messages are dispatched on a dedicated thread or on a threadpool. The threadpool is configurable through stasis.conf. Review: https://reviewboard.asterisk.org/r/4193 ASTERISK-24533 #close Reported by: xrobau Tested by: xrobau ........ Merged revisions 428681 from http://svn.asterisk.org/svn/asterisk/branches/12 ........ Merged revisions 428687 from http://svn.asterisk.org/svn/asterisk/branches/13 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@428688 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-12-01 17:59:21 +00:00
#include "asterisk/threadpool.h"
#include "asterisk/utils.h"
#include "asterisk/uuid.h"
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
#include "asterisk/vector.h"
#include "asterisk/stasis_channels.h"
#include "asterisk/stasis_bridges.h"
#include "asterisk/stasis_endpoints.h"
#include "asterisk/config_options.h"
#include "asterisk/cli.h"
/*** DOCUMENTATION
<managerEvent language="en_US" name="UserEvent">
<managerEventInstance class="EVENT_FLAG_USER">
<synopsis>A user defined event raised from the dialplan.</synopsis>
<syntax>
<channel_snapshot/>
<parameter name="UserEvent">
<para>The event name, as specified in the dialplan.</para>
</parameter>
</syntax>
<description>
<para>Event may contain additional arbitrary parameters in addition to optional bridge and endpoint snapshots. Multiple snapshots of the same type are prefixed with a numeric value.</para>
</description>
<see-also>
<ref type="application">UserEvent</ref>
<ref type="managerEvent">UserEvent</ref>
</see-also>
</managerEventInstance>
</managerEvent>
<configInfo name="stasis" language="en_US">
<configFile name="stasis.conf">
main/stasis: Allow subscriptions to use a threadpool for message delivery Prior to this patch, all Stasis subscriptions would receive a dedicated thread for servicing published messages. In contrast, prior to r400178 (see review https://reviewboard.asterisk.org/r/2881/), the subscriptions shared a thread pool. It was discovered during some initial work on Stasis that, for a low subscription count with high message throughput, the threadpool was not as performant as simply having a dedicated thread per subscriber. For situations where a subscriber receives a substantial number of messages and is always present, the model of having a dedicated thread per subscriber makes sense. While we still have plenty of subscriptions that would follow this model, e.g., AMI, CDRs, CEL, etc., there are plenty that also fall into the following two categories: * Large number of subscriptions, specifically those tied to endpoints/peers. * Low number of messages. Some subscriptions exist specifically to coordinate a single message - the subscription is created, a message is published, the delivery is synchronized, and the subscription is destroyed. In both of the latter two cases, creating a dedicated thread is wasteful (and in the case of a large number of peers/endpoints, harmful). In those cases, having shared delivery threads is far more performant. This patch adds the ability of a subscriber to Stasis to choose whether or not their messages are dispatched on a dedicated thread or on a threadpool. The threadpool is configurable through stasis.conf. Review: https://reviewboard.asterisk.org/r/4193 ASTERISK-24533 #close Reported by: xrobau Tested by: xrobau ........ Merged revisions 428681 from http://svn.asterisk.org/svn/asterisk/branches/12 ........ Merged revisions 428687 from http://svn.asterisk.org/svn/asterisk/branches/13 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@428688 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-12-01 17:59:21 +00:00
<configObject name="threadpool">
<synopsis>Settings that configure the threadpool Stasis uses to deliver some messages.</synopsis>
<configOption name="initial_size" default="5">
<synopsis>Initial number of threads in the message bus threadpool.</synopsis>
</configOption>
<configOption name="idle_timeout_sec" default="20">
<synopsis>Number of seconds before an idle thread is disposed of.</synopsis>
</configOption>
<configOption name="max_size" default="50">
<synopsis>Maximum number of threads in the threadpool.</synopsis>
</configOption>
</configObject>
<configObject name="declined_message_types">
<synopsis>Stasis message types for which to decline creation.</synopsis>
<configOption name="decline">
<synopsis>The message type to decline.</synopsis>
<description>
<para>This configuration option defines the name of the Stasis
message type that Asterisk is forbidden from creating and can be
specified as many times as necessary to achieve the desired result.</para>
<enumlist>
<enum name="stasis_app_recording_snapshot_type" />
<enum name="stasis_app_playback_snapshot_type" />
<enum name="stasis_test_message_type" />
<enum name="confbridge_start_type" />
<enum name="confbridge_end_type" />
<enum name="confbridge_join_type" />
<enum name="confbridge_leave_type" />
<enum name="confbridge_start_record_type" />
<enum name="confbridge_stop_record_type" />
<enum name="confbridge_mute_type" />
<enum name="confbridge_unmute_type" />
<enum name="confbridge_talking_type" />
<enum name="cel_generic_type" />
<enum name="ast_bridge_snapshot_type" />
<enum name="ast_bridge_merge_message_type" />
<enum name="ast_channel_entered_bridge_type" />
<enum name="ast_channel_left_bridge_type" />
<enum name="ast_blind_transfer_type" />
<enum name="ast_attended_transfer_type" />
<enum name="ast_endpoint_snapshot_type" />
<enum name="ast_endpoint_state_type" />
<enum name="ast_device_state_message_type" />
<enum name="ast_test_suite_message_type" />
<enum name="ast_mwi_state_type" />
<enum name="ast_mwi_vm_app_type" />
<enum name="ast_format_register_type" />
<enum name="ast_format_unregister_type" />
<enum name="ast_manager_get_generic_type" />
<enum name="ast_parked_call_type" />
<enum name="ast_channel_snapshot_type" />
<enum name="ast_channel_dial_type" />
<enum name="ast_channel_varset_type" />
<enum name="ast_channel_hangup_request_type" />
<enum name="ast_channel_dtmf_begin_type" />
<enum name="ast_channel_dtmf_end_type" />
<enum name="ast_channel_flash_type" />
<enum name="ast_channel_wink_type" />
<enum name="ast_channel_hold_type" />
<enum name="ast_channel_unhold_type" />
<enum name="ast_channel_chanspy_start_type" />
<enum name="ast_channel_chanspy_stop_type" />
<enum name="ast_channel_fax_type" />
<enum name="ast_channel_hangup_handler_type" />
<enum name="ast_channel_moh_start_type" />
<enum name="ast_channel_moh_stop_type" />
<enum name="ast_channel_mixmonitor_start_type" />
<enum name="ast_channel_mixmonitor_stop_type" />
<enum name="ast_channel_mixmonitor_mute_type" />
<enum name="ast_channel_agent_login_type" />
<enum name="ast_channel_agent_logoff_type" />
<enum name="ast_channel_talking_start" />
<enum name="ast_channel_talking_stop" />
<enum name="ast_security_event_type" />
<enum name="ast_named_acl_change_type" />
<enum name="ast_local_bridge_type" />
<enum name="ast_local_optimization_begin_type" />
<enum name="ast_local_optimization_end_type" />
<enum name="stasis_subscription_change_type" />
<enum name="ast_multi_user_event_type" />
<enum name="stasis_cache_clear_type" />
<enum name="stasis_cache_update_type" />
<enum name="ast_network_change_type" />
<enum name="ast_system_registry_type" />
<enum name="ast_cc_available_type" />
<enum name="ast_cc_offertimerstart_type" />
<enum name="ast_cc_requested_type" />
<enum name="ast_cc_requestacknowledged_type" />
<enum name="ast_cc_callerstopmonitoring_type" />
<enum name="ast_cc_callerstartmonitoring_type" />
<enum name="ast_cc_callerrecalling_type" />
<enum name="ast_cc_recallcomplete_type" />
<enum name="ast_cc_failure_type" />
<enum name="ast_cc_monitorfailed_type" />
<enum name="ast_presence_state_message_type" />
<enum name="ast_rtp_rtcp_sent_type" />
<enum name="ast_rtp_rtcp_received_type" />
<enum name="ast_call_pickup_type" />
<enum name="aoc_s_type" />
<enum name="aoc_d_type" />
<enum name="aoc_e_type" />
<enum name="dahdichannel_type" />
<enum name="mcid_type" />
<enum name="session_timeout_type" />
<enum name="cdr_read_message_type" />
<enum name="cdr_write_message_type" />
<enum name="cdr_prop_write_message_type" />
<enum name="corosync_ping_message_type" />
<enum name="agi_exec_start_type" />
<enum name="agi_exec_end_type" />
<enum name="agi_async_start_type" />
<enum name="agi_async_exec_type" />
<enum name="agi_async_end_type" />
<enum name="queue_caller_join_type" />
<enum name="queue_caller_leave_type" />
<enum name="queue_caller_abandon_type" />
<enum name="queue_member_status_type" />
<enum name="queue_member_added_type" />
<enum name="queue_member_removed_type" />
<enum name="queue_member_pause_type" />
<enum name="queue_member_penalty_type" />
<enum name="queue_member_ringinuse_type" />
<enum name="queue_agent_called_type" />
<enum name="queue_agent_connect_type" />
<enum name="queue_agent_complete_type" />
<enum name="queue_agent_dump_type" />
<enum name="queue_agent_ringnoanswer_type" />
<enum name="meetme_join_type" />
<enum name="meetme_leave_type" />
<enum name="meetme_end_type" />
<enum name="meetme_mute_type" />
<enum name="meetme_talking_type" />
<enum name="meetme_talk_request_type" />
<enum name="appcdr_message_type" />
<enum name="forkcdr_message_type" />
<enum name="cdr_sync_message_type" />
</enumlist>
</description>
</configOption>
</configObject>
</configFile>
</configInfo>
***/
/*!
* \page stasis-impl Stasis Implementation Notes
*
* \par Reference counting
*
* Stasis introduces a number of objects, which are tightly related to one
* another. Because we rely on ref-counting for memory management, understanding
* these relationships is important to understanding this code.
*
* \code{.txt}
*
* stasis_topic <----> stasis_subscription
* ^ ^
* \ /
* \ /
* dispatch
* |
* |
* v
* stasis_message
* |
* |
* v
* stasis_message_type
*
* \endcode
*
* The most troubling thing in this chart is the cyclic reference between
* stasis_topic and stasis_subscription. This is both unfortunate, and
* necessary. Topics need the subscription in order to dispatch messages;
* subscriptions need the topics to unsubscribe and check subscription status.
*
* The cycle is broken by stasis_unsubscribe(). The unsubscribe will remove the
* topic's reference to a subscription. When the subcription is destroyed, it
* will remove its reference to the topic.
*
* This means that until a subscription has be explicitly unsubscribed, it will
* not be destroyed. Neither will a topic be destroyed while it has subscribers.
* The destructors of both have assertions regarding this to catch ref-counting
* problems where a subscription or topic has had an extra ao2_cleanup().
*
* The \ref dispatch_exec_sync object is a transient object, which is posted to
* a subscription's taskprocessor to send a message to the subscriber. They have
* short life cycles, allocated on one thread, destroyed on another.
*
* During shutdown, or the deletion of a domain object, there are a flurry of
* ao2_cleanup()s on subscriptions and topics, as the final in-flight messages
* are processed. Any one of these cleanups could be the one to actually destroy
* a given object, so care must be taken to ensure that an object isn't
* referenced after an ao2_cleanup(). This includes the implicit ao2_unlock()
* that might happen when a RAII_VAR() goes out of scope.
*
* \par Typical life cycles
*
* \li stasis_topic - There are several topics which live for the duration of
* the Asterisk process (ast_channel_topic_all(), etc.) but most of these
* are actually fed by shorter-lived topics whose lifetime is associated
* with some domain object (like ast_channel_topic() for a given
* ast_channel).
*
* \li stasis_subscription - Subscriptions have a similar mix of lifetimes as
* topics, for similar reasons.
*
* \li dispatch - Very short lived; just long enough to post a message to a
* subscriber.
*
* \li stasis_message - Short to intermediate lifetimes, but that is mostly
* irrelevant. Messages are strictly data and have no behavior associated
* with them, so it doesn't really matter if/when they are destroyed. By
* design, a component could hold a ref to a message forever without any
* ill consequences (aside from consuming more memory).
*
* \li stasis_message_type - Long life cycles, typically only destroyed on
* module unloading or _clean_ process exit.
*
* \par Subscriber shutdown sequencing
*
* Subscribers are sensitive to shutdown sequencing, specifically in how the
* reference message types. This is fully detailed in the documentation at
* https://docs.asterisk.org/Development/Roadmap/Asterisk-12-Projects/Asterisk-12-API-Improvements/Stasis-Message-Bus/Using-the-Stasis-Message-Bus/Stasis-Subscriber-Shutdown-Problem/.
*
* In short, the lifetime of the \a data (and \a callback, if in a module) must
* be held until the stasis_subscription_final_message() has been received.
* Depending on the structure of the subscriber code, this can be handled by
* using stasis_subscription_final_message() to free resources on the final
* message, or using stasis_subscription_join()/stasis_unsubscribe_and_join() to
* block until the unsubscribe has completed.
*/
/*! Initial size of the subscribers list. */
#define INITIAL_SUBSCRIBERS_MAX 4
/*! The number of buckets to use for topic pools */
#define TOPIC_POOL_BUCKETS 57
main/stasis: Allow subscriptions to use a threadpool for message delivery Prior to this patch, all Stasis subscriptions would receive a dedicated thread for servicing published messages. In contrast, prior to r400178 (see review https://reviewboard.asterisk.org/r/2881/), the subscriptions shared a thread pool. It was discovered during some initial work on Stasis that, for a low subscription count with high message throughput, the threadpool was not as performant as simply having a dedicated thread per subscriber. For situations where a subscriber receives a substantial number of messages and is always present, the model of having a dedicated thread per subscriber makes sense. While we still have plenty of subscriptions that would follow this model, e.g., AMI, CDRs, CEL, etc., there are plenty that also fall into the following two categories: * Large number of subscriptions, specifically those tied to endpoints/peers. * Low number of messages. Some subscriptions exist specifically to coordinate a single message - the subscription is created, a message is published, the delivery is synchronized, and the subscription is destroyed. In both of the latter two cases, creating a dedicated thread is wasteful (and in the case of a large number of peers/endpoints, harmful). In those cases, having shared delivery threads is far more performant. This patch adds the ability of a subscriber to Stasis to choose whether or not their messages are dispatched on a dedicated thread or on a threadpool. The threadpool is configurable through stasis.conf. Review: https://reviewboard.asterisk.org/r/4193 ASTERISK-24533 #close Reported by: xrobau Tested by: xrobau ........ Merged revisions 428681 from http://svn.asterisk.org/svn/asterisk/branches/12 ........ Merged revisions 428687 from http://svn.asterisk.org/svn/asterisk/branches/13 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@428688 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-12-01 17:59:21 +00:00
/*! Thread pool for topics that don't want a dedicated taskprocessor */
static struct ast_threadpool *threadpool;
main/stasis: Allow subscriptions to use a threadpool for message delivery Prior to this patch, all Stasis subscriptions would receive a dedicated thread for servicing published messages. In contrast, prior to r400178 (see review https://reviewboard.asterisk.org/r/2881/), the subscriptions shared a thread pool. It was discovered during some initial work on Stasis that, for a low subscription count with high message throughput, the threadpool was not as performant as simply having a dedicated thread per subscriber. For situations where a subscriber receives a substantial number of messages and is always present, the model of having a dedicated thread per subscriber makes sense. While we still have plenty of subscriptions that would follow this model, e.g., AMI, CDRs, CEL, etc., there are plenty that also fall into the following two categories: * Large number of subscriptions, specifically those tied to endpoints/peers. * Low number of messages. Some subscriptions exist specifically to coordinate a single message - the subscription is created, a message is published, the delivery is synchronized, and the subscription is destroyed. In both of the latter two cases, creating a dedicated thread is wasteful (and in the case of a large number of peers/endpoints, harmful). In those cases, having shared delivery threads is far more performant. This patch adds the ability of a subscriber to Stasis to choose whether or not their messages are dispatched on a dedicated thread or on a threadpool. The threadpool is configurable through stasis.conf. Review: https://reviewboard.asterisk.org/r/4193 ASTERISK-24533 #close Reported by: xrobau Tested by: xrobau ........ Merged revisions 428681 from http://svn.asterisk.org/svn/asterisk/branches/12 ........ Merged revisions 428687 from http://svn.asterisk.org/svn/asterisk/branches/13 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@428688 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-12-01 17:59:21 +00:00
STASIS_MESSAGE_TYPE_DEFN(stasis_subscription_change_type);
#if defined(LOW_MEMORY)
#define TOPIC_ALL_BUCKETS 257
#else
#define TOPIC_ALL_BUCKETS 997
#endif
#ifdef AST_DEVMODE
/*! The number of buckets to use for topic statistics */
#define TOPIC_STATISTICS_BUCKETS 57
/*! The number of buckets to use for subscription statistics */
#define SUBSCRIPTION_STATISTICS_BUCKETS 57
/*! Global container which stores statistics for topics */
static AO2_GLOBAL_OBJ_STATIC(topic_statistics);
/*! Global container which stores statistics for subscriptions */
static AO2_GLOBAL_OBJ_STATIC(subscription_statistics);
/*! \internal */
struct stasis_message_type_statistics {
/*! \brief The number of messages of this published */
int published;
/*! \brief The number of messages of this that did not reach a subscriber */
int unused;
/*! \brief The stasis message type */
struct stasis_message_type *message_type;
};
/*! Lock to protect the message types vector */
AST_MUTEX_DEFINE_STATIC(message_type_statistics_lock);
/*! Vector containing message type information */
static AST_VECTOR(, struct stasis_message_type_statistics) message_type_statistics;
/*! \internal */
struct stasis_topic_statistics {
/*! \brief Highest time spent dispatching messages to subscribers */
