linux/kernel/locking/rwsem.c

1655 lines
44 KiB
C
Raw Normal View History

License cleanup: add SPDX GPL-2.0 license identifier to files with no license Many source files in the tree are missing licensing information, which makes it harder for compliance tools to determine the correct license. By default all files without license information are under the default license of the kernel, which is GPL version 2. Update the files which contain no license information with the 'GPL-2.0' SPDX license identifier. The SPDX identifier is a legally binding shorthand, which can be used instead of the full boiler plate text. This patch is based on work done by Thomas Gleixner and Kate Stewart and Philippe Ombredanne. How this work was done: Patches were generated and checked against linux-4.14-rc6 for a subset of the use cases: - file had no licensing information it it. - file was a */uapi/* one with no licensing information in it, - file was a */uapi/* one with existing licensing information, Further patches will be generated in subsequent months to fix up cases where non-standard license headers were used, and references to license had to be inferred by heuristics based on keywords. The analysis to determine which SPDX License Identifier to be applied to a file was done in a spreadsheet of side by side results from of the output of two independent scanners (ScanCode & Windriver) producing SPDX tag:value files created by Philippe Ombredanne. Philippe prepared the base worksheet, and did an initial spot review of a few 1000 files. The 4.13 kernel was the starting point of the analysis with 60,537 files assessed. Kate Stewart did a file by file comparison of the scanner results in the spreadsheet to determine which SPDX license identifier(s) to be applied to the file. She confirmed any determination that was not immediately clear with lawyers working with the Linux Foundation. Criteria used to select files for SPDX license identifier tagging was: - Files considered eligible had to be source code files. - Make and config files were included as candidates if they contained >5 lines of source - File already had some variant of a license header in it (even if <5 lines). All documentation files were explicitly excluded. The following heuristics were used to determine which SPDX license identifiers to apply. - when both scanners couldn't find any license traces, file was considered to have no license information in it, and the top level COPYING file license applied. For non */uapi/* files that summary was: SPDX license identifier # files ---------------------------------------------------|------- GPL-2.0 11139 and resulted in the first patch in this series. If that file was a */uapi/* path one, it was "GPL-2.0 WITH Linux-syscall-note" otherwise it was "GPL-2.0". Results of that was: SPDX license identifier # files ---------------------------------------------------|------- GPL-2.0 WITH Linux-syscall-note 930 and resulted in the second patch in this series. - if a file had some form of licensing information in it, and was one of the */uapi/* ones, it was denoted with the Linux-syscall-note if any GPL family license was found in the file or had no licensing in it (per prior point). Results summary: SPDX license identifier # files ---------------------------------------------------|------ GPL-2.0 WITH Linux-syscall-note 270 GPL-2.0+ WITH Linux-syscall-note 169 ((GPL-2.0 WITH Linux-syscall-note) OR BSD-2-Clause) 21 ((GPL-2.0 WITH Linux-syscall-note) OR BSD-3-Clause) 17 LGPL-2.1+ WITH Linux-syscall-note 15 GPL-1.0+ WITH Linux-syscall-note 14 ((GPL-2.0+ WITH Linux-syscall-note) OR BSD-3-Clause) 5 LGPL-2.0+ WITH Linux-syscall-note 4 LGPL-2.1 WITH Linux-syscall-note 3 ((GPL-2.0 WITH Linux-syscall-note) OR MIT) 3 ((GPL-2.0 WITH Linux-syscall-note) AND MIT) 1 and that resulted in the third patch in this series. - when the two scanners agreed on the detected license(s), that became the concluded license(s). - when there was disagreement between the two scanners (one detected a license but the other didn't, or they both detected different licenses) a manual inspection of the file occurred. - In most cases a manual inspection of the information in the file resulted in a clear resolution of the license that should apply (and which scanner probably needed to revisit its heuristics). - When it was not immediately clear, the license identifier was confirmed with lawyers working with the Linux Foundation. - If there was any question as to the appropriate license identifier, the file was flagged for further research and to be revisited later in time. In total, over 70 hours of logged manual review was done on the spreadsheet to determine the SPDX license identifiers to apply to the source files by Kate, Philippe, Thomas and, in some cases, confirmation by lawyers working with the Linux Foundation. Kate also obtained a third independent scan of the 4.13 code base from FOSSology, and compared selected files where the other two scanners disagreed against that SPDX file, to see if there was new insights. The Windriver scanner is based on an older version of FOSSology in part, so they are related. Thomas did random spot checks in about 500 files from the spreadsheets for the uapi headers and agreed with SPDX license identifier in the files he inspected. For the non-uapi files Thomas did random spot checks in about 15000 files. In initial set of patches against 4.14-rc6, 3 files were found to have copy/paste license identifier errors, and have been fixed to reflect the correct identifier. Additionally Philippe spent 10 hours this week doing a detailed manual inspection and review of the 12,461 patched files from the initial patch version early this week with: - a full scancode scan run, collecting the matched texts, detected license ids and scores - reviewing anything where there was a license detected (about 500+ files) to ensure that the applied SPDX license was correct - reviewing anything where there was no detection but the patch license was not GPL-2.0 WITH Linux-syscall-note to ensure that the applied SPDX license was correct This produced a worksheet with 20 files needing minor correction. This worksheet was then exported into 3 different .csv files for the different types of files to be modified. These .csv files were then reviewed by Greg. Thomas wrote a script to parse the csv files and add the proper SPDX tag to the file, in the format that the file expected. This script was further refined by Greg based on the output to detect more types of files automatically and to distinguish between header and source .c files (which need different comment types.) Finally Greg ran the script using the .csv files to generate the patches. Reviewed-by: Kate Stewart <kstewart@linuxfoundation.org> Reviewed-by: Philippe Ombredanne <pombredanne@nexb.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2017-11-01 17:07:57 +03:00
// SPDX-License-Identifier: GPL-2.0
/* kernel/rwsem.c: R/W semaphores, public implementation
*
* Written by David Howells (dhowells@redhat.com).
* Derived from asm-i386/semaphore.h
*
* Writer lock-stealing by Alex Shi <alex.shi@intel.com>
* and Michel Lespinasse <walken@google.com>
*
* Optimistic spinning by Tim Chen <tim.c.chen@intel.com>
* and Davidlohr Bueso <davidlohr@hp.com>. Based on mutexes.
*
locking/rwsem: Implement lock handoff to prevent lock starvation Because of writer lock stealing, it is possible that a constant stream of incoming writers will cause a waiting writer or reader to wait indefinitely leading to lock starvation. This patch implements a lock handoff mechanism to disable lock stealing and force lock handoff to the first waiter or waiters (for readers) in the queue after at least a 4ms waiting period unless it is a RT writer task which doesn't need to wait. The waiting period is used to avoid discouraging lock stealing too much to affect performance. The setting and clearing of the handoff bit is serialized by the wait_lock. So racing is not possible. A rwsem microbenchmark was run for 5 seconds on a 2-socket 40-core 80-thread Skylake system with a v5.1 based kernel and 240 write_lock threads with 5us sleep critical section. Before the patch, the min/mean/max numbers of locking operations for the locking threads were 1/7,792/173,696. After the patch, the figures became 5,842/6,542/7,458. It can be seen that the rwsem became much more fair, though there was a drop of about 16% in the mean locking operations done which was a tradeoff of having better fairness. Making the waiter set the handoff bit right after the first wakeup can impact performance especially with a mixed reader/writer workload. With the same microbenchmark with short critical section and equal number of reader and writer threads (40/40), the reader/writer locking operation counts with the current patch were: 40 readers, Iterations Min/Mean/Max = 1,793/1,794/1,796 40 writers, Iterations Min/Mean/Max = 1,793/34,956/86,081 By making waiter set handoff bit immediately after wakeup: 40 readers, Iterations Min/Mean/Max = 43/44/46 40 writers, Iterations Min/Mean/Max = 43/1,263/3,191 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-8-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:06 +03:00
* Rwsem count bit fields re-definition and rwsem rearchitecture by
* Waiman Long <longman@redhat.com> and
* Peter Zijlstra <peterz@infradead.org>.
*/
#include <linux/types.h>
#include <linux/kernel.h>
#include <linux/sched.h>
#include <linux/sched/rt.h>
#include <linux/sched/task.h>
#include <linux/sched/debug.h>
#include <linux/sched/wake_q.h>
#include <linux/sched/signal.h>
locking/rwsem: Enable time-based spinning on reader-owned rwsem When the rwsem is owned by reader, writers stop optimistic spinning simply because there is no easy way to figure out if all the readers are actively running or not. However, there are scenarios where the readers are unlikely to sleep and optimistic spinning can help performance. This patch provides a simple mechanism for spinning on a reader-owned rwsem by a writer. It is a time threshold based spinning where the allowable spinning time can vary from 10us to 25us depending on the condition of the rwsem. When the time threshold is exceeded, the nonspinnable bits will be set in the owner field to indicate that no more optimistic spinning will be allowed on this rwsem until it becomes writer owned again. Not even readers is allowed to acquire the reader-locked rwsem by optimistic spinning for fairness. We also want a writer to acquire the lock after the readers hold the lock for a relatively long time. In order to give preference to writers under such a circumstance, the single RWSEM_NONSPINNABLE bit is now split into two - one for reader and one for writer. When optimistic spinning is disabled, both bits will be set. When the reader count drop down to 0, the writer nonspinnable bit will be cleared to allow writers to spin on the lock, but not the readers. When a writer acquires the lock, it will write its own task structure pointer into sem->owner and clear the reader nonspinnable bit in the process. The time taken for each iteration of the reader-owned rwsem spinning loop varies. Below are sample minimum elapsed times for 16 iterations of the loop. System Time for 16 Iterations ------ ---------------------- 1-socket Skylake ~800ns 4-socket Broadwell ~300ns 2-socket ThunderX2 (arm64) ~250ns When the lock cacheline is contended, we can see up to almost 10X increase in elapsed time. So 25us will be at most 500, 1300 and 1600 iterations for each of the above systems. With a locking microbenchmark running on 5.1 based kernel, the total locking rates (in kops/s) on a 8-socket IvyBridge-EX system with equal numbers of readers and writers before and after this patch were as follows: # of Threads Pre-patch Post-patch ------------ --------- ---------- 2 1,759 6,684 4 1,684 6,738 8 1,074 7,222 16 900 7,163 32 458 7,316 64 208 520 128 168 425 240 143 474 This patch gives a big boost in performance for mixed reader/writer workloads. With 32 locking threads, the rwsem lock event data were: rwsem_opt_fail=79850 rwsem_opt_nospin=5069 rwsem_opt_rlock=597484 rwsem_opt_wlock=957339 rwsem_sleep_reader=57782 rwsem_sleep_writer=55663 With 64 locking threads, the data looked like: rwsem_opt_fail=346723 rwsem_opt_nospin=6293 rwsem_opt_rlock=1127119 rwsem_opt_wlock=1400628 rwsem_sleep_reader=308201 rwsem_sleep_writer=72281 So a lot more threads acquired the lock in the slowpath and more threads went to sleep. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-15-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:13 +03:00
#include <linux/sched/clock.h>
#include <linux/export.h>
#include <linux/rwsem.h>
#include <linux/atomic.h>
locking/rwsem: Add rtmutex based R/W semaphore implementation The RT specific R/W semaphore implementation used to restrict the number of readers to one, because a writer cannot block on multiple readers and inherit its priority or budget. The single reader restricting was painful in various ways: - Performance bottleneck for multi-threaded applications in the page fault path (mmap sem) - Progress blocker for drivers which are carefully crafted to avoid the potential reader/writer deadlock in mainline. The analysis of the writer code paths shows that properly written RT tasks should not take them. Syscalls like mmap(), file access which take mmap sem write locked have unbound latencies, which are completely unrelated to mmap sem. Other R/W sem users like graphics drivers are not suitable for RT tasks either. So there is little risk to hurt RT tasks when the RT rwsem implementation is done in the following way: - Allow concurrent readers - Make writers block until the last reader left the critical section. This blocking is not subject to priority/budget inheritance. - Readers blocked on a writer inherit their priority/budget in the normal way. There is a drawback with this scheme: R/W semaphores become writer unfair though the applications which have triggered writer starvation (mostly on mmap_sem) in the past are not really the typical workloads running on a RT system. So while it's unlikely to hit writer starvation, it's possible. If there are unexpected workloads on RT systems triggering it, the problem has to be revisited. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Link: https://lore.kernel.org/r/20210815211303.016885947@linutronix.de
2021-08-16 00:28:05 +03:00
#ifndef CONFIG_PREEMPT_RT
#include "lock_events.h"
/*
locking/rwsem: Remove reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It also improves the chance that a reader can get the lock as writer optimistic spinning disproportionally favors writers much more than readers. Since commit d3681e269fff ("locking/rwsem: Wake up almost all readers in wait queue"), all the waiting readers are woken up so that they can all get the read lock and run in parallel. When the number of contending readers is large, allowing reader optimistic spinning will likely cause reader fragmentation where multiple smaller groups of readers can get the read lock in a sequential manner separated by writers. That reduces reader parallelism. One possible way to address that drawback is to limit the number of readers (preferably one) that can do optimistic spinning. These readers act as representatives of all the waiting readers in the wait queue as they will wake up all those waiting readers once they get the lock. Alternatively, as reader optimistic lock stealing has already enhanced fairness to readers, it may be easier to just remove reader optimistic spinning and simplifying the optimistic spinning code as a result. Performance measurements (locking throughput kops/s) using a locking microbenchmark with 50/50 reader/writer distribution and turbo-boost disabled was done on a 2-socket Cascade Lake system (48-core 96-thread) to see the impacts of these changes: 1) Vanilla - 5.10-rc3 kernel 2) Before - 5.10-rc3 kernel with previous patches in this series 2) limit-rspin - 5.10-rc3 kernel with limited reader spinning patch 3) no-rspin - 5.10-rc3 kernel with reader spinning disabled # of threads CS Load Vanilla Before limit-rspin no-rspin ------------ ------- ------- ------ ----------- -------- 2 1 5,185 5,662 5,214 5,077 4 1 5,107 4,983 5,188 4,760 8 1 4,782 4,564 4,720 4,628 16 1 4,680 4,053 4,567 3,402 32 1 4,299 1,115 1,118 1,098 64 1 3,218 983 1,001 957 96 1 1,938 944 957 930 2 20 2,008 2,128 2,264 1,665 4 20 1,390 1,033 1,046 1,101 8 20 1,472 1,155 1,098 1,213 16 20 1,332 1,077 1,089 1,122 32 20 967 914 917 980 64 20 787 874 891 858 96 20 730 836 847 844 2 100 372 356 360 355 4 100 492 425 434 392 8 100 533 537 529 538 16 100 548 572 568 598 32 100 499 520 527 537 64 100 466 517 526 512 96 100 406 497 506 509 The column "CS Load" represents the number of pause instructions issued in the locking critical section. A CS load of 1 is extremely short and is not likey in real situations. A load of 20 (moderate) and 100 (long) are more realistic. It can be seen that the previous patches in this series have reduced performance in general except in highly contended cases with moderate or long critical sections that performance improves a bit. This change is mostly caused by the "Prevent potential lock starvation" patch that reduce reader optimistic spinning and hence reduce reader fragmentation. The patch that further limit reader optimistic spinning doesn't seem to have too much impact on overall performance as shown in the benchmark data. The patch that disables reader optimistic spinning shows reduced performance at lightly loaded cases, but comparable or slightly better performance on with heavier contention. This patch just removes reader optimistic spinning for now. As readers are not going to do optimistic spinning anymore, we don't need to consider if the OSQ is empty or not when doing lock stealing. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-6-longman@redhat.com
2020-11-21 07:14:16 +03:00
* The least significant 2 bits of the owner value has the following
* meanings when set.
* - Bit 0: RWSEM_READER_OWNED - The rwsem is owned by readers
locking/rwsem: Remove reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It also improves the chance that a reader can get the lock as writer optimistic spinning disproportionally favors writers much more than readers. Since commit d3681e269fff ("locking/rwsem: Wake up almost all readers in wait queue"), all the waiting readers are woken up so that they can all get the read lock and run in parallel. When the number of contending readers is large, allowing reader optimistic spinning will likely cause reader fragmentation where multiple smaller groups of readers can get the read lock in a sequential manner separated by writers. That reduces reader parallelism. One possible way to address that drawback is to limit the number of readers (preferably one) that can do optimistic spinning. These readers act as representatives of all the waiting readers in the wait queue as they will wake up all those waiting readers once they get the lock. Alternatively, as reader optimistic lock stealing has already enhanced fairness to readers, it may be easier to just remove reader optimistic spinning and simplifying the optimistic spinning code as a result. Performance measurements (locking throughput kops/s) using a locking microbenchmark with 50/50 reader/writer distribution and turbo-boost disabled was done on a 2-socket Cascade Lake system (48-core 96-thread) to see the impacts of these changes: 1) Vanilla - 5.10-rc3 kernel 2) Before - 5.10-rc3 kernel with previous patches in this series 2) limit-rspin - 5.10-rc3 kernel with limited reader spinning patch 3) no-rspin - 5.10-rc3 kernel with reader spinning disabled # of threads CS Load Vanilla Before limit-rspin no-rspin ------------ ------- ------- ------ ----------- -------- 2 1 5,185 5,662 5,214 5,077 4 1 5,107 4,983 5,188 4,760 8 1 4,782 4,564 4,720 4,628 16 1 4,680 4,053 4,567 3,402 32 1 4,299 1,115 1,118 1,098 64 1 3,218 983 1,001 957 96 1 1,938 944 957 930 2 20 2,008 2,128 2,264 1,665 4 20 1,390 1,033 1,046 1,101 8 20 1,472 1,155 1,098 1,213 16 20 1,332 1,077 1,089 1,122 32 20 967 914 917 980 64 20 787 874 891 858 96 20 730 836 847 844 2 100 372 356 360 355 4 100 492 425 434 392 8 100 533 537 529 538 16 100 548 572 568 598 32 100 499 520 527 537 64 100 466 517 526 512 96 100 406 497 506 509 The column "CS Load" represents the number of pause instructions issued in the locking critical section. A CS load of 1 is extremely short and is not likey in real situations. A load of 20 (moderate) and 100 (long) are more realistic. It can be seen that the previous patches in this series have reduced performance in general except in highly contended cases with moderate or long critical sections that performance improves a bit. This change is mostly caused by the "Prevent potential lock starvation" patch that reduce reader optimistic spinning and hence reduce reader fragmentation. The patch that further limit reader optimistic spinning doesn't seem to have too much impact on overall performance as shown in the benchmark data. The patch that disables reader optimistic spinning shows reduced performance at lightly loaded cases, but comparable or slightly better performance on with heavier contention. This patch just removes reader optimistic spinning for now. As readers are not going to do optimistic spinning anymore, we don't need to consider if the OSQ is empty or not when doing lock stealing. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-6-longman@redhat.com
2020-11-21 07:14:16 +03:00
* - Bit 1: RWSEM_NONSPINNABLE - Cannot spin on a reader-owned lock
*
locking/rwsem: Remove reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It also improves the chance that a reader can get the lock as writer optimistic spinning disproportionally favors writers much more than readers. Since commit d3681e269fff ("locking/rwsem: Wake up almost all readers in wait queue"), all the waiting readers are woken up so that they can all get the read lock and run in parallel. When the number of contending readers is large, allowing reader optimistic spinning will likely cause reader fragmentation where multiple smaller groups of readers can get the read lock in a sequential manner separated by writers. That reduces reader parallelism. One possible way to address that drawback is to limit the number of readers (preferably one) that can do optimistic spinning. These readers act as representatives of all the waiting readers in the wait queue as they will wake up all those waiting readers once they get the lock. Alternatively, as reader optimistic lock stealing has already enhanced fairness to readers, it may be easier to just remove reader optimistic spinning and simplifying the optimistic spinning code as a result. Performance measurements (locking throughput kops/s) using a locking microbenchmark with 50/50 reader/writer distribution and turbo-boost disabled was done on a 2-socket Cascade Lake system (48-core 96-thread) to see the impacts of these changes: 1) Vanilla - 5.10-rc3 kernel 2) Before - 5.10-rc3 kernel with previous patches in this series 2) limit-rspin - 5.10-rc3 kernel with limited reader spinning patch 3) no-rspin - 5.10-rc3 kernel with reader spinning disabled # of threads CS Load Vanilla Before limit-rspin no-rspin ------------ ------- ------- ------ ----------- -------- 2 1 5,185 5,662 5,214 5,077 4 1 5,107 4,983 5,188 4,760 8 1 4,782 4,564 4,720 4,628 16 1 4,680 4,053 4,567 3,402 32 1 4,299 1,115 1,118 1,098 64 1 3,218 983 1,001 957 96 1 1,938 944 957 930 2 20 2,008 2,128 2,264 1,665 4 20 1,390 1,033 1,046 1,101 8 20 1,472 1,155 1,098 1,213 16 20 1,332 1,077 1,089 1,122 32 20 967 914 917 980 64 20 787 874 891 858 96 20 730 836 847 844 2 100 372 356 360 355 4 100 492 425 434 392 8 100 533 537 529 538 16 100 548 572 568 598 32 100 499 520 527 537 64 100 466 517 526 512 96 100 406 497 506 509 The column "CS Load" represents the number of pause instructions issued in the locking critical section. A CS load of 1 is extremely short and is not likey in real situations. A load of 20 (moderate) and 100 (long) are more realistic. It can be seen that the previous patches in this series have reduced performance in general except in highly contended cases with moderate or long critical sections that performance improves a bit. This change is mostly caused by the "Prevent potential lock starvation" patch that reduce reader optimistic spinning and hence reduce reader fragmentation. The patch that further limit reader optimistic spinning doesn't seem to have too much impact on overall performance as shown in the benchmark data. The patch that disables reader optimistic spinning shows reduced performance at lightly loaded cases, but comparable or slightly better performance on with heavier contention. This patch just removes reader optimistic spinning for now. As readers are not going to do optimistic spinning anymore, we don't need to consider if the OSQ is empty or not when doing lock stealing. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-6-longman@redhat.com
2020-11-21 07:14:16 +03:00
* When the rwsem is reader-owned and a spinning writer has timed out,
* the nonspinnable bit will be set to disable optimistic spinning.
locking/rwsem: Enable time-based spinning on reader-owned rwsem When the rwsem is owned by reader, writers stop optimistic spinning simply because there is no easy way to figure out if all the readers are actively running or not. However, there are scenarios where the readers are unlikely to sleep and optimistic spinning can help performance. This patch provides a simple mechanism for spinning on a reader-owned rwsem by a writer. It is a time threshold based spinning where the allowable spinning time can vary from 10us to 25us depending on the condition of the rwsem. When the time threshold is exceeded, the nonspinnable bits will be set in the owner field to indicate that no more optimistic spinning will be allowed on this rwsem until it becomes writer owned again. Not even readers is allowed to acquire the reader-locked rwsem by optimistic spinning for fairness. We also want a writer to acquire the lock after the readers hold the lock for a relatively long time. In order to give preference to writers under such a circumstance, the single RWSEM_NONSPINNABLE bit is now split into two - one for reader and one for writer. When optimistic spinning is disabled, both bits will be set. When the reader count drop down to 0, the writer nonspinnable bit will be cleared to allow writers to spin on the lock, but not the readers. When a writer acquires the lock, it will write its own task structure pointer into sem->owner and clear the reader nonspinnable bit in the process. The time taken for each iteration of the reader-owned rwsem spinning loop varies. Below are sample minimum elapsed times for 16 iterations of the loop. System Time for 16 Iterations ------ ---------------------- 1-socket Skylake ~800ns 4-socket Broadwell ~300ns 2-socket ThunderX2 (arm64) ~250ns When the lock cacheline is contended, we can see up to almost 10X increase in elapsed time. So 25us will be at most 500, 1300 and 1600 iterations for each of the above systems. With a locking microbenchmark running on 5.1 based kernel, the total locking rates (in kops/s) on a 8-socket IvyBridge-EX system with equal numbers of readers and writers before and after this patch were as follows: # of Threads Pre-patch Post-patch ------------ --------- ---------- 2 1,759 6,684 4 1,684 6,738 8 1,074 7,222 16 900 7,163 32 458 7,316 64 208 520 128 168 425 240 143 474 This patch gives a big boost in performance for mixed reader/writer workloads. With 32 locking threads, the rwsem lock event data were: rwsem_opt_fail=79850 rwsem_opt_nospin=5069 rwsem_opt_rlock=597484 rwsem_opt_wlock=957339 rwsem_sleep_reader=57782 rwsem_sleep_writer=55663 With 64 locking threads, the data looked like: rwsem_opt_fail=346723 rwsem_opt_nospin=6293 rwsem_opt_rlock=1127119 rwsem_opt_wlock=1400628 rwsem_sleep_reader=308201 rwsem_sleep_writer=72281 So a lot more threads acquired the lock in the slowpath and more threads went to sleep. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-15-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:13 +03:00
* When a writer acquires a rwsem, it puts its task_struct pointer
* into the owner field. It is cleared after an unlock.
*
* When a reader acquires a rwsem, it will also puts its task_struct
locking/rwsem: Enable time-based spinning on reader-owned rwsem When the rwsem is owned by reader, writers stop optimistic spinning simply because there is no easy way to figure out if all the readers are actively running or not. However, there are scenarios where the readers are unlikely to sleep and optimistic spinning can help performance. This patch provides a simple mechanism for spinning on a reader-owned rwsem by a writer. It is a time threshold based spinning where the allowable spinning time can vary from 10us to 25us depending on the condition of the rwsem. When the time threshold is exceeded, the nonspinnable bits will be set in the owner field to indicate that no more optimistic spinning will be allowed on this rwsem until it becomes writer owned again. Not even readers is allowed to acquire the reader-locked rwsem by optimistic spinning for fairness. We also want a writer to acquire the lock after the readers hold the lock for a relatively long time. In order to give preference to writers under such a circumstance, the single RWSEM_NONSPINNABLE bit is now split into two - one for reader and one for writer. When optimistic spinning is disabled, both bits will be set. When the reader count drop down to 0, the writer nonspinnable bit will be cleared to allow writers to spin on the lock, but not the readers. When a writer acquires the lock, it will write its own task structure pointer into sem->owner and clear the reader nonspinnable bit in the process. The time taken for each iteration of the reader-owned rwsem spinning loop varies. Below are sample minimum elapsed times for 16 iterations of the loop. System Time for 16 Iterations ------ ---------------------- 1-socket Skylake ~800ns 4-socket Broadwell ~300ns 2-socket ThunderX2 (arm64) ~250ns When the lock cacheline is contended, we can see up to almost 10X increase in elapsed time. So 25us will be at most 500, 1300 and 1600 iterations for each of the above systems. With a locking microbenchmark running on 5.1 based kernel, the total locking rates (in kops/s) on a 8-socket IvyBridge-EX system with equal numbers of readers and writers before and after this patch were as follows: # of Threads Pre-patch Post-patch ------------ --------- ---------- 2 1,759 6,684 4 1,684 6,738 8 1,074 7,222 16 900 7,163 32 458 7,316 64 208 520 128 168 425 240 143 474 This patch gives a big boost in performance for mixed reader/writer workloads. With 32 locking threads, the rwsem lock event data were: rwsem_opt_fail=79850 rwsem_opt_nospin=5069 rwsem_opt_rlock=597484 rwsem_opt_wlock=957339 rwsem_sleep_reader=57782 rwsem_sleep_writer=55663 With 64 locking threads, the data looked like: rwsem_opt_fail=346723 rwsem_opt_nospin=6293 rwsem_opt_rlock=1127119 rwsem_opt_wlock=1400628 rwsem_sleep_reader=308201 rwsem_sleep_writer=72281 So a lot more threads acquired the lock in the slowpath and more threads went to sleep. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-15-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:13 +03:00
* pointer into the owner field with the RWSEM_READER_OWNED bit set.
* On unlock, the owner field will largely be left untouched. So
* for a free or reader-owned rwsem, the owner value may contain
* information about the last reader that acquires the rwsem.
*
* That information may be helpful in debugging cases where the system
* seems to hang on a reader owned rwsem especially if only one reader
* is involved. Ideally we would like to track all the readers that own
* a rwsem, but the overhead is simply too big.
locking/rwsem: Adaptive disabling of reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It makes readers relatively more preferred than writers. When a writer times out spinning on a reader-owned lock and set the nospinnable bits, there are two main reasons for that. 1) The reader critical section is long, perhaps the task sleeps after acquiring the read lock. 2) There are just too many readers contending the lock causing it to take a while to service all of them. In the former case, long reader critical section will impede the progress of writers which is usually more important for system performance. In the later case, reader optimistic spinning tends to make the reader groups that contain readers that acquire the lock together smaller leading to more of them. That may hurt performance in some cases. In other words, the setting of nonspinnable bits indicates that reader optimistic spinning may not be helpful for those workloads that cause it. Therefore, any writers that have observed the setting of the writer nonspinnable bit for a given rwsem after they fail to acquire the lock via optimistic spinning will set the reader nonspinnable bit once they acquire the write lock. Similarly, readers that observe the setting of reader nonspinnable bit at slowpath entry will also set the reader nonspinnable bit when they acquire the read lock via the wakeup path. Once the reader nonspinnable bit is on, it will only be reset when a writer is able to acquire the rwsem in the fast path or somehow a reader or writer in the slowpath doesn't observe the nonspinable bit. This is to discourage reader optmistic spinning on that particular rwsem and make writers more preferred. This adaptive disabling of reader optimistic spinning will alleviate some of the negative side effect of this feature. In addition, this patch tries to make readers in the spinning queue follow the phase-fair principle after quitting optimistic spinning by checking if another reader has somehow acquired a read lock after this reader enters the optimistic spinning queue. If so and the rwsem is still reader-owned, this reader is in the right read-phase and can attempt to acquire the lock. On a 2-socket 40-core 80-thread Skylake system, the page_fault1 test of the will-it-scale benchmark was run with various number of threads. The number of operations done before reader optimistic spinning patches, this patch and after this patch were: Threads Before rspin Before patch After patch %change ------- ------------ ------------ ----------- ------- 20 5541068 5345484 5455667 -3.5%/ +2.1% 40 10185150 7292313 9219276 -28.5%/+26.4% 60 8196733 6460517 7181209 -21.2%/+11.2% 80 9508864 6739559 8107025 -29.1%/+20.3% This patch doesn't recover all the lost performance, but it is more than half. Given the fact that reader optimistic spinning does benefit some workloads, this is a good compromise. Using the rwsem locking microbenchmark with very short critical section, this patch doesn't have too much impact on locking performance as shown by the locking rates (kops/s) below with equal numbers of readers and writers before and after this patch: # of Threads Pre-patch Post-patch ------------ --------- ---------- 2 4,730 4,969 4 4,814 4,786 8 4,866 4,815 16 4,715 4,511 32 3,338 3,500 64 3,212 3,389 80 3,110 3,044 When running the locking microbenchmark with 40 dedicated reader and writer threads, however, the reader performance is curtailed to favor the writer. Before patch: 40 readers, Iterations Min/Mean/Max = 204,026/234,309/254,816 40 writers, Iterations Min/Mean/Max = 88,515/95,884/115,644 After patch: 40 readers, Iterations Min/Mean/Max = 33,813/35,260/36,791 40 writers, Iterations Min/Mean/Max = 95,368/96,565/97,798 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-16-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:14 +03:00
*
locking/rwsem: Remove reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It also improves the chance that a reader can get the lock as writer optimistic spinning disproportionally favors writers much more than readers. Since commit d3681e269fff ("locking/rwsem: Wake up almost all readers in wait queue"), all the waiting readers are woken up so that they can all get the read lock and run in parallel. When the number of contending readers is large, allowing reader optimistic spinning will likely cause reader fragmentation where multiple smaller groups of readers can get the read lock in a sequential manner separated by writers. That reduces reader parallelism. One possible way to address that drawback is to limit the number of readers (preferably one) that can do optimistic spinning. These readers act as representatives of all the waiting readers in the wait queue as they will wake up all those waiting readers once they get the lock. Alternatively, as reader optimistic lock stealing has already enhanced fairness to readers, it may be easier to just remove reader optimistic spinning and simplifying the optimistic spinning code as a result. Performance measurements (locking throughput kops/s) using a locking microbenchmark with 50/50 reader/writer distribution and turbo-boost disabled was done on a 2-socket Cascade Lake system (48-core 96-thread) to see the impacts of these changes: 1) Vanilla - 5.10-rc3 kernel 2) Before - 5.10-rc3 kernel with previous patches in this series 2) limit-rspin - 5.10-rc3 kernel with limited reader spinning patch 3) no-rspin - 5.10-rc3 kernel with reader spinning disabled # of threads CS Load Vanilla Before limit-rspin no-rspin ------------ ------- ------- ------ ----------- -------- 2 1 5,185 5,662 5,214 5,077 4 1 5,107 4,983 5,188 4,760 8 1 4,782 4,564 4,720 4,628 16 1 4,680 4,053 4,567 3,402 32 1 4,299 1,115 1,118 1,098 64 1 3,218 983 1,001 957 96 1 1,938 944 957 930 2 20 2,008 2,128 2,264 1,665 4 20 1,390 1,033 1,046 1,101 8 20 1,472 1,155 1,098 1,213 16 20 1,332 1,077 1,089 1,122 32 20 967 914 917 980 64 20 787 874 891 858 96 20 730 836 847 844 2 100 372 356 360 355 4 100 492 425 434 392 8 100 533 537 529 538 16 100 548 572 568 598 32 100 499 520 527 537 64 100 466 517 526 512 96 100 406 497 506 509 The column "CS Load" represents the number of pause instructions issued in the locking critical section. A CS load of 1 is extremely short and is not likey in real situations. A load of 20 (moderate) and 100 (long) are more realistic. It can be seen that the previous patches in this series have reduced performance in general except in highly contended cases with moderate or long critical sections that performance improves a bit. This change is mostly caused by the "Prevent potential lock starvation" patch that reduce reader optimistic spinning and hence reduce reader fragmentation. The patch that further limit reader optimistic spinning doesn't seem to have too much impact on overall performance as shown in the benchmark data. The patch that disables reader optimistic spinning shows reduced performance at lightly loaded cases, but comparable or slightly better performance on with heavier contention. This patch just removes reader optimistic spinning for now. As readers are not going to do optimistic spinning anymore, we don't need to consider if the OSQ is empty or not when doing lock stealing. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-6-longman@redhat.com
2020-11-21 07:14:16 +03:00
* A fast path reader optimistic lock stealing is supported when the rwsem
* is previously owned by a writer and the following conditions are met:
* - OSQ is empty
* - rwsem is not currently writer owned
* - the handoff isn't set.
*/
#define RWSEM_READER_OWNED (1UL << 0)
locking/rwsem: Remove reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It also improves the chance that a reader can get the lock as writer optimistic spinning disproportionally favors writers much more than readers. Since commit d3681e269fff ("locking/rwsem: Wake up almost all readers in wait queue"), all the waiting readers are woken up so that they can all get the read lock and run in parallel. When the number of contending readers is large, allowing reader optimistic spinning will likely cause reader fragmentation where multiple smaller groups of readers can get the read lock in a sequential manner separated by writers. That reduces reader parallelism. One possible way to address that drawback is to limit the number of readers (preferably one) that can do optimistic spinning. These readers act as representatives of all the waiting readers in the wait queue as they will wake up all those waiting readers once they get the lock. Alternatively, as reader optimistic lock stealing has already enhanced fairness to readers, it may be easier to just remove reader optimistic spinning and simplifying the optimistic spinning code as a result. Performance measurements (locking throughput kops/s) using a locking microbenchmark with 50/50 reader/writer distribution and turbo-boost disabled was done on a 2-socket Cascade Lake system (48-core 96-thread) to see the impacts of these changes: 1) Vanilla - 5.10-rc3 kernel 2) Before - 5.10-rc3 kernel with previous patches in this series 2) limit-rspin - 5.10-rc3 kernel with limited reader spinning patch 3) no-rspin - 5.10-rc3 kernel with reader spinning disabled # of threads CS Load Vanilla Before limit-rspin no-rspin ------------ ------- ------- ------ ----------- -------- 2 1 5,185 5,662 5,214 5,077 4 1 5,107 4,983 5,188 4,760 8 1 4,782 4,564 4,720 4,628 16 1 4,680 4,053 4,567 3,402 32 1 4,299 1,115 1,118 1,098 64 1 3,218 983 1,001 957 96 1 1,938 944 957 930 2 20 2,008 2,128 2,264 1,665 4 20 1,390 1,033 1,046 1,101 8 20 1,472 1,155 1,098 1,213 16 20 1,332 1,077 1,089 1,122 32 20 967 914 917 980 64 20 787 874 891 858 96 20 730 836 847 844 2 100 372 356 360 355 4 100 492 425 434 392 8 100 533 537 529 538 16 100 548 572 568 598 32 100 499 520 527 537 64 100 466 517 526 512 96 100 406 497 506 509 The column "CS Load" represents the number of pause instructions issued in the locking critical section. A CS load of 1 is extremely short and is not likey in real situations. A load of 20 (moderate) and 100 (long) are more realistic. It can be seen that the previous patches in this series have reduced performance in general except in highly contended cases with moderate or long critical sections that performance improves a bit. This change is mostly caused by the "Prevent potential lock starvation" patch that reduce reader optimistic spinning and hence reduce reader fragmentation. The patch that further limit reader optimistic spinning doesn't seem to have too much impact on overall performance as shown in the benchmark data. The patch that disables reader optimistic spinning shows reduced performance at lightly loaded cases, but comparable or slightly better performance on with heavier contention. This patch just removes reader optimistic spinning for now. As readers are not going to do optimistic spinning anymore, we don't need to consider if the OSQ is empty or not when doing lock stealing. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-6-longman@redhat.com
2020-11-21 07:14:16 +03:00
#define RWSEM_NONSPINNABLE (1UL << 1)
#define RWSEM_OWNER_FLAGS_MASK (RWSEM_READER_OWNED | RWSEM_NONSPINNABLE)
#ifdef CONFIG_DEBUG_RWSEMS
# define DEBUG_RWSEMS_WARN_ON(c, sem) do { \
if (!debug_locks_silent && \
WARN_ONCE(c, "DEBUG_RWSEMS_WARN_ON(%s): count = 0x%lx, magic = 0x%lx, owner = 0x%lx, curr 0x%lx, list %sempty\n",\
#c, atomic_long_read(&(sem)->count), \
(unsigned long) sem->magic, \
atomic_long_read(&(sem)->owner), (long)current, \
list_empty(&(sem)->wait_list) ? "" : "not ")) \
debug_locks_off(); \
} while (0)
#else
# define DEBUG_RWSEMS_WARN_ON(c, sem)
#endif
/*
* On 64-bit architectures, the bit definitions of the count are:
*
* Bit 0 - writer locked bit
* Bit 1 - waiters present bit
* Bit 2 - lock handoff bit
* Bits 3-7 - reserved
* Bits 8-62 - 55-bit reader count
* Bit 63 - read fail bit
*
* On 32-bit architectures, the bit definitions of the count are:
*
* Bit 0 - writer locked bit
* Bit 1 - waiters present bit
* Bit 2 - lock handoff bit
* Bits 3-7 - reserved
* Bits 8-30 - 23-bit reader count
* Bit 31 - read fail bit
*
* It is not likely that the most significant bit (read fail bit) will ever
* be set. This guard bit is still checked anyway in the down_read() fastpath
* just in case we need to use up more of the reader bits for other purpose
* in the future.
