asterisk/main/taskprocessor.c

1459 lines
39 KiB
C
Raw Permalink Normal View History

/*
* Asterisk -- An open source telephony toolkit.
*
* Copyright (C) 2007-2013, Digium, Inc.
*
* 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.
*/
/*!
* \file
* \brief Maintain a container of uniquely-named taskprocessor threads that can be shared across modules.
*
* \author Dwayne Hubbard <dhubbard@digium.com>
*/
/*** MODULEINFO
<support_level>core</support_level>
***/
#include "asterisk.h"
#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"
/*!
* \brief tps_task structure is queued to a taskprocessor
*
* tps_tasks are processed in FIFO order and freed by the taskprocessing
* thread after the task handler returns. The callback function that is assigned
* to the execute() function pointer is responsible for releasing datap resources if necessary.
*/
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;
/*! \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;
};
/*! \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 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;
/*! \brief Taskprocessor current queue size */
long tps_queue_size;
/*! \brief Taskprocessor low water clear alert level */
long tps_queue_low;
/*! \brief Taskprocessor high water alert trigger level */
long tps_queue_high;
/*! \brief Taskprocessor queue */
AST_LIST_HEAD_NOLOCK(tps_queue, tps_task) tps_queue;
struct ast_taskprocessor_listener *listener;
/*! Current thread executing the tasks */
pthread_t thread;
/*! Indicates if the taskprocessor is currently executing a task */
unsigned int executing:1;
/*! Indicates that a high water warning has been issued on this task processor */
unsigned int high_water_warned:1;
/*! Indicates that a high water alert is active on this taskprocessor */
unsigned int high_water_alert:1;
/*! Indicates if the taskprocessor is currently suspended */
unsigned int suspended:1;
taskprocessor: Enable subsystems and overload by subsystem To prevent one subsystem's taskprocessors from causing others to stall, new capabilities have been added to taskprocessors. * Any taskprocessor name that has a '/' will have the part before the '/' saved as its "subsystem". Examples: "sorcery/acl-0000006a" and "sorcery/aor-00000019" will be grouped to subsystem "sorcery". "pjsip/distributor-00000025" and "pjsip/distributor-00000026" will bn grouped to subsystem "pjsip". Taskprocessors with no '/' have an empty subsystem. * When a taskprocessor enters high-water alert status and it has a non-empty subsystem, the subsystem alert count will be incremented. * When a taskprocessor leaves high-water alert status and it has a non-empty subsystem, the subsystem alert count will be decremented. * A new api ast_taskprocessor_get_subsystem_alert() has been added that returns the number of taskprocessors in alert for the subsystem. * A new CLI command "core show taskprocessor alerted subsystems" has been added. * A new unit test was addded. REMINDER: The taskprocessor code itself doesn't take any action based on high-water alerts or overloading. It's up to taskprocessor users to check and take action themselves. Currently only the pjsip distributor does this. * A new pjsip/global option "taskprocessor_overload_trigger" has been added that allows the user to select the trigger mechanism the distributor uses to pause accepting new requests. "none": Don't pause on any overload condition. "global": Pause on ANY taskprocessor overload (the default and current behavior) "pjsip_only": Pause only on pjsip taskprocessor overloads. * The core pjsip pool was renamed from "SIP" to "pjsip" so it can be properly grouped into the "pjsip" subsystem. * stasis taskprocessor names were changed to "stasis" as the subsystem. * Sorcery core taskprocessor names were changed to "sorcery" to match the object taskprocessors. Change-Id: I8c19068bb2fc26610a9f0b8624bdf577a04fcd56
2019-02-15 18:53:50 +00:00
/*! \brief Anything before the first '/' in the name (if there is one) */
char *subsystem;
/*! \brief Friendly name of the taskprocessor.
* Subsystem is appended after the name's NULL terminator.
*/
char name[0];
};
/*!
* \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;
};
taskprocessor: Enable subsystems and overload by subsystem To prevent one subsystem's taskprocessors from causing others to stall, new capabilities have been added to taskprocessors. * Any taskprocessor name that has a '/' will have the part before the '/' saved as its "subsystem". Examples: "sorcery/acl-0000006a" and "sorcery/aor-00000019" will be grouped to subsystem "sorcery". "pjsip/distributor-00000025" and "pjsip/distributor-00000026" will bn grouped to subsystem "pjsip". Taskprocessors with no '/' have an empty subsystem. * When a taskprocessor enters high-water alert status and it has a non-empty subsystem, the subsystem alert count will be incremented. * When a taskprocessor leaves high-water alert status and it has a non-empty subsystem, the subsystem alert count will be decremented. * A new api ast_taskprocessor_get_subsystem_alert() has been added that returns the number of taskprocessors in alert for the subsystem. * A new CLI command "core show taskprocessor alerted subsystems" has been added. * A new unit test was addded. REMINDER: The taskprocessor code itself doesn't take any action based on high-water alerts or overloading. It's up to taskprocessor users to check and take action themselves. Currently only the pjsip distributor does this. * A new pjsip/global option "taskprocessor_overload_trigger" has been added that allows the user to select the trigger mechanism the distributor uses to pause accepting new requests. "none": Don't pause on any overload condition. "global": Pause on ANY taskprocessor overload (the default and current behavior) "pjsip_only": Pause only on pjsip taskprocessor overloads. * The core pjsip pool was renamed from "SIP" to "pjsip" so it can be properly grouped into the "pjsip" subsystem. * stasis taskprocessor names were changed to "stasis" as the subsystem. * Sorcery core taskprocessor names were changed to "sorcery" to match the object taskprocessors. Change-Id: I8c19068bb2fc26610a9f0b8624bdf577a04fcd56
2019-02-15 18:53:50 +00:00
/*!
* Keep track of which subsystems are in alert
* and how many of their taskprocessors are overloaded.