long highest_time_dispatched;
/*! \brief Lowest time spent dispatching messages to subscribers */
long lowest_time_dispatched;
/*! \brief The number of messages that were not dispatched to any subscriber */
int messages_not_dispatched;
/*! \brief The number of messages that were dispatched to at least 1 subscriber */
int messages_dispatched;
/*! \brief The ids of the subscribers to this topic */
struct ao2_container *subscribers;
/*! \brief Pointer to the topic (NOT refcounted, and must NOT be accessed) */
struct stasis_topic *topic;
/*! \brief Name of the topic */
char name[0];
};
#endif
/*! \internal */
struct stasis_topic {
/*! Variable length array of the subscribers */
AST_VECTOR(, struct stasis_subscription *) subscribers;
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
/*! Topics forwarding into this topic */
AST_VECTOR(, struct stasis_topic *) upstream_topics;
#ifdef AST_DEVMODE
struct stasis_topic_statistics *statistics;
#endif
/*! Unique incrementing integer for subscriber ids */
int subscriber_id;
/*! Name of the topic */
char *name;
/*! Detail of the topic */
char *detail;
/*! Creation time */
struct timeval *creationtime;
};
struct ao2_container *topic_all;
struct topic_proxy {
AO2_WEAKPROXY();
char *name;
char *detail;
struct timeval creationtime;
char buf[0];
};
AO2_STRING_FIELD_HASH_FN(topic_proxy, name);
AO2_STRING_FIELD_CMP_FN(topic_proxy, name);
AO2_STRING_FIELD_CASE_SORT_FN(topic_proxy, name);
static void proxy_dtor(void *weakproxy, void *container)
{
ao2_unlink(container, weakproxy);
ao2_cleanup(container);
}
/* Forward declarations for the tightly-coupled subscription object */
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
static int topic_add_subscription(struct stasis_topic *topic,
struct stasis_subscription *sub);
static int topic_remove_subscription(struct stasis_topic *topic, struct stasis_subscription *sub);
/*! \brief Lock two topics. */
#define topic_lock_both(topic1, topic2) \
do { \
ao2_lock(topic1); \
while (ao2_trylock(topic2)) { \
AO2_DEADLOCK_AVOIDANCE(topic1); \
} \
} while (0)
static void topic_dtor(void *obj)
{
struct stasis_topic *topic = obj;
#ifdef AST_DEVMODE
struct ao2_container *topic_stats;
#endif
ast_debug(2, "Destroying topic. name: %s, detail: %s\n",
topic->name, topic->detail);
/* Subscribers hold a reference to topics, so they should all be
* unsubscribed before we get here. */
ast_assert(AST_VECTOR_SIZE(&topic->subscribers) == 0);
AST_VECTOR_FREE(&topic->subscribers);
AST_VECTOR_FREE(&topic->upstream_topics);
ast_debug(1, "Topic '%s': %p destroyed\n", topic->name, topic);
#ifdef AST_DEVMODE
if (topic->statistics) {
topic_stats = ao2_global_obj_ref(topic_statistics);
if (topic_stats) {
ao2_unlink(topic_stats, topic->statistics);
ao2_ref(topic_stats, -1);
}
ao2_ref(topic->statistics, -1);
}
#endif
}
#ifdef AST_DEVMODE
static void topic_statistics_destroy(void *obj)
{
struct stasis_topic_statistics *statistics = obj;
ao2_cleanup(statistics->subscribers);
}
static struct stasis_topic_statistics *stasis_topic_statistics_create(struct stasis_topic *topic)
{
struct stasis_topic_statistics *statistics;
RAII_VAR(struct ao2_container *, topic_stats, ao2_global_obj_ref(topic_statistics), ao2_cleanup);
if (!topic_stats) {
return NULL;
}
statistics = ao2_alloc(sizeof(*statistics) + strlen(topic->name) + 1, topic_statistics_destroy);
if (!statistics) {
return NULL;
}
statistics->subscribers = ast_str_container_alloc(1);
if (!statistics->subscribers) {
ao2_ref(statistics, -1);
return NULL;
}
/* This is strictly used for the pointer address when showing the topic */
statistics->topic = topic;
strcpy(statistics->name, topic->name); /* SAFE */
ao2_link(topic_stats, statistics);
return statistics;
}
#endif
static int link_topic_proxy(struct stasis_topic *topic, const char *name, const char *detail)
{
struct topic_proxy *proxy;
struct stasis_topic* topic_tmp;
size_t detail_len;
if (!topic || !name || !strlen(name) || !detail) {
return -1;
}
ao2_wrlock(topic_all);
topic_tmp = stasis_topic_get(name);
if (topic_tmp) {
ast_log(LOG_ERROR, "The same topic is already exist. name: %s\n", name);
ao2_ref(topic_tmp, -1);
ao2_unlock(topic_all);
return -1;
}
detail_len = strlen(detail) + 1;
proxy = ao2_t_weakproxy_alloc(
sizeof(*proxy) + strlen(name) + 1 + detail_len, NULL, name);
if (!proxy) {
ao2_unlock(topic_all);
return -1;
}
/* set the proxy info */
proxy->name = proxy->buf;
proxy->detail = proxy->name + strlen(name) + 1;
strcpy(proxy->name, name); /* SAFE */
ast_copy_string(proxy->detail, detail, detail_len); /* SAFE */
proxy->creationtime = ast_tvnow();
/* We have exclusive access to proxy, no need for locking here. */
if (ao2_t_weakproxy_set_object(proxy, topic, OBJ_NOLOCK, "weakproxy link")) {
ao2_cleanup(proxy);
ao2_unlock(topic_all);
return -1;
}
if (ao2_weakproxy_subscribe(proxy, proxy_dtor, ao2_bump(topic_all), OBJ_NOLOCK)) {
ao2_cleanup(proxy);
ao2_unlock(topic_all);
ao2_cleanup(topic_all);
return -1;
}
/* setting the topic point to the proxy */
topic->name = proxy->name;
topic->detail = proxy->detail;
topic->creationtime = &(proxy->creationtime);
ao2_link_flags(topic_all, proxy, OBJ_NOLOCK);
ao2_ref(proxy, -1);
ao2_unlock(topic_all);
return 0;
}
struct stasis_topic *stasis_topic_create_with_detail(
const char *name, const char* detail
)
{
struct stasis_topic *topic;
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
int res = 0;
if (!name|| !strlen(name) || !detail) {
return NULL;
}
ast_debug(2, "Creating topic. name: %s, detail: %s\n", name, detail);
topic = stasis_topic_get(name);
if (topic) {
ast_debug(2, "Topic is already exist. name: %s, detail: %s\n",
name, detail);
return topic;
}
topic = ao2_t_alloc(sizeof(*topic), topic_dtor, name);
if (!topic) {
return NULL;
}
res |= AST_VECTOR_INIT(&topic->subscribers, INITIAL_SUBSCRIBERS_MAX);
res |= AST_VECTOR_INIT(&topic->upstream_topics, 0);
if (res) {
ao2_ref(topic, -1);
return NULL;
}
/* link to the proxy */
if (link_topic_proxy(topic, name, detail)) {
ao2_ref(topic, -1);
return NULL;
}
#ifdef AST_DEVMODE
topic->statistics = stasis_topic_statistics_create(topic);
if (!topic->statistics) {
ao2_ref(topic, -1);
return NULL;
}
#endif
ast_debug(1, "Topic '%s': %p created\n", topic->name, topic);
return topic;
}
struct stasis_topic *stasis_topic_create(const char *name)
{
return stasis_topic_create_with_detail(name, "");
}
struct stasis_topic *stasis_topic_get(const char *name)
{
return ao2_weakproxy_find(topic_all, name, OBJ_SEARCH_KEY, "");
}
const char *stasis_topic_name(const struct stasis_topic *topic)
{
if (!topic) {
return NULL;
}
return topic->name;
}
const char *stasis_topic_detail(const struct stasis_topic *topic)
{
if (!topic) {
return NULL;
}
return topic->detail;
}
size_t stasis_topic_subscribers(const struct stasis_topic *topic)
{
return AST_VECTOR_SIZE(&topic->subscribers);
}
#ifdef AST_DEVMODE
struct stasis_subscription_statistics {
/*! \brief The filename where the subscription originates */
const char *file;
/*! \brief The function where the subscription originates */
const char *func;
/*! \brief Names of the topics we are subscribed to */
struct ao2_container *topics;
/*! \brief The message type that currently took the longest to process */
struct stasis_message_type *highest_time_message_type;
/*! \brief Highest time spent invoking a message */
long highest_time_invoked;
/*! \brief Lowest time spent invoking a message */
long lowest_time_invoked;
/*! \brief The number of messages that were filtered out */
int messages_dropped;
/*! \brief The number of messages that passed filtering */
int messages_passed;
/*! \brief Using a mailbox to queue messages */
int uses_mailbox;
/*! \brief Using stasis threadpool for handling messages */
int uses_threadpool;
/*! \brief The line number where the subscription originates */
int lineno;
/*! \brief Pointer to the subscription (NOT refcounted, and must NOT be accessed) */
struct stasis_subscription *sub;
/*! \brief Unique ID of the subscription */
char uniqueid[0];
};
#endif
/*! \internal */
struct stasis_subscription {
/*! Unique ID for this subscription */
char *uniqueid;
/*! Topic subscribed to. */
struct stasis_topic *topic;
/*! Mailbox for processing incoming messages. */
struct ast_taskprocessor *mailbox;
/*! Callback function for incoming message processing. */
stasis_subscription_cb callback;
/*! Data pointer to be handed to the callback. */
void *data;
/*! Condition for joining with subscription. */
ast_cond_t join_cond;
/*! Flag set when final message for sub has been received.
* Be sure join_lock is held before reading/setting. */
int final_message_rxed;
/*! Flag set when final message for sub has been processed.
* Be sure join_lock is held before reading/setting. */
int final_message_processed;
/*! The message types this subscription is accepting */
AST_VECTOR(, char) accepted_message_types;
/*! The message formatters this subscription is accepting */
enum stasis_subscription_message_formatters accepted_formatters;
/*! The message filter currently in use */
enum stasis_subscription_message_filter filter;
#ifdef AST_DEVMODE
/*! Statistics information */
struct stasis_subscription_statistics *statistics;
#endif
};
static void subscription_dtor(void *obj)
{
struct stasis_subscription *sub = obj;
#ifdef AST_DEVMODE
struct ao2_container *subscription_stats;
#endif
/* Subscriptions need to be manually unsubscribed before destruction
* b/c there's a cyclic reference between topics and subscriptions */
ast_assert(!stasis_subscription_is_subscribed(sub));
/* If there are any messages in flight to this subscription; that would
* be bad. */
ast_assert(stasis_subscription_is_done(sub));
ast_free(sub->uniqueid);
ao2_cleanup(sub->topic);
sub->topic = NULL;
ast_taskprocessor_unreference(sub->mailbox);
sub->mailbox = NULL;
ast_cond_destroy(&sub->join_cond);
AST_VECTOR_FREE(&sub->accepted_message_types);
#ifdef AST_DEVMODE
if (sub->statistics) {
subscription_stats = ao2_global_obj_ref(subscription_statistics);
if (subscription_stats) {
ao2_unlink(subscription_stats, sub->statistics);
ao2_ref(subscription_stats, -1);
}
ao2_ref(sub->statistics, -1);
}
#endif
}
/*!
* \brief Invoke the subscription's callback.
* \param sub Subscription to invoke.
* \param message Message to send.
*/
static void subscription_invoke(struct stasis_subscription *sub,
struct stasis_message *message)
{
unsigned int final = stasis_subscription_final_message(sub, message);
int message_type_id = stasis_message_type_id(stasis_subscription_change_type());
#ifdef AST_DEVMODE
struct timeval start;
long elapsed;
start = ast_tvnow();
#endif
/* Notify that the final message has been received */
if (final) {
ao2_lock(sub);
sub->final_message_rxed = 1;
ast_cond_signal(&sub->join_cond);
ao2_unlock(sub);
}
/*
* If filtering is turned on and this is a 'final' message, we only invoke the callback
* if the subscriber accepts subscription_change message types.
*/
if (!final || sub->filter != STASIS_SUBSCRIPTION_FILTER_SELECTIVE ||
(message_type_id < AST_VECTOR_SIZE(&sub->accepted_message_types) && AST_VECTOR_GET(&sub->accepted_message_types, message_type_id))) {
/* Since sub is mostly immutable, no need to lock sub */
sub->callback(sub->data, sub, message);
}
/* Notify that the final message has been processed */
if (final) {
ao2_lock(sub);
sub->final_message_processed = 1;
ast_cond_signal(&sub->join_cond);
ao2_unlock(sub);
}
#ifdef AST_DEVMODE
elapsed = ast_tvdiff_ms(ast_tvnow(), start);
if (elapsed > sub->statistics->highest_time_invoked) {
sub->statistics->highest_time_invoked = elapsed;
ao2_lock(sub->statistics);
sub->statistics->highest_time_message_type = stasis_message_type(message);
ao2_unlock(sub->statistics);
}
if (elapsed < sub->statistics->lowest_time_invoked) {
sub->statistics->lowest_time_invoked = elapsed;
}
#endif
}
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
static void send_subscription_subscribe(struct stasis_topic *topic, struct stasis_subscription *sub);
static void send_subscription_unsubscribe(struct stasis_topic *topic, struct stasis_subscription *sub);
void stasis_subscription_cb_noop(void *data, struct stasis_subscription *sub, struct stasis_message *message)
{
}
#ifdef AST_DEVMODE
static void subscription_statistics_destroy(void *obj)
{
struct stasis_subscription_statistics *statistics = obj;
ao2_cleanup(statistics->topics);
}
static struct stasis_subscription_statistics *stasis_subscription_statistics_create(struct stasis_subscription *sub,
int needs_mailbox, int use_thread_pool, const char *file, int lineno,
const char *func)
{
struct stasis_subscription_statistics *statistics;
RAII_VAR(struct ao2_container *, subscription_stats, ao2_global_obj_ref(subscription_statistics), ao2_cleanup);
if (!subscription_stats) {
return NULL;
}
statistics = ao2_alloc(sizeof(*statistics) + strlen(sub->uniqueid) + 1, subscription_statistics_destroy);
if (!statistics) {
return NULL;
}
statistics->topics = ast_str_container_alloc(1);
if (!statistics->topics) {
ao2_ref(statistics, -1);
return NULL;
}
statistics->file = file;
statistics->lineno = lineno;
statistics->func = func;
statistics->uses_mailbox = needs_mailbox;
statistics->uses_threadpool = use_thread_pool;
strcpy(statistics->uniqueid, sub->uniqueid); /* SAFE */
statistics->sub = sub;
ao2_link(subscription_stats, statistics);
return statistics;
}
#endif
struct stasis_subscription *internal_stasis_subscribe(
struct stasis_topic *topic,
stasis_subscription_cb callback,
void *data,
int needs_mailbox,
int use_thread_pool,
const char *file,
int lineno,
const char *func)
{
struct stasis_subscription *sub;
int ret;
if (!topic) {
return NULL;
}
/* The ao2 lock is used for join_cond. */
sub = ao2_t_alloc(sizeof(*sub), subscription_dtor, stasis_topic_name(topic));
if (!sub) {
return NULL;
}
#ifdef AST_DEVMODE
ret = ast_asprintf(&sub->uniqueid, "%s:%s-%d", file, stasis_topic_name(topic), ast_atomic_fetchadd_int(&topic->subscriber_id, +1));
sub->statistics = stasis_subscription_statistics_create(sub, needs_mailbox, use_thread_pool, file, lineno, func);
if (ret < 0 || !sub->statistics) {
ao2_ref(sub, -1);
return NULL;
}
#else
ret = ast_asprintf(&sub->uniqueid, "%s-%d", stasis_topic_name(topic), ast_atomic_fetchadd_int(&topic->subscriber_id, +1));
if (ret < 0) {
ao2_ref(sub, -1);
return NULL;
}
#endif
if (needs_mailbox) {
char tps_name[AST_TASKPROCESSOR_MAX_NAME + 1];
/* Create name with seq number appended. */
taskprocessor: Enable subsystems and overload by subsystem To prevent one subsystem's taskprocessors from causing others to stall, new capabilities have been added to taskprocessors. * Any taskprocessor name that has a '/' will have the part before the '/' saved as its "subsystem". Examples: "sorcery/acl-0000006a" and "sorcery/aor-00000019" will be grouped to subsystem "sorcery". "pjsip/distributor-00000025" and "pjsip/distributor-00000026" will bn grouped to subsystem "pjsip". Taskprocessors with no '/' have an empty subsystem. * When a taskprocessor enters high-water alert status and it has a non-empty subsystem, the subsystem alert count will be incremented. * When a taskprocessor leaves high-water alert status and it has a non-empty subsystem, the subsystem alert count will be decremented. * A new api ast_taskprocessor_get_subsystem_alert() has been added that returns the number of taskprocessors in alert for the subsystem. * A new CLI command "core show taskprocessor alerted subsystems" has been added. * A new unit test was addded. REMINDER: The taskprocessor code itself doesn't take any action based on high-water alerts or overloading. It's up to taskprocessor users to check and take action themselves. Currently only the pjsip distributor does this. * A new pjsip/global option "taskprocessor_overload_trigger" has been added that allows the user to select the trigger mechanism the distributor uses to pause accepting new requests. "none": Don't pause on any overload condition. "global": Pause on ANY taskprocessor overload (the default and current behavior) "pjsip_only": Pause only on pjsip taskprocessor overloads. * The core pjsip pool was renamed from "SIP" to "pjsip" so it can be properly grouped into the "pjsip" subsystem. * stasis taskprocessor names were changed to "stasis" as the subsystem. * Sorcery core taskprocessor names were changed to "sorcery" to match the object taskprocessors. Change-Id: I8c19068bb2fc26610a9f0b8624bdf577a04fcd56
2019-02-15 18:53:50 +00:00
ast_taskprocessor_build_name(tps_name, sizeof(tps_name), "stasis/%c:%s",
use_thread_pool ? 'p' : 'm',
stasis_topic_name(topic));
/*
* With a small number of subscribers, a thread-per-sub is
* acceptable. For a large number of subscribers, a thread
main/stasis: Allow subscriptions to use a threadpool for message delivery Prior to this patch, all Stasis subscriptions would receive a dedicated thread for servicing published messages. In contrast, prior to r400178 (see review https://reviewboard.asterisk.org/r/2881/), the subscriptions shared a thread pool. It was discovered during some initial work on Stasis that, for a low subscription count with high message throughput, the threadpool was not as performant as simply having a dedicated thread per subscriber. For situations where a subscriber receives a substantial number of messages and is always present, the model of having a dedicated thread per subscriber makes sense. While we still have plenty of subscriptions that would follow this model, e.g., AMI, CDRs, CEL, etc., there are plenty that also fall into the following two categories: * Large number of subscriptions, specifically those tied to endpoints/peers. * Low number of messages. Some subscriptions exist specifically to coordinate a single message - the subscription is created, a message is published, the delivery is synchronized, and the subscription is destroyed. In both of the latter two cases, creating a dedicated thread is wasteful (and in the case of a large number of peers/endpoints, harmful). In those cases, having shared delivery threads is far more performant. This patch adds the ability of a subscriber to Stasis to choose whether or not their messages are dispatched on a dedicated thread or on a threadpool. The threadpool is configurable through stasis.conf. Review: https://reviewboard.asterisk.org/r/4193 ASTERISK-24533 #close Reported by: xrobau Tested by: xrobau ........ Merged revisions 428681 from http://svn.asterisk.org/svn/asterisk/branches/12 ........ Merged revisions 428687 from http://svn.asterisk.org/svn/asterisk/branches/13 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@428688 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-12-01 17:59:21 +00:00
* pool should be used.