*
* atomic_long_fetch_add() is used to obtain reader lock, whereas
* atomic_long_cmpxchg() will be used to obtain writer lock.
locking/rwsem: Implement lock handoff to prevent lock starvation Because of writer lock stealing, it is possible that a constant stream of incoming writers will cause a waiting writer or reader to wait indefinitely leading to lock starvation. This patch implements a lock handoff mechanism to disable lock stealing and force lock handoff to the first waiter or waiters (for readers) in the queue after at least a 4ms waiting period unless it is a RT writer task which doesn't need to wait. The waiting period is used to avoid discouraging lock stealing too much to affect performance. The setting and clearing of the handoff bit is serialized by the wait_lock. So racing is not possible. A rwsem microbenchmark was run for 5 seconds on a 2-socket 40-core 80-thread Skylake system with a v5.1 based kernel and 240 write_lock threads with 5us sleep critical section. Before the patch, the min/mean/max numbers of locking operations for the locking threads were 1/7,792/173,696. After the patch, the figures became 5,842/6,542/7,458. It can be seen that the rwsem became much more fair, though there was a drop of about 16% in the mean locking operations done which was a tradeoff of having better fairness. Making the waiter set the handoff bit right after the first wakeup can impact performance especially with a mixed reader/writer workload. With the same microbenchmark with short critical section and equal number of reader and writer threads (40/40), the reader/writer locking operation counts with the current patch were: 40 readers, Iterations Min/Mean/Max = 1,793/1,794/1,796 40 writers, Iterations Min/Mean/Max = 1,793/34,956/86,081 By making waiter set handoff bit immediately after wakeup: 40 readers, Iterations Min/Mean/Max = 43/44/46 40 writers, Iterations Min/Mean/Max = 43/1,263/3,191 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-8-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:06 +03:00
*
* There are three places where the lock handoff bit may be set or cleared.
* 1) rwsem_mark_wake() for readers.
* 2) rwsem_try_write_lock() for writers.
* 3) Error path of rwsem_down_write_slowpath().
*
* For all the above cases, wait_lock will be held. A writer must also
* be the first one in the wait_list to be eligible for setting the handoff
* bit. So concurrent setting/clearing of handoff bit is not possible.
*/
#define RWSEM_WRITER_LOCKED (1UL << 0)
#define RWSEM_FLAG_WAITERS (1UL << 1)
locking/rwsem: Implement lock handoff to prevent lock starvation Because of writer lock stealing, it is possible that a constant stream of incoming writers will cause a waiting writer or reader to wait indefinitely leading to lock starvation. This patch implements a lock handoff mechanism to disable lock stealing and force lock handoff to the first waiter or waiters (for readers) in the queue after at least a 4ms waiting period unless it is a RT writer task which doesn't need to wait. The waiting period is used to avoid discouraging lock stealing too much to affect performance. The setting and clearing of the handoff bit is serialized by the wait_lock. So racing is not possible. A rwsem microbenchmark was run for 5 seconds on a 2-socket 40-core 80-thread Skylake system with a v5.1 based kernel and 240 write_lock threads with 5us sleep critical section. Before the patch, the min/mean/max numbers of locking operations for the locking threads were 1/7,792/173,696. After the patch, the figures became 5,842/6,542/7,458. It can be seen that the rwsem became much more fair, though there was a drop of about 16% in the mean locking operations done which was a tradeoff of having better fairness. Making the waiter set the handoff bit right after the first wakeup can impact performance especially with a mixed reader/writer workload. With the same microbenchmark with short critical section and equal number of reader and writer threads (40/40), the reader/writer locking operation counts with the current patch were: 40 readers, Iterations Min/Mean/Max = 1,793/1,794/1,796 40 writers, Iterations Min/Mean/Max = 1,793/34,956/86,081 By making waiter set handoff bit immediately after wakeup: 40 readers, Iterations Min/Mean/Max = 43/44/46 40 writers, Iterations Min/Mean/Max = 43/1,263/3,191 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-8-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:06 +03:00
#define RWSEM_FLAG_HANDOFF (1UL << 2)
#define RWSEM_FLAG_READFAIL (1UL << (BITS_PER_LONG - 1))
locking/rwsem: Implement lock handoff to prevent lock starvation Because of writer lock stealing, it is possible that a constant stream of incoming writers will cause a waiting writer or reader to wait indefinitely leading to lock starvation. This patch implements a lock handoff mechanism to disable lock stealing and force lock handoff to the first waiter or waiters (for readers) in the queue after at least a 4ms waiting period unless it is a RT writer task which doesn't need to wait. The waiting period is used to avoid discouraging lock stealing too much to affect performance. The setting and clearing of the handoff bit is serialized by the wait_lock. So racing is not possible. A rwsem microbenchmark was run for 5 seconds on a 2-socket 40-core 80-thread Skylake system with a v5.1 based kernel and 240 write_lock threads with 5us sleep critical section. Before the patch, the min/mean/max numbers of locking operations for the locking threads were 1/7,792/173,696. After the patch, the figures became 5,842/6,542/7,458. It can be seen that the rwsem became much more fair, though there was a drop of about 16% in the mean locking operations done which was a tradeoff of having better fairness. Making the waiter set the handoff bit right after the first wakeup can impact performance especially with a mixed reader/writer workload. With the same microbenchmark with short critical section and equal number of reader and writer threads (40/40), the reader/writer locking operation counts with the current patch were: 40 readers, Iterations Min/Mean/Max = 1,793/1,794/1,796 40 writers, Iterations Min/Mean/Max = 1,793/34,956/86,081 By making waiter set handoff bit immediately after wakeup: 40 readers, Iterations Min/Mean/Max = 43/44/46 40 writers, Iterations Min/Mean/Max = 43/1,263/3,191 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-8-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:06 +03:00
#define RWSEM_READER_SHIFT 8
#define RWSEM_READER_BIAS (1UL << RWSEM_READER_SHIFT)
#define RWSEM_READER_MASK (~(RWSEM_READER_BIAS - 1))
#define RWSEM_WRITER_MASK RWSEM_WRITER_LOCKED
#define RWSEM_LOCK_MASK (RWSEM_WRITER_MASK|RWSEM_READER_MASK)
locking/rwsem: Implement lock handoff to prevent lock starvation Because of writer lock stealing, it is possible that a constant stream of incoming writers will cause a waiting writer or reader to wait indefinitely leading to lock starvation. This patch implements a lock handoff mechanism to disable lock stealing and force lock handoff to the first waiter or waiters (for readers) in the queue after at least a 4ms waiting period unless it is a RT writer task which doesn't need to wait. The waiting period is used to avoid discouraging lock stealing too much to affect performance. The setting and clearing of the handoff bit is serialized by the wait_lock. So racing is not possible. A rwsem microbenchmark was run for 5 seconds on a 2-socket 40-core 80-thread Skylake system with a v5.1 based kernel and 240 write_lock threads with 5us sleep critical section. Before the patch, the min/mean/max numbers of locking operations for the locking threads were 1/7,792/173,696. After the patch, the figures became 5,842/6,542/7,458. It can be seen that the rwsem became much more fair, though there was a drop of about 16% in the mean locking operations done which was a tradeoff of having better fairness. Making the waiter set the handoff bit right after the first wakeup can impact performance especially with a mixed reader/writer workload. With the same microbenchmark with short critical section and equal number of reader and writer threads (40/40), the reader/writer locking operation counts with the current patch were: 40 readers, Iterations Min/Mean/Max = 1,793/1,794/1,796 40 writers, Iterations Min/Mean/Max = 1,793/34,956/86,081 By making waiter set handoff bit immediately after wakeup: 40 readers, Iterations Min/Mean/Max = 43/44/46 40 writers, Iterations Min/Mean/Max = 43/1,263/3,191 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-8-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:06 +03:00
#define RWSEM_READ_FAILED_MASK (RWSEM_WRITER_MASK|RWSEM_FLAG_WAITERS|\
RWSEM_FLAG_HANDOFF|RWSEM_FLAG_READFAIL)
/*
* All writes to owner are protected by WRITE_ONCE() to make sure that
* store tearing can't happen as optimistic spinners may read and use
* the owner value concurrently without lock. Read from owner, however,
* may not need READ_ONCE() as long as the pointer value is only used
* for comparison and isn't being dereferenced.
*/
static inline void rwsem_set_owner(struct rw_semaphore *sem)
{
atomic_long_set(&sem->owner, (long)current);
}
static inline void rwsem_clear_owner(struct rw_semaphore *sem)
{
atomic_long_set(&sem->owner, 0);
}
/*
* Test the flags in the owner field.
*/
static inline bool rwsem_test_oflags(struct rw_semaphore *sem, long flags)
{
return atomic_long_read(&sem->owner) & flags;
}
/*
* The task_struct pointer of the last owning reader will be left in
* the owner field.
*
* Note that the owner value just indicates the task has owned the rwsem
* previously, it may not be the real owner or one of the real owners
* anymore when that field is examined, so take it with a grain of salt.
locking/rwsem: Adaptive disabling of reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It makes readers relatively more preferred than writers. When a writer times out spinning on a reader-owned lock and set the nospinnable bits, there are two main reasons for that. 1) The reader critical section is long, perhaps the task sleeps after acquiring the read lock. 2) There are just too many readers contending the lock causing it to take a while to service all of them. In the former case, long reader critical section will impede the progress of writers which is usually more important for system performance. In the later case, reader optimistic spinning tends to make the reader groups that contain readers that acquire the lock together smaller leading to more of them. That may hurt performance in some cases. In other words, the setting of nonspinnable bits indicates that reader optimistic spinning may not be helpful for those workloads that cause it. Therefore, any writers that have observed the setting of the writer nonspinnable bit for a given rwsem after they fail to acquire the lock via optimistic spinning will set the reader nonspinnable bit once they acquire the write lock. Similarly, readers that observe the setting of reader nonspinnable bit at slowpath entry will also set the reader nonspinnable bit when they acquire the read lock via the wakeup path. Once the reader nonspinnable bit is on, it will only be reset when a writer is able to acquire the rwsem in the fast path or somehow a reader or writer in the slowpath doesn't observe the nonspinable bit. This is to discourage reader optmistic spinning on that particular rwsem and make writers more preferred. This adaptive disabling of reader optimistic spinning will alleviate some of the negative side effect of this feature. In addition, this patch tries to make readers in the spinning queue follow the phase-fair principle after quitting optimistic spinning by checking if another reader has somehow acquired a read lock after this reader enters the optimistic spinning queue. If so and the rwsem is still reader-owned, this reader is in the right read-phase and can attempt to acquire the lock. On a 2-socket 40-core 80-thread Skylake system, the page_fault1 test of the will-it-scale benchmark was run with various number of threads. The number of operations done before reader optimistic spinning patches, this patch and after this patch were: Threads Before rspin Before patch After patch %change ------- ------------ ------------ ----------- ------- 20 5541068 5345484 5455667 -3.5%/ +2.1% 40 10185150 7292313 9219276 -28.5%/+26.4% 60 8196733 6460517 7181209 -21.2%/+11.2% 80 9508864 6739559 8107025 -29.1%/+20.3% This patch doesn't recover all the lost performance, but it is more than half. Given the fact that reader optimistic spinning does benefit some workloads, this is a good compromise. Using the rwsem locking microbenchmark with very short critical section, this patch doesn't have too much impact on locking performance as shown by the locking rates (kops/s) below with equal numbers of readers and writers before and after this patch: # of Threads Pre-patch Post-patch ------------ --------- ---------- 2 4,730 4,969 4 4,814 4,786 8 4,866 4,815 16 4,715 4,511 32 3,338 3,500 64 3,212 3,389 80 3,110 3,044 When running the locking microbenchmark with 40 dedicated reader and writer threads, however, the reader performance is curtailed to favor the writer. Before patch: 40 readers, Iterations Min/Mean/Max = 204,026/234,309/254,816 40 writers, Iterations Min/Mean/Max = 88,515/95,884/115,644 After patch: 40 readers, Iterations Min/Mean/Max = 33,813/35,260/36,791 40 writers, Iterations Min/Mean/Max = 95,368/96,565/97,798 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-16-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:14 +03:00
*
* The reader non-spinnable bit is preserved.
*/
static inline void __rwsem_set_reader_owned(struct rw_semaphore *sem,
struct task_struct *owner)
{
locking/rwsem: Adaptive disabling of reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It makes readers relatively more preferred than writers. When a writer times out spinning on a reader-owned lock and set the nospinnable bits, there are two main reasons for that. 1) The reader critical section is long, perhaps the task sleeps after acquiring the read lock. 2) There are just too many readers contending the lock causing it to take a while to service all of them. In the former case, long reader critical section will impede the progress of writers which is usually more important for system performance. In the later case, reader optimistic spinning tends to make the reader groups that contain readers that acquire the lock together smaller leading to more of them. That may hurt performance in some cases. In other words, the setting of nonspinnable bits indicates that reader optimistic spinning may not be helpful for those workloads that cause it. Therefore, any writers that have observed the setting of the writer nonspinnable bit for a given rwsem after they fail to acquire the lock via optimistic spinning will set the reader nonspinnable bit once they acquire the write lock. Similarly, readers that observe the setting of reader nonspinnable bit at slowpath entry will also set the reader nonspinnable bit when they acquire the read lock via the wakeup path. Once the reader nonspinnable bit is on, it will only be reset when a writer is able to acquire the rwsem in the fast path or somehow a reader or writer in the slowpath doesn't observe the nonspinable bit. This is to discourage reader optmistic spinning on that particular rwsem and make writers more preferred. This adaptive disabling of reader optimistic spinning will alleviate some of the negative side effect of this feature. In addition, this patch tries to make readers in the spinning queue follow the phase-fair principle after quitting optimistic spinning by checking if another reader has somehow acquired a read lock after this reader enters the optimistic spinning queue. If so and the rwsem is still reader-owned, this reader is in the right read-phase and can attempt to acquire the lock. On a 2-socket 40-core 80-thread Skylake system, the page_fault1 test of the will-it-scale benchmark was run with various number of threads. The number of operations done before reader optimistic spinning patches, this patch and after this patch were: Threads Before rspin Before patch After patch %change ------- ------------ ------------ ----------- ------- 20 5541068 5345484 5455667 -3.5%/ +2.1% 40 10185150 7292313 9219276 -28.5%/+26.4% 60 8196733 6460517 7181209 -21.2%/+11.2% 80 9508864 6739559 8107025 -29.1%/+20.3% This patch doesn't recover all the lost performance, but it is more than half. Given the fact that reader optimistic spinning does benefit some workloads, this is a good compromise. Using the rwsem locking microbenchmark with very short critical section, this patch doesn't have too much impact on locking performance as shown by the locking rates (kops/s) below with equal numbers of readers and writers before and after this patch: # of Threads Pre-patch Post-patch ------------ --------- ---------- 2 4,730 4,969 4 4,814 4,786 8 4,866 4,815 16 4,715 4,511 32 3,338 3,500 64 3,212 3,389 80 3,110 3,044 When running the locking microbenchmark with 40 dedicated reader and writer threads, however, the reader performance is curtailed to favor the writer. Before patch: 40 readers, Iterations Min/Mean/Max = 204,026/234,309/254,816 40 writers, Iterations Min/Mean/Max = 88,515/95,884/115,644 After patch: 40 readers, Iterations Min/Mean/Max = 33,813/35,260/36,791 40 writers, Iterations Min/Mean/Max = 95,368/96,565/97,798 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-16-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:14 +03:00
unsigned long val = (unsigned long)owner | RWSEM_READER_OWNED |
locking/rwsem: Remove reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It also improves the chance that a reader can get the lock as writer optimistic spinning disproportionally favors writers much more than readers. Since commit d3681e269fff ("locking/rwsem: Wake up almost all readers in wait queue"), all the waiting readers are woken up so that they can all get the read lock and run in parallel. When the number of contending readers is large, allowing reader optimistic spinning will likely cause reader fragmentation where multiple smaller groups of readers can get the read lock in a sequential manner separated by writers. That reduces reader parallelism. One possible way to address that drawback is to limit the number of readers (preferably one) that can do optimistic spinning. These readers act as representatives of all the waiting readers in the wait queue as they will wake up all those waiting readers once they get the lock. Alternatively, as reader optimistic lock stealing has already enhanced fairness to readers, it may be easier to just remove reader optimistic spinning and simplifying the optimistic spinning code as a result. Performance measurements (locking throughput kops/s) using a locking microbenchmark with 50/50 reader/writer distribution and turbo-boost disabled was done on a 2-socket Cascade Lake system (48-core 96-thread) to see the impacts of these changes: 1) Vanilla - 5.10-rc3 kernel 2) Before - 5.10-rc3 kernel with previous patches in this series 2) limit-rspin - 5.10-rc3 kernel with limited reader spinning patch 3) no-rspin - 5.10-rc3 kernel with reader spinning disabled # of threads CS Load Vanilla Before limit-rspin no-rspin ------------ ------- ------- ------ ----------- -------- 2 1 5,185 5,662 5,214 5,077 4 1 5,107 4,983 5,188 4,760 8 1 4,782 4,564 4,720 4,628 16 1 4,680 4,053 4,567 3,402 32 1 4,299 1,115 1,118 1,098 64 1 3,218 983 1,001 957 96 1 1,938 944 957 930 2 20 2,008 2,128 2,264 1,665 4 20 1,390 1,033 1,046 1,101 8 20 1,472 1,155 1,098 1,213 16 20 1,332 1,077 1,089 1,122 32 20 967 914 917 980 64 20 787 874 891 858 96 20 730 836 847 844 2 100 372 356 360 355 4 100 492 425 434 392 8 100 533 537 529 538 16 100 548 572 568 598 32 100 499 520 527 537 64 100 466 517 526 512 96 100 406 497 506 509 The column "CS Load" represents the number of pause instructions issued in the locking critical section. A CS load of 1 is extremely short and is not likey in real situations. A load of 20 (moderate) and 100 (long) are more realistic. It can be seen that the previous patches in this series have reduced performance in general except in highly contended cases with moderate or long critical sections that performance improves a bit. This change is mostly caused by the "Prevent potential lock starvation" patch that reduce reader optimistic spinning and hence reduce reader fragmentation. The patch that further limit reader optimistic spinning doesn't seem to have too much impact on overall performance as shown in the benchmark data. The patch that disables reader optimistic spinning shows reduced performance at lightly loaded cases, but comparable or slightly better performance on with heavier contention. This patch just removes reader optimistic spinning for now. As readers are not going to do optimistic spinning anymore, we don't need to consider if the OSQ is empty or not when doing lock stealing. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-6-longman@redhat.com
2020-11-21 07:14:16 +03:00
(atomic_long_read(&sem->owner) & RWSEM_NONSPINNABLE);
atomic_long_set(&sem->owner, val);
}
static inline void rwsem_set_reader_owned(struct rw_semaphore *sem)
{
__rwsem_set_reader_owned(sem, current);
}
/*
* Return true if the rwsem is owned by a reader.
*/
static inline bool is_rwsem_reader_owned(struct rw_semaphore *sem)
{
#ifdef CONFIG_DEBUG_RWSEMS
/*
* Check the count to see if it is write-locked.
*/
long count = atomic_long_read(&sem->count);
if (count & RWSEM_WRITER_MASK)
return false;
#endif
return rwsem_test_oflags(sem, RWSEM_READER_OWNED);
}
#ifdef CONFIG_DEBUG_RWSEMS
/*
* With CONFIG_DEBUG_RWSEMS configured, it will make sure that if there
* is a task pointer in owner of a reader-owned rwsem, it will be the
* real owner or one of the real owners. The only exception is when the
* unlock is done by up_read_non_owner().
*/
static inline void rwsem_clear_reader_owned(struct rw_semaphore *sem)
{
unsigned long val = atomic_long_read(&sem->owner);
while ((val & ~RWSEM_OWNER_FLAGS_MASK) == (unsigned long)current) {
if (atomic_long_try_cmpxchg(&sem->owner, &val,
val & RWSEM_OWNER_FLAGS_MASK))
return;
}
}
#else
static inline void rwsem_clear_reader_owned(struct rw_semaphore *sem)
{
}
#endif
locking/rwsem: Enable time-based spinning on reader-owned rwsem When the rwsem is owned by reader, writers stop optimistic spinning simply because there is no easy way to figure out if all the readers are actively running or not. However, there are scenarios where the readers are unlikely to sleep and optimistic spinning can help performance. This patch provides a simple mechanism for spinning on a reader-owned rwsem by a writer. It is a time threshold based spinning where the allowable spinning time can vary from 10us to 25us depending on the condition of the rwsem. When the time threshold is exceeded, the nonspinnable bits will be set in the owner field to indicate that no more optimistic spinning will be allowed on this rwsem until it becomes writer owned again. Not even readers is allowed to acquire the reader-locked rwsem by optimistic spinning for fairness. We also want a writer to acquire the lock after the readers hold the lock for a relatively long time. In order to give preference to writers under such a circumstance, the single RWSEM_NONSPINNABLE bit is now split into two - one for reader and one for writer. When optimistic spinning is disabled, both bits will be set. When the reader count drop down to 0, the writer nonspinnable bit will be cleared to allow writers to spin on the lock, but not the readers. When a writer acquires the lock, it will write its own task structure pointer into sem->owner and clear the reader nonspinnable bit in the process. The time taken for each iteration of the reader-owned rwsem spinning loop varies. Below are sample minimum elapsed times for 16 iterations of the loop. System Time for 16 Iterations ------ ---------------------- 1-socket Skylake ~800ns 4-socket Broadwell ~300ns 2-socket ThunderX2 (arm64) ~250ns When the lock cacheline is contended, we can see up to almost 10X increase in elapsed time. So 25us will be at most 500, 1300 and 1600 iterations for each of the above systems. With a locking microbenchmark running on 5.1 based kernel, the total locking rates (in kops/s) on a 8-socket IvyBridge-EX system with equal numbers of readers and writers before and after this patch were as follows: # of Threads Pre-patch Post-patch ------------ --------- ---------- 2 1,759 6,684 4 1,684 6,738 8 1,074 7,222 16 900 7,163 32 458 7,316 64 208 520 128 168 425 240 143 474 This patch gives a big boost in performance for mixed reader/writer workloads. With 32 locking threads, the rwsem lock event data were: rwsem_opt_fail=79850 rwsem_opt_nospin=5069 rwsem_opt_rlock=597484 rwsem_opt_wlock=957339 rwsem_sleep_reader=57782 rwsem_sleep_writer=55663 With 64 locking threads, the data looked like: rwsem_opt_fail=346723 rwsem_opt_nospin=6293 rwsem_opt_rlock=1127119 rwsem_opt_wlock=1400628 rwsem_sleep_reader=308201 rwsem_sleep_writer=72281 So a lot more threads acquired the lock in the slowpath and more threads went to sleep. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-15-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:13 +03:00
/*
* Set the RWSEM_NONSPINNABLE bits if the RWSEM_READER_OWNED flag
* remains set. Otherwise, the operation will be aborted.
*/
static inline void rwsem_set_nonspinnable(struct rw_semaphore *sem)
{
unsigned long owner = atomic_long_read(&sem->owner);
do {
if (!(owner & RWSEM_READER_OWNED))
break;
if (owner & RWSEM_NONSPINNABLE)
break;
} while (!atomic_long_try_cmpxchg(&sem->owner, &owner,
owner | RWSEM_NONSPINNABLE));
}
static inline bool rwsem_read_trylock(struct rw_semaphore *sem, long *cntp)
{
*cntp = atomic_long_add_return_acquire(RWSEM_READER_BIAS, &sem->count);
if (WARN_ON_ONCE(*cntp < 0))
rwsem_set_nonspinnable(sem);
if (!(*cntp & RWSEM_READ_FAILED_MASK)) {
rwsem_set_reader_owned(sem);
return true;
}
return false;
}
static inline bool rwsem_write_trylock(struct rw_semaphore *sem)
{
long tmp = RWSEM_UNLOCKED_VALUE;
if (atomic_long_try_cmpxchg_acquire(&sem->count, &tmp, RWSEM_WRITER_LOCKED)) {
rwsem_set_owner(sem);
return true;
}
return false;
}
/*
* Return just the real task structure pointer of the owner
*/
static inline struct task_struct *rwsem_owner(struct rw_semaphore *sem)
{
return (struct task_struct *)
(atomic_long_read(&sem->owner) & ~RWSEM_OWNER_FLAGS_MASK);
}
/*
* Return the real task structure pointer of the owner and the embedded
* flags in the owner. pflags must be non-NULL.
*/
static inline struct task_struct *
rwsem_owner_flags(struct rw_semaphore *sem, unsigned long *pflags)
{
unsigned long owner = atomic_long_read(&sem->owner);
*pflags = owner & RWSEM_OWNER_FLAGS_MASK;
return (struct task_struct *)(owner & ~RWSEM_OWNER_FLAGS_MASK);
}
/*
* Guide to the rw_semaphore's count field.
*
* When the RWSEM_WRITER_LOCKED bit in count is set, the lock is owned
* by a writer.
*
* The lock is owned by readers when
* (1) the RWSEM_WRITER_LOCKED isn't set in count,
* (2) some of the reader bits are set in count, and
* (3) the owner field has RWSEM_READ_OWNED bit set.
*
* Having some reader bits set is not enough to guarantee a readers owned
* lock as the readers may be in the process of backing out from the count
* and a writer has just released the lock. So another writer may steal
* the lock immediately after that.
*/
/*
* Initialize an rwsem:
*/
void __init_rwsem(struct rw_semaphore *sem, const char *name,
struct lock_class_key *key)
{
#ifdef CONFIG_DEBUG_LOCK_ALLOC
/*
* Make sure we are not reinitializing a held semaphore:
*/
debug_check_no_locks_freed((void *)sem, sizeof(*sem));
lockdep: Introduce wait-type checks Extend lockdep to validate lock wait-type context. The current wait-types are: LD_WAIT_FREE, /* wait free, rcu etc.. */ LD_WAIT_SPIN, /* spin loops, raw_spinlock_t etc.. */ LD_WAIT_CONFIG, /* CONFIG_PREEMPT_LOCK, spinlock_t etc.. */ LD_WAIT_SLEEP, /* sleeping locks, mutex_t etc.. */ Where lockdep validates that the current lock (the one being acquired) fits in the current wait-context (as generated by the held stack). This ensures that there is no attempt to acquire mutexes while holding spinlocks, to acquire spinlocks while holding raw_spinlocks and so on. In other words, its a more fancy might_sleep(). Obviously RCU made the entire ordeal more complex than a simple single value test because RCU can be acquired in (pretty much) any context and while it presents a context to nested locks it is not the same as it got acquired in. Therefore its necessary to split the wait_type into two values, one representing the acquire (outer) and one representing the nested context (inner). For most 'normal' locks these two are the same. [ To make static initialization easier we have the rule that: .outer == INV means .outer == .inner; because INV == 0. ] It further means that its required to find the minimal .inner of the held stack to compare against the outer of the new lock; because while 'normal' RCU presents a CONFIG type to nested locks, if it is taken while already holding a SPIN type it obviously doesn't relax the rules. Below is an example output generated by the trivial test code: raw_spin_lock(&foo); spin_lock(&bar); spin_unlock(&bar); raw_spin_unlock(&foo); [ BUG: Invalid wait context ] ----------------------------- swapper/0/1 is trying to lock: ffffc90000013f20 (&bar){....}-{3:3}, at: kernel_init+0xdb/0x187 other info that might help us debug this: 1 lock held by swapper/0/1: #0: ffffc90000013ee0 (&foo){+.+.}-{2:2}, at: kernel_init+0xd1/0x187 The way to read it is to look at the new -{n,m} part in the lock description; -{3:3} for the attempted lock, and try and match that up to the held locks, which in this case is the one: -{2,2}. This tells that the acquiring lock requires a more relaxed environment than presented by the lock stack. Currently only the normal locks and RCU are converted, the rest of the lockdep users defaults to .inner = INV which is ignored. More conversions can be done when desired. The check for spinlock_t nesting is not enabled by default. It's a separate config option for now as there are known problems which are currently addressed. The config option allows to identify these problems and to verify that the solutions found are indeed solving them. The config switch will be removed and the checks will permanently enabled once the vast majority of issues has been addressed. [ bigeasy: Move LD_WAIT_FREE,… out of CONFIG_LOCKDEP to avoid compile failure with CONFIG_DEBUG_SPINLOCK + !CONFIG_LOCKDEP] [ tglx: Add the config option ] Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/20200321113242.427089655@linutronix.de
2020-03-21 14:26:01 +03:00
lockdep_init_map_wait(&sem->dep_map, name, key, 0, LD_WAIT_SLEEP);
#endif
#ifdef CONFIG_DEBUG_RWSEMS
sem->magic = sem;
#endif
atomic_long_set(&sem->count, RWSEM_UNLOCKED_VALUE);
raw_spin_lock_init(&sem->wait_lock);
INIT_LIST_HEAD(&sem->wait_list);
atomic_long_set(&sem->owner, 0L);
#ifdef CONFIG_RWSEM_SPIN_ON_OWNER
osq_lock_init(&sem->osq);
#endif
}
EXPORT_SYMBOL(__init_rwsem);
enum rwsem_waiter_type {
RWSEM_WAITING_FOR_WRITE,
RWSEM_WAITING_FOR_READ
};
struct rwsem_waiter {
struct list_head list;
struct task_struct *task;
enum rwsem_waiter_type type;
locking/rwsem: Implement lock handoff to prevent lock starvation Because of writer lock stealing, it is possible that a constant stream of incoming writers will cause a waiting writer or reader to wait indefinitely leading to lock starvation. This patch implements a lock handoff mechanism to disable lock stealing and force lock handoff to the first waiter or waiters (for readers) in the queue after at least a 4ms waiting period unless it is a RT writer task which doesn't need to wait. The waiting period is used to avoid discouraging lock stealing too much to affect performance. The setting and clearing of the handoff bit is serialized by the wait_lock. So racing is not possible. A rwsem microbenchmark was run for 5 seconds on a 2-socket 40-core 80-thread Skylake system with a v5.1 based kernel and 240 write_lock threads with 5us sleep critical section. Before the patch, the min/mean/max numbers of locking operations for the locking threads were 1/7,792/173,696. After the patch, the figures became 5,842/6,542/7,458. It can be seen that the rwsem became much more fair, though there was a drop of about 16% in the mean locking operations done which was a tradeoff of having better fairness. Making the waiter set the handoff bit right after the first wakeup can impact performance especially with a mixed reader/writer workload. With the same microbenchmark with short critical section and equal number of reader and writer threads (40/40), the reader/writer locking operation counts with the current patch were: 40 readers, Iterations Min/Mean/Max = 1,793/1,794/1,796 40 writers, Iterations Min/Mean/Max = 1,793/34,956/86,081 By making waiter set handoff bit immediately after wakeup: 40 readers, Iterations Min/Mean/Max = 43/44/46 40 writers, Iterations Min/Mean/Max = 43/1,263/3,191 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-8-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:06 +03:00
unsigned long timeout;
};
locking/rwsem: Implement lock handoff to prevent lock starvation Because of writer lock stealing, it is possible that a constant stream of incoming writers will cause a waiting writer or reader to wait indefinitely leading to lock starvation. This patch implements a lock handoff mechanism to disable lock stealing and force lock handoff to the first waiter or waiters (for readers) in the queue after at least a 4ms waiting period unless it is a RT writer task which doesn't need to wait. The waiting period is used to avoid discouraging lock stealing too much to affect performance. The setting and clearing of the handoff bit is serialized by the wait_lock. So racing is not possible. A rwsem microbenchmark was run for 5 seconds on a 2-socket 40-core 80-thread Skylake system with a v5.1 based kernel and 240 write_lock threads with 5us sleep critical section. Before the patch, the min/mean/max numbers of locking operations for the locking threads were 1/7,792/173,696. After the patch, the figures became 5,842/6,542/7,458. It can be seen that the rwsem became much more fair, though there was a drop of about 16% in the mean locking operations done which was a tradeoff of having better fairness. Making the waiter set the handoff bit right after the first wakeup can impact performance especially with a mixed reader/writer workload. With the same microbenchmark with short critical section and equal number of reader and writer threads (40/40), the reader/writer locking operation counts with the current patch were: 40 readers, Iterations Min/Mean/Max = 1,793/1,794/1,796 40 writers, Iterations Min/Mean/Max = 1,793/34,956/86,081 By making waiter set handoff bit immediately after wakeup: 40 readers, Iterations Min/Mean/Max = 43/44/46 40 writers, Iterations Min/Mean/Max = 43/1,263/3,191 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-8-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:06 +03:00
#define rwsem_first_waiter(sem) \
list_first_entry(&sem->wait_list, struct rwsem_waiter, list)
enum rwsem_wake_type {
RWSEM_WAKE_ANY, /* Wake whatever's at head of wait list */
RWSEM_WAKE_READERS, /* Wake readers only */
RWSEM_WAKE_READ_OWNED /* Waker thread holds the read lock */
};
locking/rwsem: Implement lock handoff to prevent lock starvation Because of writer lock stealing, it is possible that a constant stream of incoming writers will cause a waiting writer or reader to wait indefinitely leading to lock starvation. This patch implements a lock handoff mechanism to disable lock stealing and force lock handoff to the first waiter or waiters (for readers) in the queue after at least a 4ms waiting period unless it is a RT writer task which doesn't need to wait. The waiting period is used to avoid discouraging lock stealing too much to affect performance. The setting and clearing of the handoff bit is serialized by the wait_lock. So racing is not possible. A rwsem microbenchmark was run for 5 seconds on a 2-socket 40-core 80-thread Skylake system with a v5.1 based kernel and 240 write_lock threads with 5us sleep critical section. Before the patch, the min/mean/max numbers of locking operations for the locking threads were 1/7,792/173,696. After the patch, the figures became 5,842/6,542/7,458. It can be seen that the rwsem became much more fair, though there was a drop of about 16% in the mean locking operations done which was a tradeoff of having better fairness. Making the waiter set the handoff bit right after the first wakeup can impact performance especially with a mixed reader/writer workload. With the same microbenchmark with short critical section and equal number of reader and writer threads (40/40), the reader/writer locking operation counts with the current patch were: 40 readers, Iterations Min/Mean/Max = 1,793/1,794/1,796 40 writers, Iterations Min/Mean/Max = 1,793/34,956/86,081 By making waiter set handoff bit immediately after wakeup: 40 readers, Iterations Min/Mean/Max = 43/44/46 40 writers, Iterations Min/Mean/Max = 43/1,263/3,191 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-8-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:06 +03:00
enum writer_wait_state {
WRITER_NOT_FIRST, /* Writer is not first in wait list */
WRITER_FIRST, /* Writer is first in wait list */
WRITER_HANDOFF /* Writer is first & handoff needed */
};
/*
* The typical HZ value is either 250 or 1000. So set the minimum waiting
* time to at least 4ms or 1 jiffy (if it is higher than 4ms) in the wait
* queue before initiating the handoff protocol.
*/
#define RWSEM_WAIT_TIMEOUT DIV_ROUND_UP(HZ, 250)
locking/rwsem: Wake up almost all readers in wait queue When the front of the wait queue is a reader, other readers immediately following the first reader will also be woken up at the same time. However, if there is a writer in between. Those readers behind the writer will not be woken up. Because of optimistic spinning, the lock acquisition order is not FIFO anyway. The lock handoff mechanism will ensure that lock starvation will not happen. Assuming that the lock hold times of the other readers still in the queue will be about the same as the readers that are being woken up, there is really not much additional cost other than the additional latency due to the wakeup of additional tasks by the waker. Therefore all the readers up to a maximum of 256 in the queue are woken up when the first waiter is a reader to improve reader throughput. This is somewhat similar in concept to a phase-fair R/W lock. With a locking microbenchmark running on 5.1 based kernel, the total locking rates (in kops/s) on a 8-socket IvyBridge-EX system with equal numbers of readers and writers before and after this patch were as follows: # of Threads Pre-Patch Post-patch ------------ --------- ---------- 4 1,641 1,674 8 731 1,062 16 564 924 32 78 300 64 38 195 240 50 149 There is no performance gain at low contention level. At high contention level, however, this patch gives a pretty decent performance boost. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-11-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:09 +03:00
/*
* Magic number to batch-wakeup waiting readers, even when writers are
* also present in the queue. This both limits the amount of work the
* waking thread must do and also prevents any potential counter overflow,
* however unlikely.
*/
#define MAX_READERS_WAKEUP 0x100
/*
* handle the lock release when processes blocked on it that can now run
* - if we come here from up_xxxx(), then the RWSEM_FLAG_WAITERS bit must
* have been set.
* - there must be someone on the queue
* - the wait_lock must be held by the caller
* - tasks are marked for wakeup, the caller must later invoke wake_up_q()
* to actually wakeup the blocked task(s) and drop the reference count,
* preferably when the wait_lock is released
* - woken process blocks are discarded from the list after having task zeroed
* - writers are only marked woken if downgrading is false
*/
locking/rwsem: Code cleanup after files merging After merging all the relevant rwsem code into one single file, there are a number of optimizations and cleanups that can be done: 1) Remove all the EXPORT_SYMBOL() calls for functions that are not accessed elsewhere. 2) Remove all the __visible tags as none of the functions will be called from assembly code anymore. 3) Make all the internal functions static. 4) Remove some unneeded blank lines. 5) Remove the intermediate rwsem_down_{read|write}_failed*() functions and rename __rwsem_down_{read|write}_failed_common() to rwsem_down_{read|write}_slowpath(). 6) Remove "__" prefix of __rwsem_mark_wake(). 7) Use atomic_long_try_cmpxchg_acquire() as much as possible. 8) Remove the rwsem_rtrylock and rwsem_wtrylock lock events as they are not that useful. That enables the compiler to do better optimization and reduce code size. The text+data size of rwsem.o on an x86-64 machine with gcc8 was reduced from 10237 bytes to 5030 bytes with this change. Suggested-by: Peter Zijlstra <peterz@infradead.org> Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-6-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:04 +03:00
static void rwsem_mark_wake(struct rw_semaphore *sem,
enum rwsem_wake_type wake_type,
struct wake_q_head *wake_q)
{
struct rwsem_waiter *waiter, *tmp;
long oldcount, woken = 0, adjustment = 0;
struct list_head wlist;
locking/rwsem: Implement lock handoff to prevent lock starvation Because of writer lock stealing, it is possible that a constant stream of incoming writers will cause a waiting writer or reader to wait indefinitely leading to lock starvation. This patch implements a lock handoff mechanism to disable lock stealing and force lock handoff to the first waiter or waiters (for readers) in the queue after at least a 4ms waiting period unless it is a RT writer task which doesn't need to wait. The waiting period is used to avoid discouraging lock stealing too much to affect performance. The setting and clearing of the handoff bit is serialized by the wait_lock. So racing is not possible. A rwsem microbenchmark was run for 5 seconds on a 2-socket 40-core 80-thread Skylake system with a v5.1 based kernel and 240 write_lock threads with 5us sleep critical section. Before the patch, the min/mean/max numbers of locking operations for the locking threads were 1/7,792/173,696. After the patch, the figures became 5,842/6,542/7,458. It can be seen that the rwsem became much more fair, though there was a drop of about 16% in the mean locking operations done which was a tradeoff of having better fairness. Making the waiter set the handoff bit right after the first wakeup can impact performance especially with a mixed reader/writer workload. With the same microbenchmark with short critical section and equal number of reader and writer threads (40/40), the reader/writer locking operation counts with the current patch were: 40 readers, Iterations Min/Mean/Max = 1,793/1,794/1,796 40 writers, Iterations Min/Mean/Max = 1,793/34,956/86,081 By making waiter set handoff bit immediately after wakeup: 40 readers, Iterations Min/Mean/Max = 43/44/46 40 writers, Iterations Min/Mean/Max = 43/1,263/3,191 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-8-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:06 +03:00
lockdep_assert_held(&sem->wait_lock);
/*
* Take a peek at the queue head waiter such that we can determine
* the wakeup(s) to perform.
