2008-05-03 03:40:32 +00:00
|
|
|
/*
|
|
|
|
* Asterisk -- An open source telephony toolkit.
|
|
|
|
*
|
2013-01-15 20:15:00 +00:00
|
|
|
* Copyright (C) 2007-2013, Digium, Inc.
|
2008-05-03 03:40:32 +00:00
|
|
|
*
|
|
|
|
* Dwayne M. Hubbard <dhubbard@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.
|
|
|
|
*/
|
2010-02-22 06:45:52 +00:00
|
|
|
|
|
|
|
/*!
|
|
|
|
* \file
|
2008-05-03 03:40:32 +00:00
|
|
|
* \brief Maintain a container of uniquely-named taskprocessor threads that can be shared across modules.
|
|
|
|
*
|
|
|
|
* \author Dwayne Hubbard <dhubbard@digium.com>
|
|
|
|
*/
|
|
|
|
|
2012-06-15 16:20:16 +00:00
|
|
|
/*** MODULEINFO
|
|
|
|
<support_level>core</support_level>
|
|
|
|
***/
|
|
|
|
|
2008-05-03 03:40:32 +00:00
|
|
|
#include "asterisk.h"
|
2008-06-10 00:29:01 +00:00
|
|
|
|
git migration: Refactor the ASTERISK_FILE_VERSION macro
Git does not support the ability to replace a token with a version
string during check-in. While it does have support for replacing a
token on clone, this is somewhat sub-optimal: the token is replaced
with the object hash, which is not particularly easy for human
consumption. What's more, in practice, the source file version was often
not terribly useful. Generally, when triaging bugs, the overall version
of Asterisk is far more useful than an individual SVN version of a file. As a
result, this patch removes Asterisk's support for showing source file
versions.
Specifically, it does the following:
* Rename ASTERISK_FILE_VERSION macro to ASTERISK_REGISTER_FILE, and
remove passing the version in with the macro. Other facilities
than 'core show file version' make use of the file names, such as
setting a debug level only on a specific file. As such, the act of
registering source files with the Asterisk core still has use. The
macro rename now reflects the new macro purpose.
* main/asterisk:
- Refactor the file_version structure to reflect that it no longer
tracks a version field.
- Remove the "core show file version" CLI command. Without the file
version, it is no longer useful.
- Remove the ast_file_version_find function. The file version is no
longer tracked.
- Rename ast_register_file_version/ast_unregister_file_version to
ast_register_file/ast_unregister_file, respectively.
* main/manager: Remove value from the Version key of the ModuleCheck
Action. The actual key itself has not been removed, as doing so would
absolutely constitute a backwards incompatible change. However, since
the file version is no longer tracked, there is no need to attempt to
include it in the Version key.
* UPGRADE: Add notes for:
- Modification to the ModuleCheck AMI Action
- Removal of the "core show file version" CLI command
Change-Id: I6cf0ff280e1668bf4957dc21f32a5ff43444a40e
2015-04-12 02:38:22 +00:00
|
|
|
ASTERISK_REGISTER_FILE()
|
2008-05-29 20:26:42 +00:00
|
|
|
|
2008-05-03 03:40:32 +00:00
|
|
|
#include "asterisk/_private.h"
|
|
|
|
#include "asterisk/module.h"
|
|
|
|
#include "asterisk/time.h"
|
|
|
|
#include "asterisk/astobj2.h"
|
|
|
|
#include "asterisk/cli.h"
|
|
|
|
#include "asterisk/taskprocessor.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/sem.h"
|
2008-05-03 03:40:32 +00:00
|
|
|
|
2010-02-22 06:45:52 +00:00
|
|
|
/*!
|
|
|
|
* \brief tps_task structure is queued to a taskprocessor
|
2008-05-03 03:40:32 +00:00
|
|
|
*
|
|
|
|
* tps_tasks are processed in FIFO order and freed by the taskprocessing
|
|
|
|
* thread after the task handler returns. The callback function that is assigned
|
2010-02-22 06:45:52 +00:00
|
|
|
* to the execute() function pointer is responsible for releasing datap resources if necessary.
|
|
|
|
*/
|
2008-05-03 03:40:32 +00:00
|
|
|
struct tps_task {
|
|
|
|
/*! \brief The execute() task callback function pointer */
|
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
|
|
|
union {
|
|
|
|
int (*execute)(void *datap);
|
|
|
|
int (*execute_local)(struct ast_taskprocessor_local *local);
|
|
|
|
} callback;
|
2008-05-03 03:40:32 +00:00
|
|
|
/*! \brief The data pointer for the task execute() function */
|
|
|
|
void *datap;
|
|
|
|
/*! \brief AST_LIST_ENTRY overhead */
|
|
|
|
AST_LIST_ENTRY(tps_task) list;
|
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
|
|
|
unsigned int wants_local:1;
|
2008-05-03 03:40:32 +00:00
|
|
|
};
|
|
|
|
|
|
|
|
/*! \brief tps_taskprocessor_stats maintain statistics for a taskprocessor. */
|
|
|
|
struct tps_taskprocessor_stats {
|
|
|
|
/*! \brief This is the maximum number of tasks queued at any one time */
|
|
|
|
unsigned long max_qsize;
|
|
|
|
/*! \brief This is the current number of tasks processed */
|
|
|
|
unsigned long _tasks_processed_count;
|
|
|
|
};
|
|
|
|
|
|
|
|
/*! \brief A ast_taskprocessor structure is a singleton by name */
|
|
|
|
struct ast_taskprocessor {
|
|
|
|
/*! \brief Friendly name of the taskprocessor */
|
2009-05-21 21:13:09 +00:00
|
|
|
const char *name;
|
2008-05-03 03:40:32 +00:00
|
|
|
/*! \brief Taskprocessor statistics */
|
|
|
|
struct tps_taskprocessor_stats *stats;
|
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
|
|
|
void *local_data;
|
2008-05-03 03:40:32 +00:00
|
|
|
/*! \brief Taskprocessor current queue size */
|
2008-06-23 15:24:05 +00:00
|
|
|
long tps_queue_size;
|
2016-06-02 21:08:19 +00:00
|
|
|
/*! \brief Taskprocessor low water clear alert level */
|
|
|
|
long tps_queue_low;
|
|
|
|
/*! \brief Taskprocessor high water alert trigger level */
|
|
|
|
long tps_queue_high;
|
2008-05-03 03:40:32 +00:00
|
|
|
/*! \brief Taskprocessor queue */
|
2008-06-23 15:24:05 +00:00
|
|
|
AST_LIST_HEAD_NOLOCK(tps_queue, tps_task) tps_queue;
|
2012-11-08 23:27:16 +00:00
|
|
|
struct ast_taskprocessor_listener *listener;
|
2014-10-03 17:47:42 +00:00
|
|
|
/*! Current thread executing the tasks */
|
|
|
|
pthread_t thread;
|
2013-02-12 21:45:59 +00:00
|
|
|
/*! Indicates if the taskprocessor is currently executing a task */
|
|
|
|
unsigned int executing:1;
|
2015-11-03 22:19:43 +00:00
|
|
|
/*! Indicates that a high water warning has been issued on this task processor */
|
|
|
|
unsigned int high_water_warned:1;
|
2016-06-02 21:08:19 +00:00
|
|
|
/*! Indicates that a high water alert is active on this taskprocessor */
|
|
|
|
unsigned int high_water_alert:1;
|
2008-05-03 03:40:32 +00:00
|
|
|
};
|
2013-01-15 20:48:45 +00:00
|
|
|
|
|
|
|
/*!
|
|
|
|
* \brief A listener for taskprocessors
|
|
|
|
*
|
|
|
|
* \since 12.0.0
|
|
|
|
*
|
|
|
|
* When a taskprocessor's state changes, the listener
|
|
|
|
* is notified of the change. This allows for tasks
|
|
|
|
* to be addressed in whatever way is appropriate for
|
|
|
|
* the module using the taskprocessor.