*/
struct subsystem_alert {
unsigned int alert_count;
char subsystem[0];
};
static AST_VECTOR_RW(subsystem_alert_vector, struct subsystem_alert *) overloaded_subsystems;
#ifdef LOW_MEMORY
#define TPS_MAX_BUCKETS 61
#else
/*! \brief Number of buckets in the tps_singletons container. */
#define TPS_MAX_BUCKETS 1567
#endif
/*! \brief tps_singletons is the astobj2 container for taskprocessor singletons */
static struct ao2_container *tps_singletons;
/*! \brief CLI <example>taskprocessor ping &lt;blah&gt;</example> operation requires a ping condition */
static ast_cond_t cli_ping_cond;
/*! \brief CLI <example>taskprocessor ping &lt;blah&gt;</example> operation requires a ping condition lock */
AST_MUTEX_DEFINE_STATIC(cli_ping_cond_lock);
/*! \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 */
static int tps_cmp_cb(void *obj, void *arg, int flags);
/*! \brief CLI <example>taskprocessor ping &lt;blah&gt;</example> handler function */
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);
taskprocessor: Enable subsystems and overload by subsystem To prevent one subsystem's taskprocessors from causing others to stall, new capabilities have been added to taskprocessors. * Any taskprocessor name that has a '/' will have the part before the '/' saved as its "subsystem". Examples: "sorcery/acl-0000006a" and "sorcery/aor-00000019" will be grouped to subsystem "sorcery". "pjsip/distributor-00000025" and "pjsip/distributor-00000026" will bn grouped to subsystem "pjsip". Taskprocessors with no '/' have an empty subsystem. * When a taskprocessor enters high-water alert status and it has a non-empty subsystem, the subsystem alert count will be incremented. * When a taskprocessor leaves high-water alert status and it has a non-empty subsystem, the subsystem alert count will be decremented. * A new api ast_taskprocessor_get_subsystem_alert() has been added that returns the number of taskprocessors in alert for the subsystem. * A new CLI command "core show taskprocessor alerted subsystems" has been added. * A new unit test was addded. REMINDER: The taskprocessor code itself doesn't take any action based on high-water alerts or overloading. It's up to taskprocessor users to check and take action themselves. Currently only the pjsip distributor does this. * A new pjsip/global option "taskprocessor_overload_trigger" has been added that allows the user to select the trigger mechanism the distributor uses to pause accepting new requests. "none": Don't pause on any overload condition. "global": Pause on ANY taskprocessor overload (the default and current behavior) "pjsip_only": Pause only on pjsip taskprocessor overloads. * The core pjsip pool was renamed from "SIP" to "pjsip" so it can be properly grouped into the "pjsip" subsystem. * stasis taskprocessor names were changed to "stasis" as the subsystem. * Sorcery core taskprocessor names were changed to "sorcery" to match the object taskprocessors. Change-Id: I8c19068bb2fc26610a9f0b8624bdf577a04fcd56
2019-02-15 18:53:50 +00:00
static char *cli_subsystem_alert_report(struct ast_cli_entry *e, int cmd, struct ast_cli_args *a);
static char *cli_tps_reset_stats(struct ast_cli_entry *e, int cmd, struct ast_cli_args *a);
static char *cli_tps_reset_stats_all(struct ast_cli_entry *e, int cmd, struct ast_cli_args *a);
static int tps_sort_cb(const void *obj_left, const void *obj_right, int flags);
static struct ast_cli_entry taskprocessor_clis[] = {
AST_CLI_DEFINE(cli_tps_ping, "Ping a named task processor"),
AST_CLI_DEFINE(cli_tps_report, "List instantiated task processors and statistics"),
taskprocessor: Enable subsystems and overload by subsystem To prevent one subsystem's taskprocessors from causing others to stall, new capabilities have been added to taskprocessors. * Any taskprocessor name that has a '/' will have the part before the '/' saved as its "subsystem". Examples: "sorcery/acl-0000006a" and "sorcery/aor-00000019" will be grouped to subsystem "sorcery". "pjsip/distributor-00000025" and "pjsip/distributor-00000026" will bn grouped to subsystem "pjsip". Taskprocessors with no '/' have an empty subsystem. * When a taskprocessor enters high-water alert status and it has a non-empty subsystem, the subsystem alert count will be incremented. * When a taskprocessor leaves high-water alert status and it has a non-empty subsystem, the subsystem alert count will be decremented. * A new api ast_taskprocessor_get_subsystem_alert() has been added that returns the number of taskprocessors in alert for the subsystem. * A new CLI command "core show taskprocessor alerted subsystems" has been added. * A new unit test was addded. REMINDER: The taskprocessor code itself doesn't take any action based on high-water alerts or overloading. It's up to taskprocessor users to check and take action themselves. Currently only the pjsip distributor does this. * A new pjsip/global option "taskprocessor_overload_trigger" has been added that allows the user to select the trigger mechanism the distributor uses to pause accepting new requests. "none": Don't pause on any overload condition. "global": Pause on ANY taskprocessor overload (the default and current behavior) "pjsip_only": Pause only on pjsip taskprocessor overloads. * The core pjsip pool was renamed from "SIP" to "pjsip" so it can be properly grouped into the "pjsip" subsystem. * stasis taskprocessor names were changed to "stasis" as the subsystem. * Sorcery core taskprocessor names were changed to "sorcery" to match the object taskprocessors. Change-Id: I8c19068bb2fc26610a9f0b8624bdf577a04fcd56
2019-02-15 18:53:50 +00:00
AST_CLI_DEFINE(cli_subsystem_alert_report, "List task processor subsystems in alert"),
AST_CLI_DEFINE(cli_tps_reset_stats, "Reset a named task processor's stats"),
AST_CLI_DEFINE(cli_tps_reset_stats_all, "Reset all task processors' stats"),
};
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;
};
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)
{
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);
}
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)
{
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;
}
/*!
* \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)
{
struct ast_taskprocessor_listener *listener = data;
struct ast_taskprocessor *tps = listener->tps;
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;
}
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);
}
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");
return NULL;
}
static int default_listener_start(struct ast_taskprocessor_listener *listener)
{
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
if (ast_pthread_create(&pvt->poll_thread, NULL, default_tps_processing_function, listener)) {
return -1;
}
return 0;
}
static void default_task_pushed(struct ast_taskprocessor_listener *listener, int was_empty)
{
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
if (ast_sem_post(&pvt->sem) != 0) {
ast_log(LOG_ERROR, "Failed to notify of enqueued task: %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
static int default_listener_die(void *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
struct default_taskprocessor_listener_pvt *pvt = data;
pvt->dead = 1;
return 0;
}
static void default_listener_shutdown(struct ast_taskprocessor_listener *listener)
{
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");
if (ast_taskprocessor_push(listener->tps, default_listener_die, pvt)) {
/* This will cause the thread to exit early without completing tasks already
* in the queue. This is probably the least bad option in this situation. */
default_listener_die(pvt);
}
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->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) {
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) {
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
}
}
pvt->poll_thread = AST_PTHREADT_NULL;
}
static const struct ast_taskprocessor_listener_callbacks default_listener_callbacks = {
.start = default_listener_start,
.task_pushed = default_task_pushed,
.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,
};
/*! \brief How many seconds to wait for running taskprocessors to finish on shutdown. */
#define AST_TASKPROCESSOR_SHUTDOWN_MAX_WAIT 10
/*!