*/
locking/rwsem: Implement lock handoff to prevent lock starvation Because of writer lock stealing, it is possible that a constant stream of incoming writers will cause a waiting writer or reader to wait indefinitely leading to lock starvation. This patch implements a lock handoff mechanism to disable lock stealing and force lock handoff to the first waiter or waiters (for readers) in the queue after at least a 4ms waiting period unless it is a RT writer task which doesn't need to wait. The waiting period is used to avoid discouraging lock stealing too much to affect performance. The setting and clearing of the handoff bit is serialized by the wait_lock. So racing is not possible. A rwsem microbenchmark was run for 5 seconds on a 2-socket 40-core 80-thread Skylake system with a v5.1 based kernel and 240 write_lock threads with 5us sleep critical section. Before the patch, the min/mean/max numbers of locking operations for the locking threads were 1/7,792/173,696. After the patch, the figures became 5,842/6,542/7,458. It can be seen that the rwsem became much more fair, though there was a drop of about 16% in the mean locking operations done which was a tradeoff of having better fairness. Making the waiter set the handoff bit right after the first wakeup can impact performance especially with a mixed reader/writer workload. With the same microbenchmark with short critical section and equal number of reader and writer threads (40/40), the reader/writer locking operation counts with the current patch were: 40 readers, Iterations Min/Mean/Max = 1,793/1,794/1,796 40 writers, Iterations Min/Mean/Max = 1,793/34,956/86,081 By making waiter set handoff bit immediately after wakeup: 40 readers, Iterations Min/Mean/Max = 43/44/46 40 writers, Iterations Min/Mean/Max = 43/1,263/3,191 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-8-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:06 +03:00
waiter = rwsem_first_waiter(sem);
if (waiter->type == RWSEM_WAITING_FOR_WRITE) {
if (wake_type == RWSEM_WAKE_ANY) {
/*
* Mark writer at the front of the queue for wakeup.
* Until the task is actually later awoken later by
* the caller, other writers are able to steal it.
* Readers, on the other hand, will block as they
* will notice the queued writer.
*/
wake_q_add(wake_q, waiter->task);
lockevent_inc(rwsem_wake_writer);
}
return;
}
/*
* No reader wakeup if there are too many of them already.
*/
if (unlikely(atomic_long_read(&sem->count) < 0))
return;
/*
* Writers might steal the lock before we grant it to the next reader.
* We prefer to do the first reader grant before counting readers
* so we can bail out early if a writer stole the lock.
*/
if (wake_type != RWSEM_WAKE_READ_OWNED) {
locking/rwsem: Adaptive disabling of reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It makes readers relatively more preferred than writers. When a writer times out spinning on a reader-owned lock and set the nospinnable bits, there are two main reasons for that. 1) The reader critical section is long, perhaps the task sleeps after acquiring the read lock. 2) There are just too many readers contending the lock causing it to take a while to service all of them. In the former case, long reader critical section will impede the progress of writers which is usually more important for system performance. In the later case, reader optimistic spinning tends to make the reader groups that contain readers that acquire the lock together smaller leading to more of them. That may hurt performance in some cases. In other words, the setting of nonspinnable bits indicates that reader optimistic spinning may not be helpful for those workloads that cause it. Therefore, any writers that have observed the setting of the writer nonspinnable bit for a given rwsem after they fail to acquire the lock via optimistic spinning will set the reader nonspinnable bit once they acquire the write lock. Similarly, readers that observe the setting of reader nonspinnable bit at slowpath entry will also set the reader nonspinnable bit when they acquire the read lock via the wakeup path. Once the reader nonspinnable bit is on, it will only be reset when a writer is able to acquire the rwsem in the fast path or somehow a reader or writer in the slowpath doesn't observe the nonspinable bit. This is to discourage reader optmistic spinning on that particular rwsem and make writers more preferred. This adaptive disabling of reader optimistic spinning will alleviate some of the negative side effect of this feature. In addition, this patch tries to make readers in the spinning queue follow the phase-fair principle after quitting optimistic spinning by checking if another reader has somehow acquired a read lock after this reader enters the optimistic spinning queue. If so and the rwsem is still reader-owned, this reader is in the right read-phase and can attempt to acquire the lock. On a 2-socket 40-core 80-thread Skylake system, the page_fault1 test of the will-it-scale benchmark was run with various number of threads. The number of operations done before reader optimistic spinning patches, this patch and after this patch were: Threads Before rspin Before patch After patch %change ------- ------------ ------------ ----------- ------- 20 5541068 5345484 5455667 -3.5%/ +2.1% 40 10185150 7292313 9219276 -28.5%/+26.4% 60 8196733 6460517 7181209 -21.2%/+11.2% 80 9508864 6739559 8107025 -29.1%/+20.3% This patch doesn't recover all the lost performance, but it is more than half. Given the fact that reader optimistic spinning does benefit some workloads, this is a good compromise. Using the rwsem locking microbenchmark with very short critical section, this patch doesn't have too much impact on locking performance as shown by the locking rates (kops/s) below with equal numbers of readers and writers before and after this patch: # of Threads Pre-patch Post-patch ------------ --------- ---------- 2 4,730 4,969 4 4,814 4,786 8 4,866 4,815 16 4,715 4,511 32 3,338 3,500 64 3,212 3,389 80 3,110 3,044 When running the locking microbenchmark with 40 dedicated reader and writer threads, however, the reader performance is curtailed to favor the writer. Before patch: 40 readers, Iterations Min/Mean/Max = 204,026/234,309/254,816 40 writers, Iterations Min/Mean/Max = 88,515/95,884/115,644 After patch: 40 readers, Iterations Min/Mean/Max = 33,813/35,260/36,791 40 writers, Iterations Min/Mean/Max = 95,368/96,565/97,798 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-16-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:14 +03:00
struct task_struct *owner;
adjustment = RWSEM_READER_BIAS;
oldcount = atomic_long_fetch_add(adjustment, &sem->count);
if (unlikely(oldcount & RWSEM_WRITER_MASK)) {
locking/rwsem: Implement lock handoff to prevent lock starvation Because of writer lock stealing, it is possible that a constant stream of incoming writers will cause a waiting writer or reader to wait indefinitely leading to lock starvation. This patch implements a lock handoff mechanism to disable lock stealing and force lock handoff to the first waiter or waiters (for readers) in the queue after at least a 4ms waiting period unless it is a RT writer task which doesn't need to wait. The waiting period is used to avoid discouraging lock stealing too much to affect performance. The setting and clearing of the handoff bit is serialized by the wait_lock. So racing is not possible. A rwsem microbenchmark was run for 5 seconds on a 2-socket 40-core 80-thread Skylake system with a v5.1 based kernel and 240 write_lock threads with 5us sleep critical section. Before the patch, the min/mean/max numbers of locking operations for the locking threads were 1/7,792/173,696. After the patch, the figures became 5,842/6,542/7,458. It can be seen that the rwsem became much more fair, though there was a drop of about 16% in the mean locking operations done which was a tradeoff of having better fairness. Making the waiter set the handoff bit right after the first wakeup can impact performance especially with a mixed reader/writer workload. With the same microbenchmark with short critical section and equal number of reader and writer threads (40/40), the reader/writer locking operation counts with the current patch were: 40 readers, Iterations Min/Mean/Max = 1,793/1,794/1,796 40 writers, Iterations Min/Mean/Max = 1,793/34,956/86,081 By making waiter set handoff bit immediately after wakeup: 40 readers, Iterations Min/Mean/Max = 43/44/46 40 writers, Iterations Min/Mean/Max = 43/1,263/3,191 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-8-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:06 +03:00
/*
* When we've been waiting "too" long (for writers
* to give up the lock), request a HANDOFF to
* force the issue.
*/
if (!(oldcount & RWSEM_FLAG_HANDOFF) &&
time_after(jiffies, waiter->timeout)) {
adjustment -= RWSEM_FLAG_HANDOFF;
lockevent_inc(rwsem_rlock_handoff);
}
atomic_long_add(-adjustment, &sem->count);
return;
}
/*
* Set it to reader-owned to give spinners an early
* indication that readers now have the lock.
locking/rwsem: Adaptive disabling of reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It makes readers relatively more preferred than writers. When a writer times out spinning on a reader-owned lock and set the nospinnable bits, there are two main reasons for that. 1) The reader critical section is long, perhaps the task sleeps after acquiring the read lock. 2) There are just too many readers contending the lock causing it to take a while to service all of them. In the former case, long reader critical section will impede the progress of writers which is usually more important for system performance. In the later case, reader optimistic spinning tends to make the reader groups that contain readers that acquire the lock together smaller leading to more of them. That may hurt performance in some cases. In other words, the setting of nonspinnable bits indicates that reader optimistic spinning may not be helpful for those workloads that cause it. Therefore, any writers that have observed the setting of the writer nonspinnable bit for a given rwsem after they fail to acquire the lock via optimistic spinning will set the reader nonspinnable bit once they acquire the write lock. Similarly, readers that observe the setting of reader nonspinnable bit at slowpath entry will also set the reader nonspinnable bit when they acquire the read lock via the wakeup path. Once the reader nonspinnable bit is on, it will only be reset when a writer is able to acquire the rwsem in the fast path or somehow a reader or writer in the slowpath doesn't observe the nonspinable bit. This is to discourage reader optmistic spinning on that particular rwsem and make writers more preferred. This adaptive disabling of reader optimistic spinning will alleviate some of the negative side effect of this feature. In addition, this patch tries to make readers in the spinning queue follow the phase-fair principle after quitting optimistic spinning by checking if another reader has somehow acquired a read lock after this reader enters the optimistic spinning queue. If so and the rwsem is still reader-owned, this reader is in the right read-phase and can attempt to acquire the lock. On a 2-socket 40-core 80-thread Skylake system, the page_fault1 test of the will-it-scale benchmark was run with various number of threads. The number of operations done before reader optimistic spinning patches, this patch and after this patch were: Threads Before rspin Before patch After patch %change ------- ------------ ------------ ----------- ------- 20 5541068 5345484 5455667 -3.5%/ +2.1% 40 10185150 7292313 9219276 -28.5%/+26.4% 60 8196733 6460517 7181209 -21.2%/+11.2% 80 9508864 6739559 8107025 -29.1%/+20.3% This patch doesn't recover all the lost performance, but it is more than half. Given the fact that reader optimistic spinning does benefit some workloads, this is a good compromise. Using the rwsem locking microbenchmark with very short critical section, this patch doesn't have too much impact on locking performance as shown by the locking rates (kops/s) below with equal numbers of readers and writers before and after this patch: # of Threads Pre-patch Post-patch ------------ --------- ---------- 2 4,730 4,969 4 4,814 4,786 8 4,866 4,815 16 4,715 4,511 32 3,338 3,500 64 3,212 3,389 80 3,110 3,044 When running the locking microbenchmark with 40 dedicated reader and writer threads, however, the reader performance is curtailed to favor the writer. Before patch: 40 readers, Iterations Min/Mean/Max = 204,026/234,309/254,816 40 writers, Iterations Min/Mean/Max = 88,515/95,884/115,644 After patch: 40 readers, Iterations Min/Mean/Max = 33,813/35,260/36,791 40 writers, Iterations Min/Mean/Max = 95,368/96,565/97,798 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-16-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:14 +03:00
* The reader nonspinnable bit seen at slowpath entry of
* the reader is copied over.
*/
locking/rwsem: Adaptive disabling of reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It makes readers relatively more preferred than writers. When a writer times out spinning on a reader-owned lock and set the nospinnable bits, there are two main reasons for that. 1) The reader critical section is long, perhaps the task sleeps after acquiring the read lock. 2) There are just too many readers contending the lock causing it to take a while to service all of them. In the former case, long reader critical section will impede the progress of writers which is usually more important for system performance. In the later case, reader optimistic spinning tends to make the reader groups that contain readers that acquire the lock together smaller leading to more of them. That may hurt performance in some cases. In other words, the setting of nonspinnable bits indicates that reader optimistic spinning may not be helpful for those workloads that cause it. Therefore, any writers that have observed the setting of the writer nonspinnable bit for a given rwsem after they fail to acquire the lock via optimistic spinning will set the reader nonspinnable bit once they acquire the write lock. Similarly, readers that observe the setting of reader nonspinnable bit at slowpath entry will also set the reader nonspinnable bit when they acquire the read lock via the wakeup path. Once the reader nonspinnable bit is on, it will only be reset when a writer is able to acquire the rwsem in the fast path or somehow a reader or writer in the slowpath doesn't observe the nonspinable bit. This is to discourage reader optmistic spinning on that particular rwsem and make writers more preferred. This adaptive disabling of reader optimistic spinning will alleviate some of the negative side effect of this feature. In addition, this patch tries to make readers in the spinning queue follow the phase-fair principle after quitting optimistic spinning by checking if another reader has somehow acquired a read lock after this reader enters the optimistic spinning queue. If so and the rwsem is still reader-owned, this reader is in the right read-phase and can attempt to acquire the lock. On a 2-socket 40-core 80-thread Skylake system, the page_fault1 test of the will-it-scale benchmark was run with various number of threads. The number of operations done before reader optimistic spinning patches, this patch and after this patch were: Threads Before rspin Before patch After patch %change ------- ------------ ------------ ----------- ------- 20 5541068 5345484 5455667 -3.5%/ +2.1% 40 10185150 7292313 9219276 -28.5%/+26.4% 60 8196733 6460517 7181209 -21.2%/+11.2% 80 9508864 6739559 8107025 -29.1%/+20.3% This patch doesn't recover all the lost performance, but it is more than half. Given the fact that reader optimistic spinning does benefit some workloads, this is a good compromise. Using the rwsem locking microbenchmark with very short critical section, this patch doesn't have too much impact on locking performance as shown by the locking rates (kops/s) below with equal numbers of readers and writers before and after this patch: # of Threads Pre-patch Post-patch ------------ --------- ---------- 2 4,730 4,969 4 4,814 4,786 8 4,866 4,815 16 4,715 4,511 32 3,338 3,500 64 3,212 3,389 80 3,110 3,044 When running the locking microbenchmark with 40 dedicated reader and writer threads, however, the reader performance is curtailed to favor the writer. Before patch: 40 readers, Iterations Min/Mean/Max = 204,026/234,309/254,816 40 writers, Iterations Min/Mean/Max = 88,515/95,884/115,644 After patch: 40 readers, Iterations Min/Mean/Max = 33,813/35,260/36,791 40 writers, Iterations Min/Mean/Max = 95,368/96,565/97,798 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-16-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:14 +03:00
owner = waiter->task;
__rwsem_set_reader_owned(sem, owner);
}
/*
locking/rwsem: Wake up almost all readers in wait queue When the front of the wait queue is a reader, other readers immediately following the first reader will also be woken up at the same time. However, if there is a writer in between. Those readers behind the writer will not be woken up. Because of optimistic spinning, the lock acquisition order is not FIFO anyway. The lock handoff mechanism will ensure that lock starvation will not happen. Assuming that the lock hold times of the other readers still in the queue will be about the same as the readers that are being woken up, there is really not much additional cost other than the additional latency due to the wakeup of additional tasks by the waker. Therefore all the readers up to a maximum of 256 in the queue are woken up when the first waiter is a reader to improve reader throughput. This is somewhat similar in concept to a phase-fair R/W lock. With a locking microbenchmark running on 5.1 based kernel, the total locking rates (in kops/s) on a 8-socket IvyBridge-EX system with equal numbers of readers and writers before and after this patch were as follows: # of Threads Pre-Patch Post-patch ------------ --------- ---------- 4 1,641 1,674 8 731 1,062 16 564 924 32 78 300 64 38 195 240 50 149 There is no performance gain at low contention level. At high contention level, however, this patch gives a pretty decent performance boost. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-11-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:09 +03:00
* Grant up to MAX_READERS_WAKEUP read locks to all the readers in the
* queue. We know that the woken will be at least 1 as we accounted
* for above. Note we increment the 'active part' of the count by the
* number of readers before waking any processes up.
*
locking/rwsem: Wake up almost all readers in wait queue When the front of the wait queue is a reader, other readers immediately following the first reader will also be woken up at the same time. However, if there is a writer in between. Those readers behind the writer will not be woken up. Because of optimistic spinning, the lock acquisition order is not FIFO anyway. The lock handoff mechanism will ensure that lock starvation will not happen. Assuming that the lock hold times of the other readers still in the queue will be about the same as the readers that are being woken up, there is really not much additional cost other than the additional latency due to the wakeup of additional tasks by the waker. Therefore all the readers up to a maximum of 256 in the queue are woken up when the first waiter is a reader to improve reader throughput. This is somewhat similar in concept to a phase-fair R/W lock. With a locking microbenchmark running on 5.1 based kernel, the total locking rates (in kops/s) on a 8-socket IvyBridge-EX system with equal numbers of readers and writers before and after this patch were as follows: # of Threads Pre-Patch Post-patch ------------ --------- ---------- 4 1,641 1,674 8 731 1,062 16 564 924 32 78 300 64 38 195 240 50 149 There is no performance gain at low contention level. At high contention level, however, this patch gives a pretty decent performance boost. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-11-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:09 +03:00
* This is an adaptation of the phase-fair R/W locks where at the
* reader phase (first waiter is a reader), all readers are eligible
* to acquire the lock at the same time irrespective of their order
* in the queue. The writers acquire the lock according to their
* order in the queue.
*
* We have to do wakeup in 2 passes to prevent the possibility that
* the reader count may be decremented before it is incremented. It
* is because the to-be-woken waiter may not have slept yet. So it
* may see waiter->task got cleared, finish its critical section and
* do an unlock before the reader count increment.
*
* 1) Collect the read-waiters in a separate list, count them and
* fully increment the reader count in rwsem.
* 2) For each waiters in the new list, clear waiter->task and
* put them into wake_q to be woken up later.
*/
locking/rwsem: Wake up almost all readers in wait queue When the front of the wait queue is a reader, other readers immediately following the first reader will also be woken up at the same time. However, if there is a writer in between. Those readers behind the writer will not be woken up. Because of optimistic spinning, the lock acquisition order is not FIFO anyway. The lock handoff mechanism will ensure that lock starvation will not happen. Assuming that the lock hold times of the other readers still in the queue will be about the same as the readers that are being woken up, there is really not much additional cost other than the additional latency due to the wakeup of additional tasks by the waker. Therefore all the readers up to a maximum of 256 in the queue are woken up when the first waiter is a reader to improve reader throughput. This is somewhat similar in concept to a phase-fair R/W lock. With a locking microbenchmark running on 5.1 based kernel, the total locking rates (in kops/s) on a 8-socket IvyBridge-EX system with equal numbers of readers and writers before and after this patch were as follows: # of Threads Pre-Patch Post-patch ------------ --------- ---------- 4 1,641 1,674 8 731 1,062 16 564 924 32 78 300 64 38 195 240 50 149 There is no performance gain at low contention level. At high contention level, however, this patch gives a pretty decent performance boost. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-11-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:09 +03:00
INIT_LIST_HEAD(&wlist);
list_for_each_entry_safe(waiter, tmp, &sem->wait_list, list) {
if (waiter->type == RWSEM_WAITING_FOR_WRITE)
locking/rwsem: Wake up almost all readers in wait queue When the front of the wait queue is a reader, other readers immediately following the first reader will also be woken up at the same time. However, if there is a writer in between. Those readers behind the writer will not be woken up. Because of optimistic spinning, the lock acquisition order is not FIFO anyway. The lock handoff mechanism will ensure that lock starvation will not happen. Assuming that the lock hold times of the other readers still in the queue will be about the same as the readers that are being woken up, there is really not much additional cost other than the additional latency due to the wakeup of additional tasks by the waker. Therefore all the readers up to a maximum of 256 in the queue are woken up when the first waiter is a reader to improve reader throughput. This is somewhat similar in concept to a phase-fair R/W lock. With a locking microbenchmark running on 5.1 based kernel, the total locking rates (in kops/s) on a 8-socket IvyBridge-EX system with equal numbers of readers and writers before and after this patch were as follows: # of Threads Pre-Patch Post-patch ------------ --------- ---------- 4 1,641 1,674 8 731 1,062 16 564 924 32 78 300 64 38 195 240 50 149 There is no performance gain at low contention level. At high contention level, however, this patch gives a pretty decent performance boost. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-11-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:09 +03:00
continue;
woken++;
locking/rwsem: Wake up almost all readers in wait queue When the front of the wait queue is a reader, other readers immediately following the first reader will also be woken up at the same time. However, if there is a writer in between. Those readers behind the writer will not be woken up. Because of optimistic spinning, the lock acquisition order is not FIFO anyway. The lock handoff mechanism will ensure that lock starvation will not happen. Assuming that the lock hold times of the other readers still in the queue will be about the same as the readers that are being woken up, there is really not much additional cost other than the additional latency due to the wakeup of additional tasks by the waker. Therefore all the readers up to a maximum of 256 in the queue are woken up when the first waiter is a reader to improve reader throughput. This is somewhat similar in concept to a phase-fair R/W lock. With a locking microbenchmark running on 5.1 based kernel, the total locking rates (in kops/s) on a 8-socket IvyBridge-EX system with equal numbers of readers and writers before and after this patch were as follows: # of Threads Pre-Patch Post-patch ------------ --------- ---------- 4 1,641 1,674 8 731 1,062 16 564 924 32 78 300 64 38 195 240 50 149 There is no performance gain at low contention level. At high contention level, however, this patch gives a pretty decent performance boost. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-11-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:09 +03:00
list_move_tail(&waiter->list, &wlist);
/*
* Limit # of readers that can be woken up per wakeup call.
*/
if (woken >= MAX_READERS_WAKEUP)
break;
}
adjustment = woken * RWSEM_READER_BIAS - adjustment;
lockevent_cond_inc(rwsem_wake_reader, woken);
if (list_empty(&sem->wait_list)) {
/* hit end of list above */
adjustment -= RWSEM_FLAG_WAITERS;
}
locking/rwsem: Implement lock handoff to prevent lock starvation Because of writer lock stealing, it is possible that a constant stream of incoming writers will cause a waiting writer or reader to wait indefinitely leading to lock starvation. This patch implements a lock handoff mechanism to disable lock stealing and force lock handoff to the first waiter or waiters (for readers) in the queue after at least a 4ms waiting period unless it is a RT writer task which doesn't need to wait. The waiting period is used to avoid discouraging lock stealing too much to affect performance. The setting and clearing of the handoff bit is serialized by the wait_lock. So racing is not possible. A rwsem microbenchmark was run for 5 seconds on a 2-socket 40-core 80-thread Skylake system with a v5.1 based kernel and 240 write_lock threads with 5us sleep critical section. Before the patch, the min/mean/max numbers of locking operations for the locking threads were 1/7,792/173,696. After the patch, the figures became 5,842/6,542/7,458. It can be seen that the rwsem became much more fair, though there was a drop of about 16% in the mean locking operations done which was a tradeoff of having better fairness. Making the waiter set the handoff bit right after the first wakeup can impact performance especially with a mixed reader/writer workload. With the same microbenchmark with short critical section and equal number of reader and writer threads (40/40), the reader/writer locking operation counts with the current patch were: 40 readers, Iterations Min/Mean/Max = 1,793/1,794/1,796 40 writers, Iterations Min/Mean/Max = 1,793/34,956/86,081 By making waiter set handoff bit immediately after wakeup: 40 readers, Iterations Min/Mean/Max = 43/44/46 40 writers, Iterations Min/Mean/Max = 43/1,263/3,191 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-8-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:06 +03:00
/*
* When we've woken a reader, we no longer need to force writers
* to give up the lock and we can clear HANDOFF.
*/
if (woken && (atomic_long_read(&sem->count) & RWSEM_FLAG_HANDOFF))
adjustment -= RWSEM_FLAG_HANDOFF;
if (adjustment)
atomic_long_add(adjustment, &sem->count);
/* 2nd pass */
list_for_each_entry_safe(waiter, tmp, &wlist, list) {
struct task_struct *tsk;
tsk = waiter->task;
get_task_struct(tsk);
/*
* Ensure calling get_task_struct() before setting the reader
locking/rwsem: Code cleanup after files merging After merging all the relevant rwsem code into one single file, there are a number of optimizations and cleanups that can be done: 1) Remove all the EXPORT_SYMBOL() calls for functions that are not accessed elsewhere. 2) Remove all the __visible tags as none of the functions will be called from assembly code anymore. 3) Make all the internal functions static. 4) Remove some unneeded blank lines. 5) Remove the intermediate rwsem_down_{read|write}_failed*() functions and rename __rwsem_down_{read|write}_failed_common() to rwsem_down_{read|write}_slowpath(). 6) Remove "__" prefix of __rwsem_mark_wake(). 7) Use atomic_long_try_cmpxchg_acquire() as much as possible. 8) Remove the rwsem_rtrylock and rwsem_wtrylock lock events as they are not that useful. That enables the compiler to do better optimization and reduce code size. The text+data size of rwsem.o on an x86-64 machine with gcc8 was reduced from 10237 bytes to 5030 bytes with this change. Suggested-by: Peter Zijlstra <peterz@infradead.org> Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-6-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:04 +03:00
* waiter to nil such that rwsem_down_read_slowpath() cannot
* race with do_exit() by always holding a reference count
* to the task to wakeup.
*/
smp_store_release(&waiter->task, NULL);
/*
* Ensure issuing the wakeup (either by us or someone else)
* after setting the reader waiter to nil.
*/
wake_q_add_safe(wake_q, tsk);
}
}
/*
* This function must be called with the sem->wait_lock held to prevent
* race conditions between checking the rwsem wait list and setting the
* sem->count accordingly.
locking/rwsem: Implement lock handoff to prevent lock starvation Because of writer lock stealing, it is possible that a constant stream of incoming writers will cause a waiting writer or reader to wait indefinitely leading to lock starvation. This patch implements a lock handoff mechanism to disable lock stealing and force lock handoff to the first waiter or waiters (for readers) in the queue after at least a 4ms waiting period unless it is a RT writer task which doesn't need to wait. The waiting period is used to avoid discouraging lock stealing too much to affect performance. The setting and clearing of the handoff bit is serialized by the wait_lock. So racing is not possible. A rwsem microbenchmark was run for 5 seconds on a 2-socket 40-core 80-thread Skylake system with a v5.1 based kernel and 240 write_lock threads with 5us sleep critical section. Before the patch, the min/mean/max numbers of locking operations for the locking threads were 1/7,792/173,696. After the patch, the figures became 5,842/6,542/7,458. It can be seen that the rwsem became much more fair, though there was a drop of about 16% in the mean locking operations done which was a tradeoff of having better fairness. Making the waiter set the handoff bit right after the first wakeup can impact performance especially with a mixed reader/writer workload. With the same microbenchmark with short critical section and equal number of reader and writer threads (40/40), the reader/writer locking operation counts with the current patch were: 40 readers, Iterations Min/Mean/Max = 1,793/1,794/1,796 40 writers, Iterations Min/Mean/Max = 1,793/34,956/86,081 By making waiter set handoff bit immediately after wakeup: 40 readers, Iterations Min/Mean/Max = 43/44/46 40 writers, Iterations Min/Mean/Max = 43/1,263/3,191 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-8-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:06 +03:00
*
* If wstate is WRITER_HANDOFF, it will make sure that either the handoff
* bit is set or the lock is acquired with handoff bit cleared.
*/
static inline bool rwsem_try_write_lock(struct rw_semaphore *sem,
locking/rwsem: Implement lock handoff to prevent lock starvation Because of writer lock stealing, it is possible that a constant stream of incoming writers will cause a waiting writer or reader to wait indefinitely leading to lock starvation. This patch implements a lock handoff mechanism to disable lock stealing and force lock handoff to the first waiter or waiters (for readers) in the queue after at least a 4ms waiting period unless it is a RT writer task which doesn't need to wait. The waiting period is used to avoid discouraging lock stealing too much to affect performance. The setting and clearing of the handoff bit is serialized by the wait_lock. So racing is not possible. A rwsem microbenchmark was run for 5 seconds on a 2-socket 40-core 80-thread Skylake system with a v5.1 based kernel and 240 write_lock threads with 5us sleep critical section. Before the patch, the min/mean/max numbers of locking operations for the locking threads were 1/7,792/173,696. After the patch, the figures became 5,842/6,542/7,458. It can be seen that the rwsem became much more fair, though there was a drop of about 16% in the mean locking operations done which was a tradeoff of having better fairness. Making the waiter set the handoff bit right after the first wakeup can impact performance especially with a mixed reader/writer workload. With the same microbenchmark with short critical section and equal number of reader and writer threads (40/40), the reader/writer locking operation counts with the current patch were: 40 readers, Iterations Min/Mean/Max = 1,793/1,794/1,796 40 writers, Iterations Min/Mean/Max = 1,793/34,956/86,081 By making waiter set handoff bit immediately after wakeup: 40 readers, Iterations Min/Mean/Max = 43/44/46 40 writers, Iterations Min/Mean/Max = 43/1,263/3,191 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-8-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:06 +03:00
enum writer_wait_state wstate)
{
long count, new;
locking/rwsem: Implement lock handoff to prevent lock starvation Because of writer lock stealing, it is possible that a constant stream of incoming writers will cause a waiting writer or reader to wait indefinitely leading to lock starvation. This patch implements a lock handoff mechanism to disable lock stealing and force lock handoff to the first waiter or waiters (for readers) in the queue after at least a 4ms waiting period unless it is a RT writer task which doesn't need to wait. The waiting period is used to avoid discouraging lock stealing too much to affect performance. The setting and clearing of the handoff bit is serialized by the wait_lock. So racing is not possible. A rwsem microbenchmark was run for 5 seconds on a 2-socket 40-core 80-thread Skylake system with a v5.1 based kernel and 240 write_lock threads with 5us sleep critical section. Before the patch, the min/mean/max numbers of locking operations for the locking threads were 1/7,792/173,696. After the patch, the figures became 5,842/6,542/7,458. It can be seen that the rwsem became much more fair, though there was a drop of about 16% in the mean locking operations done which was a tradeoff of having better fairness. Making the waiter set the handoff bit right after the first wakeup can impact performance especially with a mixed reader/writer workload. With the same microbenchmark with short critical section and equal number of reader and writer threads (40/40), the reader/writer locking operation counts with the current patch were: 40 readers, Iterations Min/Mean/Max = 1,793/1,794/1,796 40 writers, Iterations Min/Mean/Max = 1,793/34,956/86,081 By making waiter set handoff bit immediately after wakeup: 40 readers, Iterations Min/Mean/Max = 43/44/46 40 writers, Iterations Min/Mean/Max = 43/1,263/3,191 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-8-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:06 +03:00
lockdep_assert_held(&sem->wait_lock);
count = atomic_long_read(&sem->count);
locking/rwsem: Implement lock handoff to prevent lock starvation Because of writer lock stealing, it is possible that a constant stream of incoming writers will cause a waiting writer or reader to wait indefinitely leading to lock starvation. This patch implements a lock handoff mechanism to disable lock stealing and force lock handoff to the first waiter or waiters (for readers) in the queue after at least a 4ms waiting period unless it is a RT writer task which doesn't need to wait. The waiting period is used to avoid discouraging lock stealing too much to affect performance. The setting and clearing of the handoff bit is serialized by the wait_lock. So racing is not possible. A rwsem microbenchmark was run for 5 seconds on a 2-socket 40-core 80-thread Skylake system with a v5.1 based kernel and 240 write_lock threads with 5us sleep critical section. Before the patch, the min/mean/max numbers of locking operations for the locking threads were 1/7,792/173,696. After the patch, the figures became 5,842/6,542/7,458. It can be seen that the rwsem became much more fair, though there was a drop of about 16% in the mean locking operations done which was a tradeoff of having better fairness. Making the waiter set the handoff bit right after the first wakeup can impact performance especially with a mixed reader/writer workload. With the same microbenchmark with short critical section and equal number of reader and writer threads (40/40), the reader/writer locking operation counts with the current patch were: 40 readers, Iterations Min/Mean/Max = 1,793/1,794/1,796 40 writers, Iterations Min/Mean/Max = 1,793/34,956/86,081 By making waiter set handoff bit immediately after wakeup: 40 readers, Iterations Min/Mean/Max = 43/44/46 40 writers, Iterations Min/Mean/Max = 43/1,263/3,191 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-8-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:06 +03:00
do {
bool has_handoff = !!(count & RWSEM_FLAG_HANDOFF);
locking/rwsem: Implement lock handoff to prevent lock starvation Because of writer lock stealing, it is possible that a constant stream of incoming writers will cause a waiting writer or reader to wait indefinitely leading to lock starvation. This patch implements a lock handoff mechanism to disable lock stealing and force lock handoff to the first waiter or waiters (for readers) in the queue after at least a 4ms waiting period unless it is a RT writer task which doesn't need to wait. The waiting period is used to avoid discouraging lock stealing too much to affect performance. The setting and clearing of the handoff bit is serialized by the wait_lock. So racing is not possible. A rwsem microbenchmark was run for 5 seconds on a 2-socket 40-core 80-thread Skylake system with a v5.1 based kernel and 240 write_lock threads with 5us sleep critical section. Before the patch, the min/mean/max numbers of locking operations for the locking threads were 1/7,792/173,696. After the patch, the figures became 5,842/6,542/7,458. It can be seen that the rwsem became much more fair, though there was a drop of about 16% in the mean locking operations done which was a tradeoff of having better fairness. Making the waiter set the handoff bit right after the first wakeup can impact performance especially with a mixed reader/writer workload. With the same microbenchmark with short critical section and equal number of reader and writer threads (40/40), the reader/writer locking operation counts with the current patch were: 40 readers, Iterations Min/Mean/Max = 1,793/1,794/1,796 40 writers, Iterations Min/Mean/Max = 1,793/34,956/86,081 By making waiter set handoff bit immediately after wakeup: 40 readers, Iterations Min/Mean/Max = 43/44/46 40 writers, Iterations Min/Mean/Max = 43/1,263/3,191 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-8-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:06 +03:00
if (has_handoff && wstate == WRITER_NOT_FIRST)
return false;
locking/rwsem: Implement lock handoff to prevent lock starvation Because of writer lock stealing, it is possible that a constant stream of incoming writers will cause a waiting writer or reader to wait indefinitely leading to lock starvation. This patch implements a lock handoff mechanism to disable lock stealing and force lock handoff to the first waiter or waiters (for readers) in the queue after at least a 4ms waiting period unless it is a RT writer task which doesn't need to wait. The waiting period is used to avoid discouraging lock stealing too much to affect performance. The setting and clearing of the handoff bit is serialized by the wait_lock. So racing is not possible. A rwsem microbenchmark was run for 5 seconds on a 2-socket 40-core 80-thread Skylake system with a v5.1 based kernel and 240 write_lock threads with 5us sleep critical section. Before the patch, the min/mean/max numbers of locking operations for the locking threads were 1/7,792/173,696. After the patch, the figures became 5,842/6,542/7,458. It can be seen that the rwsem became much more fair, though there was a drop of about 16% in the mean locking operations done which was a tradeoff of having better fairness. Making the waiter set the handoff bit right after the first wakeup can impact performance especially with a mixed reader/writer workload. With the same microbenchmark with short critical section and equal number of reader and writer threads (40/40), the reader/writer locking operation counts with the current patch were: 40 readers, Iterations Min/Mean/Max = 1,793/1,794/1,796 40 writers, Iterations Min/Mean/Max = 1,793/34,956/86,081 By making waiter set handoff bit immediately after wakeup: 40 readers, Iterations Min/Mean/Max = 43/44/46 40 writers, Iterations Min/Mean/Max = 43/1,263/3,191 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-8-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:06 +03:00
new = count;
if (count & RWSEM_LOCK_MASK) {
if (has_handoff || (wstate != WRITER_HANDOFF))
return false;
new |= RWSEM_FLAG_HANDOFF;
} else {
new |= RWSEM_WRITER_LOCKED;
new &= ~RWSEM_FLAG_HANDOFF;
if (list_is_singular(&sem->wait_list))
new &= ~RWSEM_FLAG_WAITERS;
}
} while (!atomic_long_try_cmpxchg_acquire(&sem->count, &count, new));
/*
* We have either acquired the lock with handoff bit cleared or
* set the handoff bit.
*/
if (new & RWSEM_FLAG_HANDOFF)
return false;
rwsem_set_owner(sem);
return true;
}
#ifdef CONFIG_RWSEM_SPIN_ON_OWNER
/*
* Try to acquire write lock before the writer has been put on wait queue.