|
|
|
|
*/
|
|
|
|
struct ast_taskprocessor_listener {
|
|
|
|
/*! The callbacks the taskprocessor calls into to notify of state changes */
|
|
|
|
const struct ast_taskprocessor_listener_callbacks *callbacks;
|
|
|
|
/*! The taskprocessor that the listener is listening to */
|
|
|
|
struct ast_taskprocessor *tps;
|
|
|
|
/*! Data private to the listener */
|
|
|
|
void *user_data;
|
|
|
|
};
|
|
|
|
|
2008-05-03 03:40:32 +00:00
|
|
|
#define TPS_MAX_BUCKETS 7
|
|
|
|
/*! \brief tps_singletons is the astobj2 container for taskprocessor singletons */
|
|
|
|
static struct ao2_container *tps_singletons;
|
|
|
|
|
2009-04-01 20:13:28 +00:00
|
|
|
/*! \brief CLI <example>taskprocessor ping <blah></example> operation requires a ping condition */
|
2008-05-03 03:40:32 +00:00
|
|
|
static ast_cond_t cli_ping_cond;
|
2008-06-10 00:31:30 +00:00
|
|
|
|
2009-04-01 20:13:28 +00:00
|
|
|
/*! \brief CLI <example>taskprocessor ping <blah></example> operation requires a ping condition lock */
|
2008-06-10 00:31:30 +00:00
|
|
|
AST_MUTEX_DEFINE_STATIC(cli_ping_cond_lock);
|
2008-05-03 03:40:32 +00:00
|
|
|
|
|
|
|
/*! \brief The astobj2 hash callback for taskprocessors */
|
|
|
|
static int tps_hash_cb(const void *obj, const int flags);
|
|
|
|
/*! \brief The astobj2 compare callback for taskprocessors */
|
2008-11-25 01:01:49 +00:00
|
|
|
static int tps_cmp_cb(void *obj, void *arg, int flags);
|
2008-05-03 03:40:32 +00:00
|
|
|
|
2009-04-01 20:13:28 +00:00
|
|
|
/*! \brief CLI <example>taskprocessor ping <blah></example> handler function */
|
2008-05-03 03:40:32 +00:00
|
|
|
static int tps_ping_handler(void *datap);
|
|
|
|
|
|
|
|
static char *cli_tps_ping(struct ast_cli_entry *e, int cmd, struct ast_cli_args *a);
|
|
|
|
static char *cli_tps_report(struct ast_cli_entry *e, int cmd, struct ast_cli_args *a);
|
|
|
|
|
|
|
|
static struct ast_cli_entry taskprocessor_clis[] = {
|
2009-01-13 16:02:13 +00:00
|
|
|
AST_CLI_DEFINE(cli_tps_ping, "Ping a named task processor"),
|
2008-05-03 03:40:32 +00:00
|
|
|
AST_CLI_DEFINE(cli_tps_report, "List instantiated task processors and statistics"),
|
|
|
|
};
|
|
|
|
|
2012-11-08 23:27:16 +00:00
|
|
|
struct default_taskprocessor_listener_pvt {
|
|
|
|
pthread_t poll_thread;
|
|
|
|
int dead;
|
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 ast_sem sem;
|
2012-11-08 23:27:16 +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
|
|
|
static void default_listener_pvt_destroy(struct default_taskprocessor_listener_pvt *pvt)
|
2012-11-08 23:27:16 +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
|
|
|
ast_assert(pvt->dead);
|
|
|
|
ast_sem_destroy(&pvt->sem);
|
|
|
|
ast_free(pvt);
|
2012-11-08 23:27:16 +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
|
|
|
static void default_listener_pvt_dtor(struct ast_taskprocessor_listener *listener)
|
2012-11-08 23:27:16 +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
|
|
|
struct default_taskprocessor_listener_pvt *pvt = listener->user_data;
|
|
|
|
|
|
|
|
default_listener_pvt_destroy(pvt);
|
|
|
|
|
|
|
|
listener->user_data = NULL;
|
2012-11-08 23:27:16 +00:00
|
|
|
}
|
|
|
|
|
2013-01-07 22:56:37 +00:00
|
|
|
/*!
|
|
|
|
* \brief Function that processes tasks in the taskprocessor
|
|
|
|
* \internal
|
|
|
|
*/
|
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 *default_tps_processing_function(void *data)
|
2012-11-08 23:27:16 +00:00
|
|
|
{
|
|
|
|
struct ast_taskprocessor_listener *listener = data;
|
|
|
|
struct ast_taskprocessor *tps = listener->tps;
|
2013-01-15 18:40:36 +00:00
|
|
|
struct default_taskprocessor_listener_pvt *pvt = listener->user_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
|
|
|
int sem_value;
|
|
|
|
int res;
|
|
|
|
|
|
|
|
while (!pvt->dead) {
|
|
|
|
res = ast_sem_wait(&pvt->sem);
|
|
|
|
if (res != 0 && errno != EINTR) {
|
|
|
|
ast_log(LOG_ERROR, "ast_sem_wait(): %s\n",
|
|
|
|
strerror(errno));
|
|
|
|
/* Just give up */
|
|
|
|
break;
|
2012-11-08 23:27:16 +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
|
|
|
ast_taskprocessor_execute(tps);
|
2012-11-08 23:27:16 +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
|
|
|
|
|
|
|
/* No posting to a dead taskprocessor! */
|
|
|
|
res = ast_sem_getvalue(&pvt->sem, &sem_value);
|
|
|
|
ast_assert(res == 0 && sem_value == 0);
|
|
|
|
|
|
|
|
/* Free the shutdown reference (see default_listener_shutdown) */
|
|
|
|
ao2_t_ref(listener->tps, -1, "tps-shutdown");
|
|
|
|
|
2012-11-08 23:27:16 +00:00
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
2012-12-07 00:30:35 +00:00
|
|
|
static int default_listener_start(struct ast_taskprocessor_listener *listener)
|
|
|
|
{
|
2013-01-15 18:40:36 +00:00
|
|
|
struct default_taskprocessor_listener_pvt *pvt = listener->user_data;
|
2012-12-07 00:30:35 +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
|
|
|
if (ast_pthread_create(&pvt->poll_thread, NULL, default_tps_processing_function, listener)) {
|
2012-12-07 00:30:35 +00:00
|
|
|
return -1;
|
|
|
|
}
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2012-11-19 21:31:32 +00:00
|
|
|
static void default_task_pushed(struct ast_taskprocessor_listener *listener, int was_empty)
|
|
|
|
{
|
2013-01-15 18:40:36 +00:00
|
|
|
struct default_taskprocessor_listener_pvt *pvt = listener->user_data;
|
2012-11-19 21:31:32 +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
|
|
|
if (ast_sem_post(&pvt->sem) != 0) {
|
|
|
|
ast_log(LOG_ERROR, "Failed to notify of enqueued task: %s\n",
|
|
|
|
strerror(errno));
|
2012-11-19 21:31:32 +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
|
|
|
static int default_listener_die(void *data)
|
2012-11-19 21:31:32 +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
|
|
|
struct default_taskprocessor_listener_pvt *pvt = data;
|
|
|
|
pvt->dead = 1;
|
|
|
|
return 0;
|
2012-11-19 21:31:32 +00:00
|
|
|
}
|
|
|
|
|
2012-11-16 04:33:53 +00:00
|
|
|
static void default_listener_shutdown(struct ast_taskprocessor_listener *listener)
|
2012-11-09 22:28:10 +00:00
|
|
|
{
|
2013-01-15 18:40:36 +00:00
|
|
|
struct default_taskprocessor_listener_pvt *pvt = listener->user_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
|
|
|
int res;
|
|
|
|
|
|
|
|
/* Hold a reference during shutdown */
|
|
|
|
ao2_t_ref(listener->tps, +1, "tps-shutdown");
|
|
|
|
|
|
|
|
ast_taskprocessor_push(listener->tps, default_listener_die, pvt);
|
|
|
|
|
2013-10-25 23:58:32 +00:00
|
|
|
ast_assert(pvt->poll_thread != AST_PTHREADT_NULL);
|
|
|
|
|
|
|
|
if (pthread_equal(pthread_self(), pvt->poll_thread)) {
|
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 = pthread_detach(pvt->poll_thread);
|
|
|
|
if (res != 0) {
|
2013-10-25 23:58:32 +00:00
|
|
|
ast_log(LOG_ERROR, "pthread_detach(): %s\n", strerror(errno));
|
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 {
|
|
|
|
res = pthread_join(pvt->poll_thread, NULL);
|
|
|
|
if (res != 0) {
|
2013-10-25 23:58:32 +00:00
|
|
|
ast_log(LOG_ERROR, "pthread_join(): %s\n", strerror(errno));
|
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
|
|
|
}
|
|
|
|
}
|
2012-11-09 22:28:10 +00:00
|
|
|
pvt->poll_thread = AST_PTHREADT_NULL;
|
|
|
|
}
|
|
|
|
|
2012-11-09 22:49:25 +00:00
|
|
|
static const struct ast_taskprocessor_listener_callbacks default_listener_callbacks = {
|
2012-12-07 00:30:35 +00:00
|
|
|
.start = default_listener_start,
|
2012-11-08 23:27:16 +00:00
|
|
|
.task_pushed = default_task_pushed,
|
2012-11-16 04:33:53 +00:00
|
|
|
.shutdown = default_listener_shutdown,
|
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
|
|
|
.dtor = default_listener_pvt_dtor,
|
2012-11-08 23:27:16 +00:00
|
|
|
};
|
|
|
|
|
2012-12-13 16:39:40 +00:00
|
|
|
/*!
