linux_dsm_epyc7002/kernel/locking/mutex.c

915 lines
24 KiB
C
Raw Normal View History

/*
* kernel/locking/mutex.c
*
* Mutexes: blocking mutual exclusion locks
*
* Started by Ingo Molnar:
*
* Copyright (C) 2004, 2005, 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
*
* Many thanks to Arjan van de Ven, Thomas Gleixner, Steven Rostedt and
* David Howells for suggestions and improvements.
*
mutex: implement adaptive spinning Change mutex contention behaviour such that it will sometimes busy wait on acquisition - moving its behaviour closer to that of spinlocks. This concept got ported to mainline from the -rt tree, where it was originally implemented for rtmutexes by Steven Rostedt, based on work by Gregory Haskins. Testing with Ingo's test-mutex application (http://lkml.org/lkml/2006/1/8/50) gave a 345% boost for VFS scalability on my testbox: # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 296604 # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 85870 The key criteria for the busy wait is that the lock owner has to be running on a (different) cpu. The idea is that as long as the owner is running, there is a fair chance it'll release the lock soon, and thus we'll be better off spinning instead of blocking/scheduling. Since regular mutexes (as opposed to rtmutexes) do not atomically track the owner, we add the owner in a non-atomic fashion and deal with the races in the slowpath. Furthermore, to ease the testing of the performance impact of this new code, there is means to disable this behaviour runtime (without having to reboot the system), when scheduler debugging is enabled (CONFIG_SCHED_DEBUG=y), by issuing the following command: # echo NO_OWNER_SPIN > /debug/sched_features This command re-enables spinning again (this is also the default): # echo OWNER_SPIN > /debug/sched_features Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-01-12 20:01:47 +07:00
* - Adaptive spinning for mutexes by Peter Zijlstra. (Ported to mainline
* from the -rt tree, where it was originally implemented for rtmutexes
* by Steven Rostedt, based on work by Gregory Haskins, Peter Morreale
* and Sven Dietrich.
*
* Also see Documentation/mutex-design.txt.
*/
#include <linux/mutex.h>
#include <linux/ww_mutex.h>
#include <linux/sched.h>
#include <linux/sched/rt.h>
#include <linux/export.h>
#include <linux/spinlock.h>
#include <linux/interrupt.h>
#include <linux/debug_locks.h>
#include "mcs_spinlock.h"
/*
* In the DEBUG case we are using the "NULL fastpath" for mutexes,
* which forces all calls into the slowpath:
*/
#ifdef CONFIG_DEBUG_MUTEXES
# include "mutex-debug.h"
# include <asm-generic/mutex-null.h>
#else
# include "mutex.h"
# include <asm/mutex.h>
#endif
mutex: Make more scalable by doing less atomic operations In the __mutex_lock_common() function, an initial entry into the lock slow path will cause two atomic_xchg instructions to be issued. Together with the atomic decrement in the fast path, a total of three atomic read-modify-write instructions will be issued in rapid succession. This can cause a lot of cache bouncing when many tasks are trying to acquire the mutex at the same time. This patch will reduce the number of atomic_xchg instructions used by checking the counter value first before issuing the instruction. The atomic_read() function is just a simple memory read. The atomic_xchg() function, on the other hand, can be up to 2 order of magnitude or even more in cost when compared with atomic_read(). By using atomic_read() to check the value first before calling atomic_xchg(), we can avoid a lot of unnecessary cache coherency traffic. The only downside with this change is that a task on the slow path will have a tiny bit less chance of getting the mutex when competing with another task in the fast path. The same is true for the atomic_cmpxchg() function in the mutex-spin-on-owner loop. So an atomic_read() is also performed before calling atomic_cmpxchg(). The mutex locking and unlocking code for the x86 architecture can allow any negative number to be used in the mutex count to indicate that some tasks are waiting for the mutex. I am not so sure if that is the case for the other architectures. So the default is to avoid atomic_xchg() if the count has already been set to -1. For x86, the check is modified to include all negative numbers to cover a larger case. The following table shows the jobs per minutes (JPM) scalability data on an 8-node 80-core Westmere box with a 3.7.10 kernel. The numactl command is used to restrict the running of the high_systime workloads to 1/2/4/8 nodes with hyperthreading on and off. +-----------------+-----------+------------+----------+ | Configuration | Mean JPM | Mean JPM | % Change | | | w/o patch | with patch | | +-----------------+-----------------------------------+ | | User Range 1100 - 2000 | +-----------------+-----------------------------------+ | 8 nodes, HT on | 36980 | 148590 | +301.8% | | 8 nodes, HT off | 42799 | 145011 | +238.8% | | 4 nodes, HT on | 61318 | 118445 | +51.1% | | 4 nodes, HT off | 158481 | 158592 | +0.1% | | 2 nodes, HT on | 180602 | 173967 | -3.7% | | 2 nodes, HT off | 198409 | 198073 | -0.2% | | 1 node , HT on | 149042 | 147671 | -0.9% | | 1 node , HT off | 126036 | 126533 | +0.4% | +-----------------+-----------------------------------+ | | User Range 200 - 1000 | +-----------------+-----------------------------------+ | 8 nodes, HT on | 41525 | 122349 | +194.6% | | 8 nodes, HT off | 49866 | 124032 | +148.7% | | 4 nodes, HT on | 66409 | 106984 | +61.1% | | 4 nodes, HT off | 119880 | 130508 | +8.9% | | 2 nodes, HT on | 138003 | 133948 | -2.9% | | 2 nodes, HT off | 132792 | 131997 | -0.6% | | 1 node , HT on | 116593 | 115859 | -0.6% | | 1 node , HT off | 104499 | 104597 | +0.1% | +-----------------+------------+-----------+----------+ At low user range 10-100, the JPM differences were within +/-1%. So they are not that interesting. AIM7 benchmark run has a pretty large run-to-run variance due to random nature of the subtests executed. So a difference of less than +-5% may not be really significant. This patch improves high_systime workload performance at 4 nodes and up by maintaining transaction rates without significant drop-off at high node count. The patch has practically no impact on 1 and 2 nodes system. The table below shows the percentage time (as reported by perf record -a -s -g) spent on the __mutex_lock_slowpath() function by the high_systime workload at 1500 users for 2/4/8-node configurations with hyperthreading off. +---------------+-----------------+------------------+---------+ | Configuration | %Time w/o patch | %Time with patch | %Change | +---------------+-----------------+------------------+---------+ | 8 nodes | 65.34% | 0.69% | -99% | | 4 nodes | 8.70% | 1.02% | -88% | | 2 nodes | 0.41% | 0.32% | -22% | +---------------+-----------------+------------------+---------+ It is obvious that the dramatic performance improvement at 8 nodes was due to the drastic cut in the time spent within the __mutex_lock_slowpath() function. The table below show the improvements in other AIM7 workloads (at 8 nodes, hyperthreading off). +--------------+---------------+----------------+-----------------+ | Workload | mean % change | mean % change | mean % change | | | 10-100 users | 200-1000 users | 1100-2000 users | +--------------+---------------+----------------+-----------------+ | alltests | +0.6% | +104.2% | +185.9% | | five_sec | +1.9% | +0.9% | +0.9% | | fserver | +1.4% | -7.7% | +5.1% | | new_fserver | -0.5% | +3.2% | +3.1% | | shared | +13.1% | +146.1% | +181.5% | | short | +7.4% | +5.0% | +4.2% | +--------------+---------------+----------------+-----------------+ Signed-off-by: Waiman Long <Waiman.Long@hp.com> Reviewed-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Reviewed-by: Rik van Riel <riel@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Chandramouleeswaran Aswin <aswin@hp.com> Cc: Norton: Scott J <scott.norton@hp.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: David Howells <dhowells@redhat.com> Cc: Dave Jones <davej@redhat.com> Cc: Clark Williams <williams@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1366226594-5506-3-git-send-email-Waiman.Long@hp.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-04-18 02:23:12 +07:00
/*
* A negative mutex count indicates that waiters are sleeping waiting for the
* mutex.
mutex: Make more scalable by doing less atomic operations In the __mutex_lock_common() function, an initial entry into the lock slow path will cause two atomic_xchg instructions to be issued. Together with the atomic decrement in the fast path, a total of three atomic read-modify-write instructions will be issued in rapid succession. This can cause a lot of cache bouncing when many tasks are trying to acquire the mutex at the same time. This patch will reduce the number of atomic_xchg instructions used by checking the counter value first before issuing the instruction. The atomic_read() function is just a simple memory read. The atomic_xchg() function, on the other hand, can be up to 2 order of magnitude or even more in cost when compared with atomic_read(). By using atomic_read() to check the value first before calling atomic_xchg(), we can avoid a lot of unnecessary cache coherency traffic. The only downside with this change is that a task on the slow path will have a tiny bit less chance of getting the mutex when competing with another task in the fast path. The same is true for the atomic_cmpxchg() function in the mutex-spin-on-owner loop. So an atomic_read() is also performed before calling atomic_cmpxchg(). The mutex locking and unlocking code for the x86 architecture can allow any negative number to be used in the mutex count to indicate that some tasks are waiting for the mutex. I am not so sure if that is the case for the other architectures. So the default is to avoid atomic_xchg() if the count has already been set to -1. For x86, the check is modified to include all negative numbers to cover a larger case. The following table shows the jobs per minutes (JPM) scalability data on an 8-node 80-core Westmere box with a 3.7.10 kernel. The numactl command is used to restrict the running of the high_systime workloads to 1/2/4/8 nodes with hyperthreading on and off. +-----------------+-----------+------------+----------+ | Configuration | Mean JPM | Mean JPM | % Change | | | w/o patch | with patch | | +-----------------+-----------------------------------+ | | User Range 1100 - 2000 | +-----------------+-----------------------------------+ | 8 nodes, HT on | 36980 | 148590 | +301.8% | | 8 nodes, HT off | 42799 | 145011 | +238.8% | | 4 nodes, HT on | 61318 | 118445 | +51.1% | | 4 nodes, HT off | 158481 | 158592 | +0.1% | | 2 nodes, HT on | 180602 | 173967 | -3.7% | | 2 nodes, HT off | 198409 | 198073 | -0.2% | | 1 node , HT on | 149042 | 147671 | -0.9% | | 1 node , HT off | 126036 | 126533 | +0.4% | +-----------------+-----------------------------------+ | | User Range 200 - 1000 | +-----------------+-----------------------------------+ | 8 nodes, HT on | 41525 | 122349 | +194.6% | | 8 nodes, HT off | 49866 | 124032 | +148.7% | | 4 nodes, HT on | 66409 | 106984 | +61.1% | | 4 nodes, HT off | 119880 | 130508 | +8.9% | | 2 nodes, HT on | 138003 | 133948 | -2.9% | | 2 nodes, HT off | 132792 | 131997 | -0.6% | | 1 node , HT on | 116593 | 115859 | -0.6% | | 1 node , HT off | 104499 | 104597 | +0.1% | +-----------------+------------+-----------+----------+ At low user range 10-100, the JPM differences were within +/-1%. So they are not that interesting. AIM7 benchmark run has a pretty large run-to-run variance due to random nature of the subtests executed. So a difference of less than +-5% may not be really significant. This patch improves high_systime workload performance at 4 nodes and up by maintaining transaction rates without significant drop-off at high node count. The patch has practically no impact on 1 and 2 nodes system. The table below shows the percentage time (as reported by perf record -a -s -g) spent on the __mutex_lock_slowpath() function by the high_systime workload at 1500 users for 2/4/8-node configurations with hyperthreading off. +---------------+-----------------+------------------+---------+ | Configuration | %Time w/o patch | %Time with patch | %Change | +---------------+-----------------+------------------+---------+ | 8 nodes | 65.34% | 0.69% | -99% | | 4 nodes | 8.70% | 1.02% | -88% | | 2 nodes | 0.41% | 0.32% | -22% | +---------------+-----------------+------------------+---------+ It is obvious that the dramatic performance improvement at 8 nodes was due to the drastic cut in the time spent within the __mutex_lock_slowpath() function. The table below show the improvements in other AIM7 workloads (at 8 nodes, hyperthreading off). +--------------+---------------+----------------+-----------------+ | Workload | mean % change | mean % change | mean % change | | | 10-100 users | 200-1000 users | 1100-2000 users | +--------------+---------------+----------------+-----------------+ | alltests | +0.6% | +104.2% | +185.9% | | five_sec | +1.9% | +0.9% | +0.9% | | fserver | +1.4% | -7.7% | +5.1% | | new_fserver | -0.5% | +3.2% | +3.1% | | shared | +13.1% | +146.1% | +181.5% | | short | +7.4% | +5.0% | +4.2% | +--------------+---------------+----------------+-----------------+ Signed-off-by: Waiman Long <Waiman.Long@hp.com> Reviewed-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Reviewed-by: Rik van Riel <riel@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Chandramouleeswaran Aswin <aswin@hp.com> Cc: Norton: Scott J <scott.norton@hp.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: David Howells <dhowells@redhat.com> Cc: Dave Jones <davej@redhat.com> Cc: Clark Williams <williams@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1366226594-5506-3-git-send-email-Waiman.Long@hp.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-04-18 02:23:12 +07:00
*/
#define MUTEX_SHOW_NO_WAITER(mutex) (atomic_read(&(mutex)->count) >= 0)
void
__mutex_init(struct mutex *lock, const char *name, struct lock_class_key *key)
{
atomic_set(&lock->count, 1);
spin_lock_init(&lock->wait_lock);
INIT_LIST_HEAD(&lock->wait_list);
mutex: implement adaptive spinning Change mutex contention behaviour such that it will sometimes busy wait on acquisition - moving its behaviour closer to that of spinlocks. This concept got ported to mainline from the -rt tree, where it was originally implemented for rtmutexes by Steven Rostedt, based on work by Gregory Haskins. Testing with Ingo's test-mutex application (http://lkml.org/lkml/2006/1/8/50) gave a 345% boost for VFS scalability on my testbox: # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 296604 # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 85870 The key criteria for the busy wait is that the lock owner has to be running on a (different) cpu. The idea is that as long as the owner is running, there is a fair chance it'll release the lock soon, and thus we'll be better off spinning instead of blocking/scheduling. Since regular mutexes (as opposed to rtmutexes) do not atomically track the owner, we add the owner in a non-atomic fashion and deal with the races in the slowpath. Furthermore, to ease the testing of the performance impact of this new code, there is means to disable this behaviour runtime (without having to reboot the system), when scheduler debugging is enabled (CONFIG_SCHED_DEBUG=y), by issuing the following command: # echo NO_OWNER_SPIN > /debug/sched_features This command re-enables spinning again (this is also the default): # echo OWNER_SPIN > /debug/sched_features Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-01-12 20:01:47 +07:00
mutex_clear_owner(lock);
mutex: Queue mutex spinners with MCS lock to reduce cacheline contention The current mutex spinning code (with MUTEX_SPIN_ON_OWNER option turned on) allow multiple tasks to spin on a single mutex concurrently. A potential problem with the current approach is that when the mutex becomes available, all the spinning tasks will try to acquire the mutex more or less simultaneously. As a result, there will be a lot of cacheline bouncing especially on systems with a large number of CPUs. This patch tries to reduce this kind of contention by putting the mutex spinners into a queue so that only the first one in the queue will try to acquire the mutex. This will reduce contention and allow all the tasks to move forward faster. The queuing of mutex spinners is done using an MCS lock based implementation which will further reduce contention on the mutex cacheline than a similar ticket spinlock based implementation. This patch will add a new field into the mutex data structure for holding the MCS lock. This expands the mutex size by 8 bytes for 64-bit system and 4 bytes for 32-bit system. This overhead will be avoid if the MUTEX_SPIN_ON_OWNER option is turned off. The following table shows the jobs per minute (JPM) scalability data on an 8-node 80-core Westmere box with a 3.7.10 kernel. The numactl command is used to restrict the running of the fserver workloads to 1/2/4/8 nodes with hyperthreading off. +-----------------+-----------+-----------+-------------+----------+ | Configuration | Mean JPM | Mean JPM | Mean JPM | % Change | | | w/o patch | patch 1 | patches 1&2 | 1->1&2 | +-----------------+------------------------------------------------+ | | User Range 1100 - 2000 | +-----------------+------------------------------------------------+ | 8 nodes, HT off | 227972 | 227237 | 305043 | +34.2% | | 4 nodes, HT off | 393503 | 381558 | 394650 | +3.4% | | 2 nodes, HT off | 334957 | 325240 | 338853 | +4.2% | | 1 node , HT off | 198141 | 197972 | 198075 | +0.1% | +-----------------+------------------------------------------------+ | | User Range 200 - 1000 | +-----------------+------------------------------------------------+ | 8 nodes, HT off | 282325 | 312870 | 332185 | +6.2% | | 4 nodes, HT off | 390698 | 378279 | 393419 | +4.0% | | 2 nodes, HT off | 336986 | 326543 | 340260 | +4.2% | | 1 node , HT off | 197588 | 197622 | 197582 | 0.0% | +-----------------+-----------+-----------+-------------+----------+ At low user range 10-100, the JPM differences were within +/-1%. So they are not that interesting. The fserver workload uses mutex spinning extensively. With just the mutex change in the first patch, there is no noticeable change in performance. Rather, there is a slight drop in performance. This mutex spinning patch more than recovers the lost performance and show a significant increase of +30% at high user load with the full 8 nodes. Similar improvements were also seen in a 3.8 kernel. The table below shows the %time spent by different kernel functions as reported by perf when running the fserver workload at 1500 users with all 8 nodes. +-----------------------+-----------+---------+-------------+ | Function | % time | % time | % time | | | w/o patch | patch 1 | patches 1&2 | +-----------------------+-----------+---------+-------------+ | __read_lock_failed | 34.96% | 34.91% | 29.14% | | __write_lock_failed | 10.14% | 10.68% | 7.51% | | mutex_spin_on_owner | 3.62% | 3.42% | 2.33% | | mspin_lock | N/A | N/A | 9.90% | | __mutex_lock_slowpath | 1.46% | 0.81% | 0.14% | | _raw_spin_lock | 2.25% | 2.50% | 1.10% | +-----------------------+-----------+---------+-------------+ The fserver workload for an 8-node system is dominated by the contention in the read/write lock. Mutex contention also plays a role. With the first patch only, mutex contention is down (as shown by the __mutex_lock_slowpath figure) which help a little bit. We saw only a few percents improvement with that. By applying patch 2 as well, the single mutex_spin_on_owner figure is now split out into an additional mspin_lock figure. The time increases from 3.42% to 11.23%. It shows a great reduction in contention among the spinners leading to a 30% improvement. The time ratio 9.9/2.33=4.3 indicates that there are on average 4+ spinners waiting in the spin_lock loop for each spinner in the mutex_spin_on_owner loop. Contention in other locking functions also go down by quite a lot. The table below shows the performance change of both patches 1 & 2 over patch 1 alone in other AIM7 workloads (at 8 nodes, hyperthreading off). +--------------+---------------+----------------+-----------------+ | Workload | mean % change | mean % change | mean % change | | | 10-100 users | 200-1000 users | 1100-2000 users | +--------------+---------------+----------------+-----------------+ | alltests | 0.0% | -0.8% | +0.6% | | five_sec | -0.3% | +0.8% | +0.8% | | high_systime | +0.4% | +2.4% | +2.1% | | new_fserver | +0.1% | +14.1% | +34.2% | | shared | -0.5% | -0.3% | -0.4% | | short | -1.7% | -9.8% | -8.3% | +--------------+---------------+----------------+-----------------+ The short workload is the only one that shows a decline in performance probably due to the spinner locking and queuing overhead. Signed-off-by: Waiman Long <Waiman.Long@hp.com> Reviewed-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Acked-by: Rik van Riel <riel@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Chandramouleeswaran Aswin <aswin@hp.com> Cc: Norton Scott J <scott.norton@hp.