* \internal
* \brief Clean up resources on Asterisk shutdown
*/
static void tps_shutdown(void)
{
int objcount;
int tries;
struct ao2_container *sorted_tps;
struct ast_taskprocessor *tps;
struct ao2_iterator iter;
struct timespec delay = {1, 0};
/* During shutdown there may still be taskprocessor threads running and those
* tasprocessors reference tps_singletons. When those taskprocessors finish
* they will call ast_taskprocessor_unreference, creating a race condition which
* can result in tps_singletons being referenced after being deleted. To try and
* avoid this we check the container count and if greater than zero, give the
* running taskprocessors a chance to finish */
objcount = ao2_container_count(tps_singletons);
if (objcount > 0) {
ast_log(LOG_DEBUG,
"waiting for taskprocessor shutdown, %d tps object(s) still allocated.\n",
objcount);
/* give the running taskprocessors a chance to finish, up to
* AST_TASKPROCESSOR_SHUTDOWN_MAX_WAIT seconds */
for (tries = 0; tries < AST_TASKPROCESSOR_SHUTDOWN_MAX_WAIT; tries++) {
while (nanosleep(&delay, &delay));
objcount = ao2_container_count(tps_singletons);
/* if count is 0, we are done waiting */
if (objcount == 0) {
break;
}
delay.tv_sec = 1;
delay.tv_nsec = 0;
ast_log(LOG_DEBUG,
"waiting for taskprocessor shutdown, %d tps object(s) still allocated.\n",
objcount);
}
}
/* rather than try forever, risk an assertion on shutdown. This probably indicates
* a taskprocessor was not cleaned up somewhere */
if (objcount > 0) {
ast_log(LOG_ERROR,
"Asertion may occur, the following taskprocessors are still runing:\n");
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)) {
ast_log(LOG_ERROR, "unable to get sorted list of taskprocessors");
}
else {
iter = ao2_iterator_init(sorted_tps, AO2_ITERATOR_UNLINK);
while ((tps = ao2_iterator_next(&iter))) {
ast_log(LOG_ERROR, "taskprocessor '%s'\n", tps->name);
}
}
ao2_cleanup(sorted_tps);
}
else {
ast_log(LOG_DEBUG,
"All waiting taskprocessors cleared!\n");
}
ast_cli_unregister_multiple(taskprocessor_clis, ARRAY_LEN(taskprocessor_clis));
taskprocessor: Enable subsystems and overload by subsystem To prevent one subsystem's taskprocessors from causing others to stall, new capabilities have been added to taskprocessors. * Any taskprocessor name that has a '/' will have the part before the '/' saved as its "subsystem". Examples: "sorcery/acl-0000006a" and "sorcery/aor-00000019" will be grouped to subsystem "sorcery". "pjsip/distributor-00000025" and "pjsip/distributor-00000026" will bn grouped to subsystem "pjsip". Taskprocessors with no '/' have an empty subsystem. * When a taskprocessor enters high-water alert status and it has a non-empty subsystem, the subsystem alert count will be incremented. * When a taskprocessor leaves high-water alert status and it has a non-empty subsystem, the subsystem alert count will be decremented. * A new api ast_taskprocessor_get_subsystem_alert() has been added that returns the number of taskprocessors in alert for the subsystem. * A new CLI command "core show taskprocessor alerted subsystems" has been added. * A new unit test was addded. REMINDER: The taskprocessor code itself doesn't take any action based on high-water alerts or overloading. It's up to taskprocessor users to check and take action themselves. Currently only the pjsip distributor does this. * A new pjsip/global option "taskprocessor_overload_trigger" has been added that allows the user to select the trigger mechanism the distributor uses to pause accepting new requests. "none": Don't pause on any overload condition. "global": Pause on ANY taskprocessor overload (the default and current behavior) "pjsip_only": Pause only on pjsip taskprocessor overloads. * The core pjsip pool was renamed from "SIP" to "pjsip" so it can be properly grouped into the "pjsip" subsystem. * stasis taskprocessor names were changed to "stasis" as the subsystem. * Sorcery core taskprocessor names were changed to "sorcery" to match the object taskprocessors. Change-Id: I8c19068bb2fc26610a9f0b8624bdf577a04fcd56
2019-02-15 18:53:50 +00:00
AST_VECTOR_CALLBACK_VOID(&overloaded_subsystems, ast_free);
AST_VECTOR_RW_FREE(&overloaded_subsystems);
ao2_t_ref(tps_singletons, -1, "Unref tps_singletons in shutdown");
tps_singletons = NULL;
}
/* initialize the taskprocessor container and register CLI operations */
int ast_tps_init(void)
{
tps_singletons = ao2_container_alloc_hash(AO2_ALLOC_OPT_LOCK_MUTEX, 0,
TPS_MAX_BUCKETS, tps_hash_cb, NULL, tps_cmp_cb);
if (!tps_singletons) {
ast_log(LOG_ERROR, "taskprocessor container failed to initialize!\n");
return -1;
}
taskprocessor: Enable subsystems and overload by subsystem To prevent one subsystem's taskprocessors from causing others to stall, new capabilities have been added to taskprocessors. * Any taskprocessor name that has a '/' will have the part before the '/' saved as its "subsystem". Examples: "sorcery/acl-0000006a" and "sorcery/aor-00000019" will be grouped to subsystem "sorcery". "pjsip/distributor-00000025" and "pjsip/distributor-00000026" will bn grouped to subsystem "pjsip". Taskprocessors with no '/' have an empty subsystem. * When a taskprocessor enters high-water alert status and it has a non-empty subsystem, the subsystem alert count will be incremented. * When a taskprocessor leaves high-water alert status and it has a non-empty subsystem, the subsystem alert count will be decremented. * A new api ast_taskprocessor_get_subsystem_alert() has been added that returns the number of taskprocessors in alert for the subsystem. * A new CLI command "core show taskprocessor alerted subsystems" has been added. * A new unit test was addded. REMINDER: The taskprocessor code itself doesn't take any action based on high-water alerts or overloading. It's up to taskprocessor users to check and take action themselves. Currently only the pjsip distributor does this. * A new pjsip/global option "taskprocessor_overload_trigger" has been added that allows the user to select the trigger mechanism the distributor uses to pause accepting new requests. "none": Don't pause on any overload condition. "global": Pause on ANY taskprocessor overload (the default and current behavior) "pjsip_only": Pause only on pjsip taskprocessor overloads. * The core pjsip pool was renamed from "SIP" to "pjsip" so it can be properly grouped into the "pjsip" subsystem. * stasis taskprocessor names were changed to "stasis" as the subsystem. * Sorcery core taskprocessor names were changed to "sorcery" to match the object taskprocessors. Change-Id: I8c19068bb2fc26610a9f0b8624bdf577a04fcd56
2019-02-15 18:53:50 +00:00
if (AST_VECTOR_RW_INIT(&overloaded_subsystems, 10)) {
ao2_ref(tps_singletons, -1);
ast_log(LOG_ERROR, "taskprocessor subsystems vector failed to initialize!\n");
return -1;
}
ast_cond_init(&cli_ping_cond, NULL);
ast_cli_register_multiple(taskprocessor_clis, ARRAY_LEN(taskprocessor_clis));
ast_register_cleanup(tps_shutdown);
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;
}
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;
return t;
}
/* release task resources */
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);
return NULL;
}
/* Taskprocessor tab completion.