|
|
|
|
* \internal
|
|
|
|
* \brief Clean up resources on Asterisk shutdown
|
|
|
|
*/
|
2012-10-02 01:47:16 +00:00
|
|
|
static void tps_shutdown(void)
|
|
|
|
{
|
2012-12-13 16:39:40 +00:00
|
|
|
ast_cli_unregister_multiple(taskprocessor_clis, ARRAY_LEN(taskprocessor_clis));
|
2012-10-02 01:47:16 +00:00
|
|
|
ao2_t_ref(tps_singletons, -1, "Unref tps_singletons in shutdown");
|
2012-12-13 16:39:40 +00:00
|
|
|
tps_singletons = NULL;
|
2012-10-02 01:47:16 +00:00
|
|
|
}
|
|
|
|
|
2008-05-03 03:40:32 +00:00
|
|
|
/* initialize the taskprocessor container and register CLI operations */
|
|
|
|
int ast_tps_init(void)
|
|
|
|
{
|
|
|
|
if (!(tps_singletons = ao2_container_alloc(TPS_MAX_BUCKETS, tps_hash_cb, tps_cmp_cb))) {
|
|
|
|
ast_log(LOG_ERROR, "taskprocessor container failed to initialize!\n");
|
|
|
|
return -1;
|
|
|
|
}
|
2008-06-10 00:31:30 +00:00
|
|
|
|
|
|
|
ast_cond_init(&cli_ping_cond, NULL);
|
|
|
|
|
2008-05-03 03:40:32 +00:00
|
|
|
ast_cli_register_multiple(taskprocessor_clis, ARRAY_LEN(taskprocessor_clis));
|
2012-10-02 01:47:16 +00:00
|
|
|
|
2015-03-26 22:24:26 +00:00
|
|
|
ast_register_cleanup(tps_shutdown);
|
2012-10-02 01:47:16 +00:00
|
|
|
|
2008-05-03 03:40:32 +00:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* allocate resources for the task */
|
|
|
|
static struct tps_task *tps_task_alloc(int (*task_exe)(void *datap), void *datap)
|
|
|
|
{
|
|
|
|
struct tps_task *t;
|
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 (!task_exe) {
|
|
|
|
ast_log(LOG_ERROR, "task_exe is NULL!\n");
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
t = ast_calloc(1, sizeof(*t));
|
|
|
|
if (!t) {
|
|
|
|
ast_log(LOG_ERROR, "failed to allocate task!\n");
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
t->callback.execute = task_exe;
|
|
|
|
t->datap = datap;
|
|
|
|
|
|
|
|
return t;
|
|
|
|
}
|
|
|
|
|
|
|
|
static struct tps_task *tps_task_alloc_local(int (*task_exe)(struct ast_taskprocessor_local *local), void *datap)
|
|
|
|
{
|
|
|
|
struct tps_task *t;
|
|
|
|
if (!task_exe) {
|
|
|
|
ast_log(LOG_ERROR, "task_exe is NULL!\n");
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
t = ast_calloc(1, sizeof(*t));
|
|
|
|
if (!t) {
|
|
|
|
ast_log(LOG_ERROR, "failed to allocate task!\n");
|
|
|
|
return NULL;
|
2008-05-03 03:40:32 +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
|
|
|
|
|
|
|
t->callback.execute_local = task_exe;
|
|
|
|
t->datap = datap;
|
|
|
|
t->wants_local = 1;
|
|
|
|
|
2008-05-03 03:40:32 +00:00
|
|
|
return t;
|
|
|
|
}
|
|
|
|
|
2012-03-22 19:51:16 +00:00
|
|
|
/* release task resources */
|
2008-05-03 03:40:32 +00:00
|
|
|
static void *tps_task_free(struct tps_task *task)
|
|
|
|
{
|
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_free(task);
|
2008-05-03 03:40:32 +00:00
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* taskprocessor tab completion */
|
2016-01-19 01:43:41 +00:00
|
|
|
static char *tps_taskprocessor_tab_complete(struct ast_cli_args *a)
|
2008-05-03 03:40:32 +00:00
|
|
|
{
|
|
|
|
int tklen;
|
|
|
|
int wordnum = 0;
|
2016-01-19 01:43:41 +00:00
|
|
|
struct ast_taskprocessor *p;
|
2008-05-03 03:40:32 +00:00
|
|
|
char *name = NULL;
|
|
|
|
struct ao2_iterator i;
|
|
|
|
|
2008-08-06 21:22:56 +00:00
|
|
|
if (a->pos != 3)
|
2008-05-03 03:40:32 +00:00
|
|
|
return NULL;
|
|
|
|
|
|
|
|
tklen = strlen(a->word);
|
|
|
|
i = ao2_iterator_init(tps_singletons, 0);
|
|
|
|
while ((p = ao2_iterator_next(&i))) {
|
|
|
|
if (!strncasecmp(a->word, p->name, tklen) && ++wordnum > a->n) {
|
|
|
|
name = ast_strdup(p->name);
|
2016-01-19 01:43:41 +00:00
|
|
|
ast_taskprocessor_unreference(p);
|
2008-05-03 03:40:32 +00:00
|
|
|
break;
|
|
|
|
}
|
2016-01-19 01:43:41 +00:00
|
|
|
ast_taskprocessor_unreference(p);
|
2008-05-03 03:40:32 +00:00
|
|
|
}
|
2012-01-27 18:47:16 +00:00
|
|
|
ao2_iterator_destroy(&i);
|
2008-05-03 03:40:32 +00:00
|
|
|
return name;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* ping task handling function */
|
|
|
|
static int tps_ping_handler(void *datap)
|
|
|
|
{
|
|
|
|
ast_mutex_lock(&cli_ping_cond_lock);
|
|
|
|
ast_cond_signal(&cli_ping_cond);
|
|
|
|
ast_mutex_unlock(&cli_ping_cond_lock);
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* ping the specified taskprocessor and display the ping time on the CLI */
|
|
|
|
static char *cli_tps_ping(struct ast_cli_entry *e, int cmd, struct ast_cli_args *a)
|
|
|
|
{
|
|
|
|
struct timeval begin, end, delta;
|
2009-05-21 21:13:09 +00:00
|
|
|
const char *name;
|
2008-08-10 20:57:25 +00:00
|
|
|
struct timeval when;
|
2008-05-03 03:40:32 +00:00
|
|
|
struct timespec ts;
|
2016-01-19 01:43:41 +00:00
|
|
|
struct ast_taskprocessor *tps;
|
2008-05-03 03:40:32 +00:00
|
|
|
|
|
|
|
switch (cmd) {
|
|
|
|
case CLI_INIT:
|
2008-09-28 23:32:14 +00:00
|
|
|
e->command = "core ping taskprocessor";
|
2012-03-22 19:51:16 +00:00
|
|
|
e->usage =
|
2008-09-28 23:32:14 +00:00
|
|
|
"Usage: core ping taskprocessor <taskprocessor>\n"
|
2008-08-06 21:22:56 +00:00
|
|
|
" Displays the time required for a task to be processed\n";
|
2008-05-03 03:40:32 +00:00
|
|
|
return NULL;
|
|
|
|
case CLI_GENERATE:
|
2016-01-19 01:43:41 +00:00
|
|
|
return tps_taskprocessor_tab_complete(a);
|
2008-05-03 03:40:32 +00:00
|
|
|
}
|
|
|
|
|
2008-08-06 21:22:56 +00:00
|
|
|
if (a->argc != 4)
|
2008-05-03 03:40:32 +00:00
|
|
|
return CLI_SHOWUSAGE;
|
|
|
|
|
2008-08-06 21:22:56 +00:00
|
|
|
name = a->argv[3];
|
2008-05-03 03:40:32 +00:00
|
|
|
if (!(tps = ast_taskprocessor_get(name, TPS_REF_IF_EXISTS))) {
|
|
|
|
ast_cli(a->fd, "\nping failed: %s not found\n\n", name);
|
|
|
|
return CLI_SUCCESS;
|
|
|
|
}
|
|
|
|
ast_cli(a->fd, "\npinging %s ...", name);
|
2016-01-19 01:44:24 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* Wait up to 5 seconds for a ping reply.
