kernel_optimize_test/kernel/rcu/srcutree.c

1380 lines
44 KiB
C
Raw Normal View History

// SPDX-License-Identifier: GPL-2.0+
/*
* Sleepable Read-Copy Update mechanism for mutual exclusion.
*
* Copyright (C) IBM Corporation, 2006
* Copyright (C) Fujitsu, 2012
*
* Author: Paul McKenney <paulmck@linux.ibm.com>
* Lai Jiangshan <laijs@cn.fujitsu.com>
*
* For detailed explanation of Read-Copy Update mechanism see -
* Documentation/RCU/ *.txt
*
*/
#define pr_fmt(fmt) "rcu: " fmt
#include <linux/export.h>
#include <linux/mutex.h>
#include <linux/percpu.h>
#include <linux/preempt.h>
#include <linux/rcupdate_wait.h>
#include <linux/sched.h>
#include <linux/smp.h>
#include <linux/delay.h>
#include <linux/module.h>
#include <linux/srcu.h>
#include "rcu.h"
#include "rcu_segcblist.h"
/* Holdoff in nanoseconds for auto-expediting. */
#define DEFAULT_SRCU_EXP_HOLDOFF (25 * 1000)
static ulong exp_holdoff = DEFAULT_SRCU_EXP_HOLDOFF;
module_param(exp_holdoff, ulong, 0444);
/* Overflow-check frequency. N bits roughly says every 2**N grace periods. */
static ulong counter_wrap_check = (ULONG_MAX >> 2);
module_param(counter_wrap_check, ulong, 0444);
srcu: Make call_srcu() available during very early boot Event tracing is moving to SRCU in order to take advantage of the fact that SRCU may be safely used from idle and even offline CPUs. However, event tracing can invoke call_srcu() very early in the boot process, even before workqueue_init_early() is invoked (let alone rcu_init()). Therefore, call_srcu()'s attempts to queue work fail miserably. This commit therefore detects this situation, and refrains from attempting to queue work before rcu_init() time, but does everything else that it would have done, and in addition, adds the srcu_struct to a global list. The rcu_init() function now invokes a new srcu_init() function, which is empty if CONFIG_SRCU=n. Otherwise, srcu_init() queues work for each srcu_struct on the list. This all happens early enough in boot that there is but a single CPU with interrupts disabled, which allows synchronization to be dispensed with. Of course, the queued work won't actually be invoked until after workqueue_init() is invoked, which happens shortly after the scheduler is up and running. This means that although call_srcu() may be invoked any time after per-CPU variables have been set up, there is still a very narrow window when synchronize_srcu() won't work, and this window extends from the time that the scheduler starts until the time that workqueue_init() returns. This can be fixed in a manner similar to the fix for synchronize_rcu_expedited() and friends, but until someone actually needs to use synchronize_srcu() during this window, this fix is added churn for no benefit. Finally, note that Tree SRCU's new srcu_init() function invokes queue_work() rather than the queue_delayed_work() function that is invoked post-boot. The reason is that queue_delayed_work() will (as you would expect) post a timer, and timers have not yet been initialized. So use of queue_work() avoids the complaints about use of uninitialized spinlocks that would otherwise result. Besides, some delay is already provide by the aforementioned fact that the queued work won't actually be invoked until after the scheduler is up and running. Requested-by: Steven Rostedt <rostedt@goodmis.org> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Tested-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
2018-08-14 23:45:54 +08:00
/* Early-boot callback-management, so early that no lock is required! */
static LIST_HEAD(srcu_boot_list);
static bool __read_mostly srcu_init_done;
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
static void srcu_invoke_callbacks(struct work_struct *work);
static void srcu_reschedule(struct srcu_struct *ssp, unsigned long delay);
static void process_srcu(struct work_struct *work);
static void srcu_delay_timer(struct timer_list *t);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
/* Wrappers for lock acquisition and release, see raw_spin_lock_rcu_node(). */
#define spin_lock_rcu_node(p) \
do { \
spin_lock(&ACCESS_PRIVATE(p, lock)); \
smp_mb__after_unlock_lock(); \
} while (0)
#define spin_unlock_rcu_node(p) spin_unlock(&ACCESS_PRIVATE(p, lock))
#define spin_lock_irq_rcu_node(p) \
do { \
spin_lock_irq(&ACCESS_PRIVATE(p, lock)); \
smp_mb__after_unlock_lock(); \
} while (0)
#define spin_unlock_irq_rcu_node(p) \
spin_unlock_irq(&ACCESS_PRIVATE(p, lock))
#define spin_lock_irqsave_rcu_node(p, flags) \
do { \
spin_lock_irqsave(&ACCESS_PRIVATE(p, lock), flags); \
smp_mb__after_unlock_lock(); \
} while (0)
#define spin_unlock_irqrestore_rcu_node(p, flags) \
spin_unlock_irqrestore(&ACCESS_PRIVATE(p, lock), flags) \
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
/*
* Initialize SRCU combining tree. Note that statically allocated
* srcu_struct structures might already have srcu_read_lock() and
* srcu_read_unlock() running against them. So if the is_static parameter
* is set, don't initialize ->srcu_lock_count[] and ->srcu_unlock_count[].
*/
static void init_srcu_struct_nodes(struct srcu_struct *ssp, bool is_static)
{
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
int cpu;
int i;
int level = 0;
int levelspread[RCU_NUM_LVLS];
struct srcu_data *sdp;
struct srcu_node *snp;
struct srcu_node *snp_first;
/* Work out the overall tree geometry. */
ssp->level[0] = &ssp->node[0];
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
for (i = 1; i < rcu_num_lvls; i++)
ssp->level[i] = ssp->level[i - 1] + num_rcu_lvl[i - 1];
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
rcu_init_levelspread(levelspread, num_rcu_lvl);
/* Each pass through this loop initializes one srcu_node structure. */
srcu_for_each_node_breadth_first(ssp, snp) {
spin_lock_init(&ACCESS_PRIVATE(snp, lock));
WARN_ON_ONCE(ARRAY_SIZE(snp->srcu_have_cbs) !=
ARRAY_SIZE(snp->srcu_data_have_cbs));
for (i = 0; i < ARRAY_SIZE(snp->srcu_have_cbs); i++) {
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
snp->srcu_have_cbs[i] = 0;
snp->srcu_data_have_cbs[i] = 0;
}
snp->srcu_gp_seq_needed_exp = 0;
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
snp->grplo = -1;
snp->grphi = -1;
if (snp == &ssp->node[0]) {
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
/* Root node, special case. */
snp->srcu_parent = NULL;
continue;
}
/* Non-root node. */
if (snp == ssp->level[level + 1])
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
level++;
snp->srcu_parent = ssp->level[level - 1] +
(snp - ssp->level[level]) /
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
levelspread[level - 1];
}
/*
* Initialize the per-CPU srcu_data array, which feeds into the
* leaves of the srcu_node tree.
*/
WARN_ON_ONCE(ARRAY_SIZE(sdp->srcu_lock_count) !=
ARRAY_SIZE(sdp->srcu_unlock_count));
level = rcu_num_lvls - 1;
snp_first = ssp->level[level];
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
for_each_possible_cpu(cpu) {
sdp = per_cpu_ptr(ssp->sda, cpu);
spin_lock_init(&ACCESS_PRIVATE(sdp, lock));
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
rcu_segcblist_init(&sdp->srcu_cblist);
sdp->srcu_cblist_invoking = false;
sdp->srcu_gp_seq_needed = ssp->srcu_gp_seq;
sdp->srcu_gp_seq_needed_exp = ssp->srcu_gp_seq;
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
sdp->mynode = &snp_first[cpu / levelspread[level]];
for (snp = sdp->mynode; snp != NULL; snp = snp->srcu_parent) {
if (snp->grplo < 0)
snp->grplo = cpu;
snp->grphi = cpu;
}
sdp->cpu = cpu;
INIT_WORK(&sdp->work, srcu_invoke_callbacks);
timer_setup(&sdp->delay_work, srcu_delay_timer, 0);
sdp->ssp = ssp;
sdp->grpmask = 1 << (cpu - sdp->mynode->grplo);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
if (is_static)
continue;
/* Dynamically allocated, better be no srcu_read_locks()! */
for (i = 0; i < ARRAY_SIZE(sdp->srcu_lock_count); i++) {
sdp->srcu_lock_count[i] = 0;
sdp->srcu_unlock_count[i] = 0;
}
}
}
/*
* Initialize non-compile-time initialized fields, including the
* associated srcu_node and srcu_data structures. The is_static
* parameter is passed through to init_srcu_struct_nodes(), and
* also tells us that ->sda has already been wired up to srcu_data.
*/
static int init_srcu_struct_fields(struct srcu_struct *ssp, bool is_static)
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
{
mutex_init(&ssp->srcu_cb_mutex);
mutex_init(&ssp->srcu_gp_mutex);
ssp->srcu_idx = 0;
ssp->srcu_gp_seq = 0;
ssp->srcu_barrier_seq = 0;
mutex_init(&ssp->srcu_barrier_mutex);
atomic_set(&ssp->srcu_barrier_cpu_cnt, 0);
INIT_DELAYED_WORK(&ssp->work, process_srcu);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
if (!is_static)
ssp->sda = alloc_percpu(struct srcu_data);
init_srcu_struct_nodes(ssp, is_static);
ssp->srcu_gp_seq_needed_exp = 0;
ssp->srcu_last_gp_end = ktime_get_mono_fast_ns();
smp_store_release(&ssp->srcu_gp_seq_needed, 0); /* Init done. */
return ssp->sda ? 0 : -ENOMEM;
}
#ifdef CONFIG_DEBUG_LOCK_ALLOC
int __init_srcu_struct(struct srcu_struct *ssp, const char *name,
struct lock_class_key *key)
{
/* Don't re-initialize a lock while it is held. */
debug_check_no_locks_freed((void *)ssp, sizeof(*ssp));
lockdep_init_map(&ssp->dep_map, name, key, 0);
spin_lock_init(&ACCESS_PRIVATE(ssp, lock));
return init_srcu_struct_fields(ssp, false);
}
EXPORT_SYMBOL_GPL(__init_srcu_struct);
#else /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
/**
* init_srcu_struct - initialize a sleep-RCU structure
* @ssp: structure to initialize.