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
*/
main/stasis: Allow subscriptions to use a threadpool for message delivery Prior to this patch, all Stasis subscriptions would receive a dedicated thread for servicing published messages. In contrast, prior to r400178 (see review https://reviewboard.asterisk.org/r/2881/), the subscriptions shared a thread pool. It was discovered during some initial work on Stasis that, for a low subscription count with high message throughput, the threadpool was not as performant as simply having a dedicated thread per subscriber. For situations where a subscriber receives a substantial number of messages and is always present, the model of having a dedicated thread per subscriber makes sense. While we still have plenty of subscriptions that would follow this model, e.g., AMI, CDRs, CEL, etc., there are plenty that also fall into the following two categories: * Large number of subscriptions, specifically those tied to endpoints/peers. * Low number of messages. Some subscriptions exist specifically to coordinate a single message - the subscription is created, a message is published, the delivery is synchronized, and the subscription is destroyed. In both of the latter two cases, creating a dedicated thread is wasteful (and in the case of a large number of peers/endpoints, harmful). In those cases, having shared delivery threads is far more performant. This patch adds the ability of a subscriber to Stasis to choose whether or not their messages are dispatched on a dedicated thread or on a threadpool. The threadpool is configurable through stasis.conf. Review: https://reviewboard.asterisk.org/r/4193 ASTERISK-24533 #close Reported by: xrobau Tested by: xrobau ........ Merged revisions 428681 from http://svn.asterisk.org/svn/asterisk/branches/12 ........ Merged revisions 428687 from http://svn.asterisk.org/svn/asterisk/branches/13 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@428688 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-12-01 17:59:21 +00:00
if (use_thread_pool) {
sub->mailbox = ast_threadpool_serializer(tps_name, threadpool);
main/stasis: Allow subscriptions to use a threadpool for message delivery Prior to this patch, all Stasis subscriptions would receive a dedicated thread for servicing published messages. In contrast, prior to r400178 (see review https://reviewboard.asterisk.org/r/2881/), the subscriptions shared a thread pool. It was discovered during some initial work on Stasis that, for a low subscription count with high message throughput, the threadpool was not as performant as simply having a dedicated thread per subscriber. For situations where a subscriber receives a substantial number of messages and is always present, the model of having a dedicated thread per subscriber makes sense. While we still have plenty of subscriptions that would follow this model, e.g., AMI, CDRs, CEL, etc., there are plenty that also fall into the following two categories: * Large number of subscriptions, specifically those tied to endpoints/peers. * Low number of messages. Some subscriptions exist specifically to coordinate a single message - the subscription is created, a message is published, the delivery is synchronized, and the subscription is destroyed. In both of the latter two cases, creating a dedicated thread is wasteful (and in the case of a large number of peers/endpoints, harmful). In those cases, having shared delivery threads is far more performant. This patch adds the ability of a subscriber to Stasis to choose whether or not their messages are dispatched on a dedicated thread or on a threadpool. The threadpool is configurable through stasis.conf. Review: https://reviewboard.asterisk.org/r/4193 ASTERISK-24533 #close Reported by: xrobau Tested by: xrobau ........ Merged revisions 428681 from http://svn.asterisk.org/svn/asterisk/branches/12 ........ Merged revisions 428687 from http://svn.asterisk.org/svn/asterisk/branches/13 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@428688 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-12-01 17:59:21 +00:00
} else {
sub->mailbox = ast_taskprocessor_get(tps_name, TPS_REF_DEFAULT);
main/stasis: Allow subscriptions to use a threadpool for message delivery Prior to this patch, all Stasis subscriptions would receive a dedicated thread for servicing published messages. In contrast, prior to r400178 (see review https://reviewboard.asterisk.org/r/2881/), the subscriptions shared a thread pool. It was discovered during some initial work on Stasis that, for a low subscription count with high message throughput, the threadpool was not as performant as simply having a dedicated thread per subscriber. For situations where a subscriber receives a substantial number of messages and is always present, the model of having a dedicated thread per subscriber makes sense. While we still have plenty of subscriptions that would follow this model, e.g., AMI, CDRs, CEL, etc., there are plenty that also fall into the following two categories: * Large number of subscriptions, specifically those tied to endpoints/peers. * Low number of messages. Some subscriptions exist specifically to coordinate a single message - the subscription is created, a message is published, the delivery is synchronized, and the subscription is destroyed. In both of the latter two cases, creating a dedicated thread is wasteful (and in the case of a large number of peers/endpoints, harmful). In those cases, having shared delivery threads is far more performant. This patch adds the ability of a subscriber to Stasis to choose whether or not their messages are dispatched on a dedicated thread or on a threadpool. The threadpool is configurable through stasis.conf. Review: https://reviewboard.asterisk.org/r/4193 ASTERISK-24533 #close Reported by: xrobau Tested by: xrobau ........ Merged revisions 428681 from http://svn.asterisk.org/svn/asterisk/branches/12 ........ Merged revisions 428687 from http://svn.asterisk.org/svn/asterisk/branches/13 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@428688 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-12-01 17:59:21 +00:00
}
if (!sub->mailbox) {
ao2_ref(sub, -1);
return NULL;
}
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
ast_taskprocessor_set_local(sub->mailbox, sub);
/* Taskprocessor has a reference */
ao2_ref(sub, +1);
}
ao2_ref(topic, +1);
sub->topic = topic;
sub->callback = callback;
sub->data = data;
ast_cond_init(&sub->join_cond, NULL);
sub->filter = STASIS_SUBSCRIPTION_FILTER_NONE;
AST_VECTOR_INIT(&sub->accepted_message_types, 0);
sub->accepted_formatters = STASIS_SUBSCRIPTION_FORMATTER_NONE;
if (topic_add_subscription(topic, sub) != 0) {
ao2_ref(sub, -1);
ao2_ref(topic, -1);
return NULL;
}
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
send_subscription_subscribe(topic, sub);
return sub;
}
struct stasis_subscription *__stasis_subscribe(
struct stasis_topic *topic,
stasis_subscription_cb callback,
void *data,
const char *file,
int lineno,
const char *func)
{
return internal_stasis_subscribe(topic, callback, data, 1, 0, file, lineno, func);
}
main/stasis: Allow subscriptions to use a threadpool for message delivery Prior to this patch, all Stasis subscriptions would receive a dedicated thread for servicing published messages. In contrast, prior to r400178 (see review https://reviewboard.asterisk.org/r/2881/), the subscriptions shared a thread pool. It was discovered during some initial work on Stasis that, for a low subscription count with high message throughput, the threadpool was not as performant as simply having a dedicated thread per subscriber. For situations where a subscriber receives a substantial number of messages and is always present, the model of having a dedicated thread per subscriber makes sense. While we still have plenty of subscriptions that would follow this model, e.g., AMI, CDRs, CEL, etc., there are plenty that also fall into the following two categories: * Large number of subscriptions, specifically those tied to endpoints/peers. * Low number of messages. Some subscriptions exist specifically to coordinate a single message - the subscription is created, a message is published, the delivery is synchronized, and the subscription is destroyed. In both of the latter two cases, creating a dedicated thread is wasteful (and in the case of a large number of peers/endpoints, harmful). In those cases, having shared delivery threads is far more performant. This patch adds the ability of a subscriber to Stasis to choose whether or not their messages are dispatched on a dedicated thread or on a threadpool. The threadpool is configurable through stasis.conf. Review: https://reviewboard.asterisk.org/r/4193 ASTERISK-24533 #close Reported by: xrobau Tested by: xrobau ........ Merged revisions 428681 from http://svn.asterisk.org/svn/asterisk/branches/12 ........ Merged revisions 428687 from http://svn.asterisk.org/svn/asterisk/branches/13 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@428688 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-12-01 17:59:21 +00:00
struct stasis_subscription *__stasis_subscribe_pool(
struct stasis_topic *topic,
stasis_subscription_cb callback,
void *data,
const char *file,
int lineno,
const char *func)
{
return internal_stasis_subscribe(topic, callback, data, 1, 1, file, lineno, func);
}
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
static int sub_cleanup(void *data)
{
struct stasis_subscription *sub = data;
ao2_cleanup(sub);
return 0;
}
struct stasis_subscription *stasis_unsubscribe(struct stasis_subscription *sub)
{
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
/* The subscription may be the last ref to this topic. Hold
* the topic ref open until after the unlock. */
struct stasis_topic *topic;
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
if (!sub) {
return NULL;
}
topic = ao2_bump(sub->topic);
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
/* We have to remove the subscription first, to ensure the unsubscribe
* is the final message */
if (topic_remove_subscription(sub->topic, sub) != 0) {
ast_log(LOG_ERROR,
"Internal error: subscription has invalid topic\n");
ao2_cleanup(topic);
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
return NULL;
}
/* Now let everyone know about the unsubscribe */
send_subscription_unsubscribe(topic, sub);
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
/* When all that's done, remove the ref the mailbox has on the sub */
if (sub->mailbox) {
if (ast_taskprocessor_push(sub->mailbox, sub_cleanup, sub)) {
/* Nothing we can do here, the conditional is just to keep
* the compiler happy that we're not ignoring the result. */
}
}
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
/* Unsubscribing unrefs the subscription */
ao2_cleanup(sub);
ao2_cleanup(topic);
return NULL;
}
int stasis_subscription_set_congestion_limits(struct stasis_subscription *subscription,
long low_water, long high_water)
{
int res = -1;
if (subscription) {
res = ast_taskprocessor_alert_set_levels(subscription->mailbox,
low_water, high_water);
}
return res;
}
int stasis_subscription_accept_message_type(struct stasis_subscription *subscription,
const struct stasis_message_type *type)
{
if (!subscription) {
return -1;
}
ast_assert(type != NULL);
ast_assert(stasis_message_type_name(type) != NULL);
if (!type || !stasis_message_type_name(type)) {
/* Filtering is unreliable as this message type is not yet initialized
* so force all messages through.
*/
subscription->filter = STASIS_SUBSCRIPTION_FILTER_FORCED_NONE;
return 0;
}
ao2_lock(subscription->topic);
if (AST_VECTOR_REPLACE(&subscription->accepted_message_types, stasis_message_type_id(type), 1)) {
/* We do this for the same reason as above. The subscription can still operate, so allow
* it to do so by forcing all messages through.
*/
subscription->filter = STASIS_SUBSCRIPTION_FILTER_FORCED_NONE;
}
ao2_unlock(subscription->topic);
return 0;
}
int stasis_subscription_decline_message_type(struct stasis_subscription *subscription,
const struct stasis_message_type *type)
{
if (!subscription) {
return -1;
}
ast_assert(type != NULL);
ast_assert(stasis_message_type_name(type) != NULL);
if (!type || !stasis_message_type_name(type)) {
return 0;
}
ao2_lock(subscription->topic);
if (stasis_message_type_id(type) < AST_VECTOR_SIZE(&subscription->accepted_message_types)) {
/* The memory is already allocated so this can't fail */
AST_VECTOR_REPLACE(&subscription->accepted_message_types, stasis_message_type_id(type), 0);
}
ao2_unlock(subscription->topic);
return 0;
}
int stasis_subscription_set_filter(struct stasis_subscription *subscription,
enum stasis_subscription_message_filter filter)
{
if (!subscription) {
return -1;
}
ao2_lock(subscription->topic);
if (subscription->filter != STASIS_SUBSCRIPTION_FILTER_FORCED_NONE) {
subscription->filter = filter;
}
ao2_unlock(subscription->topic);
return 0;
}
void stasis_subscription_accept_formatters(struct stasis_subscription *subscription,
enum stasis_subscription_message_formatters formatters)
{
ast_assert(subscription != NULL);
ao2_lock(subscription->topic);
subscription->accepted_formatters = formatters;
ao2_unlock(subscription->topic);
return;
}
void stasis_subscription_join(struct stasis_subscription *subscription)
{
if (subscription) {
ao2_lock(subscription);
/* Wait until the processed flag has been set */
while (!subscription->final_message_processed) {
ast_cond_wait(&subscription->join_cond,
ao2_object_get_lockaddr(subscription));
}
ao2_unlock(subscription);
}
}
int stasis_subscription_is_done(struct stasis_subscription *subscription)
{
if (subscription) {
int ret;
ao2_lock(subscription);
ret = subscription->final_message_rxed;
ao2_unlock(subscription);
return ret;
}
/* Null subscription is about as done as you can get */
return 1;
}
struct stasis_subscription *stasis_unsubscribe_and_join(
struct stasis_subscription *subscription)
{
if (!subscription) {
return NULL;
}
/* Bump refcount to hold it past the unsubscribe */
ao2_ref(subscription, +1);
stasis_unsubscribe(subscription);
stasis_subscription_join(subscription);
/* Now decrement the refcount back */
ao2_cleanup(subscription);
return NULL;
}
int stasis_subscription_is_subscribed(const struct stasis_subscription *sub)
{
if (sub) {
size_t i;
struct stasis_topic *topic = sub->topic;
ao2_lock(topic);
for (i = 0; i < AST_VECTOR_SIZE(&topic->subscribers); ++i) {
if (AST_VECTOR_GET(&topic->subscribers, i) == sub) {
ao2_unlock(topic);
return 1;
}
}
ao2_unlock(topic);
}
return 0;
}
const char *stasis_subscription_uniqueid(const struct stasis_subscription *sub)
{
return sub->uniqueid;
}
int stasis_subscription_final_message(struct stasis_subscription *sub, struct stasis_message *msg)
{
struct stasis_subscription_change *change;
if (stasis_message_type(msg) != stasis_subscription_change_type()) {
return 0;
}
change = stasis_message_data(msg);
if (strcmp("Unsubscribe", change->description)) {
return 0;
}
if (strcmp(stasis_subscription_uniqueid(sub), change->uniqueid)) {
return 0;
}
return 1;
}
/*!
* \brief Add a subscriber to a topic.
* \param topic Topic
* \param sub Subscriber
* \return 0 on success
* \return Non-zero on error
*/
static int topic_add_subscription(struct stasis_topic *topic, struct stasis_subscription *sub)
{
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
size_t idx;
ao2_lock(topic);
/* The reference from the topic to the subscription is shared with
* the owner of the subscription, which will explicitly unsubscribe
* to release it.
*
* If we bumped the refcount here, the owner would have to unsubscribe
* and cleanup, which is a bit awkward. */
AST_VECTOR_APPEND(&topic->subscribers, sub);
for (idx = 0; idx < AST_VECTOR_SIZE(&topic->upstream_topics); ++idx) {
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
topic_add_subscription(
AST_VECTOR_GET(&topic->upstream_topics, idx), sub);
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
}
#ifdef AST_DEVMODE
ast_str_container_add(topic->statistics->subscribers, stasis_subscription_uniqueid(sub));
ast_str_container_add(sub->statistics->topics, stasis_topic_name(topic));
#endif
ao2_unlock(topic);
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
return 0;
}
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
static int topic_remove_subscription(struct stasis_topic *topic, struct stasis_subscription *sub)
{
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
size_t idx;
int res;
ao2_lock(topic);
for (idx = 0; idx < AST_VECTOR_SIZE(&topic->upstream_topics); ++idx) {
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
topic_remove_subscription(
AST_VECTOR_GET(&topic->upstream_topics, idx), sub);
}
res = AST_VECTOR_REMOVE_ELEM_UNORDERED(&topic->subscribers, sub,
AST_VECTOR_ELEM_CLEANUP_NOOP);
#ifdef AST_DEVMODE
if (!res) {
ast_str_container_remove(topic->statistics->subscribers, stasis_subscription_uniqueid(sub));
ast_str_container_remove(sub->statistics->topics, stasis_topic_name(topic));
}
#endif
ao2_unlock(topic);
return res;
}
/*!
* \internal \brief Dispatch a message to a subscriber asynchronously
* \param local \ref ast_taskprocessor_local object
* \return 0
*/
static int dispatch_exec_async(struct ast_taskprocessor_local *local)
{
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
struct stasis_subscription *sub = local->local_data;
struct stasis_message *message = local->data;
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
subscription_invoke(sub, message);
ao2_cleanup(message);
return 0;
}
/*!
* \internal \brief Data passed to \ref dispatch_exec_sync to synchronize
* a published message to a subscriber
*/
struct sync_task_data {
ast_mutex_t lock;
ast_cond_t cond;
int complete;
void *task_data;
};
/*!
* \internal \brief Dispatch a message to a subscriber synchronously
* \param local \ref ast_taskprocessor_local object
* \return 0
*/
static int dispatch_exec_sync(struct ast_taskprocessor_local *local)
{
struct stasis_subscription *sub = local->local_data;
struct sync_task_data *std = local->data;
struct stasis_message *message = std->task_data;
subscription_invoke(sub, message);
ao2_cleanup(message);
ast_mutex_lock(&std->lock);
std->complete = 1;
ast_cond_signal(&std->cond);
ast_mutex_unlock(&std->lock);
return 0;
}
/*!
* \internal \brief Dispatch a message to a subscriber
* \param sub The subscriber to dispatch to
* \param message The message to send
* \param synchronous If non-zero, synchronize on the subscriber receiving
* the message
* \retval 0 if message was not dispatched
* \retval 1 if message was dispatched
*/
static unsigned int dispatch_message(struct stasis_subscription *sub,
struct stasis_message *message,
int synchronous)
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
{
int is_final = stasis_subscription_final_message(sub, message);
/*
* The 'do while' gives us an easy way to skip remaining logic once
* we determine the message should be accepted.
* The code looks more verbose than it needs to be but it optimizes
* down very nicely. It's just easier to understand and debug this way.
*/
do {
struct stasis_message_type *message_type = stasis_message_type(message);
int type_id = stasis_message_type_id(message_type);
int type_filter_specified = 0;
int formatter_filter_specified = 0;
int type_filter_passed = 0;
int formatter_filter_passed = 0;
/* We always accept final messages so only run the filter logic if not final */
if (is_final) {
break;
}
type_filter_specified = sub->filter & STASIS_SUBSCRIPTION_FILTER_SELECTIVE;
formatter_filter_specified = sub->accepted_formatters != STASIS_SUBSCRIPTION_FORMATTER_NONE;
/* Accept if no filters of either type were specified */
if (!type_filter_specified && !formatter_filter_specified) {
break;
}
type_filter_passed = type_filter_specified
&& type_id < AST_VECTOR_SIZE(&sub->accepted_message_types)
&& AST_VECTOR_GET(&sub->accepted_message_types, type_id);
/*
* Since the type and formatter filters are OR'd, we can skip
* the formatter check if the type check passes.
*/
if (type_filter_passed) {
break;
}
formatter_filter_passed = formatter_filter_specified
&& (sub->accepted_formatters & stasis_message_type_available_formatters(message_type));
if (formatter_filter_passed) {
break;
}
#ifdef AST_DEVMODE
ast_atomic_fetchadd_int(&sub->statistics->messages_dropped, +1);
#endif
return 0;
} while (0);
#ifdef AST_DEVMODE
ast_atomic_fetchadd_int(&sub->statistics->messages_passed, +1);
#endif
if (!sub->mailbox) {
/* Dispatch directly */
subscription_invoke(sub, message);
return 1;
}
/* Bump the message for the taskprocessor push. This will get de-ref'd
* by the task processor callback.
*/
ao2_bump(message);
if (!synchronous) {
if (ast_taskprocessor_push_local(sub->mailbox, dispatch_exec_async, message)) {
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
/* Push failed; ugh. */
ast_log(LOG_ERROR, "Dropping async dispatch\n");
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
ao2_cleanup(message);
return 0;
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
}
} else {
struct sync_task_data std;
ast_mutex_init(&std.lock);
ast_cond_init(&std.cond, NULL);
std.complete = 0;
std.task_data = message;
if (ast_taskprocessor_push_local(sub->mailbox, dispatch_exec_sync, &std)) {
/* Push failed; ugh. */
ast_log(LOG_ERROR, "Dropping sync dispatch\n");
ao2_cleanup(message);
ast_mutex_destroy(&std.lock);
ast_cond_destroy(&std.cond);
return 0;
}
ast_mutex_lock(&std.lock);
while (!std.complete) {
ast_cond_wait(&std.cond, &std.lock);
}
ast_mutex_unlock(&std.lock);
ast_mutex_destroy(&std.lock);
ast_cond_destroy(&std.cond);
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
}
return 1;
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
}
/*!
* \internal \brief Publish a message to a topic's subscribers
* \brief topic The topic to publish to
* \brief message The message to publish
* \brief sync_sub An optional subscriber of the topic to publish synchronously
* to
*/
static void publish_msg(struct stasis_topic *topic,
struct stasis_message *message, struct stasis_subscription *sync_sub)
{
size_t i;
#ifdef AST_DEVMODE
unsigned int dispatched = 0;
int message_type_id = stasis_message_type_id(stasis_message_type(message));
struct stasis_message_type_statistics *statistics;
struct timeval start;
long elapsed;
#endif
ast_assert(topic != NULL);
ast_assert(message != NULL);
#ifdef AST_DEVMODE
ast_mutex_lock(&message_type_statistics_lock);
if (message_type_id >= AST_VECTOR_SIZE(&message_type_statistics)) {
struct stasis_message_type_statistics new_statistics = {
.published = 0,
};
if (AST_VECTOR_REPLACE(&message_type_statistics, message_type_id, new_statistics)) {
ast_mutex_unlock(&message_type_statistics_lock);
return;
}
}
statistics = AST_VECTOR_GET_ADDR(&message_type_statistics, message_type_id);
statistics->message_type = stasis_message_type(message);
ast_mutex_unlock(&message_type_statistics_lock);
ast_atomic_fetchadd_int(&statistics->published, +1);
#endif
/* If there are no subscribers don't bother */
if (!stasis_topic_subscribers(topic)) {
#ifdef AST_DEVMODE
ast_atomic_fetchadd_int(&statistics->unused, +1);
ast_atomic_fetchadd_int(&topic->statistics->messages_not_dispatched, +1);
#endif
return;
}
/*
* The topic may be unref'ed by the subscription invocation.
* Make sure we hold onto a reference while dispatching.