*
* The caller of this function is responsible for argument
* position checks prior to calling.
*/
static char *tps_taskprocessor_tab_complete(struct ast_cli_args *a)
{
int tklen;
struct ast_taskprocessor *p;
struct ao2_iterator i;
tklen = strlen(a->word);
i = ao2_iterator_init(tps_singletons, 0);
while ((p = ao2_iterator_next(&i))) {
if (!strncasecmp(a->word, p->name, tklen)) {
if (ast_cli_completion_add(ast_strdup(p->name))) {
ast_taskprocessor_unreference(p);
break;
}
}
ast_taskprocessor_unreference(p);
}
ao2_iterator_destroy(&i);
return NULL;
}
/* 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;
const char *name;
struct timeval when;
struct timespec ts;
struct ast_taskprocessor *tps;
switch (cmd) {
case CLI_INIT:
e->command = "core ping taskprocessor";
e->usage =
"Usage: core ping taskprocessor <taskprocessor>\n"
" Displays the time required for a task to be processed\n";
return NULL;
case CLI_GENERATE:
if (a->pos == 3) {
return tps_taskprocessor_tab_complete(a);
} else {
return NULL;
}
}
if (a->argc != 4)
return CLI_SHOWUSAGE;
name = a->argv[3];
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);
/*
* 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));
ts.tv_sec = when.tv_sec;
ts.tv_nsec = when.tv_usec * 1000;
ast_mutex_lock(&cli_ping_cond_lock);
if (ast_taskprocessor_push(tps, tps_ping_handler, 0) < 0) {
ast_mutex_unlock(&cli_ping_cond_lock);
ast_cli(a->fd, "\nping failed: could not push task to %s\n\n", name);
ast_taskprocessor_unreference(tps);
return CLI_FAILURE;
}
ast_cond_timedwait(&cli_ping_cond, &cli_ping_cond_lock, &ts);
ast_mutex_unlock(&cli_ping_cond_lock);
end = ast_tvnow();
delta = ast_tvsub(end, begin);
ast_cli(a->fd, "\n\t%24s ping time: %.1ld.%.6ld sec\n\n", name, (long)delta.tv_sec, (long int)delta.tv_usec);
ast_taskprocessor_unreference(tps);
return CLI_SUCCESS;
}
/*!
* \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 negative if obj_left < obj_right
* \retval 0 if obj_left == obj_right
* \retval positive 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;
}
#define FMT_HEADERS "%-70s %10s %10s %10s %10s %10s\n"
#define FMT_FIELDS "%-70s %10lu %10lu %10lu %10lu %10lu\n"
/*!
* \internal
* \brief Print taskprocessor information to CLI.
* \since 13.30.0
*
* \param fd the file descriptor
* \param tps the taskprocessor
*/
static void tps_report_taskprocessor_list_helper(int fd, struct ast_taskprocessor *tps)
{
ast_cli(fd, FMT_FIELDS, tps->name, tps->stats._tasks_processed_count,
tps->tps_queue_size, tps->stats.max_qsize, tps->tps_queue_low,
tps->tps_queue_high);
}
/*!
* \internal
* \brief Prints an optionally narrowed down list of taskprocessors to the CLI.
* \since 13.30.0
*
* \param fd the file descriptor
* \param like the string we are matching on
*
* \return number of taskprocessors on success
* \retval 0 otherwise
*/
static int tps_report_taskprocessor_list(int fd, const char *like)
{
int tps_count = 0;
int word_len;
struct ao2_container *sorted_tps;
struct ast_taskprocessor *tps;
struct ao2_iterator iter;
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)) {
ast_debug(1, "Failed to retrieve sorted taskprocessors\n");
ao2_cleanup(sorted_tps);
return 0;
}
word_len = strlen(like);
iter = ao2_iterator_init(sorted_tps, AO2_ITERATOR_UNLINK);
while ((tps = ao2_iterator_next(&iter))) {
if (like) {
if (!strncasecmp(like, tps->name, word_len)) {
tps_report_taskprocessor_list_helper(fd, tps);
tps_count++;
}
} else {
tps_report_taskprocessor_list_helper(fd, tps);
tps_count++;
}
ast_taskprocessor_unreference(tps);
}
ao2_iterator_destroy(&iter);
ao2_ref(sorted_tps, -1);
return tps_count;
}
static char *cli_tps_report(struct ast_cli_entry *e, int cmd, struct ast_cli_args *a)
{
const char *like;
switch (cmd) {
case CLI_INIT:
e->command = "core show taskprocessors [like]";
e->usage =
"Usage: core show taskprocessors [like keyword]\n"
" Shows a list of instantiated task processors and their statistics\n";
return NULL;
case CLI_GENERATE:
if (a->pos == e->args) {
return tps_taskprocessor_tab_complete(a);
} else {
return NULL;
}
}
if (a->argc == e->args - 1) {
like = "";
} else if (a->argc == e->args + 1 && !strcasecmp(a->argv[e->args-1], "like")) {
like = a->argv[e->args];
} else {
return CLI_SHOWUSAGE;
}
ast_cli(a->fd, "\n" FMT_HEADERS, "Processor", "Processed", "In Queue", "Max Depth", "Low water", "High water");
ast_cli(a->fd, "\n%d taskprocessors\n\n", tps_report_taskprocessor_list(a->fd, like));
return CLI_SUCCESS;
}
/* hash callback for astobj2 */
static int tps_hash_cb(const void *obj, const int flags)
{
const struct ast_taskprocessor *tps = obj;
const char *name = flags & OBJ_KEY ? obj : tps->name;
return ast_str_case_hash(name);
}
/* compare callback for astobj2 */
static int tps_cmp_cb(void *obj, void *arg, int flags)
{
struct ast_taskprocessor *lhs = obj, *rhs = arg;
const char *rhsname = flags & OBJ_KEY ? arg : rhs->name;
return !strcasecmp(lhs->name, rhsname) ? CMP_MATCH | CMP_STOP : 0;
}
taskprocessor: Enable subsystems and overload by subsystem To prevent one subsystem's taskprocessors from causing others to stall, new capabilities have been added to taskprocessors. * Any taskprocessor name that has a '/' will have the part before the '/' saved as its "subsystem". Examples: "sorcery/acl-0000006a" and "sorcery/aor-00000019" will be grouped to subsystem "sorcery". "pjsip/distributor-00000025" and "pjsip/distributor-00000026" will bn grouped to subsystem "pjsip". Taskprocessors with no '/' have an empty subsystem. * When a taskprocessor enters high-water alert status and it has a non-empty subsystem, the subsystem alert count will be incremented. * When a taskprocessor leaves high-water alert status and it has a non-empty subsystem, the subsystem alert count will be decremented. * A new api ast_taskprocessor_get_subsystem_alert() has been added that returns the number of taskprocessors in alert for the subsystem. * A new CLI command "core show taskprocessor alerted subsystems" has been added. * A new unit test was addded. REMINDER: The taskprocessor code itself doesn't take any action based on high-water alerts or overloading. It's up to taskprocessor users to check and take action themselves. Currently only the pjsip distributor does this. * A new pjsip/global option "taskprocessor_overload_trigger" has been added that allows the user to select the trigger mechanism the distributor uses to pause accepting new requests. "none": Don't pause on any overload condition. "global": Pause on ANY taskprocessor overload (the default and current behavior) "pjsip_only": Pause only on pjsip taskprocessor overloads. * The core pjsip pool was renamed from "SIP" to "pjsip" so it can be properly grouped into the "pjsip" subsystem. * stasis taskprocessor names were changed to "stasis" as the subsystem. * Sorcery core taskprocessor names were changed to "sorcery" to match the object taskprocessors. Change-Id: I8c19068bb2fc26610a9f0b8624bdf577a04fcd56