*
* Must invoke this on a given srcu_struct before passing that srcu_struct
* to any other function. Each srcu_struct represents a separate domain
* of SRCU protection.
*/
int init_srcu_struct(struct srcu_struct *ssp)
{
spin_lock_init(&ACCESS_PRIVATE(ssp, lock));
return init_srcu_struct_fields(ssp, false);
}
EXPORT_SYMBOL_GPL(init_srcu_struct);
#endif /* #else #ifdef CONFIG_DEBUG_LOCK_ALLOC */
/*
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
* First-use initialization of statically allocated srcu_struct
* structure. Wiring up the combining tree is more than can be
* done with compile-time initialization, so this check is added
* to each update-side SRCU primitive. Use ssp->lock, which -is-
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
* compile-time initialized, to resolve races involving multiple
* CPUs trying to garner first-use privileges.
*/
static void check_init_srcu_struct(struct srcu_struct *ssp)
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
{
unsigned long flags;
/* The smp_load_acquire() pairs with the smp_store_release(). */
if (!rcu_seq_state(smp_load_acquire(&ssp->srcu_gp_seq_needed))) /*^^^*/
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
return; /* Already initialized. */
spin_lock_irqsave_rcu_node(ssp, flags);
if (!rcu_seq_state(ssp->srcu_gp_seq_needed)) {
spin_unlock_irqrestore_rcu_node(ssp, flags);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
return;
}
init_srcu_struct_fields(ssp, true);
spin_unlock_irqrestore_rcu_node(ssp, flags);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
}
/*
* Returns approximate total of the readers' ->srcu_lock_count[] values
* for the rank of per-CPU counters specified by idx.
*/
static unsigned long srcu_readers_lock_idx(struct srcu_struct *ssp, int idx)
{
int cpu;
unsigned long sum = 0;
for_each_possible_cpu(cpu) {
struct srcu_data *cpuc = per_cpu_ptr(ssp->sda, cpu);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
sum += READ_ONCE(cpuc->srcu_lock_count[idx]);
}
return sum;
}
/*
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
* Returns approximate total of the readers' ->srcu_unlock_count[] values
* for the rank of per-CPU counters specified by idx.
*/
static unsigned long srcu_readers_unlock_idx(struct srcu_struct *ssp, int idx)
{
int cpu;
unsigned long sum = 0;
for_each_possible_cpu(cpu) {
struct srcu_data *cpuc = per_cpu_ptr(ssp->sda, cpu);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
sum += READ_ONCE(cpuc->srcu_unlock_count[idx]);
}
return sum;
}
/*
* Return true if the number of pre-existing readers is determined to
* be zero.
*/
static bool srcu_readers_active_idx_check(struct srcu_struct *ssp, int idx)
{
unsigned long unlocks;
unlocks = srcu_readers_unlock_idx(ssp, idx);
/*
* Make sure that a lock is always counted if the corresponding
* unlock is counted. Needs to be a smp_mb() as the read side may
* contain a read from a variable that is written to before the
* synchronize_srcu() in the write side. In this case smp_mb()s
* A and B act like the store buffering pattern.
*
* This smp_mb() also pairs with smp_mb() C to prevent accesses
* after the synchronize_srcu() from being executed before the
* grace period ends.
*/
smp_mb(); /* A */
/*
* If the locks are the same as the unlocks, then there must have
* been no readers on this index at some time in between. This does
* not mean that there are no more readers, as one could have read
* the current index but not have incremented the lock counter yet.
*
* So suppose that the updater is preempted here for so long
* that more than ULONG_MAX non-nested readers come and go in
* the meantime. It turns out that this cannot result in overflow
* because if a reader modifies its unlock count after we read it
* above, then that reader's next load of ->srcu_idx is guaranteed
* to get the new value, which will cause it to operate on the
* other bank of counters, where it cannot contribute to the
* overflow of these counters. This means that there is a maximum
* of 2*NR_CPUS increments, which cannot overflow given current
* systems, especially not on 64-bit systems.
*
* OK, how about nesting? This does impose a limit on nesting
* of floor(ULONG_MAX/NR_CPUS/2), which should be sufficient,
* especially on 64-bit systems.
*/
return srcu_readers_lock_idx(ssp, idx) == unlocks;
}
/**
* srcu_readers_active - returns true if there are readers. and false
* otherwise
* @ssp: which srcu_struct to count active readers (holding srcu_read_lock).
*
* Note that this is not an atomic primitive, and can therefore suffer
* severe errors when invoked on an active srcu_struct. That said, it
* can be useful as an error check at cleanup time.
*/
static bool srcu_readers_active(struct srcu_struct *ssp)
{
int cpu;
unsigned long sum = 0;
for_each_possible_cpu(cpu) {
struct srcu_data *cpuc = per_cpu_ptr(ssp->sda, cpu);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
sum += READ_ONCE(cpuc->srcu_lock_count[0]);
sum += READ_ONCE(cpuc->srcu_lock_count[1]);
sum -= READ_ONCE(cpuc->srcu_unlock_count[0]);
sum -= READ_ONCE(cpuc->srcu_unlock_count[1]);
}
return sum;
}
#define SRCU_INTERVAL 1
/*
* Return grace-period delay, zero if there are expedited grace
* periods pending, SRCU_INTERVAL otherwise.
*/
static unsigned long srcu_get_delay(struct srcu_struct *ssp)
{
if (ULONG_CMP_LT(READ_ONCE(ssp->srcu_gp_seq),
READ_ONCE(ssp->srcu_gp_seq_needed_exp)))
return 0;
return SRCU_INTERVAL;
}
srcu: Remove cleanup_srcu_struct_quiesced() The cleanup_srcu_struct_quiesced() function was added because NVME used WQ_MEM_RECLAIM workqueues and SRCU did not, which meant that NVME workqueues waiting on SRCU workqueues could result in deadlocks during low-memory conditions. However, SRCU now also has WQ_MEM_RECLAIM workqueues, so there is no longer a potential for deadlock. Furthermore, it turns out to be extremely hard to use cleanup_srcu_struct_quiesced() correctly due to the fact that SRCU callback invocation accesses the srcu_struct structure's per-CPU data area just after callbacks are invoked. Therefore, the usual practice of using srcu_barrier() to wait for callbacks to be invoked before invoking cleanup_srcu_struct_quiesced() fails because SRCU's callback-invocation workqueue handler might be delayed, which can result in cleanup_srcu_struct_quiesced() being invoked (and thus freeing the per-CPU data) before the SRCU's callback-invocation workqueue handler is finished using that per-CPU data. Nor is this a theoretical problem: KASAN emitted use-after-free warnings because of this problem on actual runs. In short, NVME can now safely invoke cleanup_srcu_struct(), which avoids the use-after-free scenario. And cleanup_srcu_struct_quiesced() is quite difficult to use safely. This commit therefore removes cleanup_srcu_struct_quiesced(), switching its sole user back to cleanup_srcu_struct(). This effectively reverts the following pair of commits: f7194ac32ca2 ("srcu: Add cleanup_srcu_struct_quiesced()") 4317228ad9b8 ("nvme: Avoid flush dependency in delete controller flow") Reported-by: Bart Van Assche <bvanassche@acm.org> Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com> Reviewed-by: Bart Van Assche <bvanassche@acm.org> Tested-by: Bart Van Assche <bvanassche@acm.org>
2019-02-14 05:54:37 +08:00
/**
* cleanup_srcu_struct - deconstruct a sleep-RCU structure
* @ssp: structure to clean up.
*
* Must invoke this after you are finished using a given srcu_struct that
* was initialized via init_srcu_struct(), else you leak memory.