*/
ao2_ref(topic, +1);
#ifdef AST_DEVMODE
start = ast_tvnow();
#endif
ao2_lock(topic);
for (i = 0; i < AST_VECTOR_SIZE(&topic->subscribers); ++i) {
struct stasis_subscription *sub = AST_VECTOR_GET(&topic->subscribers, i);
ast_assert(sub != NULL);
#ifdef AST_DEVMODE
dispatched +=
#endif
dispatch_message(sub, message, (sub == sync_sub));
}
ao2_unlock(topic);
#ifdef AST_DEVMODE
elapsed = ast_tvdiff_ms(ast_tvnow(), start);
if (elapsed > topic->statistics->highest_time_dispatched) {
topic->statistics->highest_time_dispatched = elapsed;
}
if (elapsed < topic->statistics->lowest_time_dispatched) {
topic->statistics->lowest_time_dispatched = elapsed;
}
if (dispatched) {
ast_atomic_fetchadd_int(&topic->statistics->messages_dispatched, +1);
} else {
ast_atomic_fetchadd_int(&statistics->unused, +1);
ast_atomic_fetchadd_int(&topic->statistics->messages_not_dispatched, +1);
}
#endif
ao2_ref(topic, -1);
}
void stasis_publish(struct stasis_topic *topic, struct stasis_message *message)
{
publish_msg(topic, message, NULL);
}
void stasis_publish_sync(struct stasis_subscription *sub, struct stasis_message *message)
{
ast_assert(sub != NULL);
publish_msg(sub->topic, message, sub);
}
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
/*!
* \brief Forwarding information
*
* Any message posted to \a from_topic is forwarded to \a to_topic.
*
* In cases where both the \a from_topic and \a to_topic need to be locked,
* always lock the \a to_topic first, then the \a from_topic. Lest you deadlock.
*/
struct stasis_forward {
/*! Originating topic */
struct stasis_topic *from_topic;
/*! Destination topic */
struct stasis_topic *to_topic;
};
static void forward_dtor(void *obj)
{
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
struct stasis_forward *forward = obj;
ao2_cleanup(forward->from_topic);
forward->from_topic = NULL;
ao2_cleanup(forward->to_topic);
forward->to_topic = NULL;
}
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
struct stasis_forward *stasis_forward_cancel(struct stasis_forward *forward)
{
int idx;
struct stasis_topic *from;
struct stasis_topic *to;
if (!forward) {
return NULL;
}
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
from = forward->from_topic;
to = forward->to_topic;
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
res_corosync: Update module to work with Stasis (and compile) This patch fixes res_corosync such that it works with Asterisk 12. This restores the functionality that was present in previous versions of Asterisk, and ensures compatibility with those versions by restoring the binary message format needed to pass information from/to them. The following changes were made in the core to support this: * The event system has been partially restored. All event definition and event types in this patch were pulled from Asterisk 11. Previously, we had hoped that this information would live in res_corosync; however, the approach in this patch seems to be better for a few reasons: (1) Theoretically, ast_events can be used by any module as a binary representation of a Stasis message. Given the structure of an ast_event object, that information has to live in the core to be used universally. For example, defining the payload of a device state ast_event in res_corosync could result in an incompatible device state representation in another module. (2) Much of this representation already lived in the core, and was not easily extensible. (3) The code already existed. :-) * Stasis message types now have a message formatter that converts their payload to an ast_event object. * Stasis message forwarders now handle forwarding to themselves. Previously this would result in an infinite recursive call. Now, this simply creates a new forwarding object with no forwards set up (as it is the thing it is forwarding to). This is advantageous for res_corosync, as returning NULL would also imply an unrecoverable error. Returning a subscription in this case allows for easier handling of message types that are published directly to an aggregate topic that has forwarders. Review: https://reviewboard.asterisk.org/r/3486/ ASTERISK-22912 #close ASTERISK-22372 #close ........ Merged revisions 414330 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@414331 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-05-22 12:01:37 +00:00
if (from && to) {
topic_lock_both(to, from);
AST_VECTOR_REMOVE_ELEM_UNORDERED(&to->upstream_topics, from,
AST_VECTOR_ELEM_CLEANUP_NOOP);
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
res_corosync: Update module to work with Stasis (and compile) This patch fixes res_corosync such that it works with Asterisk 12. This restores the functionality that was present in previous versions of Asterisk, and ensures compatibility with those versions by restoring the binary message format needed to pass information from/to them. The following changes were made in the core to support this: * The event system has been partially restored. All event definition and event types in this patch were pulled from Asterisk 11. Previously, we had hoped that this information would live in res_corosync; however, the approach in this patch seems to be better for a few reasons: (1) Theoretically, ast_events can be used by any module as a binary representation of a Stasis message. Given the structure of an ast_event object, that information has to live in the core to be used universally. For example, defining the payload of a device state ast_event in res_corosync could result in an incompatible device state representation in another module. (2) Much of this representation already lived in the core, and was not easily extensible. (3) The code already existed. :-) * Stasis message types now have a message formatter that converts their payload to an ast_event object. * Stasis message forwarders now handle forwarding to themselves. Previously this would result in an infinite recursive call. Now, this simply creates a new forwarding object with no forwards set up (as it is the thing it is forwarding to). This is advantageous for res_corosync, as returning NULL would also imply an unrecoverable error. Returning a subscription in this case allows for easier handling of message types that are published directly to an aggregate topic that has forwarders. Review: https://reviewboard.asterisk.org/r/3486/ ASTERISK-22912 #close ASTERISK-22372 #close ........ Merged revisions 414330 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@414331 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-05-22 12:01:37 +00:00
for (idx = 0; idx < AST_VECTOR_SIZE(&to->subscribers); ++idx) {
topic_remove_subscription(from, AST_VECTOR_GET(&to->subscribers, idx));
}
ao2_unlock(from);
ao2_unlock(to);
}
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
ao2_cleanup(forward);
return NULL;
}
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
struct stasis_forward *stasis_forward_all(struct stasis_topic *from_topic,
struct stasis_topic *to_topic)
{
int res;
size_t idx;
struct stasis_forward *forward;
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
if (!from_topic || !to_topic) {
return NULL;
}
forward = ao2_alloc_options(sizeof(*forward), forward_dtor, AO2_ALLOC_OPT_LOCK_NOLOCK);
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
if (!forward) {
return NULL;
}
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
res_corosync: Update module to work with Stasis (and compile) This patch fixes res_corosync such that it works with Asterisk 12. This restores the functionality that was present in previous versions of Asterisk, and ensures compatibility with those versions by restoring the binary message format needed to pass information from/to them. The following changes were made in the core to support this: * The event system has been partially restored. All event definition and event types in this patch were pulled from Asterisk 11. Previously, we had hoped that this information would live in res_corosync; however, the approach in this patch seems to be better for a few reasons: (1) Theoretically, ast_events can be used by any module as a binary representation of a Stasis message. Given the structure of an ast_event object, that information has to live in the core to be used universally. For example, defining the payload of a device state ast_event in res_corosync could result in an incompatible device state representation in another module. (2) Much of this representation already lived in the core, and was not easily extensible. (3) The code already existed. :-) * Stasis message types now have a message formatter that converts their payload to an ast_event object. * Stasis message forwarders now handle forwarding to themselves. Previously this would result in an infinite recursive call. Now, this simply creates a new forwarding object with no forwards set up (as it is the thing it is forwarding to). This is advantageous for res_corosync, as returning NULL would also imply an unrecoverable error. Returning a subscription in this case allows for easier handling of message types that are published directly to an aggregate topic that has forwarders. Review: https://reviewboard.asterisk.org/r/3486/ ASTERISK-22912 #close ASTERISK-22372 #close ........ Merged revisions 414330 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@414331 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-05-22 12:01:37 +00:00
/* Forwards to ourselves are implicit. */
if (to_topic == from_topic) {
return forward;
res_corosync: Update module to work with Stasis (and compile) This patch fixes res_corosync such that it works with Asterisk 12. This restores the functionality that was present in previous versions of Asterisk, and ensures compatibility with those versions by restoring the binary message format needed to pass information from/to them. The following changes were made in the core to support this: * The event system has been partially restored. All event definition and event types in this patch were pulled from Asterisk 11. Previously, we had hoped that this information would live in res_corosync; however, the approach in this patch seems to be better for a few reasons: (1) Theoretically, ast_events can be used by any module as a binary representation of a Stasis message. Given the structure of an ast_event object, that information has to live in the core to be used universally. For example, defining the payload of a device state ast_event in res_corosync could result in an incompatible device state representation in another module. (2) Much of this representation already lived in the core, and was not easily extensible. (3) The code already existed. :-) * Stasis message types now have a message formatter that converts their payload to an ast_event object. * Stasis message forwarders now handle forwarding to themselves. Previously this would result in an infinite recursive call. Now, this simply creates a new forwarding object with no forwards set up (as it is the thing it is forwarding to). This is advantageous for res_corosync, as returning NULL would also imply an unrecoverable error. Returning a subscription in this case allows for easier handling of message types that are published directly to an aggregate topic that has forwarders. Review: https://reviewboard.asterisk.org/r/3486/ ASTERISK-22912 #close ASTERISK-22372 #close ........ Merged revisions 414330 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@414331 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-05-22 12:01:37 +00:00
}
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
forward->from_topic = ao2_bump(from_topic);
forward->to_topic = ao2_bump(to_topic);
topic_lock_both(to_topic, from_topic);
res = AST_VECTOR_APPEND(&to_topic->upstream_topics, from_topic);
if (res != 0) {
ao2_unlock(from_topic);
ao2_unlock(to_topic);
ao2_ref(forward, -1);
return NULL;
}
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
for (idx = 0; idx < AST_VECTOR_SIZE(&to_topic->subscribers); ++idx) {
topic_add_subscription(from_topic, AST_VECTOR_GET(&to_topic->subscribers, idx));
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
}
ao2_unlock(from_topic);
ao2_unlock(to_topic);
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
return forward;
}
static void subscription_change_dtor(void *obj)
{
struct stasis_subscription_change *change = obj;
ao2_cleanup(change->topic);
}
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
static struct stasis_subscription_change *subscription_change_alloc(struct stasis_topic *topic, const char *uniqueid, const char *description)
{
size_t description_len = strlen(description) + 1;
size_t uniqueid_len = strlen(uniqueid) + 1;
struct stasis_subscription_change *change;
change = ao2_alloc_options(sizeof(*change) + description_len + uniqueid_len,
subscription_change_dtor, AO2_ALLOC_OPT_LOCK_NOLOCK);
if (!change) {
return NULL;
}
strcpy(change->description, description); /* SAFE */
change->uniqueid = change->description + description_len;
ast_copy_string(change->uniqueid, uniqueid, uniqueid_len); /* SAFE */
ao2_ref(topic, +1);
change->topic = topic;
return change;
}
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
static void send_subscription_subscribe(struct stasis_topic *topic, struct stasis_subscription *sub)
{
struct stasis_subscription_change *change;
struct stasis_message *msg;
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
/* This assumes that we have already unsubscribed */
ast_assert(stasis_subscription_is_subscribed(sub));
if (!stasis_subscription_change_type()) {
return;
}
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
change = subscription_change_alloc(topic, sub->uniqueid, "Subscribe");
if (!change) {
return;
}
msg = stasis_message_create(stasis_subscription_change_type(), change);
if (!msg) {
ao2_cleanup(change);
return;
}
stasis_publish(topic, msg);
ao2_cleanup(msg);
ao2_cleanup(change);
}
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
static void send_subscription_unsubscribe(struct stasis_topic *topic,
struct stasis_subscription *sub)
{
struct stasis_subscription_change *change;
struct stasis_message *msg;
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
/* This assumes that we have already unsubscribed */
ast_assert(!stasis_subscription_is_subscribed(sub));
if (!stasis_subscription_change_type()) {
return;
}
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
change = subscription_change_alloc(topic, sub->uniqueid, "Unsubscribe");
if (!change) {
return;
}
msg = stasis_message_create(stasis_subscription_change_type(), change);
if (!msg) {
ao2_cleanup(change);
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
return;
}
stasis_publish(topic, msg);
/* Now we have to dispatch to the subscription itself */
dispatch_message(sub, msg, 0);
ao2_cleanup(msg);
ao2_cleanup(change);
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
}
struct topic_pool_entry {
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
struct stasis_forward *forward;
struct stasis_topic *topic;
char name[0];
};
static void topic_pool_entry_dtor(void *obj)
{
struct topic_pool_entry *entry = obj;
Multiple revisions 399887,400138,400178,400180-400181 ........ r399887 | dlee | 2013-09-26 10:41:47 -0500 (Thu, 26 Sep 2013) | 1 line Minor performance bump by not allocate manager variable struct if we don't need it ........ r400138 | dlee | 2013-09-30 10:24:00 -0500 (Mon, 30 Sep 2013) | 23 lines Stasis performance improvements This patch addresses several performance problems that were found in the initial performance testing of Asterisk 12. The Stasis dispatch object was allocated as an AO2 object, even though it has a very confined lifecycle. This was replaced with a straight ast_malloc(). The Stasis message router was spending an inordinate amount of time searching hash tables. In this case, most of our routers had 6 or fewer routes in them to begin with. This was replaced with an array that's searched linearly for the route. We more heavily rely on AO2 objects in Asterisk 12, and the memset() in ao2_ref() actually became noticeable on the profile. This was #ifdef'ed to only run when AO2_DEBUG was enabled. After being misled by an erroneous comment in taskprocessor.c during profiling, the wrong comment was removed. Review: https://reviewboard.asterisk.org/r/2873/ ........ r400178 | dlee | 2013-09-30 13:26:27 -0500 (Mon, 30 Sep 2013) | 24 lines Taskprocessor optimization; switch Stasis to use taskprocessors This patch optimizes taskprocessor to use a semaphore for signaling, which the OS can do a better job at managing contention and waiting that we can with a mutex and condition. The taskprocessor execution was also slightly optimized to reduce the number of locks taken. The only observable difference in the taskprocessor implementation is that when the final reference to the taskprocessor goes away, it will execute all tasks to completion instead of discarding the unexecuted tasks. For systems where unnamed semaphores are not supported, a really simple semaphore implementation is provided. (Which gives identical performance as the original taskprocessor implementation). The way we ended up implementing Stasis caused the threadpool to be a burden instead of a boost to performance. This was switched to just use taskprocessors directly for subscriptions. Review: https://reviewboard.asterisk.org/r/2881/ ........ r400180 | dlee | 2013-09-30 13:39:34 -0500 (Mon, 30 Sep 2013) | 28 lines Optimize how Stasis forwards are dispatched This patch optimizes how forwards are dispatched in Stasis. Originally, forwards were dispatched as subscriptions that are invoked on the publishing thread. This did not account for the vast number of forwards we would end up having in the system, and the amount of work it would take to walk though the forward subscriptions. This patch modifies Stasis so that rather than walking the tree of forwards on every dispatch, when forwards and subscriptions are changed, the subscriber list for every topic in the tree is changed. This has a couple of benefits. First, this reduces the workload of dispatching messages. It also reduces contention when dispatching to different topics that happen to forward to the same aggregation topic (as happens with all of the channel, bridge and endpoint topics). Since forwards are no longer subscriptions, the bulk of this patch is simply changing stasis_subscription objects to stasis_forward objects (which, admittedly, I should have done in the first place.) Since this required me to yet again put in a growing array, I finally abstracted that out into a set of ast_vector macros in asterisk/vector.h. Review: https://reviewboard.asterisk.org/r/2883/ ........ r400181 | dlee | 2013-09-30 13:48:57 -0500 (Mon, 30 Sep 2013) | 28 lines Remove dispatch object allocation from Stasis publishing While looking for areas for performance improvement, I realized that an unused feature in Stasis was negatively impacting performance. When a message is sent to a subscriber, a dispatch object is allocated for the dispatch, containing the topic the message was published to, the subscriber the message is being sent to, and the message itself. The topic is actually unused by any subscriber in Asterisk today. And the subscriber is associated with the taskprocessor the message is being dispatched to. First, this patch removes the unused topic parameter from Stasis subscription callbacks. Second, this patch introduces the concept of taskprocessor local data, data that may be set on a taskprocessor and provided along with the data pointer when a task is pushed using the ast_taskprocessor_push_local() call. This allows the task to have both data specific to that taskprocessor, in addition to data specific to that invocation. With those two changes, the dispatch object can be removed completely, and the message is simply refcounted and sent directly to the taskprocessor. Review: https://reviewboard.asterisk.org/r/2884/ ........ Merged revisions 399887,400138,400178,400180-400181 from http://svn.asterisk.org/svn/asterisk/branches/12 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@400186 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-09-30 18:55:27 +00:00
entry->forward = stasis_forward_cancel(entry->forward);
ao2_cleanup(entry->topic);
entry->topic = NULL;
}
static struct topic_pool_entry *topic_pool_entry_alloc(const char *topic_name)
{
struct topic_pool_entry *topic_pool_entry;
topic_pool_entry = ao2_alloc_options(sizeof(*topic_pool_entry) + strlen(topic_name) + 1,
topic_pool_entry_dtor, AO2_ALLOC_OPT_LOCK_NOLOCK);
if (!topic_pool_entry) {
return NULL;
}
strcpy(topic_pool_entry->name, topic_name); /* Safe */
return topic_pool_entry;
}
struct stasis_topic_pool {
struct ao2_container *pool_container;
struct stasis_topic *pool_topic;
};
static void topic_pool_dtor(void *obj)
{
struct stasis_topic_pool *pool = obj;
#ifdef AO2_DEBUG
{
char *container_name =
ast_alloca(strlen(stasis_topic_name(pool->pool_topic)) + strlen("-pool") + 1);
sprintf(container_name, "%s-pool", stasis_topic_name(pool->pool_topic));
ao2_container_unregister(container_name);
}
#endif
ao2_cleanup(pool->pool_container);
pool->pool_container = NULL;
ao2_cleanup(pool->pool_topic);
pool->pool_topic = NULL;
}
static int topic_pool_entry_hash(const void *obj, const int flags)
{
const struct topic_pool_entry *object;
const char *key;
switch (flags & OBJ_SEARCH_MASK) {
case OBJ_SEARCH_KEY:
key = obj;
break;
case OBJ_SEARCH_OBJECT:
object = obj;
key = object->name;
break;
default:
/* Hash can only work on something with a full key. */
ast_assert(0);
return 0;
}
return ast_str_case_hash(key);
}
static int topic_pool_entry_cmp(void *obj, void *arg, int flags)
{
const struct topic_pool_entry *object_left = obj;
const struct topic_pool_entry *object_right = arg;
const char *right_key = arg;
int cmp;
switch (flags & OBJ_SEARCH_MASK) {
case OBJ_SEARCH_OBJECT:
right_key = object_right->name;
/* Fall through */
case OBJ_SEARCH_KEY:
cmp = strcasecmp(object_left->name, right_key);
break;
case OBJ_SEARCH_PARTIAL_KEY:
/* Not supported by container */
ast_assert(0);
cmp = -1;
break;
default:
/*
* What arg points to is specific to this traversal callback
* and has no special meaning to astobj2.
*/
cmp = 0;
break;
}
if (cmp) {
return 0;
}
/*
* At this point the traversal callback is identical to a sorted
* container.
*/
return CMP_MATCH;
}
#ifdef AO2_DEBUG
static void topic_pool_prnt_obj(void *v_obj, void *where, ao2_prnt_fn *prnt)
{
struct topic_pool_entry *entry = v_obj;
if (!entry) {
return;
}
prnt(where, "%s", stasis_topic_name(entry->topic));
}
#endif
struct stasis_topic_pool *stasis_topic_pool_create(struct stasis_topic *pooled_topic)
{
struct stasis_topic_pool *pool;
pool = ao2_alloc_options(sizeof(*pool), topic_pool_dtor, AO2_ALLOC_OPT_LOCK_NOLOCK);
if (!pool) {
return NULL;
}
pool->pool_container = ao2_container_alloc_hash(AO2_ALLOC_OPT_LOCK_MUTEX, 0,
TOPIC_POOL_BUCKETS, topic_pool_entry_hash, NULL, topic_pool_entry_cmp);
if (!pool->pool_container) {
ao2_cleanup(pool);
return NULL;
}
#ifdef AO2_DEBUG
{
char *container_name =
ast_alloca(strlen(stasis_topic_name(pooled_topic)) + strlen("-pool") + 1);
sprintf(container_name, "%s-pool", stasis_topic_name(pooled_topic));
ao2_container_register(container_name, pool->pool_container, topic_pool_prnt_obj);
}
#endif
ao2_ref(pooled_topic, +1);
pool->pool_topic = pooled_topic;
return pool;
}
void stasis_topic_pool_delete_topic(struct stasis_topic_pool *pool, const char *topic_name)
{
/*
* The topic_name passed in could be a fully-qualified name like <pool_topic_name>/<topic_name>
* or just <topic_name>. If it's fully qualified, we need to skip past <pool_topic_name>
* name and search only on <topic_name>.