2019-02-15 18:53:50 +00:00
static int subsystem_match(struct subsystem_alert *alert, const char *subsystem)
{
return !strcmp(alert->subsystem, subsystem);
}
static int subsystem_cmp(struct subsystem_alert *a, struct subsystem_alert *b)
{
return strcmp(a->subsystem, b->subsystem);
}
unsigned int ast_taskprocessor_get_subsystem_alert(const char *subsystem)
{
struct subsystem_alert *alert;
unsigned int count = 0;
int idx;
AST_VECTOR_RW_RDLOCK(&overloaded_subsystems);
idx = AST_VECTOR_GET_INDEX(&overloaded_subsystems, subsystem, subsystem_match);
if (idx >= 0) {
alert = AST_VECTOR_GET(&overloaded_subsystems, idx);
count = alert->alert_count;
}
AST_VECTOR_RW_UNLOCK(&overloaded_subsystems);
return count;
}
static void subsystem_alert_increment(const char *subsystem)
{
struct subsystem_alert *alert;
int idx;
if (ast_strlen_zero(subsystem)) {
return;
}
AST_VECTOR_RW_WRLOCK(&overloaded_subsystems);
idx = AST_VECTOR_GET_INDEX(&overloaded_subsystems, subsystem, subsystem_match);
if (idx >= 0) {
alert = AST_VECTOR_GET(&overloaded_subsystems, idx);
alert->alert_count++;
AST_VECTOR_RW_UNLOCK(&overloaded_subsystems);
return;
}
alert = ast_malloc(sizeof(*alert) + strlen(subsystem) + 1);
if (!alert) {
AST_VECTOR_RW_UNLOCK(&overloaded_subsystems);
return;
}
alert->alert_count = 1;
strcpy(alert->subsystem, subsystem); /* Safe */
if (AST_VECTOR_APPEND(&overloaded_subsystems, alert)) {
ast_free(alert);
}
AST_VECTOR_RW_UNLOCK(&overloaded_subsystems);
}
static void subsystem_alert_decrement(const char *subsystem)
{
struct subsystem_alert *alert;
int idx;
if (ast_strlen_zero(subsystem)) {
return;
}
AST_VECTOR_RW_WRLOCK(&overloaded_subsystems);
idx = AST_VECTOR_GET_INDEX(&overloaded_subsystems, subsystem, subsystem_match);
if (idx < 0) {
ast_log(LOG_ERROR,
"Can't decrement alert count for subsystem '%s' as it wasn't in alert\n", subsystem);
AST_VECTOR_RW_UNLOCK(&overloaded_subsystems);
return;
}
alert = AST_VECTOR_GET(&overloaded_subsystems, idx);
alert->alert_count--;
if (alert->alert_count <= 0) {
AST_VECTOR_REMOVE(&overloaded_subsystems, idx, 0);
ast_free(alert);
}
AST_VECTOR_RW_UNLOCK(&overloaded_subsystems);
}
static void subsystem_copy(struct subsystem_alert *alert,
struct subsystem_alert_vector *vector)
{
struct subsystem_alert *alert_copy;
alert_copy = ast_malloc(sizeof(*alert_copy) + strlen(alert->subsystem) + 1);
if (!alert_copy) {
return;
}
alert_copy->alert_count = alert->alert_count;
strcpy(alert_copy->subsystem, alert->subsystem); /* Safe */
if (AST_VECTOR_ADD_SORTED(vector, alert_copy, subsystem_cmp)) {
ast_free(alert_copy);
}
}
static char *cli_subsystem_alert_report(struct ast_cli_entry *e, int cmd, struct ast_cli_args *a)
{
struct subsystem_alert_vector sorted_subsystems;
int i;
#define FMT_HEADERS_SUBSYSTEM "%-32s %12s\n"
#define FMT_FIELDS_SUBSYSTEM "%-32s %12u\n"
switch (cmd) {
case CLI_INIT:
e->command = "core show taskprocessor alerted subsystems";
e->usage =
"Usage: core show taskprocessor alerted subsystems\n"
" Shows a list of task processor subsystems that are currently alerted\n";
return NULL;
case CLI_GENERATE:
return NULL;
}
if (a->argc != e->args) {
return CLI_SHOWUSAGE;
}
if (AST_VECTOR_INIT(&sorted_subsystems, AST_VECTOR_SIZE(&overloaded_subsystems))) {
return CLI_FAILURE;
}
AST_VECTOR_RW_RDLOCK(&overloaded_subsystems);
for (i = 0; i < AST_VECTOR_SIZE(&overloaded_subsystems); i++) {
subsystem_copy(AST_VECTOR_GET(&overloaded_subsystems, i), &sorted_subsystems);
}
AST_VECTOR_RW_UNLOCK(&overloaded_subsystems);
ast_cli(a->fd, "\n" FMT_HEADERS_SUBSYSTEM, "Subsystem", "Alert Count");
for (i = 0; i < AST_VECTOR_SIZE(&sorted_subsystems); i++) {
struct subsystem_alert *alert = AST_VECTOR_GET(&sorted_subsystems, i);
ast_cli(a->fd, FMT_FIELDS_SUBSYSTEM, alert->subsystem, alert->alert_count);
}
ast_cli(a->fd, "\n%zu subsystems\n\n", AST_VECTOR_SIZE(&sorted_subsystems));
taskprocessor: Enable subsystems and overload by subsystem To prevent one subsystem's taskprocessors from causing others to stall, new capabilities have been added to taskprocessors. * Any taskprocessor name that has a '/' will have the part before the '/' saved as its "subsystem". Examples: "sorcery/acl-0000006a" and "sorcery/aor-00000019" will be grouped to subsystem "sorcery". "pjsip/distributor-00000025" and "pjsip/distributor-00000026" will bn grouped to subsystem "pjsip". Taskprocessors with no '/' have an empty subsystem. * When a taskprocessor enters high-water alert status and it has a non-empty subsystem, the subsystem alert count will be incremented. * When a taskprocessor leaves high-water alert status and it has a non-empty subsystem, the subsystem alert count will be decremented. * A new api ast_taskprocessor_get_subsystem_alert() has been added that returns the number of taskprocessors in alert for the subsystem. * A new CLI command "core show taskprocessor alerted subsystems" has been added. * A new unit test was addded. REMINDER: The taskprocessor code itself doesn't take any action based on high-water alerts or overloading. It's up to taskprocessor users to check and take action themselves. Currently only the pjsip distributor does this. * A new pjsip/global option "taskprocessor_overload_trigger" has been added that allows the user to select the trigger mechanism the distributor uses to pause accepting new requests. "none": Don't pause on any overload condition. "global": Pause on ANY taskprocessor overload (the default and current behavior) "pjsip_only": Pause only on pjsip taskprocessor overloads. * The core pjsip pool was renamed from "SIP" to "pjsip" so it can be properly grouped into the "pjsip" subsystem. * stasis taskprocessor names were changed to "stasis" as the subsystem. * Sorcery core taskprocessor names were changed to "sorcery" to match the object taskprocessors. Change-Id: I8c19068bb2fc26610a9f0b8624bdf577a04fcd56
2019-02-15 18:53:50 +00:00
AST_VECTOR_CALLBACK_VOID(&sorted_subsystems, ast_free);
AST_VECTOR_FREE(&sorted_subsystems);
return CLI_SUCCESS;
}
/*! 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.