|
|
|
|
*
|
|
|
|
* On a very busy system it could take awhile to get a
|
|
|
|
* ping response from some taskprocessors.
|
|
|
|
*/
|
|
|
|
begin = ast_tvnow();
|
|
|
|
when = ast_tvadd(begin, ast_samp2tv(5000, 1000));
|
2008-08-10 20:57:25 +00:00
|
|
|
ts.tv_sec = when.tv_sec;
|
|
|
|
ts.tv_nsec = when.tv_usec * 1000;
|
2016-01-19 01:44:24 +00:00
|
|
|
|
2008-05-03 03:40:32 +00:00
|
|
|
ast_mutex_lock(&cli_ping_cond_lock);
|
|
|
|
if (ast_taskprocessor_push(tps, tps_ping_handler, 0) < 0) {
|
2016-01-05 22:54:06 +00:00
|
|
|
ast_mutex_unlock(&cli_ping_cond_lock);
|
2008-05-03 03:40:32 +00:00
|
|
|
ast_cli(a->fd, "\nping failed: could not push task to %s\n\n", name);
|
2016-01-19 01:43:41 +00:00
|
|
|
ast_taskprocessor_unreference(tps);
|
2008-05-03 03:40:32 +00:00
|
|
|
return CLI_FAILURE;
|
|
|
|
}
|
|
|
|
ast_cond_timedwait(&cli_ping_cond, &cli_ping_cond_lock, &ts);
|
|
|
|
ast_mutex_unlock(&cli_ping_cond_lock);
|
2016-01-19 01:44:24 +00:00
|
|
|
|
2008-05-03 03:40:32 +00:00
|
|
|
end = ast_tvnow();
|
|
|
|
delta = ast_tvsub(end, begin);
|
2008-05-15 10:56:29 +00:00
|
|
|
ast_cli(a->fd, "\n\t%24s ping time: %.1ld.%.6ld sec\n\n", name, (long)delta.tv_sec, (long int)delta.tv_usec);
|
2016-01-19 01:43:41 +00:00
|
|
|
ast_taskprocessor_unreference(tps);
|
2012-03-22 19:51:16 +00:00
|
|
|
return CLI_SUCCESS;
|
2008-05-03 03:40:32 +00:00
|
|
|
}
|
|
|
|
|
2016-01-08 02:44:18 +00:00
|
|
|
/*!
|
|
|
|
* \internal
|
|
|
|
* \brief Taskprocessor ao2 container sort function.
|
|
|
|
* \since 13.8.0
|
|
|
|
*
|
|
|
|
* \param obj_left pointer to the (user-defined part) of an object.
|
|
|
|
* \param obj_right pointer to the (user-defined part) of an object.
|
|
|
|
* \param flags flags from ao2_callback()
|
|
|
|
* OBJ_SEARCH_OBJECT - if set, 'obj_right', is an object.
|
|
|
|
* OBJ_SEARCH_KEY - if set, 'obj_right', is a search key item that is not an object.
|
|
|
|
* OBJ_SEARCH_PARTIAL_KEY - if set, 'obj_right', is a partial search key item that is not an object.
|
|
|
|
*
|
|
|
|
* \retval <0 if obj_left < obj_right
|
|
|
|
* \retval =0 if obj_left == obj_right
|
|
|
|
* \retval >0 if obj_left > obj_right
|
|
|
|
*/
|
|
|
|
static int tps_sort_cb(const void *obj_left, const void *obj_right, int flags)
|
|
|
|
{
|
|
|
|
const struct ast_taskprocessor *tps_left = obj_left;
|
|
|
|
const struct ast_taskprocessor *tps_right = obj_right;
|
|
|
|
const char *right_key = obj_right;
|
|
|
|
int cmp;
|
|
|
|
|
|
|
|
switch (flags & OBJ_SEARCH_MASK) {
|
|
|
|
default:
|
|
|
|
case OBJ_SEARCH_OBJECT:
|
|
|
|
right_key = tps_right->name;
|
|
|
|
/* Fall through */
|
|
|
|
case OBJ_SEARCH_KEY:
|
|
|
|
cmp = strcasecmp(tps_left->name, right_key);
|
|
|
|
break;
|
|
|
|
case OBJ_SEARCH_PARTIAL_KEY:
|
|
|
|
cmp = strncasecmp(tps_left->name, right_key, strlen(right_key));
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
return cmp;
|
|
|
|
}
|
|
|
|
|
2008-05-03 03:40:32 +00:00
|
|
|
static char *cli_tps_report(struct ast_cli_entry *e, int cmd, struct ast_cli_args *a)
|
|
|
|
{
|
|
|
|
char name[256];
|
|
|
|
int tcount;
|
|
|
|
unsigned long qsize;
|
|
|
|
unsigned long maxqsize;
|
|
|
|
unsigned long processed;
|
2016-01-08 02:44:18 +00:00
|
|
|
struct ao2_container *sorted_tps;
|
|
|
|
struct ast_taskprocessor *tps;
|
2016-01-06 23:19:19 +00:00
|
|
|
struct ao2_iterator iter;
|
2016-06-02 21:08:19 +00:00
|
|
|
#define FMT_HEADERS "%-45s %10s %10s %10s %10s %10s\n"
|
|
|
|
#define FMT_FIELDS "%-45s %10lu %10lu %10lu %10lu %10lu\n"
|
2008-05-03 03:40:32 +00:00
|
|
|
|
|
|
|
switch (cmd) {
|
|
|
|
case CLI_INIT:
|
2008-08-06 21:22:56 +00:00
|
|
|
e->command = "core show taskprocessors";
|
2012-03-22 19:51:16 +00:00
|
|
|
e->usage =
|
2008-08-06 21:22:56 +00:00
|
|
|
"Usage: core show taskprocessors\n"
|
2008-05-03 03:40:32 +00:00
|
|
|
" Shows a list of instantiated task processors and their statistics\n";
|
|
|
|
return NULL;
|
|
|
|
case CLI_GENERATE:
|
2012-03-22 19:51:16 +00:00
|
|
|
return NULL;
|
2008-05-03 03:40:32 +00:00
|
|
|
}
|
|
|
|
|
2016-01-08 02:44:18 +00:00
|
|
|
if (a->argc != e->args) {
|
2008-05-03 03:40:32 +00:00
|
|
|
return CLI_SHOWUSAGE;
|
2016-01-08 02:44:18 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
sorted_tps = ao2_container_alloc_rbtree(AO2_ALLOC_OPT_LOCK_NOLOCK, 0, tps_sort_cb,
|
|
|
|
NULL);
|
|
|
|
if (!sorted_tps
|
|
|
|
|| ao2_container_dup(sorted_tps, tps_singletons, 0)) {
|
|
|
|
ao2_cleanup(sorted_tps);
|
|
|
|
return CLI_FAILURE;
|
|
|
|
}
|
2008-05-03 03:40:32 +00:00
|
|
|
|
2016-06-02 21:08:19 +00:00
|
|
|
ast_cli(a->fd, "\n" FMT_HEADERS, "Processor", "Processed", "In Queue", "Max Depth", "Low water", "High water");
|
2016-01-06 23:19:19 +00:00
|
|
|
tcount = 0;
|
2016-01-08 02:44:18 +00:00
|
|
|