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: David Howells <dhowells@redhat.com> Cc: Dave Jones <davej@redhat.com> Cc: Clark Williams <williams@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1366226594-5506-4-git-send-email-Waiman.Long@hp.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-04-18 02:23:13 +07:00
#ifdef CONFIG_MUTEX_SPIN_ON_OWNER
lock->mcs_lock = NULL;
mutex: Queue mutex spinners with MCS lock to reduce cacheline contention The current mutex spinning code (with MUTEX_SPIN_ON_OWNER option turned on) allow multiple tasks to spin on a single mutex concurrently. A potential problem with the current approach is that when the mutex becomes available, all the spinning tasks will try to acquire the mutex more or less simultaneously. As a result, there will be a lot of cacheline bouncing especially on systems with a large number of CPUs. This patch tries to reduce this kind of contention by putting the mutex spinners into a queue so that only the first one in the queue will try to acquire the mutex. This will reduce contention and allow all the tasks to move forward faster. The queuing of mutex spinners is done using an MCS lock based implementation which will further reduce contention on the mutex cacheline than a similar ticket spinlock based implementation. This patch will add a new field into the mutex data structure for holding the MCS lock. This expands the mutex size by 8 bytes for 64-bit system and 4 bytes for 32-bit system. This overhead will be avoid if the MUTEX_SPIN_ON_OWNER option is turned off. The following table shows the jobs per minute (JPM) scalability data on an 8-node 80-core Westmere box with a 3.7.10 kernel. The numactl command is used to restrict the running of the fserver workloads to 1/2/4/8 nodes with hyperthreading off. +-----------------+-----------+-----------+-------------+----------+ | Configuration | Mean JPM | Mean JPM | Mean JPM | % Change | | | w/o patch | patch 1 | patches 1&2 | 1->1&2 | +-----------------+------------------------------------------------+ | | User Range 1100 - 2000 | +-----------------+------------------------------------------------+ | 8 nodes, HT off | 227972 | 227237 | 305043 | +34.2% | | 4 nodes, HT off | 393503 | 381558 | 394650 | +3.4% | | 2 nodes, HT off | 334957 | 325240 | 338853 | +4.2% | | 1 node , HT off | 198141 | 197972 | 198075 | +0.1% | +-----------------+------------------------------------------------+ | | User Range 200 - 1000 | +-----------------+------------------------------------------------+ | 8 nodes, HT off | 282325 | 312870 | 332185 | +6.2% | | 4 nodes, HT off | 390698 | 378279 | 393419 | +4.0% | | 2 nodes, HT off | 336986 | 326543 | 340260 | +4.2% | | 1 node , HT off | 197588 | 197622 | 197582 | 0.0% | +-----------------+-----------+-----------+-------------+----------+ At low user range 10-100, the JPM differences were within +/-1%. So they are not that interesting. The fserver workload uses mutex spinning extensively. With just the mutex change in the first patch, there is no noticeable change in performance. Rather, there is a slight drop in performance. This mutex spinning patch more than recovers the lost performance and show a significant increase of +30% at high user load with the full 8 nodes. Similar improvements were also seen in a 3.8 kernel. The table below shows the %time spent by different kernel functions as reported by perf when running the fserver workload at 1500 users with all 8 nodes. +-----------------------+-----------+---------+-------------+ | Function | % time | % time | % time | | | w/o patch | patch 1 | patches 1&2 | +-----------------------+-----------+---------+-------------+ | __read_lock_failed | 34.96% | 34.91% | 29.14% | | __write_lock_failed | 10.14% | 10.68% | 7.51% | | mutex_spin_on_owner | 3.62% | 3.42% | 2.33% | | mspin_lock | N/A | N/A | 9.90% | | __mutex_lock_slowpath | 1.46% | 0.81% | 0.14% | | _raw_spin_lock | 2.25% | 2.50% | 1.10% | +-----------------------+-----------+---------+-------------+ The fserver workload for an 8-node system is dominated by the contention in the read/write lock. Mutex contention also plays a role. With the first patch only, mutex contention is down (as shown by the __mutex_lock_slowpath figure) which help a little bit. We saw only a few percents improvement with that. By applying patch 2 as well, the single mutex_spin_on_owner figure is now split out into an additional mspin_lock figure. The time increases from 3.42% to 11.23%. It shows a great reduction in contention among the spinners leading to a 30% improvement. The time ratio 9.9/2.33=4.3 indicates that there are on average 4+ spinners waiting in the spin_lock loop for each spinner in the mutex_spin_on_owner loop. Contention in other locking functions also go down by quite a lot. The table below shows the performance change of both patches 1 & 2 over patch 1 alone in other AIM7 workloads (at 8 nodes, hyperthreading off). +--------------+---------------+----------------+-----------------+ | Workload | mean % change | mean % change | mean % change | | | 10-100 users | 200-1000 users | 1100-2000 users | +--------------+---------------+----------------+-----------------+ | alltests | 0.0% | -0.8% | +0.6% | | five_sec | -0.3% | +0.8% | +0.8% | | high_systime | +0.4% | +2.4% | +2.1% | | new_fserver | +0.1% | +14.1% | +34.2% | | shared | -0.5% | -0.3% | -0.4% | | short | -1.7% | -9.8% | -8.3% | +--------------+---------------+----------------+-----------------+ The short workload is the only one that shows a decline in performance probably due to the spinner locking and queuing overhead. Signed-off-by: Waiman Long <Waiman.Long@hp.com> Reviewed-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Acked-by: Rik van Riel <riel@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Chandramouleeswaran Aswin <aswin@hp.com> Cc: Norton Scott J <scott.norton@hp.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: David Howells <dhowells@redhat.com> Cc: Dave Jones <davej@redhat.com> Cc: Clark Williams <williams@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1366226594-5506-4-git-send-email-Waiman.Long@hp.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-04-18 02:23:13 +07:00
#endif
debug_mutex_init(lock, name, key);
}
EXPORT_SYMBOL(__mutex_init);
#ifndef CONFIG_DEBUG_LOCK_ALLOC
/*
* We split the mutex lock/unlock logic into separate fastpath and
* slowpath functions, to reduce the register pressure on the fastpath.
* We also put the fastpath first in the kernel image, to make sure the
* branch is predicted by the CPU as default-untaken.
*/
static __used noinline void __sched
__mutex_lock_slowpath(atomic_t *lock_count);
/**
* mutex_lock - acquire the mutex
* @lock: the mutex to be acquired
*
* Lock the mutex exclusively for this task. If the mutex is not
* available right now, it will sleep until it can get it.
*
* The mutex must later on be released by the same task that
* acquired it. Recursive locking is not allowed. The task
* may not exit without first unlocking the mutex. Also, kernel
* memory where the mutex resides mutex must not be freed with
* the mutex still locked. The mutex must first be initialized
* (or statically defined) before it can be locked. memset()-ing
* the mutex to 0 is not allowed.
*
* ( The CONFIG_DEBUG_MUTEXES .config option turns on debugging
* checks that will enforce the restrictions and will also do
* deadlock debugging. )
*
* This function is similar to (but not equivalent to) down().
*/
void __sched mutex_lock(struct mutex *lock)
{
might_sleep();
/*
* The locking fastpath is the 1->0 transition from
* 'unlocked' into 'locked' state.
*/
__mutex_fastpath_lock(&lock->count, __mutex_lock_slowpath);
mutex: implement adaptive spinning Change mutex contention behaviour such that it will sometimes busy wait on acquisition - moving its behaviour closer to that of spinlocks. This concept got ported to mainline from the -rt tree, where it was originally implemented for rtmutexes by Steven Rostedt, based on work by Gregory Haskins. Testing with Ingo's test-mutex application (http://lkml.org/lkml/2006/1/8/50) gave a 345% boost for VFS scalability on my testbox: # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 296604 # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 85870 The key criteria for the busy wait is that the lock owner has to be running on a (different) cpu. The idea is that as long as the owner is running, there is a fair chance it'll release the lock soon, and thus we'll be better off spinning instead of blocking/scheduling. Since regular mutexes (as opposed to rtmutexes) do not atomically track the owner, we add the owner in a non-atomic fashion and deal with the races in the slowpath. Furthermore, to ease the testing of the performance impact of this new code, there is means to disable this behaviour runtime (without having to reboot the system), when scheduler debugging is enabled (CONFIG_SCHED_DEBUG=y), by issuing the following command: # echo NO_OWNER_SPIN > /debug/sched_features This command re-enables spinning again (this is also the default): # echo OWNER_SPIN > /debug/sched_features Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-01-12 20:01:47 +07:00
mutex_set_owner(lock);
}
EXPORT_SYMBOL(mutex_lock);
#endif
#ifdef CONFIG_MUTEX_SPIN_ON_OWNER
mutex: Queue mutex spinners with MCS lock to reduce cacheline contention The current mutex spinning code (with MUTEX_SPIN_ON_OWNER option turned on) allow multiple tasks to spin on a single mutex concurrently. A potential problem with the current approach is that when the mutex becomes available, all the spinning tasks will try to acquire the mutex more or less simultaneously. As a result, there will be a lot of cacheline bouncing especially on systems with a large number of CPUs. This patch tries to reduce this kind of contention by putting the mutex spinners into a queue so that only the first one in the queue will try to acquire the mutex. This will reduce contention and allow all the tasks to move forward faster. The queuing of mutex spinners is done using an MCS lock based implementation which will further reduce contention on the mutex cacheline than a similar ticket spinlock based implementation. This patch will add a new field into the mutex data structure for holding the MCS lock. This expands the mutex size by 8 bytes for 64-bit system and 4 bytes for 32-bit system. This overhead will be avoid if the MUTEX_SPIN_ON_OWNER option is turned off. The following table shows the jobs per minute (JPM) scalability data on an 8-node 80-core Westmere box with a 3.7.10 kernel. The numactl command is used to restrict the running of the fserver workloads to 1/2/4/8 nodes with hyperthreading off. +-----------------+-----------+-----------+-------------+----------+ | Configuration | Mean JPM | Mean JPM | Mean JPM | % Change | | | w/o patch | patch 1 | patches 1&2 | 1->1&2 | +-----------------+------------------------------------------------+ | | User Range 1100 - 2000 | +-----------------+------------------------------------------------+ | 8 nodes, HT off | 227972 | 227237 | 305043 | +34.2% | | 4 nodes, HT off | 393503 | 381558 | 394650 | +3.4% | | 2 nodes, HT off | 334957 | 325240 | 338853 | +4.2% | | 1 node , HT off | 198141 | 197972 | 198075 | +0.1% | +-----------------+------------------------------------------------+ | | User Range 200 - 1000 | +-----------------+------------------------------------------------+ | 8 nodes, HT off | 282325 | 312870 | 332185 | +6.2% | | 4 nodes, HT off | 390698 | 378279 | 393419 | +4.0% | | 2 nodes, HT off | 336986 | 326543 | 340260 | +4.2% | | 1 node , HT off | 197588 | 197622 | 197582 | 0.0% | +-----------------+-----------+-----------+-------------+----------+ At low user range 10-100, the JPM differences were within +/-1%. So they are not that interesting. The fserver workload uses mutex spinning extensively. With just the mutex change in the first patch, there is no noticeable change in performance. Rather, there is a slight drop in performance. This mutex spinning patch more than recovers the lost performance and show a significant increase of +30% at high user load with the full 8 nodes. Similar improvements were also seen in a 3.8 kernel. The table below shows the %time spent by different kernel functions as reported by perf when running the fserver workload at 1500 users with all 8 nodes. +-----------------------+-----------+---------+-------------+ | Function | % time | % time | % time | | | w/o patch | patch 1 | patches 1&2 | +-----------------------+-----------+---------+-------------+ | __read_lock_failed | 34.96% | 34.91% | 29.14% | | __write_lock_failed | 10.14% | 10.68% | 7.51% | | mutex_spin_on_owner | 3.62% | 3.42% | 2.33% | | mspin_lock | N/A | N/A | 9.90% | | __mutex_lock_slowpath | 1.46% | 0.81% | 0.14% | | _raw_spin_lock | 2.25% | 2.50% | 1.10% | +-----------------------+-----------+---------+-------------+ The fserver workload for an 8-node system is dominated by the contention in the read/write lock. Mutex contention also plays a role. With the first patch only, mutex contention is down (as shown by the __mutex_lock_slowpath figure) which help a little bit. We saw only a few percents improvement with that. By applying patch 2 as well, the single mutex_spin_on_owner figure is now split out into an additional mspin_lock figure. The time increases from 3.42% to 11.23%. It shows a great reduction in contention among the spinners leading to a 30% improvement. The time ratio 9.9/2.33=4.3 indicates that there are on average 4+ spinners waiting in the spin_lock loop for each spinner in the mutex_spin_on_owner loop. Contention in other locking functions also go down by quite a lot. The table below shows the performance change of both patches 1 & 2 over patch 1 alone in other AIM7 workloads (at 8 nodes, hyperthreading off). +--------------+---------------+----------------+-----------------+ | Workload | mean % change | mean % change | mean % change | | | 10-100 users | 200-1000 users | 1100-2000 users | +--------------+---------------+----------------+-----------------+ | alltests | 0.0% | -0.8% | +0.6% | | five_sec | -0.3% | +0.8% | +0.8% | | high_systime | +0.4% | +2.4% | +2.1% | | new_fserver | +0.1% | +14.1% | +34.2% | | shared | -0.5% | -0.3% | -0.4% | | short | -1.7% | -9.8% | -8.3% | +--------------+---------------+----------------+-----------------+ The short workload is the only one that shows a decline in performance probably due to the spinner locking and queuing overhead. Signed-off-by: Waiman Long <Waiman.Long@hp.com> Reviewed-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Acked-by: Rik van Riel <riel@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Chandramouleeswaran Aswin <aswin@hp.com> Cc: Norton Scott J <scott.norton@hp.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: David Howells <dhowells@redhat.com> Cc: Dave Jones <davej@redhat.com> Cc: Clark Williams <williams@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1366226594-5506-4-git-send-email-Waiman.Long@hp.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-04-18 02:23:13 +07:00
/*
* In order to avoid a stampede of mutex spinners from acquiring the mutex
* more or less simultaneously, the spinners need to acquire a MCS lock
* first before spinning on the owner field.
*
*/
/*
* Mutex spinning code migrated from kernel/sched/core.c
*/
static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
{
if (lock->owner != owner)
return false;
/*
* Ensure we emit the owner->on_cpu, dereference _after_ checking
* lock->owner still matches owner, if that fails, owner might
* point to free()d memory, if it still matches, the rcu_read_lock()
* ensures the memory stays valid.
*/
barrier();
return owner->on_cpu;
}
/*
* Look out! "owner" is an entirely speculative pointer
* access and not reliable.
*/
static noinline
int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
{
rcu_read_lock();
while (owner_running(lock, owner)) {
if (need_resched())
break;
arch_mutex_cpu_relax();
}
rcu_read_unlock();
/*
* We break out the loop above on need_resched() and when the
* owner changed, which is a sign for heavy contention. Return
* success only when lock->owner is NULL.