*/
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");
}
taskprocessor: Enable subsystems and overload by subsystem To prevent one subsystem's taskprocessors from causing others to stall, new capabilities have been added to taskprocessors. * Any taskprocessor name that has a '/' will have the part before the '/' saved as its "subsystem". Examples: "sorcery/acl-0000006a" and "sorcery/aor-00000019" will be grouped to subsystem "sorcery". "pjsip/distributor-00000025" and "pjsip/distributor-00000026" will bn grouped to subsystem "pjsip". Taskprocessors with no '/' have an empty subsystem. * When a taskprocessor enters high-water alert status and it has a non-empty subsystem, the subsystem alert count will be incremented. * When a taskprocessor leaves high-water alert status and it has a non-empty subsystem, the subsystem alert count will be decremented. * A new api ast_taskprocessor_get_subsystem_alert() has been added that returns the number of taskprocessors in alert for the subsystem. * A new CLI command "core show taskprocessor alerted subsystems" has been added. * A new unit test was addded. REMINDER: The taskprocessor code itself doesn't take any action based on high-water alerts or overloading. It's up to taskprocessor users to check and take action themselves. Currently only the pjsip distributor does this. * A new pjsip/global option "taskprocessor_overload_trigger" has been added that allows the user to select the trigger mechanism the distributor uses to pause accepting new requests. "none": Don't pause on any overload condition. "global": Pause on ANY taskprocessor overload (the default and current behavior) "pjsip_only": Pause only on pjsip taskprocessor overloads. * The core pjsip pool was renamed from "SIP" to "pjsip" so it can be properly grouped into the "pjsip" subsystem. * stasis taskprocessor names were changed to "stasis" as the subsystem. * Sorcery core taskprocessor names were changed to "sorcery" to match the object taskprocessors. Change-Id: I8c19068bb2fc26610a9f0b8624bdf577a04fcd56
2019-02-15 18:53:50 +00:00
if (tps->subsystem[0] != '\0') {
if (delta > 0) {
subsystem_alert_increment(tps->subsystem);
} else {
subsystem_alert_decrement(tps->subsystem);
}
}
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;
}
/* destroy the taskprocessor */
static void tps_taskprocessor_dtor(void *tps)
{
struct ast_taskprocessor *t = tps;
struct tps_task *task;
while ((task = AST_LIST_REMOVE_HEAD(&t->tps_queue, list))) {
tps_task_free(task);
}
t->tps_queue_size = 0;
if (t->high_water_alert) {
t->high_water_alert = 0;
tps_alert_add(t, -1);
}
ao2_cleanup(t->listener);
t->listener = NULL;
}
/* pop the front task and return it */
static struct tps_task *tps_taskprocessor_pop(struct ast_taskprocessor *tps)
{
struct tps_task *task;
if ((task = AST_LIST_REMOVE_HEAD(&tps->tps_queue, list))) {
--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);
}
}
return task;
}
long ast_taskprocessor_size(struct ast_taskprocessor *tps)
{
return (tps) ? tps->tps_queue_size : -1;
}
/* 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;
}
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);
}
}
struct ast_taskprocessor_listener *ast_taskprocessor_listener_alloc(const struct ast_taskprocessor_listener_callbacks *callbacks, void *user_data)
{
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);
if (!listener) {
return NULL;
}
listener->callbacks = callbacks;
listener->user_data = user_data;
return listener;
}
struct ast_taskprocessor *ast_taskprocessor_listener_get_tps(const struct ast_taskprocessor_listener *listener)
{
ao2_ref(listener->tps, +1);
return listener->tps;
}
void *ast_taskprocessor_listener_get_user_data(const struct ast_taskprocessor_listener *listener)
{
return listener->user_data;
}
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;
}
return pvt;
}
/*!
* \internal
* \brief Allocate a task processor structure
*
* \param name Name of the task processor.
* \param listener Listener to associate with the task processor.
*
* \return The newly allocated task processor.
*
* \pre tps_singletons must be locked by the caller.