iter = ao2_iterator_init(sorted_tps, AO2_ITERATOR_UNLINK);
|
|
|
|
while ((tps = ao2_iterator_next(&iter))) {
|
|
|
|
ast_copy_string(name, tps->name, sizeof(name));
|
|
|
|
qsize = tps->tps_queue_size;
|
|
|
|
if (tps->stats) {
|
|
|
|
maxqsize = tps->stats->max_qsize;
|
|
|
|
processed = tps->stats->_tasks_processed_count;
|
2016-01-06 23:19:19 +00:00
|
|
|
} else {
|
|
|
|
maxqsize = 0;
|
|
|
|
processed = 0;
|
|
|
|
}
|
2016-06-02 21:08:19 +00:00
|
|
|
ast_cli(a->fd, FMT_FIELDS, name, processed, qsize, maxqsize,
|
|
|
|
tps->tps_queue_low, tps->tps_queue_high);
|
2016-01-08 02:44:18 +00:00
|
|
|
ast_taskprocessor_unreference(tps);
|
2016-01-06 23:19:19 +00:00
|
|
|
++tcount;
|
2008-05-03 03:40:32 +00:00
|
|
|
}
|
2016-01-06 23:19:19 +00:00
|
|
|
ao2_iterator_destroy(&iter);
|
|
|
|
ast_cli(a->fd, "\n%d taskprocessors\n\n", tcount);
|
2016-01-08 02:44:18 +00:00
|
|
|
ao2_ref(sorted_tps, -1);
|
2012-03-22 19:51:16 +00:00
|
|
|
return CLI_SUCCESS;
|
2008-05-03 03:40:32 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/* hash callback for astobj2 */
|
|
|
|
static int tps_hash_cb(const void *obj, const int flags)
|
|
|
|
{
|
|
|
|
const struct ast_taskprocessor *tps = obj;
|
2012-11-08 23:27:16 +00:00
|
|
|
const char *name = flags & OBJ_KEY ? obj : tps->name;
|
2008-05-03 03:40:32 +00:00
|
|
|
|
2012-11-08 23:27:16 +00:00
|
|
|
return ast_str_case_hash(name);
|
2008-05-03 03:40:32 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/* compare callback for astobj2 */
|
2008-11-25 01:01:49 +00:00
|
|
|
static int tps_cmp_cb(void *obj, void *arg, int flags)
|
2008-05-03 03:40:32 +00:00
|
|
|
{
|
|
|
|
struct ast_taskprocessor *lhs = obj, *rhs = arg;
|
2012-11-08 23:27:16 +00:00
|
|
|
const char *rhsname = flags & OBJ_KEY ? arg : rhs->name;
|
2008-05-03 03:40:32 +00:00
|
|
|
|
2012-11-08 23:27:16 +00:00
|
|
|
return !strcasecmp(lhs->name, rhsname) ? CMP_MATCH | CMP_STOP : 0;
|
2008-05-03 03:40:32 +00:00
|
|
|
}
|
|
|
|
|
2016-06-02 21:08:19 +00:00
|
|
|
/*! Count of the number of taskprocessors in high water alert. */
|
|
|
|
static unsigned int tps_alert_count;
|
|
|
|
|
|
|
|
/*! Access protection for tps_alert_count */
|
|
|
|
AST_RWLOCK_DEFINE_STATIC(tps_alert_lock);
|
|
|
|
|
|
|
|
/*!
|
|
|
|
* \internal
|
|
|
|
* \brief Add a delta to tps_alert_count with protection.
|
|
|
|
* \since 13.10.0
|
|
|
|
*
|
|
|
|
* \param tps Taskprocessor updating queue water mark alert trigger.
|
|
|
|
* \param delta The amount to add to tps_alert_count.
|
|
|
|
*
|
|
|
|
* \return Nothing
|
|
|
|
*/
|
|
|
|
static void tps_alert_add(struct ast_taskprocessor *tps, int delta)
|
|
|
|
{
|
|
|
|
unsigned int old;
|
|
|
|
|
|
|
|
ast_rwlock_wrlock(&tps_alert_lock);
|
|
|
|
old = tps_alert_count;
|
|
|
|
tps_alert_count += delta;
|
|
|
|
if (DEBUG_ATLEAST(3)
|
|
|
|
/* and tps_alert_count becomes zero or non-zero */
|
|
|
|
&& !old != !tps_alert_count) {
|
|
|
|
ast_log(LOG_DEBUG, "Taskprocessor '%s' %s the high water alert.\n",
|
|
|
|
tps->name, tps_alert_count ? "triggered" : "cleared");
|
|
|
|
}
|
|
|
|
ast_rwlock_unlock(&tps_alert_lock);
|
|
|
|
}
|
|
|
|
|
|
|
|
unsigned int ast_taskprocessor_alert_get(void)
|
|
|
|
{
|
|
|
|
unsigned int count;
|
|
|
|
|
|
|
|
ast_rwlock_rdlock(&tps_alert_lock);
|
|
|
|
count = tps_alert_count;
|
|
|
|
ast_rwlock_unlock(&tps_alert_lock);
|
|
|
|
|
|
|
|
return count;
|
|
|
|
}
|
|
|
|
|
|
|
|
int ast_taskprocessor_alert_set_levels(struct ast_taskprocessor *tps, long low_water, long high_water)
|
|
|
|
{
|
|
|
|
if (!tps || high_water < 0 || high_water < low_water) {
|
|
|
|
return -1;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (low_water < 0) {
|
|
|
|
/* Set low water level to 90% of high water level */
|
|
|
|
low_water = (high_water * 9) / 10;
|
|
|
|
}
|
|
|
|
|
|
|
|
ao2_lock(tps);
|
|
|
|
|
|
|
|
tps->tps_queue_low = low_water;
|
|
|
|
tps->tps_queue_high = high_water;
|
|
|
|
|
|
|
|
if (tps->high_water_alert) {
|
|
|
|
if (!tps->tps_queue_size || tps->tps_queue_size < low_water) {
|
|
|
|
/* Update water mark alert immediately */
|
|
|
|
tps->high_water_alert = 0;
|
|
|
|
tps_alert_add(tps, -1);
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
if (high_water <= tps->tps_queue_size) {
|
|
|
|
/* Update water mark alert immediately */
|
|
|
|
tps->high_water_alert = 1;
|
|
|
|
tps_alert_add(tps, +1);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
ao2_unlock(tps);
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2008-05-03 03:40:32 +00:00
|
|
|
/* destroy the taskprocessor */
|
2016-06-02 21:08:19 +00:00
|
|
|
static void tps_taskprocessor_dtor(void *tps)
|
2008-05-03 03:40:32 +00:00
|
|
|
{
|
|
|
|
struct ast_taskprocessor *t = tps;
|
2013-01-07 23:11:41 +00:00
|
|
|
struct tps_task *task;