*/
return lock->owner == NULL;
}
mutex: Queue mutex spinners with MCS lock to reduce cacheline contention The current mutex spinning code (with MUTEX_SPIN_ON_OWNER option turned on) allow multiple tasks to spin on a single mutex concurrently. A potential problem with the current approach is that when the mutex becomes available, all the spinning tasks will try to acquire the mutex more or less simultaneously. As a result, there will be a lot of cacheline bouncing especially on systems with a large number of CPUs. This patch tries to reduce this kind of contention by putting the mutex spinners into a queue so that only the first one in the queue will try to acquire the mutex. This will reduce contention and allow all the tasks to move forward faster. The queuing of mutex spinners is done using an MCS lock based implementation which will further reduce contention on the mutex cacheline than a similar ticket spinlock based implementation. This patch will add a new field into the mutex data structure for holding the MCS lock. This expands the mutex size by 8 bytes for 64-bit system and 4 bytes for 32-bit system. This overhead will be avoid if the MUTEX_SPIN_ON_OWNER option is turned off. The following table shows the jobs per minute (JPM) scalability data on an 8-node 80-core Westmere box with a 3.7.10 kernel. The numactl command is used to restrict the running of the fserver workloads to 1/2/4/8 nodes with hyperthreading off. +-----------------+-----------+-----------+-------------+----------+ | Configuration | Mean JPM | Mean JPM | Mean JPM | % Change | | | w/o patch | patch 1 | patches 1&2 | 1->1&2 | +-----------------+------------------------------------------------+ | | User Range 1100 - 2000 | +-----------------+------------------------------------------------+ | 8 nodes, HT off | 227972 | 227237 | 305043 | +34.2% | | 4 nodes, HT off | 393503 | 381558 | 394650 | +3.4% | | 2 nodes, HT off | 334957 | 325240 | 338853 | +4.2% | | 1 node , HT off | 198141 | 197972 | 198075 | +0.1% | +-----------------+------------------------------------------------+ | | User Range 200 - 1000 | +-----------------+------------------------------------------------+ | 8 nodes, HT off | 282325 | 312870 | 332185 | +6.2% | | 4 nodes, HT off | 390698 | 378279 | 393419 | +4.0% | | 2 nodes, HT off | 336986 | 326543 | 340260 | +4.2% | | 1 node , HT off | 197588 | 197622 | 197582 | 0.0% | +-----------------+-----------+-----------+-------------+----------+ At low user range 10-100, the JPM differences were within +/-1%. So they are not that interesting. The fserver workload uses mutex spinning extensively. With just the mutex change in the first patch, there is no noticeable change in performance. Rather, there is a slight drop in performance. This mutex spinning patch more than recovers the lost performance and show a significant increase of +30% at high user load with the full 8 nodes. Similar improvements were also seen in a 3.8 kernel. The table below shows the %time spent by different kernel functions as reported by perf when running the fserver workload at 1500 users with all 8 nodes. +-----------------------+-----------+---------+-------------+ | Function | % time | % time | % time | | | w/o patch | patch 1 | patches 1&2 | +-----------------------+-----------+---------+-------------+ | __read_lock_failed | 34.96% | 34.91% | 29.14% | | __write_lock_failed | 10.14% | 10.68% | 7.51% | | mutex_spin_on_owner | 3.62% | 3.42% | 2.33% | | mspin_lock | N/A | N/A | 9.90% | | __mutex_lock_slowpath | 1.46% | 0.81% | 0.14% | | _raw_spin_lock | 2.25% | 2.50% | 1.10% | +-----------------------+-----------+---------+-------------+ The fserver workload for an 8-node system is dominated by the contention in the read/write lock. Mutex contention also plays a role. With the first patch only, mutex contention is down (as shown by the __mutex_lock_slowpath figure) which help a little bit. We saw only a few percents improvement with that. By applying patch 2 as well, the single mutex_spin_on_owner figure is now split out into an additional mspin_lock figure. The time increases from 3.42% to 11.23%. It shows a great reduction in contention among the spinners leading to a 30% improvement. The time ratio 9.9/2.33=4.3 indicates that there are on average 4+ spinners waiting in the spin_lock loop for each spinner in the mutex_spin_on_owner loop. Contention in other locking functions also go down by quite a lot. The table below shows the performance change of both patches 1 & 2 over patch 1 alone in other AIM7 workloads (at 8 nodes, hyperthreading off). +--------------+---------------+----------------+-----------------+ | Workload | mean % change | mean % change | mean % change | | | 10-100 users | 200-1000 users | 1100-2000 users | +--------------+---------------+----------------+-----------------+ | alltests | 0.0% | -0.8% | +0.6% | | five_sec | -0.3% | +0.8% | +0.8% | | high_systime | +0.4% | +2.4% | +2.1% | | new_fserver | +0.1% | +14.1% | +34.2% | | shared | -0.5% | -0.3% | -0.4% | | short | -1.7% | -9.8% | -8.3% | +--------------+---------------+----------------+-----------------+ The short workload is the only one that shows a decline in performance probably due to the spinner locking and queuing overhead. Signed-off-by: Waiman Long <Waiman.Long@hp.com> Reviewed-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Acked-by: Rik van Riel <riel@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Chandramouleeswaran Aswin <aswin@hp.com> Cc: Norton Scott J <scott.norton@hp.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: David Howells <dhowells@redhat.com> Cc: Dave Jones <davej@redhat.com> Cc: Clark Williams <williams@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1366226594-5506-4-git-send-email-Waiman.Long@hp.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-04-18 02:23:13 +07:00
/*
* Initial check for entering the mutex spinning loop
*/
static inline int mutex_can_spin_on_owner(struct mutex *lock)
{
struct task_struct *owner;
mutex: Queue mutex spinners with MCS lock to reduce cacheline contention The current mutex spinning code (with MUTEX_SPIN_ON_OWNER option turned on) allow multiple tasks to spin on a single mutex concurrently. A potential problem with the current approach is that when the mutex becomes available, all the spinning tasks will try to acquire the mutex more or less simultaneously. As a result, there will be a lot of cacheline bouncing especially on systems with a large number of CPUs. This patch tries to reduce this kind of contention by putting the mutex spinners into a queue so that only the first one in the queue will try to acquire the mutex. This will reduce contention and allow all the tasks to move forward faster. The queuing of mutex spinners is done using an MCS lock based implementation which will further reduce contention on the mutex cacheline than a similar ticket spinlock based implementation. This patch will add a new field into the mutex data structure for holding the MCS lock. This expands the mutex size by 8 bytes for 64-bit system and 4 bytes for 32-bit system. This overhead will be avoid if the MUTEX_SPIN_ON_OWNER option is turned off. The following table shows the jobs per minute (JPM) scalability data on an 8-node 80-core Westmere box with a 3.7.10 kernel. The numactl command is used to restrict the running of the fserver workloads to 1/2/4/8 nodes with hyperthreading off. +-----------------+-----------+-----------+-------------+----------+ | Configuration | Mean JPM | Mean JPM | Mean JPM | % Change | | | w/o patch | patch 1 | patches 1&2 | 1->1&2 | +-----------------+------------------------------------------------+ | | User Range 1100 - 2000 | +-----------------+------------------------------------------------+ | 8 nodes, HT off | 227972 | 227237 | 305043 | +34.2% | | 4 nodes, HT off | 393503 | 381558 | 394650 | +3.4% | | 2 nodes, HT off | 334957 | 325240 | 338853 | +4.2% | | 1 node , HT off | 198141 | 197972 | 198075 | +0.1% | +-----------------+------------------------------------------------+ | | User Range 200 - 1000 | +-----------------+------------------------------------------------+ | 8 nodes, HT off | 282325 | 312870 | 332185 | +6.2% | | 4 nodes, HT off | 390698 | 378279 | 393419 | +4.0% | | 2 nodes, HT off | 336986 | 326543 | 340260 | +4.2% | | 1 node , HT off | 197588 | 197622 | 197582 | 0.0% | +-----------------+-----------+-----------+-------------+----------+ At low user range 10-100, the JPM differences were within +/-1%. So they are not that interesting. The fserver workload uses mutex spinning extensively. With just the mutex change in the first patch, there is no noticeable change in performance. Rather, there is a slight drop in performance. This mutex spinning patch more than recovers the lost performance and show a significant increase of +30% at high user load with the full 8 nodes. Similar improvements were also seen in a 3.8 kernel. The table below shows the %time spent by different kernel functions as reported by perf when running the fserver workload at 1500 users with all 8 nodes. +-----------------------+-----------+---------+-------------+ | Function | % time | % time | % time | | | w/o patch | patch 1 | patches 1&2 | +-----------------------+-----------+---------+-------------+ | __read_lock_failed | 34.96% | 34.91% | 29.14% | | __write_lock_failed | 10.14% | 10.68% | 7.51% | | mutex_spin_on_owner | 3.62% | 3.42% | 2.33% | | mspin_lock | N/A | N/A | 9.90% | | __mutex_lock_slowpath | 1.46% | 0.81% | 0.14% | | _raw_spin_lock | 2.25% | 2.50% | 1.10% | +-----------------------+-----------+---------+-------------+ The fserver workload for an 8-node system is dominated by the contention in the read/write lock. Mutex contention also plays a role. With the first patch only, mutex contention is down (as shown by the __mutex_lock_slowpath figure) which help a little bit. We saw only a few percents improvement with that. By applying patch 2 as well, the single mutex_spin_on_owner figure is now split out into an additional mspin_lock figure. The time increases from 3.42% to 11.23%. It shows a great reduction in contention among the spinners leading to a 30% improvement. The time ratio 9.9/2.33=4.3 indicates that there are on average 4+ spinners waiting in the spin_lock loop for each spinner in the mutex_spin_on_owner loop. Contention in other locking functions also go down by quite a lot. The table below shows the performance change of both patches 1 & 2 over patch 1 alone in other AIM7 workloads (at 8 nodes, hyperthreading off). +--------------+---------------+----------------+-----------------+ | Workload | mean % change | mean % change | mean % change | | | 10-100 users | 200-1000 users | 1100-2000 users | +--------------+---------------+----------------+-----------------+ | alltests | 0.0% | -0.8% | +0.6% | | five_sec | -0.3% | +0.8% | +0.8% | | high_systime | +0.4% | +2.4% | +2.1% | | new_fserver | +0.1% | +14.1% | +34.2% | | shared | -0.5% | -0.3% | -0.4% | | short | -1.7% | -9.8% | -8.3% | +--------------+---------------+----------------+-----------------+ The short workload is the only one that shows a decline in performance probably due to the spinner locking and queuing overhead. Signed-off-by: Waiman Long <Waiman.Long@hp.com> Reviewed-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Acked-by: Rik van Riel <riel@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Chandramouleeswaran Aswin <aswin@hp.com> Cc: Norton Scott J <scott.norton@hp.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: David Howells <dhowells@redhat.com> Cc: Dave Jones <davej@redhat.com> Cc: Clark Williams <williams@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1366226594-5506-4-git-send-email-Waiman.Long@hp.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-04-18 02:23:13 +07:00
int retval = 1;
if (need_resched())
return 0;
mutex: Queue mutex spinners with MCS lock to reduce cacheline contention The current mutex spinning code (with MUTEX_SPIN_ON_OWNER option turned on) allow multiple tasks to spin on a single mutex concurrently. A potential problem with the current approach is that when the mutex becomes available, all the spinning tasks will try to acquire the mutex more or less simultaneously. As a result, there will be a lot of cacheline bouncing especially on systems with a large number of CPUs. This patch tries to reduce this kind of contention by putting the mutex spinners into a queue so that only the first one in the queue will try to acquire the mutex. This will reduce contention and allow all the tasks to move forward faster. The queuing of mutex spinners is done using an MCS lock based implementation which will further reduce contention on the mutex cacheline than a similar ticket spinlock based implementation. This patch will add a new field into the mutex data structure for holding the MCS lock. This expands the mutex size by 8 bytes for 64-bit system and 4 bytes for 32-bit system. This overhead will be avoid if the MUTEX_SPIN_ON_OWNER option is turned off. The following table shows the jobs per minute (JPM) scalability data on an 8-node 80-core Westmere box with a 3.7.10 kernel. The numactl command is used to restrict the running of the fserver workloads to 1/2/4/8 nodes with hyperthreading off. +-----------------+-----------+-----------+-------------+----------+ | Configuration | Mean JPM | Mean JPM | Mean JPM | % Change | | | w/o patch | patch 1 | patches 1&2 | 1->1&2 | +-----------------+------------------------------------------------+ | | User Range 1100 - 2000 | +-----------------+------------------------------------------------+ | 8 nodes, HT off | 227972 | 227237 | 305043 | +34.2% | | 4 nodes, HT off | 393503 | 381558 | 394650 | +3.4% | | 2 nodes, HT off | 334957 | 325240 | 338853 | +4.2% | | 1 node , HT off | 198141 | 197972 | 198075 | +0.1% | +-----------------+------------------------------------------------+ | | User Range 200 - 1000 | +-----------------+------------------------------------------------+ | 8 nodes, HT off | 282325 | 312870 | 332185 | +6.2% | | 4 nodes, HT off | 390698 | 378279 | 393419 | +4.0% | | 2 nodes, HT off | 336986 | 326543 | 340260 | +4.2% | | 1 node , HT off | 197588 | 197622 | 197582 | 0.0% | +-----------------+-----------+-----------+-------------+----------+ At low user range 10-100, the JPM differences were within +/-1%. So they are not that interesting. The fserver workload uses mutex spinning extensively. With just the mutex change in the first patch, there is no noticeable change in performance. Rather, there is a slight drop in performance. This mutex spinning patch more than recovers the lost performance and show a significant increase of +30% at high user load with the full 8 nodes. Similar improvements were also seen in a 3.8 kernel. The table below shows the %time spent by different kernel functions as reported by perf when running the fserver workload at 1500 users with all 8 nodes. +-----------------------+-----------+---------+-------------+ | Function | % time | % time | % time | | | w/o patch | patch 1 | patches 1&2 | +-----------------------+-----------+---------+-------------+ | __read_lock_failed | 34.96% | 34.91% | 29.14% | | __write_lock_failed | 10.14% | 10.68% | 7.51% | | mutex_spin_on_owner | 3.62% | 3.42% | 2.33% | | mspin_lock | N/A | N/A | 9.90% | | __mutex_lock_slowpath | 1.46% | 0.81% | 0.14% | | _raw_spin_lock | 2.25% | 2.50% | 1.10% | +-----------------------+-----------+---------+-------------+ The fserver workload for an 8-node system is dominated by the contention in the read/write lock. Mutex contention also plays a role. With the first patch only, mutex contention is down (as shown by the __mutex_lock_slowpath figure) which help a little bit. We saw only a few percents improvement with that. By applying patch 2 as well, the single mutex_spin_on_owner figure is now split out into an additional mspin_lock figure. The time increases from 3.42% to 11.23%. It shows a great reduction in contention among the spinners leading to a 30% improvement. The time ratio 9.9/2.33=4.3 indicates that there are on average 4+ spinners waiting in the spin_lock loop for each spinner in the mutex_spin_on_owner loop. Contention in other locking functions also go down by quite a lot. The table below shows the performance change of both patches 1 & 2 over patch 1 alone in other AIM7 workloads (at 8 nodes, hyperthreading off). +--------------+---------------+----------------+-----------------+ | Workload | mean % change | mean % change | mean % change | | | 10-100 users | 200-1000 users | 1100-2000 users | +--------------+---------------+----------------+-----------------+ | alltests | 0.0% | -0.8% | +0.6% | | five_sec | -0.3% | +0.8% | +0.8% | | high_systime | +0.4% | +2.4% | +2.1% | | new_fserver | +0.1% | +14.1% | +34.2% | | shared | -0.5% | -0.3% | -0.4% | | short | -1.7% | -9.8% | -8.3% | +--------------+---------------+----------------+-----------------+ The short workload is the only one that shows a decline in performance probably due to the spinner locking and queuing overhead. Signed-off-by: Waiman Long <Waiman.Long@hp.com> Reviewed-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Acked-by: Rik van Riel <riel@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Chandramouleeswaran Aswin <aswin@hp.com> Cc: Norton Scott J <scott.norton@hp.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: David Howells <dhowells@redhat.com> Cc: Dave Jones <davej@redhat.com> Cc: Clark Williams <williams@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1366226594-5506-4-git-send-email-Waiman.Long@hp.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-04-18 02:23:13 +07:00
rcu_read_lock();
owner = ACCESS_ONCE(lock->owner);
if (owner)
retval = owner->on_cpu;
mutex: Queue mutex spinners with MCS lock to reduce cacheline contention The current mutex spinning code (with MUTEX_SPIN_ON_OWNER option turned on) allow multiple tasks to spin on a single mutex concurrently. A potential problem with the current approach is that when the mutex becomes available, all the spinning tasks will try to acquire the mutex more or less simultaneously. As a result, there will be a lot of cacheline bouncing especially on systems with a large number of CPUs. This patch tries to reduce this kind of contention by putting the mutex spinners into a queue so that only the first one in the queue will try to acquire the mutex. This will reduce contention and allow all the tasks to move forward faster. The queuing of mutex spinners is done using an MCS lock based implementation which will further reduce contention on the mutex cacheline than a similar ticket spinlock based implementation. This patch will add a new field into the mutex data structure for holding the MCS lock. This expands the mutex size by 8 bytes for 64-bit system and 4 bytes for 32-bit system. This overhead will be avoid if the MUTEX_SPIN_ON_OWNER option is turned off. The following table shows the jobs per minute (JPM) scalability data on an 8-node 80-core Westmere box with a 3.7.10 kernel. The numactl command is used to restrict the running of the fserver workloads to 1/2/4/8 nodes with hyperthreading off. +-----------------+-----------+-----------+-------------+----------+ | Configuration | Mean JPM | Mean JPM | Mean JPM | % Change | | | w/o patch | patch 1 | patches 1&2 | 1->1&2 | +-----------------+------------------------------------------------+ | | User Range 1100 - 2000 | +-----------------+------------------------------------------------+ | 8 nodes, HT off | 227972 | 227237 | 305043 | +34.2% | | 4 nodes, HT off | 393503 | 381558 | 394650 | +3.4% | | 2 nodes, HT off | 334957 | 325240 | 338853 | +4.2% | | 1 node , HT off | 198141 | 197972 | 198075 | +0.1% | +-----------------+------------------------------------------------+ | | User Range 200 - 1000 | +-----------------+------------------------------------------------+ | 8 nodes, HT off | 282325 | 312870 | 332185 | +6.2% | | 4 nodes, HT off | 390698 | 378279 | 393419 | +4.0% | | 2 nodes, HT off | 336986 | 326543 | 340260 | +4.2% | | 1 node , HT off | 197588 | 197622 | 197582 | 0.0% | +-----------------+-----------+-----------+-------------+----------+ At low user range 10-100, the JPM differences were within +/-1%. So they are not that interesting. The fserver workload uses mutex spinning extensively. With just the mutex change in the first patch, there is no noticeable change in performance. Rather, there is a slight drop in performance. This mutex spinning patch more than recovers the lost performance and show a significant increase of +30% at high user load with the full 8 nodes. Similar improvements were also seen in a 3.8 kernel. The table below shows the %time spent by different kernel functions as reported by perf when running the fserver workload at 1500 users with all 8 nodes. +-----------------------+-----------+---------+-------------+ | Function | % time | % time | % time | | | w/o patch | patch 1 | patches 1&2 | +-----------------------+-----------+---------+-------------+ | __read_lock_failed | 34.96% | 34.91% | 29.14% | | __write_lock_failed | 10.14% | 10.68% | 7.51% | | mutex_spin_on_owner | 3.62% | 3.42% | 2.33% | | mspin_lock | N/A | N/A | 9.90% | | __mutex_lock_slowpath | 1.46% | 0.81% | 0.14% | | _raw_spin_lock | 2.25% | 2.50% | 1.10% | +-----------------------+-----------+---------+-------------+ The fserver workload for an 8-node system is dominated by the contention in the read/write lock. Mutex contention also plays a role. With the first patch only, mutex contention is down (as shown by the __mutex_lock_slowpath figure) which help a little bit. We saw only a few percents improvement with that. By applying patch 2 as well, the single mutex_spin_on_owner figure is now split out into an additional mspin_lock figure. The time increases from 3.42% to 11.23%. It shows a great reduction in contention among the spinners leading to a 30% improvement. The time ratio 9.9/2.33=4.3 indicates that there are on average 4+ spinners waiting in the spin_lock loop for each spinner in the mutex_spin_on_owner loop. Contention in other locking functions also go down by quite a lot. The table below shows the performance change of both patches 1 & 2 over patch 1 alone in other AIM7 workloads (at 8 nodes, hyperthreading off). +--------------+---------------+----------------+-----------------+ | Workload | mean % change | mean % change | mean % change | | | 10-100 users | 200-1000 users | 1100-2000 users | +--------------+---------------+----------------+-----------------+ | alltests | 0.0% | -0.8% | +0.6% | | five_sec | -0.3% | +0.8% | +0.8% | | high_systime | +0.4% | +2.4% | +2.1% | | new_fserver | +0.1% | +14.1% | +34.2% | | shared | -0.5% | -0.3% | -0.4% | | short | -1.7% | -9.8% | -8.3% | +--------------+---------------+----------------+-----------------+ The short workload is the only one that shows a decline in performance probably due to the spinner locking and queuing overhead. Signed-off-by: Waiman Long <Waiman.Long@hp.com> Reviewed-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Acked-by: Rik van Riel <riel@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Chandramouleeswaran Aswin <aswin@hp.com> Cc: Norton Scott J <scott.norton@hp.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: David Howells <dhowells@redhat.com> Cc: Dave Jones <davej@redhat.com> Cc: Clark Williams <williams@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1366226594-5506-4-git-send-email-Waiman.Long@hp.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-04-18 02:23:13 +07:00
rcu_read_unlock();
/*
* if lock->owner is not set, the mutex owner may have just acquired
* it and not set the owner yet or the mutex has been released.