*/
static struct ast_taskprocessor *__allocate_taskprocessor(const char *name, struct ast_taskprocessor_listener *listener)
{
struct ast_taskprocessor *p;
taskprocessor: Enable subsystems and overload by subsystem To prevent one subsystem's taskprocessors from causing others to stall, new capabilities have been added to taskprocessors. * Any taskprocessor name that has a '/' will have the part before the '/' saved as its "subsystem". Examples: "sorcery/acl-0000006a" and "sorcery/aor-00000019" will be grouped to subsystem "sorcery". "pjsip/distributor-00000025" and "pjsip/distributor-00000026" will bn grouped to subsystem "pjsip". Taskprocessors with no '/' have an empty subsystem. * When a taskprocessor enters high-water alert status and it has a non-empty subsystem, the subsystem alert count will be incremented. * When a taskprocessor leaves high-water alert status and it has a non-empty subsystem, the subsystem alert count will be decremented. * A new api ast_taskprocessor_get_subsystem_alert() has been added that returns the number of taskprocessors in alert for the subsystem. * A new CLI command "core show taskprocessor alerted subsystems" has been added. * A new unit test was addded. REMINDER: The taskprocessor code itself doesn't take any action based on high-water alerts or overloading. It's up to taskprocessor users to check and take action themselves. Currently only the pjsip distributor does this. * A new pjsip/global option "taskprocessor_overload_trigger" has been added that allows the user to select the trigger mechanism the distributor uses to pause accepting new requests. "none": Don't pause on any overload condition. "global": Pause on ANY taskprocessor overload (the default and current behavior) "pjsip_only": Pause only on pjsip taskprocessor overloads. * The core pjsip pool was renamed from "SIP" to "pjsip" so it can be properly grouped into the "pjsip" subsystem. * stasis taskprocessor names were changed to "stasis" as the subsystem. * Sorcery core taskprocessor names were changed to "sorcery" to match the object taskprocessors. Change-Id: I8c19068bb2fc26610a9f0b8624bdf577a04fcd56
2019-02-15 18:53:50 +00:00
char *subsystem_separator;
size_t subsystem_length = 0;
size_t name_length;
name_length = strlen(name);
subsystem_separator = strchr(name, '/');
if (subsystem_separator) {
subsystem_length = subsystem_separator - name;
}
taskprocessor: Enable subsystems and overload by subsystem To prevent one subsystem's taskprocessors from causing others to stall, new capabilities have been added to taskprocessors. * Any taskprocessor name that has a '/' will have the part before the '/' saved as its "subsystem". Examples: "sorcery/acl-0000006a" and "sorcery/aor-00000019" will be grouped to subsystem "sorcery". "pjsip/distributor-00000025" and "pjsip/distributor-00000026" will bn grouped to subsystem "pjsip". Taskprocessors with no '/' have an empty subsystem. * When a taskprocessor enters high-water alert status and it has a non-empty subsystem, the subsystem alert count will be incremented. * When a taskprocessor leaves high-water alert status and it has a non-empty subsystem, the subsystem alert count will be decremented. * A new api ast_taskprocessor_get_subsystem_alert() has been added that returns the number of taskprocessors in alert for the subsystem. * A new CLI command "core show taskprocessor alerted subsystems" has been added. * A new unit test was addded. REMINDER: The taskprocessor code itself doesn't take any action based on high-water alerts or overloading. It's up to taskprocessor users to check and take action themselves. Currently only the pjsip distributor does this. * A new pjsip/global option "taskprocessor_overload_trigger" has been added that allows the user to select the trigger mechanism the distributor uses to pause accepting new requests. "none": Don't pause on any overload condition. "global": Pause on ANY taskprocessor overload (the default and current behavior) "pjsip_only": Pause only on pjsip taskprocessor overloads. * The core pjsip pool was renamed from "SIP" to "pjsip" so it can be properly grouped into the "pjsip" subsystem. * stasis taskprocessor names were changed to "stasis" as the subsystem. * Sorcery core taskprocessor names were changed to "sorcery" to match the object taskprocessors. Change-Id: I8c19068bb2fc26610a9f0b8624bdf577a04fcd56
2019-02-15 18:53:50 +00:00
p = ao2_alloc(sizeof(*p) + name_length + subsystem_length + 2, tps_taskprocessor_dtor);
if (!p) {
ast_log(LOG_WARNING, "failed to create taskprocessor '%s'\n", name);
return NULL;
}
/* 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;
taskprocessor: Enable subsystems and overload by subsystem To prevent one subsystem's taskprocessors from causing others to stall, new capabilities have been added to taskprocessors. * Any taskprocessor name that has a '/' will have the part before the '/' saved as its "subsystem". Examples: "sorcery/acl-0000006a" and "sorcery/aor-00000019" will be grouped to subsystem "sorcery". "pjsip/distributor-00000025" and "pjsip/distributor-00000026" will bn grouped to subsystem "pjsip". Taskprocessors with no '/' have an empty subsystem. * When a taskprocessor enters high-water alert status and it has a non-empty subsystem, the subsystem alert count will be incremented. * When a taskprocessor leaves high-water alert status and it has a non-empty subsystem, the subsystem alert count will be decremented. * A new api ast_taskprocessor_get_subsystem_alert() has been added that returns the number of taskprocessors in alert for the subsystem. * A new CLI command "core show taskprocessor alerted subsystems" has been added. * A new unit test was addded. REMINDER: The taskprocessor code itself doesn't take any action based on high-water alerts or overloading. It's up to taskprocessor users to check and take action themselves. Currently only the pjsip distributor does this. * A new pjsip/global option "taskprocessor_overload_trigger" has been added that allows the user to select the trigger mechanism the distributor uses to pause accepting new requests. "none": Don't pause on any overload condition. "global": Pause on ANY taskprocessor overload (the default and current behavior) "pjsip_only": Pause only on pjsip taskprocessor overloads. * The core pjsip pool was renamed from "SIP" to "pjsip" so it can be properly grouped into the "pjsip" subsystem. * stasis taskprocessor names were changed to "stasis" as the subsystem. * Sorcery core taskprocessor names were changed to "sorcery" to match the object taskprocessors. Change-Id: I8c19068bb2fc26610a9f0b8624bdf577a04fcd56
2019-02-15 18:53:50 +00:00
strcpy(p->name, name); /* Safe */
p->subsystem = p->name + name_length + 1;
ast_copy_string(p->subsystem, name, subsystem_length + 1);
ao2_ref(listener, +1);
p->listener = listener;
p->thread = AST_PTHREADT_NULL;
ao2_ref(p, +1);
listener->tps = p;
if (!(ao2_link_flags(tps_singletons, p, OBJ_NOLOCK))) {
ast_log(LOG_ERROR, "Failed to add taskprocessor '%s' to container\n", p->name);
listener->tps = NULL;
ao2_ref(p, -2);
return NULL;
}
return p;
}