*/
void cleanup_srcu_struct(struct srcu_struct *ssp)
{
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
int cpu;
if (WARN_ON(!srcu_get_delay(ssp)))
return; /* Just leak it! */
if (WARN_ON(srcu_readers_active(ssp)))
return; /* Just leak it! */
srcu: Remove cleanup_srcu_struct_quiesced() The cleanup_srcu_struct_quiesced() function was added because NVME used WQ_MEM_RECLAIM workqueues and SRCU did not, which meant that NVME workqueues waiting on SRCU workqueues could result in deadlocks during low-memory conditions. However, SRCU now also has WQ_MEM_RECLAIM workqueues, so there is no longer a potential for deadlock. Furthermore, it turns out to be extremely hard to use cleanup_srcu_struct_quiesced() correctly due to the fact that SRCU callback invocation accesses the srcu_struct structure's per-CPU data area just after callbacks are invoked. Therefore, the usual practice of using srcu_barrier() to wait for callbacks to be invoked before invoking cleanup_srcu_struct_quiesced() fails because SRCU's callback-invocation workqueue handler might be delayed, which can result in cleanup_srcu_struct_quiesced() being invoked (and thus freeing the per-CPU data) before the SRCU's callback-invocation workqueue handler is finished using that per-CPU data. Nor is this a theoretical problem: KASAN emitted use-after-free warnings because of this problem on actual runs. In short, NVME can now safely invoke cleanup_srcu_struct(), which avoids the use-after-free scenario. And cleanup_srcu_struct_quiesced() is quite difficult to use safely. This commit therefore removes cleanup_srcu_struct_quiesced(), switching its sole user back to cleanup_srcu_struct(). This effectively reverts the following pair of commits: f7194ac32ca2 ("srcu: Add cleanup_srcu_struct_quiesced()") 4317228ad9b8 ("nvme: Avoid flush dependency in delete controller flow") Reported-by: Bart Van Assche <bvanassche@acm.org> Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com> Reviewed-by: Bart Van Assche <bvanassche@acm.org> Tested-by: Bart Van Assche <bvanassche@acm.org>
2019-02-14 05:54:37 +08:00
flush_delayed_work(&ssp->work);
for_each_possible_cpu(cpu) {
struct srcu_data *sdp = per_cpu_ptr(ssp->sda, cpu);
srcu: Remove cleanup_srcu_struct_quiesced() The cleanup_srcu_struct_quiesced() function was added because NVME used WQ_MEM_RECLAIM workqueues and SRCU did not, which meant that NVME workqueues waiting on SRCU workqueues could result in deadlocks during low-memory conditions. However, SRCU now also has WQ_MEM_RECLAIM workqueues, so there is no longer a potential for deadlock. Furthermore, it turns out to be extremely hard to use cleanup_srcu_struct_quiesced() correctly due to the fact that SRCU callback invocation accesses the srcu_struct structure's per-CPU data area just after callbacks are invoked. Therefore, the usual practice of using srcu_barrier() to wait for callbacks to be invoked before invoking cleanup_srcu_struct_quiesced() fails because SRCU's callback-invocation workqueue handler might be delayed, which can result in cleanup_srcu_struct_quiesced() being invoked (and thus freeing the per-CPU data) before the SRCU's callback-invocation workqueue handler is finished using that per-CPU data. Nor is this a theoretical problem: KASAN emitted use-after-free warnings because of this problem on actual runs. In short, NVME can now safely invoke cleanup_srcu_struct(), which avoids the use-after-free scenario. And cleanup_srcu_struct_quiesced() is quite difficult to use safely. This commit therefore removes cleanup_srcu_struct_quiesced(), switching its sole user back to cleanup_srcu_struct(). This effectively reverts the following pair of commits: f7194ac32ca2 ("srcu: Add cleanup_srcu_struct_quiesced()") 4317228ad9b8 ("nvme: Avoid flush dependency in delete controller flow") Reported-by: Bart Van Assche <bvanassche@acm.org> Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com> Reviewed-by: Bart Van Assche <bvanassche@acm.org> Tested-by: Bart Van Assche <bvanassche@acm.org>
2019-02-14 05:54:37 +08:00
del_timer_sync(&sdp->delay_work);
flush_work(&sdp->work);
if (WARN_ON(rcu_segcblist_n_cbs(&sdp->srcu_cblist)))
return; /* Forgot srcu_barrier(), so just leak it! */
}
if (WARN_ON(rcu_seq_state(READ_ONCE(ssp->srcu_gp_seq)) != SRCU_STATE_IDLE) ||
WARN_ON(srcu_readers_active(ssp))) {
pr_info("%s: Active srcu_struct %p state: %d\n",
__func__, ssp, rcu_seq_state(READ_ONCE(ssp->srcu_gp_seq)));
return; /* Caller forgot to stop doing call_srcu()? */
}
free_percpu(ssp->sda);
ssp->sda = NULL;
}
srcu: Remove cleanup_srcu_struct_quiesced() The cleanup_srcu_struct_quiesced() function was added because NVME used WQ_MEM_RECLAIM workqueues and SRCU did not, which meant that NVME workqueues waiting on SRCU workqueues could result in deadlocks during low-memory conditions. However, SRCU now also has WQ_MEM_RECLAIM workqueues, so there is no longer a potential for deadlock. Furthermore, it turns out to be extremely hard to use cleanup_srcu_struct_quiesced() correctly due to the fact that SRCU callback invocation accesses the srcu_struct structure's per-CPU data area just after callbacks are invoked. Therefore, the usual practice of using srcu_barrier() to wait for callbacks to be invoked before invoking cleanup_srcu_struct_quiesced() fails because SRCU's callback-invocation workqueue handler might be delayed, which can result in cleanup_srcu_struct_quiesced() being invoked (and thus freeing the per-CPU data) before the SRCU's callback-invocation workqueue handler is finished using that per-CPU data. Nor is this a theoretical problem: KASAN emitted use-after-free warnings because of this problem on actual runs. In short, NVME can now safely invoke cleanup_srcu_struct(), which avoids the use-after-free scenario. And cleanup_srcu_struct_quiesced() is quite difficult to use safely. This commit therefore removes cleanup_srcu_struct_quiesced(), switching its sole user back to cleanup_srcu_struct(). This effectively reverts the following pair of commits: f7194ac32ca2 ("srcu: Add cleanup_srcu_struct_quiesced()") 4317228ad9b8 ("nvme: Avoid flush dependency in delete controller flow") Reported-by: Bart Van Assche <bvanassche@acm.org> Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com> Reviewed-by: Bart Van Assche <bvanassche@acm.org> Tested-by: Bart Van Assche <bvanassche@acm.org>
2019-02-14 05:54:37 +08:00
EXPORT_SYMBOL_GPL(cleanup_srcu_struct);
/*
* Counts the new reader in the appropriate per-CPU element of the
srcu: Allow use of Tiny/Tree SRCU from both process and interrupt context Linu Cherian reported a WARN in cleanup_srcu_struct() when shutting down a guest running iperf on a VFIO assigned device. This happens because irqfd_wakeup() calls srcu_read_lock(&kvm->irq_srcu) in interrupt context, while a worker thread does the same inside kvm_set_irq(). If the interrupt happens while the worker thread is executing __srcu_read_lock(), updates to the Classic SRCU ->lock_count[] field or the Tree SRCU ->srcu_lock_count[] field can be lost. The docs say you are not supposed to call srcu_read_lock() and srcu_read_unlock() from irq context, but KVM interrupt injection happens from (host) interrupt context and it would be nice if SRCU supported the use case. KVM is using SRCU here not really for the "sleepable" part, but rather due to its IPI-free fast detection of grace periods. It is therefore not desirable to switch back to RCU, which would effectively revert commit 719d93cd5f5c ("kvm/irqchip: Speed up KVM_SET_GSI_ROUTING", 2014-01-16). However, the docs are overly conservative. You can have an SRCU instance only has users in irq context, and you can mix process and irq context as long as process context users disable interrupts. In addition, __srcu_read_unlock() actually uses this_cpu_dec() on both Tree SRCU and Classic SRCU. For those two implementations, only srcu_read_lock() is unsafe. When Classic SRCU's __srcu_read_unlock() was changed to use this_cpu_dec(), in commit 5a41344a3d83 ("srcu: Simplify __srcu_read_unlock() via this_cpu_dec()", 2012-11-29), __srcu_read_lock() did two increments. Therefore it kept __this_cpu_inc(), with preempt_disable/enable in the caller. Tree SRCU however only does one increment, so on most architectures it is more efficient for __srcu_read_lock() to use this_cpu_inc(), and any performance differences appear to be down in the noise. Unlike Classic and Tree SRCU, Tiny SRCU does increments and decrements on a single variable. Therefore, as Peter Zijlstra pointed out, Tiny SRCU's implementation already supports mixed-context use of srcu_read_lock() and srcu_read_unlock(), at least as long as uses of srcu_read_lock() and srcu_read_unlock() in each handler are nested and paired properly. In other words, it is still illegal to (say) invoke srcu_read_lock() in an interrupt handler and to invoke the matching srcu_read_unlock() in a softirq handler. Therefore, the only change required for Tiny SRCU is to its comments. Fixes: 719d93cd5f5c ("kvm/irqchip: Speed up KVM_SET_GSI_ROUTING") Reported-by: Linu Cherian <linuc.decode@gmail.com> Suggested-by: Linu Cherian <linuc.decode@gmail.com> Cc: kvm@vger.kernel.org Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Tested-by: Paolo Bonzini <pbonzini@redhat.com>
2017-05-31 20:03:10 +08:00
* srcu_struct.
* Returns an index that must be passed to the matching srcu_read_unlock().
*/
int __srcu_read_lock(struct srcu_struct *ssp)
{
int idx;
idx = READ_ONCE(ssp->srcu_idx) & 0x1;
this_cpu_inc(ssp->sda->srcu_lock_count[idx]);
smp_mb(); /* B */ /* Avoid leaking the critical section. */
return idx;
}
EXPORT_SYMBOL_GPL(__srcu_read_lock);
/*
* Removes the count for the old reader from the appropriate per-CPU
* element of the srcu_struct. Note that this may well be a different
* CPU than that which was incremented by the corresponding srcu_read_lock().
*/
void __srcu_read_unlock(struct srcu_struct *ssp, int idx)
{
smp_mb(); /* C */ /* Avoid leaking the critical section. */
this_cpu_inc(ssp->sda->srcu_unlock_count[idx]);
}
EXPORT_SYMBOL_GPL(__srcu_read_unlock);
/*
* We use an adaptive strategy for synchronize_srcu() and especially for
* synchronize_srcu_expedited(). We spin for a fixed time period
* (defined below) to allow SRCU readers to exit their read-side critical
* sections. If there are still some readers after a few microseconds,
* we repeatedly block for 1-millisecond time periods.
*/
#define SRCU_RETRY_CHECK_DELAY 5
/*
* Start an SRCU grace period.
*/
static void srcu_gp_start(struct srcu_struct *ssp)
{
struct srcu_data *sdp = this_cpu_ptr(ssp->sda);
int state;
lockdep_assert_held(&ACCESS_PRIVATE(ssp, lock));
WARN_ON_ONCE(ULONG_CMP_GE(ssp->srcu_gp_seq, ssp->srcu_gp_seq_needed));
spin_lock_rcu_node(sdp); /* Interrupts already disabled. */
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
rcu_segcblist_advance(&sdp->srcu_cblist,
rcu_seq_current(&ssp->srcu_gp_seq));
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
(void)rcu_segcblist_accelerate(&sdp->srcu_cblist,
rcu_seq_snap(&ssp->srcu_gp_seq));
spin_unlock_rcu_node(sdp); /* Interrupts remain disabled. */
smp_mb(); /* Order prior store to ->srcu_gp_seq_needed vs. GP start. */
rcu_seq_start(&ssp->srcu_gp_seq);
state = rcu_seq_state(READ_ONCE(ssp->srcu_gp_seq));
WARN_ON_ONCE(state != SRCU_STATE_SCAN1);
}
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
static void srcu_delay_timer(struct timer_list *t)
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
{
struct srcu_data *sdp = container_of(t, struct srcu_data, delay_work);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
queue_work_on(sdp->cpu, rcu_gp_wq, &sdp->work);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
}
static void srcu_queue_delayed_work_on(struct srcu_data *sdp,
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
unsigned long delay)
{
if (!delay) {
queue_work_on(sdp->cpu, rcu_gp_wq, &sdp->work);
return;
}
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
timer_reduce(&sdp->delay_work, jiffies + delay);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
}
/*
* Schedule callback invocation for the specified srcu_data structure,
* if possible, on the corresponding CPU.