*/
static inline bool rwsem_try_write_lock_unqueued(struct rw_semaphore *sem)
{
long count = atomic_long_read(&sem->count);
locking/rwsem: Implement lock handoff to prevent lock starvation Because of writer lock stealing, it is possible that a constant stream of incoming writers will cause a waiting writer or reader to wait indefinitely leading to lock starvation. This patch implements a lock handoff mechanism to disable lock stealing and force lock handoff to the first waiter or waiters (for readers) in the queue after at least a 4ms waiting period unless it is a RT writer task which doesn't need to wait. The waiting period is used to avoid discouraging lock stealing too much to affect performance. The setting and clearing of the handoff bit is serialized by the wait_lock. So racing is not possible. A rwsem microbenchmark was run for 5 seconds on a 2-socket 40-core 80-thread Skylake system with a v5.1 based kernel and 240 write_lock threads with 5us sleep critical section. Before the patch, the min/mean/max numbers of locking operations for the locking threads were 1/7,792/173,696. After the patch, the figures became 5,842/6,542/7,458. It can be seen that the rwsem became much more fair, though there was a drop of about 16% in the mean locking operations done which was a tradeoff of having better fairness. Making the waiter set the handoff bit right after the first wakeup can impact performance especially with a mixed reader/writer workload. With the same microbenchmark with short critical section and equal number of reader and writer threads (40/40), the reader/writer locking operation counts with the current patch were: 40 readers, Iterations Min/Mean/Max = 1,793/1,794/1,796 40 writers, Iterations Min/Mean/Max = 1,793/34,956/86,081 By making waiter set handoff bit immediately after wakeup: 40 readers, Iterations Min/Mean/Max = 43/44/46 40 writers, Iterations Min/Mean/Max = 43/1,263/3,191 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-8-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:06 +03:00
while (!(count & (RWSEM_LOCK_MASK|RWSEM_FLAG_HANDOFF))) {
if (atomic_long_try_cmpxchg_acquire(&sem->count, &count,
locking/rwsem: Implement lock handoff to prevent lock starvation Because of writer lock stealing, it is possible that a constant stream of incoming writers will cause a waiting writer or reader to wait indefinitely leading to lock starvation. This patch implements a lock handoff mechanism to disable lock stealing and force lock handoff to the first waiter or waiters (for readers) in the queue after at least a 4ms waiting period unless it is a RT writer task which doesn't need to wait. The waiting period is used to avoid discouraging lock stealing too much to affect performance. The setting and clearing of the handoff bit is serialized by the wait_lock. So racing is not possible. A rwsem microbenchmark was run for 5 seconds on a 2-socket 40-core 80-thread Skylake system with a v5.1 based kernel and 240 write_lock threads with 5us sleep critical section. Before the patch, the min/mean/max numbers of locking operations for the locking threads were 1/7,792/173,696. After the patch, the figures became 5,842/6,542/7,458. It can be seen that the rwsem became much more fair, though there was a drop of about 16% in the mean locking operations done which was a tradeoff of having better fairness. Making the waiter set the handoff bit right after the first wakeup can impact performance especially with a mixed reader/writer workload. With the same microbenchmark with short critical section and equal number of reader and writer threads (40/40), the reader/writer locking operation counts with the current patch were: 40 readers, Iterations Min/Mean/Max = 1,793/1,794/1,796 40 writers, Iterations Min/Mean/Max = 1,793/34,956/86,081 By making waiter set handoff bit immediately after wakeup: 40 readers, Iterations Min/Mean/Max = 43/44/46 40 writers, Iterations Min/Mean/Max = 43/1,263/3,191 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-8-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:06 +03:00
count | RWSEM_WRITER_LOCKED)) {
rwsem_set_owner(sem);
locking/rwsem: Remove reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It also improves the chance that a reader can get the lock as writer optimistic spinning disproportionally favors writers much more than readers. Since commit d3681e269fff ("locking/rwsem: Wake up almost all readers in wait queue"), all the waiting readers are woken up so that they can all get the read lock and run in parallel. When the number of contending readers is large, allowing reader optimistic spinning will likely cause reader fragmentation where multiple smaller groups of readers can get the read lock in a sequential manner separated by writers. That reduces reader parallelism. One possible way to address that drawback is to limit the number of readers (preferably one) that can do optimistic spinning. These readers act as representatives of all the waiting readers in the wait queue as they will wake up all those waiting readers once they get the lock. Alternatively, as reader optimistic lock stealing has already enhanced fairness to readers, it may be easier to just remove reader optimistic spinning and simplifying the optimistic spinning code as a result. Performance measurements (locking throughput kops/s) using a locking microbenchmark with 50/50 reader/writer distribution and turbo-boost disabled was done on a 2-socket Cascade Lake system (48-core 96-thread) to see the impacts of these changes: 1) Vanilla - 5.10-rc3 kernel 2) Before - 5.10-rc3 kernel with previous patches in this series 2) limit-rspin - 5.10-rc3 kernel with limited reader spinning patch 3) no-rspin - 5.10-rc3 kernel with reader spinning disabled # of threads CS Load Vanilla Before limit-rspin no-rspin ------------ ------- ------- ------ ----------- -------- 2 1 5,185 5,662 5,214 5,077 4 1 5,107 4,983 5,188 4,760 8 1 4,782 4,564 4,720 4,628 16 1 4,680 4,053 4,567 3,402 32 1 4,299 1,115 1,118 1,098 64 1 3,218 983 1,001 957 96 1 1,938 944 957 930 2 20 2,008 2,128 2,264 1,665 4 20 1,390 1,033 1,046 1,101 8 20 1,472 1,155 1,098 1,213 16 20 1,332 1,077 1,089 1,122 32 20 967 914 917 980 64 20 787 874 891 858 96 20 730 836 847 844 2 100 372 356 360 355 4 100 492 425 434 392 8 100 533 537 529 538 16 100 548 572 568 598 32 100 499 520 527 537 64 100 466 517 526 512 96 100 406 497 506 509 The column "CS Load" represents the number of pause instructions issued in the locking critical section. A CS load of 1 is extremely short and is not likey in real situations. A load of 20 (moderate) and 100 (long) are more realistic. It can be seen that the previous patches in this series have reduced performance in general except in highly contended cases with moderate or long critical sections that performance improves a bit. This change is mostly caused by the "Prevent potential lock starvation" patch that reduce reader optimistic spinning and hence reduce reader fragmentation. The patch that further limit reader optimistic spinning doesn't seem to have too much impact on overall performance as shown in the benchmark data. The patch that disables reader optimistic spinning shows reduced performance at lightly loaded cases, but comparable or slightly better performance on with heavier contention. This patch just removes reader optimistic spinning for now. As readers are not going to do optimistic spinning anymore, we don't need to consider if the OSQ is empty or not when doing lock stealing. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-6-longman@redhat.com
2020-11-21 07:14:16 +03:00
lockevent_inc(rwsem_opt_lock);
return true;
}
}
return false;
}
static inline bool owner_on_cpu(struct task_struct *owner)
{
/*
* As lock holder preemption issue, we both skip spinning if
* task is not on cpu or its cpu is preempted
*/
return owner->on_cpu && !vcpu_is_preempted(task_cpu(owner));
}
locking/rwsem: Remove reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It also improves the chance that a reader can get the lock as writer optimistic spinning disproportionally favors writers much more than readers. Since commit d3681e269fff ("locking/rwsem: Wake up almost all readers in wait queue"), all the waiting readers are woken up so that they can all get the read lock and run in parallel. When the number of contending readers is large, allowing reader optimistic spinning will likely cause reader fragmentation where multiple smaller groups of readers can get the read lock in a sequential manner separated by writers. That reduces reader parallelism. One possible way to address that drawback is to limit the number of readers (preferably one) that can do optimistic spinning. These readers act as representatives of all the waiting readers in the wait queue as they will wake up all those waiting readers once they get the lock. Alternatively, as reader optimistic lock stealing has already enhanced fairness to readers, it may be easier to just remove reader optimistic spinning and simplifying the optimistic spinning code as a result. Performance measurements (locking throughput kops/s) using a locking microbenchmark with 50/50 reader/writer distribution and turbo-boost disabled was done on a 2-socket Cascade Lake system (48-core 96-thread) to see the impacts of these changes: 1) Vanilla - 5.10-rc3 kernel 2) Before - 5.10-rc3 kernel with previous patches in this series 2) limit-rspin - 5.10-rc3 kernel with limited reader spinning patch 3) no-rspin - 5.10-rc3 kernel with reader spinning disabled # of threads CS Load Vanilla Before limit-rspin no-rspin ------------ ------- ------- ------ ----------- -------- 2 1 5,185 5,662 5,214 5,077 4 1 5,107 4,983 5,188 4,760 8 1 4,782 4,564 4,720 4,628 16 1 4,680 4,053 4,567 3,402 32 1 4,299 1,115 1,118 1,098 64 1 3,218 983 1,001 957 96 1 1,938 944 957 930 2 20 2,008 2,128 2,264 1,665 4 20 1,390 1,033 1,046 1,101 8 20 1,472 1,155 1,098 1,213 16 20 1,332 1,077 1,089 1,122 32 20 967 914 917 980 64 20 787 874 891 858 96 20 730 836 847 844 2 100 372 356 360 355 4 100 492 425 434 392 8 100 533 537 529 538 16 100 548 572 568 598 32 100 499 520 527 537 64 100 466 517 526 512 96 100 406 497 506 509 The column "CS Load" represents the number of pause instructions issued in the locking critical section. A CS load of 1 is extremely short and is not likey in real situations. A load of 20 (moderate) and 100 (long) are more realistic. It can be seen that the previous patches in this series have reduced performance in general except in highly contended cases with moderate or long critical sections that performance improves a bit. This change is mostly caused by the "Prevent potential lock starvation" patch that reduce reader optimistic spinning and hence reduce reader fragmentation. The patch that further limit reader optimistic spinning doesn't seem to have too much impact on overall performance as shown in the benchmark data. The patch that disables reader optimistic spinning shows reduced performance at lightly loaded cases, but comparable or slightly better performance on with heavier contention. This patch just removes reader optimistic spinning for now. As readers are not going to do optimistic spinning anymore, we don't need to consider if the OSQ is empty or not when doing lock stealing. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-6-longman@redhat.com
2020-11-21 07:14:16 +03:00
static inline bool rwsem_can_spin_on_owner(struct rw_semaphore *sem)
{
struct task_struct *owner;
unsigned long flags;
bool ret = true;
if (need_resched()) {
lockevent_inc(rwsem_opt_fail);
return false;
}
preempt_disable();
rcu_read_lock();
owner = rwsem_owner_flags(sem, &flags);
locking/rwsem: Don't call owner_on_cpu() on read-owner For writer, the owner value is cleared on unlock. For reader, it is left intact on unlock for providing better debugging aid on crash dump and the unlock of one reader may not mean the lock is free. As a result, the owner_on_cpu() shouldn't be used on read-owner as the task pointer value may not be valid and it might have been freed. That is the case in rwsem_spin_on_owner(), but not in rwsem_can_spin_on_owner(). This can lead to use-after-free error from KASAN. For example, BUG: KASAN: use-after-free in rwsem_down_write_slowpath (/home/miguel/kernel/linux/kernel/locking/rwsem.c:669 /home/miguel/kernel/linux/kernel/locking/rwsem.c:1125) Fix this by checking for RWSEM_READER_OWNED flag before calling owner_on_cpu(). Reported-by: Luis Henriques <lhenriques@suse.com> Tested-by: Luis Henriques <lhenriques@suse.com> Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jeff Layton <jlayton@kernel.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Fixes: 94a9717b3c40e ("locking/rwsem: Make rwsem->owner an atomic_long_t") Link: https://lkml.kernel.org/r/81e82d5b-5074-77e8-7204-28479bbe0df0@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-07-20 18:04:10 +03:00
/*
* Don't check the read-owner as the entry may be stale.
*/
locking/rwsem: Remove reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It also improves the chance that a reader can get the lock as writer optimistic spinning disproportionally favors writers much more than readers. Since commit d3681e269fff ("locking/rwsem: Wake up almost all readers in wait queue"), all the waiting readers are woken up so that they can all get the read lock and run in parallel. When the number of contending readers is large, allowing reader optimistic spinning will likely cause reader fragmentation where multiple smaller groups of readers can get the read lock in a sequential manner separated by writers. That reduces reader parallelism. One possible way to address that drawback is to limit the number of readers (preferably one) that can do optimistic spinning. These readers act as representatives of all the waiting readers in the wait queue as they will wake up all those waiting readers once they get the lock. Alternatively, as reader optimistic lock stealing has already enhanced fairness to readers, it may be easier to just remove reader optimistic spinning and simplifying the optimistic spinning code as a result. Performance measurements (locking throughput kops/s) using a locking microbenchmark with 50/50 reader/writer distribution and turbo-boost disabled was done on a 2-socket Cascade Lake system (48-core 96-thread) to see the impacts of these changes: 1) Vanilla - 5.10-rc3 kernel 2) Before - 5.10-rc3 kernel with previous patches in this series 2) limit-rspin - 5.10-rc3 kernel with limited reader spinning patch 3) no-rspin - 5.10-rc3 kernel with reader spinning disabled # of threads CS Load Vanilla Before limit-rspin no-rspin ------------ ------- ------- ------ ----------- -------- 2 1 5,185 5,662 5,214 5,077 4 1 5,107 4,983 5,188 4,760 8 1 4,782 4,564 4,720 4,628 16 1 4,680 4,053 4,567 3,402 32 1 4,299 1,115 1,118 1,098 64 1 3,218 983 1,001 957 96 1 1,938 944 957 930 2 20 2,008 2,128 2,264 1,665 4 20 1,390 1,033 1,046 1,101 8 20 1,472 1,155 1,098 1,213 16 20 1,332 1,077 1,089 1,122 32 20 967 914 917 980 64 20 787 874 891 858 96 20 730 836 847 844 2 100 372 356 360 355 4 100 492 425 434 392 8 100 533 537 529 538 16 100 548 572 568 598 32 100 499 520 527 537 64 100 466 517 526 512 96 100 406 497 506 509 The column "CS Load" represents the number of pause instructions issued in the locking critical section. A CS load of 1 is extremely short and is not likey in real situations. A load of 20 (moderate) and 100 (long) are more realistic. It can be seen that the previous patches in this series have reduced performance in general except in highly contended cases with moderate or long critical sections that performance improves a bit. This change is mostly caused by the "Prevent potential lock starvation" patch that reduce reader optimistic spinning and hence reduce reader fragmentation. The patch that further limit reader optimistic spinning doesn't seem to have too much impact on overall performance as shown in the benchmark data. The patch that disables reader optimistic spinning shows reduced performance at lightly loaded cases, but comparable or slightly better performance on with heavier contention. This patch just removes reader optimistic spinning for now. As readers are not going to do optimistic spinning anymore, we don't need to consider if the OSQ is empty or not when doing lock stealing. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-6-longman@redhat.com
2020-11-21 07:14:16 +03:00
if ((flags & RWSEM_NONSPINNABLE) ||
locking/rwsem: Don't call owner_on_cpu() on read-owner For writer, the owner value is cleared on unlock. For reader, it is left intact on unlock for providing better debugging aid on crash dump and the unlock of one reader may not mean the lock is free. As a result, the owner_on_cpu() shouldn't be used on read-owner as the task pointer value may not be valid and it might have been freed. That is the case in rwsem_spin_on_owner(), but not in rwsem_can_spin_on_owner(). This can lead to use-after-free error from KASAN. For example, BUG: KASAN: use-after-free in rwsem_down_write_slowpath (/home/miguel/kernel/linux/kernel/locking/rwsem.c:669 /home/miguel/kernel/linux/kernel/locking/rwsem.c:1125) Fix this by checking for RWSEM_READER_OWNED flag before calling owner_on_cpu(). Reported-by: Luis Henriques <lhenriques@suse.com> Tested-by: Luis Henriques <lhenriques@suse.com> Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jeff Layton <jlayton@kernel.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Fixes: 94a9717b3c40e ("locking/rwsem: Make rwsem->owner an atomic_long_t") Link: https://lkml.kernel.org/r/81e82d5b-5074-77e8-7204-28479bbe0df0@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-07-20 18:04:10 +03:00
(owner && !(flags & RWSEM_READER_OWNED) && !owner_on_cpu(owner)))
ret = false;
rcu_read_unlock();
preempt_enable();
lockevent_cond_inc(rwsem_opt_fail, !ret);
return ret;
}
/*
* The rwsem_spin_on_owner() function returns the following 4 values
* depending on the lock owner state.
* OWNER_NULL : owner is currently NULL
* OWNER_WRITER: when owner changes and is a writer
* OWNER_READER: when owner changes and the new owner may be a reader.
* OWNER_NONSPINNABLE:
* when optimistic spinning has to stop because either the
* owner stops running, is unknown, or its timeslice has
* been used up.
*/
enum owner_state {
OWNER_NULL = 1 << 0,
OWNER_WRITER = 1 << 1,
OWNER_READER = 1 << 2,
OWNER_NONSPINNABLE = 1 << 3,
};
locking/rwsem: Enable time-based spinning on reader-owned rwsem When the rwsem is owned by reader, writers stop optimistic spinning simply because there is no easy way to figure out if all the readers are actively running or not. However, there are scenarios where the readers are unlikely to sleep and optimistic spinning can help performance. This patch provides a simple mechanism for spinning on a reader-owned rwsem by a writer. It is a time threshold based spinning where the allowable spinning time can vary from 10us to 25us depending on the condition of the rwsem. When the time threshold is exceeded, the nonspinnable bits will be set in the owner field to indicate that no more optimistic spinning will be allowed on this rwsem until it becomes writer owned again. Not even readers is allowed to acquire the reader-locked rwsem by optimistic spinning for fairness. We also want a writer to acquire the lock after the readers hold the lock for a relatively long time. In order to give preference to writers under such a circumstance, the single RWSEM_NONSPINNABLE bit is now split into two - one for reader and one for writer. When optimistic spinning is disabled, both bits will be set. When the reader count drop down to 0, the writer nonspinnable bit will be cleared to allow writers to spin on the lock, but not the readers. When a writer acquires the lock, it will write its own task structure pointer into sem->owner and clear the reader nonspinnable bit in the process. The time taken for each iteration of the reader-owned rwsem spinning loop varies. Below are sample minimum elapsed times for 16 iterations of the loop. System Time for 16 Iterations ------ ---------------------- 1-socket Skylake ~800ns 4-socket Broadwell ~300ns 2-socket ThunderX2 (arm64) ~250ns When the lock cacheline is contended, we can see up to almost 10X increase in elapsed time. So 25us will be at most 500, 1300 and 1600 iterations for each of the above systems. With a locking microbenchmark running on 5.1 based kernel, the total locking rates (in kops/s) on a 8-socket IvyBridge-EX system with equal numbers of readers and writers before and after this patch were as follows: # of Threads Pre-patch Post-patch ------------ --------- ---------- 2 1,759 6,684 4 1,684 6,738 8 1,074 7,222 16 900 7,163 32 458 7,316 64 208 520 128 168 425 240 143 474 This patch gives a big boost in performance for mixed reader/writer workloads. With 32 locking threads, the rwsem lock event data were: rwsem_opt_fail=79850 rwsem_opt_nospin=5069 rwsem_opt_rlock=597484 rwsem_opt_wlock=957339 rwsem_sleep_reader=57782 rwsem_sleep_writer=55663 With 64 locking threads, the data looked like: rwsem_opt_fail=346723 rwsem_opt_nospin=6293 rwsem_opt_rlock=1127119 rwsem_opt_wlock=1400628 rwsem_sleep_reader=308201 rwsem_sleep_writer=72281 So a lot more threads acquired the lock in the slowpath and more threads went to sleep. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-15-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:13 +03:00
#define OWNER_SPINNABLE (OWNER_NULL | OWNER_WRITER | OWNER_READER)
static inline enum owner_state
locking/rwsem: Remove reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It also improves the chance that a reader can get the lock as writer optimistic spinning disproportionally favors writers much more than readers. Since commit d3681e269fff ("locking/rwsem: Wake up almost all readers in wait queue"), all the waiting readers are woken up so that they can all get the read lock and run in parallel. When the number of contending readers is large, allowing reader optimistic spinning will likely cause reader fragmentation where multiple smaller groups of readers can get the read lock in a sequential manner separated by writers. That reduces reader parallelism. One possible way to address that drawback is to limit the number of readers (preferably one) that can do optimistic spinning. These readers act as representatives of all the waiting readers in the wait queue as they will wake up all those waiting readers once they get the lock. Alternatively, as reader optimistic lock stealing has already enhanced fairness to readers, it may be easier to just remove reader optimistic spinning and simplifying the optimistic spinning code as a result. Performance measurements (locking throughput kops/s) using a locking microbenchmark with 50/50 reader/writer distribution and turbo-boost disabled was done on a 2-socket Cascade Lake system (48-core 96-thread) to see the impacts of these changes: 1) Vanilla - 5.10-rc3 kernel 2) Before - 5.10-rc3 kernel with previous patches in this series 2) limit-rspin - 5.10-rc3 kernel with limited reader spinning patch 3) no-rspin - 5.10-rc3 kernel with reader spinning disabled # of threads CS Load Vanilla Before limit-rspin no-rspin ------------ ------- ------- ------ ----------- -------- 2 1 5,185 5,662 5,214 5,077 4 1 5,107 4,983 5,188 4,760 8 1 4,782 4,564 4,720 4,628 16 1 4,680 4,053 4,567 3,402 32 1 4,299 1,115 1,118 1,098 64 1 3,218 983 1,001 957 96 1 1,938 944 957 930 2 20 2,008 2,128 2,264 1,665 4 20 1,390 1,033 1,046 1,101 8 20 1,472 1,155 1,098 1,213 16 20 1,332 1,077 1,089 1,122 32 20 967 914 917 980 64 20 787 874 891 858 96 20 730 836 847 844 2 100 372 356 360 355 4 100 492 425 434 392 8 100 533 537 529 538 16 100 548 572 568 598 32 100 499 520 527 537 64 100 466 517 526 512 96 100 406 497 506 509 The column "CS Load" represents the number of pause instructions issued in the locking critical section. A CS load of 1 is extremely short and is not likey in real situations. A load of 20 (moderate) and 100 (long) are more realistic. It can be seen that the previous patches in this series have reduced performance in general except in highly contended cases with moderate or long critical sections that performance improves a bit. This change is mostly caused by the "Prevent potential lock starvation" patch that reduce reader optimistic spinning and hence reduce reader fragmentation. The patch that further limit reader optimistic spinning doesn't seem to have too much impact on overall performance as shown in the benchmark data. The patch that disables reader optimistic spinning shows reduced performance at lightly loaded cases, but comparable or slightly better performance on with heavier contention. This patch just removes reader optimistic spinning for now. As readers are not going to do optimistic spinning anymore, we don't need to consider if the OSQ is empty or not when doing lock stealing. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-6-longman@redhat.com
2020-11-21 07:14:16 +03:00
rwsem_owner_state(struct task_struct *owner, unsigned long flags)
{
locking/rwsem: Remove reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It also improves the chance that a reader can get the lock as writer optimistic spinning disproportionally favors writers much more than readers. Since commit d3681e269fff ("locking/rwsem: Wake up almost all readers in wait queue"), all the waiting readers are woken up so that they can all get the read lock and run in parallel. When the number of contending readers is large, allowing reader optimistic spinning will likely cause reader fragmentation where multiple smaller groups of readers can get the read lock in a sequential manner separated by writers. That reduces reader parallelism. One possible way to address that drawback is to limit the number of readers (preferably one) that can do optimistic spinning. These readers act as representatives of all the waiting readers in the wait queue as they will wake up all those waiting readers once they get the lock. Alternatively, as reader optimistic lock stealing has already enhanced fairness to readers, it may be easier to just remove reader optimistic spinning and simplifying the optimistic spinning code as a result. Performance measurements (locking throughput kops/s) using a locking microbenchmark with 50/50 reader/writer distribution and turbo-boost disabled was done on a 2-socket Cascade Lake system (48-core 96-thread) to see the impacts of these changes: 1) Vanilla - 5.10-rc3 kernel 2) Before - 5.10-rc3 kernel with previous patches in this series 2) limit-rspin - 5.10-rc3 kernel with limited reader spinning patch 3) no-rspin - 5.10-rc3 kernel with reader spinning disabled # of threads CS Load Vanilla Before limit-rspin no-rspin ------------ ------- ------- ------ ----------- -------- 2 1 5,185 5,662 5,214 5,077 4 1 5,107 4,983 5,188 4,760 8 1 4,782 4,564 4,720 4,628 16 1 4,680 4,053 4,567 3,402 32 1 4,299 1,115 1,118 1,098 64 1 3,218 983 1,001 957 96 1 1,938 944 957 930 2 20 2,008 2,128 2,264 1,665 4 20 1,390 1,033 1,046 1,101 8 20 1,472 1,155 1,098 1,213 16 20 1,332 1,077 1,089 1,122 32 20 967 914 917 980 64 20 787 874 891 858 96 20 730 836 847 844 2 100 372 356 360 355 4 100 492 425 434 392 8 100 533 537 529 538 16 100 548 572 568 598 32 100 499 520 527 537 64 100 466 517 526 512 96 100 406 497 506 509 The column "CS Load" represents the number of pause instructions issued in the locking critical section. A CS load of 1 is extremely short and is not likey in real situations. A load of 20 (moderate) and 100 (long) are more realistic. It can be seen that the previous patches in this series have reduced performance in general except in highly contended cases with moderate or long critical sections that performance improves a bit. This change is mostly caused by the "Prevent potential lock starvation" patch that reduce reader optimistic spinning and hence reduce reader fragmentation. The patch that further limit reader optimistic spinning doesn't seem to have too much impact on overall performance as shown in the benchmark data. The patch that disables reader optimistic spinning shows reduced performance at lightly loaded cases, but comparable or slightly better performance on with heavier contention. This patch just removes reader optimistic spinning for now. As readers are not going to do optimistic spinning anymore, we don't need to consider if the OSQ is empty or not when doing lock stealing. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-6-longman@redhat.com
2020-11-21 07:14:16 +03:00
if (flags & RWSEM_NONSPINNABLE)
return OWNER_NONSPINNABLE;
if (flags & RWSEM_READER_OWNED)
return OWNER_READER;
return owner ? OWNER_WRITER : OWNER_NULL;
}
locking/rwsem: Enable time-based spinning on reader-owned rwsem When the rwsem is owned by reader, writers stop optimistic spinning simply because there is no easy way to figure out if all the readers are actively running or not. However, there are scenarios where the readers are unlikely to sleep and optimistic spinning can help performance. This patch provides a simple mechanism for spinning on a reader-owned rwsem by a writer. It is a time threshold based spinning where the allowable spinning time can vary from 10us to 25us depending on the condition of the rwsem. When the time threshold is exceeded, the nonspinnable bits will be set in the owner field to indicate that no more optimistic spinning will be allowed on this rwsem until it becomes writer owned again. Not even readers is allowed to acquire the reader-locked rwsem by optimistic spinning for fairness. We also want a writer to acquire the lock after the readers hold the lock for a relatively long time. In order to give preference to writers under such a circumstance, the single RWSEM_NONSPINNABLE bit is now split into two - one for reader and one for writer. When optimistic spinning is disabled, both bits will be set. When the reader count drop down to 0, the writer nonspinnable bit will be cleared to allow writers to spin on the lock, but not the readers. When a writer acquires the lock, it will write its own task structure pointer into sem->owner and clear the reader nonspinnable bit in the process. The time taken for each iteration of the reader-owned rwsem spinning loop varies. Below are sample minimum elapsed times for 16 iterations of the loop. System Time for 16 Iterations ------ ---------------------- 1-socket Skylake ~800ns 4-socket Broadwell ~300ns 2-socket ThunderX2 (arm64) ~250ns When the lock cacheline is contended, we can see up to almost 10X increase in elapsed time. So 25us will be at most 500, 1300 and 1600 iterations for each of the above systems. With a locking microbenchmark running on 5.1 based kernel, the total locking rates (in kops/s) on a 8-socket IvyBridge-EX system with equal numbers of readers and writers before and after this patch were as follows: # of Threads Pre-patch Post-patch ------------ --------- ---------- 2 1,759 6,684 4 1,684 6,738 8 1,074 7,222 16 900 7,163 32 458 7,316 64 208 520 128 168 425 240 143 474 This patch gives a big boost in performance for mixed reader/writer workloads. With 32 locking threads, the rwsem lock event data were: rwsem_opt_fail=79850 rwsem_opt_nospin=5069 rwsem_opt_rlock=597484 rwsem_opt_wlock=957339 rwsem_sleep_reader=57782 rwsem_sleep_writer=55663 With 64 locking threads, the data looked like: rwsem_opt_fail=346723 rwsem_opt_nospin=6293 rwsem_opt_rlock=1127119 rwsem_opt_wlock=1400628 rwsem_sleep_reader=308201 rwsem_sleep_writer=72281 So a lot more threads acquired the lock in the slowpath and more threads went to sleep. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-15-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:13 +03:00
static noinline enum owner_state
locking/rwsem: Remove reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It also improves the chance that a reader can get the lock as writer optimistic spinning disproportionally favors writers much more than readers. Since commit d3681e269fff ("locking/rwsem: Wake up almost all readers in wait queue"), all the waiting readers are woken up so that they can all get the read lock and run in parallel. When the number of contending readers is large, allowing reader optimistic spinning will likely cause reader fragmentation where multiple smaller groups of readers can get the read lock in a sequential manner separated by writers. That reduces reader parallelism. One possible way to address that drawback is to limit the number of readers (preferably one) that can do optimistic spinning. These readers act as representatives of all the waiting readers in the wait queue as they will wake up all those waiting readers once they get the lock. Alternatively, as reader optimistic lock stealing has already enhanced fairness to readers, it may be easier to just remove reader optimistic spinning and simplifying the optimistic spinning code as a result. Performance measurements (locking throughput kops/s) using a locking microbenchmark with 50/50 reader/writer distribution and turbo-boost disabled was done on a 2-socket Cascade Lake system (48-core 96-thread) to see the impacts of these changes: 1) Vanilla - 5.10-rc3 kernel 2) Before - 5.10-rc3 kernel with previous patches in this series 2) limit-rspin - 5.10-rc3 kernel with limited reader spinning patch 3) no-rspin - 5.10-rc3 kernel with reader spinning disabled # of threads CS Load Vanilla Before limit-rspin no-rspin ------------ ------- ------- ------ ----------- -------- 2 1 5,185 5,662 5,214 5,077 4 1 5,107 4,983 5,188 4,760 8 1 4,782 4,564 4,720 4,628 16 1 4,680 4,053 4,567 3,402 32 1 4,299 1,115 1,118 1,098 64 1 3,218 983 1,001 957 96 1 1,938 944 957 930 2 20 2,008 2,128 2,264 1,665 4 20 1,390 1,033 1,046 1,101 8 20 1,472 1,155 1,098 1,213 16 20 1,332 1,077 1,089 1,122 32 20 967 914 917 980 64 20 787 874 891 858 96 20 730 836 847 844 2 100 372 356 360 355 4 100 492 425 434 392 8 100 533 537 529 538 16 100 548 572 568 598 32 100 499 520 527 537 64 100 466 517 526 512 96 100 406 497 506 509 The column "CS Load" represents the number of pause instructions issued in the locking critical section. A CS load of 1 is extremely short and is not likey in real situations. A load of 20 (moderate) and 100 (long) are more realistic. It can be seen that the previous patches in this series have reduced performance in general except in highly contended cases with moderate or long critical sections that performance improves a bit. This change is mostly caused by the "Prevent potential lock starvation" patch that reduce reader optimistic spinning and hence reduce reader fragmentation. The patch that further limit reader optimistic spinning doesn't seem to have too much impact on overall performance as shown in the benchmark data. The patch that disables reader optimistic spinning shows reduced performance at lightly loaded cases, but comparable or slightly better performance on with heavier contention. This patch just removes reader optimistic spinning for now. As readers are not going to do optimistic spinning anymore, we don't need to consider if the OSQ is empty or not when doing lock stealing. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-6-longman@redhat.com
2020-11-21 07:14:16 +03:00
rwsem_spin_on_owner(struct rw_semaphore *sem)
{
struct task_struct *new, *owner;
unsigned long flags, new_flags;
enum owner_state state;
owner = rwsem_owner_flags(sem, &flags);
locking/rwsem: Remove reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It also improves the chance that a reader can get the lock as writer optimistic spinning disproportionally favors writers much more than readers. Since commit d3681e269fff ("locking/rwsem: Wake up almost all readers in wait queue"), all the waiting readers are woken up so that they can all get the read lock and run in parallel. When the number of contending readers is large, allowing reader optimistic spinning will likely cause reader fragmentation where multiple smaller groups of readers can get the read lock in a sequential manner separated by writers. That reduces reader parallelism. One possible way to address that drawback is to limit the number of readers (preferably one) that can do optimistic spinning. These readers act as representatives of all the waiting readers in the wait queue as they will wake up all those waiting readers once they get the lock. Alternatively, as reader optimistic lock stealing has already enhanced fairness to readers, it may be easier to just remove reader optimistic spinning and simplifying the optimistic spinning code as a result. Performance measurements (locking throughput kops/s) using a locking microbenchmark with 50/50 reader/writer distribution and turbo-boost disabled was done on a 2-socket Cascade Lake system (48-core 96-thread) to see the impacts of these changes: 1) Vanilla - 5.10-rc3 kernel 2) Before - 5.10-rc3 kernel with previous patches in this series 2) limit-rspin - 5.10-rc3 kernel with limited reader spinning patch 3) no-rspin - 5.10-rc3 kernel with reader spinning disabled # of threads CS Load Vanilla Before limit-rspin no-rspin ------------ ------- ------- ------ ----------- -------- 2 1 5,185 5,662 5,214 5,077 4 1 5,107 4,983 5,188 4,760 8 1 4,782 4,564 4,720 4,628 16 1 4,680 4,053 4,567 3,402 32 1 4,299 1,115 1,118 1,098 64 1 3,218 983 1,001 957 96 1 1,938 944 957 930 2 20 2,008 2,128 2,264 1,665 4 20 1,390 1,033 1,046 1,101 8 20 1,472 1,155 1,098 1,213 16 20 1,332 1,077 1,089 1,122 32 20 967 914 917 980 64 20 787 874 891 858 96 20 730 836 847 844 2 100 372 356 360 355 4 100 492 425 434 392 8 100 533 537 529 538 16 100 548 572 568 598 32 100 499 520 527 537 64 100 466 517 526 512 96 100 406 497 506 509 The column "CS Load" represents the number of pause instructions issued in the locking critical section. A CS load of 1 is extremely short and is not likey in real situations. A load of 20 (moderate) and 100 (long) are more realistic. It can be seen that the previous patches in this series have reduced performance in general except in highly contended cases with moderate or long critical sections that performance improves a bit. This change is mostly caused by the "Prevent potential lock starvation" patch that reduce reader optimistic spinning and hence reduce reader fragmentation. The patch that further limit reader optimistic spinning doesn't seem to have too much impact on overall performance as shown in the benchmark data. The patch that disables reader optimistic spinning shows reduced performance at lightly loaded cases, but comparable or slightly better performance on with heavier contention. This patch just removes reader optimistic spinning for now. As readers are not going to do optimistic spinning anymore, we don't need to consider if the OSQ is empty or not when doing lock stealing. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-6-longman@redhat.com
2020-11-21 07:14:16 +03:00
state = rwsem_owner_state(owner, flags);
if (state != OWNER_WRITER)
return state;
rcu_read_lock();
for (;;) {
locking/rwsem: Make handoff writer optimistically spin on owner When the handoff bit is set by a writer, no other tasks other than the setting writer itself is allowed to acquire the lock. If the to-be-handoff'ed writer goes to sleep, there will be a wakeup latency period where the lock is free, but no one can acquire it. That is less than ideal. To reduce that latency, the handoff writer will now optimistically spin on the owner if it happens to be a on-cpu writer. It will spin until it releases the lock and the to-be-handoff'ed writer can then acquire the lock immediately without any delay. Of course, if the owner is not a on-cpu writer, the to-be-handoff'ed writer will have to sleep anyway. The optimistic spinning code is also modified to not stop spinning when the handoff bit is set. This will prevent an occasional setting of handoff bit from causing a bunch of optimistic spinners from entering into the wait queue causing significant reduction in throughput. On a 1-socket 22-core 44-thread Skylake system, the AIM7 shared_memory workload was run with 7000 users. The throughput (jobs/min) of the following kernels were as follows: 1) 5.2-rc6 - 8,092,486 2) 5.2-rc6 + tip's rwsem patches - 7,567,568 3) 5.2-rc6 + tip's rwsem patches + this patch - 7,954,545 Using perf-record(1), the %cpu time used by rwsem_down_write_slowpath(), rwsem_down_write_failed() and their callees for the 3 kernels were 1.70%, 5.46% and 2.08% respectively. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: x86@kernel.org Cc: Ingo Molnar <mingo@redhat.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: "H. Peter Anvin" <hpa@zytor.com> Link: https://lkml.kernel.org/r/20190625143913.24154-1-longman@redhat.com
2019-06-25 17:39:13 +03:00
/*
* When a waiting writer set the handoff flag, it may spin
* on the owner as well. Once that writer acquires the lock,
* we can spin on it. So we don't need to quit even when the
* handoff bit is set.
*/
new = rwsem_owner_flags(sem, &new_flags);
if ((new != owner) || (new_flags != flags)) {
locking/rwsem: Remove reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It also improves the chance that a reader can get the lock as writer optimistic spinning disproportionally favors writers much more than readers. Since commit d3681e269fff ("locking/rwsem: Wake up almost all readers in wait queue"), all the waiting readers are woken up so that they can all get the read lock and run in parallel. When the number of contending readers is large, allowing reader optimistic spinning will likely cause reader fragmentation where multiple smaller groups of readers can get the read lock in a sequential manner separated by writers. That reduces reader parallelism. One possible way to address that drawback is to limit the number of readers (preferably one) that can do optimistic spinning. These readers act as representatives of all the waiting readers in the wait queue as they will wake up all those waiting readers once they get the lock. Alternatively, as reader optimistic lock stealing has already enhanced fairness to readers, it may be easier to just remove reader optimistic spinning and simplifying the optimistic spinning code as a result. Performance measurements (locking throughput kops/s) using a locking microbenchmark with 50/50 reader/writer distribution and turbo-boost disabled was done on a 2-socket Cascade Lake system (48-core 96-thread) to see the impacts of these changes: 1) Vanilla - 5.10-rc3 kernel 2) Before - 5.10-rc3 kernel with previous patches in this series 2) limit-rspin - 5.10-rc3 kernel with limited reader spinning patch 3) no-rspin - 5.10-rc3 kernel with reader spinning disabled # of threads CS Load Vanilla Before limit-rspin no-rspin ------------ ------- ------- ------ ----------- -------- 2 1 5,185 5,662 5,214 5,077 4 1 5,107 4,983 5,188 4,760 8 1 4,782 4,564 4,720 4,628 16 1 4,680 4,053 4,567 3,402 32 1 4,299 1,115 1,118 1,098 64 1 3,218 983 1,001 957 96 1 1,938 944 957 930 2 20 2,008 2,128 2,264 1,665 4 20 1,390 1,033 1,046 1,101 8 20 1,472 1,155 1,098 1,213 16 20 1,332 1,077 1,089 1,122 32 20 967 914 917 980 64 20 787 874 891 858 96 20 730 836 847 844 2 100 372 356 360 355 4 100 492 425 434 392 8 100 533 537 529 538 16 100 548 572 568 598 32 100 499 520 527 537 64 100 466 517 526 512 96 100 406 497 506 509 The column "CS Load" represents the number of pause instructions issued in the locking critical section. A CS load of 1 is extremely short and is not likey in real situations. A load of 20 (moderate) and 100 (long) are more realistic. It can be seen that the previous patches in this series have reduced performance in general except in highly contended cases with moderate or long critical sections that performance improves a bit. This change is mostly caused by the "Prevent potential lock starvation" patch that reduce reader optimistic spinning and hence reduce reader fragmentation. The patch that further limit reader optimistic spinning doesn't seem to have too much impact on overall performance as shown in the benchmark data. The patch that disables reader optimistic spinning shows reduced performance at lightly loaded cases, but comparable or slightly better performance on with heavier contention. This patch just removes reader optimistic spinning for now. As readers are not going to do optimistic spinning anymore, we don't need to consider if the OSQ is empty or not when doing lock stealing. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-6-longman@redhat.com
2020-11-21 07:14:16 +03:00
state = rwsem_owner_state(new, new_flags);
break;
}
/*
* Ensure we emit the owner->on_cpu, dereference _after_
* checking sem->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();
if (need_resched() || !owner_on_cpu(owner)) {
state = OWNER_NONSPINNABLE;
break;
}
cpu_relax();
}
rcu_read_unlock();
return state;
}
locking/rwsem: Enable time-based spinning on reader-owned rwsem When the rwsem is owned by reader, writers stop optimistic spinning simply because there is no easy way to figure out if all the readers are actively running or not. However, there are scenarios where the readers are unlikely to sleep and optimistic spinning can help performance. This patch provides a simple mechanism for spinning on a reader-owned rwsem by a writer. It is a time threshold based spinning where the allowable spinning time can vary from 10us to 25us depending on the condition of the rwsem. When the time threshold is exceeded, the nonspinnable bits will be set in the owner field to indicate that no more optimistic spinning will be allowed on this rwsem until it becomes writer owned again. Not even readers is allowed to acquire the reader-locked rwsem by optimistic spinning for fairness. We also want a writer to acquire the lock after the readers hold the lock for a relatively long time. In order to give preference to writers under such a circumstance, the single RWSEM_NONSPINNABLE bit is now split into two - one for reader and one for writer. When optimistic spinning is disabled, both bits will be set. When the reader count drop down to 0, the writer nonspinnable bit will be cleared to allow writers to spin on the lock, but not the readers. When a writer acquires the lock, it will write its own task structure pointer into sem->owner and clear the reader nonspinnable bit in the process. The time taken for each iteration of the reader-owned rwsem spinning loop varies. Below are sample minimum elapsed times for 16 iterations of the loop. System Time for 16 Iterations ------ ---------------------- 1-socket Skylake ~800ns 4-socket Broadwell ~300ns 2-socket ThunderX2 (arm64) ~250ns When the lock cacheline is contended, we can see up to almost 10X increase in elapsed time. So 25us will be at most 500, 1300 and 1600 iterations for each of the above systems. With a locking microbenchmark running on 5.1 based kernel, the total locking rates (in kops/s) on a 8-socket IvyBridge-EX system with equal numbers of readers and writers before and after this patch were as follows: # of Threads Pre-patch Post-patch ------------ --------- ---------- 2 1,759 6,684 4 1,684 6,738 8 1,074 7,222 16 900 7,163 32 458 7,316 64 208 520 128 168 425 240 143 474 This patch gives a big boost in performance for mixed reader/writer workloads. With 32 locking threads, the rwsem lock event data were: rwsem_opt_fail=79850 rwsem_opt_nospin=5069 rwsem_opt_rlock=597484 rwsem_opt_wlock=957339 rwsem_sleep_reader=57782 rwsem_sleep_writer=55663 With 64 locking threads, the data looked like: rwsem_opt_fail=346723 rwsem_opt_nospin=6293 rwsem_opt_rlock=1127119 rwsem_opt_wlock=1400628 rwsem_sleep_reader=308201 rwsem_sleep_writer=72281 So a lot more threads acquired the lock in the slowpath and more threads went to sleep. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-15-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:13 +03:00
/*
* Calculate reader-owned rwsem spinning threshold for writer
*
* The more readers own the rwsem, the longer it will take for them to
* wind down and free the rwsem. So the empirical formula used to
* determine the actual spinning time limit here is:
*
* Spinning threshold = (10 + nr_readers/2)us
*
* The limit is capped to a maximum of 25us (30 readers). This is just
* a heuristic and is subjected to change in the future.