*/
const char *pool_topic_name = stasis_topic_name(pool->pool_topic);
int pool_topic_name_len = strlen(pool_topic_name);
const char *search_topic_name;
if (strncmp(pool_topic_name, topic_name, pool_topic_name_len) == 0) {
search_topic_name = topic_name + pool_topic_name_len + 1;
} else {
search_topic_name = topic_name;
}
ao2_find(pool->pool_container, search_topic_name, OBJ_SEARCH_KEY | OBJ_NODATA | OBJ_UNLINK);
}
struct stasis_topic *stasis_topic_pool_get_topic(struct stasis_topic_pool *pool, const char *topic_name)
{
RAII_VAR(struct topic_pool_entry *, topic_pool_entry, NULL, ao2_cleanup);
SCOPED_AO2LOCK(topic_container_lock, pool->pool_container);
char *new_topic_name;
int ret;
topic_pool_entry = ao2_find(pool->pool_container, topic_name, OBJ_SEARCH_KEY | OBJ_NOLOCK);
if (topic_pool_entry) {
return topic_pool_entry->topic;
}
topic_pool_entry = topic_pool_entry_alloc(topic_name);
if (!topic_pool_entry) {
return NULL;
}
/* To provide further detail and to ensure that the topic is unique within the scope of the
* system we prefix it with the pooling topic name, which should itself already be unique.
*/
ret = ast_asprintf(&new_topic_name, "%s/%s", stasis_topic_name(pool->pool_topic), topic_name);
if (ret < 0) {
return NULL;
}
topic_pool_entry->topic = stasis_topic_create(new_topic_name);
ast_free(new_topic_name);
if (!topic_pool_entry->topic) {
return NULL;
}
topic_pool_entry->forward = stasis_forward_all(topic_pool_entry->topic, pool->pool_topic);
if (!topic_pool_entry->forward) {
return NULL;
}
if (!ao2_link_flags(pool->pool_container, topic_pool_entry, OBJ_NOLOCK)) {
return NULL;
}
return topic_pool_entry->topic;
}
int stasis_topic_pool_topic_exists(const struct stasis_topic_pool *pool, const char *topic_name)
{
struct topic_pool_entry *topic_pool_entry;
topic_pool_entry = ao2_find(pool->pool_container, topic_name, OBJ_SEARCH_KEY);
if (!topic_pool_entry) {
return 0;
}
ao2_ref(topic_pool_entry, -1);
return 1;
}
Avoid unnecessary cleanups during immediate shutdown This patch addresses issues during immediate shutdowns, where modules are not unloaded, but Asterisk atexit handlers are run. In the typical case, this usually isn't a big deal. But the introduction of the Stasis message bus makes it much more likely for asynchronous activity to be happening off in some thread during shutdown. During an immediate shutdown, Asterisk skips unloading modules. But while it is processing the atexit handlers, there is a window of time where some of the core message types have been cleaned up, but the message bus is still running. Specifically, it's still running module subscriptions that might be using the core message types. If a message is received by that subscription in that window, it will attempt to use a message type that has been cleaned up. To solve this problem, this patch introduces ast_register_cleanup(). This function operates identically to ast_register_atexit(), except that cleanup calls are not invoked on an immediate shutdown. All of the core message type and topic cleanup was moved from atexit handlers to cleanup handlers. This ensures that core type and topic cleanup only happens if the modules that used them are first unloaded. This patch also changes the ast_assert() when accessing a cleaned up or uninitialized message type to an error log message. Message type functions are actually NULL safe across the board, so the assert was a bit heavy handed. Especially for anyone with DO_CRASH enabled. Review: https://reviewboard.asterisk.org/r/2562/ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@390122 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-05-30 17:05:53 +00:00
void stasis_log_bad_type_access(const char *name)
{
#ifdef AST_DEVMODE
if (!stasis_message_type_declined(name)) {
ast_log(LOG_ERROR, "Use of %s() before init/after destruction\n", name);
}
#endif
Avoid unnecessary cleanups during immediate shutdown This patch addresses issues during immediate shutdowns, where modules are not unloaded, but Asterisk atexit handlers are run. In the typical case, this usually isn't a big deal. But the introduction of the Stasis message bus makes it much more likely for asynchronous activity to be happening off in some thread during shutdown. During an immediate shutdown, Asterisk skips unloading modules. But while it is processing the atexit handlers, there is a window of time where some of the core message types have been cleaned up, but the message bus is still running. Specifically, it's still running module subscriptions that might be using the core message types. If a message is received by that subscription in that window, it will attempt to use a message type that has been cleaned up. To solve this problem, this patch introduces ast_register_cleanup(). This function operates identically to ast_register_atexit(), except that cleanup calls are not invoked on an immediate shutdown. All of the core message type and topic cleanup was moved from atexit handlers to cleanup handlers. This ensures that core type and topic cleanup only happens if the modules that used them are first unloaded. This patch also changes the ast_assert() when accessing a cleaned up or uninitialized message type to an error log message. Message type functions are actually NULL safe across the board, so the assert was a bit heavy handed. Especially for anyone with DO_CRASH enabled. Review: https://reviewboard.asterisk.org/r/2562/ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@390122 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-05-30 17:05:53 +00:00
}
/*! \brief A multi object blob data structure to carry user event stasis messages */
struct ast_multi_object_blob {
struct ast_json *blob; /*< A blob of JSON data */
AST_VECTOR(, void *) snapshots[STASIS_UMOS_MAX]; /*< Vector of snapshots for each type */
};
/*!
* \internal
* \brief Destructor for \ref ast_multi_object_blob objects
*/
static void multi_object_blob_dtor(void *obj)
{
struct ast_multi_object_blob *multi = obj;
int type;
int i;
for (type = 0; type < STASIS_UMOS_MAX; ++type) {
for (i = 0; i < AST_VECTOR_SIZE(&multi->snapshots[type]); ++i) {
ao2_cleanup(AST_VECTOR_GET(&multi->snapshots[type], i));
}
AST_VECTOR_FREE(&multi->snapshots[type]);
}
ast_json_unref(multi->blob);
}
/*! \brief Create a stasis user event multi object blob */
struct ast_multi_object_blob *ast_multi_object_blob_create(struct ast_json *blob)
{
int type;
struct ast_multi_object_blob *multi;
ast_assert(blob != NULL);
multi = ao2_alloc(sizeof(*multi), multi_object_blob_dtor);
if (!multi) {
return NULL;
}
for (type = 0; type < STASIS_UMOS_MAX; ++type) {
if (AST_VECTOR_INIT(&multi->snapshots[type], 0)) {
ao2_ref(multi, -1);
return NULL;
}
}
multi->blob = ast_json_ref(blob);
return multi;
}
/*! \brief Add an object (snapshot) to the blob */
void ast_multi_object_blob_add(struct ast_multi_object_blob *multi,
enum stasis_user_multi_object_snapshot_type type, void *object)
{
if (!multi || !object || AST_VECTOR_APPEND(&multi->snapshots[type], object)) {
ao2_cleanup(object);
}
}
/*! \brief Publish single channel user event (for app_userevent compatibility) */
void ast_multi_object_blob_single_channel_publish(struct ast_channel *chan,
struct stasis_message_type *type, struct ast_json *blob)
{
struct stasis_message *message;
struct ast_channel_snapshot *channel_snapshot;
struct ast_multi_object_blob *multi;
if (!type) {
return;
}
multi = ast_multi_object_blob_create(blob);
if (!multi) {
return;
}
channel_snapshot = ast_channel_snapshot_create(chan);
if (!channel_snapshot) {
ao2_ref(multi, -1);
return;
}
/* this call steals the channel_snapshot reference */
ast_multi_object_blob_add(multi, STASIS_UMOS_CHANNEL, channel_snapshot);
message = stasis_message_create(type, multi);
ao2_ref(multi, -1);
if (message) {
/* app_userevent still publishes to channel */
stasis_publish(ast_channel_topic(chan), message);
ao2_ref(message, -1);
}
}
/*! \internal \brief convert multi object blob to ari json */
static struct ast_json *multi_user_event_to_json(
struct stasis_message *message,
const struct stasis_message_sanitizer *sanitize)
{
struct ast_json *out;
struct ast_multi_object_blob *multi = stasis_message_data(message);
struct ast_json *blob = multi->blob;
const struct timeval *tv = stasis_message_timestamp(message);
enum stasis_user_multi_object_snapshot_type type;
int i;
out = ast_json_object_create();
if (!out) {
return NULL;
}
ast_json_object_set(out, "type", ast_json_string_create("ChannelUserevent"));
ast_json_object_set(out, "timestamp", ast_json_timeval(*tv, NULL));
ast_json_object_set(out, "eventname", ast_json_ref(ast_json_object_get(blob, "eventname")));
ast_json_object_set(out, "userevent", ast_json_ref(blob));
for (type = 0; type < STASIS_UMOS_MAX; ++type) {
for (i = 0; i < AST_VECTOR_SIZE(&multi->snapshots[type]); ++i) {
struct ast_json *json_object = NULL;
char *name = NULL;
void *snapshot = AST_VECTOR_GET(&multi->snapshots[type], i);
switch (type) {
case STASIS_UMOS_CHANNEL:
json_object = ast_channel_snapshot_to_json(snapshot, sanitize);
name = "channel";
break;
case STASIS_UMOS_BRIDGE:
json_object = ast_bridge_snapshot_to_json(snapshot, sanitize);
name = "bridge";
break;
case STASIS_UMOS_ENDPOINT:
json_object = ast_endpoint_snapshot_to_json(snapshot, sanitize);
name = "endpoint";
break;
}
if (json_object) {
ast_json_object_set(out, name, json_object);
}
}
}
return out;
}
/*! \internal \brief convert multi object blob to ami string */
static struct ast_str *multi_object_blob_to_ami(void *obj)
{
struct ast_str *ami_str=ast_str_create(1024);
struct ast_str *ami_snapshot;
const struct ast_multi_object_blob *multi = obj;
enum stasis_user_multi_object_snapshot_type type;
int i;
if (!ami_str) {
return NULL;
}
if (!multi) {
ast_free(ami_str);
return NULL;
}
for (type = 0; type < STASIS_UMOS_MAX; ++type) {
for (i = 0; i < AST_VECTOR_SIZE(&multi->snapshots[type]); ++i) {
char *name = NULL;
void *snapshot = AST_VECTOR_GET(&multi->snapshots[type], i);
ami_snapshot = NULL;
if (i > 0) {
ast_asprintf(&name, "%d", i + 1);
}
switch (type) {
case STASIS_UMOS_CHANNEL:
ami_snapshot = ast_manager_build_channel_state_string_prefix(snapshot, name ?: "");
break;
case STASIS_UMOS_BRIDGE:
ami_snapshot = ast_manager_build_bridge_state_string_prefix(snapshot, name ?: "");
break;
case STASIS_UMOS_ENDPOINT:
/* currently not sending endpoint snapshots to AMI */
break;
}
if (ami_snapshot) {
ast_str_append(&ami_str, 0, "%s", ast_str_buffer(ami_snapshot));
ast_free(ami_snapshot);
}
ast_free(name);
}
}
return ami_str;
}
/*! \internal \brief Callback to pass only user defined parameters from blob */
static int userevent_exclusion_cb(const char *key)
{
if (!strcmp("eventname", key)) {
return 1;
}
return 0;
}
static struct ast_manager_event_blob *multi_user_event_to_ami(
struct stasis_message *message)
{
RAII_VAR(struct ast_str *, object_string, NULL, ast_free);
RAII_VAR(struct ast_str *, body, NULL, ast_free);
struct ast_multi_object_blob *multi = stasis_message_data(message);
const char *eventname;
eventname = ast_json_string_get(ast_json_object_get(multi->blob, "eventname"));
body = ast_manager_str_from_json_object(multi->blob, userevent_exclusion_cb);
object_string = multi_object_blob_to_ami(multi);
if (!object_string || !body) {
return NULL;
}
return ast_manager_event_blob_create(EVENT_FLAG_USER, "UserEvent",
"%s"
"UserEvent: %s\r\n"
"%s",
ast_str_buffer(object_string),
eventname,
ast_str_buffer(body));
}
/*! \brief A structure to hold global configuration-related options */
struct stasis_declined_config {
/*! The list of message types to decline */
struct ao2_container *declined;
};
main/stasis: Allow subscriptions to use a threadpool for message delivery Prior to this patch, all Stasis subscriptions would receive a dedicated thread for servicing published messages. In contrast, prior to r400178 (see review https://reviewboard.asterisk.org/r/2881/), the subscriptions shared a thread pool. It was discovered during some initial work on Stasis that, for a low subscription count with high message throughput, the threadpool was not as performant as simply having a dedicated thread per subscriber. For situations where a subscriber receives a substantial number of messages and is always present, the model of having a dedicated thread per subscriber makes sense. While we still have plenty of subscriptions that would follow this model, e.g., AMI, CDRs, CEL, etc., there are plenty that also fall into the following two categories: * Large number of subscriptions, specifically those tied to endpoints/peers. * Low number of messages. Some subscriptions exist specifically to coordinate a single message - the subscription is created, a message is published, the delivery is synchronized, and the subscription is destroyed. In both of the latter two cases, creating a dedicated thread is wasteful (and in the case of a large number of peers/endpoints, harmful). In those cases, having shared delivery threads is far more performant. This patch adds the ability of a subscriber to Stasis to choose whether or not their messages are dispatched on a dedicated thread or on a threadpool. The threadpool is configurable through stasis.conf. Review: https://reviewboard.asterisk.org/r/4193 ASTERISK-24533 #close Reported by: xrobau Tested by: xrobau ........ Merged revisions 428681 from http://svn.asterisk.org/svn/asterisk/branches/12 ........ Merged revisions 428687 from http://svn.asterisk.org/svn/asterisk/branches/13 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@428688 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-12-01 17:59:21 +00:00
/*! \brief Threadpool configuration options */
struct stasis_threadpool_conf {
/*! Initial size of the thread pool */
int initial_size;
/*! Time, in seconds, before we expire a thread */
int idle_timeout_sec;
/*! Maximum number of thread to allow */
int max_size;
};
struct stasis_config {
main/stasis: Allow subscriptions to use a threadpool for message delivery Prior to this patch, all Stasis subscriptions would receive a dedicated thread for servicing published messages. In contrast, prior to r400178 (see review https://reviewboard.asterisk.org/r/2881/), the subscriptions shared a thread pool. It was discovered during some initial work on Stasis that, for a low subscription count with high message throughput, the threadpool was not as performant as simply having a dedicated thread per subscriber. For situations where a subscriber receives a substantial number of messages and is always present, the model of having a dedicated thread per subscriber makes sense. While we still have plenty of subscriptions that would follow this model, e.g., AMI, CDRs, CEL, etc., there are plenty that also fall into the following two categories: * Large number of subscriptions, specifically those tied to endpoints/peers. * Low number of messages. Some subscriptions exist specifically to coordinate a single message - the subscription is created, a message is published, the delivery is synchronized, and the subscription is destroyed. In both of the latter two cases, creating a dedicated thread is wasteful (and in the case of a large number of peers/endpoints, harmful). In those cases, having shared delivery threads is far more performant. This patch adds the ability of a subscriber to Stasis to choose whether or not their messages are dispatched on a dedicated thread or on a threadpool. The threadpool is configurable through stasis.conf. Review: https://reviewboard.asterisk.org/r/4193 ASTERISK-24533 #close Reported by: xrobau Tested by: xrobau ........ Merged revisions 428681 from http://svn.asterisk.org/svn/asterisk/branches/12 ........ Merged revisions 428687 from http://svn.asterisk.org/svn/asterisk/branches/13 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@428688 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-12-01 17:59:21 +00:00
/*! Thread pool configuration options */
struct stasis_threadpool_conf *threadpool_options;
/*! Declined message types */
struct stasis_declined_config *declined_message_types;
};
main/stasis: Allow subscriptions to use a threadpool for message delivery Prior to this patch, all Stasis subscriptions would receive a dedicated thread for servicing published messages. In contrast, prior to r400178 (see review https://reviewboard.asterisk.org/r/2881/), the subscriptions shared a thread pool. It was discovered during some initial work on Stasis that, for a low subscription count with high message throughput, the threadpool was not as performant as simply having a dedicated thread per subscriber. For situations where a subscriber receives a substantial number of messages and is always present, the model of having a dedicated thread per subscriber makes sense. While we still have plenty of subscriptions that would follow this model, e.g., AMI, CDRs, CEL, etc., there are plenty that also fall into the following two categories: * Large number of subscriptions, specifically those tied to endpoints/peers. * Low number of messages. Some subscriptions exist specifically to coordinate a single message - the subscription is created, a message is published, the delivery is synchronized, and the subscription is destroyed. In both of the latter two cases, creating a dedicated thread is wasteful (and in the case of a large number of peers/endpoints, harmful). In those cases, having shared delivery threads is far more performant. This patch adds the ability of a subscriber to Stasis to choose whether or not their messages are dispatched on a dedicated thread or on a threadpool. The threadpool is configurable through stasis.conf. Review: https://reviewboard.asterisk.org/r/4193 ASTERISK-24533 #close Reported by: xrobau Tested by: xrobau ........ Merged revisions 428681 from http://svn.asterisk.org/svn/asterisk/branches/12 ........ Merged revisions 428687 from http://svn.asterisk.org/svn/asterisk/branches/13 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@428688 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-12-01 17:59:21 +00:00
static struct aco_type threadpool_option = {
.type = ACO_GLOBAL,
.name = "threadpool",
.item_offset = offsetof(struct stasis_config, threadpool_options),
.category = "threadpool",
.category_match = ACO_WHITELIST_EXACT,
main/stasis: Allow subscriptions to use a threadpool for message delivery Prior to this patch, all Stasis subscriptions would receive a dedicated thread for servicing published messages. In contrast, prior to r400178 (see review https://reviewboard.asterisk.org/r/2881/), the subscriptions shared a thread pool. It was discovered during some initial work on Stasis that, for a low subscription count with high message throughput, the threadpool was not as performant as simply having a dedicated thread per subscriber. For situations where a subscriber receives a substantial number of messages and is always present, the model of having a dedicated thread per subscriber makes sense. While we still have plenty of subscriptions that would follow this model, e.g., AMI, CDRs, CEL, etc., there are plenty that also fall into the following two categories: * Large number of subscriptions, specifically those tied to endpoints/peers. * Low number of messages. Some subscriptions exist specifically to coordinate a single message - the subscription is created, a message is published, the delivery is synchronized, and the subscription is destroyed. In both of the latter two cases, creating a dedicated thread is wasteful (and in the case of a large number of peers/endpoints, harmful). In those cases, having shared delivery threads is far more performant. This patch adds the ability of a subscriber to Stasis to choose whether or not their messages are dispatched on a dedicated thread or on a threadpool. The threadpool is configurable through stasis.conf. Review: https://reviewboard.asterisk.org/r/4193 ASTERISK-24533 #close Reported by: xrobau Tested by: xrobau ........ Merged revisions 428681 from http://svn.asterisk.org/svn/asterisk/branches/12 ........ Merged revisions 428687 from http://svn.asterisk.org/svn/asterisk/branches/13 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@428688 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-12-01 17:59:21 +00:00
};
static struct aco_type *threadpool_options[] = ACO_TYPES(&threadpool_option);
/*! \brief An aco_type structure to link the "declined_message_types" category to the stasis_declined_config type */