*/
return retval;
}
#endif
static __used noinline void __sched __mutex_unlock_slowpath(atomic_t *lock_count);
/**
* mutex_unlock - release the mutex
* @lock: the mutex to be released
*
* Unlock a mutex that has been locked by this task previously.
*
* This function must not be used in interrupt context. Unlocking
* of a not locked mutex is not allowed.
*
* This function is similar to (but not equivalent to) up().
*/
void __sched mutex_unlock(struct mutex *lock)
{
/*
* The unlocking fastpath is the 0->1 transition from 'locked'
* into 'unlocked' state:
*/
mutex: implement adaptive spinning Change mutex contention behaviour such that it will sometimes busy wait on acquisition - moving its behaviour closer to that of spinlocks. This concept got ported to mainline from the -rt tree, where it was originally implemented for rtmutexes by Steven Rostedt, based on work by Gregory Haskins. Testing with Ingo's test-mutex application (http://lkml.org/lkml/2006/1/8/50) gave a 345% boost for VFS scalability on my testbox: # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 296604 # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 85870 The key criteria for the busy wait is that the lock owner has to be running on a (different) cpu. The idea is that as long as the owner is running, there is a fair chance it'll release the lock soon, and thus we'll be better off spinning instead of blocking/scheduling. Since regular mutexes (as opposed to rtmutexes) do not atomically track the owner, we add the owner in a non-atomic fashion and deal with the races in the slowpath. Furthermore, to ease the testing of the performance impact of this new code, there is means to disable this behaviour runtime (without having to reboot the system), when scheduler debugging is enabled (CONFIG_SCHED_DEBUG=y), by issuing the following command: # echo NO_OWNER_SPIN > /debug/sched_features This command re-enables spinning again (this is also the default): # echo OWNER_SPIN > /debug/sched_features Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-01-12 20:01:47 +07:00
#ifndef CONFIG_DEBUG_MUTEXES
/*
* When debugging is enabled we must not clear the owner before time,
* the slow path will always be taken, and that clears the owner field
* after verifying that it was indeed current.
*/
mutex_clear_owner(lock);
#endif
__mutex_fastpath_unlock(&lock->count, __mutex_unlock_slowpath);
}
EXPORT_SYMBOL(mutex_unlock);
/**
* ww_mutex_unlock - release the w/w mutex
* @lock: the mutex to be released
*
* Unlock a mutex that has been locked by this task previously with any of the
* ww_mutex_lock* functions (with or without an acquire context). It is
* forbidden to release the locks after releasing the acquire context.
*
* This function must not be used in interrupt context. Unlocking
* of a unlocked mutex is not allowed.
*/
void __sched ww_mutex_unlock(struct ww_mutex *lock)
{
/*
* The unlocking fastpath is the 0->1 transition from 'locked'
* into 'unlocked' state:
*/
if (lock->ctx) {
#ifdef CONFIG_DEBUG_MUTEXES
DEBUG_LOCKS_WARN_ON(!lock->ctx->acquired);
#endif
if (lock->ctx->acquired > 0)
lock->ctx->acquired--;
lock->ctx = NULL;
}
#ifndef CONFIG_DEBUG_MUTEXES
/*
* When debugging is enabled we must not clear the owner before time,
* the slow path will always be taken, and that clears the owner field
* after verifying that it was indeed current.
*/
mutex_clear_owner(&lock->base);
#endif
__mutex_fastpath_unlock(&lock->base.count, __mutex_unlock_slowpath);
}
EXPORT_SYMBOL(ww_mutex_unlock);
static inline int __sched
__mutex_lock_check_stamp(struct mutex *lock, struct ww_acquire_ctx *ctx)
{
struct ww_mutex *ww = container_of(lock, struct ww_mutex, base);
struct ww_acquire_ctx *hold_ctx = ACCESS_ONCE(ww->ctx);
if (!hold_ctx)
return 0;
if (unlikely(ctx == hold_ctx))
return -EALREADY;
if (ctx->stamp - hold_ctx->stamp <= LONG_MAX &&
(ctx->stamp != hold_ctx->stamp || ctx > hold_ctx)) {
#ifdef CONFIG_DEBUG_MUTEXES
DEBUG_LOCKS_WARN_ON(ctx->contending_lock);
ctx->contending_lock = ww;
#endif
return -EDEADLK;
}
return 0;
}
static __always_inline void ww_mutex_lock_acquired(struct ww_mutex *ww,
struct ww_acquire_ctx *ww_ctx)
{
#ifdef CONFIG_DEBUG_MUTEXES
/*
* If this WARN_ON triggers, you used ww_mutex_lock to acquire,
* but released with a normal mutex_unlock in this call.
*
* This should never happen, always use ww_mutex_unlock.
*/
DEBUG_LOCKS_WARN_ON(ww->ctx);
/*
* Not quite done after calling ww_acquire_done() ?
*/
DEBUG_LOCKS_WARN_ON(ww_ctx->done_acquire);
if (ww_ctx->contending_lock) {
/*
* After -EDEADLK you tried to
* acquire a different ww_mutex? Bad!
*/
DEBUG_LOCKS_WARN_ON(ww_ctx->contending_lock != ww);
/*
* You called ww_mutex_lock after receiving -EDEADLK,
* but 'forgot' to unlock everything else first?
*/
DEBUG_LOCKS_WARN_ON(ww_ctx->acquired > 0);
ww_ctx->contending_lock = NULL;
}
/*
* Naughty, using a different class will lead to undefined behavior!
*/
DEBUG_LOCKS_WARN_ON(ww_ctx->ww_class != ww->ww_class);
#endif
ww_ctx->acquired++;
}
/*
* after acquiring lock with fastpath or when we lost out in contested
* slowpath, set ctx and wake up any waiters so they can recheck.
*
* This function is never called when CONFIG_DEBUG_LOCK_ALLOC is set,
* as the fastpath and opportunistic spinning are disabled in that case.
*/
static __always_inline void
ww_mutex_set_context_fastpath(struct ww_mutex *lock,
struct ww_acquire_ctx *ctx)
{
unsigned long flags;
struct mutex_waiter *cur;
ww_mutex_lock_acquired(lock, ctx);
lock->ctx = ctx;
/*
* The lock->ctx update should be visible on all cores before
* the atomic read is done, otherwise contended waiters might be
* missed. The contended waiters will either see ww_ctx == NULL
* and keep spinning, or it will acquire wait_lock, add itself
* to waiter list and sleep.
*/
smp_mb(); /* ^^^ */
/*
* Check if lock is contended, if not there is nobody to wake up
*/
if (likely(atomic_read(&lock->base.count) == 0))
return;
/*
* Uh oh, we raced in fastpath, wake up everyone in this case,
* so they can see the new lock->ctx.
*/
spin_lock_mutex(&lock->base.wait_lock, flags);
list_for_each_entry(cur, &lock->base.wait_list, list) {
debug_mutex_wake_waiter(&lock->base, cur);
wake_up_process(cur->task);
}
spin_unlock_mutex(&lock->base.wait_lock, flags);
}
/*
* Lock a mutex (possibly interruptible), slowpath:
*/
static __always_inline int __sched
__mutex_lock_common(struct mutex *lock, long state, unsigned int subclass,
struct lockdep_map *nest_lock, unsigned long ip,
struct ww_acquire_ctx *ww_ctx, const bool use_ww_ctx)
{
struct task_struct *task = current;
struct mutex_waiter waiter;
unsigned long flags;
int ret;
preempt_disable();
mutex_acquire_nest(&lock->dep_map, subclass, 0, nest_lock, ip);
#ifdef CONFIG_MUTEX_SPIN_ON_OWNER
mutex: implement adaptive spinning Change mutex contention behaviour such that it will sometimes busy wait on acquisition - moving its behaviour closer to that of spinlocks. This concept got ported to mainline from the -rt tree, where it was originally implemented for rtmutexes by Steven Rostedt, based on work by Gregory Haskins. Testing with Ingo's test-mutex application (http://lkml.org/lkml/2006/1/8/50) gave a 345% boost for VFS scalability on my testbox: # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 296604 # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 85870 The key criteria for the busy wait is that the lock owner has to be running on a (different) cpu. The idea is that as long as the owner is running, there is a fair chance it'll release the lock soon, and thus we'll be better off spinning instead of blocking/scheduling. Since regular mutexes (as opposed to rtmutexes) do not atomically track the owner, we add the owner in a non-atomic fashion and deal with the races in the slowpath. Furthermore, to ease the testing of the performance impact of this new code, there is means to disable this behaviour runtime (without having to reboot the system), when scheduler debugging is enabled (CONFIG_SCHED_DEBUG=y), by issuing the following command: # echo NO_OWNER_SPIN > /debug/sched_features This command re-enables spinning again (this is also the default): # echo OWNER_SPIN > /debug/sched_features Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-01-12 20:01:47 +07:00
/*
* Optimistic spinning.
*
* We try to spin for acquisition when we find that there are no
* pending waiters and the lock owner is currently running on a
* (different) CPU.
*
* The rationale is that if the lock owner is running, it is likely to
* release the lock soon.
*
* Since this needs the lock owner, and this mutex implementation
* doesn't track the owner atomically in the lock field, we need to
* track it non-atomically.
*
* We can't do this for DEBUG_MUTEXES because that relies on wait_lock
* to serialize everything.
mutex: Queue mutex spinners with MCS lock to reduce cacheline contention The current mutex spinning code (with MUTEX_SPIN_ON_OWNER option turned on) allow multiple tasks to spin on a single mutex concurrently. A potential problem with the current approach is that when the mutex becomes available, all the spinning tasks will try to acquire the mutex more or less simultaneously. As a result, there will be a lot of cacheline bouncing especially on systems with a large number of CPUs. This patch tries to reduce this kind of contention by putting the mutex spinners into a queue so that only the first one in the queue will try to acquire the mutex. This will reduce contention and allow all the tasks to move forward faster. The queuing of mutex spinners is done using an MCS lock based implementation which will further reduce contention on the mutex cacheline than a similar ticket spinlock based implementation. This patch will add a new field into the mutex data structure for holding the MCS lock. This expands the mutex size by 8 bytes for 64-bit system and 4 bytes for 32-bit system. This overhead will be avoid if the MUTEX_SPIN_ON_OWNER option is turned off. The following table shows the jobs per minute (JPM) scalability data on an 8-node 80-core Westmere box with a 3.7.10 kernel. The numactl command is used to restrict the running of the fserver workloads to 1/2/4/8 nodes with hyperthreading off. +-----------------+-----------+-----------+-------------+----------+ | Configuration | Mean JPM | Mean JPM | Mean JPM | % Change | | | w/o patch | patch 1 | patches 1&2 | 1->1&2 | +-----------------+------------------------------------------------+ | | User Range 1100 - 2000 | +-----------------+------------------------------------------------+ | 8 nodes, HT off | 227972 | 227237 | 305043 | +34.2% | | 4 nodes, HT off | 393503 | 381558 | 394650 | +3.4% | | 2 nodes, HT off | 334957 | 325240 | 338853 | +4.2% | | 1 node , HT off | 198141 | 197972 | 198075 | +0.1% | +-----------------+------------------------------------------------+ | | User Range 200 - 1000 | +-----------------+------------------------------------------------+ | 8 nodes, HT off | 282325 | 312870 | 332185 | +6.2% | | 4 nodes, HT off | 390698 | 378279 | 393419 | +4.0% | | 2 nodes, HT off | 336986 | 326543 | 340260 | +4.2% | | 1 node , HT off | 197588 | 197622 | 197582 | 0.0% | +-----------------+-----------+-----------+-------------+----------+ At low user range 10-100, the JPM differences were within +/-1%. So they are not that interesting. The fserver workload uses mutex spinning extensively. With just the mutex change in the first patch, there is no noticeable change in performance. Rather, there is a slight drop in performance. This mutex spinning patch more than recovers the lost performance and show a significant increase of +30% at high user load with the full 8 nodes. Similar improvements were also seen in a 3.8 kernel. The table below shows the %time spent by different kernel functions as reported by perf when running the fserver workload at 1500 users with all 8 nodes. +-----------------------+-----------+---------+-------------+ | Function | % time | % time | % time | | | w/o patch | patch 1 | patches 1&2 | +-----------------------+-----------+---------+-------------+ | __read_lock_failed | 34.96% | 34.91% | 29.14% | | __write_lock_failed | 10.14% | 10.68% | 7.51% | | mutex_spin_on_owner | 3.62% | 3.42% | 2.33% | | mspin_lock | N/A | N/A | 9.90% | | __mutex_lock_slowpath | 1.46% | 0.81% | 0.14% | | _raw_spin_lock | 2.25% | 2.50% | 1.10% | +-----------------------+-----------+---------+-------------+ The fserver workload for an 8-node system is dominated by the contention in the read/write lock. Mutex contention also plays a role. With the first patch only, mutex contention is down (as shown by the __mutex_lock_slowpath figure) which help a little bit. We saw only a few percents improvement with that. By applying patch 2 as well, the single mutex_spin_on_owner figure is now split out into an additional mspin_lock figure. The time increases from 3.42% to 11.23%. It shows a great reduction in contention among the spinners leading to a 30% improvement. The time ratio 9.9/2.33=4.3 indicates that there are on average 4+ spinners waiting in the spin_lock loop for each spinner in the mutex_spin_on_owner loop. Contention in other locking functions also go down by quite a lot. The table below shows the performance change of both patches 1 & 2 over patch 1 alone in other AIM7 workloads (at 8 nodes, hyperthreading off). +--------------+---------------+----------------+-----------------+ | Workload | mean % change | mean % change | mean % change | | | 10-100 users | 200-1000 users | 1100-2000 users | +--------------+---------------+----------------+-----------------+ | alltests | 0.0% | -0.8% | +0.6% | | five_sec | -0.3% | +0.8% | +0.8% | | high_systime | +0.4% | +2.4% | +2.1% | | new_fserver | +0.1% | +14.1% | +34.2% | | shared | -0.5% | -0.3% | -0.4% | | short | -1.7% | -9.8% | -8.3% | +--------------+---------------+----------------+-----------------+ The short workload is the only one that shows a decline in performance probably due to the spinner locking and queuing overhead. Signed-off-by: Waiman Long <Waiman.Long@hp.com> Reviewed-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Acked-by: Rik van Riel <riel@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Chandramouleeswaran Aswin <aswin@hp.com> Cc: Norton Scott J <scott.norton@hp.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: David Howells <dhowells@redhat.com> Cc: Dave Jones <davej@redhat.com> Cc: Clark Williams <williams@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1366226594-5506-4-git-send-email-Waiman.Long@hp.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-04-18 02:23:13 +07:00
*
* The mutex spinners are queued up using MCS lock so that only one
* spinner can compete for the mutex. However, if mutex spinning isn't
* going to happen, there is no point in going through the lock/unlock
* overhead.