*/
static inline u64 rwsem_rspin_threshold(struct rw_semaphore *sem)
{
long count = atomic_long_read(&sem->count);
int readers = count >> RWSEM_READER_SHIFT;
u64 delta;
if (readers > 30)
readers = 30;
delta = (20 + readers) * NSEC_PER_USEC / 2;
return sched_clock() + delta;
}
locking/rwsem: Remove reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It also improves the chance that a reader can get the lock as writer optimistic spinning disproportionally favors writers much more than readers. Since commit d3681e269fff ("locking/rwsem: Wake up almost all readers in wait queue"), all the waiting readers are woken up so that they can all get the read lock and run in parallel. When the number of contending readers is large, allowing reader optimistic spinning will likely cause reader fragmentation where multiple smaller groups of readers can get the read lock in a sequential manner separated by writers. That reduces reader parallelism. One possible way to address that drawback is to limit the number of readers (preferably one) that can do optimistic spinning. These readers act as representatives of all the waiting readers in the wait queue as they will wake up all those waiting readers once they get the lock. Alternatively, as reader optimistic lock stealing has already enhanced fairness to readers, it may be easier to just remove reader optimistic spinning and simplifying the optimistic spinning code as a result. Performance measurements (locking throughput kops/s) using a locking microbenchmark with 50/50 reader/writer distribution and turbo-boost disabled was done on a 2-socket Cascade Lake system (48-core 96-thread) to see the impacts of these changes: 1) Vanilla - 5.10-rc3 kernel 2) Before - 5.10-rc3 kernel with previous patches in this series 2) limit-rspin - 5.10-rc3 kernel with limited reader spinning patch 3) no-rspin - 5.10-rc3 kernel with reader spinning disabled # of threads CS Load Vanilla Before limit-rspin no-rspin ------------ ------- ------- ------ ----------- -------- 2 1 5,185 5,662 5,214 5,077 4 1 5,107 4,983 5,188 4,760 8 1 4,782 4,564 4,720 4,628 16 1 4,680 4,053 4,567 3,402 32 1 4,299 1,115 1,118 1,098 64 1 3,218 983 1,001 957 96 1 1,938 944 957 930 2 20 2,008 2,128 2,264 1,665 4 20 1,390 1,033 1,046 1,101 8 20 1,472 1,155 1,098 1,213 16 20 1,332 1,077 1,089 1,122 32 20 967 914 917 980 64 20 787 874 891 858 96 20 730 836 847 844 2 100 372 356 360 355 4 100 492 425 434 392 8 100 533 537 529 538 16 100 548 572 568 598 32 100 499 520 527 537 64 100 466 517 526 512 96 100 406 497 506 509 The column "CS Load" represents the number of pause instructions issued in the locking critical section. A CS load of 1 is extremely short and is not likey in real situations. A load of 20 (moderate) and 100 (long) are more realistic. It can be seen that the previous patches in this series have reduced performance in general except in highly contended cases with moderate or long critical sections that performance improves a bit. This change is mostly caused by the "Prevent potential lock starvation" patch that reduce reader optimistic spinning and hence reduce reader fragmentation. The patch that further limit reader optimistic spinning doesn't seem to have too much impact on overall performance as shown in the benchmark data. The patch that disables reader optimistic spinning shows reduced performance at lightly loaded cases, but comparable or slightly better performance on with heavier contention. This patch just removes reader optimistic spinning for now. As readers are not going to do optimistic spinning anymore, we don't need to consider if the OSQ is empty or not when doing lock stealing. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-6-longman@redhat.com
2020-11-21 07:14:16 +03:00
static bool rwsem_optimistic_spin(struct rw_semaphore *sem)
{
bool taken = false;
locking/rwsem: More optimal RT task handling of null owner An RT task can do optimistic spinning only if the lock holder is actually running. If the state of the lock holder isn't known, there is a possibility that high priority of the RT task may block forward progress of the lock holder if it happens to reside on the same CPU. This will lead to deadlock. So we have to make sure that an RT task will not spin on a reader-owned rwsem. When the owner is temporarily set to NULL, there are two cases where we may want to continue spinning: 1) The lock owner is in the process of releasing the lock, sem->owner is cleared but the lock has not been released yet. 2) The lock was free and owner cleared, but another task just comes in and acquire the lock before we try to get it. The new owner may be a spinnable writer. So an RT task is now made to retry one more time to see if it can acquire the lock or continue spinning on the new owning writer. When testing on a 8-socket IvyBridge-EX system, the one additional retry seems to improve locking performance of RT write locking threads under heavy contentions. The table below shows the locking rates (in kops/s) with various write locking threads before and after the patch. Locking threads Pre-patch Post-patch --------------- --------- ----------- 4 2,753 2,608 8 2,529 2,520 16 1,727 1,918 32 1,263 1,956 64 889 1,343 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-10-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:08 +03:00
int prev_owner_state = OWNER_NULL;
locking/rwsem: Enable time-based spinning on reader-owned rwsem When the rwsem is owned by reader, writers stop optimistic spinning simply because there is no easy way to figure out if all the readers are actively running or not. However, there are scenarios where the readers are unlikely to sleep and optimistic spinning can help performance. This patch provides a simple mechanism for spinning on a reader-owned rwsem by a writer. It is a time threshold based spinning where the allowable spinning time can vary from 10us to 25us depending on the condition of the rwsem. When the time threshold is exceeded, the nonspinnable bits will be set in the owner field to indicate that no more optimistic spinning will be allowed on this rwsem until it becomes writer owned again. Not even readers is allowed to acquire the reader-locked rwsem by optimistic spinning for fairness. We also want a writer to acquire the lock after the readers hold the lock for a relatively long time. In order to give preference to writers under such a circumstance, the single RWSEM_NONSPINNABLE bit is now split into two - one for reader and one for writer. When optimistic spinning is disabled, both bits will be set. When the reader count drop down to 0, the writer nonspinnable bit will be cleared to allow writers to spin on the lock, but not the readers. When a writer acquires the lock, it will write its own task structure pointer into sem->owner and clear the reader nonspinnable bit in the process. The time taken for each iteration of the reader-owned rwsem spinning loop varies. Below are sample minimum elapsed times for 16 iterations of the loop. System Time for 16 Iterations ------ ---------------------- 1-socket Skylake ~800ns 4-socket Broadwell ~300ns 2-socket ThunderX2 (arm64) ~250ns When the lock cacheline is contended, we can see up to almost 10X increase in elapsed time. So 25us will be at most 500, 1300 and 1600 iterations for each of the above systems. With a locking microbenchmark running on 5.1 based kernel, the total locking rates (in kops/s) on a 8-socket IvyBridge-EX system with equal numbers of readers and writers before and after this patch were as follows: # of Threads Pre-patch Post-patch ------------ --------- ---------- 2 1,759 6,684 4 1,684 6,738 8 1,074 7,222 16 900 7,163 32 458 7,316 64 208 520 128 168 425 240 143 474 This patch gives a big boost in performance for mixed reader/writer workloads. With 32 locking threads, the rwsem lock event data were: rwsem_opt_fail=79850 rwsem_opt_nospin=5069 rwsem_opt_rlock=597484 rwsem_opt_wlock=957339 rwsem_sleep_reader=57782 rwsem_sleep_writer=55663 With 64 locking threads, the data looked like: rwsem_opt_fail=346723 rwsem_opt_nospin=6293 rwsem_opt_rlock=1127119 rwsem_opt_wlock=1400628 rwsem_sleep_reader=308201 rwsem_sleep_writer=72281 So a lot more threads acquired the lock in the slowpath and more threads went to sleep. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-15-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:13 +03:00
int loop = 0;
u64 rspin_threshold = 0;
preempt_disable();
/* sem->wait_lock should not be held when doing optimistic spinning */
if (!osq_lock(&sem->osq))
goto done;
/*
* Optimistically spin on the owner field and attempt to acquire the
* lock whenever the owner changes. Spinning will be stopped when:
* 1) the owning writer isn't running; or
locking/rwsem: Enable time-based spinning on reader-owned rwsem When the rwsem is owned by reader, writers stop optimistic spinning simply because there is no easy way to figure out if all the readers are actively running or not. However, there are scenarios where the readers are unlikely to sleep and optimistic spinning can help performance. This patch provides a simple mechanism for spinning on a reader-owned rwsem by a writer. It is a time threshold based spinning where the allowable spinning time can vary from 10us to 25us depending on the condition of the rwsem. When the time threshold is exceeded, the nonspinnable bits will be set in the owner field to indicate that no more optimistic spinning will be allowed on this rwsem until it becomes writer owned again. Not even readers is allowed to acquire the reader-locked rwsem by optimistic spinning for fairness. We also want a writer to acquire the lock after the readers hold the lock for a relatively long time. In order to give preference to writers under such a circumstance, the single RWSEM_NONSPINNABLE bit is now split into two - one for reader and one for writer. When optimistic spinning is disabled, both bits will be set. When the reader count drop down to 0, the writer nonspinnable bit will be cleared to allow writers to spin on the lock, but not the readers. When a writer acquires the lock, it will write its own task structure pointer into sem->owner and clear the reader nonspinnable bit in the process. The time taken for each iteration of the reader-owned rwsem spinning loop varies. Below are sample minimum elapsed times for 16 iterations of the loop. System Time for 16 Iterations ------ ---------------------- 1-socket Skylake ~800ns 4-socket Broadwell ~300ns 2-socket ThunderX2 (arm64) ~250ns When the lock cacheline is contended, we can see up to almost 10X increase in elapsed time. So 25us will be at most 500, 1300 and 1600 iterations for each of the above systems. With a locking microbenchmark running on 5.1 based kernel, the total locking rates (in kops/s) on a 8-socket IvyBridge-EX system with equal numbers of readers and writers before and after this patch were as follows: # of Threads Pre-patch Post-patch ------------ --------- ---------- 2 1,759 6,684 4 1,684 6,738 8 1,074 7,222 16 900 7,163 32 458 7,316 64 208 520 128 168 425 240 143 474 This patch gives a big boost in performance for mixed reader/writer workloads. With 32 locking threads, the rwsem lock event data were: rwsem_opt_fail=79850 rwsem_opt_nospin=5069 rwsem_opt_rlock=597484 rwsem_opt_wlock=957339 rwsem_sleep_reader=57782 rwsem_sleep_writer=55663 With 64 locking threads, the data looked like: rwsem_opt_fail=346723 rwsem_opt_nospin=6293 rwsem_opt_rlock=1127119 rwsem_opt_wlock=1400628 rwsem_sleep_reader=308201 rwsem_sleep_writer=72281 So a lot more threads acquired the lock in the slowpath and more threads went to sleep. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-15-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:13 +03:00
* 2) readers own the lock and spinning time has exceeded limit.
*/
locking/rwsem: More optimal RT task handling of null owner An RT task can do optimistic spinning only if the lock holder is actually running. If the state of the lock holder isn't known, there is a possibility that high priority of the RT task may block forward progress of the lock holder if it happens to reside on the same CPU. This will lead to deadlock. So we have to make sure that an RT task will not spin on a reader-owned rwsem. When the owner is temporarily set to NULL, there are two cases where we may want to continue spinning: 1) The lock owner is in the process of releasing the lock, sem->owner is cleared but the lock has not been released yet. 2) The lock was free and owner cleared, but another task just comes in and acquire the lock before we try to get it. The new owner may be a spinnable writer. So an RT task is now made to retry one more time to see if it can acquire the lock or continue spinning on the new owning writer. When testing on a 8-socket IvyBridge-EX system, the one additional retry seems to improve locking performance of RT write locking threads under heavy contentions. The table below shows the locking rates (in kops/s) with various write locking threads before and after the patch. Locking threads Pre-patch Post-patch --------------- --------- ----------- 4 2,753 2,608 8 2,529 2,520 16 1,727 1,918 32 1,263 1,956 64 889 1,343 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-10-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:08 +03:00
for (;;) {
locking/rwsem: Enable time-based spinning on reader-owned rwsem When the rwsem is owned by reader, writers stop optimistic spinning simply because there is no easy way to figure out if all the readers are actively running or not. However, there are scenarios where the readers are unlikely to sleep and optimistic spinning can help performance. This patch provides a simple mechanism for spinning on a reader-owned rwsem by a writer. It is a time threshold based spinning where the allowable spinning time can vary from 10us to 25us depending on the condition of the rwsem. When the time threshold is exceeded, the nonspinnable bits will be set in the owner field to indicate that no more optimistic spinning will be allowed on this rwsem until it becomes writer owned again. Not even readers is allowed to acquire the reader-locked rwsem by optimistic spinning for fairness. We also want a writer to acquire the lock after the readers hold the lock for a relatively long time. In order to give preference to writers under such a circumstance, the single RWSEM_NONSPINNABLE bit is now split into two - one for reader and one for writer. When optimistic spinning is disabled, both bits will be set. When the reader count drop down to 0, the writer nonspinnable bit will be cleared to allow writers to spin on the lock, but not the readers. When a writer acquires the lock, it will write its own task structure pointer into sem->owner and clear the reader nonspinnable bit in the process. The time taken for each iteration of the reader-owned rwsem spinning loop varies. Below are sample minimum elapsed times for 16 iterations of the loop. System Time for 16 Iterations ------ ---------------------- 1-socket Skylake ~800ns 4-socket Broadwell ~300ns 2-socket ThunderX2 (arm64) ~250ns When the lock cacheline is contended, we can see up to almost 10X increase in elapsed time. So 25us will be at most 500, 1300 and 1600 iterations for each of the above systems. With a locking microbenchmark running on 5.1 based kernel, the total locking rates (in kops/s) on a 8-socket IvyBridge-EX system with equal numbers of readers and writers before and after this patch were as follows: # of Threads Pre-patch Post-patch ------------ --------- ---------- 2 1,759 6,684 4 1,684 6,738 8 1,074 7,222 16 900 7,163 32 458 7,316 64 208 520 128 168 425 240 143 474 This patch gives a big boost in performance for mixed reader/writer workloads. With 32 locking threads, the rwsem lock event data were: rwsem_opt_fail=79850 rwsem_opt_nospin=5069 rwsem_opt_rlock=597484 rwsem_opt_wlock=957339 rwsem_sleep_reader=57782 rwsem_sleep_writer=55663 With 64 locking threads, the data looked like: rwsem_opt_fail=346723 rwsem_opt_nospin=6293 rwsem_opt_rlock=1127119 rwsem_opt_wlock=1400628 rwsem_sleep_reader=308201 rwsem_sleep_writer=72281 So a lot more threads acquired the lock in the slowpath and more threads went to sleep. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-15-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:13 +03:00
enum owner_state owner_state;
locking/rwsem: More optimal RT task handling of null owner An RT task can do optimistic spinning only if the lock holder is actually running. If the state of the lock holder isn't known, there is a possibility that high priority of the RT task may block forward progress of the lock holder if it happens to reside on the same CPU. This will lead to deadlock. So we have to make sure that an RT task will not spin on a reader-owned rwsem. When the owner is temporarily set to NULL, there are two cases where we may want to continue spinning: 1) The lock owner is in the process of releasing the lock, sem->owner is cleared but the lock has not been released yet. 2) The lock was free and owner cleared, but another task just comes in and acquire the lock before we try to get it. The new owner may be a spinnable writer. So an RT task is now made to retry one more time to see if it can acquire the lock or continue spinning on the new owning writer. When testing on a 8-socket IvyBridge-EX system, the one additional retry seems to improve locking performance of RT write locking threads under heavy contentions. The table below shows the locking rates (in kops/s) with various write locking threads before and after the patch. Locking threads Pre-patch Post-patch --------------- --------- ----------- 4 2,753 2,608 8 2,529 2,520 16 1,727 1,918 32 1,263 1,956 64 889 1,343 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-10-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:08 +03:00
locking/rwsem: Remove reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It also improves the chance that a reader can get the lock as writer optimistic spinning disproportionally favors writers much more than readers. Since commit d3681e269fff ("locking/rwsem: Wake up almost all readers in wait queue"), all the waiting readers are woken up so that they can all get the read lock and run in parallel. When the number of contending readers is large, allowing reader optimistic spinning will likely cause reader fragmentation where multiple smaller groups of readers can get the read lock in a sequential manner separated by writers. That reduces reader parallelism. One possible way to address that drawback is to limit the number of readers (preferably one) that can do optimistic spinning. These readers act as representatives of all the waiting readers in the wait queue as they will wake up all those waiting readers once they get the lock. Alternatively, as reader optimistic lock stealing has already enhanced fairness to readers, it may be easier to just remove reader optimistic spinning and simplifying the optimistic spinning code as a result. Performance measurements (locking throughput kops/s) using a locking microbenchmark with 50/50 reader/writer distribution and turbo-boost disabled was done on a 2-socket Cascade Lake system (48-core 96-thread) to see the impacts of these changes: 1) Vanilla - 5.10-rc3 kernel 2) Before - 5.10-rc3 kernel with previous patches in this series 2) limit-rspin - 5.10-rc3 kernel with limited reader spinning patch 3) no-rspin - 5.10-rc3 kernel with reader spinning disabled # of threads CS Load Vanilla Before limit-rspin no-rspin ------------ ------- ------- ------ ----------- -------- 2 1 5,185 5,662 5,214 5,077 4 1 5,107 4,983 5,188 4,760 8 1 4,782 4,564 4,720 4,628 16 1 4,680 4,053 4,567 3,402 32 1 4,299 1,115 1,118 1,098 64 1 3,218 983 1,001 957 96 1 1,938 944 957 930 2 20 2,008 2,128 2,264 1,665 4 20 1,390 1,033 1,046 1,101 8 20 1,472 1,155 1,098 1,213 16 20 1,332 1,077 1,089 1,122 32 20 967 914 917 980 64 20 787 874 891 858 96 20 730 836 847 844 2 100 372 356 360 355 4 100 492 425 434 392 8 100 533 537 529 538 16 100 548 572 568 598 32 100 499 520 527 537 64 100 466 517 526 512 96 100 406 497 506 509 The column "CS Load" represents the number of pause instructions issued in the locking critical section. A CS load of 1 is extremely short and is not likey in real situations. A load of 20 (moderate) and 100 (long) are more realistic. It can be seen that the previous patches in this series have reduced performance in general except in highly contended cases with moderate or long critical sections that performance improves a bit. This change is mostly caused by the "Prevent potential lock starvation" patch that reduce reader optimistic spinning and hence reduce reader fragmentation. The patch that further limit reader optimistic spinning doesn't seem to have too much impact on overall performance as shown in the benchmark data. The patch that disables reader optimistic spinning shows reduced performance at lightly loaded cases, but comparable or slightly better performance on with heavier contention. This patch just removes reader optimistic spinning for now. As readers are not going to do optimistic spinning anymore, we don't need to consider if the OSQ is empty or not when doing lock stealing. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-6-longman@redhat.com
2020-11-21 07:14:16 +03:00
owner_state = rwsem_spin_on_owner(sem);
locking/rwsem: More optimal RT task handling of null owner An RT task can do optimistic spinning only if the lock holder is actually running. If the state of the lock holder isn't known, there is a possibility that high priority of the RT task may block forward progress of the lock holder if it happens to reside on the same CPU. This will lead to deadlock. So we have to make sure that an RT task will not spin on a reader-owned rwsem. When the owner is temporarily set to NULL, there are two cases where we may want to continue spinning: 1) The lock owner is in the process of releasing the lock, sem->owner is cleared but the lock has not been released yet. 2) The lock was free and owner cleared, but another task just comes in and acquire the lock before we try to get it. The new owner may be a spinnable writer. So an RT task is now made to retry one more time to see if it can acquire the lock or continue spinning on the new owning writer. When testing on a 8-socket IvyBridge-EX system, the one additional retry seems to improve locking performance of RT write locking threads under heavy contentions. The table below shows the locking rates (in kops/s) with various write locking threads before and after the patch. Locking threads Pre-patch Post-patch --------------- --------- ----------- 4 2,753 2,608 8 2,529 2,520 16 1,727 1,918 32 1,263 1,956 64 889 1,343 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-10-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:08 +03:00
if (!(owner_state & OWNER_SPINNABLE))
break;
/*
* Try to acquire the lock
*/
locking/rwsem: Remove reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It also improves the chance that a reader can get the lock as writer optimistic spinning disproportionally favors writers much more than readers. Since commit d3681e269fff ("locking/rwsem: Wake up almost all readers in wait queue"), all the waiting readers are woken up so that they can all get the read lock and run in parallel. When the number of contending readers is large, allowing reader optimistic spinning will likely cause reader fragmentation where multiple smaller groups of readers can get the read lock in a sequential manner separated by writers. That reduces reader parallelism. One possible way to address that drawback is to limit the number of readers (preferably one) that can do optimistic spinning. These readers act as representatives of all the waiting readers in the wait queue as they will wake up all those waiting readers once they get the lock. Alternatively, as reader optimistic lock stealing has already enhanced fairness to readers, it may be easier to just remove reader optimistic spinning and simplifying the optimistic spinning code as a result. Performance measurements (locking throughput kops/s) using a locking microbenchmark with 50/50 reader/writer distribution and turbo-boost disabled was done on a 2-socket Cascade Lake system (48-core 96-thread) to see the impacts of these changes: 1) Vanilla - 5.10-rc3 kernel 2) Before - 5.10-rc3 kernel with previous patches in this series 2) limit-rspin - 5.10-rc3 kernel with limited reader spinning patch 3) no-rspin - 5.10-rc3 kernel with reader spinning disabled # of threads CS Load Vanilla Before limit-rspin no-rspin ------------ ------- ------- ------ ----------- -------- 2 1 5,185 5,662 5,214 5,077 4 1 5,107 4,983 5,188 4,760 8 1 4,782 4,564 4,720 4,628 16 1 4,680 4,053 4,567 3,402 32 1 4,299 1,115 1,118 1,098 64 1 3,218 983 1,001 957 96 1 1,938 944 957 930 2 20 2,008 2,128 2,264 1,665 4 20 1,390 1,033 1,046 1,101 8 20 1,472 1,155 1,098 1,213 16 20 1,332 1,077 1,089 1,122 32 20 967 914 917 980 64 20 787 874 891 858 96 20 730 836 847 844 2 100 372 356 360 355 4 100 492 425 434 392 8 100 533 537 529 538 16 100 548 572 568 598 32 100 499 520 527 537 64 100 466 517 526 512 96 100 406 497 506 509 The column "CS Load" represents the number of pause instructions issued in the locking critical section. A CS load of 1 is extremely short and is not likey in real situations. A load of 20 (moderate) and 100 (long) are more realistic. It can be seen that the previous patches in this series have reduced performance in general except in highly contended cases with moderate or long critical sections that performance improves a bit. This change is mostly caused by the "Prevent potential lock starvation" patch that reduce reader optimistic spinning and hence reduce reader fragmentation. The patch that further limit reader optimistic spinning doesn't seem to have too much impact on overall performance as shown in the benchmark data. The patch that disables reader optimistic spinning shows reduced performance at lightly loaded cases, but comparable or slightly better performance on with heavier contention. This patch just removes reader optimistic spinning for now. As readers are not going to do optimistic spinning anymore, we don't need to consider if the OSQ is empty or not when doing lock stealing. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-6-longman@redhat.com
2020-11-21 07:14:16 +03:00
taken = rwsem_try_write_lock_unqueued(sem);
if (taken)
break;
locking/rwsem: Enable time-based spinning on reader-owned rwsem When the rwsem is owned by reader, writers stop optimistic spinning simply because there is no easy way to figure out if all the readers are actively running or not. However, there are scenarios where the readers are unlikely to sleep and optimistic spinning can help performance. This patch provides a simple mechanism for spinning on a reader-owned rwsem by a writer. It is a time threshold based spinning where the allowable spinning time can vary from 10us to 25us depending on the condition of the rwsem. When the time threshold is exceeded, the nonspinnable bits will be set in the owner field to indicate that no more optimistic spinning will be allowed on this rwsem until it becomes writer owned again. Not even readers is allowed to acquire the reader-locked rwsem by optimistic spinning for fairness. We also want a writer to acquire the lock after the readers hold the lock for a relatively long time. In order to give preference to writers under such a circumstance, the single RWSEM_NONSPINNABLE bit is now split into two - one for reader and one for writer. When optimistic spinning is disabled, both bits will be set. When the reader count drop down to 0, the writer nonspinnable bit will be cleared to allow writers to spin on the lock, but not the readers. When a writer acquires the lock, it will write its own task structure pointer into sem->owner and clear the reader nonspinnable bit in the process. The time taken for each iteration of the reader-owned rwsem spinning loop varies. Below are sample minimum elapsed times for 16 iterations of the loop. System Time for 16 Iterations ------ ---------------------- 1-socket Skylake ~800ns 4-socket Broadwell ~300ns 2-socket ThunderX2 (arm64) ~250ns When the lock cacheline is contended, we can see up to almost 10X increase in elapsed time. So 25us will be at most 500, 1300 and 1600 iterations for each of the above systems. With a locking microbenchmark running on 5.1 based kernel, the total locking rates (in kops/s) on a 8-socket IvyBridge-EX system with equal numbers of readers and writers before and after this patch were as follows: # of Threads Pre-patch Post-patch ------------ --------- ---------- 2 1,759 6,684 4 1,684 6,738 8 1,074 7,222 16 900 7,163 32 458 7,316 64 208 520 128 168 425 240 143 474 This patch gives a big boost in performance for mixed reader/writer workloads. With 32 locking threads, the rwsem lock event data were: rwsem_opt_fail=79850 rwsem_opt_nospin=5069 rwsem_opt_rlock=597484 rwsem_opt_wlock=957339 rwsem_sleep_reader=57782 rwsem_sleep_writer=55663 With 64 locking threads, the data looked like: rwsem_opt_fail=346723 rwsem_opt_nospin=6293 rwsem_opt_rlock=1127119 rwsem_opt_wlock=1400628 rwsem_sleep_reader=308201 rwsem_sleep_writer=72281 So a lot more threads acquired the lock in the slowpath and more threads went to sleep. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-15-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:13 +03:00
/*
* Time-based reader-owned rwsem optimistic spinning
*/
locking/rwsem: Remove reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It also improves the chance that a reader can get the lock as writer optimistic spinning disproportionally favors writers much more than readers. Since commit d3681e269fff ("locking/rwsem: Wake up almost all readers in wait queue"), all the waiting readers are woken up so that they can all get the read lock and run in parallel. When the number of contending readers is large, allowing reader optimistic spinning will likely cause reader fragmentation where multiple smaller groups of readers can get the read lock in a sequential manner separated by writers. That reduces reader parallelism. One possible way to address that drawback is to limit the number of readers (preferably one) that can do optimistic spinning. These readers act as representatives of all the waiting readers in the wait queue as they will wake up all those waiting readers once they get the lock. Alternatively, as reader optimistic lock stealing has already enhanced fairness to readers, it may be easier to just remove reader optimistic spinning and simplifying the optimistic spinning code as a result. Performance measurements (locking throughput kops/s) using a locking microbenchmark with 50/50 reader/writer distribution and turbo-boost disabled was done on a 2-socket Cascade Lake system (48-core 96-thread) to see the impacts of these changes: 1) Vanilla - 5.10-rc3 kernel 2) Before - 5.10-rc3 kernel with previous patches in this series 2) limit-rspin - 5.10-rc3 kernel with limited reader spinning patch 3) no-rspin - 5.10-rc3 kernel with reader spinning disabled # of threads CS Load Vanilla Before limit-rspin no-rspin ------------ ------- ------- ------ ----------- -------- 2 1 5,185 5,662 5,214 5,077 4 1 5,107 4,983 5,188 4,760 8 1 4,782 4,564 4,720 4,628 16 1 4,680 4,053 4,567 3,402 32 1 4,299 1,115 1,118 1,098 64 1 3,218 983 1,001 957 96 1 1,938 944 957 930 2 20 2,008 2,128 2,264 1,665 4 20 1,390 1,033 1,046 1,101 8 20 1,472 1,155 1,098 1,213 16 20 1,332 1,077 1,089 1,122 32 20 967 914 917 980 64 20 787 874 891 858 96 20 730 836 847 844 2 100 372 356 360 355 4 100 492 425 434 392 8 100 533 537 529 538 16 100 548 572 568 598 32 100 499 520 527 537 64 100 466 517 526 512 96 100 406 497 506 509 The column "CS Load" represents the number of pause instructions issued in the locking critical section. A CS load of 1 is extremely short and is not likey in real situations. A load of 20 (moderate) and 100 (long) are more realistic. It can be seen that the previous patches in this series have reduced performance in general except in highly contended cases with moderate or long critical sections that performance improves a bit. This change is mostly caused by the "Prevent potential lock starvation" patch that reduce reader optimistic spinning and hence reduce reader fragmentation. The patch that further limit reader optimistic spinning doesn't seem to have too much impact on overall performance as shown in the benchmark data. The patch that disables reader optimistic spinning shows reduced performance at lightly loaded cases, but comparable or slightly better performance on with heavier contention. This patch just removes reader optimistic spinning for now. As readers are not going to do optimistic spinning anymore, we don't need to consider if the OSQ is empty or not when doing lock stealing. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-6-longman@redhat.com
2020-11-21 07:14:16 +03:00
if (owner_state == OWNER_READER) {
locking/rwsem: Enable time-based spinning on reader-owned rwsem When the rwsem is owned by reader, writers stop optimistic spinning simply because there is no easy way to figure out if all the readers are actively running or not. However, there are scenarios where the readers are unlikely to sleep and optimistic spinning can help performance. This patch provides a simple mechanism for spinning on a reader-owned rwsem by a writer. It is a time threshold based spinning where the allowable spinning time can vary from 10us to 25us depending on the condition of the rwsem. When the time threshold is exceeded, the nonspinnable bits will be set in the owner field to indicate that no more optimistic spinning will be allowed on this rwsem until it becomes writer owned again. Not even readers is allowed to acquire the reader-locked rwsem by optimistic spinning for fairness. We also want a writer to acquire the lock after the readers hold the lock for a relatively long time. In order to give preference to writers under such a circumstance, the single RWSEM_NONSPINNABLE bit is now split into two - one for reader and one for writer. When optimistic spinning is disabled, both bits will be set. When the reader count drop down to 0, the writer nonspinnable bit will be cleared to allow writers to spin on the lock, but not the readers. When a writer acquires the lock, it will write its own task structure pointer into sem->owner and clear the reader nonspinnable bit in the process. The time taken for each iteration of the reader-owned rwsem spinning loop varies. Below are sample minimum elapsed times for 16 iterations of the loop. System Time for 16 Iterations ------ ---------------------- 1-socket Skylake ~800ns 4-socket Broadwell ~300ns 2-socket ThunderX2 (arm64) ~250ns When the lock cacheline is contended, we can see up to almost 10X increase in elapsed time. So 25us will be at most 500, 1300 and 1600 iterations for each of the above systems. With a locking microbenchmark running on 5.1 based kernel, the total locking rates (in kops/s) on a 8-socket IvyBridge-EX system with equal numbers of readers and writers before and after this patch were as follows: # of Threads Pre-patch Post-patch ------------ --------- ---------- 2 1,759 6,684 4 1,684 6,738 8 1,074 7,222 16 900 7,163 32 458 7,316 64 208 520 128 168 425 240 143 474 This patch gives a big boost in performance for mixed reader/writer workloads. With 32 locking threads, the rwsem lock event data were: rwsem_opt_fail=79850 rwsem_opt_nospin=5069 rwsem_opt_rlock=597484 rwsem_opt_wlock=957339 rwsem_sleep_reader=57782 rwsem_sleep_writer=55663 With 64 locking threads, the data looked like: rwsem_opt_fail=346723 rwsem_opt_nospin=6293 rwsem_opt_rlock=1127119 rwsem_opt_wlock=1400628 rwsem_sleep_reader=308201 rwsem_sleep_writer=72281 So a lot more threads acquired the lock in the slowpath and more threads went to sleep. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-15-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:13 +03:00
/*
* Re-initialize rspin_threshold every time when
* the owner state changes from non-reader to reader.
* This allows a writer to steal the lock in between
* 2 reader phases and have the threshold reset at
* the beginning of the 2nd reader phase.
*/
if (prev_owner_state != OWNER_READER) {
locking/rwsem: Remove reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It also improves the chance that a reader can get the lock as writer optimistic spinning disproportionally favors writers much more than readers. Since commit d3681e269fff ("locking/rwsem: Wake up almost all readers in wait queue"), all the waiting readers are woken up so that they can all get the read lock and run in parallel. When the number of contending readers is large, allowing reader optimistic spinning will likely cause reader fragmentation where multiple smaller groups of readers can get the read lock in a sequential manner separated by writers. That reduces reader parallelism. One possible way to address that drawback is to limit the number of readers (preferably one) that can do optimistic spinning. These readers act as representatives of all the waiting readers in the wait queue as they will wake up all those waiting readers once they get the lock. Alternatively, as reader optimistic lock stealing has already enhanced fairness to readers, it may be easier to just remove reader optimistic spinning and simplifying the optimistic spinning code as a result. Performance measurements (locking throughput kops/s) using a locking microbenchmark with 50/50 reader/writer distribution and turbo-boost disabled was done on a 2-socket Cascade Lake system (48-core 96-thread) to see the impacts of these changes: 1) Vanilla - 5.10-rc3 kernel 2) Before - 5.10-rc3 kernel with previous patches in this series 2) limit-rspin - 5.10-rc3 kernel with limited reader spinning patch 3) no-rspin - 5.10-rc3 kernel with reader spinning disabled # of threads CS Load Vanilla Before limit-rspin no-rspin ------------ ------- ------- ------ ----------- -------- 2 1 5,185 5,662 5,214 5,077 4 1 5,107 4,983 5,188 4,760 8 1 4,782 4,564 4,720 4,628 16 1 4,680 4,053 4,567 3,402 32 1 4,299 1,115 1,118 1,098 64 1 3,218 983 1,001 957 96 1 1,938 944 957 930 2 20 2,008 2,128 2,264 1,665 4 20 1,390 1,033 1,046 1,101 8 20 1,472 1,155 1,098 1,213 16 20 1,332 1,077 1,089 1,122 32 20 967 914 917 980 64 20 787 874 891 858 96 20 730 836 847 844 2 100 372 356 360 355 4 100 492 425 434 392 8 100 533 537 529 538 16 100 548 572 568 598 32 100 499 520 527 537 64 100 466 517 526 512 96 100 406 497 506 509 The column "CS Load" represents the number of pause instructions issued in the locking critical section. A CS load of 1 is extremely short and is not likey in real situations. A load of 20 (moderate) and 100 (long) are more realistic. It can be seen that the previous patches in this series have reduced performance in general except in highly contended cases with moderate or long critical sections that performance improves a bit. This change is mostly caused by the "Prevent potential lock starvation" patch that reduce reader optimistic spinning and hence reduce reader fragmentation. The patch that further limit reader optimistic spinning doesn't seem to have too much impact on overall performance as shown in the benchmark data. The patch that disables reader optimistic spinning shows reduced performance at lightly loaded cases, but comparable or slightly better performance on with heavier contention. This patch just removes reader optimistic spinning for now. As readers are not going to do optimistic spinning anymore, we don't need to consider if the OSQ is empty or not when doing lock stealing. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-6-longman@redhat.com
2020-11-21 07:14:16 +03:00
if (rwsem_test_oflags(sem, RWSEM_NONSPINNABLE))
locking/rwsem: Enable time-based spinning on reader-owned rwsem When the rwsem is owned by reader, writers stop optimistic spinning simply because there is no easy way to figure out if all the readers are actively running or not. However, there are scenarios where the readers are unlikely to sleep and optimistic spinning can help performance. This patch provides a simple mechanism for spinning on a reader-owned rwsem by a writer. It is a time threshold based spinning where the allowable spinning time can vary from 10us to 25us depending on the condition of the rwsem. When the time threshold is exceeded, the nonspinnable bits will be set in the owner field to indicate that no more optimistic spinning will be allowed on this rwsem until it becomes writer owned again. Not even readers is allowed to acquire the reader-locked rwsem by optimistic spinning for fairness. We also want a writer to acquire the lock after the readers hold the lock for a relatively long time. In order to give preference to writers under such a circumstance, the single RWSEM_NONSPINNABLE bit is now split into two - one for reader and one for writer. When optimistic spinning is disabled, both bits will be set. When the reader count drop down to 0, the writer nonspinnable bit will be cleared to allow writers to spin on the lock, but not the readers. When a writer acquires the lock, it will write its own task structure pointer into sem->owner and clear the reader nonspinnable bit in the process. The time taken for each iteration of the reader-owned rwsem spinning loop varies. Below are sample minimum elapsed times for 16 iterations of the loop. System Time for 16 Iterations ------ ---------------------- 1-socket Skylake ~800ns 4-socket Broadwell ~300ns 2-socket ThunderX2 (arm64) ~250ns When the lock cacheline is contended, we can see up to almost 10X increase in elapsed time. So 25us will be at most 500, 1300 and 1600 iterations for each of the above systems. With a locking microbenchmark running on 5.1 based kernel, the total locking rates (in kops/s) on a 8-socket IvyBridge-EX system with equal numbers of readers and writers before and after this patch were as follows: # of Threads Pre-patch Post-patch ------------ --------- ---------- 2 1,759 6,684 4 1,684 6,738 8 1,074 7,222 16 900 7,163 32 458 7,316 64 208 520 128 168 425 240 143 474 This patch gives a big boost in performance for mixed reader/writer workloads. With 32 locking threads, the rwsem lock event data were: rwsem_opt_fail=79850 rwsem_opt_nospin=5069 rwsem_opt_rlock=597484 rwsem_opt_wlock=957339 rwsem_sleep_reader=57782 rwsem_sleep_writer=55663 With 64 locking threads, the data looked like: rwsem_opt_fail=346723 rwsem_opt_nospin=6293 rwsem_opt_rlock=1127119 rwsem_opt_wlock=1400628 rwsem_sleep_reader=308201 rwsem_sleep_writer=72281 So a lot more threads acquired the lock in the slowpath and more threads went to sleep. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-15-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:13 +03:00
break;
rspin_threshold = rwsem_rspin_threshold(sem);
loop = 0;
}
/*
* Check time threshold once every 16 iterations to
* avoid calling sched_clock() too frequently so
* as to reduce the average latency between the times
* when the lock becomes free and when the spinner
* is ready to do a trylock.