*/
static void srcu_schedule_cbs_sdp(struct srcu_data *sdp, unsigned long delay)
{
srcu_queue_delayed_work_on(sdp, delay);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
}
/*
* Schedule callback invocation for all srcu_data structures associated
* with the specified srcu_node structure that have callbacks for the
* just-completed grace period, the one corresponding to idx. If possible,
* schedule this invocation on the corresponding CPUs.
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
*/
static void srcu_schedule_cbs_snp(struct srcu_struct *ssp, struct srcu_node *snp,
unsigned long mask, unsigned long delay)
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
{
int cpu;
for (cpu = snp->grplo; cpu <= snp->grphi; cpu++) {
if (!(mask & (1 << (cpu - snp->grplo))))
continue;
srcu_schedule_cbs_sdp(per_cpu_ptr(ssp->sda, cpu), delay);
}
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
}
/*
* Note the end of an SRCU grace period. Initiates callback invocation
* and starts a new grace period if needed.
*
* The ->srcu_cb_mutex acquisition does not protect any data, but
* instead prevents more than one grace period from starting while we
* are initiating callback invocation. This allows the ->srcu_have_cbs[]
* array to have a finite number of elements.
*/
static void srcu_gp_end(struct srcu_struct *ssp)
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
{
unsigned long cbdelay;
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
bool cbs;
bool last_lvl;
int cpu;
unsigned long flags;
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
unsigned long gpseq;
int idx;
unsigned long mask;
struct srcu_data *sdp;
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
struct srcu_node *snp;
/* Prevent more than one additional grace period. */
mutex_lock(&ssp->srcu_cb_mutex);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
/* End the current grace period. */
spin_lock_irq_rcu_node(ssp);
idx = rcu_seq_state(ssp->srcu_gp_seq);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
WARN_ON_ONCE(idx != SRCU_STATE_SCAN2);
cbdelay = srcu_get_delay(ssp);
WRITE_ONCE(ssp->srcu_last_gp_end, ktime_get_mono_fast_ns());
rcu_seq_end(&ssp->srcu_gp_seq);
gpseq = rcu_seq_current(&ssp->srcu_gp_seq);
if (ULONG_CMP_LT(ssp->srcu_gp_seq_needed_exp, gpseq))
ssp->srcu_gp_seq_needed_exp = gpseq;
spin_unlock_irq_rcu_node(ssp);
mutex_unlock(&ssp->srcu_gp_mutex);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
/* A new grace period can start at this point. But only one. */
/* Initiate callback invocation as needed. */
idx = rcu_seq_ctr(gpseq) % ARRAY_SIZE(snp->srcu_have_cbs);
srcu_for_each_node_breadth_first(ssp, snp) {
spin_lock_irq_rcu_node(snp);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
cbs = false;
last_lvl = snp >= ssp->level[rcu_num_lvls - 1];
if (last_lvl)
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
cbs = snp->srcu_have_cbs[idx] == gpseq;
snp->srcu_have_cbs[idx] = gpseq;
rcu_seq_set_state(&snp->srcu_have_cbs[idx], 1);
if (ULONG_CMP_LT(snp->srcu_gp_seq_needed_exp, gpseq))
snp->srcu_gp_seq_needed_exp = gpseq;
mask = snp->srcu_data_have_cbs[idx];
snp->srcu_data_have_cbs[idx] = 0;
spin_unlock_irq_rcu_node(snp);
if (cbs)
srcu_schedule_cbs_snp(ssp, snp, mask, cbdelay);
/* Occasionally prevent srcu_data counter wrap. */
if (!(gpseq & counter_wrap_check) && last_lvl)
for (cpu = snp->grplo; cpu <= snp->grphi; cpu++) {
sdp = per_cpu_ptr(ssp->sda, cpu);
spin_lock_irqsave_rcu_node(sdp, flags);
if (ULONG_CMP_GE(gpseq,
sdp->srcu_gp_seq_needed + 100))
sdp->srcu_gp_seq_needed = gpseq;
if (ULONG_CMP_GE(gpseq,
sdp->srcu_gp_seq_needed_exp + 100))
sdp->srcu_gp_seq_needed_exp = gpseq;
spin_unlock_irqrestore_rcu_node(sdp, flags);
}
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
}
/* Callback initiation done, allow grace periods after next. */
mutex_unlock(&ssp->srcu_cb_mutex);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
/* Start a new grace period if needed. */
spin_lock_irq_rcu_node(ssp);
gpseq = rcu_seq_current(&ssp->srcu_gp_seq);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
if (!rcu_seq_state(gpseq) &&
ULONG_CMP_LT(gpseq, ssp->srcu_gp_seq_needed)) {
srcu_gp_start(ssp);
spin_unlock_irq_rcu_node(ssp);
srcu_reschedule(ssp, 0);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
} else {
spin_unlock_irq_rcu_node(ssp);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
}
}
/*
* Funnel-locking scheme to scalably mediate many concurrent expedited
* grace-period requests. This function is invoked for the first known
* expedited request for a grace period that has already been requested,
* but without expediting. To start a completely new grace period,
* whether expedited or not, use srcu_funnel_gp_start() instead.
*/
static void srcu_funnel_exp_start(struct srcu_struct *ssp, struct srcu_node *snp,
unsigned long s)
{
unsigned long flags;
for (; snp != NULL; snp = snp->srcu_parent) {
if (rcu_seq_done(&ssp->srcu_gp_seq, s) ||
ULONG_CMP_GE(READ_ONCE(snp->srcu_gp_seq_needed_exp), s))
return;
spin_lock_irqsave_rcu_node(snp, flags);
if (ULONG_CMP_GE(snp->srcu_gp_seq_needed_exp, s)) {
spin_unlock_irqrestore_rcu_node(snp, flags);
return;
}
WRITE_ONCE(snp->srcu_gp_seq_needed_exp, s);
spin_unlock_irqrestore_rcu_node(snp, flags);
}
spin_lock_irqsave_rcu_node(ssp, flags);
if (ULONG_CMP_LT(ssp->srcu_gp_seq_needed_exp, s))
ssp->srcu_gp_seq_needed_exp = s;
spin_unlock_irqrestore_rcu_node(ssp, flags);
}
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
/*
* Funnel-locking scheme to scalably mediate many concurrent grace-period
* requests. The winner has to do the work of actually starting grace
* period s. Losers must either ensure that their desired grace-period
* number is recorded on at least their leaf srcu_node structure, or they
* must take steps to invoke their own callbacks.
*
* Note that this function also does the work of srcu_funnel_exp_start(),
* in some cases by directly invoking it.
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
*/
static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp,
unsigned long s, bool do_norm)
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
{
unsigned long flags;
int idx = rcu_seq_ctr(s) % ARRAY_SIZE(sdp->mynode->srcu_have_cbs);
struct srcu_node *snp = sdp->mynode;
unsigned long snp_seq;
/* Each pass through the loop does one level of the srcu_node tree. */
for (; snp != NULL; snp = snp->srcu_parent) {
if (rcu_seq_done(&ssp->srcu_gp_seq, s) && snp != sdp->mynode)
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
return; /* GP already done and CBs recorded. */
spin_lock_irqsave_rcu_node(snp, flags);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
if (ULONG_CMP_GE(snp->srcu_have_cbs[idx], s)) {
snp_seq = snp->srcu_have_cbs[idx];
if (snp == sdp->mynode && snp_seq == s)
snp->srcu_data_have_cbs[idx] |= sdp->grpmask;
spin_unlock_irqrestore_rcu_node(snp, flags);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
if (snp == sdp->mynode && snp_seq != s) {
srcu_schedule_cbs_sdp(sdp, do_norm
? SRCU_INTERVAL
: 0);
return;
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
}
if (!do_norm)
srcu_funnel_exp_start(ssp, snp, s);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
return;
}
snp->srcu_have_cbs[idx] = s;
if (snp == sdp->mynode)
snp->srcu_data_have_cbs[idx] |= sdp->grpmask;
if (!do_norm && ULONG_CMP_LT(snp->srcu_gp_seq_needed_exp, s))
snp->srcu_gp_seq_needed_exp = s;
spin_unlock_irqrestore_rcu_node(snp, flags);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
}
/* Top of tree, must ensure the grace period will be started. */
spin_lock_irqsave_rcu_node(ssp, flags);
if (ULONG_CMP_LT(ssp->srcu_gp_seq_needed, s)) {
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
/*
* Record need for grace period s. Pair with load
* acquire setting up for initialization.