mutex: implement adaptive spinning Change mutex contention behaviour such that it will sometimes busy wait on acquisition - moving its behaviour closer to that of spinlocks. This concept got ported to mainline from the -rt tree, where it was originally implemented for rtmutexes by Steven Rostedt, based on work by Gregory Haskins. Testing with Ingo's test-mutex application (http://lkml.org/lkml/2006/1/8/50) gave a 345% boost for VFS scalability on my testbox: # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 296604 # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 85870 The key criteria for the busy wait is that the lock owner has to be running on a (different) cpu. The idea is that as long as the owner is running, there is a fair chance it'll release the lock soon, and thus we'll be better off spinning instead of blocking/scheduling. Since regular mutexes (as opposed to rtmutexes) do not atomically track the owner, we add the owner in a non-atomic fashion and deal with the races in the slowpath. Furthermore, to ease the testing of the performance impact of this new code, there is means to disable this behaviour runtime (without having to reboot the system), when scheduler debugging is enabled (CONFIG_SCHED_DEBUG=y), by issuing the following command: # echo NO_OWNER_SPIN > /debug/sched_features This command re-enables spinning again (this is also the default): # echo OWNER_SPIN > /debug/sched_features Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-01-12 20:01:47 +07:00
*/
mutex: Queue mutex spinners with MCS lock to reduce cacheline contention The current mutex spinning code (with MUTEX_SPIN_ON_OWNER option turned on) allow multiple tasks to spin on a single mutex concurrently. A potential problem with the current approach is that when the mutex becomes available, all the spinning tasks will try to acquire the mutex more or less simultaneously. As a result, there will be a lot of cacheline bouncing especially on systems with a large number of CPUs. This patch tries to reduce this kind of contention by putting the mutex spinners into a queue so that only the first one in the queue will try to acquire the mutex. This will reduce contention and allow all the tasks to move forward faster. The queuing of mutex spinners is done using an MCS lock based implementation which will further reduce contention on the mutex cacheline than a similar ticket spinlock based implementation. This patch will add a new field into the mutex data structure for holding the MCS lock. This expands the mutex size by 8 bytes for 64-bit system and 4 bytes for 32-bit system. This overhead will be avoid if the MUTEX_SPIN_ON_OWNER option is turned off. The following table shows the jobs per minute (JPM) scalability data on an 8-node 80-core Westmere box with a 3.7.10 kernel. The numactl command is used to restrict the running of the fserver workloads to 1/2/4/8 nodes with hyperthreading off. +-----------------+-----------+-----------+-------------+----------+ | Configuration | Mean JPM | Mean JPM | Mean JPM | % Change | | | w/o patch | patch 1 | patches 1&2 | 1->1&2 | +-----------------+------------------------------------------------+ | | User Range 1100 - 2000 | +-----------------+------------------------------------------------+ | 8 nodes, HT off | 227972 | 227237 | 305043 | +34.2% | | 4 nodes, HT off | 393503 | 381558 | 394650 | +3.4% | | 2 nodes, HT off | 334957 | 325240 | 338853 | +4.2% | | 1 node , HT off | 198141 | 197972 | 198075 | +0.1% | +-----------------+------------------------------------------------+ | | User Range 200 - 1000 | +-----------------+------------------------------------------------+ | 8 nodes, HT off | 282325 | 312870 | 332185 | +6.2% | | 4 nodes, HT off | 390698 | 378279 | 393419 | +4.0% | | 2 nodes, HT off | 336986 | 326543 | 340260 | +4.2% | | 1 node , HT off | 197588 | 197622 | 197582 | 0.0% | +-----------------+-----------+-----------+-------------+----------+ At low user range 10-100, the JPM differences were within +/-1%. So they are not that interesting. The fserver workload uses mutex spinning extensively. With just the mutex change in the first patch, there is no noticeable change in performance. Rather, there is a slight drop in performance. This mutex spinning patch more than recovers the lost performance and show a significant increase of +30% at high user load with the full 8 nodes. Similar improvements were also seen in a 3.8 kernel. The table below shows the %time spent by different kernel functions as reported by perf when running the fserver workload at 1500 users with all 8 nodes. +-----------------------+-----------+---------+-------------+ | Function | % time | % time | % time | | | w/o patch | patch 1 | patches 1&2 | +-----------------------+-----------+---------+-------------+ | __read_lock_failed | 34.96% | 34.91% | 29.14% | | __write_lock_failed | 10.14% | 10.68% | 7.51% | | mutex_spin_on_owner | 3.62% | 3.42% | 2.33% | | mspin_lock | N/A | N/A | 9.90% | | __mutex_lock_slowpath | 1.46% | 0.81% | 0.14% | | _raw_spin_lock | 2.25% | 2.50% | 1.10% | +-----------------------+-----------+---------+-------------+ The fserver workload for an 8-node system is dominated by the contention in the read/write lock. Mutex contention also plays a role. With the first patch only, mutex contention is down (as shown by the __mutex_lock_slowpath figure) which help a little bit. We saw only a few percents improvement with that. By applying patch 2 as well, the single mutex_spin_on_owner figure is now split out into an additional mspin_lock figure. The time increases from 3.42% to 11.23%. It shows a great reduction in contention among the spinners leading to a 30% improvement. The time ratio 9.9/2.33=4.3 indicates that there are on average 4+ spinners waiting in the spin_lock loop for each spinner in the mutex_spin_on_owner loop. Contention in other locking functions also go down by quite a lot. The table below shows the performance change of both patches 1 & 2 over patch 1 alone in other AIM7 workloads (at 8 nodes, hyperthreading off). +--------------+---------------+----------------+-----------------+ | Workload | mean % change | mean % change | mean % change | | | 10-100 users | 200-1000 users | 1100-2000 users | +--------------+---------------+----------------+-----------------+ | alltests | 0.0% | -0.8% | +0.6% | | five_sec | -0.3% | +0.8% | +0.8% | | high_systime | +0.4% | +2.4% | +2.1% | | new_fserver | +0.1% | +14.1% | +34.2% | | shared | -0.5% | -0.3% | -0.4% | | short | -1.7% | -9.8% | -8.3% | +--------------+---------------+----------------+-----------------+ The short workload is the only one that shows a decline in performance probably due to the spinner locking and queuing overhead. Signed-off-by: Waiman Long <Waiman.Long@hp.com> Reviewed-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Acked-by: Rik van Riel <riel@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Chandramouleeswaran Aswin <aswin@hp.com> Cc: Norton Scott J <scott.norton@hp.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: David Howells <dhowells@redhat.com> Cc: Dave Jones <davej@redhat.com> Cc: Clark Williams <williams@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1366226594-5506-4-git-send-email-Waiman.Long@hp.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-04-18 02:23:13 +07:00
if (!mutex_can_spin_on_owner(lock))
goto slowpath;
mutex: implement adaptive spinning Change mutex contention behaviour such that it will sometimes busy wait on acquisition - moving its behaviour closer to that of spinlocks. This concept got ported to mainline from the -rt tree, where it was originally implemented for rtmutexes by Steven Rostedt, based on work by Gregory Haskins. Testing with Ingo's test-mutex application (http://lkml.org/lkml/2006/1/8/50) gave a 345% boost for VFS scalability on my testbox: # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 296604 # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 85870 The key criteria for the busy wait is that the lock owner has to be running on a (different) cpu. The idea is that as long as the owner is running, there is a fair chance it'll release the lock soon, and thus we'll be better off spinning instead of blocking/scheduling. Since regular mutexes (as opposed to rtmutexes) do not atomically track the owner, we add the owner in a non-atomic fashion and deal with the races in the slowpath. Furthermore, to ease the testing of the performance impact of this new code, there is means to disable this behaviour runtime (without having to reboot the system), when scheduler debugging is enabled (CONFIG_SCHED_DEBUG=y), by issuing the following command: # echo NO_OWNER_SPIN > /debug/sched_features This command re-enables spinning again (this is also the default): # echo OWNER_SPIN > /debug/sched_features Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-01-12 20:01:47 +07:00
mcs_spin_lock(&lock->mcs_lock, &node);
mutex: implement adaptive spinning Change mutex contention behaviour such that it will sometimes busy wait on acquisition - moving its behaviour closer to that of spinlocks. This concept got ported to mainline from the -rt tree, where it was originally implemented for rtmutexes by Steven Rostedt, based on work by Gregory Haskins. Testing with Ingo's test-mutex application (http://lkml.org/lkml/2006/1/8/50) gave a 345% boost for VFS scalability on my testbox: # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 296604 # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 85870 The key criteria for the busy wait is that the lock owner has to be running on a (different) cpu. The idea is that as long as the owner is running, there is a fair chance it'll release the lock soon, and thus we'll be better off spinning instead of blocking/scheduling. Since regular mutexes (as opposed to rtmutexes) do not atomically track the owner, we add the owner in a non-atomic fashion and deal with the races in the slowpath. Furthermore, to ease the testing of the performance impact of this new code, there is means to disable this behaviour runtime (without having to reboot the system), when scheduler debugging is enabled (CONFIG_SCHED_DEBUG=y), by issuing the following command: # echo NO_OWNER_SPIN > /debug/sched_features This command re-enables spinning again (this is also the default): # echo OWNER_SPIN > /debug/sched_features Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-01-12 20:01:47 +07:00
for (;;) {
struct task_struct *owner;
mutex: implement adaptive spinning Change mutex contention behaviour such that it will sometimes busy wait on acquisition - moving its behaviour closer to that of spinlocks. This concept got ported to mainline from the -rt tree, where it was originally implemented for rtmutexes by Steven Rostedt, based on work by Gregory Haskins. Testing with Ingo's test-mutex application (http://lkml.org/lkml/2006/1/8/50) gave a 345% boost for VFS scalability on my testbox: # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 296604 # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 85870 The key criteria for the busy wait is that the lock owner has to be running on a (different) cpu. The idea is that as long as the owner is running, there is a fair chance it'll release the lock soon, and thus we'll be better off spinning instead of blocking/scheduling. Since regular mutexes (as opposed to rtmutexes) do not atomically track the owner, we add the owner in a non-atomic fashion and deal with the races in the slowpath. Furthermore, to ease the testing of the performance impact of this new code, there is means to disable this behaviour runtime (without having to reboot the system), when scheduler debugging is enabled (CONFIG_SCHED_DEBUG=y), by issuing the following command: # echo NO_OWNER_SPIN > /debug/sched_features This command re-enables spinning again (this is also the default): # echo OWNER_SPIN > /debug/sched_features Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-01-12 20:01:47 +07:00
if (use_ww_ctx && ww_ctx->acquired > 0) {
struct ww_mutex *ww;
ww = container_of(lock, struct ww_mutex, base);
/*
* If ww->ctx is set the contents are undefined, only
* by acquiring wait_lock there is a guarantee that
* they are not invalid when reading.
*
* As such, when deadlock detection needs to be
* performed the optimistic spinning cannot be done.
*/
if (ACCESS_ONCE(ww->ctx))
break;
}
mutex: implement adaptive spinning Change mutex contention behaviour such that it will sometimes busy wait on acquisition - moving its behaviour closer to that of spinlocks. This concept got ported to mainline from the -rt tree, where it was originally implemented for rtmutexes by Steven Rostedt, based on work by Gregory Haskins. Testing with Ingo's test-mutex application (http://lkml.org/lkml/2006/1/8/50) gave a 345% boost for VFS scalability on my testbox: # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 296604 # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 85870 The key criteria for the busy wait is that the lock owner has to be running on a (different) cpu. The idea is that as long as the owner is running, there is a fair chance it'll release the lock soon, and thus we'll be better off spinning instead of blocking/scheduling. Since regular mutexes (as opposed to rtmutexes) do not atomically track the owner, we add the owner in a non-atomic fashion and deal with the races in the slowpath. Furthermore, to ease the testing of the performance impact of this new code, there is means to disable this behaviour runtime (without having to reboot the system), when scheduler debugging is enabled (CONFIG_SCHED_DEBUG=y), by issuing the following command: # echo NO_OWNER_SPIN > /debug/sched_features This command re-enables spinning again (this is also the default): # echo OWNER_SPIN > /debug/sched_features Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-01-12 20:01:47 +07:00
/*
* If there's an owner, wait for it to either
* release the lock or go to sleep.
*/
owner = ACCESS_ONCE(lock->owner);
if (owner && !mutex_spin_on_owner(lock, owner))
break;
mutex: implement adaptive spinning Change mutex contention behaviour such that it will sometimes busy wait on acquisition - moving its behaviour closer to that of spinlocks. This concept got ported to mainline from the -rt tree, where it was originally implemented for rtmutexes by Steven Rostedt, based on work by Gregory Haskins. Testing with Ingo's test-mutex application (http://lkml.org/lkml/2006/1/8/50) gave a 345% boost for VFS scalability on my testbox: # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 296604 # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 85870 The key criteria for the busy wait is that the lock owner has to be running on a (different) cpu. The idea is that as long as the owner is running, there is a fair chance it'll release the lock soon, and thus we'll be better off spinning instead of blocking/scheduling. Since regular mutexes (as opposed to rtmutexes) do not atomically track the owner, we add the owner in a non-atomic fashion and deal with the races in the slowpath. Furthermore, to ease the testing of the performance impact of this new code, there is means to disable this behaviour runtime (without having to reboot the system), when scheduler debugging is enabled (CONFIG_SCHED_DEBUG=y), by issuing the following command: # echo NO_OWNER_SPIN > /debug/sched_features This command re-enables spinning again (this is also the default): # echo OWNER_SPIN > /debug/sched_features Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-01-12 20:01:47 +07:00
mutex: Make more scalable by doing less atomic operations In the __mutex_lock_common() function, an initial entry into the lock slow path will cause two atomic_xchg instructions to be issued. Together with the atomic decrement in the fast path, a total of three atomic read-modify-write instructions will be issued in rapid succession. This can cause a lot of cache bouncing when many tasks are trying to acquire the mutex at the same time. This patch will reduce the number of atomic_xchg instructions used by checking the counter value first before issuing the instruction. The atomic_read() function is just a simple memory read. The atomic_xchg() function, on the other hand, can be up to 2 order of magnitude or even more in cost when compared with atomic_read(). By using atomic_read() to check the value first before calling atomic_xchg(), we can avoid a lot of unnecessary cache coherency traffic. The only downside with this change is that a task on the slow path will have a tiny bit less chance of getting the mutex when competing with another task in the fast path. The same is true for the atomic_cmpxchg() function in the mutex-spin-on-owner loop. So an atomic_read() is also performed before calling atomic_cmpxchg(). The mutex locking and unlocking code for the x86 architecture can allow any negative number to be used in the mutex count to indicate that some tasks are waiting for the mutex. I am not so sure if that is the case for the other architectures. So the default is to avoid atomic_xchg() if the count has already been set to -1. For x86, the check is modified to include all negative numbers to cover a larger case. The following table shows the jobs per minutes (JPM) scalability data on an 8-node 80-core Westmere box with a 3.7.10 kernel. The numactl command is used to restrict the running of the high_systime workloads to 1/2/4/8 nodes with hyperthreading on and off. +-----------------+-----------+------------+----------+ | Configuration | Mean JPM | Mean JPM | % Change | | | w/o patch | with patch | | +-----------------+-----------------------------------+ | | User Range 1100 - 2000 | +-----------------+-----------------------------------+ | 8 nodes, HT on | 36980 | 148590 | +301.8% | | 8 nodes, HT off | 42799 | 145011 | +238.8% | | 4 nodes, HT on | 61318 | 118445 | +51.1% | | 4 nodes, HT off | 158481 | 158592 | +0.1% | | 2 nodes, HT on | 180602 | 173967 | -3.7% | | 2 nodes, HT off | 198409 | 198073 | -0.2% | | 1 node , HT on | 149042 | 147671 | -0.9% | | 1 node , HT off | 126036 | 126533 | +0.4% | +-----------------+-----------------------------------+ | | User Range 200 - 1000 | +-----------------+-----------------------------------+ | 8 nodes, HT on | 41525 | 122349 | +194.6% | | 8 nodes, HT off | 49866 | 124032 | +148.7% | | 4 nodes, HT on | 66409 | 106984 | +61.1% | | 4 nodes, HT off | 119880 | 130508 | +8.9% | | 2 nodes, HT on | 138003 | 133948 | -2.9% | | 2 nodes, HT off | 132792 | 131997 | -0.6% | | 1 node , HT on | 116593 | 115859 | -0.6% | | 1 node , HT off | 104499 | 104597 | +0.1% | +-----------------+------------+-----------+----------+ At low user range 10-100, the JPM differences were within +/-1%. So they are not that interesting. AIM7 benchmark run has a pretty large run-to-run variance due to random nature of the subtests executed. So a difference of less than +-5% may not be really significant. This patch improves high_systime workload performance at 4 nodes and up by maintaining transaction rates without significant drop-off at high node count. The patch has practically no impact on 1 and 2 nodes system. The table below shows the percentage time (as reported by perf record -a -s -g) spent on the __mutex_lock_slowpath() function by the high_systime workload at 1500 users for 2/4/8-node configurations with hyperthreading off. +---------------+-----------------+------------------+---------+ | Configuration | %Time w/o patch | %Time with patch | %Change | +---------------+-----------------+------------------+---------+ | 8 nodes | 65.34% | 0.69% | -99% | | 4 nodes | 8.70% | 1.02% | -88% | | 2 nodes | 0.41% | 0.32% | -22% | +---------------+-----------------+------------------+---------+ It is obvious that the dramatic performance improvement at 8 nodes was due to the drastic cut in the time spent within the __mutex_lock_slowpath() function. The table below show the improvements in other AIM7 workloads (at 8 nodes, hyperthreading off). +--------------+---------------+----------------+-----------------+ | Workload | mean % change | mean % change | mean % change | | | 10-100 users | 200-1000 users | 1100-2000 users | +--------------+---------------+----------------+-----------------+ | alltests | +0.6% | +104.2% | +185.9% | | five_sec | +1.9% | +0.9% | +0.9% | | fserver | +1.4% | -7.7% | +5.1% | | new_fserver | -0.5% | +3.2% | +3.1% | | shared | +13.1% | +146.1% | +181.5% | | short | +7.4% | +5.0% | +4.2% | +--------------+---------------+----------------+-----------------+ Signed-off-by: Waiman Long <Waiman.Long@hp.com> Reviewed-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Reviewed-by: Rik van Riel <riel@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Chandramouleeswaran Aswin <aswin@hp.com> Cc: Norton: Scott J <scott.norton@hp.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: David Howells <dhowells@redhat.com> Cc: Dave Jones <davej@redhat.com> Cc: Clark Williams <williams@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1366226594-5506-3-git-send-email-Waiman.Long@hp.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-04-18 02:23:12 +07:00
if ((atomic_read(&lock->count) == 1) &&
(atomic_cmpxchg(&lock->count, 1, 0) == 1)) {
lock_acquired(&lock->dep_map, ip);
if (use_ww_ctx) {
struct ww_mutex *ww;
ww = container_of(lock, struct ww_mutex, base);
ww_mutex_set_context_fastpath(ww, ww_ctx);
}
mutex_set_owner(lock);
mcs_spin_unlock(&lock->mcs_lock, &node);
preempt_enable();
return 0;
}
mutex: implement adaptive spinning Change mutex contention behaviour such that it will sometimes busy wait on acquisition - moving its behaviour closer to that of spinlocks. This concept got ported to mainline from the -rt tree, where it was originally implemented for rtmutexes by Steven Rostedt, based on work by Gregory Haskins. Testing with Ingo's test-mutex application (http://lkml.org/lkml/2006/1/8/50) gave a 345% boost for VFS scalability on my testbox: # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 296604 # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 85870 The key criteria for the busy wait is that the lock owner has to be running on a (different) cpu. The idea is that as long as the owner is running, there is a fair chance it'll release the lock soon, and thus we'll be better off spinning instead of blocking/scheduling. Since regular mutexes (as opposed to rtmutexes) do not atomically track the owner, we add the owner in a non-atomic fashion and deal with the races in the slowpath. Furthermore, to ease the testing of the performance impact of this new code, there is means to disable this behaviour runtime (without having to reboot the system), when scheduler debugging is enabled (CONFIG_SCHED_DEBUG=y), by issuing the following command: # echo NO_OWNER_SPIN > /debug/sched_features This command re-enables spinning again (this is also the default): # echo OWNER_SPIN > /debug/sched_features Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-01-12 20:01:47 +07:00
/*
* When there's no owner, we might have preempted between the
* owner acquiring the lock and setting the owner field. If
* we're an RT task that will live-lock because we won't let
* the owner complete.
*/
if (!owner && (need_resched() || rt_task(task)))
break;
mutex: implement adaptive spinning Change mutex contention behaviour such that it will sometimes busy wait on acquisition - moving its behaviour closer to that of spinlocks. This concept got ported to mainline from the -rt tree, where it was originally implemented for rtmutexes by Steven Rostedt, based on work by Gregory Haskins. Testing with Ingo's test-mutex application (http://lkml.org/lkml/2006/1/8/50) gave a 345% boost for VFS scalability on my testbox: # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 296604 # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 85870 The key criteria for the busy wait is that the lock owner has to be running on a (different) cpu. The idea is that as long as the owner is running, there is a fair chance it'll release the lock soon, and thus we'll be better off spinning instead of blocking/scheduling. Since regular mutexes (as opposed to rtmutexes) do not atomically track the owner, we add the owner in a non-atomic fashion and deal with the races in the slowpath. Furthermore, to ease the testing of the performance impact of this new code, there is means to disable this behaviour runtime (without having to reboot the system), when scheduler debugging is enabled (CONFIG_SCHED_DEBUG=y), by issuing the following command: # echo NO_OWNER_SPIN > /debug/sched_features This command re-enables spinning again (this is also the default): # echo OWNER_SPIN > /debug/sched_features Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-01-12 20:01:47 +07:00
/*
* The cpu_relax() call is a compiler barrier which forces
* everything in this loop to be re-loaded. We don't need
* memory barriers as we'll eventually observe the right
* values at the cost of a few extra spins.