*/
else if (!(++loop & 0xf) && (sched_clock() > rspin_threshold)) {
rwsem_set_nonspinnable(sem);
lockevent_inc(rwsem_opt_nospin);
break;
}
}
/*
locking/rwsem: More optimal RT task handling of null owner An RT task can do optimistic spinning only if the lock holder is actually running. If the state of the lock holder isn't known, there is a possibility that high priority of the RT task may block forward progress of the lock holder if it happens to reside on the same CPU. This will lead to deadlock. So we have to make sure that an RT task will not spin on a reader-owned rwsem. When the owner is temporarily set to NULL, there are two cases where we may want to continue spinning: 1) The lock owner is in the process of releasing the lock, sem->owner is cleared but the lock has not been released yet. 2) The lock was free and owner cleared, but another task just comes in and acquire the lock before we try to get it. The new owner may be a spinnable writer. So an RT task is now made to retry one more time to see if it can acquire the lock or continue spinning on the new owning writer. When testing on a 8-socket IvyBridge-EX system, the one additional retry seems to improve locking performance of RT write locking threads under heavy contentions. The table below shows the locking rates (in kops/s) with various write locking threads before and after the patch. Locking threads Pre-patch Post-patch --------------- --------- ----------- 4 2,753 2,608 8 2,529 2,520 16 1,727 1,918 32 1,263 1,956 64 889 1,343 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-10-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:08 +03:00
* An RT task cannot do optimistic spinning if it cannot
* be sure the lock holder is running or live-lock may
* happen if the current task and the lock holder happen
* to run in the same CPU. However, aborting optimistic
* spinning while a NULL owner is detected may miss some
* opportunity where spinning can continue without causing
* problem.
*
* There are 2 possible cases where an RT task may be able
* to continue spinning.
*
* 1) The lock owner is in the process of releasing the
* lock, sem->owner is cleared but the lock has not
* been released yet.
* 2) The lock was free and owner cleared, but another
* task just comes in and acquire the lock before
* we try to get it. The new owner may be a spinnable
* writer.
*
* To take advantage of two scenarios listed above, the RT
locking/rwsem: More optimal RT task handling of null owner An RT task can do optimistic spinning only if the lock holder is actually running. If the state of the lock holder isn't known, there is a possibility that high priority of the RT task may block forward progress of the lock holder if it happens to reside on the same CPU. This will lead to deadlock. So we have to make sure that an RT task will not spin on a reader-owned rwsem. When the owner is temporarily set to NULL, there are two cases where we may want to continue spinning: 1) The lock owner is in the process of releasing the lock, sem->owner is cleared but the lock has not been released yet. 2) The lock was free and owner cleared, but another task just comes in and acquire the lock before we try to get it. The new owner may be a spinnable writer. So an RT task is now made to retry one more time to see if it can acquire the lock or continue spinning on the new owning writer. When testing on a 8-socket IvyBridge-EX system, the one additional retry seems to improve locking performance of RT write locking threads under heavy contentions. The table below shows the locking rates (in kops/s) with various write locking threads before and after the patch. Locking threads Pre-patch Post-patch --------------- --------- ----------- 4 2,753 2,608 8 2,529 2,520 16 1,727 1,918 32 1,263 1,956 64 889 1,343 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-10-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:08 +03:00
* task is made to retry one more time to see if it can
* acquire the lock or continue spinning on the new owning
* writer. Of course, if the time lag is long enough or the
* new owner is not a writer or spinnable, the RT task will
* quit spinning.
*
* If the owner is a writer, the need_resched() check is
* done inside rwsem_spin_on_owner(). If the owner is not
* a writer, need_resched() check needs to be done here.
*/
locking/rwsem: More optimal RT task handling of null owner An RT task can do optimistic spinning only if the lock holder is actually running. If the state of the lock holder isn't known, there is a possibility that high priority of the RT task may block forward progress of the lock holder if it happens to reside on the same CPU. This will lead to deadlock. So we have to make sure that an RT task will not spin on a reader-owned rwsem. When the owner is temporarily set to NULL, there are two cases where we may want to continue spinning: 1) The lock owner is in the process of releasing the lock, sem->owner is cleared but the lock has not been released yet. 2) The lock was free and owner cleared, but another task just comes in and acquire the lock before we try to get it. The new owner may be a spinnable writer. So an RT task is now made to retry one more time to see if it can acquire the lock or continue spinning on the new owning writer. When testing on a 8-socket IvyBridge-EX system, the one additional retry seems to improve locking performance of RT write locking threads under heavy contentions. The table below shows the locking rates (in kops/s) with various write locking threads before and after the patch. Locking threads Pre-patch Post-patch --------------- --------- ----------- 4 2,753 2,608 8 2,529 2,520 16 1,727 1,918 32 1,263 1,956 64 889 1,343 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-10-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:08 +03:00
if (owner_state != OWNER_WRITER) {
if (need_resched())
break;
if (rt_task(current) &&
(prev_owner_state != OWNER_WRITER))
break;
}
prev_owner_state = owner_state;
/*
* 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.
*/
cpu_relax();
}
osq_unlock(&sem->osq);
done:
preempt_enable();
lockevent_cond_inc(rwsem_opt_fail, !taken);
return taken;
}
locking/rwsem: Enable time-based spinning on reader-owned rwsem When the rwsem is owned by reader, writers stop optimistic spinning simply because there is no easy way to figure out if all the readers are actively running or not. However, there are scenarios where the readers are unlikely to sleep and optimistic spinning can help performance. This patch provides a simple mechanism for spinning on a reader-owned rwsem by a writer. It is a time threshold based spinning where the allowable spinning time can vary from 10us to 25us depending on the condition of the rwsem. When the time threshold is exceeded, the nonspinnable bits will be set in the owner field to indicate that no more optimistic spinning will be allowed on this rwsem until it becomes writer owned again. Not even readers is allowed to acquire the reader-locked rwsem by optimistic spinning for fairness. We also want a writer to acquire the lock after the readers hold the lock for a relatively long time. In order to give preference to writers under such a circumstance, the single RWSEM_NONSPINNABLE bit is now split into two - one for reader and one for writer. When optimistic spinning is disabled, both bits will be set. When the reader count drop down to 0, the writer nonspinnable bit will be cleared to allow writers to spin on the lock, but not the readers. When a writer acquires the lock, it will write its own task structure pointer into sem->owner and clear the reader nonspinnable bit in the process. The time taken for each iteration of the reader-owned rwsem spinning loop varies. Below are sample minimum elapsed times for 16 iterations of the loop. System Time for 16 Iterations ------ ---------------------- 1-socket Skylake ~800ns 4-socket Broadwell ~300ns 2-socket ThunderX2 (arm64) ~250ns When the lock cacheline is contended, we can see up to almost 10X increase in elapsed time. So 25us will be at most 500, 1300 and 1600 iterations for each of the above systems. With a locking microbenchmark running on 5.1 based kernel, the total locking rates (in kops/s) on a 8-socket IvyBridge-EX system with equal numbers of readers and writers before and after this patch were as follows: # of Threads Pre-patch Post-patch ------------ --------- ---------- 2 1,759 6,684 4 1,684 6,738 8 1,074 7,222 16 900 7,163 32 458 7,316 64 208 520 128 168 425 240 143 474 This patch gives a big boost in performance for mixed reader/writer workloads. With 32 locking threads, the rwsem lock event data were: rwsem_opt_fail=79850 rwsem_opt_nospin=5069 rwsem_opt_rlock=597484 rwsem_opt_wlock=957339 rwsem_sleep_reader=57782 rwsem_sleep_writer=55663 With 64 locking threads, the data looked like: rwsem_opt_fail=346723 rwsem_opt_nospin=6293 rwsem_opt_rlock=1127119 rwsem_opt_wlock=1400628 rwsem_sleep_reader=308201 rwsem_sleep_writer=72281 So a lot more threads acquired the lock in the slowpath and more threads went to sleep. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-15-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:13 +03:00
/*
locking/rwsem: Remove reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It also improves the chance that a reader can get the lock as writer optimistic spinning disproportionally favors writers much more than readers. Since commit d3681e269fff ("locking/rwsem: Wake up almost all readers in wait queue"), all the waiting readers are woken up so that they can all get the read lock and run in parallel. When the number of contending readers is large, allowing reader optimistic spinning will likely cause reader fragmentation where multiple smaller groups of readers can get the read lock in a sequential manner separated by writers. That reduces reader parallelism. One possible way to address that drawback is to limit the number of readers (preferably one) that can do optimistic spinning. These readers act as representatives of all the waiting readers in the wait queue as they will wake up all those waiting readers once they get the lock. Alternatively, as reader optimistic lock stealing has already enhanced fairness to readers, it may be easier to just remove reader optimistic spinning and simplifying the optimistic spinning code as a result. Performance measurements (locking throughput kops/s) using a locking microbenchmark with 50/50 reader/writer distribution and turbo-boost disabled was done on a 2-socket Cascade Lake system (48-core 96-thread) to see the impacts of these changes: 1) Vanilla - 5.10-rc3 kernel 2) Before - 5.10-rc3 kernel with previous patches in this series 2) limit-rspin - 5.10-rc3 kernel with limited reader spinning patch 3) no-rspin - 5.10-rc3 kernel with reader spinning disabled # of threads CS Load Vanilla Before limit-rspin no-rspin ------------ ------- ------- ------ ----------- -------- 2 1 5,185 5,662 5,214 5,077 4 1 5,107 4,983 5,188 4,760 8 1 4,782 4,564 4,720 4,628 16 1 4,680 4,053 4,567 3,402 32 1 4,299 1,115 1,118 1,098 64 1 3,218 983 1,001 957 96 1 1,938 944 957 930 2 20 2,008 2,128 2,264 1,665 4 20 1,390 1,033 1,046 1,101 8 20 1,472 1,155 1,098 1,213 16 20 1,332 1,077 1,089 1,122 32 20 967 914 917 980 64 20 787 874 891 858 96 20 730 836 847 844 2 100 372 356 360 355 4 100 492 425 434 392 8 100 533 537 529 538 16 100 548 572 568 598 32 100 499 520 527 537 64 100 466 517 526 512 96 100 406 497 506 509 The column "CS Load" represents the number of pause instructions issued in the locking critical section. A CS load of 1 is extremely short and is not likey in real situations. A load of 20 (moderate) and 100 (long) are more realistic. It can be seen that the previous patches in this series have reduced performance in general except in highly contended cases with moderate or long critical sections that performance improves a bit. This change is mostly caused by the "Prevent potential lock starvation" patch that reduce reader optimistic spinning and hence reduce reader fragmentation. The patch that further limit reader optimistic spinning doesn't seem to have too much impact on overall performance as shown in the benchmark data. The patch that disables reader optimistic spinning shows reduced performance at lightly loaded cases, but comparable or slightly better performance on with heavier contention. This patch just removes reader optimistic spinning for now. As readers are not going to do optimistic spinning anymore, we don't need to consider if the OSQ is empty or not when doing lock stealing. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-6-longman@redhat.com
2020-11-21 07:14:16 +03:00
* Clear the owner's RWSEM_NONSPINNABLE bit if it is set. This should
locking/rwsem: Enable time-based spinning on reader-owned rwsem When the rwsem is owned by reader, writers stop optimistic spinning simply because there is no easy way to figure out if all the readers are actively running or not. However, there are scenarios where the readers are unlikely to sleep and optimistic spinning can help performance. This patch provides a simple mechanism for spinning on a reader-owned rwsem by a writer. It is a time threshold based spinning where the allowable spinning time can vary from 10us to 25us depending on the condition of the rwsem. When the time threshold is exceeded, the nonspinnable bits will be set in the owner field to indicate that no more optimistic spinning will be allowed on this rwsem until it becomes writer owned again. Not even readers is allowed to acquire the reader-locked rwsem by optimistic spinning for fairness. We also want a writer to acquire the lock after the readers hold the lock for a relatively long time. In order to give preference to writers under such a circumstance, the single RWSEM_NONSPINNABLE bit is now split into two - one for reader and one for writer. When optimistic spinning is disabled, both bits will be set. When the reader count drop down to 0, the writer nonspinnable bit will be cleared to allow writers to spin on the lock, but not the readers. When a writer acquires the lock, it will write its own task structure pointer into sem->owner and clear the reader nonspinnable bit in the process. The time taken for each iteration of the reader-owned rwsem spinning loop varies. Below are sample minimum elapsed times for 16 iterations of the loop. System Time for 16 Iterations ------ ---------------------- 1-socket Skylake ~800ns 4-socket Broadwell ~300ns 2-socket ThunderX2 (arm64) ~250ns When the lock cacheline is contended, we can see up to almost 10X increase in elapsed time. So 25us will be at most 500, 1300 and 1600 iterations for each of the above systems. With a locking microbenchmark running on 5.1 based kernel, the total locking rates (in kops/s) on a 8-socket IvyBridge-EX system with equal numbers of readers and writers before and after this patch were as follows: # of Threads Pre-patch Post-patch ------------ --------- ---------- 2 1,759 6,684 4 1,684 6,738 8 1,074 7,222 16 900 7,163 32 458 7,316 64 208 520 128 168 425 240 143 474 This patch gives a big boost in performance for mixed reader/writer workloads. With 32 locking threads, the rwsem lock event data were: rwsem_opt_fail=79850 rwsem_opt_nospin=5069 rwsem_opt_rlock=597484 rwsem_opt_wlock=957339 rwsem_sleep_reader=57782 rwsem_sleep_writer=55663 With 64 locking threads, the data looked like: rwsem_opt_fail=346723 rwsem_opt_nospin=6293 rwsem_opt_rlock=1127119 rwsem_opt_wlock=1400628 rwsem_sleep_reader=308201 rwsem_sleep_writer=72281 So a lot more threads acquired the lock in the slowpath and more threads went to sleep. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-15-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:13 +03:00
* only be called when the reader count reaches 0.
locking/rwsem: Adaptive disabling of reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It makes readers relatively more preferred than writers. When a writer times out spinning on a reader-owned lock and set the nospinnable bits, there are two main reasons for that. 1) The reader critical section is long, perhaps the task sleeps after acquiring the read lock. 2) There are just too many readers contending the lock causing it to take a while to service all of them. In the former case, long reader critical section will impede the progress of writers which is usually more important for system performance. In the later case, reader optimistic spinning tends to make the reader groups that contain readers that acquire the lock together smaller leading to more of them. That may hurt performance in some cases. In other words, the setting of nonspinnable bits indicates that reader optimistic spinning may not be helpful for those workloads that cause it. Therefore, any writers that have observed the setting of the writer nonspinnable bit for a given rwsem after they fail to acquire the lock via optimistic spinning will set the reader nonspinnable bit once they acquire the write lock. Similarly, readers that observe the setting of reader nonspinnable bit at slowpath entry will also set the reader nonspinnable bit when they acquire the read lock via the wakeup path. Once the reader nonspinnable bit is on, it will only be reset when a writer is able to acquire the rwsem in the fast path or somehow a reader or writer in the slowpath doesn't observe the nonspinable bit. This is to discourage reader optmistic spinning on that particular rwsem and make writers more preferred. This adaptive disabling of reader optimistic spinning will alleviate some of the negative side effect of this feature. In addition, this patch tries to make readers in the spinning queue follow the phase-fair principle after quitting optimistic spinning by checking if another reader has somehow acquired a read lock after this reader enters the optimistic spinning queue. If so and the rwsem is still reader-owned, this reader is in the right read-phase and can attempt to acquire the lock. On a 2-socket 40-core 80-thread Skylake system, the page_fault1 test of the will-it-scale benchmark was run with various number of threads. The number of operations done before reader optimistic spinning patches, this patch and after this patch were: Threads Before rspin Before patch After patch %change ------- ------------ ------------ ----------- ------- 20 5541068 5345484 5455667 -3.5%/ +2.1% 40 10185150 7292313 9219276 -28.5%/+26.4% 60 8196733 6460517 7181209 -21.2%/+11.2% 80 9508864 6739559 8107025 -29.1%/+20.3% This patch doesn't recover all the lost performance, but it is more than half. Given the fact that reader optimistic spinning does benefit some workloads, this is a good compromise. Using the rwsem locking microbenchmark with very short critical section, this patch doesn't have too much impact on locking performance as shown by the locking rates (kops/s) below with equal numbers of readers and writers before and after this patch: # of Threads Pre-patch Post-patch ------------ --------- ---------- 2 4,730 4,969 4 4,814 4,786 8 4,866 4,815 16 4,715 4,511 32 3,338 3,500 64 3,212 3,389 80 3,110 3,044 When running the locking microbenchmark with 40 dedicated reader and writer threads, however, the reader performance is curtailed to favor the writer. Before patch: 40 readers, Iterations Min/Mean/Max = 204,026/234,309/254,816 40 writers, Iterations Min/Mean/Max = 88,515/95,884/115,644 After patch: 40 readers, Iterations Min/Mean/Max = 33,813/35,260/36,791 40 writers, Iterations Min/Mean/Max = 95,368/96,565/97,798 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-16-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:14 +03:00
*/
locking/rwsem: Remove reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It also improves the chance that a reader can get the lock as writer optimistic spinning disproportionally favors writers much more than readers. Since commit d3681e269fff ("locking/rwsem: Wake up almost all readers in wait queue"), all the waiting readers are woken up so that they can all get the read lock and run in parallel. When the number of contending readers is large, allowing reader optimistic spinning will likely cause reader fragmentation where multiple smaller groups of readers can get the read lock in a sequential manner separated by writers. That reduces reader parallelism. One possible way to address that drawback is to limit the number of readers (preferably one) that can do optimistic spinning. These readers act as representatives of all the waiting readers in the wait queue as they will wake up all those waiting readers once they get the lock. Alternatively, as reader optimistic lock stealing has already enhanced fairness to readers, it may be easier to just remove reader optimistic spinning and simplifying the optimistic spinning code as a result. Performance measurements (locking throughput kops/s) using a locking microbenchmark with 50/50 reader/writer distribution and turbo-boost disabled was done on a 2-socket Cascade Lake system (48-core 96-thread) to see the impacts of these changes: 1) Vanilla - 5.10-rc3 kernel 2) Before - 5.10-rc3 kernel with previous patches in this series 2) limit-rspin - 5.10-rc3 kernel with limited reader spinning patch 3) no-rspin - 5.10-rc3 kernel with reader spinning disabled # of threads CS Load Vanilla Before limit-rspin no-rspin ------------ ------- ------- ------ ----------- -------- 2 1 5,185 5,662 5,214 5,077 4 1 5,107 4,983 5,188 4,760 8 1 4,782 4,564 4,720 4,628 16 1 4,680 4,053 4,567 3,402 32 1 4,299 1,115 1,118 1,098 64 1 3,218 983 1,001 957 96 1 1,938 944 957 930 2 20 2,008 2,128 2,264 1,665 4 20 1,390 1,033 1,046 1,101 8 20 1,472 1,155 1,098 1,213 16 20 1,332 1,077 1,089 1,122 32 20 967 914 917 980 64 20 787 874 891 858 96 20 730 836 847 844 2 100 372 356 360 355 4 100 492 425 434 392 8 100 533 537 529 538 16 100 548 572 568 598 32 100 499 520 527 537 64 100 466 517 526 512 96 100 406 497 506 509 The column "CS Load" represents the number of pause instructions issued in the locking critical section. A CS load of 1 is extremely short and is not likey in real situations. A load of 20 (moderate) and 100 (long) are more realistic. It can be seen that the previous patches in this series have reduced performance in general except in highly contended cases with moderate or long critical sections that performance improves a bit. This change is mostly caused by the "Prevent potential lock starvation" patch that reduce reader optimistic spinning and hence reduce reader fragmentation. The patch that further limit reader optimistic spinning doesn't seem to have too much impact on overall performance as shown in the benchmark data. The patch that disables reader optimistic spinning shows reduced performance at lightly loaded cases, but comparable or slightly better performance on with heavier contention. This patch just removes reader optimistic spinning for now. As readers are not going to do optimistic spinning anymore, we don't need to consider if the OSQ is empty or not when doing lock stealing. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-6-longman@redhat.com
2020-11-21 07:14:16 +03:00
static inline void clear_nonspinnable(struct rw_semaphore *sem)
locking/rwsem: Adaptive disabling of reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It makes readers relatively more preferred than writers. When a writer times out spinning on a reader-owned lock and set the nospinnable bits, there are two main reasons for that. 1) The reader critical section is long, perhaps the task sleeps after acquiring the read lock. 2) There are just too many readers contending the lock causing it to take a while to service all of them. In the former case, long reader critical section will impede the progress of writers which is usually more important for system performance. In the later case, reader optimistic spinning tends to make the reader groups that contain readers that acquire the lock together smaller leading to more of them. That may hurt performance in some cases. In other words, the setting of nonspinnable bits indicates that reader optimistic spinning may not be helpful for those workloads that cause it. Therefore, any writers that have observed the setting of the writer nonspinnable bit for a given rwsem after they fail to acquire the lock via optimistic spinning will set the reader nonspinnable bit once they acquire the write lock. Similarly, readers that observe the setting of reader nonspinnable bit at slowpath entry will also set the reader nonspinnable bit when they acquire the read lock via the wakeup path. Once the reader nonspinnable bit is on, it will only be reset when a writer is able to acquire the rwsem in the fast path or somehow a reader or writer in the slowpath doesn't observe the nonspinable bit. This is to discourage reader optmistic spinning on that particular rwsem and make writers more preferred. This adaptive disabling of reader optimistic spinning will alleviate some of the negative side effect of this feature. In addition, this patch tries to make readers in the spinning queue follow the phase-fair principle after quitting optimistic spinning by checking if another reader has somehow acquired a read lock after this reader enters the optimistic spinning queue. If so and the rwsem is still reader-owned, this reader is in the right read-phase and can attempt to acquire the lock. On a 2-socket 40-core 80-thread Skylake system, the page_fault1 test of the will-it-scale benchmark was run with various number of threads. The number of operations done before reader optimistic spinning patches, this patch and after this patch were: Threads Before rspin Before patch After patch %change ------- ------------ ------------ ----------- ------- 20 5541068 5345484 5455667 -3.5%/ +2.1% 40 10185150 7292313 9219276 -28.5%/+26.4% 60 8196733 6460517 7181209 -21.2%/+11.2% 80 9508864 6739559 8107025 -29.1%/+20.3% This patch doesn't recover all the lost performance, but it is more than half. Given the fact that reader optimistic spinning does benefit some workloads, this is a good compromise. Using the rwsem locking microbenchmark with very short critical section, this patch doesn't have too much impact on locking performance as shown by the locking rates (kops/s) below with equal numbers of readers and writers before and after this patch: # of Threads Pre-patch Post-patch ------------ --------- ---------- 2 4,730 4,969 4 4,814 4,786 8 4,866 4,815 16 4,715 4,511 32 3,338 3,500 64 3,212 3,389 80 3,110 3,044 When running the locking microbenchmark with 40 dedicated reader and writer threads, however, the reader performance is curtailed to favor the writer. Before patch: 40 readers, Iterations Min/Mean/Max = 204,026/234,309/254,816 40 writers, Iterations Min/Mean/Max = 88,515/95,884/115,644 After patch: 40 readers, Iterations Min/Mean/Max = 33,813/35,260/36,791 40 writers, Iterations Min/Mean/Max = 95,368/96,565/97,798 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-16-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:14 +03:00
{
locking/rwsem: Remove reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It also improves the chance that a reader can get the lock as writer optimistic spinning disproportionally favors writers much more than readers. Since commit d3681e269fff ("locking/rwsem: Wake up almost all readers in wait queue"), all the waiting readers are woken up so that they can all get the read lock and run in parallel. When the number of contending readers is large, allowing reader optimistic spinning will likely cause reader fragmentation where multiple smaller groups of readers can get the read lock in a sequential manner separated by writers. That reduces reader parallelism. One possible way to address that drawback is to limit the number of readers (preferably one) that can do optimistic spinning. These readers act as representatives of all the waiting readers in the wait queue as they will wake up all those waiting readers once they get the lock. Alternatively, as reader optimistic lock stealing has already enhanced fairness to readers, it may be easier to just remove reader optimistic spinning and simplifying the optimistic spinning code as a result. Performance measurements (locking throughput kops/s) using a locking microbenchmark with 50/50 reader/writer distribution and turbo-boost disabled was done on a 2-socket Cascade Lake system (48-core 96-thread) to see the impacts of these changes: 1) Vanilla - 5.10-rc3 kernel 2) Before - 5.10-rc3 kernel with previous patches in this series 2) limit-rspin - 5.10-rc3 kernel with limited reader spinning patch 3) no-rspin - 5.10-rc3 kernel with reader spinning disabled # of threads CS Load Vanilla Before limit-rspin no-rspin ------------ ------- ------- ------ ----------- -------- 2 1 5,185 5,662 5,214 5,077 4 1 5,107 4,983 5,188 4,760 8 1 4,782 4,564 4,720 4,628 16 1 4,680 4,053 4,567 3,402 32 1 4,299 1,115 1,118 1,098 64 1 3,218 983 1,001 957 96 1 1,938 944 957 930 2 20 2,008 2,128 2,264 1,665 4 20 1,390 1,033 1,046 1,101 8 20 1,472 1,155 1,098 1,213 16 20 1,332 1,077 1,089 1,122 32 20 967 914 917 980 64 20 787 874 891 858 96 20 730 836 847 844 2 100 372 356 360 355 4 100 492 425 434 392 8 100 533 537 529 538 16 100 548 572 568 598 32 100 499 520 527 537 64 100 466 517 526 512 96 100 406 497 506 509 The column "CS Load" represents the number of pause instructions issued in the locking critical section. A CS load of 1 is extremely short and is not likey in real situations. A load of 20 (moderate) and 100 (long) are more realistic. It can be seen that the previous patches in this series have reduced performance in general except in highly contended cases with moderate or long critical sections that performance improves a bit. This change is mostly caused by the "Prevent potential lock starvation" patch that reduce reader optimistic spinning and hence reduce reader fragmentation. The patch that further limit reader optimistic spinning doesn't seem to have too much impact on overall performance as shown in the benchmark data. The patch that disables reader optimistic spinning shows reduced performance at lightly loaded cases, but comparable or slightly better performance on with heavier contention. This patch just removes reader optimistic spinning for now. As readers are not going to do optimistic spinning anymore, we don't need to consider if the OSQ is empty or not when doing lock stealing. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-6-longman@redhat.com
2020-11-21 07:14:16 +03:00
if (rwsem_test_oflags(sem, RWSEM_NONSPINNABLE))
atomic_long_andnot(RWSEM_NONSPINNABLE, &sem->owner);
}
#else
locking/rwsem: Remove reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It also improves the chance that a reader can get the lock as writer optimistic spinning disproportionally favors writers much more than readers. Since commit d3681e269fff ("locking/rwsem: Wake up almost all readers in wait queue"), all the waiting readers are woken up so that they can all get the read lock and run in parallel. When the number of contending readers is large, allowing reader optimistic spinning will likely cause reader fragmentation where multiple smaller groups of readers can get the read lock in a sequential manner separated by writers. That reduces reader parallelism. One possible way to address that drawback is to limit the number of readers (preferably one) that can do optimistic spinning. These readers act as representatives of all the waiting readers in the wait queue as they will wake up all those waiting readers once they get the lock. Alternatively, as reader optimistic lock stealing has already enhanced fairness to readers, it may be easier to just remove reader optimistic spinning and simplifying the optimistic spinning code as a result. Performance measurements (locking throughput kops/s) using a locking microbenchmark with 50/50 reader/writer distribution and turbo-boost disabled was done on a 2-socket Cascade Lake system (48-core 96-thread) to see the impacts of these changes: 1) Vanilla - 5.10-rc3 kernel 2) Before - 5.10-rc3 kernel with previous patches in this series 2) limit-rspin - 5.10-rc3 kernel with limited reader spinning patch 3) no-rspin - 5.10-rc3 kernel with reader spinning disabled # of threads CS Load Vanilla Before limit-rspin no-rspin ------------ ------- ------- ------ ----------- -------- 2 1 5,185 5,662 5,214 5,077 4 1 5,107 4,983 5,188 4,760 8 1 4,782 4,564 4,720 4,628 16 1 4,680 4,053 4,567 3,402 32 1 4,299 1,115 1,118 1,098 64 1 3,218 983 1,001 957 96 1 1,938 944 957 930 2 20 2,008 2,128 2,264 1,665 4 20 1,390 1,033 1,046 1,101 8 20 1,472 1,155 1,098 1,213 16 20 1,332 1,077 1,089 1,122 32 20 967 914 917 980 64 20 787 874 891 858 96 20 730 836 847 844 2 100 372 356 360 355 4 100 492 425 434 392 8 100 533 537 529 538 16 100 548 572 568 598 32 100 499 520 527 537 64 100 466 517 526 512 96 100 406 497 506 509 The column "CS Load" represents the number of pause instructions issued in the locking critical section. A CS load of 1 is extremely short and is not likey in real situations. A load of 20 (moderate) and 100 (long) are more realistic. It can be seen that the previous patches in this series have reduced performance in general except in highly contended cases with moderate or long critical sections that performance improves a bit. This change is mostly caused by the "Prevent potential lock starvation" patch that reduce reader optimistic spinning and hence reduce reader fragmentation. The patch that further limit reader optimistic spinning doesn't seem to have too much impact on overall performance as shown in the benchmark data. The patch that disables reader optimistic spinning shows reduced performance at lightly loaded cases, but comparable or slightly better performance on with heavier contention. This patch just removes reader optimistic spinning for now. As readers are not going to do optimistic spinning anymore, we don't need to consider if the OSQ is empty or not when doing lock stealing. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-6-longman@redhat.com
2020-11-21 07:14:16 +03:00
static inline bool rwsem_can_spin_on_owner(struct rw_semaphore *sem)
{
return false;
}
locking/rwsem: Remove reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It also improves the chance that a reader can get the lock as writer optimistic spinning disproportionally favors writers much more than readers. Since commit d3681e269fff ("locking/rwsem: Wake up almost all readers in wait queue"), all the waiting readers are woken up so that they can all get the read lock and run in parallel. When the number of contending readers is large, allowing reader optimistic spinning will likely cause reader fragmentation where multiple smaller groups of readers can get the read lock in a sequential manner separated by writers. That reduces reader parallelism. One possible way to address that drawback is to limit the number of readers (preferably one) that can do optimistic spinning. These readers act as representatives of all the waiting readers in the wait queue as they will wake up all those waiting readers once they get the lock. Alternatively, as reader optimistic lock stealing has already enhanced fairness to readers, it may be easier to just remove reader optimistic spinning and simplifying the optimistic spinning code as a result. Performance measurements (locking throughput kops/s) using a locking microbenchmark with 50/50 reader/writer distribution and turbo-boost disabled was done on a 2-socket Cascade Lake system (48-core 96-thread) to see the impacts of these changes: 1) Vanilla - 5.10-rc3 kernel 2) Before - 5.10-rc3 kernel with previous patches in this series 2) limit-rspin - 5.10-rc3 kernel with limited reader spinning patch 3) no-rspin - 5.10-rc3 kernel with reader spinning disabled # of threads CS Load Vanilla Before limit-rspin no-rspin ------------ ------- ------- ------ ----------- -------- 2 1 5,185 5,662 5,214 5,077 4 1 5,107 4,983 5,188 4,760 8 1 4,782 4,564 4,720 4,628 16 1 4,680 4,053 4,567 3,402 32 1 4,299 1,115 1,118 1,098 64 1 3,218 983 1,001 957 96 1 1,938 944 957 930 2 20 2,008 2,128 2,264 1,665 4 20 1,390 1,033 1,046 1,101 8 20 1,472 1,155 1,098 1,213 16 20 1,332 1,077 1,089 1,122 32 20 967 914 917 980 64 20 787 874 891 858 96 20 730 836 847 844 2 100 372 356 360 355 4 100 492 425 434 392 8 100 533 537 529 538 16 100 548 572 568 598 32 100 499 520 527 537 64 100 466 517 526 512 96 100 406 497 506 509 The column "CS Load" represents the number of pause instructions issued in the locking critical section. A CS load of 1 is extremely short and is not likey in real situations. A load of 20 (moderate) and 100 (long) are more realistic. It can be seen that the previous patches in this series have reduced performance in general except in highly contended cases with moderate or long critical sections that performance improves a bit. This change is mostly caused by the "Prevent potential lock starvation" patch that reduce reader optimistic spinning and hence reduce reader fragmentation. The patch that further limit reader optimistic spinning doesn't seem to have too much impact on overall performance as shown in the benchmark data. The patch that disables reader optimistic spinning shows reduced performance at lightly loaded cases, but comparable or slightly better performance on with heavier contention. This patch just removes reader optimistic spinning for now. As readers are not going to do optimistic spinning anymore, we don't need to consider if the OSQ is empty or not when doing lock stealing. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-6-longman@redhat.com
2020-11-21 07:14:16 +03:00
static inline bool rwsem_optimistic_spin(struct rw_semaphore *sem)
{
return false;
}
locking/rwsem: Enable time-based spinning on reader-owned rwsem When the rwsem is owned by reader, writers stop optimistic spinning simply because there is no easy way to figure out if all the readers are actively running or not. However, there are scenarios where the readers are unlikely to sleep and optimistic spinning can help performance. This patch provides a simple mechanism for spinning on a reader-owned rwsem by a writer. It is a time threshold based spinning where the allowable spinning time can vary from 10us to 25us depending on the condition of the rwsem. When the time threshold is exceeded, the nonspinnable bits will be set in the owner field to indicate that no more optimistic spinning will be allowed on this rwsem until it becomes writer owned again. Not even readers is allowed to acquire the reader-locked rwsem by optimistic spinning for fairness. We also want a writer to acquire the lock after the readers hold the lock for a relatively long time. In order to give preference to writers under such a circumstance, the single RWSEM_NONSPINNABLE bit is now split into two - one for reader and one for writer. When optimistic spinning is disabled, both bits will be set. When the reader count drop down to 0, the writer nonspinnable bit will be cleared to allow writers to spin on the lock, but not the readers. When a writer acquires the lock, it will write its own task structure pointer into sem->owner and clear the reader nonspinnable bit in the process. The time taken for each iteration of the reader-owned rwsem spinning loop varies. Below are sample minimum elapsed times for 16 iterations of the loop. System Time for 16 Iterations ------ ---------------------- 1-socket Skylake ~800ns 4-socket Broadwell ~300ns 2-socket ThunderX2 (arm64) ~250ns When the lock cacheline is contended, we can see up to almost 10X increase in elapsed time. So 25us will be at most 500, 1300 and 1600 iterations for each of the above systems. With a locking microbenchmark running on 5.1 based kernel, the total locking rates (in kops/s) on a 8-socket IvyBridge-EX system with equal numbers of readers and writers before and after this patch were as follows: # of Threads Pre-patch Post-patch ------------ --------- ---------- 2 1,759 6,684 4 1,684 6,738 8 1,074 7,222 16 900 7,163 32 458 7,316 64 208 520 128 168 425 240 143 474 This patch gives a big boost in performance for mixed reader/writer workloads. With 32 locking threads, the rwsem lock event data were: rwsem_opt_fail=79850 rwsem_opt_nospin=5069 rwsem_opt_rlock=597484 rwsem_opt_wlock=957339 rwsem_sleep_reader=57782 rwsem_sleep_writer=55663 With 64 locking threads, the data looked like: rwsem_opt_fail=346723 rwsem_opt_nospin=6293 rwsem_opt_rlock=1127119 rwsem_opt_wlock=1400628 rwsem_sleep_reader=308201 rwsem_sleep_writer=72281 So a lot more threads acquired the lock in the slowpath and more threads went to sleep. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-15-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:13 +03:00
locking/rwsem: Remove reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It also improves the chance that a reader can get the lock as writer optimistic spinning disproportionally favors writers much more than readers. Since commit d3681e269fff ("locking/rwsem: Wake up almost all readers in wait queue"), all the waiting readers are woken up so that they can all get the read lock and run in parallel. When the number of contending readers is large, allowing reader optimistic spinning will likely cause reader fragmentation where multiple smaller groups of readers can get the read lock in a sequential manner separated by writers. That reduces reader parallelism. One possible way to address that drawback is to limit the number of readers (preferably one) that can do optimistic spinning. These readers act as representatives of all the waiting readers in the wait queue as they will wake up all those waiting readers once they get the lock. Alternatively, as reader optimistic lock stealing has already enhanced fairness to readers, it may be easier to just remove reader optimistic spinning and simplifying the optimistic spinning code as a result. Performance measurements (locking throughput kops/s) using a locking microbenchmark with 50/50 reader/writer distribution and turbo-boost disabled was done on a 2-socket Cascade Lake system (48-core 96-thread) to see the impacts of these changes: 1) Vanilla - 5.10-rc3 kernel 2) Before - 5.10-rc3 kernel with previous patches in this series 2) limit-rspin - 5.10-rc3 kernel with limited reader spinning patch 3) no-rspin - 5.10-rc3 kernel with reader spinning disabled # of threads CS Load Vanilla Before limit-rspin no-rspin ------------ ------- ------- ------ ----------- -------- 2 1 5,185 5,662 5,214 5,077 4 1 5,107 4,983 5,188 4,760 8 1 4,782 4,564 4,720 4,628 16 1 4,680 4,053 4,567 3,402 32 1 4,299 1,115 1,118 1,098 64 1 3,218 983 1,001 957 96 1 1,938 944 957 930 2 20 2,008 2,128 2,264 1,665 4 20 1,390 1,033 1,046 1,101 8 20 1,472 1,155 1,098 1,213 16 20 1,332 1,077 1,089 1,122 32 20 967 914 917 980 64 20 787 874 891 858 96 20 730 836 847 844 2 100 372 356 360 355 4 100 492 425 434 392 8 100 533 537 529 538 16 100 548 572 568 598 32 100 499 520 527 537 64 100 466 517 526 512 96 100 406 497 506 509 The column "CS Load" represents the number of pause instructions issued in the locking critical section. A CS load of 1 is extremely short and is not likey in real situations. A load of 20 (moderate) and 100 (long) are more realistic. It can be seen that the previous patches in this series have reduced performance in general except in highly contended cases with moderate or long critical sections that performance improves a bit. This change is mostly caused by the "Prevent potential lock starvation" patch that reduce reader optimistic spinning and hence reduce reader fragmentation. The patch that further limit reader optimistic spinning doesn't seem to have too much impact on overall performance as shown in the benchmark data. The patch that disables reader optimistic spinning shows reduced performance at lightly loaded cases, but comparable or slightly better performance on with heavier contention. This patch just removes reader optimistic spinning for now. As readers are not going to do optimistic spinning anymore, we don't need to consider if the OSQ is empty or not when doing lock stealing. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-6-longman@redhat.com