*/
smp_store_release(&ssp->srcu_gp_seq_needed, s); /*^^^*/
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
}
if (!do_norm && ULONG_CMP_LT(ssp->srcu_gp_seq_needed_exp, s))
ssp->srcu_gp_seq_needed_exp = s;
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
/* If grace period not already done and none in progress, start it. */
if (!rcu_seq_done(&ssp->srcu_gp_seq, s) &&
rcu_seq_state(ssp->srcu_gp_seq) == SRCU_STATE_IDLE) {
WARN_ON_ONCE(ULONG_CMP_GE(ssp->srcu_gp_seq, ssp->srcu_gp_seq_needed));
srcu_gp_start(ssp);
srcu: Make call_srcu() available during very early boot Event tracing is moving to SRCU in order to take advantage of the fact that SRCU may be safely used from idle and even offline CPUs. However, event tracing can invoke call_srcu() very early in the boot process, even before workqueue_init_early() is invoked (let alone rcu_init()). Therefore, call_srcu()'s attempts to queue work fail miserably. This commit therefore detects this situation, and refrains from attempting to queue work before rcu_init() time, but does everything else that it would have done, and in addition, adds the srcu_struct to a global list. The rcu_init() function now invokes a new srcu_init() function, which is empty if CONFIG_SRCU=n. Otherwise, srcu_init() queues work for each srcu_struct on the list. This all happens early enough in boot that there is but a single CPU with interrupts disabled, which allows synchronization to be dispensed with. Of course, the queued work won't actually be invoked until after workqueue_init() is invoked, which happens shortly after the scheduler is up and running. This means that although call_srcu() may be invoked any time after per-CPU variables have been set up, there is still a very narrow window when synchronize_srcu() won't work, and this window extends from the time that the scheduler starts until the time that workqueue_init() returns. This can be fixed in a manner similar to the fix for synchronize_rcu_expedited() and friends, but until someone actually needs to use synchronize_srcu() during this window, this fix is added churn for no benefit. Finally, note that Tree SRCU's new srcu_init() function invokes queue_work() rather than the queue_delayed_work() function that is invoked post-boot. The reason is that queue_delayed_work() will (as you would expect) post a timer, and timers have not yet been initialized. So use of queue_work() avoids the complaints about use of uninitialized spinlocks that would otherwise result. Besides, some delay is already provide by the aforementioned fact that the queued work won't actually be invoked until after the scheduler is up and running. Requested-by: Steven Rostedt <rostedt@goodmis.org> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Tested-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
2018-08-14 23:45:54 +08:00
if (likely(srcu_init_done))
queue_delayed_work(rcu_gp_wq, &ssp->work,
srcu_get_delay(ssp));
else if (list_empty(&ssp->work.work.entry))
list_add(&ssp->work.work.entry, &srcu_boot_list);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
}
spin_unlock_irqrestore_rcu_node(ssp, flags);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
}
/*
* Wait until all readers counted by array index idx complete, but
* loop an additional time if there is an expedited grace period pending.
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
* The caller must ensure that ->srcu_idx is not changed while checking.
*/
static bool try_check_zero(struct srcu_struct *ssp, int idx, int trycount)
{
for (;;) {
if (srcu_readers_active_idx_check(ssp, idx))
return true;
if (--trycount + !srcu_get_delay(ssp) <= 0)
return false;
udelay(SRCU_RETRY_CHECK_DELAY);
}
}
/*
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
* Increment the ->srcu_idx counter so that future SRCU readers will
* use the other rank of the ->srcu_(un)lock_count[] arrays. This allows
* us to wait for pre-existing readers in a starvation-free manner.
*/
static void srcu_flip(struct srcu_struct *ssp)
{
/*
* Ensure that if this updater saw a given reader's increment
* from __srcu_read_lock(), that reader was using an old value
* of ->srcu_idx. Also ensure that if a given reader sees the
* new value of ->srcu_idx, this updater's earlier scans cannot
* have seen that reader's increments (which is OK, because this
* grace period need not wait on that reader).
*/
smp_mb(); /* E */ /* Pairs with B and C. */
WRITE_ONCE(ssp->srcu_idx, ssp->srcu_idx + 1);
/*
* Ensure that if the updater misses an __srcu_read_unlock()
* increment, that task's next __srcu_read_lock() will see the
* above counter update. Note that both this memory barrier
* and the one in srcu_readers_active_idx_check() provide the
* guarantee for __srcu_read_lock().
*/
smp_mb(); /* D */ /* Pairs with C. */
}
/*
* If SRCU is likely idle, return true, otherwise return false.
*
* Note that it is OK for several current from-idle requests for a new
* grace period from idle to specify expediting because they will all end
* up requesting the same grace period anyhow. So no loss.
*
* Note also that if any CPU (including the current one) is still invoking
* callbacks, this function will nevertheless say "idle". This is not
* ideal, but the overhead of checking all CPUs' callback lists is even
* less ideal, especially on large systems. Furthermore, the wakeup
* can happen before the callback is fully removed, so we have no choice
* but to accept this type of error.
*
* This function is also subject to counter-wrap errors, but let's face
* it, if this function was preempted for enough time for the counters
* to wrap, it really doesn't matter whether or not we expedite the grace
* period. The extra overhead of a needlessly expedited grace period is
* negligible when amoritized over that time period, and the extra latency
* of a needlessly non-expedited grace period is similarly negligible.
*/
static bool srcu_might_be_idle(struct srcu_struct *ssp)
{
unsigned long curseq;
unsigned long flags;
struct srcu_data *sdp;
unsigned long t;
unsigned long tlast;
/* If the local srcu_data structure has callbacks, not idle. */
local_irq_save(flags);
sdp = this_cpu_ptr(ssp->sda);
if (rcu_segcblist_pend_cbs(&sdp->srcu_cblist)) {
local_irq_restore(flags);
return false; /* Callbacks already present, so not idle. */
}
local_irq_restore(flags);
/*
* No local callbacks, so probabalistically probe global state.
* Exact information would require acquiring locks, which would
* kill scalability, hence the probabalistic nature of the probe.
*/
/* First, see if enough time has passed since the last GP. */
t = ktime_get_mono_fast_ns();
tlast = READ_ONCE(ssp->srcu_last_gp_end);
if (exp_holdoff == 0 ||
time_in_range_open(t, tlast, tlast + exp_holdoff))
return false; /* Too soon after last GP. */
/* Next, check for probable idleness. */
curseq = rcu_seq_current(&ssp->srcu_gp_seq);
smp_mb(); /* Order ->srcu_gp_seq with ->srcu_gp_seq_needed. */
if (ULONG_CMP_LT(curseq, READ_ONCE(ssp->srcu_gp_seq_needed)))
return false; /* Grace period in progress, so not idle. */
smp_mb(); /* Order ->srcu_gp_seq with prior access. */
if (curseq != rcu_seq_current(&ssp->srcu_gp_seq))
return false; /* GP # changed, so not idle. */
return true; /* With reasonable probability, idle! */
}
/*
* SRCU callback function to leak a callback.
*/
static void srcu_leak_callback(struct rcu_head *rhp)
{
}
/*
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
* Enqueue an SRCU callback on the srcu_data structure associated with
* the current CPU and the specified srcu_struct structure, initiating
* grace-period processing if it is not already running.
*
* Note that all CPUs must agree that the grace period extended beyond
* all pre-existing SRCU read-side critical section. On systems with
* more than one CPU, this means that when "func()" is invoked, each CPU
* is guaranteed to have executed a full memory barrier since the end of
* its last corresponding SRCU read-side critical section whose beginning
* preceded the call to call_srcu(). It also means that each CPU executing
* an SRCU read-side critical section that continues beyond the start of
* "func()" must have executed a memory barrier after the call_srcu()
* but before the beginning of that SRCU read-side critical section.
* Note that these guarantees include CPUs that are offline, idle, or
* executing in user mode, as well as CPUs that are executing in the kernel.
*
* Furthermore, if CPU A invoked call_srcu() and CPU B invoked the
* resulting SRCU callback function "func()", then both CPU A and CPU
* B are guaranteed to execute a full memory barrier during the time
* interval between the call to call_srcu() and the invocation of "func()".
* This guarantee applies even if CPU A and CPU B are the same CPU (but
* again only if the system has more than one CPU).
*
* Of course, these guarantees apply only for invocations of call_srcu(),
* srcu_read_lock(), and srcu_read_unlock() that are all passed the same
* srcu_struct structure.
*/
static void __call_srcu(struct srcu_struct *ssp, struct rcu_head *rhp,
rcu_callback_t func, bool do_norm)
{
unsigned long flags;
int idx;
bool needexp = false;
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
bool needgp = false;
unsigned long s;
struct srcu_data *sdp;
check_init_srcu_struct(ssp);
if (debug_rcu_head_queue(rhp)) {
/* Probable double call_srcu(), so leak the callback. */
WRITE_ONCE(rhp->func, srcu_leak_callback);
WARN_ONCE(1, "call_srcu(): Leaked duplicate callback\n");
return;
}
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
rhp->func = func;
idx = srcu_read_lock(ssp);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
local_irq_save(flags);
sdp = this_cpu_ptr(ssp->sda);
spin_lock_rcu_node(sdp);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
rcu_segcblist_enqueue(&sdp->srcu_cblist, rhp, false);
rcu_segcblist_advance(&sdp->srcu_cblist,
rcu_seq_current(&ssp->srcu_gp_seq));
s = rcu_seq_snap(&ssp->srcu_gp_seq);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
(void)rcu_segcblist_accelerate(&sdp->srcu_cblist, s);
if (ULONG_CMP_LT(sdp->srcu_gp_seq_needed, s)) {
sdp->srcu_gp_seq_needed = s;
needgp = true;
}
if (!do_norm && ULONG_CMP_LT(sdp->srcu_gp_seq_needed_exp, s)) {
sdp->srcu_gp_seq_needed_exp = s;
needexp = true;
}
spin_unlock_irqrestore_rcu_node(sdp, flags);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
if (needgp)
srcu_funnel_gp_start(ssp, sdp, s, do_norm);
else if (needexp)
srcu_funnel_exp_start(ssp, sdp->mynode, s);
srcu_read_unlock(ssp, idx);
}
/**
* call_srcu() - Queue a callback for invocation after an SRCU grace period
* @ssp: srcu_struct in queue the callback
* @rhp: structure to be used for queueing the SRCU callback.
* @func: function to be invoked after the SRCU grace period
*
* The callback function will be invoked some time after a full SRCU
* grace period elapses, in other words after all pre-existing SRCU
* read-side critical sections have completed. However, the callback
* function might well execute concurrently with other SRCU read-side
* critical sections that started after call_srcu() was invoked. SRCU
* read-side critical sections are delimited by srcu_read_lock() and
* srcu_read_unlock(), and may be nested.
*
* The callback will be invoked from process context, but must nevertheless
* be fast and must not block.
*/
void call_srcu(struct srcu_struct *ssp, struct rcu_head *rhp,
rcu_callback_t func)
{
__call_srcu(ssp, rhp, func, true);
}
EXPORT_SYMBOL_GPL(call_srcu);
/*
* Helper function for synchronize_srcu() and synchronize_srcu_expedited().