|
2011-02-04 16:55:39 +00:00
|
|
|
|
2016-06-02 21:08:19 +00:00
|
|
|
while ((task = AST_LIST_REMOVE_HEAD(&t->tps_queue, list))) {
|
|
|
|
tps_task_free(task);
|
2008-05-03 03:40:32 +00:00
|
|
|
}
|
2016-06-02 21:08:19 +00:00
|
|
|
t->tps_queue_size = 0;
|
|
|
|
|
|
|
|
if (t->high_water_alert) {
|
|
|
|
t->high_water_alert = 0;
|
|
|
|
tps_alert_add(t, -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
|
|
|
ast_free(t->stats);
|
|
|
|
t->stats = NULL;
|
2009-05-21 21:13:09 +00:00
|
|
|
ast_free((char *) t->name);
|
2016-06-02 21:08:19 +00:00
|
|
|
t->name = NULL;
|
|
|
|
ao2_cleanup(t->listener);
|
|
|
|
t->listener = NULL;
|
2008-05-03 03:40:32 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/* pop the front task and return it */
|
|
|
|
static struct tps_task *tps_taskprocessor_pop(struct ast_taskprocessor *tps)
|
|
|
|
{
|
|
|
|
struct tps_task *task;
|
|
|
|
|
2008-06-23 15:24:05 +00:00
|
|
|
if ((task = AST_LIST_REMOVE_HEAD(&tps->tps_queue, list))) {
|
2016-06-02 21:08:19 +00:00
|
|
|
--tps->tps_queue_size;
|
|
|
|
if (tps->high_water_alert && tps->tps_queue_size <= tps->tps_queue_low) {
|
|
|
|
tps->high_water_alert = 0;
|
|
|
|
tps_alert_add(tps, -1);
|
|
|
|
}
|
2008-05-03 03:40:32 +00:00
|
|
|
}
|
|
|
|
return task;
|
|
|
|
}
|
|
|
|
|
2015-11-11 23:11:53 +00:00
|
|
|
long ast_taskprocessor_size(struct ast_taskprocessor *tps)
|
2008-05-03 03:40:32 +00:00
|
|
|
{
|
2008-06-23 15:24:05 +00:00
|
|
|
return (tps) ? tps->tps_queue_size : -1;
|
2008-05-03 03:40:32 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/* taskprocessor name accessor */
|
|
|
|
const char *ast_taskprocessor_name(struct ast_taskprocessor *tps)
|
|
|
|
{
|
|
|
|
if (!tps) {
|
|
|
|
ast_log(LOG_ERROR, "no taskprocessor specified!\n");
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
return tps->name;
|
|
|
|
}
|
|
|
|
|
2012-11-19 21:31:32 +00:00
|
|
|
static void listener_shutdown(struct ast_taskprocessor_listener *listener)
|
|
|
|
{
|
|
|
|
listener->callbacks->shutdown(listener);
|
|
|
|
ao2_ref(listener->tps, -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
|
|
|
static void taskprocessor_listener_dtor(void *obj)
|
|
|
|
{
|
|
|
|
struct ast_taskprocessor_listener *listener = obj;
|
|
|
|
|
|
|
|
if (listener->callbacks->dtor) {
|
|
|
|
listener->callbacks->dtor(listener);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-01-15 18:40:36 +00:00
|
|
|
struct ast_taskprocessor_listener *ast_taskprocessor_listener_alloc(const struct ast_taskprocessor_listener_callbacks *callbacks, void *user_data)
|
2012-11-08 23:27:16 +00:00
|
|
|
{
|
2013-10-25 23:58:32 +00:00
|
|
|
struct ast_taskprocessor_listener *listener;
|
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
|
|
|
|
|
|
|
listener = ao2_alloc(sizeof(*listener), taskprocessor_listener_dtor);
|
2012-11-08 23:27:16 +00:00
|
|
|
if (!listener) {
|
|
|
|
return NULL;
|
|
|
|
}
|
2012-11-09 22:28:10 +00:00
|
|
|
listener->callbacks = callbacks;
|
2013-01-15 18:40:36 +00:00
|
|
|
listener->user_data = user_data;
|
2012-11-09 22:28:10 +00:00
|
|
|
|
2012-11-08 23:27:16 +00:00
|
|
|
return listener;
|
|
|
|
}
|
|
|
|
|
2013-01-15 20:48:45 +00:00
|
|
|
struct ast_taskprocessor *ast_taskprocessor_listener_get_tps(const struct ast_taskprocessor_listener *listener)
|
|
|
|
{
|
2013-01-15 21:15:04 +00:00
|
|
|
ao2_ref(listener->tps, +1);
|
2013-01-15 20:48:45 +00:00
|
|
|
return listener->tps;
|
|
|
|
}
|
|
|
|
|
|
|
|
void *ast_taskprocessor_listener_get_user_data(const struct ast_taskprocessor_listener *listener)
|
|
|
|
{
|
|
|
|
return listener->user_data;
|
|
|
|
}
|
|
|
|
|
2013-01-15 18:40:36 +00:00
|
|
|
static void *default_listener_pvt_alloc(void)
|
|
|
|
{
|
|
|
|
struct default_taskprocessor_listener_pvt *pvt;
|
|
|
|
|
|
|
|
pvt = ast_calloc(1, sizeof(*pvt));
|
|
|
|
if (!pvt) {
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
pvt->poll_thread = AST_PTHREADT_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
|
|
|
if (ast_sem_init(&pvt->sem, 0, 0) != 0) {
|
|
|
|
ast_log(LOG_ERROR, "ast_sem_init(): %s\n", strerror(errno));
|
|
|
|
ast_free(pvt);
|
|
|
|
return NULL;
|
|
|
|
}
|
2013-01-15 18:40:36 +00:00
|
|
|
return pvt;
|
|
|
|
}
|
|
|
|
|
2013-01-15 20:15:00 +00:00
|
|
|
static struct ast_taskprocessor *__allocate_taskprocessor(const char *name, struct ast_taskprocessor_listener *listener)
|
|
|
|
{
|
2016-06-02 21:08:19 +00:00
|
|
|
struct ast_taskprocessor *p;
|
2013-01-15 20:15:00 +00:00
|
|
|
|
2016-06-02 21:08:19 +00:00
|
|
|
p = ao2_alloc(sizeof(*p), tps_taskprocessor_dtor);
|
2013-01-15 20:15:00 +00:00
|
|
|
if (!p) {
|
|
|
|
ast_log(LOG_WARNING, "failed to create taskprocessor '%s'\n", name);
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