2020-11-21 07:14:16 +03:00
static inline void clear_nonspinnable(struct rw_semaphore *sem) { }
locking/rwsem: Make handoff writer optimistically spin on owner When the handoff bit is set by a writer, no other tasks other than the setting writer itself is allowed to acquire the lock. If the to-be-handoff'ed writer goes to sleep, there will be a wakeup latency period where the lock is free, but no one can acquire it. That is less than ideal. To reduce that latency, the handoff writer will now optimistically spin on the owner if it happens to be a on-cpu writer. It will spin until it releases the lock and the to-be-handoff'ed writer can then acquire the lock immediately without any delay. Of course, if the owner is not a on-cpu writer, the to-be-handoff'ed writer will have to sleep anyway. The optimistic spinning code is also modified to not stop spinning when the handoff bit is set. This will prevent an occasional setting of handoff bit from causing a bunch of optimistic spinners from entering into the wait queue causing significant reduction in throughput. On a 1-socket 22-core 44-thread Skylake system, the AIM7 shared_memory workload was run with 7000 users. The throughput (jobs/min) of the following kernels were as follows: 1) 5.2-rc6 - 8,092,486 2) 5.2-rc6 + tip's rwsem patches - 7,567,568 3) 5.2-rc6 + tip's rwsem patches + this patch - 7,954,545 Using perf-record(1), the %cpu time used by rwsem_down_write_slowpath(), rwsem_down_write_failed() and their callees for the 3 kernels were 1.70%, 5.46% and 2.08% respectively. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: x86@kernel.org Cc: Ingo Molnar <mingo@redhat.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: "H. Peter Anvin" <hpa@zytor.com> Link: https://lkml.kernel.org/r/20190625143913.24154-1-longman@redhat.com
2019-06-25 17:39:13 +03:00
static inline int
locking/rwsem: Remove reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It also improves the chance that a reader can get the lock as writer optimistic spinning disproportionally favors writers much more than readers. Since commit d3681e269fff ("locking/rwsem: Wake up almost all readers in wait queue"), all the waiting readers are woken up so that they can all get the read lock and run in parallel. When the number of contending readers is large, allowing reader optimistic spinning will likely cause reader fragmentation where multiple smaller groups of readers can get the read lock in a sequential manner separated by writers. That reduces reader parallelism. One possible way to address that drawback is to limit the number of readers (preferably one) that can do optimistic spinning. These readers act as representatives of all the waiting readers in the wait queue as they will wake up all those waiting readers once they get the lock. Alternatively, as reader optimistic lock stealing has already enhanced fairness to readers, it may be easier to just remove reader optimistic spinning and simplifying the optimistic spinning code as a result. Performance measurements (locking throughput kops/s) using a locking microbenchmark with 50/50 reader/writer distribution and turbo-boost disabled was done on a 2-socket Cascade Lake system (48-core 96-thread) to see the impacts of these changes: 1) Vanilla - 5.10-rc3 kernel 2) Before - 5.10-rc3 kernel with previous patches in this series 2) limit-rspin - 5.10-rc3 kernel with limited reader spinning patch 3) no-rspin - 5.10-rc3 kernel with reader spinning disabled # of threads CS Load Vanilla Before limit-rspin no-rspin ------------ ------- ------- ------ ----------- -------- 2 1 5,185 5,662 5,214 5,077 4 1 5,107 4,983 5,188 4,760 8 1 4,782 4,564 4,720 4,628 16 1 4,680 4,053 4,567 3,402 32 1 4,299 1,115 1,118 1,098 64 1 3,218 983 1,001 957 96 1 1,938 944 957 930 2 20 2,008 2,128 2,264 1,665 4 20 1,390 1,033 1,046 1,101 8 20 1,472 1,155 1,098 1,213 16 20 1,332 1,077 1,089 1,122 32 20 967 914 917 980 64 20 787 874 891 858 96 20 730 836 847 844 2 100 372 356 360 355 4 100 492 425 434 392 8 100 533 537 529 538 16 100 548 572 568 598 32 100 499 520 527 537 64 100 466 517 526 512 96 100 406 497 506 509 The column "CS Load" represents the number of pause instructions issued in the locking critical section. A CS load of 1 is extremely short and is not likey in real situations. A load of 20 (moderate) and 100 (long) are more realistic. It can be seen that the previous patches in this series have reduced performance in general except in highly contended cases with moderate or long critical sections that performance improves a bit. This change is mostly caused by the "Prevent potential lock starvation" patch that reduce reader optimistic spinning and hence reduce reader fragmentation. The patch that further limit reader optimistic spinning doesn't seem to have too much impact on overall performance as shown in the benchmark data. The patch that disables reader optimistic spinning shows reduced performance at lightly loaded cases, but comparable or slightly better performance on with heavier contention. This patch just removes reader optimistic spinning for now. As readers are not going to do optimistic spinning anymore, we don't need to consider if the OSQ is empty or not when doing lock stealing. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-6-longman@redhat.com
2020-11-21 07:14:16 +03:00
rwsem_spin_on_owner(struct rw_semaphore *sem)
locking/rwsem: Make handoff writer optimistically spin on owner When the handoff bit is set by a writer, no other tasks other than the setting writer itself is allowed to acquire the lock. If the to-be-handoff'ed writer goes to sleep, there will be a wakeup latency period where the lock is free, but no one can acquire it. That is less than ideal. To reduce that latency, the handoff writer will now optimistically spin on the owner if it happens to be a on-cpu writer. It will spin until it releases the lock and the to-be-handoff'ed writer can then acquire the lock immediately without any delay. Of course, if the owner is not a on-cpu writer, the to-be-handoff'ed writer will have to sleep anyway. The optimistic spinning code is also modified to not stop spinning when the handoff bit is set. This will prevent an occasional setting of handoff bit from causing a bunch of optimistic spinners from entering into the wait queue causing significant reduction in throughput. On a 1-socket 22-core 44-thread Skylake system, the AIM7 shared_memory workload was run with 7000 users. The throughput (jobs/min) of the following kernels were as follows: 1) 5.2-rc6 - 8,092,486 2) 5.2-rc6 + tip's rwsem patches - 7,567,568 3) 5.2-rc6 + tip's rwsem patches + this patch - 7,954,545 Using perf-record(1), the %cpu time used by rwsem_down_write_slowpath(), rwsem_down_write_failed() and their callees for the 3 kernels were 1.70%, 5.46% and 2.08% respectively. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: x86@kernel.org Cc: Ingo Molnar <mingo@redhat.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: "H. Peter Anvin" <hpa@zytor.com> Link: https://lkml.kernel.org/r/20190625143913.24154-1-longman@redhat.com
2019-06-25 17:39:13 +03:00
{
return 0;
}
#define OWNER_NULL 1
#endif
/*
* Wait for the read lock to be granted
*/
locking/rwsem: Code cleanup after files merging After merging all the relevant rwsem code into one single file, there are a number of optimizations and cleanups that can be done: 1) Remove all the EXPORT_SYMBOL() calls for functions that are not accessed elsewhere. 2) Remove all the __visible tags as none of the functions will be called from assembly code anymore. 3) Make all the internal functions static. 4) Remove some unneeded blank lines. 5) Remove the intermediate rwsem_down_{read|write}_failed*() functions and rename __rwsem_down_{read|write}_failed_common() to rwsem_down_{read|write}_slowpath(). 6) Remove "__" prefix of __rwsem_mark_wake(). 7) Use atomic_long_try_cmpxchg_acquire() as much as possible. 8) Remove the rwsem_rtrylock and rwsem_wtrylock lock events as they are not that useful. That enables the compiler to do better optimization and reduce code size. The text+data size of rwsem.o on an x86-64 machine with gcc8 was reduced from 10237 bytes to 5030 bytes with this change. Suggested-by: Peter Zijlstra <peterz@infradead.org> Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-6-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:04 +03:00
static struct rw_semaphore __sched *
rwsem_down_read_slowpath(struct rw_semaphore *sem, long count, unsigned int state)
{
locking/rwsem: Remove reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It also improves the chance that a reader can get the lock as writer optimistic spinning disproportionally favors writers much more than readers. Since commit d3681e269fff ("locking/rwsem: Wake up almost all readers in wait queue"), all the waiting readers are woken up so that they can all get the read lock and run in parallel. When the number of contending readers is large, allowing reader optimistic spinning will likely cause reader fragmentation where multiple smaller groups of readers can get the read lock in a sequential manner separated by writers. That reduces reader parallelism. One possible way to address that drawback is to limit the number of readers (preferably one) that can do optimistic spinning. These readers act as representatives of all the waiting readers in the wait queue as they will wake up all those waiting readers once they get the lock. Alternatively, as reader optimistic lock stealing has already enhanced fairness to readers, it may be easier to just remove reader optimistic spinning and simplifying the optimistic spinning code as a result. Performance measurements (locking throughput kops/s) using a locking microbenchmark with 50/50 reader/writer distribution and turbo-boost disabled was done on a 2-socket Cascade Lake system (48-core 96-thread) to see the impacts of these changes: 1) Vanilla - 5.10-rc3 kernel 2) Before - 5.10-rc3 kernel with previous patches in this series 2) limit-rspin - 5.10-rc3 kernel with limited reader spinning patch 3) no-rspin - 5.10-rc3 kernel with reader spinning disabled # of threads CS Load Vanilla Before limit-rspin no-rspin ------------ ------- ------- ------ ----------- -------- 2 1 5,185 5,662 5,214 5,077 4 1 5,107 4,983 5,188 4,760 8 1 4,782 4,564 4,720 4,628 16 1 4,680 4,053 4,567 3,402 32 1 4,299 1,115 1,118 1,098 64 1 3,218 983 1,001 957 96 1 1,938 944 957 930 2 20 2,008 2,128 2,264 1,665 4 20 1,390 1,033 1,046 1,101 8 20 1,472 1,155 1,098 1,213 16 20 1,332 1,077 1,089 1,122 32 20 967 914 917 980 64 20 787 874 891 858 96 20 730 836 847 844 2 100 372 356 360 355 4 100 492 425 434 392 8 100 533 537 529 538 16 100 548 572 568 598 32 100 499 520 527 537 64 100 466 517 526 512 96 100 406 497 506 509 The column "CS Load" represents the number of pause instructions issued in the locking critical section. A CS load of 1 is extremely short and is not likey in real situations. A load of 20 (moderate) and 100 (long) are more realistic. It can be seen that the previous patches in this series have reduced performance in general except in highly contended cases with moderate or long critical sections that performance improves a bit. This change is mostly caused by the "Prevent potential lock starvation" patch that reduce reader optimistic spinning and hence reduce reader fragmentation. The patch that further limit reader optimistic spinning doesn't seem to have too much impact on overall performance as shown in the benchmark data. The patch that disables reader optimistic spinning shows reduced performance at lightly loaded cases, but comparable or slightly better performance on with heavier contention. This patch just removes reader optimistic spinning for now. As readers are not going to do optimistic spinning anymore, we don't need to consider if the OSQ is empty or not when doing lock stealing. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-6-longman@redhat.com
2020-11-21 07:14:16 +03:00
long adjustment = -RWSEM_READER_BIAS;
locking/rwsem: Prevent potential lock starvation The lock handoff bit is added in commit 4f23dbc1e657 ("locking/rwsem: Implement lock handoff to prevent lock starvation") to avoid lock starvation. However, allowing readers to do optimistic spinning does introduce an unlikely scenario where lock starvation can happen. The lock handoff bit may only be set when a waiter is being woken up. In the case of reader unlock, wakeup happens only when the reader count reaches 0. If there is a continuous stream of incoming readers acquiring read lock via optimistic spinning, it is possible that the reader count may never reach 0 and so the handoff bit will never be asserted. One way to prevent this scenario from happening is to disallow optimistic spinning if the rwsem is currently owned by readers. If the previous or current owner is a writer, optimistic spinning will be allowed. If the previous owner is a reader but the reader count has reached 0 before, a wakeup should have been issued. So the handoff mechanism will be kicked in to prevent lock starvation. As a result, it should be OK to do optimistic spinning in this case. This patch may have some impact on reader performance as it reduces reader optimistic spinning especially if the lock critical sections are short the number of contending readers are small. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-3-longman@redhat.com
2020-11-21 07:14:13 +03:00
long rcnt = (count >> RWSEM_READER_SHIFT);
struct rwsem_waiter waiter;
DEFINE_WAKE_Q(wake_q);
bool wake = false;
locking/rwsem: Prevent potential lock starvation The lock handoff bit is added in commit 4f23dbc1e657 ("locking/rwsem: Implement lock handoff to prevent lock starvation") to avoid lock starvation. However, allowing readers to do optimistic spinning does introduce an unlikely scenario where lock starvation can happen. The lock handoff bit may only be set when a waiter is being woken up. In the case of reader unlock, wakeup happens only when the reader count reaches 0. If there is a continuous stream of incoming readers acquiring read lock via optimistic spinning, it is possible that the reader count may never reach 0 and so the handoff bit will never be asserted. One way to prevent this scenario from happening is to disallow optimistic spinning if the rwsem is currently owned by readers. If the previous or current owner is a writer, optimistic spinning will be allowed. If the previous owner is a reader but the reader count has reached 0 before, a wakeup should have been issued. So the handoff mechanism will be kicked in to prevent lock starvation. As a result, it should be OK to do optimistic spinning in this case. This patch may have some impact on reader performance as it reduces reader optimistic spinning especially if the lock critical sections are short the number of contending readers are small. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-3-longman@redhat.com
2020-11-21 07:14:13 +03:00
/*
* To prevent a constant stream of readers from starving a sleeping
locking/rwsem: Remove reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It also improves the chance that a reader can get the lock as writer optimistic spinning disproportionally favors writers much more than readers. Since commit d3681e269fff ("locking/rwsem: Wake up almost all readers in wait queue"), all the waiting readers are woken up so that they can all get the read lock and run in parallel. When the number of contending readers is large, allowing reader optimistic spinning will likely cause reader fragmentation where multiple smaller groups of readers can get the read lock in a sequential manner separated by writers. That reduces reader parallelism. One possible way to address that drawback is to limit the number of readers (preferably one) that can do optimistic spinning. These readers act as representatives of all the waiting readers in the wait queue as they will wake up all those waiting readers once they get the lock. Alternatively, as reader optimistic lock stealing has already enhanced fairness to readers, it may be easier to just remove reader optimistic spinning and simplifying the optimistic spinning code as a result. Performance measurements (locking throughput kops/s) using a locking microbenchmark with 50/50 reader/writer distribution and turbo-boost disabled was done on a 2-socket Cascade Lake system (48-core 96-thread) to see the impacts of these changes: 1) Vanilla - 5.10-rc3 kernel 2) Before - 5.10-rc3 kernel with previous patches in this series 2) limit-rspin - 5.10-rc3 kernel with limited reader spinning patch 3) no-rspin - 5.10-rc3 kernel with reader spinning disabled # of threads CS Load Vanilla Before limit-rspin no-rspin ------------ ------- ------- ------ ----------- -------- 2 1 5,185 5,662 5,214 5,077 4 1 5,107 4,983 5,188 4,760 8 1 4,782 4,564 4,720 4,628 16 1 4,680 4,053 4,567 3,402 32 1 4,299 1,115 1,118 1,098 64 1 3,218 983 1,001 957 96 1 1,938 944 957 930 2 20 2,008 2,128 2,264 1,665 4 20 1,390 1,033 1,046 1,101 8 20 1,472 1,155 1,098 1,213 16 20 1,332 1,077 1,089 1,122 32 20 967 914 917 980 64 20 787 874 891 858 96 20 730 836 847 844 2 100 372 356 360 355 4 100 492 425 434 392 8 100 533 537 529 538 16 100 548 572 568 598 32 100 499 520 527 537 64 100 466 517 526 512 96 100 406 497 506 509 The column "CS Load" represents the number of pause instructions issued in the locking critical section. A CS load of 1 is extremely short and is not likey in real situations. A load of 20 (moderate) and 100 (long) are more realistic. It can be seen that the previous patches in this series have reduced performance in general except in highly contended cases with moderate or long critical sections that performance improves a bit. This change is mostly caused by the "Prevent potential lock starvation" patch that reduce reader optimistic spinning and hence reduce reader fragmentation. The patch that further limit reader optimistic spinning doesn't seem to have too much impact on overall performance as shown in the benchmark data. The patch that disables reader optimistic spinning shows reduced performance at lightly loaded cases, but comparable or slightly better performance on with heavier contention. This patch just removes reader optimistic spinning for now. As readers are not going to do optimistic spinning anymore, we don't need to consider if the OSQ is empty or not when doing lock stealing. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-6-longman@redhat.com
2020-11-21 07:14:16 +03:00
* waiter, don't attempt optimistic lock stealing if the lock is
* currently owned by readers.
locking/rwsem: Prevent potential lock starvation The lock handoff bit is added in commit 4f23dbc1e657 ("locking/rwsem: Implement lock handoff to prevent lock starvation") to avoid lock starvation. However, allowing readers to do optimistic spinning does introduce an unlikely scenario where lock starvation can happen. The lock handoff bit may only be set when a waiter is being woken up. In the case of reader unlock, wakeup happens only when the reader count reaches 0. If there is a continuous stream of incoming readers acquiring read lock via optimistic spinning, it is possible that the reader count may never reach 0 and so the handoff bit will never be asserted. One way to prevent this scenario from happening is to disallow optimistic spinning if the rwsem is currently owned by readers. If the previous or current owner is a writer, optimistic spinning will be allowed. If the previous owner is a reader but the reader count has reached 0 before, a wakeup should have been issued. So the handoff mechanism will be kicked in to prevent lock starvation. As a result, it should be OK to do optimistic spinning in this case. This patch may have some impact on reader performance as it reduces reader optimistic spinning especially if the lock critical sections are short the number of contending readers are small. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-3-longman@redhat.com
2020-11-21 07:14:13 +03:00
*/
locking/rwsem: Remove reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It also improves the chance that a reader can get the lock as writer optimistic spinning disproportionally favors writers much more than readers. Since commit d3681e269fff ("locking/rwsem: Wake up almost all readers in wait queue"), all the waiting readers are woken up so that they can all get the read lock and run in parallel. When the number of contending readers is large, allowing reader optimistic spinning will likely cause reader fragmentation where multiple smaller groups of readers can get the read lock in a sequential manner separated by writers. That reduces reader parallelism. One possible way to address that drawback is to limit the number of readers (preferably one) that can do optimistic spinning. These readers act as representatives of all the waiting readers in the wait queue as they will wake up all those waiting readers once they get the lock. Alternatively, as reader optimistic lock stealing has already enhanced fairness to readers, it may be easier to just remove reader optimistic spinning and simplifying the optimistic spinning code as a result. Performance measurements (locking throughput kops/s) using a locking microbenchmark with 50/50 reader/writer distribution and turbo-boost disabled was done on a 2-socket Cascade Lake system (48-core 96-thread) to see the impacts of these changes: 1) Vanilla - 5.10-rc3 kernel 2) Before - 5.10-rc3 kernel with previous patches in this series 2) limit-rspin - 5.10-rc3 kernel with limited reader spinning patch 3) no-rspin - 5.10-rc3 kernel with reader spinning disabled # of threads CS Load Vanilla Before limit-rspin no-rspin ------------ ------- ------- ------ ----------- -------- 2 1 5,185 5,662 5,214 5,077 4 1 5,107 4,983 5,188 4,760 8 1 4,782 4,564 4,720 4,628 16 1 4,680 4,053 4,567 3,402 32 1 4,299 1,115 1,118 1,098 64 1 3,218 983 1,001 957 96 1 1,938 944 957 930 2 20 2,008 2,128 2,264 1,665 4 20 1,390 1,033 1,046 1,101 8 20 1,472 1,155 1,098 1,213 16 20 1,332 1,077 1,089 1,122 32 20 967 914 917 980 64 20 787 874 891 858 96 20 730 836 847 844 2 100 372 356 360 355 4 100 492 425 434 392 8 100 533 537 529 538 16 100 548 572 568 598 32 100 499 520 527 537 64 100 466 517 526 512 96 100 406 497 506 509 The column "CS Load" represents the number of pause instructions issued in the locking critical section. A CS load of 1 is extremely short and is not likey in real situations. A load of 20 (moderate) and 100 (long) are more realistic. It can be seen that the previous patches in this series have reduced performance in general except in highly contended cases with moderate or long critical sections that performance improves a bit. This change is mostly caused by the "Prevent potential lock starvation" patch that reduce reader optimistic spinning and hence reduce reader fragmentation. The patch that further limit reader optimistic spinning doesn't seem to have too much impact on overall performance as shown in the benchmark data. The patch that disables reader optimistic spinning shows reduced performance at lightly loaded cases, but comparable or slightly better performance on with heavier contention. This patch just removes reader optimistic spinning for now. As readers are not going to do optimistic spinning anymore, we don't need to consider if the OSQ is empty or not when doing lock stealing. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-6-longman@redhat.com
2020-11-21 07:14:16 +03:00
if ((atomic_long_read(&sem->owner) & RWSEM_READER_OWNED) &&
(rcnt > 1) && !(count & RWSEM_WRITER_LOCKED))
locking/rwsem: Prevent potential lock starvation The lock handoff bit is added in commit 4f23dbc1e657 ("locking/rwsem: Implement lock handoff to prevent lock starvation") to avoid lock starvation. However, allowing readers to do optimistic spinning does introduce an unlikely scenario where lock starvation can happen. The lock handoff bit may only be set when a waiter is being woken up. In the case of reader unlock, wakeup happens only when the reader count reaches 0. If there is a continuous stream of incoming readers acquiring read lock via optimistic spinning, it is possible that the reader count may never reach 0 and so the handoff bit will never be asserted. One way to prevent this scenario from happening is to disallow optimistic spinning if the rwsem is currently owned by readers. If the previous or current owner is a writer, optimistic spinning will be allowed. If the previous owner is a reader but the reader count has reached 0 before, a wakeup should have been issued. So the handoff mechanism will be kicked in to prevent lock starvation. As a result, it should be OK to do optimistic spinning in this case. This patch may have some impact on reader performance as it reduces reader optimistic spinning especially if the lock critical sections are short the number of contending readers are small. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-3-longman@redhat.com
2020-11-21 07:14:13 +03:00
goto queue;
/*
locking/rwsem: Remove reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It also improves the chance that a reader can get the lock as writer optimistic spinning disproportionally favors writers much more than readers. Since commit d3681e269fff ("locking/rwsem: Wake up almost all readers in wait queue"), all the waiting readers are woken up so that they can all get the read lock and run in parallel. When the number of contending readers is large, allowing reader optimistic spinning will likely cause reader fragmentation where multiple smaller groups of readers can get the read lock in a sequential manner separated by writers. That reduces reader parallelism. One possible way to address that drawback is to limit the number of readers (preferably one) that can do optimistic spinning. These readers act as representatives of all the waiting readers in the wait queue as they will wake up all those waiting readers once they get the lock. Alternatively, as reader optimistic lock stealing has already enhanced fairness to readers, it may be easier to just remove reader optimistic spinning and simplifying the optimistic spinning code as a result. Performance measurements (locking throughput kops/s) using a locking microbenchmark with 50/50 reader/writer distribution and turbo-boost disabled was done on a 2-socket Cascade Lake system (48-core 96-thread) to see the impacts of these changes: 1) Vanilla - 5.10-rc3 kernel 2) Before - 5.10-rc3 kernel with previous patches in this series 2) limit-rspin - 5.10-rc3 kernel with limited reader spinning patch 3) no-rspin - 5.10-rc3 kernel with reader spinning disabled # of threads CS Load Vanilla Before limit-rspin no-rspin ------------ ------- ------- ------ ----------- -------- 2 1 5,185 5,662 5,214 5,077 4 1 5,107 4,983 5,188 4,760 8 1 4,782 4,564 4,720 4,628 16 1 4,680 4,053 4,567 3,402 32 1 4,299 1,115 1,118 1,098 64 1 3,218 983 1,001 957 96 1 1,938 944 957 930 2 20 2,008 2,128 2,264 1,665 4 20 1,390 1,033 1,046 1,101 8 20 1,472 1,155 1,098 1,213 16 20 1,332 1,077 1,089 1,122 32 20 967 914 917 980 64 20 787 874 891 858 96 20 730 836 847 844 2 100 372 356 360 355 4 100 492 425 434 392 8 100 533 537 529 538 16 100 548 572 568 598 32 100 499 520 527 537 64 100 466 517 526 512 96 100 406 497 506 509 The column "CS Load" represents the number of pause instructions issued in the locking critical section. A CS load of 1 is extremely short and is not likey in real situations. A load of 20 (moderate) and 100 (long) are more realistic. It can be seen that the previous patches in this series have reduced performance in general except in highly contended cases with moderate or long critical sections that performance improves a bit. This change is mostly caused by the "Prevent potential lock starvation" patch that reduce reader optimistic spinning and hence reduce reader fragmentation. The patch that further limit reader optimistic spinning doesn't seem to have too much impact on overall performance as shown in the benchmark data. The patch that disables reader optimistic spinning shows reduced performance at lightly loaded cases, but comparable or slightly better performance on with heavier contention. This patch just removes reader optimistic spinning for now. As readers are not going to do optimistic spinning anymore, we don't need to consider if the OSQ is empty or not when doing lock stealing. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-6-longman@redhat.com
2020-11-21 07:14:16 +03:00
* Reader optimistic lock stealing.
*/
locking/rwsem: Remove reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It also improves the chance that a reader can get the lock as writer optimistic spinning disproportionally favors writers much more than readers. Since commit d3681e269fff ("locking/rwsem: Wake up almost all readers in wait queue"), all the waiting readers are woken up so that they can all get the read lock and run in parallel. When the number of contending readers is large, allowing reader optimistic spinning will likely cause reader fragmentation where multiple smaller groups of readers can get the read lock in a sequential manner separated by writers. That reduces reader parallelism. One possible way to address that drawback is to limit the number of readers (preferably one) that can do optimistic spinning. These readers act as representatives of all the waiting readers in the wait queue as they will wake up all those waiting readers once they get the lock. Alternatively, as reader optimistic lock stealing has already enhanced fairness to readers, it may be easier to just remove reader optimistic spinning and simplifying the optimistic spinning code as a result. Performance measurements (locking throughput kops/s) using a locking microbenchmark with 50/50 reader/writer distribution and turbo-boost disabled was done on a 2-socket Cascade Lake system (48-core 96-thread) to see the impacts of these changes: 1) Vanilla - 5.10-rc3 kernel 2) Before - 5.10-rc3 kernel with previous patches in this series 2) limit-rspin - 5.10-rc3 kernel with limited reader spinning patch 3) no-rspin - 5.10-rc3 kernel with reader spinning disabled # of threads CS Load Vanilla Before limit-rspin no-rspin ------------ ------- ------- ------ ----------- -------- 2 1 5,185 5,662 5,214 5,077 4 1 5,107 4,983 5,188 4,760 8 1 4,782 4,564 4,720 4,628 16 1 4,680 4,053 4,567 3,402 32 1 4,299 1,115 1,118 1,098 64 1 3,218 983 1,001 957 96 1 1,938 944 957 930 2 20 2,008 2,128 2,264 1,665 4 20 1,390 1,033 1,046 1,101 8 20 1,472 1,155 1,098 1,213 16 20 1,332 1,077 1,089 1,122 32 20 967 914 917 980 64 20 787 874 891 858 96 20 730 836 847 844 2 100 372 356 360 355 4 100 492 425 434 392 8 100 533 537 529 538 16 100 548 572 568 598 32 100 499 520 527 537 64 100 466 517 526 512 96 100 406 497 506 509 The column "CS Load" represents the number of pause instructions issued in the locking critical section. A CS load of 1 is extremely short and is not likey in real situations. A load of 20 (moderate) and 100 (long) are more realistic. It can be seen that the previous patches in this series have reduced performance in general except in highly contended cases with moderate or long critical sections that performance improves a bit. This change is mostly caused by the "Prevent potential lock starvation" patch that reduce reader optimistic spinning and hence reduce reader fragmentation. The patch that further limit reader optimistic spinning doesn't seem to have too much impact on overall performance as shown in the benchmark data. The patch that disables reader optimistic spinning shows reduced performance at lightly loaded cases, but comparable or slightly better performance on with heavier contention. This patch just removes reader optimistic spinning for now. As readers are not going to do optimistic spinning anymore, we don't need to consider if the OSQ is empty or not when doing lock stealing. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-6-longman@redhat.com
2020-11-21 07:14:16 +03:00
if (!(count & (RWSEM_WRITER_LOCKED | RWSEM_FLAG_HANDOFF))) {
rwsem_set_reader_owned(sem);
lockevent_inc(rwsem_rlock_steal);
/*
locking/rwsem: Remove reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It also improves the chance that a reader can get the lock as writer optimistic spinning disproportionally favors writers much more than readers. Since commit d3681e269fff ("locking/rwsem: Wake up almost all readers in wait queue"), all the waiting readers are woken up so that they can all get the read lock and run in parallel. When the number of contending readers is large, allowing reader optimistic spinning will likely cause reader fragmentation where multiple smaller groups of readers can get the read lock in a sequential manner separated by writers. That reduces reader parallelism. One possible way to address that drawback is to limit the number of readers (preferably one) that can do optimistic spinning. These readers act as representatives of all the waiting readers in the wait queue as they will wake up all those waiting readers once they get the lock. Alternatively, as reader optimistic lock stealing has already enhanced fairness to readers, it may be easier to just remove reader optimistic spinning and simplifying the optimistic spinning code as a result. Performance measurements (locking throughput kops/s) using a locking microbenchmark with 50/50 reader/writer distribution and turbo-boost disabled was done on a 2-socket Cascade Lake system (48-core 96-thread) to see the impacts of these changes: 1) Vanilla - 5.10-rc3 kernel 2) Before - 5.10-rc3 kernel with previous patches in this series 2) limit-rspin - 5.10-rc3 kernel with limited reader spinning patch 3) no-rspin - 5.10-rc3 kernel with reader spinning disabled # of threads CS Load Vanilla Before limit-rspin no-rspin ------------ ------- ------- ------ ----------- -------- 2 1 5,185 5,662 5,214 5,077 4 1 5,107 4,983 5,188 4,760 8 1 4,782 4,564 4,720 4,628 16 1 4,680 4,053 4,567 3,402 32 1 4,299 1,115 1,118 1,098 64 1 3,218 983 1,001 957 96 1 1,938 944 957 930 2 20 2,008 2,128 2,264 1,665 4 20 1,390 1,033 1,046 1,101 8 20 1,472 1,155 1,098 1,213 16 20 1,332 1,077 1,089 1,122 32 20 967 914 917 980 64 20 787 874 891 858 96 20 730 836 847 844 2 100 372 356 360 355 4 100 492 425 434 392 8 100 533 537 529 538 16 100 548 572 568 598 32 100 499 520 527 537 64 100 466 517 526 512 96 100 406 497 506 509 The column "CS Load" represents the number of pause instructions issued in the locking critical section. A CS load of 1 is extremely short and is not likey in real situations. A load of 20 (moderate) and 100 (long) are more realistic. It can be seen that the previous patches in this series have reduced performance in general except in highly contended cases with moderate or long critical sections that performance improves a bit. This change is mostly caused by the "Prevent potential lock starvation" patch that reduce reader optimistic spinning and hence reduce reader fragmentation. The patch that further limit reader optimistic spinning doesn't seem to have too much impact on overall performance as shown in the benchmark data. The patch that disables reader optimistic spinning shows reduced performance at lightly loaded cases, but comparable or slightly better performance on with heavier contention. This patch just removes reader optimistic spinning for now. As readers are not going to do optimistic spinning anymore, we don't need to consider if the OSQ is empty or not when doing lock stealing. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-6-longman@redhat.com
2020-11-21 07:14:16 +03:00
* Wake up other readers in the wait queue if it is
* the first reader.
*/
locking/rwsem: Remove reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It also improves the chance that a reader can get the lock as writer optimistic spinning disproportionally favors writers much more than readers. Since commit d3681e269fff ("locking/rwsem: Wake up almost all readers in wait queue"), all the waiting readers are woken up so that they can all get the read lock and run in parallel. When the number of contending readers is large, allowing reader optimistic spinning will likely cause reader fragmentation where multiple smaller groups of readers can get the read lock in a sequential manner separated by writers. That reduces reader parallelism. One possible way to address that drawback is to limit the number of readers (preferably one) that can do optimistic spinning. These readers act as representatives of all the waiting readers in the wait queue as they will wake up all those waiting readers once they get the lock. Alternatively, as reader optimistic lock stealing has already enhanced fairness to readers, it may be easier to just remove reader optimistic spinning and simplifying the optimistic spinning code as a result. Performance measurements (locking throughput kops/s) using a locking microbenchmark with 50/50 reader/writer distribution and turbo-boost disabled was done on a 2-socket Cascade Lake system (48-core 96-thread) to see the impacts of these changes: 1) Vanilla - 5.10-rc3 kernel 2) Before - 5.10-rc3 kernel with previous patches in this series 2) limit-rspin - 5.10-rc3 kernel with limited reader spinning patch 3) no-rspin - 5.10-rc3 kernel with reader spinning disabled # of threads CS Load Vanilla Before limit-rspin no-rspin ------------ ------- ------- ------ ----------- -------- 2 1 5,185 5,662 5,214 5,077 4 1 5,107 4,983 5,188 4,760 8 1 4,782 4,564 4,720 4,628 16 1 4,680 4,053 4,567 3,402 32 1 4,299 1,115 1,118 1,098 64 1 3,218 983 1,001 957 96 1 1,938 944 957 930 2 20 2,008 2,128 2,264 1,665 4 20 1,390 1,033 1,046 1,101 8 20 1,472 1,155 1,098 1,213 16 20 1,332 1,077 1,089 1,122 32 20 967 914 917 980 64 20 787 874 891 858 96 20 730 836 847 844 2 100 372 356 360 355 4 100 492 425 434 392 8 100 533 537 529 538 16 100 548 572 568 598 32 100 499 520 527 537 64 100 466 517 526 512 96 100 406 497 506 509 The column "CS Load" represents the number of pause instructions issued in the locking critical section. A CS load of 1 is extremely short and is not likey in real situations. A load of 20 (moderate) and 100 (long) are more realistic. It can be seen that the previous patches in this series have reduced performance in general except in highly contended cases with moderate or long critical sections that performance improves a bit. This change is mostly caused by the "Prevent potential lock starvation" patch that reduce reader optimistic spinning and hence reduce reader fragmentation. The patch that further limit reader optimistic spinning doesn't seem to have too much impact on overall performance as shown in the benchmark data. The patch that disables reader optimistic spinning shows reduced performance at lightly loaded cases, but comparable or slightly better performance on with heavier contention. This patch just removes reader optimistic spinning for now. As readers are not going to do optimistic spinning anymore, we don't need to consider if the OSQ is empty or not when doing lock stealing. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-6-longman@redhat.com
2020-11-21 07:14:16 +03:00
if ((rcnt == 1) && (count & RWSEM_FLAG_WAITERS)) {
raw_spin_lock_irq(&sem->wait_lock);
if (!list_empty(&sem->wait_list))
rwsem_mark_wake(sem, RWSEM_WAKE_READ_OWNED,
&wake_q);
raw_spin_unlock_irq(&sem->wait_lock);
wake_up_q(&wake_q);
}
return sem;
}
queue:
waiter.task = current;
waiter.type = RWSEM_WAITING_FOR_READ;
locking/rwsem: Implement lock handoff to prevent lock starvation Because of writer lock stealing, it is possible that a constant stream of incoming writers will cause a waiting writer or reader to wait indefinitely leading to lock starvation. This patch implements a lock handoff mechanism to disable lock stealing and force lock handoff to the first waiter or waiters (for readers) in the queue after at least a 4ms waiting period unless it is a RT writer task which doesn't need to wait. The waiting period is used to avoid discouraging lock stealing too much to affect performance. The setting and clearing of the handoff bit is serialized by the wait_lock. So racing is not possible. A rwsem microbenchmark was run for 5 seconds on a 2-socket 40-core 80-thread Skylake system with a v5.1 based kernel and 240 write_lock threads with 5us sleep critical section. Before the patch, the min/mean/max numbers of locking operations for the locking threads were 1/7,792/173,696. After the patch, the figures became 5,842/6,542/7,458. It can be seen that the rwsem became much more fair, though there was a drop of about 16% in the mean locking operations done which was a tradeoff of having better fairness. Making the waiter set the handoff bit right after the first wakeup can impact performance especially with a mixed reader/writer workload. With the same microbenchmark with short critical section and equal number of reader and writer threads (40/40), the reader/writer locking operation counts with the current patch were: 40 readers, Iterations Min/Mean/Max = 1,793/1,794/1,796 40 writers, Iterations Min/Mean/Max = 1,793/34,956/86,081 By making waiter set handoff bit immediately after wakeup: 40 readers, Iterations Min/Mean/Max = 43/44/46 40 writers, Iterations Min/Mean/Max = 43/1,263/3,191 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-8-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:06 +03:00
waiter.timeout = jiffies + RWSEM_WAIT_TIMEOUT;
raw_spin_lock_irq(&sem->wait_lock);
if (list_empty(&sem->wait_list)) {
/*
* In case the wait queue is empty and the lock isn't owned
locking/rwsem: Implement lock handoff to prevent lock starvation Because of writer lock stealing, it is possible that a constant stream of incoming writers will cause a waiting writer or reader to wait indefinitely leading to lock starvation. This patch implements a lock handoff mechanism to disable lock stealing and force lock handoff to the first waiter or waiters (for readers) in the queue after at least a 4ms waiting period unless it is a RT writer task which doesn't need to wait. The waiting period is used to avoid discouraging lock stealing too much to affect performance. The setting and clearing of the handoff bit is serialized by the wait_lock. So racing is not possible. A rwsem microbenchmark was run for 5 seconds on a 2-socket 40-core 80-thread Skylake system with a v5.1 based kernel and 240 write_lock threads with 5us sleep critical section. Before the patch, the min/mean/max numbers of locking operations for the locking threads were 1/7,792/173,696. After the patch, the figures became 5,842/6,542/7,458. It can be seen that the rwsem became much more fair, though there was a drop of about 16% in the mean locking operations done which was a tradeoff of having better fairness. Making the waiter set the handoff bit right after the first wakeup can impact performance especially with a mixed reader/writer workload. With the same microbenchmark with short critical section and equal number of reader and writer threads (40/40), the reader/writer locking operation counts with the current patch were: 40 readers, Iterations Min/Mean/Max = 1,793/1,794/1,796 40 writers, Iterations Min/Mean/Max = 1,793/34,956/86,081 By making waiter set handoff bit immediately after wakeup: 40 readers, Iterations Min/Mean/Max = 43/44/46 40 writers, Iterations Min/Mean/Max = 43/1,263/3,191 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-8-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:06 +03:00
* by a writer or has the handoff bit set, this reader can
* exit the slowpath and return immediately as its
* RWSEM_READER_BIAS has already been set in the count.