*/
arch_mutex_cpu_relax();
mutex: implement adaptive spinning Change mutex contention behaviour such that it will sometimes busy wait on acquisition - moving its behaviour closer to that of spinlocks. This concept got ported to mainline from the -rt tree, where it was originally implemented for rtmutexes by Steven Rostedt, based on work by Gregory Haskins. Testing with Ingo's test-mutex application (http://lkml.org/lkml/2006/1/8/50) gave a 345% boost for VFS scalability on my testbox: # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 296604 # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 85870 The key criteria for the busy wait is that the lock owner has to be running on a (different) cpu. The idea is that as long as the owner is running, there is a fair chance it'll release the lock soon, and thus we'll be better off spinning instead of blocking/scheduling. Since regular mutexes (as opposed to rtmutexes) do not atomically track the owner, we add the owner in a non-atomic fashion and deal with the races in the slowpath. Furthermore, to ease the testing of the performance impact of this new code, there is means to disable this behaviour runtime (without having to reboot the system), when scheduler debugging is enabled (CONFIG_SCHED_DEBUG=y), by issuing the following command: # echo NO_OWNER_SPIN > /debug/sched_features This command re-enables spinning again (this is also the default): # echo OWNER_SPIN > /debug/sched_features Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-01-12 20:01:47 +07:00
}
mcs_spin_unlock(&lock->mcs_lock, &node);
mutex: Queue mutex spinners with MCS lock to reduce cacheline contention The current mutex spinning code (with MUTEX_SPIN_ON_OWNER option turned on) allow multiple tasks to spin on a single mutex concurrently. A potential problem with the current approach is that when the mutex becomes available, all the spinning tasks will try to acquire the mutex more or less simultaneously. As a result, there will be a lot of cacheline bouncing especially on systems with a large number of CPUs. This patch tries to reduce this kind of contention by putting the mutex spinners into a queue so that only the first one in the queue will try to acquire the mutex. This will reduce contention and allow all the tasks to move forward faster. The queuing of mutex spinners is done using an MCS lock based implementation which will further reduce contention on the mutex cacheline than a similar ticket spinlock based implementation. This patch will add a new field into the mutex data structure for holding the MCS lock. This expands the mutex size by 8 bytes for 64-bit system and 4 bytes for 32-bit system. This overhead will be avoid if the MUTEX_SPIN_ON_OWNER option is turned off. The following table shows the jobs per minute (JPM) scalability data on an 8-node 80-core Westmere box with a 3.7.10 kernel. The numactl command is used to restrict the running of the fserver workloads to 1/2/4/8 nodes with hyperthreading off. +-----------------+-----------+-----------+-------------+----------+ | Configuration | Mean JPM | Mean JPM | Mean JPM | % Change | | | w/o patch | patch 1 | patches 1&2 | 1->1&2 | +-----------------+------------------------------------------------+ | | User Range 1100 - 2000 | +-----------------+------------------------------------------------+ | 8 nodes, HT off | 227972 | 227237 | 305043 | +34.2% | | 4 nodes, HT off | 393503 | 381558 | 394650 | +3.4% | | 2 nodes, HT off | 334957 | 325240 | 338853 | +4.2% | | 1 node , HT off | 198141 | 197972 | 198075 | +0.1% | +-----------------+------------------------------------------------+ | | User Range 200 - 1000 | +-----------------+------------------------------------------------+ | 8 nodes, HT off | 282325 | 312870 | 332185 | +6.2% | | 4 nodes, HT off | 390698 | 378279 | 393419 | +4.0% | | 2 nodes, HT off | 336986 | 326543 | 340260 | +4.2% | | 1 node , HT off | 197588 | 197622 | 197582 | 0.0% | +-----------------+-----------+-----------+-------------+----------+ At low user range 10-100, the JPM differences were within +/-1%. So they are not that interesting. The fserver workload uses mutex spinning extensively. With just the mutex change in the first patch, there is no noticeable change in performance. Rather, there is a slight drop in performance. This mutex spinning patch more than recovers the lost performance and show a significant increase of +30% at high user load with the full 8 nodes. Similar improvements were also seen in a 3.8 kernel. The table below shows the %time spent by different kernel functions as reported by perf when running the fserver workload at 1500 users with all 8 nodes. +-----------------------+-----------+---------+-------------+ | Function | % time | % time | % time | | | w/o patch | patch 1 | patches 1&2 | +-----------------------+-----------+---------+-------------+ | __read_lock_failed | 34.96% | 34.91% | 29.14% | | __write_lock_failed | 10.14% | 10.68% | 7.51% | | mutex_spin_on_owner | 3.62% | 3.42% | 2.33% | | mspin_lock | N/A | N/A | 9.90% | | __mutex_lock_slowpath | 1.46% | 0.81% | 0.14% | | _raw_spin_lock | 2.25% | 2.50% | 1.10% | +-----------------------+-----------+---------+-------------+ The fserver workload for an 8-node system is dominated by the contention in the read/write lock. Mutex contention also plays a role. With the first patch only, mutex contention is down (as shown by the __mutex_lock_slowpath figure) which help a little bit. We saw only a few percents improvement with that. By applying patch 2 as well, the single mutex_spin_on_owner figure is now split out into an additional mspin_lock figure. The time increases from 3.42% to 11.23%. It shows a great reduction in contention among the spinners leading to a 30% improvement. The time ratio 9.9/2.33=4.3 indicates that there are on average 4+ spinners waiting in the spin_lock loop for each spinner in the mutex_spin_on_owner loop. Contention in other locking functions also go down by quite a lot. The table below shows the performance change of both patches 1 & 2 over patch 1 alone in other AIM7 workloads (at 8 nodes, hyperthreading off). +--------------+---------------+----------------+-----------------+ | Workload | mean % change | mean % change | mean % change | | | 10-100 users | 200-1000 users | 1100-2000 users | +--------------+---------------+----------------+-----------------+ | alltests | 0.0% | -0.8% | +0.6% | | five_sec | -0.3% | +0.8% | +0.8% | | high_systime | +0.4% | +2.4% | +2.1% | | new_fserver | +0.1% | +14.1% | +34.2% | | shared | -0.5% | -0.3% | -0.4% | | short | -1.7% | -9.8% | -8.3% | +--------------+---------------+----------------+-----------------+ The short workload is the only one that shows a decline in performance probably due to the spinner locking and queuing overhead. Signed-off-by: Waiman Long <Waiman.Long@hp.com> Reviewed-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Acked-by: Rik van Riel <riel@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Chandramouleeswaran Aswin <aswin@hp.com> Cc: Norton Scott J <scott.norton@hp.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: David Howells <dhowells@redhat.com> Cc: Dave Jones <davej@redhat.com> Cc: Clark Williams <williams@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1366226594-5506-4-git-send-email-Waiman.Long@hp.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-04-18 02:23:13 +07:00
slowpath:
mutex: implement adaptive spinning Change mutex contention behaviour such that it will sometimes busy wait on acquisition - moving its behaviour closer to that of spinlocks. This concept got ported to mainline from the -rt tree, where it was originally implemented for rtmutexes by Steven Rostedt, based on work by Gregory Haskins. Testing with Ingo's test-mutex application (http://lkml.org/lkml/2006/1/8/50) gave a 345% boost for VFS scalability on my testbox: # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 296604 # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 85870 The key criteria for the busy wait is that the lock owner has to be running on a (different) cpu. The idea is that as long as the owner is running, there is a fair chance it'll release the lock soon, and thus we'll be better off spinning instead of blocking/scheduling. Since regular mutexes (as opposed to rtmutexes) do not atomically track the owner, we add the owner in a non-atomic fashion and deal with the races in the slowpath. Furthermore, to ease the testing of the performance impact of this new code, there is means to disable this behaviour runtime (without having to reboot the system), when scheduler debugging is enabled (CONFIG_SCHED_DEBUG=y), by issuing the following command: # echo NO_OWNER_SPIN > /debug/sched_features This command re-enables spinning again (this is also the default): # echo OWNER_SPIN > /debug/sched_features Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-01-12 20:01:47 +07:00
#endif
spin_lock_mutex(&lock->wait_lock, flags);
mutex: Do not unnecessarily deal with waiters Upon entering the slowpath, we immediately attempt to acquire the lock by checking if it is already unlocked. If we are lucky enough that this is the case, then we don't need to deal with any waiter related logic. Furthermore any checks for an empty wait_list are unnecessary as we already know that count is non-negative and hence no one is waiting for the lock. Move the count check and xchg calls to be done before any waiters are setup - including waiter debugging. Upon failure to acquire the lock, the xchg sets the counter to 0, instead of -1 as it was originally. This can be done here since we set it back to -1 right at the beginning of the loop so other waiters are woken up when the lock is released. When tested on a 8-socket (80 core) system against a vanilla 3.10-rc1 kernel, this patch provides some small performance benefits (+2-6%). While it could be considered in the noise level, the average percentages were stable across multiple runs and no performance regressions were seen. Two big winners, for small amounts of users (10-100), were the short and compute workloads had a +19.36% and +%15.76% in jobs per minute. Also change some break statements to 'goto slowpath', which IMO makes a little more intuitive to read. Signed-off-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Acked-by: Rik van Riel <riel@redhat.com> Acked-by: Maarten Lankhorst <maarten.lankhorst@canonical.com> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/1372450398.2106.1.camel@buesod1.americas.hpqcorp.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-06-29 03:13:18 +07:00
/* once more, can we acquire the lock? */
if (MUTEX_SHOW_NO_WAITER(lock) && (atomic_xchg(&lock->count, 0) == 1))
goto skip_wait;
debug_mutex_lock_common(lock, &waiter);
debug_mutex_add_waiter(lock, &waiter, task_thread_info(task));
/* add waiting tasks to the end of the waitqueue (FIFO): */
list_add_tail(&waiter.list, &lock->wait_list);
waiter.task = task;
lock_contended(&lock->dep_map, ip);
for (;;) {
/*
* Lets try to take the lock again - this is needed even if
* we get here for the first time (shortly after failing to
* acquire the lock), to make sure that we get a wakeup once
* it's unlocked. Later on, if we sleep, this is the
* operation that gives us the lock. We xchg it to -1, so
* that when we release the lock, we properly wake up the
* other waiters:
*/
mutex: Make more scalable by doing less atomic operations In the __mutex_lock_common() function, an initial entry into the lock slow path will cause two atomic_xchg instructions to be issued. Together with the atomic decrement in the fast path, a total of three atomic read-modify-write instructions will be issued in rapid succession. This can cause a lot of cache bouncing when many tasks are trying to acquire the mutex at the same time. This patch will reduce the number of atomic_xchg instructions used by checking the counter value first before issuing the instruction. The atomic_read() function is just a simple memory read. The atomic_xchg() function, on the other hand, can be up to 2 order of magnitude or even more in cost when compared with atomic_read(). By using atomic_read() to check the value first before calling atomic_xchg(), we can avoid a lot of unnecessary cache coherency traffic. The only downside with this change is that a task on the slow path will have a tiny bit less chance of getting the mutex when competing with another task in the fast path. The same is true for the atomic_cmpxchg() function in the mutex-spin-on-owner loop. So an atomic_read() is also performed before calling atomic_cmpxchg(). The mutex locking and unlocking code for the x86 architecture can allow any negative number to be used in the mutex count to indicate that some tasks are waiting for the mutex. I am not so sure if that is the case for the other architectures. So the default is to avoid atomic_xchg() if the count has already been set to -1. For x86, the check is modified to include all negative numbers to cover a larger case. The following table shows the jobs per minutes (JPM) scalability data on an 8-node 80-core Westmere box with a 3.7.10 kernel. The numactl command is used to restrict the running of the high_systime workloads to 1/2/4/8 nodes with hyperthreading on and off. +-----------------+-----------+------------+----------+ | Configuration | Mean JPM | Mean JPM | % Change | | | w/o patch | with patch | | +-----------------+-----------------------------------+ | | User Range 1100 - 2000 | +-----------------+-----------------------------------+ | 8 nodes, HT on | 36980 | 148590 | +301.8% | | 8 nodes, HT off | 42799 | 145011 | +238.8% | | 4 nodes, HT on | 61318 | 118445 | +51.1% | | 4 nodes, HT off | 158481 | 158592 | +0.1% | | 2 nodes, HT on | 180602 | 173967 | -3.7% | | 2 nodes, HT off | 198409 | 198073 | -0.2% | | 1 node , HT on | 149042 | 147671 | -0.9% | | 1 node , HT off | 126036 | 126533 | +0.4% | +-----------------+-----------------------------------+ | | User Range 200 - 1000 | +-----------------+-----------------------------------+ | 8 nodes, HT on | 41525 | 122349 | +194.6% | | 8 nodes, HT off | 49866 | 124032 | +148.7% | | 4 nodes, HT on | 66409 | 106984 | +61.1% | | 4 nodes, HT off | 119880 | 130508 | +8.9% | | 2 nodes, HT on | 138003 | 133948 | -2.9% | | 2 nodes, HT off | 132792 | 131997 | -0.6% | | 1 node , HT on | 116593 | 115859 | -0.6% | | 1 node , HT off | 104499 | 104597 | +0.1% | +-----------------+------------+-----------+----------+ At low user range 10-100, the JPM differences were within +/-1%. So they are not that interesting. AIM7 benchmark run has a pretty large run-to-run variance due to random nature of the subtests executed. So a difference of less than +-5% may not be really significant. This patch improves high_systime workload performance at 4 nodes and up by maintaining transaction rates without significant drop-off at high node count. The patch has practically no impact on 1 and 2 nodes system. The table below shows the percentage time (as reported by perf record -a -s -g) spent on the __mutex_lock_slowpath() function by the high_systime workload at 1500 users for 2/4/8-node configurations with hyperthreading off. +---------------+-----------------+------------------+---------+ | Configuration | %Time w/o patch | %Time with patch | %Change | +---------------+-----------------+------------------+---------+ | 8 nodes | 65.34% | 0.69% | -99% | | 4 nodes | 8.70% | 1.02% | -88% | | 2 nodes | 0.41% | 0.32% | -22% | +---------------+-----------------+------------------+---------+ It is obvious that the dramatic performance improvement at 8 nodes was due to the drastic cut in the time spent within the __mutex_lock_slowpath() function. The table below show the improvements in other AIM7 workloads (at 8 nodes, hyperthreading off). +--------------+---------------+----------------+-----------------+ | Workload | mean % change | mean % change | mean % change | | | 10-100 users | 200-1000 users | 1100-2000 users | +--------------+---------------+----------------+-----------------+ | alltests | +0.6% | +104.2% | +185.9% | | five_sec | +1.9% | +0.9% | +0.9% | | fserver | +1.4% | -7.7% | +5.1% | | new_fserver | -0.5% | +3.2% | +3.1% | | shared | +13.1% | +146.1% | +181.5% | | short | +7.4% | +5.0% | +4.2% | +--------------+---------------+----------------+-----------------+ Signed-off-by: Waiman Long <Waiman.Long@hp.com> Reviewed-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Reviewed-by: Rik van Riel <riel@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Chandramouleeswaran Aswin <aswin@hp.com> Cc: Norton: Scott J <scott.norton@hp.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: David Howells <dhowells@redhat.com> Cc: Dave Jones <davej@redhat.com> Cc: Clark Williams <williams@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1366226594-5506-3-git-send-email-Waiman.Long@hp.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-04-18 02:23:12 +07:00
if (MUTEX_SHOW_NO_WAITER(lock) &&
mutex: Do not unnecessarily deal with waiters Upon entering the slowpath, we immediately attempt to acquire the lock by checking if it is already unlocked. If we are lucky enough that this is the case, then we don't need to deal with any waiter related logic. Furthermore any checks for an empty wait_list are unnecessary as we already know that count is non-negative and hence no one is waiting for the lock. Move the count check and xchg calls to be done before any waiters are setup - including waiter debugging. Upon failure to acquire the lock, the xchg sets the counter to 0, instead of -1 as it was originally. This can be done here since we set it back to -1 right at the beginning of the loop so other waiters are woken up when the lock is released. When tested on a 8-socket (80 core) system against a vanilla 3.10-rc1 kernel, this patch provides some small performance benefits (+2-6%). While it could be considered in the noise level, the average percentages were stable across multiple runs and no performance regressions were seen. Two big winners, for small amounts of users (10-100), were the short and compute workloads had a +19.36% and +%15.76% in jobs per minute. Also change some break statements to 'goto slowpath', which IMO makes a little more intuitive to read. Signed-off-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Acked-by: Rik van Riel <riel@redhat.com> Acked-by: Maarten Lankhorst <maarten.lankhorst@canonical.com> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/1372450398.2106.1.camel@buesod1.americas.hpqcorp.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-06-29 03:13:18 +07:00
(atomic_xchg(&lock->count, -1) == 1))
break;
/*
* got a signal? (This code gets eliminated in the
* TASK_UNINTERRUPTIBLE case.)