*/
static void __synchronize_srcu(struct srcu_struct *ssp, bool do_norm)
{
struct rcu_synchronize rcu;
RCU_LOCKDEP_WARN(lock_is_held(&ssp->dep_map) ||
lock_is_held(&rcu_bh_lock_map) ||
lock_is_held(&rcu_lock_map) ||
lock_is_held(&rcu_sched_lock_map),
"Illegal synchronize_srcu() in same-type SRCU (or in RCU) read-side critical section");
if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
return;
might_sleep();
check_init_srcu_struct(ssp);
init_completion(&rcu.completion);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
init_rcu_head_on_stack(&rcu.head);
__call_srcu(ssp, &rcu.head, wakeme_after_rcu, do_norm);
wait_for_completion(&rcu.completion);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
destroy_rcu_head_on_stack(&rcu.head);
srcu: Provide ordering for CPU not involved in grace period Tree RCU guarantees that every online CPU has a memory barrier between any given grace period and any of that CPU's RCU read-side sections that must be ordered against that grace period. Since RCU doesn't always know where read-side critical sections are, the actual implementation guarantees order against prior and subsequent non-idle non-offline code, whether in an RCU read-side critical section or not. As a result, there does not need to be a memory barrier at the end of synchronize_rcu() and friends because the ordering internal to the grace period has ordered every CPU's post-grace-period execution against each CPU's pre-grace-period execution, again for all non-idle online CPUs. In contrast, SRCU can have non-idle online CPUs that are completely uninvolved in a given SRCU grace period, for example, a CPU that never runs any SRCU read-side critical sections and took no part in the grace-period processing. It is in theory possible for a given synchronize_srcu()'s wakeup to be delivered to a CPU that was completely uninvolved in the prior SRCU grace period, which could mean that the code following that synchronize_srcu() would end up being unordered with respect to both the grace period and any pre-existing SRCU read-side critical sections. This commit therefore adds an smp_mb() to the end of __synchronize_srcu(), which prevents this scenario from occurring. Reported-by: Lance Roy <ldr709@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Acked-by: Lance Roy <ldr709@gmail.com> Cc: <stable@vger.kernel.org> # 4.12.x
2017-07-06 04:30:21 +08:00
/*
* Make sure that later code is ordered after the SRCU grace
* period. This pairs with the spin_lock_irq_rcu_node()
srcu: Provide ordering for CPU not involved in grace period Tree RCU guarantees that every online CPU has a memory barrier between any given grace period and any of that CPU's RCU read-side sections that must be ordered against that grace period. Since RCU doesn't always know where read-side critical sections are, the actual implementation guarantees order against prior and subsequent non-idle non-offline code, whether in an RCU read-side critical section or not. As a result, there does not need to be a memory barrier at the end of synchronize_rcu() and friends because the ordering internal to the grace period has ordered every CPU's post-grace-period execution against each CPU's pre-grace-period execution, again for all non-idle online CPUs. In contrast, SRCU can have non-idle online CPUs that are completely uninvolved in a given SRCU grace period, for example, a CPU that never runs any SRCU read-side critical sections and took no part in the grace-period processing. It is in theory possible for a given synchronize_srcu()'s wakeup to be delivered to a CPU that was completely uninvolved in the prior SRCU grace period, which could mean that the code following that synchronize_srcu() would end up being unordered with respect to both the grace period and any pre-existing SRCU read-side critical sections. This commit therefore adds an smp_mb() to the end of __synchronize_srcu(), which prevents this scenario from occurring. Reported-by: Lance Roy <ldr709@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Acked-by: Lance Roy <ldr709@gmail.com> Cc: <stable@vger.kernel.org> # 4.12.x
2017-07-06 04:30:21 +08:00
* in srcu_invoke_callbacks(). Unlike Tree RCU, this is needed
* because the current CPU might have been totally uninvolved with
* (and thus unordered against) that grace period.
*/
smp_mb();
}
/**
* synchronize_srcu_expedited - Brute-force SRCU grace period
* @ssp: srcu_struct with which to synchronize.
*
* Wait for an SRCU grace period to elapse, but be more aggressive about
* spinning rather than blocking when waiting.
*
* Note that synchronize_srcu_expedited() has the same deadlock and
* memory-ordering properties as does synchronize_srcu().
*/
void synchronize_srcu_expedited(struct srcu_struct *ssp)
{
__synchronize_srcu(ssp, rcu_gp_is_normal());
}
EXPORT_SYMBOL_GPL(synchronize_srcu_expedited);
/**
* synchronize_srcu - wait for prior SRCU read-side critical-section completion
* @ssp: srcu_struct with which to synchronize.
*
* Wait for the count to drain to zero of both indexes. To avoid the
* possible starvation of synchronize_srcu(), it waits for the count of
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
* the index=((->srcu_idx & 1) ^ 1) to drain to zero at first,
* and then flip the srcu_idx and wait for the count of the other index.
*
* Can block; must be called from process context.
*
* Note that it is illegal to call synchronize_srcu() from the corresponding
* SRCU read-side critical section; doing so will result in deadlock.
* However, it is perfectly legal to call synchronize_srcu() on one
* srcu_struct from some other srcu_struct's read-side critical section,
* as long as the resulting graph of srcu_structs is acyclic.
*
* There are memory-ordering constraints implied by synchronize_srcu().
* On systems with more than one CPU, when synchronize_srcu() returns,
* each CPU is guaranteed to have executed a full memory barrier since
* the end of its last corresponding SRCU read-side critical section
* whose beginning preceded the call to synchronize_srcu(). In addition,
* each CPU having an SRCU read-side critical section that extends beyond
* the return from synchronize_srcu() is guaranteed to have executed a
* full memory barrier after the beginning of synchronize_srcu() and before
* the beginning of that SRCU read-side critical section. Note that these
* guarantees include CPUs that are offline, idle, or executing in user mode,
* as well as CPUs that are executing in the kernel.
*
* Furthermore, if CPU A invoked synchronize_srcu(), which returned
* to its caller on CPU B, then both CPU A and CPU B are guaranteed
* to have executed a full memory barrier during the execution of
* synchronize_srcu(). This guarantee applies even if CPU A and CPU B
* are the same CPU, but again only if the system has more than one CPU.
*
* Of course, these memory-ordering guarantees apply only when
* synchronize_srcu(), srcu_read_lock(), and srcu_read_unlock() are
* passed the same srcu_struct structure.
*
* If SRCU is likely idle, expedite the first request. This semantic
* was provided by Classic SRCU, and is relied upon by its users, so TREE
* SRCU must also provide it. Note that detecting idleness is heuristic
* and subject to both false positives and negatives.
*/
void synchronize_srcu(struct srcu_struct *ssp)
{
if (srcu_might_be_idle(ssp) || rcu_gp_is_expedited())
synchronize_srcu_expedited(ssp);
else
__synchronize_srcu(ssp, true);
}
EXPORT_SYMBOL_GPL(synchronize_srcu);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
/*
* Callback function for srcu_barrier() use.
*/
static void srcu_barrier_cb(struct rcu_head *rhp)
{
struct srcu_data *sdp;
struct srcu_struct *ssp;
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
sdp = container_of(rhp, struct srcu_data, srcu_barrier_head);
ssp = sdp->ssp;
if (atomic_dec_and_test(&ssp->srcu_barrier_cpu_cnt))
complete(&ssp->srcu_barrier_completion);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
}
/**
* srcu_barrier - Wait until all in-flight call_srcu() callbacks complete.
* @ssp: srcu_struct on which to wait for in-flight callbacks.
*/
void srcu_barrier(struct srcu_struct *ssp)
{
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
int cpu;
struct srcu_data *sdp;
unsigned long s = rcu_seq_snap(&ssp->srcu_barrier_seq);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
check_init_srcu_struct(ssp);
mutex_lock(&ssp->srcu_barrier_mutex);
if (rcu_seq_done(&ssp->srcu_barrier_seq, s)) {
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
smp_mb(); /* Force ordering following return. */
mutex_unlock(&ssp->srcu_barrier_mutex);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
return; /* Someone else did our work for us. */
}
rcu_seq_start(&ssp->srcu_barrier_seq);
init_completion(&ssp->srcu_barrier_completion);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
/* Initial count prevents reaching zero until all CBs are posted. */
atomic_set(&ssp->srcu_barrier_cpu_cnt, 1);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
/*
* Each pass through this loop enqueues a callback, but only
* on CPUs already having callbacks enqueued. Note that if
* a CPU already has callbacks enqueue, it must have already
* registered the need for a future grace period, so all we
* need do is enqueue a callback that will use the same
* grace period as the last callback already in the queue.
*/
for_each_possible_cpu(cpu) {
sdp = per_cpu_ptr(ssp->sda, cpu);
spin_lock_irq_rcu_node(sdp);
atomic_inc(&ssp->srcu_barrier_cpu_cnt);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
sdp->srcu_barrier_head.func = srcu_barrier_cb;
debug_rcu_head_queue(&sdp->srcu_barrier_head);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
if (!rcu_segcblist_entrain(&sdp->srcu_cblist,
&sdp->srcu_barrier_head, 0)) {
debug_rcu_head_unqueue(&sdp->srcu_barrier_head);
atomic_dec(&ssp->srcu_barrier_cpu_cnt);
}
spin_unlock_irq_rcu_node(sdp);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
}
/* Remove the initial count, at which point reaching zero can happen. */
if (atomic_dec_and_test(&ssp->srcu_barrier_cpu_cnt))
complete(&ssp->srcu_barrier_completion);
wait_for_completion(&ssp->srcu_barrier_completion);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
rcu_seq_end(&ssp->srcu_barrier_seq);
mutex_unlock(&ssp->srcu_barrier_mutex);
}
EXPORT_SYMBOL_GPL(srcu_barrier);
/**
* srcu_batches_completed - return batches completed.
* @ssp: srcu_struct on which to report batch completion.
*
* Report the number of batches, correlated with, but not necessarily
* precisely the same as, the number of grace periods that have elapsed.