static struct aco_type declined_option = {
.type = ACO_GLOBAL,
.name = "declined_message_types",
.item_offset = offsetof(struct stasis_config, declined_message_types),
.category_match = ACO_WHITELIST_EXACT,
.category = "declined_message_types",
};
struct aco_type *declined_options[] = ACO_TYPES(&declined_option);
struct aco_file stasis_conf = {
.filename = "stasis.conf",
main/stasis: Allow subscriptions to use a threadpool for message delivery Prior to this patch, all Stasis subscriptions would receive a dedicated thread for servicing published messages. In contrast, prior to r400178 (see review https://reviewboard.asterisk.org/r/2881/), the subscriptions shared a thread pool. It was discovered during some initial work on Stasis that, for a low subscription count with high message throughput, the threadpool was not as performant as simply having a dedicated thread per subscriber. For situations where a subscriber receives a substantial number of messages and is always present, the model of having a dedicated thread per subscriber makes sense. While we still have plenty of subscriptions that would follow this model, e.g., AMI, CDRs, CEL, etc., there are plenty that also fall into the following two categories: * Large number of subscriptions, specifically those tied to endpoints/peers. * Low number of messages. Some subscriptions exist specifically to coordinate a single message - the subscription is created, a message is published, the delivery is synchronized, and the subscription is destroyed. In both of the latter two cases, creating a dedicated thread is wasteful (and in the case of a large number of peers/endpoints, harmful). In those cases, having shared delivery threads is far more performant. This patch adds the ability of a subscriber to Stasis to choose whether or not their messages are dispatched on a dedicated thread or on a threadpool. The threadpool is configurable through stasis.conf. Review: https://reviewboard.asterisk.org/r/4193 ASTERISK-24533 #close Reported by: xrobau Tested by: xrobau ........ Merged revisions 428681 from http://svn.asterisk.org/svn/asterisk/branches/12 ........ Merged revisions 428687 from http://svn.asterisk.org/svn/asterisk/branches/13 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@428688 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-12-01 17:59:21 +00:00
.types = ACO_TYPES(&declined_option, &threadpool_option),
};
/*! \brief A global object container that will contain the stasis_config that gets swapped out on reloads */
static AO2_GLOBAL_OBJ_STATIC(globals);
static void *stasis_config_alloc(void);
/*! \brief Register information about the configs being processed by this module */
CONFIG_INFO_CORE("stasis", cfg_info, globals, stasis_config_alloc,
.files = ACO_FILES(&stasis_conf),
);
static void stasis_declined_config_destructor(void *obj)
{
struct stasis_declined_config *declined = obj;
main/stasis: Allow subscriptions to use a threadpool for message delivery Prior to this patch, all Stasis subscriptions would receive a dedicated thread for servicing published messages. In contrast, prior to r400178 (see review https://reviewboard.asterisk.org/r/2881/), the subscriptions shared a thread pool. It was discovered during some initial work on Stasis that, for a low subscription count with high message throughput, the threadpool was not as performant as simply having a dedicated thread per subscriber. For situations where a subscriber receives a substantial number of messages and is always present, the model of having a dedicated thread per subscriber makes sense. While we still have plenty of subscriptions that would follow this model, e.g., AMI, CDRs, CEL, etc., there are plenty that also fall into the following two categories: * Large number of subscriptions, specifically those tied to endpoints/peers. * Low number of messages. Some subscriptions exist specifically to coordinate a single message - the subscription is created, a message is published, the delivery is synchronized, and the subscription is destroyed. In both of the latter two cases, creating a dedicated thread is wasteful (and in the case of a large number of peers/endpoints, harmful). In those cases, having shared delivery threads is far more performant. This patch adds the ability of a subscriber to Stasis to choose whether or not their messages are dispatched on a dedicated thread or on a threadpool. The threadpool is configurable through stasis.conf. Review: https://reviewboard.asterisk.org/r/4193 ASTERISK-24533 #close Reported by: xrobau Tested by: xrobau ........ Merged revisions 428681 from http://svn.asterisk.org/svn/asterisk/branches/12 ........ Merged revisions 428687 from http://svn.asterisk.org/svn/asterisk/branches/13 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@428688 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-12-01 17:59:21 +00:00
ao2_cleanup(declined->declined);
}
static void stasis_config_destructor(void *obj)
{
struct stasis_config *cfg = obj;
main/stasis: Allow subscriptions to use a threadpool for message delivery Prior to this patch, all Stasis subscriptions would receive a dedicated thread for servicing published messages. In contrast, prior to r400178 (see review https://reviewboard.asterisk.org/r/2881/), the subscriptions shared a thread pool. It was discovered during some initial work on Stasis that, for a low subscription count with high message throughput, the threadpool was not as performant as simply having a dedicated thread per subscriber. For situations where a subscriber receives a substantial number of messages and is always present, the model of having a dedicated thread per subscriber makes sense. While we still have plenty of subscriptions that would follow this model, e.g., AMI, CDRs, CEL, etc., there are plenty that also fall into the following two categories: * Large number of subscriptions, specifically those tied to endpoints/peers. * Low number of messages. Some subscriptions exist specifically to coordinate a single message - the subscription is created, a message is published, the delivery is synchronized, and the subscription is destroyed. In both of the latter two cases, creating a dedicated thread is wasteful (and in the case of a large number of peers/endpoints, harmful). In those cases, having shared delivery threads is far more performant. This patch adds the ability of a subscriber to Stasis to choose whether or not their messages are dispatched on a dedicated thread or on a threadpool. The threadpool is configurable through stasis.conf. Review: https://reviewboard.asterisk.org/r/4193 ASTERISK-24533 #close Reported by: xrobau Tested by: xrobau ........ Merged revisions 428681 from http://svn.asterisk.org/svn/asterisk/branches/12 ........ Merged revisions 428687 from http://svn.asterisk.org/svn/asterisk/branches/13 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@428688 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-12-01 17:59:21 +00:00
ao2_cleanup(cfg->declined_message_types);
main/stasis: Allow subscriptions to use a threadpool for message delivery Prior to this patch, all Stasis subscriptions would receive a dedicated thread for servicing published messages. In contrast, prior to r400178 (see review https://reviewboard.asterisk.org/r/2881/), the subscriptions shared a thread pool. It was discovered during some initial work on Stasis that, for a low subscription count with high message throughput, the threadpool was not as performant as simply having a dedicated thread per subscriber. For situations where a subscriber receives a substantial number of messages and is always present, the model of having a dedicated thread per subscriber makes sense. While we still have plenty of subscriptions that would follow this model, e.g., AMI, CDRs, CEL, etc., there are plenty that also fall into the following two categories: * Large number of subscriptions, specifically those tied to endpoints/peers. * Low number of messages. Some subscriptions exist specifically to coordinate a single message - the subscription is created, a message is published, the delivery is synchronized, and the subscription is destroyed. In both of the latter two cases, creating a dedicated thread is wasteful (and in the case of a large number of peers/endpoints, harmful). In those cases, having shared delivery threads is far more performant. This patch adds the ability of a subscriber to Stasis to choose whether or not their messages are dispatched on a dedicated thread or on a threadpool. The threadpool is configurable through stasis.conf. Review: https://reviewboard.asterisk.org/r/4193 ASTERISK-24533 #close Reported by: xrobau Tested by: xrobau ........ Merged revisions 428681 from http://svn.asterisk.org/svn/asterisk/branches/12 ........ Merged revisions 428687 from http://svn.asterisk.org/svn/asterisk/branches/13 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@428688 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-12-01 17:59:21 +00:00
ast_free(cfg->threadpool_options);
}
static void *stasis_config_alloc(void)
{
struct stasis_config *cfg;
if (!(cfg = ao2_alloc(sizeof(*cfg), stasis_config_destructor))) {
return NULL;
}
main/stasis: Allow subscriptions to use a threadpool for message delivery Prior to this patch, all Stasis subscriptions would receive a dedicated thread for servicing published messages. In contrast, prior to r400178 (see review https://reviewboard.asterisk.org/r/2881/), the subscriptions shared a thread pool. It was discovered during some initial work on Stasis that, for a low subscription count with high message throughput, the threadpool was not as performant as simply having a dedicated thread per subscriber. For situations where a subscriber receives a substantial number of messages and is always present, the model of having a dedicated thread per subscriber makes sense. While we still have plenty of subscriptions that would follow this model, e.g., AMI, CDRs, CEL, etc., there are plenty that also fall into the following two categories: * Large number of subscriptions, specifically those tied to endpoints/peers. * Low number of messages. Some subscriptions exist specifically to coordinate a single message - the subscription is created, a message is published, the delivery is synchronized, and the subscription is destroyed. In both of the latter two cases, creating a dedicated thread is wasteful (and in the case of a large number of peers/endpoints, harmful). In those cases, having shared delivery threads is far more performant. This patch adds the ability of a subscriber to Stasis to choose whether or not their messages are dispatched on a dedicated thread or on a threadpool. The threadpool is configurable through stasis.conf. Review: https://reviewboard.asterisk.org/r/4193 ASTERISK-24533 #close Reported by: xrobau Tested by: xrobau ........ Merged revisions 428681 from http://svn.asterisk.org/svn/asterisk/branches/12 ........ Merged revisions 428687 from http://svn.asterisk.org/svn/asterisk/branches/13 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@428688 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-12-01 17:59:21 +00:00
cfg->threadpool_options = ast_calloc(1, sizeof(*cfg->threadpool_options));
if (!cfg->threadpool_options) {
ao2_ref(cfg, -1);
return NULL;
}
cfg->declined_message_types = ao2_alloc(sizeof(*cfg->declined_message_types),
stasis_declined_config_destructor);
if (!cfg->declined_message_types) {
main/stasis: Allow subscriptions to use a threadpool for message delivery Prior to this patch, all Stasis subscriptions would receive a dedicated thread for servicing published messages. In contrast, prior to r400178 (see review https://reviewboard.asterisk.org/r/2881/), the subscriptions shared a thread pool. It was discovered during some initial work on Stasis that, for a low subscription count with high message throughput, the threadpool was not as performant as simply having a dedicated thread per subscriber. For situations where a subscriber receives a substantial number of messages and is always present, the model of having a dedicated thread per subscriber makes sense. While we still have plenty of subscriptions that would follow this model, e.g., AMI, CDRs, CEL, etc., there are plenty that also fall into the following two categories: * Large number of subscriptions, specifically those tied to endpoints/peers. * Low number of messages. Some subscriptions exist specifically to coordinate a single message - the subscription is created, a message is published, the delivery is synchronized, and the subscription is destroyed. In both of the latter two cases, creating a dedicated thread is wasteful (and in the case of a large number of peers/endpoints, harmful). In those cases, having shared delivery threads is far more performant. This patch adds the ability of a subscriber to Stasis to choose whether or not their messages are dispatched on a dedicated thread or on a threadpool. The threadpool is configurable through stasis.conf. Review: https://reviewboard.asterisk.org/r/4193 ASTERISK-24533 #close Reported by: xrobau Tested by: xrobau ........ Merged revisions 428681 from http://svn.asterisk.org/svn/asterisk/branches/12 ........ Merged revisions 428687 from http://svn.asterisk.org/svn/asterisk/branches/13 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@428688 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-12-01 17:59:21 +00:00
ao2_ref(cfg, -1);
return NULL;
}
cfg->declined_message_types->declined = ast_str_container_alloc(13);
if (!cfg->declined_message_types->declined) {
main/stasis: Allow subscriptions to use a threadpool for message delivery Prior to this patch, all Stasis subscriptions would receive a dedicated thread for servicing published messages. In contrast, prior to r400178 (see review https://reviewboard.asterisk.org/r/2881/), the subscriptions shared a thread pool. It was discovered during some initial work on Stasis that, for a low subscription count with high message throughput, the threadpool was not as performant as simply having a dedicated thread per subscriber. For situations where a subscriber receives a substantial number of messages and is always present, the model of having a dedicated thread per subscriber makes sense. While we still have plenty of subscriptions that would follow this model, e.g., AMI, CDRs, CEL, etc., there are plenty that also fall into the following two categories: * Large number of subscriptions, specifically those tied to endpoints/peers. * Low number of messages. Some subscriptions exist specifically to coordinate a single message - the subscription is created, a message is published, the delivery is synchronized, and the subscription is destroyed. In both of the latter two cases, creating a dedicated thread is wasteful (and in the case of a large number of peers/endpoints, harmful). In those cases, having shared delivery threads is far more performant. This patch adds the ability of a subscriber to Stasis to choose whether or not their messages are dispatched on a dedicated thread or on a threadpool. The threadpool is configurable through stasis.conf. Review: https://reviewboard.asterisk.org/r/4193 ASTERISK-24533 #close Reported by: xrobau Tested by: xrobau ........ Merged revisions 428681 from http://svn.asterisk.org/svn/asterisk/branches/12 ........ Merged revisions 428687 from http://svn.asterisk.org/svn/asterisk/branches/13 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@428688 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-12-01 17:59:21 +00:00
ao2_ref(cfg, -1);
return NULL;
}
return cfg;
}
int stasis_message_type_declined(const char *name)
{
struct stasis_config *cfg = ao2_global_obj_ref(globals);
char *name_in_declined;
int res;
if (!cfg || !cfg->declined_message_types) {
ao2_cleanup(cfg);
return 0;
}
name_in_declined = ao2_find(cfg->declined_message_types->declined, name, OBJ_SEARCH_KEY);
res = name_in_declined ? 1 : 0;
ao2_cleanup(name_in_declined);
ao2_ref(cfg, -1);
if (res) {
ast_debug(4, "Declining to allocate Stasis message type '%s' due to configuration\n", name);
}
return res;
}
static int declined_handler(const struct aco_option *opt, struct ast_variable *var, void *obj)
{
struct stasis_declined_config *declined = obj;
if (ast_strlen_zero(var->value)) {
return 0;
}
if (ast_str_container_add(declined->declined, var->value)) {
return -1;
}
return 0;
}
/*!
* @{ \brief Define multi user event message type(s).
*/
STASIS_MESSAGE_TYPE_DEFN(ast_multi_user_event_type,
.to_json = multi_user_event_to_json,
.to_ami = multi_user_event_to_ami,
);
/*! @} */
/*!
* \internal
* \brief CLI command implementation for 'stasis show topics'
*/
static char *stasis_show_topics(struct ast_cli_entry *e, int cmd, struct ast_cli_args *a)
{
struct ao2_iterator iter;
struct topic_proxy *topic;
struct ao2_container *tmp_container;
int count = 0;
#define FMT_HEADERS "%-64s %-64s\n"
#define FMT_FIELDS "%-64s %-64s\n"
switch (cmd) {
case CLI_INIT:
e->command = "stasis show topics";
e->usage =
"Usage: stasis show topics\n"
" Shows a list of topics\n";
return NULL;
case CLI_GENERATE:
return NULL;
}
if (a->argc != e->args) {
return CLI_SHOWUSAGE;
}
ast_cli(a->fd, "\n" FMT_HEADERS, "Name", "Detail");
tmp_container = ao2_container_alloc_list(AO2_ALLOC_OPT_LOCK_NOLOCK, 0,
topic_proxy_sort_fn, NULL);
if (!tmp_container || ao2_container_dup(tmp_container, topic_all, OBJ_SEARCH_OBJECT)) {
ao2_cleanup(tmp_container);
return NULL;
}
/* getting all topic in order */
iter = ao2_iterator_init(tmp_container, AO2_ITERATOR_UNLINK);
while ((topic = ao2_iterator_next(&iter))) {
ast_cli(a->fd, FMT_FIELDS, topic->name, topic->detail);
ao2_ref(topic, -1);
++count;
}
ao2_iterator_destroy(&iter);
ao2_cleanup(tmp_container);
ast_cli(a->fd, "\n%d Total topics\n\n", count);
#undef FMT_HEADERS
#undef FMT_FIELDS
return CLI_SUCCESS;
}
/*!
* \internal
* \brief CLI tab completion for topic names
*/
static char *topic_complete_name(const char *word)
{
struct topic_proxy *topic;
struct ao2_iterator it;
int wordlen = strlen(word);
int ret;
it = ao2_iterator_init(topic_all, 0);
while ((topic = ao2_iterator_next(&it))) {
if (!strncasecmp(word, topic->name, wordlen)) {
ret = ast_cli_completion_add(ast_strdup(topic->name));
if (ret) {
ao2_ref(topic, -1);
break;
}
}
ao2_ref(topic, -1);
}
ao2_iterator_destroy(&it);
return NULL;
}
/*!
* \internal
* \brief CLI command implementation for 'stasis show topic'
*/
static char *stasis_show_topic(struct ast_cli_entry *e, int cmd, struct ast_cli_args *a)
{
struct stasis_topic *topic;
char print_time[32];
int i;
switch (cmd) {
case CLI_INIT:
e->command = "stasis show topic";
e->usage =
"Usage: stasis show topic <name>\n"
" Show stasis topic detail info.\n";
return NULL;
case CLI_GENERATE:
if (a->pos == 3) {
return topic_complete_name(a->word);
} else {
return NULL;
}
}
if (a->argc != 4) {
return CLI_SHOWUSAGE;
}
topic = stasis_topic_get(a->argv[3]);
if (!topic) {
ast_cli(a->fd, "Specified topic '%s' does not exist\n", a->argv[3]);
return CLI_FAILURE;
}
ast_cli(a->fd, "Name: %s\n", topic->name);
ast_cli(a->fd, "Detail: %s\n", topic->detail);
ast_cli(a->fd, "Subscribers count: %zu\n", AST_VECTOR_SIZE(&topic->subscribers));
ast_cli(a->fd, "Forwarding topic count: %zu\n", AST_VECTOR_SIZE(&topic->upstream_topics));
ast_format_duration_hh_mm_ss(ast_tvnow().tv_sec - topic->creationtime->tv_sec, print_time, sizeof(print_time));
ast_cli(a->fd, "Duration time: %s\n", print_time);
ao2_lock(topic);
ast_cli(a->fd, "\nSubscribers:\n");
for (i = 0; i < AST_VECTOR_SIZE(&topic->subscribers); i++) {
struct stasis_subscription *subscription_tmp = AST_VECTOR_GET(&topic->subscribers, i);
ast_cli(a->fd, " UniqueID: %s, Topic: %s, Detail: %s\n",
subscription_tmp->uniqueid, subscription_tmp->topic->name, subscription_tmp->topic->detail);
}
ast_cli(a->fd, "\nForwarded topics:\n");
for (i = 0; i < AST_VECTOR_SIZE(&topic->upstream_topics); i++) {
struct stasis_topic *topic_tmp = AST_VECTOR_GET(&topic->upstream_topics, i);
ast_cli(a->fd, " Topic: %s, Detail: %s\n", topic_tmp->name, topic_tmp->detail);
}
ao2_unlock(topic);
ao2_ref(topic, -1);
return CLI_SUCCESS;
}
static struct ast_cli_entry cli_stasis[] = {
AST_CLI_DEFINE(stasis_show_topics, "Show all topics"),
AST_CLI_DEFINE(stasis_show_topic, "Show topic"),
};
#ifdef AST_DEVMODE
AO2_STRING_FIELD_SORT_FN(stasis_subscription_statistics, uniqueid);
/*!