*/
if (unlikely(signal_pending_state(state, task))) {
ret = -EINTR;
goto err;
}
if (use_ww_ctx && ww_ctx->acquired > 0) {
ret = __mutex_lock_check_stamp(lock, ww_ctx);
if (ret)
goto err;
}
__set_task_state(task, state);
/* didn't get the lock, go to sleep: */
spin_unlock_mutex(&lock->wait_lock, flags);
schedule_preempt_disabled();
spin_lock_mutex(&lock->wait_lock, flags);
}
mutex: Do not unnecessarily deal with waiters Upon entering the slowpath, we immediately attempt to acquire the lock by checking if it is already unlocked. If we are lucky enough that this is the case, then we don't need to deal with any waiter related logic. Furthermore any checks for an empty wait_list are unnecessary as we already know that count is non-negative and hence no one is waiting for the lock. Move the count check and xchg calls to be done before any waiters are setup - including waiter debugging. Upon failure to acquire the lock, the xchg sets the counter to 0, instead of -1 as it was originally. This can be done here since we set it back to -1 right at the beginning of the loop so other waiters are woken up when the lock is released. When tested on a 8-socket (80 core) system against a vanilla 3.10-rc1 kernel, this patch provides some small performance benefits (+2-6%). While it could be considered in the noise level, the average percentages were stable across multiple runs and no performance regressions were seen. Two big winners, for small amounts of users (10-100), were the short and compute workloads had a +19.36% and +%15.76% in jobs per minute. Also change some break statements to 'goto slowpath', which IMO makes a little more intuitive to read. Signed-off-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Acked-by: Rik van Riel <riel@redhat.com> Acked-by: Maarten Lankhorst <maarten.lankhorst@canonical.com> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/1372450398.2106.1.camel@buesod1.americas.hpqcorp.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-06-29 03:13:18 +07:00
mutex_remove_waiter(lock, &waiter, current_thread_info());
/* set it to 0 if there are no waiters left: */
if (likely(list_empty(&lock->wait_list)))
atomic_set(&lock->count, 0);
debug_mutex_free_waiter(&waiter);
mutex: Do not unnecessarily deal with waiters Upon entering the slowpath, we immediately attempt to acquire the lock by checking if it is already unlocked. If we are lucky enough that this is the case, then we don't need to deal with any waiter related logic. Furthermore any checks for an empty wait_list are unnecessary as we already know that count is non-negative and hence no one is waiting for the lock. Move the count check and xchg calls to be done before any waiters are setup - including waiter debugging. Upon failure to acquire the lock, the xchg sets the counter to 0, instead of -1 as it was originally. This can be done here since we set it back to -1 right at the beginning of the loop so other waiters are woken up when the lock is released. When tested on a 8-socket (80 core) system against a vanilla 3.10-rc1 kernel, this patch provides some small performance benefits (+2-6%). While it could be considered in the noise level, the average percentages were stable across multiple runs and no performance regressions were seen. Two big winners, for small amounts of users (10-100), were the short and compute workloads had a +19.36% and +%15.76% in jobs per minute. Also change some break statements to 'goto slowpath', which IMO makes a little more intuitive to read. Signed-off-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Acked-by: Rik van Riel <riel@redhat.com> Acked-by: Maarten Lankhorst <maarten.lankhorst@canonical.com> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/1372450398.2106.1.camel@buesod1.americas.hpqcorp.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-06-29 03:13:18 +07:00
skip_wait:
/* got the lock - cleanup and rejoice! */
lock_acquired(&lock->dep_map, ip);
mutex: implement adaptive spinning Change mutex contention behaviour such that it will sometimes busy wait on acquisition - moving its behaviour closer to that of spinlocks. This concept got ported to mainline from the -rt tree, where it was originally implemented for rtmutexes by Steven Rostedt, based on work by Gregory Haskins. Testing with Ingo's test-mutex application (http://lkml.org/lkml/2006/1/8/50) gave a 345% boost for VFS scalability on my testbox: # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 296604 # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 85870 The key criteria for the busy wait is that the lock owner has to be running on a (different) cpu. The idea is that as long as the owner is running, there is a fair chance it'll release the lock soon, and thus we'll be better off spinning instead of blocking/scheduling. Since regular mutexes (as opposed to rtmutexes) do not atomically track the owner, we add the owner in a non-atomic fashion and deal with the races in the slowpath. Furthermore, to ease the testing of the performance impact of this new code, there is means to disable this behaviour runtime (without having to reboot the system), when scheduler debugging is enabled (CONFIG_SCHED_DEBUG=y), by issuing the following command: # echo NO_OWNER_SPIN > /debug/sched_features This command re-enables spinning again (this is also the default): # echo OWNER_SPIN > /debug/sched_features Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-01-12 20:01:47 +07:00
mutex_set_owner(lock);
if (use_ww_ctx) {
mutex: Do not unnecessarily deal with waiters Upon entering the slowpath, we immediately attempt to acquire the lock by checking if it is already unlocked. If we are lucky enough that this is the case, then we don't need to deal with any waiter related logic. Furthermore any checks for an empty wait_list are unnecessary as we already know that count is non-negative and hence no one is waiting for the lock. Move the count check and xchg calls to be done before any waiters are setup - including waiter debugging. Upon failure to acquire the lock, the xchg sets the counter to 0, instead of -1 as it was originally. This can be done here since we set it back to -1 right at the beginning of the loop so other waiters are woken up when the lock is released. When tested on a 8-socket (80 core) system against a vanilla 3.10-rc1 kernel, this patch provides some small performance benefits (+2-6%). While it could be considered in the noise level, the average percentages were stable across multiple runs and no performance regressions were seen. Two big winners, for small amounts of users (10-100), were the short and compute workloads had a +19.36% and +%15.76% in jobs per minute. Also change some break statements to 'goto slowpath', which IMO makes a little more intuitive to read. Signed-off-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Acked-by: Rik van Riel <riel@redhat.com> Acked-by: Maarten Lankhorst <maarten.lankhorst@canonical.com> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/1372450398.2106.1.camel@buesod1.americas.hpqcorp.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-06-29 03:13:18 +07:00
struct ww_mutex *ww = container_of(lock, struct ww_mutex, base);
struct mutex_waiter *cur;
/*
* This branch gets optimized out for the common case,
* and is only important for ww_mutex_lock.
*/
ww_mutex_lock_acquired(ww, ww_ctx);
ww->ctx = ww_ctx;
/*
* Give any possible sleeping processes the chance to wake up,
* so they can recheck if they have to back off.
*/
list_for_each_entry(cur, &lock->wait_list, list) {
debug_mutex_wake_waiter(lock, cur);
wake_up_process(cur->task);
}
}
spin_unlock_mutex(&lock->wait_lock, flags);
preempt_enable();
return 0;
err:
mutex_remove_waiter(lock, &waiter, task_thread_info(task));
spin_unlock_mutex(&lock->wait_lock, flags);
debug_mutex_free_waiter(&waiter);
mutex_release(&lock->dep_map, 1, ip);
preempt_enable();
return ret;
}
#ifdef CONFIG_DEBUG_LOCK_ALLOC
void __sched
mutex_lock_nested(struct mutex *lock, unsigned int subclass)
{
might_sleep();
__mutex_lock_common(lock, TASK_UNINTERRUPTIBLE,
subclass, NULL, _RET_IP_, NULL, 0);
}
EXPORT_SYMBOL_GPL(mutex_lock_nested);
void __sched
_mutex_lock_nest_lock(struct mutex *lock, struct lockdep_map *nest)
{
might_sleep();
__mutex_lock_common(lock, TASK_UNINTERRUPTIBLE,
0, nest, _RET_IP_, NULL, 0);
}
EXPORT_SYMBOL_GPL(_mutex_lock_nest_lock);
int __sched
mutex_lock_killable_nested(struct mutex *lock, unsigned int subclass)
{
might_sleep();
return __mutex_lock_common(lock, TASK_KILLABLE,
subclass, NULL, _RET_IP_, NULL, 0);
}
EXPORT_SYMBOL_GPL(mutex_lock_killable_nested);
int __sched
mutex_lock_interruptible_nested(struct mutex *lock, unsigned int subclass)
{
might_sleep();
mutex: implement adaptive spinning Change mutex contention behaviour such that it will sometimes busy wait on acquisition - moving its behaviour closer to that of spinlocks. This concept got ported to mainline from the -rt tree, where it was originally implemented for rtmutexes by Steven Rostedt, based on work by Gregory Haskins. Testing with Ingo's test-mutex application (http://lkml.org/lkml/2006/1/8/50) gave a 345% boost for VFS scalability on my testbox: # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 296604 # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 85870 The key criteria for the busy wait is that the lock owner has to be running on a (different) cpu. The idea is that as long as the owner is running, there is a fair chance it'll release the lock soon, and thus we'll be better off spinning instead of blocking/scheduling. Since regular mutexes (as opposed to rtmutexes) do not atomically track the owner, we add the owner in a non-atomic fashion and deal with the races in the slowpath. Furthermore, to ease the testing of the performance impact of this new code, there is means to disable this behaviour runtime (without having to reboot the system), when scheduler debugging is enabled (CONFIG_SCHED_DEBUG=y), by issuing the following command: # echo NO_OWNER_SPIN > /debug/sched_features This command re-enables spinning again (this is also the default): # echo OWNER_SPIN > /debug/sched_features Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-01-12 20:01:47 +07:00
return __mutex_lock_common(lock, TASK_INTERRUPTIBLE,
subclass, NULL, _RET_IP_, NULL, 0);
}
EXPORT_SYMBOL_GPL(mutex_lock_interruptible_nested);
mutex: Add w/w mutex slowpath debugging Injects EDEADLK conditions at pseudo-random interval, with exponential backoff up to UINT_MAX (to ensure that every lock operation still completes in a reasonable time). This way we can test the wound slowpath even for ww mutex users where contention is never expected, and the ww deadlock avoidance algorithm is only needed for correctness against malicious userspace. An example would be protecting kernel modesetting properties, which thanks to single-threaded X isn't really expected to contend, ever. I've looked into using the CONFIG_FAULT_INJECTION infrastructure, but decided against it for two reasons: - EDEADLK handling is mandatory for ww mutex users and should never affect the outcome of a syscall. This is in contrast to -ENOMEM injection. So fine configurability isn't required. - The fault injection framework only allows to set a simple probability for failure. Now the probability that a ww mutex acquire stage with N locks will never complete (due to too many injected EDEADLK backoffs) is zero. But the expected number of ww_mutex_lock operations for the completely uncontended case would be O(exp(N)). The per-acuiqire ctx exponential backoff solution choosen here only results in O(log N) overhead due to injection and so O(log N * N) lock operations. This way we can fail with high probability (and so have good test coverage even for fancy backoff and lock acquisition paths) without running into patalogical cases. Note that EDEADLK will only ever be injected when we managed to acquire the lock. This prevents any behaviour changes for users which rely on the EALREADY semantics. Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: dri-devel@lists.freedesktop.org Cc: linaro-mm-sig@lists.linaro.org Cc: rostedt@goodmis.org Cc: daniel@ffwll.ch Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/20130620113117.4001.21681.stgit@patser Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-06-20 18:31:17 +07:00
static inline int
ww_mutex_deadlock_injection(struct ww_mutex *lock, struct ww_acquire_ctx *ctx)
{
#ifdef CONFIG_DEBUG_WW_MUTEX_SLOWPATH
unsigned tmp;
if (ctx->deadlock_inject_countdown-- == 0) {
tmp = ctx->deadlock_inject_interval;
if (tmp > UINT_MAX/4)
tmp = UINT_MAX;
else
tmp = tmp*2 + tmp + tmp/2;
ctx->deadlock_inject_interval = tmp;
ctx->deadlock_inject_countdown = tmp;
ctx->contending_lock = lock;
ww_mutex_unlock(lock);
return -EDEADLK;
}
#endif
return 0;
}
int __sched
__ww_mutex_lock(struct ww_mutex *lock, struct ww_acquire_ctx *ctx)
{
mutex: Add w/w mutex slowpath debugging Injects EDEADLK conditions at pseudo-random interval, with exponential backoff up to UINT_MAX (to ensure that every lock operation still completes in a reasonable time). This way we can test the wound slowpath even for ww mutex users where contention is never expected, and the ww deadlock avoidance algorithm is only needed for correctness against malicious userspace. An example would be protecting kernel modesetting properties, which thanks to single-threaded X isn't really expected to contend, ever. I've looked into using the CONFIG_FAULT_INJECTION infrastructure, but decided against it for two reasons: - EDEADLK handling is mandatory for ww mutex users and should never affect the outcome of a syscall. This is in contrast to -ENOMEM injection. So fine configurability isn't required. - The fault injection framework only allows to set a simple probability for failure. Now the probability that a ww mutex acquire stage with N locks will never complete (due to too many injected EDEADLK backoffs) is zero. But the expected number of ww_mutex_lock operations for the completely uncontended case would be O(exp(N)). The per-acuiqire ctx exponential backoff solution choosen here only results in O(log N) overhead due to injection and so O(log N * N) lock operations. This way we can fail with high probability (and so have good test coverage even for fancy backoff and lock acquisition paths) without running into patalogical cases. Note that EDEADLK will only ever be injected when we managed to acquire the lock. This prevents any behaviour changes for users which rely on the EALREADY semantics. Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: dri-devel@lists.freedesktop.org Cc: linaro-mm-sig@lists.linaro.org Cc: rostedt@goodmis.org Cc: daniel@ffwll.ch Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/20130620113117.4001.21681.stgit@patser Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-06-20 18:31:17 +07:00
int ret;
might_sleep();
mutex: Add w/w mutex slowpath debugging Injects EDEADLK conditions at pseudo-random interval, with exponential backoff up to UINT_MAX (to ensure that every lock operation still completes in a reasonable time). This way we can test the wound slowpath even for ww mutex users where contention is never expected, and the ww deadlock avoidance algorithm is only needed for correctness against malicious userspace. An example would be protecting kernel modesetting properties, which thanks to single-threaded X isn't really expected to contend, ever. I've looked into using the CONFIG_FAULT_INJECTION infrastructure, but decided against it for two reasons: - EDEADLK handling is mandatory for ww mutex users and should never affect the outcome of a syscall. This is in contrast to -ENOMEM injection. So fine configurability isn't required. - The fault injection framework only allows to set a simple probability for failure. Now the probability that a ww mutex acquire stage with N locks will never complete (due to too many injected EDEADLK backoffs) is zero. But the expected number of ww_mutex_lock operations for the completely uncontended case would be O(exp(N)). The per-acuiqire ctx exponential backoff solution choosen here only results in O(log N) overhead due to injection and so O(log N * N) lock operations. This way we can fail with high probability (and so have good test coverage even for fancy backoff and lock acquisition paths) without running into patalogical cases. Note that EDEADLK will only ever be injected when we managed to acquire the lock. This prevents any behaviour changes for users which rely on the EALREADY semantics. Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: dri-devel@lists.freedesktop.org Cc: linaro-mm-sig@lists.linaro.org Cc: rostedt@goodmis.org Cc: daniel@ffwll.ch Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/20130620113117.4001.21681.stgit@patser Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-06-20 18:31:17 +07:00
ret = __mutex_lock_common(&lock->base, TASK_UNINTERRUPTIBLE,
0, &ctx->dep_map, _RET_IP_, ctx, 1);
if (!ret && ctx->acquired > 1)
mutex: Add w/w mutex slowpath debugging Injects EDEADLK conditions at pseudo-random interval, with exponential backoff up to UINT_MAX (to ensure that every lock operation still completes in a reasonable time). This way we can test the wound slowpath even for ww mutex users where contention is never expected, and the ww deadlock avoidance algorithm is only needed for correctness against malicious userspace. An example would be protecting kernel modesetting properties, which thanks to single-threaded X isn't really expected to contend, ever. I've looked into using the CONFIG_FAULT_INJECTION infrastructure, but decided against it for two reasons: - EDEADLK handling is mandatory for ww mutex users and should never affect the outcome of a syscall. This is in contrast to -ENOMEM injection. So fine configurability isn't required. - The fault injection framework only allows to set a simple probability for failure. Now the probability that a ww mutex acquire stage with N locks will never complete (due to too many injected EDEADLK backoffs) is zero. But the expected number of ww_mutex_lock operations for the completely uncontended case would be O(exp(N)). The per-acuiqire ctx exponential backoff solution choosen here only results in O(log N) overhead due to injection and so O(log N * N) lock operations. This way we can fail with high probability (and so have good test coverage even for fancy backoff and lock acquisition paths) without running into patalogical cases. Note that EDEADLK will only ever be injected when we managed to acquire the lock. This prevents any behaviour changes for users which rely on the EALREADY semantics. Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: dri-devel@lists.freedesktop.org Cc: linaro-mm-sig@lists.linaro.org Cc: rostedt@goodmis.org Cc: daniel@ffwll.ch Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/20130620113117.4001.21681.stgit@patser Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-06-20 18:31:17 +07:00
return ww_mutex_deadlock_injection(lock, ctx);
return ret;
}
EXPORT_SYMBOL_GPL(__ww_mutex_lock);
int __sched
__ww_mutex_lock_interruptible(struct ww_mutex *lock, struct ww_acquire_ctx *ctx)
{
mutex: Add w/w mutex slowpath debugging Injects EDEADLK conditions at pseudo-random interval, with exponential backoff up to UINT_MAX (to ensure that every lock operation still completes in a reasonable time). This way we can test the wound slowpath even for ww mutex users where contention is never expected, and the ww deadlock avoidance algorithm is only needed for correctness against malicious userspace. An example would be protecting kernel modesetting properties, which thanks to single-threaded X isn't really expected to contend, ever. I've looked into using the CONFIG_FAULT_INJECTION infrastructure, but decided against it for two reasons: - EDEADLK handling is mandatory for ww mutex users and should never affect the outcome of a syscall. This is in contrast to -ENOMEM injection. So fine configurability isn't required. - The fault injection framework only allows to set a simple probability for failure. Now the probability that a ww mutex acquire stage with N locks will never complete (due to too many injected EDEADLK backoffs) is zero. But the expected number of ww_mutex_lock operations for the completely uncontended case would be O(exp(N)). The per-acuiqire ctx exponential backoff solution choosen here only results in O(log N) overhead due to injection and so O(log N * N) lock operations. This way we can fail with high probability (and so have good test coverage even for fancy backoff and lock acquisition paths) without running into patalogical cases. Note that EDEADLK will only ever be injected when we managed to acquire the lock. This prevents any behaviour changes for users which rely on the EALREADY semantics. Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: dri-devel@lists.freedesktop.org Cc: linaro-mm-sig@lists.linaro.org Cc: rostedt@goodmis.org Cc: daniel@ffwll.ch Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/20130620113117.4001.21681.stgit@patser Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-06-20 18:31:17 +07:00
int ret;
might_sleep();
mutex: Add w/w mutex slowpath debugging Injects EDEADLK conditions at pseudo-random interval, with exponential backoff up to UINT_MAX (to ensure that every lock operation still completes in a reasonable time). This way we can test the wound slowpath even for ww mutex users where contention is never expected, and the ww deadlock avoidance algorithm is only needed for correctness against malicious userspace. An example would be protecting kernel modesetting properties, which thanks to single-threaded X isn't really expected to contend, ever. I've looked into using the CONFIG_FAULT_INJECTION infrastructure, but decided against it for two reasons: - EDEADLK handling is mandatory for ww mutex users and should never affect the outcome of a syscall. This is in contrast to -ENOMEM injection. So fine configurability isn't required. - The fault injection framework only allows to set a simple probability for failure. Now the probability that a ww mutex acquire stage with N locks will never complete (due to too many injected EDEADLK backoffs) is zero. But the expected number of ww_mutex_lock operations for the completely uncontended case would be O(exp(N)). The per-acuiqire ctx exponential backoff solution choosen here only results in O(log N) overhead due to injection and so O(log N * N) lock operations. This way we can fail with high probability (and so have good test coverage even for fancy backoff and lock acquisition paths) without running into patalogical cases. Note that EDEADLK will only ever be injected when we managed to acquire the lock. This prevents any behaviour changes for users which rely on the EALREADY semantics. Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: dri-devel@lists.freedesktop.org Cc: linaro-mm-sig@lists.linaro.org Cc: rostedt@goodmis.org Cc: daniel@ffwll.ch Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/20130620113117.4001.21681.stgit@patser Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-06-20 18:31:17 +07:00
ret = __mutex_lock_common(&lock->base, TASK_INTERRUPTIBLE,
0, &ctx->dep_map, _RET_IP_, ctx, 1);
mutex: Add w/w mutex slowpath debugging Injects EDEADLK conditions at pseudo-random interval, with exponential backoff up to UINT_MAX (to ensure that every lock operation still completes in a reasonable time). This way we can test the wound slowpath even for ww mutex users where contention is never expected, and the ww deadlock avoidance algorithm is only needed for correctness against malicious userspace. An example would be protecting kernel modesetting properties, which thanks to single-threaded X isn't really expected to contend, ever. I've looked into using the CONFIG_FAULT_INJECTION infrastructure, but decided against it for two reasons: - EDEADLK handling is mandatory for ww mutex users and should never affect the outcome of a syscall. This is in contrast to -ENOMEM injection. So fine configurability isn't required. - The fault injection framework only allows to set a simple probability for failure. Now the probability that a ww mutex acquire stage with N locks will never complete (due to too many injected EDEADLK backoffs) is zero. But the expected number of ww_mutex_lock operations for the completely uncontended case would be O(exp(N)). The per-acuiqire ctx exponential backoff solution choosen here only results in O(log N) overhead due to injection and so O(log N * N) lock operations. This way we can fail with high probability (and so have good test coverage even for fancy backoff and lock acquisition paths) without running into patalogical cases. Note that EDEADLK will only ever be injected when we managed to acquire the lock. This prevents any behaviour changes for users which rely on the EALREADY semantics. Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: dri-devel@lists.freedesktop.org Cc: linaro-mm-sig@lists.linaro.org Cc: rostedt@goodmis.org Cc: daniel@ffwll.ch Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/20130620113117.4001.21681.stgit@patser Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-06-20 18:31:17 +07:00
if (!ret && ctx->acquired > 1)
mutex: Add w/w mutex slowpath debugging Injects EDEADLK conditions at pseudo-random interval, with exponential backoff up to UINT_MAX (to ensure that every lock operation still completes in a reasonable time). This way we can test the wound slowpath even for ww mutex users where contention is never expected, and the ww deadlock avoidance algorithm is only needed for correctness against malicious userspace. An example would be protecting kernel modesetting properties, which thanks to single-threaded X isn't really expected to contend, ever. I've looked into using the CONFIG_FAULT_INJECTION infrastructure, but decided against it for two reasons: - EDEADLK handling is mandatory for ww mutex users and should never affect the outcome of a syscall. This is in contrast to -ENOMEM injection. So fine configurability isn't required. - The fault injection framework only allows to set a simple probability for failure. Now the probability that a ww mutex acquire stage with N locks will never complete (due to too many injected EDEADLK backoffs) is zero. But the expected number of ww_mutex_lock operations for the completely uncontended case would be O(exp(N)). The per-acuiqire ctx exponential backoff solution choosen here only results in O(log N) overhead due to injection and so O(log N * N) lock operations. This way we can fail with high probability (and so have good test coverage even for fancy backoff and lock acquisition paths) without running into patalogical cases. Note that EDEADLK will only ever be injected when we managed to acquire the lock. This prevents any behaviour changes for users which rely on the EALREADY semantics. Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: dri-devel@lists.freedesktop.org Cc: linaro-mm-sig@lists.linaro.org Cc: rostedt@goodmis.org Cc: daniel@ffwll.ch Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/20130620113117.4001.21681.stgit@patser Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-06-20 18:31:17 +07:00
return ww_mutex_deadlock_injection(lock, ctx);
return ret;
}
EXPORT_SYMBOL_GPL(__ww_mutex_lock_interruptible);
#endif
/*
* Release the lock, slowpath:
*/
static inline void
__mutex_unlock_common_slowpath(atomic_t *lock_count, int nested)
{
struct mutex *lock = container_of(lock_count, struct mutex, count);
unsigned long flags;
spin_lock_mutex(&lock->wait_lock, flags);
mutex_release(&lock->dep_map, nested, _RET_IP_);
debug_mutex_unlock(lock);
/*
* some architectures leave the lock unlocked in the fastpath failure
* case, others need to leave it locked. In the later case we have to
* unlock it here
*/
if (__mutex_slowpath_needs_to_unlock())
atomic_set(&lock->count, 1);
if (!list_empty(&lock->wait_list)) {
/* get the first entry from the wait-list: */
struct mutex_waiter *waiter =
list_entry(lock->wait_list.next,
struct mutex_waiter, list);
debug_mutex_wake_waiter(lock, waiter);
wake_up_process(waiter->task);
}
spin_unlock_mutex(&lock->wait_lock, flags);
}
/*
* Release the lock, slowpath:
*/
static __used noinline void
__mutex_unlock_slowpath(atomic_t *lock_count)
{
__mutex_unlock_common_slowpath(lock_count, 1);
}
#ifndef CONFIG_DEBUG_LOCK_ALLOC
/*
* Here come the less common (and hence less performance-critical) APIs:
* mutex_lock_interruptible() and mutex_trylock().