static struct ast_taskprocessor *__start_taskprocessor(struct ast_taskprocessor *p)
{
if (p && p->listener->callbacks->start(p->listener)) {
ast_log(LOG_ERROR, "Unable to start taskprocessor listener for taskprocessor %s\n",
p->name);
ast_taskprocessor_unreference(p);
return NULL;
}
return p;
}
/* Provide a reference to a taskprocessor. Create the taskprocessor if necessary, but don't
* create the taskprocessor if we were told via ast_tps_options to return a reference only
* if it already exists */
struct ast_taskprocessor *ast_taskprocessor_get(const char *name, enum ast_tps_options create)
{
struct ast_taskprocessor *p;
struct ast_taskprocessor_listener *listener;
struct default_taskprocessor_listener_pvt *pvt;
if (ast_strlen_zero(name)) {
ast_log(LOG_ERROR, "requesting a nameless taskprocessor!!!\n");
return NULL;
}
ao2_lock(tps_singletons);
p = ao2_find(tps_singletons, name, OBJ_KEY | OBJ_NOLOCK);
if (p || (create & TPS_REF_IF_EXISTS)) {
/* calling function does not want a new taskprocessor to be created if it doesn't already exist */
ao2_unlock(tps_singletons);
return p;
}
/* Create a new taskprocessor. Start by creating a default listener */
pvt = default_listener_pvt_alloc();
if (!pvt) {
ao2_unlock(tps_singletons);
return NULL;
}
listener = ast_taskprocessor_listener_alloc(&default_listener_callbacks, pvt);
if (!listener) {
ao2_unlock(tps_singletons);
default_listener_pvt_destroy(pvt);
return NULL;
}
p = __allocate_taskprocessor(name, listener);
ao2_unlock(tps_singletons);
p = __start_taskprocessor(p);
ao2_ref(listener, -1);
return p;
}
struct ast_taskprocessor *ast_taskprocessor_create_with_listener(const char *name, struct ast_taskprocessor_listener *listener)
{
struct ast_taskprocessor *p;
ao2_lock(tps_singletons);
p = ao2_find(tps_singletons, name, OBJ_KEY | OBJ_NOLOCK);
if (p) {
ao2_unlock(tps_singletons);
ast_taskprocessor_unreference(p);
return NULL;
}
p = __allocate_taskprocessor(name, listener);
ao2_unlock(tps_singletons);
return __start_taskprocessor(p);
}
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;
}
/* decrement the taskprocessor reference count and unlink from the container if necessary */
void *ast_taskprocessor_unreference(struct ast_taskprocessor *tps)
{
if (!tps) {
return NULL;
}
/* 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);
if (ao2_ref(tps, -1) > 3) {
ao2_unlock(tps_singletons);
return NULL;
}
/* 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
*/
ao2_unlink_flags(tps_singletons, tps, OBJ_NOLOCK);
ao2_unlock(tps_singletons);
listener_shutdown(tps->listener);
return NULL;
}
/* 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)
{
int previous_size;
int was_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 (!tps) {
ast_log(LOG_ERROR, "tps is NULL!\n");
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");
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
ao2_lock(tps);
AST_LIST_INSERT_TAIL(&tps->tps_queue, t, list);
previous_size = tps->tps_queue_size++;
if (tps->tps_queue_high <= tps->tps_queue_size) {
if (!tps->high_water_alert) {
ast_log(LOG_WARNING, "The '%s' task processor queue reached %ld scheduled tasks%s.\n",
tps->name, tps->tps_queue_size, tps->high_water_warned ? " again" : "");
tps->high_water_warned = 1;
tps->high_water_alert = 1;
tps_alert_add(tps, +1);
}
}
/* The currently executing task counts as still in queue */
was_empty = tps->executing ? 0 : previous_size == 0;
ao2_unlock(tps);
tps->listener->callbacks->task_pushed(tps->listener, was_empty);
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));
}
int ast_taskprocessor_suspend(struct ast_taskprocessor *tps)
{
if (tps) {
ao2_lock(tps);
tps->suspended = 1;
ao2_unlock(tps);
return 0;
}
return -1;
}
int ast_taskprocessor_unsuspend(struct ast_taskprocessor *tps)
{
if (tps) {
ao2_lock(tps);
tps->suspended = 0;
ao2_unlock(tps);
return 0;
}
return -1;
}
int ast_taskprocessor_is_suspended(struct ast_taskprocessor *tps)
{
return tps ? tps->suspended : -1;
}
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;
struct tps_task *t;
long size;
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;
}
tps->thread = pthread_self();
tps->executing = 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->wants_local) {
local.local_data = tps->local_data;
local.data = t->datap;
}
ao2_unlock(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
if (t->wants_local) {
t->callback.execute_local(&local);
} else {
t->callback.execute(t->datap);
}
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);
ao2_lock(tps);
tps->thread = AST_PTHREADT_NULL;
/* 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;
size = ast_taskprocessor_size(tps);
/* Update the stats */
++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;
}
ao2_unlock(tps);
/* 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) {
tps->listener->callbacks->emptied(tps->listener);
}
return size > 0;
}
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;
}
unsigned int ast_taskprocessor_seq_num(void)
{
static int seq_num;
return (unsigned int) ast_atomic_fetchadd_int(&seq_num, +1);
}
#define SEQ_STR_SIZE (1 + 8 + 1) /* Dash plus 8 hex digits plus null terminator */
void ast_taskprocessor_name_append(char *buf, unsigned int size, const char *name)
{
int final_size = strlen(name) + SEQ_STR_SIZE;
ast_assert(buf != NULL && name != NULL);
ast_assert(final_size <= size);
snprintf(buf, final_size, "%s-%08x", name, ast_taskprocessor_seq_num());
}
void ast_taskprocessor_build_name(char *buf, unsigned int size, const char *format, ...)
{
va_list ap;
int user_size;
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());
}
static void tps_reset_stats(struct ast_taskprocessor *tps)
{
ao2_lock(tps);
tps->stats._tasks_processed_count = 0;
tps->stats.max_qsize = 0;
ao2_unlock(tps);
}
static char *cli_tps_reset_stats(struct ast_cli_entry *e, int cmd, struct ast_cli_args *a)
{
const char *name;
struct ast_taskprocessor *tps;
switch (cmd) {
case CLI_INIT:
e->command = "core reset taskprocessor";
e->usage =
"Usage: core reset taskprocessor <taskprocessor>\n"
" Resets stats for the specified taskprocessor\n";
return NULL;
case CLI_GENERATE:
return tps_taskprocessor_tab_complete(a);
}
if (a->argc != 4) {
return CLI_SHOWUSAGE;
}
name = a->argv[3];
if (!(tps = ast_taskprocessor_get(name, TPS_REF_IF_EXISTS))) {
ast_cli(a->fd, "\nReset failed: %s not found\n\n", name);
return CLI_SUCCESS;
}
ast_cli(a->fd, "\nResetting %s\n\n", name);
tps_reset_stats(tps);
ast_taskprocessor_unreference(tps);
return CLI_SUCCESS;
}
static char *cli_tps_reset_stats_all(struct ast_cli_entry *e, int cmd, struct ast_cli_args *a)
{
struct ast_taskprocessor *tps;
struct ao2_iterator iter;
switch (cmd) {
case CLI_INIT:
e->command = "core reset taskprocessors";
e->usage =
"Usage: core reset taskprocessors\n"
" Resets stats for all taskprocessors\n";
return NULL;
case CLI_GENERATE:
return NULL;
}
if (a->argc != e->args) {
return CLI_SHOWUSAGE;
}
ast_cli(a->fd, "\nResetting stats for all taskprocessors\n\n");
iter = ao2_iterator_init(tps_singletons, 0);
while ((tps = ao2_iterator_next(&iter))) {
tps_reset_stats(tps);
ast_taskprocessor_unreference(tps);
}
ao2_iterator_destroy(&iter);
return CLI_SUCCESS;
}