* \internal
* \brief CLI command implementation for 'stasis statistics show subscriptions'
*/
static char *statistics_show_subscriptions(struct ast_cli_entry *e, int cmd, struct ast_cli_args *a)
{
struct ao2_container *sorted_subscriptions;
struct ao2_container *subscription_stats;
struct ao2_iterator iter;
struct stasis_subscription_statistics *statistics;
int count = 0;
int dropped = 0;
int passed = 0;
#define FMT_HEADERS "%-64s %10s %10s %16s %16s\n"
#define FMT_FIELDS "%-64s %10d %10d %16ld %16ld\n"
#define FMT_FIELDS2 "%-64s %10d %10d\n"
switch (cmd) {
case CLI_INIT:
e->command = "stasis statistics show subscriptions";
e->usage =
"Usage: stasis statistics show subscriptions\n"
" Shows a list of subscriptions and their general statistics\n";
return NULL;
case CLI_GENERATE:
return NULL;
}
if (a->argc != e->args) {
return CLI_SHOWUSAGE;
}
subscription_stats = ao2_global_obj_ref(subscription_statistics);
if (!subscription_stats) {
ast_cli(a->fd, "Could not fetch subscription_statistics container\n");
return CLI_FAILURE;
}
sorted_subscriptions = ao2_container_alloc_rbtree(AO2_ALLOC_OPT_LOCK_NOLOCK, 0,
stasis_subscription_statistics_sort_fn, NULL);
if (!sorted_subscriptions) {
ao2_ref(subscription_stats, -1);
ast_cli(a->fd, "Could not create container for sorting subscription statistics\n");
return CLI_SUCCESS;
}
if (ao2_container_dup(sorted_subscriptions, subscription_stats, 0)) {
ao2_ref(sorted_subscriptions, -1);
ao2_ref(subscription_stats, -1);
ast_cli(a->fd, "Could not sort subscription statistics\n");
return CLI_SUCCESS;
}
ao2_ref(subscription_stats, -1);
ast_cli(a->fd, "\n" FMT_HEADERS, "Subscription", "Dropped", "Passed", "Lowest Invoke", "Highest Invoke");
iter = ao2_iterator_init(sorted_subscriptions, 0);
while ((statistics = ao2_iterator_next(&iter))) {
ast_cli(a->fd, FMT_FIELDS, statistics->uniqueid, statistics->messages_dropped, statistics->messages_passed,
statistics->lowest_time_invoked, statistics->highest_time_invoked);
dropped += statistics->messages_dropped;
passed += statistics->messages_passed;
ao2_ref(statistics, -1);
++count;
}
ao2_iterator_destroy(&iter);
ao2_ref(sorted_subscriptions, -1);
ast_cli(a->fd, FMT_FIELDS2, "Total", dropped, passed);
ast_cli(a->fd, "\n%d subscriptions\n\n", count);
#undef FMT_HEADERS
#undef FMT_FIELDS
#undef FMT_FIELDS2
return CLI_SUCCESS;
}
/*!
* \internal
* \brief CLI tab completion for subscription statistics names
*/
static char *subscription_statistics_complete_name(const char *word, int state)
{
struct stasis_subscription_statistics *statistics;
struct ao2_container *subscription_stats;
struct ao2_iterator it_statistics;
int wordlen = strlen(word);
int which = 0;
char *result = NULL;
subscription_stats = ao2_global_obj_ref(subscription_statistics);
if (!subscription_stats) {
return result;
}
it_statistics = ao2_iterator_init(subscription_stats, 0);
while ((statistics = ao2_iterator_next(&it_statistics))) {
if (!strncasecmp(word, statistics->uniqueid, wordlen)
&& ++which > state) {
result = ast_strdup(statistics->uniqueid);
}
ao2_ref(statistics, -1);
if (result) {
break;
}
}
ao2_iterator_destroy(&it_statistics);
ao2_ref(subscription_stats, -1);
return result;
}
/*!
* \internal
* \brief CLI command implementation for 'stasis statistics show subscription'
*/
static char *statistics_show_subscription(struct ast_cli_entry *e, int cmd, struct ast_cli_args *a)
{
struct stasis_subscription_statistics *statistics;
struct ao2_container *subscription_stats;
struct ao2_iterator i;
char *name;
switch (cmd) {
case CLI_INIT:
e->command = "stasis statistics show subscription";
e->usage =
"Usage: stasis statistics show subscription <uniqueid>\n"
" Show stasis subscription statistics.\n";
return NULL;
case CLI_GENERATE:
if (a->pos == 4) {
return subscription_statistics_complete_name(a->word, a->n);
} else {
return NULL;
}
}
if (a->argc != 5) {
return CLI_SHOWUSAGE;
}
subscription_stats = ao2_global_obj_ref(subscription_statistics);
if (!subscription_stats) {
ast_cli(a->fd, "Could not fetch subcription_statistics container\n");
return CLI_FAILURE;
}
statistics = ao2_find(subscription_stats, a->argv[4], OBJ_SEARCH_KEY);
if (!statistics) {
ao2_ref(subscription_stats, -1);
ast_cli(a->fd, "Specified subscription '%s' does not exist\n", a->argv[4]);
return CLI_FAILURE;
}
ao2_ref(subscription_stats, -1);
ast_cli(a->fd, "Subscription: %s\n", statistics->uniqueid);
ast_cli(a->fd, "Pointer Address: %p\n", statistics->sub);
ast_cli(a->fd, "Source filename: %s\n", S_OR(statistics->file, "<unavailable>"));
ast_cli(a->fd, "Source line number: %d\n", statistics->lineno);
ast_cli(a->fd, "Source function: %s\n", S_OR(statistics->func, "<unavailable>"));
ast_cli(a->fd, "Number of messages dropped due to filtering: %d\n", statistics->messages_dropped);
ast_cli(a->fd, "Number of messages passed to subscriber callback: %d\n", statistics->messages_passed);
ast_cli(a->fd, "Using mailbox to queue messages: %s\n", statistics->uses_mailbox ? "Yes" : "No");
ast_cli(a->fd, "Using stasis threadpool for handling messages: %s\n", statistics->uses_threadpool ? "Yes" : "No");
ast_cli(a->fd, "Lowest amount of time (in milliseconds) spent invoking message: %ld\n", statistics->lowest_time_invoked);
ast_cli(a->fd, "Highest amount of time (in milliseconds) spent invoking message: %ld\n", statistics->highest_time_invoked);
ao2_lock(statistics);
if (statistics->highest_time_message_type) {
ast_cli(a->fd, "Offender message type for highest invoking time: %s\n", stasis_message_type_name(statistics->highest_time_message_type));
}
ao2_unlock(statistics);
ast_cli(a->fd, "Number of topics: %d\n", ao2_container_count(statistics->topics));
ast_cli(a->fd, "Subscribed topics:\n");
i = ao2_iterator_init(statistics->topics, 0);
while ((name = ao2_iterator_next(&i))) {
ast_cli(a->fd, "\t%s\n", name);
ao2_ref(name, -1);
}
ao2_iterator_destroy(&i);
ao2_ref(statistics, -1);
return CLI_SUCCESS;
}
AO2_STRING_FIELD_SORT_FN(stasis_topic_statistics, name);
/*!
* \internal
* \brief CLI command implementation for 'stasis statistics show topics'
*/
static char *statistics_show_topics(struct ast_cli_entry *e, int cmd, struct ast_cli_args *a)
{
struct ao2_container *sorted_topics;
struct ao2_container *topic_stats;
struct ao2_iterator iter;
struct stasis_topic_statistics *statistics;
int count = 0;
int not_dispatched = 0;
int dispatched = 0;
#define FMT_HEADERS "%-64s %10s %10s %10s %16s %16s\n"
#define FMT_FIELDS "%-64s %10d %10d %10d %16ld %16ld\n"
#define FMT_FIELDS2 "%-64s %10s %10d %10d\n"
switch (cmd) {
case CLI_INIT:
e->command = "stasis statistics show topics";
e->usage =
"Usage: stasis statistics show topics\n"
" Shows a list of topics and their general statistics\n";
return NULL;
case CLI_GENERATE:
return NULL;
}
if (a->argc != e->args) {
return CLI_SHOWUSAGE;
}
topic_stats = ao2_global_obj_ref(topic_statistics);
if (!topic_stats) {
ast_cli(a->fd, "Could not fetch topic_statistics container\n");
return CLI_FAILURE;
}
sorted_topics = ao2_container_alloc_rbtree(AO2_ALLOC_OPT_LOCK_NOLOCK, 0,
stasis_topic_statistics_sort_fn, NULL);
if (!sorted_topics) {
ao2_ref(topic_stats, -1);
ast_cli(a->fd, "Could not create container for sorting topic statistics\n");
return CLI_SUCCESS;
}
if (ao2_container_dup(sorted_topics, topic_stats, 0)) {
ao2_ref(sorted_topics, -1);
ao2_ref(topic_stats, -1);
ast_cli(a->fd, "Could not sort topic statistics\n");
return CLI_SUCCESS;
}
ao2_ref(topic_stats, -1);
ast_cli(a->fd, "\n" FMT_HEADERS, "Topic", "Subscribers", "Dropped", "Dispatched", "Lowest Dispatch", "Highest Dispatch");
iter = ao2_iterator_init(sorted_topics, 0);
while ((statistics = ao2_iterator_next(&iter))) {
ast_cli(a->fd, FMT_FIELDS, statistics->name, ao2_container_count(statistics->subscribers),
statistics->messages_not_dispatched, statistics->messages_dispatched,
statistics->lowest_time_dispatched, statistics->highest_time_dispatched);
not_dispatched += statistics->messages_not_dispatched;
dispatched += statistics->messages_dispatched;
ao2_ref(statistics, -1);
++count;
}
ao2_iterator_destroy(&iter);
ao2_ref(sorted_topics, -1);
ast_cli(a->fd, FMT_FIELDS2, "Total", "", not_dispatched, dispatched);
ast_cli(a->fd, "\n%d topics\n\n", count);
#undef FMT_HEADERS
#undef FMT_FIELDS
#undef FMT_FIELDS2
return CLI_SUCCESS;
}
/*!
* \internal
* \brief CLI tab completion for topic statistics names
*/
static char *topic_statistics_complete_name(const char *word, int state)
{
struct stasis_topic_statistics *statistics;
struct ao2_container *topic_stats;
struct ao2_iterator it_statistics;
int wordlen = strlen(word);
int which = 0;
char *result = NULL;
topic_stats = ao2_global_obj_ref(topic_statistics);
if (!topic_stats) {
return result;
}
it_statistics = ao2_iterator_init(topic_stats, 0);
while ((statistics = ao2_iterator_next(&it_statistics))) {
if (!strncasecmp(word, statistics->name, wordlen)
&& ++which > state) {
result = ast_strdup(statistics->name);
}
ao2_ref(statistics, -1);
if (result) {
break;
}
}
ao2_iterator_destroy(&it_statistics);
ao2_ref(topic_stats, -1);
return result;
}
/*!
* \internal
* \brief CLI command implementation for 'stasis statistics show topic'
*/
static char *statistics_show_topic(struct ast_cli_entry *e, int cmd, struct ast_cli_args *a)
{
struct stasis_topic_statistics *statistics;
struct ao2_container *topic_stats;
struct ao2_iterator i;
char *uniqueid;
switch (cmd) {
case CLI_INIT:
e->command = "stasis statistics show topic";
e->usage =
"Usage: stasis statistics show topic <name>\n"
" Show stasis topic statistics.\n";
return NULL;
case CLI_GENERATE:
if (a->pos == 4) {
return topic_statistics_complete_name(a->word, a->n);
} else {
return NULL;
}
}
if (a->argc != 5) {
return CLI_SHOWUSAGE;
}
topic_stats = ao2_global_obj_ref(topic_statistics);
if (!topic_stats) {
ast_cli(a->fd, "Could not fetch topic_statistics container\n");
return CLI_FAILURE;
}
statistics = ao2_find(topic_stats, a->argv[4], OBJ_SEARCH_KEY);
if (!statistics) {
ao2_ref(topic_stats, -1);
ast_cli(a->fd, "Specified topic '%s' does not exist\n", a->argv[4]);
return CLI_FAILURE;
}
ao2_ref(topic_stats, -1);
ast_cli(a->fd, "Topic: %s\n", statistics->name);
ast_cli(a->fd, "Pointer Address: %p\n", statistics->topic);
ast_cli(a->fd, "Number of messages published that went to no subscriber: %d\n", statistics->messages_not_dispatched);
ast_cli(a->fd, "Number of messages that went to at least one subscriber: %d\n", statistics->messages_dispatched);
ast_cli(a->fd, "Lowest amount of time (in milliseconds) spent dispatching message: %ld\n", statistics->lowest_time_dispatched);
ast_cli(a->fd, "Highest amount of time (in milliseconds) spent dispatching messages: %ld\n", statistics->highest_time_dispatched);
ast_cli(a->fd, "Number of subscribers: %d\n", ao2_container_count(statistics->subscribers));
ast_cli(a->fd, "Subscribers:\n");
i = ao2_iterator_init(statistics->subscribers, 0);
while ((uniqueid = ao2_iterator_next(&i))) {
ast_cli(a->fd, "\t%s\n", uniqueid);
ao2_ref(uniqueid, -1);
}
ao2_iterator_destroy(&i);
ao2_ref(statistics, -1);
return CLI_SUCCESS;
}
/*!
* \internal
* \brief CLI command implementation for 'stasis statistics show messages'
*/
static char *statistics_show_messages(struct ast_cli_entry *e, int cmd, struct ast_cli_args *a)
{
int i;
int count = 0;
int published = 0;
int unused = 0;
#define FMT_HEADERS "%-64s %10s %10s\n"
#define FMT_FIELDS "%-64s %10d %10d\n"
switch (cmd) {
case CLI_INIT:
e->command = "stasis statistics show messages";
e->usage =
"Usage: stasis statistics show messages\n"
" Shows a list of message types and their general statistics\n";
return NULL;
case CLI_GENERATE:
return NULL;
}
if (a->argc != e->args) {
return CLI_SHOWUSAGE;
}
ast_cli(a->fd, "\n" FMT_HEADERS, "Message Type", "Published", "Unused");
ast_mutex_lock(&message_type_statistics_lock);
for (i = 0; i < AST_VECTOR_SIZE(&message_type_statistics); ++i) {
struct stasis_message_type_statistics *statistics = AST_VECTOR_GET_ADDR(&message_type_statistics, i);
if (!statistics->message_type) {
continue;
}
ast_cli(a->fd, FMT_FIELDS, stasis_message_type_name(statistics->message_type), statistics->published,
statistics->unused);
published += statistics->published;
unused += statistics->unused;
++count;
}
ast_mutex_unlock(&message_type_statistics_lock);
ast_cli(a->fd, FMT_FIELDS, "Total", published, unused);
ast_cli(a->fd, "\n%d seen message types\n\n", count);
#undef FMT_HEADERS
#undef FMT_FIELDS
return CLI_SUCCESS;
}
static struct ast_cli_entry cli_stasis_statistics[] = {
AST_CLI_DEFINE(statistics_show_subscriptions, "Show subscriptions with general statistics"),
AST_CLI_DEFINE(statistics_show_subscription, "Show subscription statistics"),
AST_CLI_DEFINE(statistics_show_topics, "Show topics with general statistics"),
AST_CLI_DEFINE(statistics_show_topic, "Show topic statistics"),
AST_CLI_DEFINE(statistics_show_messages, "Show message types with general statistics"),
};
static int subscription_statistics_hash(const void *obj, const int flags)
{
const struct stasis_subscription_statistics *object;
const char *key;
switch (flags & OBJ_SEARCH_MASK) {
case OBJ_SEARCH_KEY:
key = obj;
break;
case OBJ_SEARCH_OBJECT:
object = obj;
key = object->uniqueid;
break;
default:
/* Hash can only work on something with a full key. */
ast_assert(0);
return 0;
}
return ast_str_case_hash(key);
}
static int subscription_statistics_cmp(void *obj, void *arg, int flags)
{
const struct stasis_subscription_statistics *object_left = obj;
const struct stasis_subscription_statistics *object_right = arg;
const char *right_key = arg;
int cmp;
switch (flags & OBJ_SEARCH_MASK) {
case OBJ_SEARCH_OBJECT:
right_key = object_right->uniqueid;
/* Fall through */
case OBJ_SEARCH_KEY:
cmp = strcasecmp(object_left->uniqueid, right_key);
break;
case OBJ_SEARCH_PARTIAL_KEY:
/* Not supported by container */
ast_assert(0);
cmp = -1;
break;
default:
/*
* What arg points to is specific to this traversal callback
* and has no special meaning to astobj2.
*/
cmp = 0;
break;
}
if (cmp) {
return 0;
}
/*
* At this point the traversal callback is identical to a sorted
* container.
*/
return CMP_MATCH;
}
static int topic_statistics_hash(const void *obj, const int flags)
{
const struct stasis_topic_statistics *object;
const char *key;
switch (flags & OBJ_SEARCH_MASK) {
case OBJ_SEARCH_KEY:
key = obj;
break;
case OBJ_SEARCH_OBJECT:
object = obj;
key = object->name;
break;
default:
/* Hash can only work on something with a full key. */
ast_assert(0);
return 0;
}
return ast_str_case_hash(key);
}
static int topic_statistics_cmp(void *obj, void *arg, int flags)
{
const struct stasis_topic_statistics *object_left = obj;
const struct stasis_topic_statistics *object_right = arg;
const char *right_key = arg;
int cmp;
switch (flags & OBJ_SEARCH_MASK) {
case OBJ_SEARCH_OBJECT:
right_key = object_right->name;
/* Fall through */
case OBJ_SEARCH_KEY:
cmp = strcasecmp(object_left->name, right_key);
break;
case OBJ_SEARCH_PARTIAL_KEY:
/* Not supported by container */
ast_assert(0);
cmp = -1;
break;
default:
/*
* What arg points to is specific to this traversal callback
* and has no special meaning to astobj2.
*/
cmp = 0;
break;
}
if (cmp) {
return 0;
}
/*
* At this point the traversal callback is identical to a sorted
* container.