*/
unsigned long srcu_batches_completed(struct srcu_struct *ssp)
{
return ssp->srcu_idx;
}
EXPORT_SYMBOL_GPL(srcu_batches_completed);
/*
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
* Core SRCU state machine. Push state bits of ->srcu_gp_seq
* to SRCU_STATE_SCAN2, and invoke srcu_gp_end() when scan has
* completed in that state.
*/
static void srcu_advance_state(struct srcu_struct *ssp)
{
int idx;
mutex_lock(&ssp->srcu_gp_mutex);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
/*
* Because readers might be delayed for an extended period after
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
* fetching ->srcu_idx for their index, at any point in time there
* might well be readers using both idx=0 and idx=1. We therefore
* need to wait for readers to clear from both index values before
* invoking a callback.
*
* The load-acquire ensures that we see the accesses performed
* by the prior grace period.
*/
idx = rcu_seq_state(smp_load_acquire(&ssp->srcu_gp_seq)); /* ^^^ */
if (idx == SRCU_STATE_IDLE) {
spin_lock_irq_rcu_node(ssp);
if (ULONG_CMP_GE(ssp->srcu_gp_seq, ssp->srcu_gp_seq_needed)) {
WARN_ON_ONCE(rcu_seq_state(ssp->srcu_gp_seq));
spin_unlock_irq_rcu_node(ssp);
mutex_unlock(&ssp->srcu_gp_mutex);
return;
}
idx = rcu_seq_state(READ_ONCE(ssp->srcu_gp_seq));
if (idx == SRCU_STATE_IDLE)
srcu_gp_start(ssp);
spin_unlock_irq_rcu_node(ssp);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
if (idx != SRCU_STATE_IDLE) {
mutex_unlock(&ssp->srcu_gp_mutex);
return; /* Someone else started the grace period. */
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
}
}
if (rcu_seq_state(READ_ONCE(ssp->srcu_gp_seq)) == SRCU_STATE_SCAN1) {
idx = 1 ^ (ssp->srcu_idx & 1);
if (!try_check_zero(ssp, idx, 1)) {
mutex_unlock(&ssp->srcu_gp_mutex);
return; /* readers present, retry later. */
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
}
srcu_flip(ssp);
rcu_seq_set_state(&ssp->srcu_gp_seq, SRCU_STATE_SCAN2);
}
if (rcu_seq_state(READ_ONCE(ssp->srcu_gp_seq)) == SRCU_STATE_SCAN2) {
/*
* SRCU read-side critical sections are normally short,
* so check at least twice in quick succession after a flip.
*/
idx = 1 ^ (ssp->srcu_idx & 1);
if (!try_check_zero(ssp, idx, 2)) {
mutex_unlock(&ssp->srcu_gp_mutex);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
return; /* readers present, retry later. */
}
srcu_gp_end(ssp); /* Releases ->srcu_gp_mutex. */
}
}
/*
* Invoke a limited number of SRCU callbacks that have passed through
* their grace period. If there are more to do, SRCU will reschedule
* the workqueue. Note that needed memory barriers have been executed
* in this task's context by srcu_readers_active_idx_check().
*/
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
static void srcu_invoke_callbacks(struct work_struct *work)
{
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
bool more;
struct rcu_cblist ready_cbs;
struct rcu_head *rhp;
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
struct srcu_data *sdp;
struct srcu_struct *ssp;
sdp = container_of(work, struct srcu_data, work);
ssp = sdp->ssp;
rcu_cblist_init(&ready_cbs);
spin_lock_irq_rcu_node(sdp);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
rcu_segcblist_advance(&sdp->srcu_cblist,
rcu_seq_current(&ssp->srcu_gp_seq));
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
if (sdp->srcu_cblist_invoking ||
!rcu_segcblist_ready_cbs(&sdp->srcu_cblist)) {
spin_unlock_irq_rcu_node(sdp);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
return; /* Someone else on the job or nothing to do. */
}
/* We are on the job! Extract and invoke ready callbacks. */
sdp->srcu_cblist_invoking = true;
rcu_segcblist_extract_done_cbs(&sdp->srcu_cblist, &ready_cbs);
spin_unlock_irq_rcu_node(sdp);
rhp = rcu_cblist_dequeue(&ready_cbs);
for (; rhp != NULL; rhp = rcu_cblist_dequeue(&ready_cbs)) {
debug_rcu_head_unqueue(rhp);
local_bh_disable();
rhp->func(rhp);
local_bh_enable();
}
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
/*
* Update counts, accelerate new callbacks, and if needed,
* schedule another round of callback invocation.
*/
spin_lock_irq_rcu_node(sdp);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
rcu_segcblist_insert_count(&sdp->srcu_cblist, &ready_cbs);
(void)rcu_segcblist_accelerate(&sdp->srcu_cblist,
rcu_seq_snap(&ssp->srcu_gp_seq));
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
sdp->srcu_cblist_invoking = false;
more = rcu_segcblist_ready_cbs(&sdp->srcu_cblist);
spin_unlock_irq_rcu_node(sdp);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
if (more)
srcu_schedule_cbs_sdp(sdp, 0);
}
/*
* Finished one round of SRCU grace period. Start another if there are
* more SRCU callbacks queued, otherwise put SRCU into not-running state.
*/
static void srcu_reschedule(struct srcu_struct *ssp, unsigned long delay)
{
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
bool pushgp = true;
spin_lock_irq_rcu_node(ssp);
if (ULONG_CMP_GE(ssp->srcu_gp_seq, ssp->srcu_gp_seq_needed)) {
if (!WARN_ON_ONCE(rcu_seq_state(ssp->srcu_gp_seq))) {
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
/* All requests fulfilled, time to go idle. */
pushgp = false;
}
} else if (!rcu_seq_state(ssp->srcu_gp_seq)) {
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
/* Outstanding request and no GP. Start one. */
srcu_gp_start(ssp);
}
spin_unlock_irq_rcu_node(ssp);
srcu: Parallelize callback handling Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2], however, there are workloads that could result in a high volume of concurrent invocations of call_srcu(), which with current SRCU would result in excessive lock contention on the srcu_struct structure's ->queue_lock, which protects SRCU's callback lists. This commit therefore moves SRCU to per-CPU callback lists, thus greatly reducing contention. Because a given SRCU instance no longer has a single centralized callback list, starting grace periods and invoking callbacks are both more complex than in the single-list Classic SRCU implementation. Starting grace periods and handling callbacks are now handled using an srcu_node tree that is in some ways similar to the rcu_node trees used by RCU-bh, RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is controlled by exactly the same Kconfig options and boot parameters that control the shape of the rcu_node tree). In addition, the old per-CPU srcu_array structure is now named srcu_data and contains an rcu_segcblist structure named ->srcu_cblist for its callbacks (and a spinlock to protect this). The srcu_struct gets an srcu_gp_seq that is used to associate callback segments with the corresponding completion-time grace-period number. These completion-time grace-period numbers are propagated up the srcu_node tree so that the grace-period workqueue handler can determine whether additional grace periods are needed on the one hand and where to look for callbacks that are ready to be invoked. The srcu_barrier() function must now wait on all instances of the per-CPU ->srcu_cblist. Because each ->srcu_cblist is protected by ->lock, srcu_barrier() can remotely add the needed callbacks. In theory, it could also remotely start grace periods, but in practice doing so is complex and racy. And interestingly enough, it is never necessary for srcu_barrier() to start a grace period because srcu_barrier() only enqueues a callback when a callback is already present--and it turns out that a grace period has to have already been started for this pre-existing callback. Furthermore, it is only the callback that srcu_barrier() needs to wait on, not any particular grace period. Therefore, a new rcu_segcblist_entrain() function enqueues the srcu_barrier() function's callback into the same segment occupied by the last pre-existing callback in the list. The special case where all the pre-existing callbacks are on a different list (because they are in the process of being invoked) is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL segment, relying on the done-callbacks check that takes place after all callbacks are inovked. Note that the readers use the same algorithm as before. Note that there is a separate srcu_idx that tells the readers what counter to increment. This unfortunately cannot be combined with srcu_gp_seq because they need to be incremented at different times. This commit introduces some ugly #ifdefs in rcutorture. These will go away when I feel good enough about Tree SRCU to ditch Classic SRCU. Some crude performance comparisons, courtesy of a quickly hacked rcuperf asynchronous-grace-period capability: Callback Queuing Overhead ------------------------- # CPUS Classic SRCU Tree SRCU ------ ------------ --------- 2 0.349 us 0.342 us 16 31.66 us 0.4 us 41 --------- 0.417 us The times are the 90th percentiles, a statistic that was chosen to reject the overheads of the occasional srcu_barrier() call needed to avoid OOMing the test machine. The rcuperf test hangs when running Classic SRCU at 41 CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code and that statistics, this is a convincing demonstration of Tree SRCU's performance and scalability advantages. [1] https://lwn.net/Articles/309030/ [2] https://patchwork.kernel.org/patch/5108281/ Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> [ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
2017-04-06 00:01:53 +08:00
if (pushgp)
queue_delayed_work(rcu_gp_wq, &ssp->work, delay);
}
/*
* This is the work-queue function that handles SRCU grace periods.
*/
static void process_srcu(struct work_struct *work)
{
struct srcu_struct *ssp;
ssp = container_of(work, struct srcu_struct, work.work);
srcu_advance_state(ssp);
srcu_reschedule(ssp, srcu_get_delay(ssp));
}
void srcutorture_get_gp_data(enum rcutorture_type test_type,
struct srcu_struct *ssp, int *flags,
unsigned long *gp_seq)
{
if (test_type != SRCU_FLAVOR)
return;
*flags = 0;
*gp_seq = rcu_seq_current(&ssp->srcu_gp_seq);
}
EXPORT_SYMBOL_GPL(srcutorture_get_gp_data);
void srcu_torture_stats_print(struct srcu_struct *ssp, char *tt, char *tf)
{
int cpu;
int idx;
unsigned long s0 = 0, s1 = 0;
idx = ssp->srcu_idx & 0x1;
pr_alert("%s%s Tree SRCU g%ld per-CPU(idx=%d):",
tt, tf, rcu_seq_current(&ssp->srcu_gp_seq), idx);
for_each_possible_cpu(cpu) {
unsigned long l0, l1;
unsigned long u0, u1;
long c0, c1;
struct srcu_data *sdp;
sdp = per_cpu_ptr(ssp->sda, cpu);
u0 = sdp->srcu_unlock_count[!idx];
u1 = sdp->srcu_unlock_count[idx];
/*
* Make sure that a lock is always counted if the corresponding
* unlock is counted.