2016-06-02 21:08:19 +00:00
|
|
|
/* Set default congestion water level alert triggers. */
|
|
|
|
p->tps_queue_low = (AST_TASKPROCESSOR_HIGH_WATER_LEVEL * 9) / 10;
|
|
|
|
p->tps_queue_high = AST_TASKPROCESSOR_HIGH_WATER_LEVEL;
|
|
|
|
|
|
|
|
p->stats = ast_calloc(1, sizeof(*p->stats));
|
|
|
|
p->name = ast_strdup(name);
|
|
|
|
if (!p->stats || !p->name) {
|
|
|
|
ao2_ref(p, -1);
|
2013-01-15 20:15:00 +00:00
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
ao2_ref(listener, +1);
|
|
|
|
p->listener = listener;
|
|
|
|
|
2014-10-03 17:47:42 +00:00
|
|
|
p->thread = AST_PTHREADT_NULL;
|
|
|
|
|
2013-01-15 20:15:00 +00:00
|
|
|
ao2_ref(p, +1);
|
|
|
|
listener->tps = p;
|
|
|
|
|
|
|
|
if (!(ao2_link(tps_singletons, p))) {
|
|
|
|
ast_log(LOG_ERROR, "Failed to add taskprocessor '%s' to container\n", p->name);
|
2016-01-19 01:43:41 +00:00
|
|
|
listener->tps = NULL;
|
2016-06-02 21:08:19 +00:00
|
|
|
ao2_ref(p, -2);
|
2013-01-15 20:15:00 +00:00
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (p->listener->callbacks->start(p->listener)) {
|
2016-06-02 21:08:19 +00:00
|
|
|
ast_log(LOG_ERROR, "Unable to start taskprocessor listener for taskprocessor %s\n",
|
|
|
|
p->name);
|
2013-01-15 20:15:00 +00:00
|
|
|
ast_taskprocessor_unreference(p);
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
return p;
|
|
|
|
}
|
|
|
|
|
2008-05-03 03:40:32 +00:00
|
|
|
/* Provide a reference to a taskprocessor. Create the taskprocessor if necessary, but don't
|
2012-03-22 19:51:16 +00:00
|
|
|
* create the taskprocessor if we were told via ast_tps_options to return a reference only
|
2008-05-03 03:40:32 +00:00
|
|
|
* if it already exists */
|
2009-05-21 21:13:09 +00:00
|
|
|
struct ast_taskprocessor *ast_taskprocessor_get(const char *name, enum ast_tps_options create)
|
2008-05-03 03:40:32 +00:00
|
|
|
{
|
2012-11-08 23:27:16 +00:00
|
|
|
struct ast_taskprocessor *p;
|
|
|
|
struct ast_taskprocessor_listener *listener;
|
2013-01-15 18:40:36 +00:00
|
|
|
struct default_taskprocessor_listener_pvt *pvt;
|
2012-03-22 19:51:16 +00:00
|
|
|
|
2008-05-03 03:40:32 +00:00
|
|
|
if (ast_strlen_zero(name)) {
|
|
|
|
ast_log(LOG_ERROR, "requesting a nameless taskprocessor!!!\n");
|
|
|
|
return NULL;
|
|
|
|
}
|
2012-11-08 23:27:16 +00:00
|
|
|
p = ao2_find(tps_singletons, name, OBJ_KEY);
|
2008-05-03 03:40:32 +00:00
|
|
|
if (p) {
|
|
|
|
return p;
|
|
|
|
}
|
|
|
|
if (create & TPS_REF_IF_EXISTS) {
|
|
|
|
/* calling function does not want a new taskprocessor to be created if it doesn't already exist */
|
|
|
|
return NULL;
|
|
|
|
}
|
2012-11-08 23:27:16 +00:00
|
|
|
/* Create a new taskprocessor. Start by creating a default listener */
|
2013-01-15 18:40:36 +00:00
|
|
|
pvt = default_listener_pvt_alloc();
|
|
|
|
if (!pvt) {
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
listener = ast_taskprocessor_listener_alloc(&default_listener_callbacks, pvt);
|
2012-11-09 22:28:10 +00:00
|
|
|
if (!listener) {
|
2013-01-15 18:40:36 +00:00
|
|
|
default_listener_pvt_destroy(pvt);
|
2012-11-09 22:28:10 +00:00
|
|
|
return NULL;
|
|
|
|
}
|
2012-11-08 23:27:16 +00:00
|
|
|
|
2013-01-15 20:15:00 +00:00
|
|
|
p = __allocate_taskprocessor(name, listener);
|
2012-11-09 22:28:10 +00:00
|
|
|
|
2012-11-08 23:27:16 +00:00
|
|
|
ao2_ref(listener, -1);
|
|
|
|
return p;
|
|
|
|
}
|
|
|
|
|
|
|
|
struct ast_taskprocessor *ast_taskprocessor_create_with_listener(const char *name, struct ast_taskprocessor_listener *listener)
|
|
|
|
{
|
2013-01-15 20:15:00 +00:00
|
|
|
struct ast_taskprocessor *p = ao2_find(tps_singletons, name, OBJ_KEY);
|
2012-11-08 23:27:16 +00:00
|
|
|
|
2013-01-15 20:15:00 +00:00
|
|
|
if (p) {
|
2012-12-07 00:30:35 +00:00
|
|
|
ast_taskprocessor_unreference(p);
|
|
|
|
return NULL;
|
|
|
|
}
|
2013-01-15 20:15:00 +00:00
|
|
|
return __allocate_taskprocessor(name, listener);
|
2008-05-03 03:40:32 +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
|
|
|
void ast_taskprocessor_set_local(struct ast_taskprocessor *tps,
|
|
|
|
void *local_data)
|
|
|
|
{
|
|
|
|
SCOPED_AO2LOCK(lock, tps);
|
|
|
|
tps->local_data = local_data;
|
|
|
|
}
|
|
|
|
|
2008-05-03 03:40:32 +00:00
|
|
|
/* decrement the taskprocessor reference count and unlink from the container if necessary */
|
|
|
|
void *ast_taskprocessor_unreference(struct ast_taskprocessor *tps)
|
|
|
|
{
|
2012-11-08 23:27:16 +00:00
|
|
|
if (!tps) {
|
|
|
|
return NULL;
|
2008-05-03 03:40:32 +00:00
|
|
|
}
|
2012-11-08 23:27:16 +00:00
|
|
|
|
2015-08-29 15:36:35 +00:00
|
|
|
/* To prevent another thread from finding and getting a reference to this
|
|
|
|
* taskprocessor we hold the singletons lock. If we didn't do this then
|
|
|
|
* they may acquire it and find that the listener has been shut down.