*/
return CMP_MATCH;
}
#endif
Avoid unnecessary cleanups during immediate shutdown This patch addresses issues during immediate shutdowns, where modules are not unloaded, but Asterisk atexit handlers are run. In the typical case, this usually isn't a big deal. But the introduction of the Stasis message bus makes it much more likely for asynchronous activity to be happening off in some thread during shutdown. During an immediate shutdown, Asterisk skips unloading modules. But while it is processing the atexit handlers, there is a window of time where some of the core message types have been cleaned up, but the message bus is still running. Specifically, it's still running module subscriptions that might be using the core message types. If a message is received by that subscription in that window, it will attempt to use a message type that has been cleaned up. To solve this problem, this patch introduces ast_register_cleanup(). This function operates identically to ast_register_atexit(), except that cleanup calls are not invoked on an immediate shutdown. All of the core message type and topic cleanup was moved from atexit handlers to cleanup handlers. This ensures that core type and topic cleanup only happens if the modules that used them are first unloaded. This patch also changes the ast_assert() when accessing a cleaned up or uninitialized message type to an error log message. Message type functions are actually NULL safe across the board, so the assert was a bit heavy handed. Especially for anyone with DO_CRASH enabled. Review: https://reviewboard.asterisk.org/r/2562/ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@390122 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-05-30 17:05:53 +00:00
/*! \brief Cleanup function for graceful shutdowns */
static void stasis_cleanup(void)
{
#ifdef AST_DEVMODE
ast_cli_unregister_multiple(cli_stasis_statistics, ARRAY_LEN(cli_stasis_statistics));
AST_VECTOR_FREE(&message_type_statistics);
ao2_global_obj_release(subscription_statistics);
ao2_global_obj_release(topic_statistics);
#endif
ast_cli_unregister_multiple(cli_stasis, ARRAY_LEN(cli_stasis));
ao2_cleanup(topic_all);
topic_all = NULL;
ast_threadpool_shutdown(threadpool);
threadpool = NULL;
STASIS_MESSAGE_TYPE_CLEANUP(stasis_subscription_change_type);
STASIS_MESSAGE_TYPE_CLEANUP(ast_multi_user_event_type);
aco_info_destroy(&cfg_info);
ao2_global_obj_release(globals);
}
int stasis_init(void)
{
struct stasis_config *cfg;
int cache_init;
main/stasis: Allow subscriptions to use a threadpool for message delivery Prior to this patch, all Stasis subscriptions would receive a dedicated thread for servicing published messages. In contrast, prior to r400178 (see review https://reviewboard.asterisk.org/r/2881/), the subscriptions shared a thread pool. It was discovered during some initial work on Stasis that, for a low subscription count with high message throughput, the threadpool was not as performant as simply having a dedicated thread per subscriber. For situations where a subscriber receives a substantial number of messages and is always present, the model of having a dedicated thread per subscriber makes sense. While we still have plenty of subscriptions that would follow this model, e.g., AMI, CDRs, CEL, etc., there are plenty that also fall into the following two categories: * Large number of subscriptions, specifically those tied to endpoints/peers. * Low number of messages. Some subscriptions exist specifically to coordinate a single message - the subscription is created, a message is published, the delivery is synchronized, and the subscription is destroyed. In both of the latter two cases, creating a dedicated thread is wasteful (and in the case of a large number of peers/endpoints, harmful). In those cases, having shared delivery threads is far more performant. This patch adds the ability of a subscriber to Stasis to choose whether or not their messages are dispatched on a dedicated thread or on a threadpool. The threadpool is configurable through stasis.conf. Review: https://reviewboard.asterisk.org/r/4193 ASTERISK-24533 #close Reported by: xrobau Tested by: xrobau ........ Merged revisions 428681 from http://svn.asterisk.org/svn/asterisk/branches/12 ........ Merged revisions 428687 from http://svn.asterisk.org/svn/asterisk/branches/13 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@428688 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-12-01 17:59:21 +00:00
struct ast_threadpool_options threadpool_opts = { 0, };
#ifdef AST_DEVMODE
struct ao2_container *subscription_stats;
struct ao2_container *topic_stats;
#endif
Avoid unnecessary cleanups during immediate shutdown This patch addresses issues during immediate shutdowns, where modules are not unloaded, but Asterisk atexit handlers are run. In the typical case, this usually isn't a big deal. But the introduction of the Stasis message bus makes it much more likely for asynchronous activity to be happening off in some thread during shutdown. During an immediate shutdown, Asterisk skips unloading modules. But while it is processing the atexit handlers, there is a window of time where some of the core message types have been cleaned up, but the message bus is still running. Specifically, it's still running module subscriptions that might be using the core message types. If a message is received by that subscription in that window, it will attempt to use a message type that has been cleaned up. To solve this problem, this patch introduces ast_register_cleanup(). This function operates identically to ast_register_atexit(), except that cleanup calls are not invoked on an immediate shutdown. All of the core message type and topic cleanup was moved from atexit handlers to cleanup handlers. This ensures that core type and topic cleanup only happens if the modules that used them are first unloaded. This patch also changes the ast_assert() when accessing a cleaned up or uninitialized message type to an error log message. Message type functions are actually NULL safe across the board, so the assert was a bit heavy handed. Especially for anyone with DO_CRASH enabled. Review: https://reviewboard.asterisk.org/r/2562/ git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@390122 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2013-05-30 17:05:53 +00:00
/* Be sure the types are cleaned up after the message bus */
ast_register_cleanup(stasis_cleanup);
if (aco_info_init(&cfg_info)) {
return -1;
}
main/stasis: Allow subscriptions to use a threadpool for message delivery Prior to this patch, all Stasis subscriptions would receive a dedicated thread for servicing published messages. In contrast, prior to r400178 (see review https://reviewboard.asterisk.org/r/2881/), the subscriptions shared a thread pool. It was discovered during some initial work on Stasis that, for a low subscription count with high message throughput, the threadpool was not as performant as simply having a dedicated thread per subscriber. For situations where a subscriber receives a substantial number of messages and is always present, the model of having a dedicated thread per subscriber makes sense. While we still have plenty of subscriptions that would follow this model, e.g., AMI, CDRs, CEL, etc., there are plenty that also fall into the following two categories: * Large number of subscriptions, specifically those tied to endpoints/peers. * Low number of messages. Some subscriptions exist specifically to coordinate a single message - the subscription is created, a message is published, the delivery is synchronized, and the subscription is destroyed. In both of the latter two cases, creating a dedicated thread is wasteful (and in the case of a large number of peers/endpoints, harmful). In those cases, having shared delivery threads is far more performant. This patch adds the ability of a subscriber to Stasis to choose whether or not their messages are dispatched on a dedicated thread or on a threadpool. The threadpool is configurable through stasis.conf. Review: https://reviewboard.asterisk.org/r/4193 ASTERISK-24533 #close Reported by: xrobau Tested by: xrobau ........ Merged revisions 428681 from http://svn.asterisk.org/svn/asterisk/branches/12 ........ Merged revisions 428687 from http://svn.asterisk.org/svn/asterisk/branches/13 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@428688 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-12-01 17:59:21 +00:00
aco_option_register_custom(&cfg_info, "decline", ACO_EXACT,
declined_options, "", declined_handler, 0);
aco_option_register(&cfg_info, "initial_size", ACO_EXACT,
threadpool_options, "5", OPT_INT_T, PARSE_IN_RANGE,
FLDSET(struct stasis_threadpool_conf, initial_size), 0,
INT_MAX);
aco_option_register(&cfg_info, "idle_timeout_sec", ACO_EXACT,
threadpool_options, "20", OPT_INT_T, PARSE_IN_RANGE,
FLDSET(struct stasis_threadpool_conf, idle_timeout_sec), 0,
INT_MAX);
aco_option_register(&cfg_info, "max_size", ACO_EXACT,
threadpool_options, "50", OPT_INT_T, PARSE_IN_RANGE,
FLDSET(struct stasis_threadpool_conf, max_size), 0,
INT_MAX);
if (aco_process_config(&cfg_info, 0) == ACO_PROCESS_ERROR) {
main/stasis: Allow subscriptions to use a threadpool for message delivery Prior to this patch, all Stasis subscriptions would receive a dedicated thread for servicing published messages. In contrast, prior to r400178 (see review https://reviewboard.asterisk.org/r/2881/), the subscriptions shared a thread pool. It was discovered during some initial work on Stasis that, for a low subscription count with high message throughput, the threadpool was not as performant as simply having a dedicated thread per subscriber. For situations where a subscriber receives a substantial number of messages and is always present, the model of having a dedicated thread per subscriber makes sense. While we still have plenty of subscriptions that would follow this model, e.g., AMI, CDRs, CEL, etc., there are plenty that also fall into the following two categories: * Large number of subscriptions, specifically those tied to endpoints/peers. * Low number of messages. Some subscriptions exist specifically to coordinate a single message - the subscription is created, a message is published, the delivery is synchronized, and the subscription is destroyed. In both of the latter two cases, creating a dedicated thread is wasteful (and in the case of a large number of peers/endpoints, harmful). In those cases, having shared delivery threads is far more performant. This patch adds the ability of a subscriber to Stasis to choose whether or not their messages are dispatched on a dedicated thread or on a threadpool. The threadpool is configurable through stasis.conf. Review: https://reviewboard.asterisk.org/r/4193 ASTERISK-24533 #close Reported by: xrobau Tested by: xrobau ........ Merged revisions 428681 from http://svn.asterisk.org/svn/asterisk/branches/12 ........ Merged revisions 428687 from http://svn.asterisk.org/svn/asterisk/branches/13 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@428688 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-12-01 17:59:21 +00:00
struct stasis_config *default_cfg = stasis_config_alloc();
if (!default_cfg) {
return -1;
}
main/stasis: Allow subscriptions to use a threadpool for message delivery Prior to this patch, all Stasis subscriptions would receive a dedicated thread for servicing published messages. In contrast, prior to r400178 (see review https://reviewboard.asterisk.org/r/2881/), the subscriptions shared a thread pool. It was discovered during some initial work on Stasis that, for a low subscription count with high message throughput, the threadpool was not as performant as simply having a dedicated thread per subscriber. For situations where a subscriber receives a substantial number of messages and is always present, the model of having a dedicated thread per subscriber makes sense. While we still have plenty of subscriptions that would follow this model, e.g., AMI, CDRs, CEL, etc., there are plenty that also fall into the following two categories: * Large number of subscriptions, specifically those tied to endpoints/peers. * Low number of messages. Some subscriptions exist specifically to coordinate a single message - the subscription is created, a message is published, the delivery is synchronized, and the subscription is destroyed. In both of the latter two cases, creating a dedicated thread is wasteful (and in the case of a large number of peers/endpoints, harmful). In those cases, having shared delivery threads is far more performant. This patch adds the ability of a subscriber to Stasis to choose whether or not their messages are dispatched on a dedicated thread or on a threadpool. The threadpool is configurable through stasis.conf. Review: https://reviewboard.asterisk.org/r/4193 ASTERISK-24533 #close Reported by: xrobau Tested by: xrobau ........ Merged revisions 428681 from http://svn.asterisk.org/svn/asterisk/branches/12 ........ Merged revisions 428687 from http://svn.asterisk.org/svn/asterisk/branches/13 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@428688 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-12-01 17:59:21 +00:00
if (aco_set_defaults(&threadpool_option, "threadpool", default_cfg->threadpool_options)) {
ast_log(LOG_ERROR, "Failed to initialize defaults on Stasis configuration object\n");
ao2_ref(default_cfg, -1);
main/stasis: Allow subscriptions to use a threadpool for message delivery Prior to this patch, all Stasis subscriptions would receive a dedicated thread for servicing published messages. In contrast, prior to r400178 (see review https://reviewboard.asterisk.org/r/2881/), the subscriptions shared a thread pool. It was discovered during some initial work on Stasis that, for a low subscription count with high message throughput, the threadpool was not as performant as simply having a dedicated thread per subscriber. For situations where a subscriber receives a substantial number of messages and is always present, the model of having a dedicated thread per subscriber makes sense. While we still have plenty of subscriptions that would follow this model, e.g., AMI, CDRs, CEL, etc., there are plenty that also fall into the following two categories: * Large number of subscriptions, specifically those tied to endpoints/peers. * Low number of messages. Some subscriptions exist specifically to coordinate a single message - the subscription is created, a message is published, the delivery is synchronized, and the subscription is destroyed. In both of the latter two cases, creating a dedicated thread is wasteful (and in the case of a large number of peers/endpoints, harmful). In those cases, having shared delivery threads is far more performant. This patch adds the ability of a subscriber to Stasis to choose whether or not their messages are dispatched on a dedicated thread or on a threadpool. The threadpool is configurable through stasis.conf. Review: https://reviewboard.asterisk.org/r/4193 ASTERISK-24533 #close Reported by: xrobau Tested by: xrobau ........ Merged revisions 428681 from http://svn.asterisk.org/svn/asterisk/branches/12 ........ Merged revisions 428687 from http://svn.asterisk.org/svn/asterisk/branches/13 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@428688 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-12-01 17:59:21 +00:00
return -1;
}
if (aco_set_defaults(&declined_option, "declined_message_types", default_cfg->declined_message_types)) {
ast_log(LOG_ERROR, "Failed to load stasis.conf and failed to initialize defaults.\n");
ao2_ref(default_cfg, -1);
return -1;
}
main/stasis: Allow subscriptions to use a threadpool for message delivery Prior to this patch, all Stasis subscriptions would receive a dedicated thread for servicing published messages. In contrast, prior to r400178 (see review https://reviewboard.asterisk.org/r/2881/), the subscriptions shared a thread pool. It was discovered during some initial work on Stasis that, for a low subscription count with high message throughput, the threadpool was not as performant as simply having a dedicated thread per subscriber. For situations where a subscriber receives a substantial number of messages and is always present, the model of having a dedicated thread per subscriber makes sense. While we still have plenty of subscriptions that would follow this model, e.g., AMI, CDRs, CEL, etc., there are plenty that also fall into the following two categories: * Large number of subscriptions, specifically those tied to endpoints/peers. * Low number of messages. Some subscriptions exist specifically to coordinate a single message - the subscription is created, a message is published, the delivery is synchronized, and the subscription is destroyed. In both of the latter two cases, creating a dedicated thread is wasteful (and in the case of a large number of peers/endpoints, harmful). In those cases, having shared delivery threads is far more performant. This patch adds the ability of a subscriber to Stasis to choose whether or not their messages are dispatched on a dedicated thread or on a threadpool. The threadpool is configurable through stasis.conf. Review: https://reviewboard.asterisk.org/r/4193 ASTERISK-24533 #close Reported by: xrobau Tested by: xrobau ........ Merged revisions 428681 from http://svn.asterisk.org/svn/asterisk/branches/12 ........ Merged revisions 428687 from http://svn.asterisk.org/svn/asterisk/branches/13 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@428688 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-12-01 17:59:21 +00:00
ast_log(LOG_NOTICE, "Could not load Stasis configuration; using defaults\n");
ao2_global_obj_replace_unref(globals, default_cfg);
cfg = default_cfg;
} else {
cfg = ao2_global_obj_ref(globals);
if (!cfg) {
ast_log(LOG_ERROR, "Failed to obtain Stasis configuration object\n");
main/stasis: Allow subscriptions to use a threadpool for message delivery Prior to this patch, all Stasis subscriptions would receive a dedicated thread for servicing published messages. In contrast, prior to r400178 (see review https://reviewboard.asterisk.org/r/2881/), the subscriptions shared a thread pool. It was discovered during some initial work on Stasis that, for a low subscription count with high message throughput, the threadpool was not as performant as simply having a dedicated thread per subscriber. For situations where a subscriber receives a substantial number of messages and is always present, the model of having a dedicated thread per subscriber makes sense. While we still have plenty of subscriptions that would follow this model, e.g., AMI, CDRs, CEL, etc., there are plenty that also fall into the following two categories: * Large number of subscriptions, specifically those tied to endpoints/peers. * Low number of messages. Some subscriptions exist specifically to coordinate a single message - the subscription is created, a message is published, the delivery is synchronized, and the subscription is destroyed. In both of the latter two cases, creating a dedicated thread is wasteful (and in the case of a large number of peers/endpoints, harmful). In those cases, having shared delivery threads is far more performant. This patch adds the ability of a subscriber to Stasis to choose whether or not their messages are dispatched on a dedicated thread or on a threadpool. The threadpool is configurable through stasis.conf. Review: https://reviewboard.asterisk.org/r/4193 ASTERISK-24533 #close Reported by: xrobau Tested by: xrobau ........ Merged revisions 428681 from http://svn.asterisk.org/svn/asterisk/branches/12 ........ Merged revisions 428687 from http://svn.asterisk.org/svn/asterisk/branches/13 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@428688 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-12-01 17:59:21 +00:00
return -1;
}
}
threadpool_opts.version = AST_THREADPOOL_OPTIONS_VERSION;
threadpool_opts.initial_size = cfg->threadpool_options->initial_size;
threadpool_opts.auto_increment = 1;
threadpool_opts.max_size = cfg->threadpool_options->max_size;
threadpool_opts.idle_timeout = cfg->threadpool_options->idle_timeout_sec;
threadpool = ast_threadpool_create("stasis", NULL, &threadpool_opts);
ao2_ref(cfg, -1);
if (!threadpool) {
main/stasis: Allow subscriptions to use a threadpool for message delivery Prior to this patch, all Stasis subscriptions would receive a dedicated thread for servicing published messages. In contrast, prior to r400178 (see review https://reviewboard.asterisk.org/r/2881/), the subscriptions shared a thread pool. It was discovered during some initial work on Stasis that, for a low subscription count with high message throughput, the threadpool was not as performant as simply having a dedicated thread per subscriber. For situations where a subscriber receives a substantial number of messages and is always present, the model of having a dedicated thread per subscriber makes sense. While we still have plenty of subscriptions that would follow this model, e.g., AMI, CDRs, CEL, etc., there are plenty that also fall into the following two categories: * Large number of subscriptions, specifically those tied to endpoints/peers. * Low number of messages. Some subscriptions exist specifically to coordinate a single message - the subscription is created, a message is published, the delivery is synchronized, and the subscription is destroyed. In both of the latter two cases, creating a dedicated thread is wasteful (and in the case of a large number of peers/endpoints, harmful). In those cases, having shared delivery threads is far more performant. This patch adds the ability of a subscriber to Stasis to choose whether or not their messages are dispatched on a dedicated thread or on a threadpool. The threadpool is configurable through stasis.conf. Review: https://reviewboard.asterisk.org/r/4193 ASTERISK-24533 #close Reported by: xrobau Tested by: xrobau ........ Merged revisions 428681 from http://svn.asterisk.org/svn/asterisk/branches/12 ........ Merged revisions 428687 from http://svn.asterisk.org/svn/asterisk/branches/13 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@428688 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-12-01 17:59:21 +00:00
ast_log(LOG_ERROR, "Failed to create 'stasis-core' threadpool\n");
main/stasis: Allow subscriptions to use a threadpool for message delivery Prior to this patch, all Stasis subscriptions would receive a dedicated thread for servicing published messages. In contrast, prior to r400178 (see review https://reviewboard.asterisk.org/r/2881/), the subscriptions shared a thread pool. It was discovered during some initial work on Stasis that, for a low subscription count with high message throughput, the threadpool was not as performant as simply having a dedicated thread per subscriber. For situations where a subscriber receives a substantial number of messages and is always present, the model of having a dedicated thread per subscriber makes sense. While we still have plenty of subscriptions that would follow this model, e.g., AMI, CDRs, CEL, etc., there are plenty that also fall into the following two categories: * Large number of subscriptions, specifically those tied to endpoints/peers. * Low number of messages. Some subscriptions exist specifically to coordinate a single message - the subscription is created, a message is published, the delivery is synchronized, and the subscription is destroyed. In both of the latter two cases, creating a dedicated thread is wasteful (and in the case of a large number of peers/endpoints, harmful). In those cases, having shared delivery threads is far more performant. This patch adds the ability of a subscriber to Stasis to choose whether or not their messages are dispatched on a dedicated thread or on a threadpool. The threadpool is configurable through stasis.conf. Review: https://reviewboard.asterisk.org/r/4193 ASTERISK-24533 #close Reported by: xrobau Tested by: xrobau ........ Merged revisions 428681 from http://svn.asterisk.org/svn/asterisk/branches/12 ........ Merged revisions 428687 from http://svn.asterisk.org/svn/asterisk/branches/13 git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@428688 65c4cc65-6c06-0410-ace0-fbb531ad65f3
2014-12-01 17:59:21 +00:00
return -1;
}
cache_init = stasis_cache_init();
if (cache_init != 0) {
return -1;
}
if (STASIS_MESSAGE_TYPE_INIT(stasis_subscription_change_type) != 0) {
return -1;
}
if (STASIS_MESSAGE_TYPE_INIT(ast_multi_user_event_type) != 0) {
return -1;
}
topic_all = ao2_container_alloc_hash(AO2_ALLOC_OPT_LOCK_MUTEX, 0, TOPIC_ALL_BUCKETS,
topic_proxy_hash_fn, 0, topic_proxy_cmp_fn);
if (!topic_all) {
return -1;
}
if (ast_cli_register_multiple(cli_stasis, ARRAY_LEN(cli_stasis))) {
return -1;
}
#ifdef AST_DEVMODE
/* Statistics information is stored separately so that we don't alter or interrupt the lifetime of the underlying
* topic or subscripton.
*/
subscription_stats = ao2_container_alloc_hash(AO2_ALLOC_OPT_LOCK_MUTEX, 0, SUBSCRIPTION_STATISTICS_BUCKETS,
subscription_statistics_hash, 0, subscription_statistics_cmp);
if (!subscription_stats) {
return -1;
}
ao2_global_obj_replace_unref(subscription_statistics, subscription_stats);
ao2_cleanup(subscription_stats);
topic_stats = ao2_container_alloc_hash(AO2_ALLOC_OPT_LOCK_MUTEX, 0, TOPIC_STATISTICS_BUCKETS,
topic_statistics_hash, 0, topic_statistics_cmp);
if (!topic_stats) {
return -1;
}
ao2_global_obj_replace_unref(topic_statistics, topic_stats);
ao2_cleanup(topic_stats);
if (!topic_stats) {
return -1;
}
AST_VECTOR_INIT(&message_type_statistics, 0);
if (ast_cli_register_multiple(cli_stasis_statistics, ARRAY_LEN(cli_stasis_statistics))) {
return -1;
}
#endif
return 0;
}