*/
smp_rmb();
l0 = sdp->srcu_lock_count[!idx];
l1 = sdp->srcu_lock_count[idx];
c0 = l0 - u0;
c1 = l1 - u1;
pr_cont(" %d(%ld,%ld %c)",
cpu, c0, c1,
"C."[rcu_segcblist_empty(&sdp->srcu_cblist)]);
s0 += c0;
s1 += c1;
}
pr_cont(" T(%ld,%ld)\n", s0, s1);
}
EXPORT_SYMBOL_GPL(srcu_torture_stats_print);
static int __init srcu_bootup_announce(void)
{
pr_info("Hierarchical SRCU implementation.\n");
if (exp_holdoff != DEFAULT_SRCU_EXP_HOLDOFF)
pr_info("\tNon-default auto-expedite holdoff of %lu ns.\n", exp_holdoff);
return 0;
}
early_initcall(srcu_bootup_announce);
srcu: Make call_srcu() available during very early boot Event tracing is moving to SRCU in order to take advantage of the fact that SRCU may be safely used from idle and even offline CPUs. However, event tracing can invoke call_srcu() very early in the boot process, even before workqueue_init_early() is invoked (let alone rcu_init()). Therefore, call_srcu()'s attempts to queue work fail miserably. This commit therefore detects this situation, and refrains from attempting to queue work before rcu_init() time, but does everything else that it would have done, and in addition, adds the srcu_struct to a global list. The rcu_init() function now invokes a new srcu_init() function, which is empty if CONFIG_SRCU=n. Otherwise, srcu_init() queues work for each srcu_struct on the list. This all happens early enough in boot that there is but a single CPU with interrupts disabled, which allows synchronization to be dispensed with. Of course, the queued work won't actually be invoked until after workqueue_init() is invoked, which happens shortly after the scheduler is up and running. This means that although call_srcu() may be invoked any time after per-CPU variables have been set up, there is still a very narrow window when synchronize_srcu() won't work, and this window extends from the time that the scheduler starts until the time that workqueue_init() returns. This can be fixed in a manner similar to the fix for synchronize_rcu_expedited() and friends, but until someone actually needs to use synchronize_srcu() during this window, this fix is added churn for no benefit. Finally, note that Tree SRCU's new srcu_init() function invokes queue_work() rather than the queue_delayed_work() function that is invoked post-boot. The reason is that queue_delayed_work() will (as you would expect) post a timer, and timers have not yet been initialized. So use of queue_work() avoids the complaints about use of uninitialized spinlocks that would otherwise result. Besides, some delay is already provide by the aforementioned fact that the queued work won't actually be invoked until after the scheduler is up and running. Requested-by: Steven Rostedt <rostedt@goodmis.org> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Tested-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
2018-08-14 23:45:54 +08:00
void __init srcu_init(void)
{
struct srcu_struct *ssp;
srcu: Make call_srcu() available during very early boot Event tracing is moving to SRCU in order to take advantage of the fact that SRCU may be safely used from idle and even offline CPUs. However, event tracing can invoke call_srcu() very early in the boot process, even before workqueue_init_early() is invoked (let alone rcu_init()). Therefore, call_srcu()'s attempts to queue work fail miserably. This commit therefore detects this situation, and refrains from attempting to queue work before rcu_init() time, but does everything else that it would have done, and in addition, adds the srcu_struct to a global list. The rcu_init() function now invokes a new srcu_init() function, which is empty if CONFIG_SRCU=n. Otherwise, srcu_init() queues work for each srcu_struct on the list. This all happens early enough in boot that there is but a single CPU with interrupts disabled, which allows synchronization to be dispensed with. Of course, the queued work won't actually be invoked until after workqueue_init() is invoked, which happens shortly after the scheduler is up and running. This means that although call_srcu() may be invoked any time after per-CPU variables have been set up, there is still a very narrow window when synchronize_srcu() won't work, and this window extends from the time that the scheduler starts until the time that workqueue_init() returns. This can be fixed in a manner similar to the fix for synchronize_rcu_expedited() and friends, but until someone actually needs to use synchronize_srcu() during this window, this fix is added churn for no benefit. Finally, note that Tree SRCU's new srcu_init() function invokes queue_work() rather than the queue_delayed_work() function that is invoked post-boot. The reason is that queue_delayed_work() will (as you would expect) post a timer, and timers have not yet been initialized. So use of queue_work() avoids the complaints about use of uninitialized spinlocks that would otherwise result. Besides, some delay is already provide by the aforementioned fact that the queued work won't actually be invoked until after the scheduler is up and running. Requested-by: Steven Rostedt <rostedt@goodmis.org> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Tested-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
2018-08-14 23:45:54 +08:00
srcu_init_done = true;
while (!list_empty(&srcu_boot_list)) {
ssp = list_first_entry(&srcu_boot_list, struct srcu_struct,
work.work.entry);
check_init_srcu_struct(ssp);
list_del_init(&ssp->work.work.entry);
queue_work(rcu_gp_wq, &ssp->work.work);
srcu: Make call_srcu() available during very early boot Event tracing is moving to SRCU in order to take advantage of the fact that SRCU may be safely used from idle and even offline CPUs. However, event tracing can invoke call_srcu() very early in the boot process, even before workqueue_init_early() is invoked (let alone rcu_init()). Therefore, call_srcu()'s attempts to queue work fail miserably. This commit therefore detects this situation, and refrains from attempting to queue work before rcu_init() time, but does everything else that it would have done, and in addition, adds the srcu_struct to a global list. The rcu_init() function now invokes a new srcu_init() function, which is empty if CONFIG_SRCU=n. Otherwise, srcu_init() queues work for each srcu_struct on the list. This all happens early enough in boot that there is but a single CPU with interrupts disabled, which allows synchronization to be dispensed with. Of course, the queued work won't actually be invoked until after workqueue_init() is invoked, which happens shortly after the scheduler is up and running. This means that although call_srcu() may be invoked any time after per-CPU variables have been set up, there is still a very narrow window when synchronize_srcu() won't work, and this window extends from the time that the scheduler starts until the time that workqueue_init() returns. This can be fixed in a manner similar to the fix for synchronize_rcu_expedited() and friends, but until someone actually needs to use synchronize_srcu() during this window, this fix is added churn for no benefit. Finally, note that Tree SRCU's new srcu_init() function invokes queue_work() rather than the queue_delayed_work() function that is invoked post-boot. The reason is that queue_delayed_work() will (as you would expect) post a timer, and timers have not yet been initialized. So use of queue_work() avoids the complaints about use of uninitialized spinlocks that would otherwise result. Besides, some delay is already provide by the aforementioned fact that the queued work won't actually be invoked until after the scheduler is up and running. Requested-by: Steven Rostedt <rostedt@goodmis.org> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Tested-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
2018-08-14 23:45:54 +08:00
}
}
srcu: Allocate per-CPU data for DEFINE_SRCU() in modules Adding DEFINE_SRCU() or DEFINE_STATIC_SRCU() to a loadable module requires that the size of the reserved region be increased, which is not something we want to be doing all that often. One approach would be to require that loadable modules define an srcu_struct and invoke init_srcu_struct() from their module_init function and cleanup_srcu_struct() from their module_exit function. However, this is more than a bit user unfriendly. This commit therefore creates an ___srcu_struct_ptrs linker section, and pointers to srcu_struct structures created by DEFINE_SRCU() and DEFINE_STATIC_SRCU() within a module are placed into that module's ___srcu_struct_ptrs section. The required init_srcu_struct() and cleanup_srcu_struct() functions are then automatically invoked as needed when that module is loaded and unloaded, thus allowing modules to continue to use DEFINE_SRCU() and DEFINE_STATIC_SRCU() while avoiding the need to increase the size of the reserved region. Many of the algorithms and some of the code was cheerfully cherry-picked from other code making use of linker sections, perhaps most notably from tracepoints. All bugs are nevertheless the sole property of the author. Suggested-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com> [ paulmck: Use __section() and use "default" in srcu_module_notify()'s "switch" statement as suggested by Joel Fernandes. ] Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com> Tested-by: Joel Fernandes (Google) <joel@joelfernandes.org>
2019-04-06 07:15:00 +08:00
#ifdef CONFIG_MODULES
/* Initialize any global-scope srcu_struct structures used by this module. */
static int srcu_module_coming(struct module *mod)
{
int i;
struct srcu_struct **sspp = mod->srcu_struct_ptrs;
int ret;
for (i = 0; i < mod->num_srcu_structs; i++) {
ret = init_srcu_struct(*(sspp++));
if (WARN_ON_ONCE(ret))
return ret;
}
return 0;
}
/* Clean up any global-scope srcu_struct structures used by this module. */
static void srcu_module_going(struct module *mod)
{
int i;
struct srcu_struct **sspp = mod->srcu_struct_ptrs;
for (i = 0; i < mod->num_srcu_structs; i++)
cleanup_srcu_struct(*(sspp++));
}
/* Handle one module, either coming or going. */
static int srcu_module_notify(struct notifier_block *self,
unsigned long val, void *data)
{
struct module *mod = data;
int ret = 0;
switch (val) {
case MODULE_STATE_COMING:
ret = srcu_module_coming(mod);
break;
case MODULE_STATE_GOING:
srcu_module_going(mod);
break;
default:
break;
}
return ret;
}
static struct notifier_block srcu_module_nb = {
.notifier_call = srcu_module_notify,
.priority = 0,
};
static __init int init_srcu_module_notifier(void)
{
int ret;
ret = register_module_notifier(&srcu_module_nb);
if (ret)
pr_warn("Failed to register srcu module notifier\n");
return ret;
}
late_initcall(init_srcu_module_notifier);
#endif /* #ifdef CONFIG_MODULES */