*/
static noinline int __sched
__mutex_lock_killable_slowpath(struct mutex *lock);
static noinline int __sched
__mutex_lock_interruptible_slowpath(struct mutex *lock);
/**
* mutex_lock_interruptible - acquire the mutex, interruptible
* @lock: the mutex to be acquired
*
* Lock the mutex like mutex_lock(), and return 0 if the mutex has
* been acquired or sleep until the mutex becomes available. If a
* signal arrives while waiting for the lock then this function
* returns -EINTR.
*
* This function is similar to (but not equivalent to) down_interruptible().
*/
int __sched mutex_lock_interruptible(struct mutex *lock)
{
mutex: implement adaptive spinning Change mutex contention behaviour such that it will sometimes busy wait on acquisition - moving its behaviour closer to that of spinlocks. This concept got ported to mainline from the -rt tree, where it was originally implemented for rtmutexes by Steven Rostedt, based on work by Gregory Haskins. Testing with Ingo's test-mutex application (http://lkml.org/lkml/2006/1/8/50) gave a 345% boost for VFS scalability on my testbox: # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 296604 # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 85870 The key criteria for the busy wait is that the lock owner has to be running on a (different) cpu. The idea is that as long as the owner is running, there is a fair chance it'll release the lock soon, and thus we'll be better off spinning instead of blocking/scheduling. Since regular mutexes (as opposed to rtmutexes) do not atomically track the owner, we add the owner in a non-atomic fashion and deal with the races in the slowpath. Furthermore, to ease the testing of the performance impact of this new code, there is means to disable this behaviour runtime (without having to reboot the system), when scheduler debugging is enabled (CONFIG_SCHED_DEBUG=y), by issuing the following command: # echo NO_OWNER_SPIN > /debug/sched_features This command re-enables spinning again (this is also the default): # echo OWNER_SPIN > /debug/sched_features Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-01-12 20:01:47 +07:00
int ret;
might_sleep();
ret = __mutex_fastpath_lock_retval(&lock->count);
if (likely(!ret)) {
mutex: implement adaptive spinning Change mutex contention behaviour such that it will sometimes busy wait on acquisition - moving its behaviour closer to that of spinlocks. This concept got ported to mainline from the -rt tree, where it was originally implemented for rtmutexes by Steven Rostedt, based on work by Gregory Haskins. Testing with Ingo's test-mutex application (http://lkml.org/lkml/2006/1/8/50) gave a 345% boost for VFS scalability on my testbox: # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 296604 # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 85870 The key criteria for the busy wait is that the lock owner has to be running on a (different) cpu. The idea is that as long as the owner is running, there is a fair chance it'll release the lock soon, and thus we'll be better off spinning instead of blocking/scheduling. Since regular mutexes (as opposed to rtmutexes) do not atomically track the owner, we add the owner in a non-atomic fashion and deal with the races in the slowpath. Furthermore, to ease the testing of the performance impact of this new code, there is means to disable this behaviour runtime (without having to reboot the system), when scheduler debugging is enabled (CONFIG_SCHED_DEBUG=y), by issuing the following command: # echo NO_OWNER_SPIN > /debug/sched_features This command re-enables spinning again (this is also the default): # echo OWNER_SPIN > /debug/sched_features Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-01-12 20:01:47 +07:00
mutex_set_owner(lock);
return 0;
} else
return __mutex_lock_interruptible_slowpath(lock);
}
EXPORT_SYMBOL(mutex_lock_interruptible);
int __sched mutex_lock_killable(struct mutex *lock)
{
mutex: implement adaptive spinning Change mutex contention behaviour such that it will sometimes busy wait on acquisition - moving its behaviour closer to that of spinlocks. This concept got ported to mainline from the -rt tree, where it was originally implemented for rtmutexes by Steven Rostedt, based on work by Gregory Haskins. Testing with Ingo's test-mutex application (http://lkml.org/lkml/2006/1/8/50) gave a 345% boost for VFS scalability on my testbox: # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 296604 # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 85870 The key criteria for the busy wait is that the lock owner has to be running on a (different) cpu. The idea is that as long as the owner is running, there is a fair chance it'll release the lock soon, and thus we'll be better off spinning instead of blocking/scheduling. Since regular mutexes (as opposed to rtmutexes) do not atomically track the owner, we add the owner in a non-atomic fashion and deal with the races in the slowpath. Furthermore, to ease the testing of the performance impact of this new code, there is means to disable this behaviour runtime (without having to reboot the system), when scheduler debugging is enabled (CONFIG_SCHED_DEBUG=y), by issuing the following command: # echo NO_OWNER_SPIN > /debug/sched_features This command re-enables spinning again (this is also the default): # echo OWNER_SPIN > /debug/sched_features Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-01-12 20:01:47 +07:00
int ret;
might_sleep();
ret = __mutex_fastpath_lock_retval(&lock->count);
if (likely(!ret)) {
mutex: implement adaptive spinning Change mutex contention behaviour such that it will sometimes busy wait on acquisition - moving its behaviour closer to that of spinlocks. This concept got ported to mainline from the -rt tree, where it was originally implemented for rtmutexes by Steven Rostedt, based on work by Gregory Haskins. Testing with Ingo's test-mutex application (http://lkml.org/lkml/2006/1/8/50) gave a 345% boost for VFS scalability on my testbox: # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 296604 # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 85870 The key criteria for the busy wait is that the lock owner has to be running on a (different) cpu. The idea is that as long as the owner is running, there is a fair chance it'll release the lock soon, and thus we'll be better off spinning instead of blocking/scheduling. Since regular mutexes (as opposed to rtmutexes) do not atomically track the owner, we add the owner in a non-atomic fashion and deal with the races in the slowpath. Furthermore, to ease the testing of the performance impact of this new code, there is means to disable this behaviour runtime (without having to reboot the system), when scheduler debugging is enabled (CONFIG_SCHED_DEBUG=y), by issuing the following command: # echo NO_OWNER_SPIN > /debug/sched_features This command re-enables spinning again (this is also the default): # echo OWNER_SPIN > /debug/sched_features Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-01-12 20:01:47 +07:00
mutex_set_owner(lock);
return 0;
} else
return __mutex_lock_killable_slowpath(lock);
}
EXPORT_SYMBOL(mutex_lock_killable);
static __used noinline void __sched
__mutex_lock_slowpath(atomic_t *lock_count)
{
struct mutex *lock = container_of(lock_count, struct mutex, count);
__mutex_lock_common(lock, TASK_UNINTERRUPTIBLE, 0,
NULL, _RET_IP_, NULL, 0);
}
static noinline int __sched
__mutex_lock_killable_slowpath(struct mutex *lock)
{
return __mutex_lock_common(lock, TASK_KILLABLE, 0,
NULL, _RET_IP_, NULL, 0);
}
static noinline int __sched
__mutex_lock_interruptible_slowpath(struct mutex *lock)
{
return __mutex_lock_common(lock, TASK_INTERRUPTIBLE, 0,
NULL, _RET_IP_, NULL, 0);
}
static noinline int __sched
__ww_mutex_lock_slowpath(struct ww_mutex *lock, struct ww_acquire_ctx *ctx)
{
return __mutex_lock_common(&lock->base, TASK_UNINTERRUPTIBLE, 0,
NULL, _RET_IP_, ctx, 1);
}
static noinline int __sched
__ww_mutex_lock_interruptible_slowpath(struct ww_mutex *lock,
struct ww_acquire_ctx *ctx)
{
return __mutex_lock_common(&lock->base, TASK_INTERRUPTIBLE, 0,
NULL, _RET_IP_, ctx, 1);
}
#endif
/*
* Spinlock based trylock, we take the spinlock and check whether we
* can get the lock:
*/
static inline int __mutex_trylock_slowpath(atomic_t *lock_count)
{
struct mutex *lock = container_of(lock_count, struct mutex, count);
unsigned long flags;
int prev;
spin_lock_mutex(&lock->wait_lock, flags);
prev = atomic_xchg(&lock->count, -1);
if (likely(prev == 1)) {
mutex: implement adaptive spinning Change mutex contention behaviour such that it will sometimes busy wait on acquisition - moving its behaviour closer to that of spinlocks. This concept got ported to mainline from the -rt tree, where it was originally implemented for rtmutexes by Steven Rostedt, based on work by Gregory Haskins. Testing with Ingo's test-mutex application (http://lkml.org/lkml/2006/1/8/50) gave a 345% boost for VFS scalability on my testbox: # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 296604 # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 85870 The key criteria for the busy wait is that the lock owner has to be running on a (different) cpu. The idea is that as long as the owner is running, there is a fair chance it'll release the lock soon, and thus we'll be better off spinning instead of blocking/scheduling. Since regular mutexes (as opposed to rtmutexes) do not atomically track the owner, we add the owner in a non-atomic fashion and deal with the races in the slowpath. Furthermore, to ease the testing of the performance impact of this new code, there is means to disable this behaviour runtime (without having to reboot the system), when scheduler debugging is enabled (CONFIG_SCHED_DEBUG=y), by issuing the following command: # echo NO_OWNER_SPIN > /debug/sched_features This command re-enables spinning again (this is also the default): # echo OWNER_SPIN > /debug/sched_features Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-01-12 20:01:47 +07:00
mutex_set_owner(lock);
mutex_acquire(&lock->dep_map, 0, 1, _RET_IP_);
}
mutex: implement adaptive spinning Change mutex contention behaviour such that it will sometimes busy wait on acquisition - moving its behaviour closer to that of spinlocks. This concept got ported to mainline from the -rt tree, where it was originally implemented for rtmutexes by Steven Rostedt, based on work by Gregory Haskins. Testing with Ingo's test-mutex application (http://lkml.org/lkml/2006/1/8/50) gave a 345% boost for VFS scalability on my testbox: # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 296604 # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 85870 The key criteria for the busy wait is that the lock owner has to be running on a (different) cpu. The idea is that as long as the owner is running, there is a fair chance it'll release the lock soon, and thus we'll be better off spinning instead of blocking/scheduling. Since regular mutexes (as opposed to rtmutexes) do not atomically track the owner, we add the owner in a non-atomic fashion and deal with the races in the slowpath. Furthermore, to ease the testing of the performance impact of this new code, there is means to disable this behaviour runtime (without having to reboot the system), when scheduler debugging is enabled (CONFIG_SCHED_DEBUG=y), by issuing the following command: # echo NO_OWNER_SPIN > /debug/sched_features This command re-enables spinning again (this is also the default): # echo OWNER_SPIN > /debug/sched_features Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-01-12 20:01:47 +07:00
/* Set it back to 0 if there are no waiters: */
if (likely(list_empty(&lock->wait_list)))
atomic_set(&lock->count, 0);
spin_unlock_mutex(&lock->wait_lock, flags);
return prev == 1;
}
/**
* mutex_trylock - try to acquire the mutex, without waiting
* @lock: the mutex to be acquired
*
* Try to acquire the mutex atomically. Returns 1 if the mutex
* has been acquired successfully, and 0 on contention.
*
* NOTE: this function follows the spin_trylock() convention, so
* it is negated from the down_trylock() return values! Be careful
* about this when converting semaphore users to mutexes.
*
* This function must not be used in interrupt context. The
* mutex must be released by the same task that acquired it.
*/
int __sched mutex_trylock(struct mutex *lock)
{
mutex: implement adaptive spinning Change mutex contention behaviour such that it will sometimes busy wait on acquisition - moving its behaviour closer to that of spinlocks. This concept got ported to mainline from the -rt tree, where it was originally implemented for rtmutexes by Steven Rostedt, based on work by Gregory Haskins. Testing with Ingo's test-mutex application (http://lkml.org/lkml/2006/1/8/50) gave a 345% boost for VFS scalability on my testbox: # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 296604 # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 85870 The key criteria for the busy wait is that the lock owner has to be running on a (different) cpu. The idea is that as long as the owner is running, there is a fair chance it'll release the lock soon, and thus we'll be better off spinning instead of blocking/scheduling. Since regular mutexes (as opposed to rtmutexes) do not atomically track the owner, we add the owner in a non-atomic fashion and deal with the races in the slowpath. Furthermore, to ease the testing of the performance impact of this new code, there is means to disable this behaviour runtime (without having to reboot the system), when scheduler debugging is enabled (CONFIG_SCHED_DEBUG=y), by issuing the following command: # echo NO_OWNER_SPIN > /debug/sched_features This command re-enables spinning again (this is also the default): # echo OWNER_SPIN > /debug/sched_features Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-01-12 20:01:47 +07:00
int ret;
ret = __mutex_fastpath_trylock(&lock->count, __mutex_trylock_slowpath);
if (ret)
mutex_set_owner(lock);
return ret;
}
EXPORT_SYMBOL(mutex_trylock);
#ifndef CONFIG_DEBUG_LOCK_ALLOC
int __sched
__ww_mutex_lock(struct ww_mutex *lock, struct ww_acquire_ctx *ctx)
{
int ret;
might_sleep();
ret = __mutex_fastpath_lock_retval(&lock->base.count);
if (likely(!ret)) {
ww_mutex_set_context_fastpath(lock, ctx);
mutex_set_owner(&lock->base);
} else
ret = __ww_mutex_lock_slowpath(lock, ctx);
return ret;
}
EXPORT_SYMBOL(__ww_mutex_lock);
int __sched
__ww_mutex_lock_interruptible(struct ww_mutex *lock, struct ww_acquire_ctx *ctx)
{
int ret;
might_sleep();
ret = __mutex_fastpath_lock_retval(&lock->base.count);
if (likely(!ret)) {
ww_mutex_set_context_fastpath(lock, ctx);
mutex_set_owner(&lock->base);
} else
ret = __ww_mutex_lock_interruptible_slowpath(lock, ctx);
return ret;
}
EXPORT_SYMBOL(__ww_mutex_lock_interruptible);
#endif
/**
* atomic_dec_and_mutex_lock - return holding mutex if we dec to 0
* @cnt: the atomic which we are to dec
* @lock: the mutex to return holding if we dec to 0
*
* return true and hold lock if we dec to 0, return false otherwise
*/
int atomic_dec_and_mutex_lock(atomic_t *cnt, struct mutex *lock)
{
/* dec if we can't possibly hit 0 */
if (atomic_add_unless(cnt, -1, 1))
return 0;
/* we might hit 0, so take the lock */
mutex_lock(lock);
if (!atomic_dec_and_test(cnt)) {
/* when we actually did the dec, we didn't hit 0 */
mutex_unlock(lock);
return 0;
}
/* we hit 0, and we hold the lock */
return 1;
}
EXPORT_SYMBOL(atomic_dec_and_mutex_lock);