|
|
|
|
*/
|
|
|
|
ao2_lock(tps_singletons);
|
|
|
|
|
2012-11-08 23:27:16 +00:00
|
|
|
if (ao2_ref(tps, -1) > 3) {
|
2015-08-29 15:36:35 +00:00
|
|
|
ao2_unlock(tps_singletons);
|
2012-11-08 23:27:16 +00:00
|
|
|
return NULL;
|
|
|
|
}
|
2015-08-29 15:36:35 +00:00
|
|
|
|
2012-11-08 23:27:16 +00:00
|
|
|
/* If we're down to 3 references, then those must be:
|
|
|
|
* 1. The reference we just got rid of
|
|
|
|
* 2. The container
|
|
|
|
* 3. The listener
|
|
|
|
*/
|
2015-08-29 15:36:35 +00:00
|
|
|
ao2_unlink_flags(tps_singletons, tps, OBJ_NOLOCK);
|
|
|
|
ao2_unlock(tps_singletons);
|
|
|
|
|
2012-12-07 19:05:30 +00:00
|
|
|
listener_shutdown(tps->listener);
|
2008-05-03 03:40:32 +00:00
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
2012-03-22 19:51:16 +00:00
|
|
|
/* push the task into the taskprocessor queue */
|
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 taskprocessor_push(struct ast_taskprocessor *tps, struct tps_task *t)
|
2008-05-03 03:40:32 +00:00
|
|
|
{
|
2012-11-08 23:27:16 +00:00
|
|
|
int previous_size;
|
2013-02-12 21:45:59 +00:00
|
|
|
int was_empty;
|
2008-05-03 03:40:32 +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
|
|
|
if (!tps) {
|
|
|
|
ast_log(LOG_ERROR, "tps is NULL!\n");
|
2008-05-03 03:40:32 +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
|
|
|
|
|
|
|
if (!t) {
|
|
|
|
ast_log(LOG_ERROR, "t is NULL!\n");
|
2008-05-03 03:40:32 +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
|
|
|
|
2012-11-08 23:35:10 +00:00
|
|
|
ao2_lock(tps);
|
2008-06-23 15:24:05 +00:00
|
|
|
AST_LIST_INSERT_TAIL(&tps->tps_queue, t, list);
|
2012-11-08 23:27:16 +00:00
|
|
|
previous_size = tps->tps_queue_size++;
|
2015-11-03 22:19:43 +00:00
|
|
|
|
2016-06-02 21:08:19 +00:00
|
|
|
if (previous_size >= tps->tps_queue_high) {
|
|
|
|
if (!tps->high_water_warned) {
|
|
|
|
tps->high_water_warned = 1;
|
|
|
|
ast_log(LOG_WARNING, "The '%s' task processor queue reached %d scheduled tasks.\n",
|
|
|
|
tps->name, previous_size);
|
|
|
|
}
|
|
|
|
if (!tps->high_water_alert) {
|
|
|
|
tps->high_water_alert = 1;
|
|
|
|
tps_alert_add(tps, +1);
|
|
|
|
}
|
2015-11-03 22:19:43 +00:00
|
|
|
}
|
|
|
|
|
2013-02-12 21:45:59 +00:00
|
|
|
/* The currently executing task counts as still in queue */
|
|
|
|
was_empty = tps->executing ? 0 : previous_size == 0;
|
2012-11-08 23:35:10 +00:00
|
|
|
ao2_unlock(tps);
|
2013-02-12 21:45:59 +00:00
|
|
|
tps->listener->callbacks->task_pushed(tps->listener, was_empty);
|
2008-05-03 03:40:32 +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
|
|
|
int ast_taskprocessor_push(struct ast_taskprocessor *tps, int (*task_exe)(void *datap), void *datap)
|
|
|
|
{
|
|
|
|
return taskprocessor_push(tps, tps_task_alloc(task_exe, datap));
|
|
|
|
}
|
|
|
|
|
|
|
|
int ast_taskprocessor_push_local(struct ast_taskprocessor *tps, int (*task_exe)(struct ast_taskprocessor_local *datap), void *datap)
|
|
|
|
{
|
|
|
|
return taskprocessor_push(tps, tps_task_alloc_local(task_exe, datap));
|
|
|
|
}
|
|
|
|
|
2012-11-08 23:27:16 +00:00
|
|
|
int ast_taskprocessor_execute(struct ast_taskprocessor *tps)
|
|
|
|
{
|
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 ast_taskprocessor_local local;
|
2012-11-08 23:27:16 +00:00
|
|
|
struct tps_task *t;
|
2015-11-11 23:11:53 +00:00
|
|
|
long size;
|
2012-11-19 22:34:27 +00:00
|
|
|
|
2013-02-12 21:45:59 +00:00
|
|
|
ao2_lock(tps);
|
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
|
|
|
t = tps_taskprocessor_pop(tps);
|
|
|
|
if (!t) {
|
|
|
|
ao2_unlock(tps);
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2014-10-03 17:47:42 +00:00
|
|
|
tps->thread = pthread_self();
|
2013-02-12 21:45:59 +00:00
|
|
|
tps->executing = 1;
|
2012-11-08 23:27:16 +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
|
|
|
if (t->wants_local) {
|
|
|
|
local.local_data = tps->local_data;
|
|
|
|
local.data = t->datap;
|
|
|
|
}
|
|
|
|
ao2_unlock(tps);
|
2012-11-08 23:27:16 +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
|
|
|
if (t->wants_local) {
|
|
|
|
t->callback.execute_local(&local);
|
|
|
|
} else {
|
|
|
|
t->callback.execute(t->datap);
|
2013-02-12 21:45:59 +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
|
|
|
tps_task_free(t);
|
2012-11-08 23:27:16 +00:00
|
|
|
|
2012-11-08 23:35:10 +00:00
|
|
|
ao2_lock(tps);
|
2014-10-03 17:47:42 +00:00
|
|
|
tps->thread = AST_PTHREADT_NULL;
|
2013-02-12 21:45:59 +00:00
|
|
|
/* We need to check size in the same critical section where we reset the
|
|
|
|
* executing bit. Avoids a race condition where a task is pushed right
|
|
|
|
* after we pop an empty stack.
|
|
|
|
*/
|
|
|
|
tps->executing = 0;
|
2015-11-11 23:11:53 +00:00
|
|
|
size = ast_taskprocessor_size(tps);
|
2016-01-06 23:19:19 +00:00
|
|
|
|
|
|
|
/* Update the stats */
|
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 (tps->stats) {
|
2016-01-06 23:19:19 +00:00
|
|
|
++tps->stats->_tasks_processed_count;
|
|
|
|
|
|
|
|
/* Include the task we just executed as part of the queue size. */
|
|
|
|
if (size >= tps->stats->max_qsize) {
|
|
|
|
tps->stats->max_qsize = size + 1;
|
2012-11-08 23:27:16 +00:00
|
|
|
}
|
|
|
|
}
|
2012-11-08 23:35:10 +00:00
|
|
|
ao2_unlock(tps);
|
2012-11-08 23:27:16 +00:00
|
|
|
|
2013-02-12 21:45:59 +00:00
|
|
|
/* If we executed a task, check for the transition to empty */
|
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 (size == 0 && tps->listener->callbacks->emptied) {
|
2012-11-08 23:27:16 +00:00
|
|
|
tps->listener->callbacks->emptied(tps->listener);
|
|
|
|
}
|
2013-02-12 21:45:59 +00:00
|
|
|
return size > 0;
|
2012-11-08 23:27:16 +00:00
|
|
|
}
|
2014-10-03 17:47:42 +00:00
|
|
|
|
|
|
|
int ast_taskprocessor_is_task(struct ast_taskprocessor *tps)
|
|
|
|
{
|
|
|
|
int is_task;
|
|
|
|
|
|
|
|
ao2_lock(tps);
|
|
|
|
is_task = pthread_equal(tps->thread, pthread_self());
|
|
|
|
ao2_unlock(tps);
|
|
|
|
return is_task;
|
|
|
|
}
|
2016-01-07 22:15:35 +00:00
|
|
|
|
|
|
|
unsigned int ast_taskprocessor_seq_num(void)
|
|
|
|
{
|
|
|
|
static int seq_num;
|
|
|
|
|
|
|
|
return (unsigned int) ast_atomic_fetchadd_int(&seq_num, +1);
|
|
|
|
}
|
|
|
|
|
|
|
|
void ast_taskprocessor_build_name(char *buf, unsigned int size, const char *format, ...)
|
|
|
|
{
|
|
|
|
va_list ap;
|
|
|
|
int user_size;
|
|
|
|
#define SEQ_STR_SIZE (1 + 8 + 1) /* Dash plus 8 hex digits plus null terminator */
|
|
|
|
|
|
|
|
ast_assert(buf != NULL);
|
|
|
|
ast_assert(SEQ_STR_SIZE <= size);
|
|
|
|
|
|
|
|
va_start(ap, format);
|
|
|
|
user_size = vsnprintf(buf, size - (SEQ_STR_SIZE - 1), format, ap);
|
|
|
|
va_end(ap);
|
|
|
|
if (user_size < 0) {
|
|
|
|
/*
|
|
|
|
* Wow! We got an output error to a memory buffer.
|
|
|
|
* Assume no user part of name written.
|
|
|
|
*/
|
|
|
|
user_size = 0;
|
|
|
|
} else if (size < user_size + SEQ_STR_SIZE) {
|
|
|
|
/* Truncate user part of name to make sequence number fit. */
|
|
|
|
user_size = size - SEQ_STR_SIZE;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Append sequence number to end of user name. */
|
|
|
|
snprintf(buf + user_size, SEQ_STR_SIZE, "-%08x", ast_taskprocessor_seq_num());
|
|
|
|
}
|