*/
locking/rwsem: Remove reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It also improves the chance that a reader can get the lock as writer optimistic spinning disproportionally favors writers much more than readers. Since commit d3681e269fff ("locking/rwsem: Wake up almost all readers in wait queue"), all the waiting readers are woken up so that they can all get the read lock and run in parallel. When the number of contending readers is large, allowing reader optimistic spinning will likely cause reader fragmentation where multiple smaller groups of readers can get the read lock in a sequential manner separated by writers. That reduces reader parallelism. One possible way to address that drawback is to limit the number of readers (preferably one) that can do optimistic spinning. These readers act as representatives of all the waiting readers in the wait queue as they will wake up all those waiting readers once they get the lock. Alternatively, as reader optimistic lock stealing has already enhanced fairness to readers, it may be easier to just remove reader optimistic spinning and simplifying the optimistic spinning code as a result. Performance measurements (locking throughput kops/s) using a locking microbenchmark with 50/50 reader/writer distribution and turbo-boost disabled was done on a 2-socket Cascade Lake system (48-core 96-thread) to see the impacts of these changes: 1) Vanilla - 5.10-rc3 kernel 2) Before - 5.10-rc3 kernel with previous patches in this series 2) limit-rspin - 5.10-rc3 kernel with limited reader spinning patch 3) no-rspin - 5.10-rc3 kernel with reader spinning disabled # of threads CS Load Vanilla Before limit-rspin no-rspin ------------ ------- ------- ------ ----------- -------- 2 1 5,185 5,662 5,214 5,077 4 1 5,107 4,983 5,188 4,760 8 1 4,782 4,564 4,720 4,628 16 1 4,680 4,053 4,567 3,402 32 1 4,299 1,115 1,118 1,098 64 1 3,218 983 1,001 957 96 1 1,938 944 957 930 2 20 2,008 2,128 2,264 1,665 4 20 1,390 1,033 1,046 1,101 8 20 1,472 1,155 1,098 1,213 16 20 1,332 1,077 1,089 1,122 32 20 967 914 917 980 64 20 787 874 891 858 96 20 730 836 847 844 2 100 372 356 360 355 4 100 492 425 434 392 8 100 533 537 529 538 16 100 548 572 568 598 32 100 499 520 527 537 64 100 466 517 526 512 96 100 406 497 506 509 The column "CS Load" represents the number of pause instructions issued in the locking critical section. A CS load of 1 is extremely short and is not likey in real situations. A load of 20 (moderate) and 100 (long) are more realistic. It can be seen that the previous patches in this series have reduced performance in general except in highly contended cases with moderate or long critical sections that performance improves a bit. This change is mostly caused by the "Prevent potential lock starvation" patch that reduce reader optimistic spinning and hence reduce reader fragmentation. The patch that further limit reader optimistic spinning doesn't seem to have too much impact on overall performance as shown in the benchmark data. The patch that disables reader optimistic spinning shows reduced performance at lightly loaded cases, but comparable or slightly better performance on with heavier contention. This patch just removes reader optimistic spinning for now. As readers are not going to do optimistic spinning anymore, we don't need to consider if the OSQ is empty or not when doing lock stealing. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-6-longman@redhat.com
2020-11-21 07:14:16 +03:00
if (!(atomic_long_read(&sem->count) &
locking/rwsem: Implement lock handoff to prevent lock starvation Because of writer lock stealing, it is possible that a constant stream of incoming writers will cause a waiting writer or reader to wait indefinitely leading to lock starvation. This patch implements a lock handoff mechanism to disable lock stealing and force lock handoff to the first waiter or waiters (for readers) in the queue after at least a 4ms waiting period unless it is a RT writer task which doesn't need to wait. The waiting period is used to avoid discouraging lock stealing too much to affect performance. The setting and clearing of the handoff bit is serialized by the wait_lock. So racing is not possible. A rwsem microbenchmark was run for 5 seconds on a 2-socket 40-core 80-thread Skylake system with a v5.1 based kernel and 240 write_lock threads with 5us sleep critical section. Before the patch, the min/mean/max numbers of locking operations for the locking threads were 1/7,792/173,696. After the patch, the figures became 5,842/6,542/7,458. It can be seen that the rwsem became much more fair, though there was a drop of about 16% in the mean locking operations done which was a tradeoff of having better fairness. Making the waiter set the handoff bit right after the first wakeup can impact performance especially with a mixed reader/writer workload. With the same microbenchmark with short critical section and equal number of reader and writer threads (40/40), the reader/writer locking operation counts with the current patch were: 40 readers, Iterations Min/Mean/Max = 1,793/1,794/1,796 40 writers, Iterations Min/Mean/Max = 1,793/34,956/86,081 By making waiter set handoff bit immediately after wakeup: 40 readers, Iterations Min/Mean/Max = 43/44/46 40 writers, Iterations Min/Mean/Max = 43/1,263/3,191 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-8-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:06 +03:00
(RWSEM_WRITER_MASK | RWSEM_FLAG_HANDOFF))) {
locking/rwsem: Add missing ACQUIRE to read_slowpath exit when queue is empty LTP mtest06 has been observed to occasionally hit "still mapped when deleted" and following BUG_ON on arm64. The extra mapcount originated from pagefault handler, which handled pagefault for vma that has already been detached. vma is detached under mmap_sem write lock by detach_vmas_to_be_unmapped(), which also invalidates vmacache. When the pagefault handler (under mmap_sem read lock) calls find_vma(), vmacache_valid() wrongly reports vmacache as valid. After rwsem down_read() returns via 'queue empty' path (as of v5.2), it does so without an ACQUIRE on sem->count: down_read() __down_read() rwsem_down_read_failed() __rwsem_down_read_failed_common() raw_spin_lock_irq(&sem->wait_lock); if (list_empty(&sem->wait_list)) { if (atomic_long_read(&sem->count) >= 0) { raw_spin_unlock_irq(&sem->wait_lock); return sem; The problem can be reproduced by running LTP mtest06 in a loop and building the kernel (-j $NCPUS) in parallel. It does reproduces since v4.20 on arm64 HPE Apollo 70 (224 CPUs, 256GB RAM, 2 nodes). It triggers reliably in about an hour. The patched kernel ran fine for 10+ hours. Signed-off-by: Jan Stancek <jstancek@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Will Deacon <will@kernel.org> Acked-by: Waiman Long <longman@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: dbueso@suse.de Fixes: 4b486b535c33 ("locking/rwsem: Exit read lock slowpath if queue empty & no writer") Link: https://lkml.kernel.org/r/50b8914e20d1d62bb2dee42d342836c2c16ebee7.1563438048.git.jstancek@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-07-18 11:51:25 +03:00
/* Provide lock ACQUIRE */
smp_acquire__after_ctrl_dep();
raw_spin_unlock_irq(&sem->wait_lock);
rwsem_set_reader_owned(sem);
lockevent_inc(rwsem_rlock_fast);
return sem;
}
adjustment += RWSEM_FLAG_WAITERS;
}
list_add_tail(&waiter.list, &sem->wait_list);
/* we're now waiting on the lock, but no longer actively locking */
locking/rwsem: Remove reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It also improves the chance that a reader can get the lock as writer optimistic spinning disproportionally favors writers much more than readers. Since commit d3681e269fff ("locking/rwsem: Wake up almost all readers in wait queue"), all the waiting readers are woken up so that they can all get the read lock and run in parallel. When the number of contending readers is large, allowing reader optimistic spinning will likely cause reader fragmentation where multiple smaller groups of readers can get the read lock in a sequential manner separated by writers. That reduces reader parallelism. One possible way to address that drawback is to limit the number of readers (preferably one) that can do optimistic spinning. These readers act as representatives of all the waiting readers in the wait queue as they will wake up all those waiting readers once they get the lock. Alternatively, as reader optimistic lock stealing has already enhanced fairness to readers, it may be easier to just remove reader optimistic spinning and simplifying the optimistic spinning code as a result. Performance measurements (locking throughput kops/s) using a locking microbenchmark with 50/50 reader/writer distribution and turbo-boost disabled was done on a 2-socket Cascade Lake system (48-core 96-thread) to see the impacts of these changes: 1) Vanilla - 5.10-rc3 kernel 2) Before - 5.10-rc3 kernel with previous patches in this series 2) limit-rspin - 5.10-rc3 kernel with limited reader spinning patch 3) no-rspin - 5.10-rc3 kernel with reader spinning disabled # of threads CS Load Vanilla Before limit-rspin no-rspin ------------ ------- ------- ------ ----------- -------- 2 1 5,185 5,662 5,214 5,077 4 1 5,107 4,983 5,188 4,760 8 1 4,782 4,564 4,720 4,628 16 1 4,680 4,053 4,567 3,402 32 1 4,299 1,115 1,118 1,098 64 1 3,218 983 1,001 957 96 1 1,938 944 957 930 2 20 2,008 2,128 2,264 1,665 4 20 1,390 1,033 1,046 1,101 8 20 1,472 1,155 1,098 1,213 16 20 1,332 1,077 1,089 1,122 32 20 967 914 917 980 64 20 787 874 891 858 96 20 730 836 847 844 2 100 372 356 360 355 4 100 492 425 434 392 8 100 533 537 529 538 16 100 548 572 568 598 32 100 499 520 527 537 64 100 466 517 526 512 96 100 406 497 506 509 The column "CS Load" represents the number of pause instructions issued in the locking critical section. A CS load of 1 is extremely short and is not likey in real situations. A load of 20 (moderate) and 100 (long) are more realistic. It can be seen that the previous patches in this series have reduced performance in general except in highly contended cases with moderate or long critical sections that performance improves a bit. This change is mostly caused by the "Prevent potential lock starvation" patch that reduce reader optimistic spinning and hence reduce reader fragmentation. The patch that further limit reader optimistic spinning doesn't seem to have too much impact on overall performance as shown in the benchmark data. The patch that disables reader optimistic spinning shows reduced performance at lightly loaded cases, but comparable or slightly better performance on with heavier contention. This patch just removes reader optimistic spinning for now. As readers are not going to do optimistic spinning anymore, we don't need to consider if the OSQ is empty or not when doing lock stealing. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-6-longman@redhat.com
2020-11-21 07:14:16 +03:00
count = atomic_long_add_return(adjustment, &sem->count);
/*
* If there are no active locks, wake the front queued process(es).
*
* If there are no writers and we are first in the queue,
* wake our own waiter to join the existing active readers !
*/
locking/rwsem: Enable time-based spinning on reader-owned rwsem When the rwsem is owned by reader, writers stop optimistic spinning simply because there is no easy way to figure out if all the readers are actively running or not. However, there are scenarios where the readers are unlikely to sleep and optimistic spinning can help performance. This patch provides a simple mechanism for spinning on a reader-owned rwsem by a writer. It is a time threshold based spinning where the allowable spinning time can vary from 10us to 25us depending on the condition of the rwsem. When the time threshold is exceeded, the nonspinnable bits will be set in the owner field to indicate that no more optimistic spinning will be allowed on this rwsem until it becomes writer owned again. Not even readers is allowed to acquire the reader-locked rwsem by optimistic spinning for fairness. We also want a writer to acquire the lock after the readers hold the lock for a relatively long time. In order to give preference to writers under such a circumstance, the single RWSEM_NONSPINNABLE bit is now split into two - one for reader and one for writer. When optimistic spinning is disabled, both bits will be set. When the reader count drop down to 0, the writer nonspinnable bit will be cleared to allow writers to spin on the lock, but not the readers. When a writer acquires the lock, it will write its own task structure pointer into sem->owner and clear the reader nonspinnable bit in the process. The time taken for each iteration of the reader-owned rwsem spinning loop varies. Below are sample minimum elapsed times for 16 iterations of the loop. System Time for 16 Iterations ------ ---------------------- 1-socket Skylake ~800ns 4-socket Broadwell ~300ns 2-socket ThunderX2 (arm64) ~250ns When the lock cacheline is contended, we can see up to almost 10X increase in elapsed time. So 25us will be at most 500, 1300 and 1600 iterations for each of the above systems. With a locking microbenchmark running on 5.1 based kernel, the total locking rates (in kops/s) on a 8-socket IvyBridge-EX system with equal numbers of readers and writers before and after this patch were as follows: # of Threads Pre-patch Post-patch ------------ --------- ---------- 2 1,759 6,684 4 1,684 6,738 8 1,074 7,222 16 900 7,163 32 458 7,316 64 208 520 128 168 425 240 143 474 This patch gives a big boost in performance for mixed reader/writer workloads. With 32 locking threads, the rwsem lock event data were: rwsem_opt_fail=79850 rwsem_opt_nospin=5069 rwsem_opt_rlock=597484 rwsem_opt_wlock=957339 rwsem_sleep_reader=57782 rwsem_sleep_writer=55663 With 64 locking threads, the data looked like: rwsem_opt_fail=346723 rwsem_opt_nospin=6293 rwsem_opt_rlock=1127119 rwsem_opt_wlock=1400628 rwsem_sleep_reader=308201 rwsem_sleep_writer=72281 So a lot more threads acquired the lock in the slowpath and more threads went to sleep. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-15-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:13 +03:00
if (!(count & RWSEM_LOCK_MASK)) {
locking/rwsem: Remove reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It also improves the chance that a reader can get the lock as writer optimistic spinning disproportionally favors writers much more than readers. Since commit d3681e269fff ("locking/rwsem: Wake up almost all readers in wait queue"), all the waiting readers are woken up so that they can all get the read lock and run in parallel. When the number of contending readers is large, allowing reader optimistic spinning will likely cause reader fragmentation where multiple smaller groups of readers can get the read lock in a sequential manner separated by writers. That reduces reader parallelism. One possible way to address that drawback is to limit the number of readers (preferably one) that can do optimistic spinning. These readers act as representatives of all the waiting readers in the wait queue as they will wake up all those waiting readers once they get the lock. Alternatively, as reader optimistic lock stealing has already enhanced fairness to readers, it may be easier to just remove reader optimistic spinning and simplifying the optimistic spinning code as a result. Performance measurements (locking throughput kops/s) using a locking microbenchmark with 50/50 reader/writer distribution and turbo-boost disabled was done on a 2-socket Cascade Lake system (48-core 96-thread) to see the impacts of these changes: 1) Vanilla - 5.10-rc3 kernel 2) Before - 5.10-rc3 kernel with previous patches in this series 2) limit-rspin - 5.10-rc3 kernel with limited reader spinning patch 3) no-rspin - 5.10-rc3 kernel with reader spinning disabled # of threads CS Load Vanilla Before limit-rspin no-rspin ------------ ------- ------- ------ ----------- -------- 2 1 5,185 5,662 5,214 5,077 4 1 5,107 4,983 5,188 4,760 8 1 4,782 4,564 4,720 4,628 16 1 4,680 4,053 4,567 3,402 32 1 4,299 1,115 1,118 1,098 64 1 3,218 983 1,001 957 96 1 1,938 944 957 930 2 20 2,008 2,128 2,264 1,665 4 20 1,390 1,033 1,046 1,101 8 20 1,472 1,155 1,098 1,213 16 20 1,332 1,077 1,089 1,122 32 20 967 914 917 980 64 20 787 874 891 858 96 20 730 836 847 844 2 100 372 356 360 355 4 100 492 425 434 392 8 100 533 537 529 538 16 100 548 572 568 598 32 100 499 520 527 537 64 100 466 517 526 512 96 100 406 497 506 509 The column "CS Load" represents the number of pause instructions issued in the locking critical section. A CS load of 1 is extremely short and is not likey in real situations. A load of 20 (moderate) and 100 (long) are more realistic. It can be seen that the previous patches in this series have reduced performance in general except in highly contended cases with moderate or long critical sections that performance improves a bit. This change is mostly caused by the "Prevent potential lock starvation" patch that reduce reader optimistic spinning and hence reduce reader fragmentation. The patch that further limit reader optimistic spinning doesn't seem to have too much impact on overall performance as shown in the benchmark data. The patch that disables reader optimistic spinning shows reduced performance at lightly loaded cases, but comparable or slightly better performance on with heavier contention. This patch just removes reader optimistic spinning for now. As readers are not going to do optimistic spinning anymore, we don't need to consider if the OSQ is empty or not when doing lock stealing. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-6-longman@redhat.com
2020-11-21 07:14:16 +03:00
clear_nonspinnable(sem);
locking/rwsem: Enable time-based spinning on reader-owned rwsem When the rwsem is owned by reader, writers stop optimistic spinning simply because there is no easy way to figure out if all the readers are actively running or not. However, there are scenarios where the readers are unlikely to sleep and optimistic spinning can help performance. This patch provides a simple mechanism for spinning on a reader-owned rwsem by a writer. It is a time threshold based spinning where the allowable spinning time can vary from 10us to 25us depending on the condition of the rwsem. When the time threshold is exceeded, the nonspinnable bits will be set in the owner field to indicate that no more optimistic spinning will be allowed on this rwsem until it becomes writer owned again. Not even readers is allowed to acquire the reader-locked rwsem by optimistic spinning for fairness. We also want a writer to acquire the lock after the readers hold the lock for a relatively long time. In order to give preference to writers under such a circumstance, the single RWSEM_NONSPINNABLE bit is now split into two - one for reader and one for writer. When optimistic spinning is disabled, both bits will be set. When the reader count drop down to 0, the writer nonspinnable bit will be cleared to allow writers to spin on the lock, but not the readers. When a writer acquires the lock, it will write its own task structure pointer into sem->owner and clear the reader nonspinnable bit in the process. The time taken for each iteration of the reader-owned rwsem spinning loop varies. Below are sample minimum elapsed times for 16 iterations of the loop. System Time for 16 Iterations ------ ---------------------- 1-socket Skylake ~800ns 4-socket Broadwell ~300ns 2-socket ThunderX2 (arm64) ~250ns When the lock cacheline is contended, we can see up to almost 10X increase in elapsed time. So 25us will be at most 500, 1300 and 1600 iterations for each of the above systems. With a locking microbenchmark running on 5.1 based kernel, the total locking rates (in kops/s) on a 8-socket IvyBridge-EX system with equal numbers of readers and writers before and after this patch were as follows: # of Threads Pre-patch Post-patch ------------ --------- ---------- 2 1,759 6,684 4 1,684 6,738 8 1,074 7,222 16 900 7,163 32 458 7,316 64 208 520 128 168 425 240 143 474 This patch gives a big boost in performance for mixed reader/writer workloads. With 32 locking threads, the rwsem lock event data were: rwsem_opt_fail=79850 rwsem_opt_nospin=5069 rwsem_opt_rlock=597484 rwsem_opt_wlock=957339 rwsem_sleep_reader=57782 rwsem_sleep_writer=55663 With 64 locking threads, the data looked like: rwsem_opt_fail=346723 rwsem_opt_nospin=6293 rwsem_opt_rlock=1127119 rwsem_opt_wlock=1400628 rwsem_sleep_reader=308201 rwsem_sleep_writer=72281 So a lot more threads acquired the lock in the slowpath and more threads went to sleep. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-15-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:13 +03:00
wake = true;
}
if (wake || (!(count & RWSEM_WRITER_MASK) &&
(adjustment & RWSEM_FLAG_WAITERS)))
locking/rwsem: Code cleanup after files merging After merging all the relevant rwsem code into one single file, there are a number of optimizations and cleanups that can be done: 1) Remove all the EXPORT_SYMBOL() calls for functions that are not accessed elsewhere. 2) Remove all the __visible tags as none of the functions will be called from assembly code anymore. 3) Make all the internal functions static. 4) Remove some unneeded blank lines. 5) Remove the intermediate rwsem_down_{read|write}_failed*() functions and rename __rwsem_down_{read|write}_failed_common() to rwsem_down_{read|write}_slowpath(). 6) Remove "__" prefix of __rwsem_mark_wake(). 7) Use atomic_long_try_cmpxchg_acquire() as much as possible. 8) Remove the rwsem_rtrylock and rwsem_wtrylock lock events as they are not that useful. That enables the compiler to do better optimization and reduce code size. The text+data size of rwsem.o on an x86-64 machine with gcc8 was reduced from 10237 bytes to 5030 bytes with this change. Suggested-by: Peter Zijlstra <peterz@infradead.org> Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-6-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:04 +03:00
rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q);
raw_spin_unlock_irq(&sem->wait_lock);
wake_up_q(&wake_q);
/* wait to be given the lock */
for (;;) {
set_current_state(state);
if (!smp_load_acquire(&waiter.task)) {
/* Matches rwsem_mark_wake()'s smp_store_release(). */
break;
}
if (signal_pending_state(state, current)) {
raw_spin_lock_irq(&sem->wait_lock);
if (waiter.task)
goto out_nolock;
raw_spin_unlock_irq(&sem->wait_lock);
/* Ordered by sem->wait_lock against rwsem_mark_wake(). */
break;
}
schedule();
lockevent_inc(rwsem_sleep_reader);
}
__set_current_state(TASK_RUNNING);
lockevent_inc(rwsem_rlock);
return sem;
out_nolock:
list_del(&waiter.list);
locking/rwsem: Implement lock handoff to prevent lock starvation Because of writer lock stealing, it is possible that a constant stream of incoming writers will cause a waiting writer or reader to wait indefinitely leading to lock starvation. This patch implements a lock handoff mechanism to disable lock stealing and force lock handoff to the first waiter or waiters (for readers) in the queue after at least a 4ms waiting period unless it is a RT writer task which doesn't need to wait. The waiting period is used to avoid discouraging lock stealing too much to affect performance. The setting and clearing of the handoff bit is serialized by the wait_lock. So racing is not possible. A rwsem microbenchmark was run for 5 seconds on a 2-socket 40-core 80-thread Skylake system with a v5.1 based kernel and 240 write_lock threads with 5us sleep critical section. Before the patch, the min/mean/max numbers of locking operations for the locking threads were 1/7,792/173,696. After the patch, the figures became 5,842/6,542/7,458. It can be seen that the rwsem became much more fair, though there was a drop of about 16% in the mean locking operations done which was a tradeoff of having better fairness. Making the waiter set the handoff bit right after the first wakeup can impact performance especially with a mixed reader/writer workload. With the same microbenchmark with short critical section and equal number of reader and writer threads (40/40), the reader/writer locking operation counts with the current patch were: 40 readers, Iterations Min/Mean/Max = 1,793/1,794/1,796 40 writers, Iterations Min/Mean/Max = 1,793/34,956/86,081 By making waiter set handoff bit immediately after wakeup: 40 readers, Iterations Min/Mean/Max = 43/44/46 40 writers, Iterations Min/Mean/Max = 43/1,263/3,191 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-8-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:06 +03:00
if (list_empty(&sem->wait_list)) {
atomic_long_andnot(RWSEM_FLAG_WAITERS|RWSEM_FLAG_HANDOFF,
&sem->count);
}
raw_spin_unlock_irq(&sem->wait_lock);
__set_current_state(TASK_RUNNING);
lockevent_inc(rwsem_rlock_fail);
return ERR_PTR(-EINTR);
}
/*
* Wait until we successfully acquire the write lock
*/
locking/rwsem: Code cleanup after files merging After merging all the relevant rwsem code into one single file, there are a number of optimizations and cleanups that can be done: 1) Remove all the EXPORT_SYMBOL() calls for functions that are not accessed elsewhere. 2) Remove all the __visible tags as none of the functions will be called from assembly code anymore. 3) Make all the internal functions static. 4) Remove some unneeded blank lines. 5) Remove the intermediate rwsem_down_{read|write}_failed*() functions and rename __rwsem_down_{read|write}_failed_common() to rwsem_down_{read|write}_slowpath(). 6) Remove "__" prefix of __rwsem_mark_wake(). 7) Use atomic_long_try_cmpxchg_acquire() as much as possible. 8) Remove the rwsem_rtrylock and rwsem_wtrylock lock events as they are not that useful. That enables the compiler to do better optimization and reduce code size. The text+data size of rwsem.o on an x86-64 machine with gcc8 was reduced from 10237 bytes to 5030 bytes with this change. Suggested-by: Peter Zijlstra <peterz@infradead.org> Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-6-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:04 +03:00
static struct rw_semaphore *
rwsem_down_write_slowpath(struct rw_semaphore *sem, int state)
{
long count;
locking/rwsem: Implement lock handoff to prevent lock starvation Because of writer lock stealing, it is possible that a constant stream of incoming writers will cause a waiting writer or reader to wait indefinitely leading to lock starvation. This patch implements a lock handoff mechanism to disable lock stealing and force lock handoff to the first waiter or waiters (for readers) in the queue after at least a 4ms waiting period unless it is a RT writer task which doesn't need to wait. The waiting period is used to avoid discouraging lock stealing too much to affect performance. The setting and clearing of the handoff bit is serialized by the wait_lock. So racing is not possible. A rwsem microbenchmark was run for 5 seconds on a 2-socket 40-core 80-thread Skylake system with a v5.1 based kernel and 240 write_lock threads with 5us sleep critical section. Before the patch, the min/mean/max numbers of locking operations for the locking threads were 1/7,792/173,696. After the patch, the figures became 5,842/6,542/7,458. It can be seen that the rwsem became much more fair, though there was a drop of about 16% in the mean locking operations done which was a tradeoff of having better fairness. Making the waiter set the handoff bit right after the first wakeup can impact performance especially with a mixed reader/writer workload. With the same microbenchmark with short critical section and equal number of reader and writer threads (40/40), the reader/writer locking operation counts with the current patch were: 40 readers, Iterations Min/Mean/Max = 1,793/1,794/1,796 40 writers, Iterations Min/Mean/Max = 1,793/34,956/86,081 By making waiter set handoff bit immediately after wakeup: 40 readers, Iterations Min/Mean/Max = 43/44/46 40 writers, Iterations Min/Mean/Max = 43/1,263/3,191 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-8-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:06 +03:00
enum writer_wait_state wstate;
struct rwsem_waiter waiter;
struct rw_semaphore *ret = sem;
DEFINE_WAKE_Q(wake_q);
/* do optimistic spinning and steal lock if possible */
locking/rwsem: Remove reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It also improves the chance that a reader can get the lock as writer optimistic spinning disproportionally favors writers much more than readers. Since commit d3681e269fff ("locking/rwsem: Wake up almost all readers in wait queue"), all the waiting readers are woken up so that they can all get the read lock and run in parallel. When the number of contending readers is large, allowing reader optimistic spinning will likely cause reader fragmentation where multiple smaller groups of readers can get the read lock in a sequential manner separated by writers. That reduces reader parallelism. One possible way to address that drawback is to limit the number of readers (preferably one) that can do optimistic spinning. These readers act as representatives of all the waiting readers in the wait queue as they will wake up all those waiting readers once they get the lock. Alternatively, as reader optimistic lock stealing has already enhanced fairness to readers, it may be easier to just remove reader optimistic spinning and simplifying the optimistic spinning code as a result. Performance measurements (locking throughput kops/s) using a locking microbenchmark with 50/50 reader/writer distribution and turbo-boost disabled was done on a 2-socket Cascade Lake system (48-core 96-thread) to see the impacts of these changes: 1) Vanilla - 5.10-rc3 kernel 2) Before - 5.10-rc3 kernel with previous patches in this series 2) limit-rspin - 5.10-rc3 kernel with limited reader spinning patch 3) no-rspin - 5.10-rc3 kernel with reader spinning disabled # of threads CS Load Vanilla Before limit-rspin no-rspin ------------ ------- ------- ------ ----------- -------- 2 1 5,185 5,662 5,214 5,077 4 1 5,107 4,983 5,188 4,760 8 1 4,782 4,564 4,720 4,628 16 1 4,680 4,053 4,567 3,402 32 1 4,299 1,115 1,118 1,098 64 1 3,218 983 1,001 957 96 1 1,938 944 957 930 2 20 2,008 2,128 2,264 1,665 4 20 1,390 1,033 1,046 1,101 8 20 1,472 1,155 1,098 1,213 16 20 1,332 1,077 1,089 1,122 32 20 967 914 917 980 64 20 787 874 891 858 96 20 730 836 847 844 2 100 372 356 360 355 4 100 492 425 434 392 8 100 533 537 529 538 16 100 548 572 568 598 32 100 499 520 527 537 64 100 466 517 526 512 96 100 406 497 506 509 The column "CS Load" represents the number of pause instructions issued in the locking critical section. A CS load of 1 is extremely short and is not likey in real situations. A load of 20 (moderate) and 100 (long) are more realistic. It can be seen that the previous patches in this series have reduced performance in general except in highly contended cases with moderate or long critical sections that performance improves a bit. This change is mostly caused by the "Prevent potential lock starvation" patch that reduce reader optimistic spinning and hence reduce reader fragmentation. The patch that further limit reader optimistic spinning doesn't seem to have too much impact on overall performance as shown in the benchmark data. The patch that disables reader optimistic spinning shows reduced performance at lightly loaded cases, but comparable or slightly better performance on with heavier contention. This patch just removes reader optimistic spinning for now. As readers are not going to do optimistic spinning anymore, we don't need to consider if the OSQ is empty or not when doing lock stealing. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-6-longman@redhat.com
2020-11-21 07:14:16 +03:00
if (rwsem_can_spin_on_owner(sem) && rwsem_optimistic_spin(sem)) {
/* rwsem_optimistic_spin() implies ACQUIRE on success */
return sem;
}
/*
* Optimistic spinning failed, proceed to the slowpath
* and block until we can acquire the sem.
*/
waiter.task = current;
waiter.type = RWSEM_WAITING_FOR_WRITE;
locking/rwsem: Implement lock handoff to prevent lock starvation Because of writer lock stealing, it is possible that a constant stream of incoming writers will cause a waiting writer or reader to wait indefinitely leading to lock starvation. This patch implements a lock handoff mechanism to disable lock stealing and force lock handoff to the first waiter or waiters (for readers) in the queue after at least a 4ms waiting period unless it is a RT writer task which doesn't need to wait. The waiting period is used to avoid discouraging lock stealing too much to affect performance. The setting and clearing of the handoff bit is serialized by the wait_lock. So racing is not possible. A rwsem microbenchmark was run for 5 seconds on a 2-socket 40-core 80-thread Skylake system with a v5.1 based kernel and 240 write_lock threads with 5us sleep critical section. Before the patch, the min/mean/max numbers of locking operations for the locking threads were 1/7,792/173,696. After the patch, the figures became 5,842/6,542/7,458. It can be seen that the rwsem became much more fair, though there was a drop of about 16% in the mean locking operations done which was a tradeoff of having better fairness. Making the waiter set the handoff bit right after the first wakeup can impact performance especially with a mixed reader/writer workload. With the same microbenchmark with short critical section and equal number of reader and writer threads (40/40), the reader/writer locking operation counts with the current patch were: 40 readers, Iterations Min/Mean/Max = 1,793/1,794/1,796 40 writers, Iterations Min/Mean/Max = 1,793/34,956/86,081 By making waiter set handoff bit immediately after wakeup: 40 readers, Iterations Min/Mean/Max = 43/44/46 40 writers, Iterations Min/Mean/Max = 43/1,263/3,191 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-8-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:06 +03:00
waiter.timeout = jiffies + RWSEM_WAIT_TIMEOUT;
raw_spin_lock_irq(&sem->wait_lock);
/* account for this before adding a new element to the list */
locking/rwsem: Implement lock handoff to prevent lock starvation Because of writer lock stealing, it is possible that a constant stream of incoming writers will cause a waiting writer or reader to wait indefinitely leading to lock starvation. This patch implements a lock handoff mechanism to disable lock stealing and force lock handoff to the first waiter or waiters (for readers) in the queue after at least a 4ms waiting period unless it is a RT writer task which doesn't need to wait. The waiting period is used to avoid discouraging lock stealing too much to affect performance. The setting and clearing of the handoff bit is serialized by the wait_lock. So racing is not possible. A rwsem microbenchmark was run for 5 seconds on a 2-socket 40-core 80-thread Skylake system with a v5.1 based kernel and 240 write_lock threads with 5us sleep critical section. Before the patch, the min/mean/max numbers of locking operations for the locking threads were 1/7,792/173,696. After the patch, the figures became 5,842/6,542/7,458. It can be seen that the rwsem became much more fair, though there was a drop of about 16% in the mean locking operations done which was a tradeoff of having better fairness. Making the waiter set the handoff bit right after the first wakeup can impact performance especially with a mixed reader/writer workload. With the same microbenchmark with short critical section and equal number of reader and writer threads (40/40), the reader/writer locking operation counts with the current patch were: 40 readers, Iterations Min/Mean/Max = 1,793/1,794/1,796 40 writers, Iterations Min/Mean/Max = 1,793/34,956/86,081 By making waiter set handoff bit immediately after wakeup: 40 readers, Iterations Min/Mean/Max = 43/44/46 40 writers, Iterations Min/Mean/Max = 43/1,263/3,191 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-8-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:06 +03:00
wstate = list_empty(&sem->wait_list) ? WRITER_FIRST : WRITER_NOT_FIRST;
list_add_tail(&waiter.list, &sem->wait_list);
/* we're now waiting on the lock */
locking/rwsem: Implement lock handoff to prevent lock starvation Because of writer lock stealing, it is possible that a constant stream of incoming writers will cause a waiting writer or reader to wait indefinitely leading to lock starvation. This patch implements a lock handoff mechanism to disable lock stealing and force lock handoff to the first waiter or waiters (for readers) in the queue after at least a 4ms waiting period unless it is a RT writer task which doesn't need to wait. The waiting period is used to avoid discouraging lock stealing too much to affect performance. The setting and clearing of the handoff bit is serialized by the wait_lock. So racing is not possible. A rwsem microbenchmark was run for 5 seconds on a 2-socket 40-core 80-thread Skylake system with a v5.1 based kernel and 240 write_lock threads with 5us sleep critical section. Before the patch, the min/mean/max numbers of locking operations for the locking threads were 1/7,792/173,696. After the patch, the figures became 5,842/6,542/7,458. It can be seen that the rwsem became much more fair, though there was a drop of about 16% in the mean locking operations done which was a tradeoff of having better fairness. Making the waiter set the handoff bit right after the first wakeup can impact performance especially with a mixed reader/writer workload. With the same microbenchmark with short critical section and equal number of reader and writer threads (40/40), the reader/writer locking operation counts with the current patch were: 40 readers, Iterations Min/Mean/Max = 1,793/1,794/1,796 40 writers, Iterations Min/Mean/Max = 1,793/34,956/86,081 By making waiter set handoff bit immediately after wakeup: 40 readers, Iterations Min/Mean/Max = 43/44/46 40 writers, Iterations Min/Mean/Max = 43/1,263/3,191 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-8-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:06 +03:00
if (wstate == WRITER_NOT_FIRST) {
count = atomic_long_read(&sem->count);
/*
locking/rwsem: Implement lock handoff to prevent lock starvation Because of writer lock stealing, it is possible that a constant stream of incoming writers will cause a waiting writer or reader to wait indefinitely leading to lock starvation. This patch implements a lock handoff mechanism to disable lock stealing and force lock handoff to the first waiter or waiters (for readers) in the queue after at least a 4ms waiting period unless it is a RT writer task which doesn't need to wait. The waiting period is used to avoid discouraging lock stealing too much to affect performance. The setting and clearing of the handoff bit is serialized by the wait_lock. So racing is not possible. A rwsem microbenchmark was run for 5 seconds on a 2-socket 40-core 80-thread Skylake system with a v5.1 based kernel and 240 write_lock threads with 5us sleep critical section. Before the patch, the min/mean/max numbers of locking operations for the locking threads were 1/7,792/173,696. After the patch, the figures became 5,842/6,542/7,458. It can be seen that the rwsem became much more fair, though there was a drop of about 16% in the mean locking operations done which was a tradeoff of having better fairness. Making the waiter set the handoff bit right after the first wakeup can impact performance especially with a mixed reader/writer workload. With the same microbenchmark with short critical section and equal number of reader and writer threads (40/40), the reader/writer locking operation counts with the current patch were: 40 readers, Iterations Min/Mean/Max = 1,793/1,794/1,796 40 writers, Iterations Min/Mean/Max = 1,793/34,956/86,081 By making waiter set handoff bit immediately after wakeup: 40 readers, Iterations Min/Mean/Max = 43/44/46 40 writers, Iterations Min/Mean/Max = 43/1,263/3,191 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-8-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:06 +03:00
* If there were already threads queued before us and:
* 1) there are no active locks, wake the front
locking/rwsem: Implement lock handoff to prevent lock starvation Because of writer lock stealing, it is possible that a constant stream of incoming writers will cause a waiting writer or reader to wait indefinitely leading to lock starvation. This patch implements a lock handoff mechanism to disable lock stealing and force lock handoff to the first waiter or waiters (for readers) in the queue after at least a 4ms waiting period unless it is a RT writer task which doesn't need to wait. The waiting period is used to avoid discouraging lock stealing too much to affect performance. The setting and clearing of the handoff bit is serialized by the wait_lock. So racing is not possible. A rwsem microbenchmark was run for 5 seconds on a 2-socket 40-core 80-thread Skylake system with a v5.1 based kernel and 240 write_lock threads with 5us sleep critical section. Before the patch, the min/mean/max numbers of locking operations for the locking threads were 1/7,792/173,696. After the patch, the figures became 5,842/6,542/7,458. It can be seen that the rwsem became much more fair, though there was a drop of about 16% in the mean locking operations done which was a tradeoff of having better fairness. Making the waiter set the handoff bit right after the first wakeup can impact performance especially with a mixed reader/writer workload. With the same microbenchmark with short critical section and equal number of reader and writer threads (40/40), the reader/writer locking operation counts with the current patch were: 40 readers, Iterations Min/Mean/Max = 1,793/1,794/1,796 40 writers, Iterations Min/Mean/Max = 1,793/34,956/86,081 By making waiter set handoff bit immediately after wakeup: 40 readers, Iterations Min/Mean/Max = 43/44/46 40 writers, Iterations Min/Mean/Max = 43/1,263/3,191 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-8-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:06 +03:00
* queued process(es) as the handoff bit might be set.
* 2) there are no active writers and some readers, the lock
* must be read owned; so we try to wake any read lock
* waiters that were queued ahead of us.
*/
locking/rwsem: Implement lock handoff to prevent lock starvation Because of writer lock stealing, it is possible that a constant stream of incoming writers will cause a waiting writer or reader to wait indefinitely leading to lock starvation. This patch implements a lock handoff mechanism to disable lock stealing and force lock handoff to the first waiter or waiters (for readers) in the queue after at least a 4ms waiting period unless it is a RT writer task which doesn't need to wait. The waiting period is used to avoid discouraging lock stealing too much to affect performance. The setting and clearing of the handoff bit is serialized by the wait_lock. So racing is not possible. A rwsem microbenchmark was run for 5 seconds on a 2-socket 40-core 80-thread Skylake system with a v5.1 based kernel and 240 write_lock threads with 5us sleep critical section. Before the patch, the min/mean/max numbers of locking operations for the locking threads were 1/7,792/173,696. After the patch, the figures became 5,842/6,542/7,458. It can be seen that the rwsem became much more fair, though there was a drop of about 16% in the mean locking operations done which was a tradeoff of having better fairness. Making the waiter set the handoff bit right after the first wakeup can impact performance especially with a mixed reader/writer workload. With the same microbenchmark with short critical section and equal number of reader and writer threads (40/40), the reader/writer locking operation counts with the current patch were: 40 readers, Iterations Min/Mean/Max = 1,793/1,794/1,796 40 writers, Iterations Min/Mean/Max = 1,793/34,956/86,081 By making waiter set handoff bit immediately after wakeup: 40 readers, Iterations Min/Mean/Max = 43/44/46 40 writers, Iterations Min/Mean/Max = 43/1,263/3,191 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-8-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:06 +03:00
if (count & RWSEM_WRITER_MASK)
goto wait;
locking/rwsem: Implement lock handoff to prevent lock starvation Because of writer lock stealing, it is possible that a constant stream of incoming writers will cause a waiting writer or reader to wait indefinitely leading to lock starvation. This patch implements a lock handoff mechanism to disable lock stealing and force lock handoff to the first waiter or waiters (for readers) in the queue after at least a 4ms waiting period unless it is a RT writer task which doesn't need to wait. The waiting period is used to avoid discouraging lock stealing too much to affect performance. The setting and clearing of the handoff bit is serialized by the wait_lock. So racing is not possible. A rwsem microbenchmark was run for 5 seconds on a 2-socket 40-core 80-thread Skylake system with a v5.1 based kernel and 240 write_lock threads with 5us sleep critical section. Before the patch, the min/mean/max numbers of locking operations for the locking threads were 1/7,792/173,696. After the patch, the figures became 5,842/6,542/7,458. It can be seen that the rwsem became much more fair, though there was a drop of about 16% in the mean locking operations done which was a tradeoff of having better fairness. Making the waiter set the handoff bit right after the first wakeup can impact performance especially with a mixed reader/writer workload. With the same microbenchmark with short critical section and equal number of reader and writer threads (40/40), the reader/writer locking operation counts with the current patch were: 40 readers, Iterations Min/Mean/Max = 1,793/1,794/1,796 40 writers, Iterations Min/Mean/Max = 1,793/34,956/86,081 By making waiter set handoff bit immediately after wakeup: 40 readers, Iterations Min/Mean/Max = 43/44/46 40 writers, Iterations Min/Mean/Max = 43/1,263/3,191 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-8-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:06 +03:00
rwsem_mark_wake(sem, (count & RWSEM_READER_MASK)
? RWSEM_WAKE_READERS
: RWSEM_WAKE_ANY, &wake_q);
if (!wake_q_empty(&wake_q)) {
/*
* We want to minimize wait_lock hold time especially
* when a large number of readers are to be woken up.
*/
raw_spin_unlock_irq(&sem->wait_lock);
wake_up_q(&wake_q);
wake_q_init(&wake_q); /* Used again, reinit */
raw_spin_lock_irq(&sem->wait_lock);
}
} else {
atomic_long_or(RWSEM_FLAG_WAITERS, &sem->count);
}
locking/rwsem: Implement lock handoff to prevent lock starvation Because of writer lock stealing, it is possible that a constant stream of incoming writers will cause a waiting writer or reader to wait indefinitely leading to lock starvation. This patch implements a lock handoff mechanism to disable lock stealing and force lock handoff to the first waiter or waiters (for readers) in the queue after at least a 4ms waiting period unless it is a RT writer task which doesn't need to wait. The waiting period is used to avoid discouraging lock stealing too much to affect performance. The setting and clearing of the handoff bit is serialized by the wait_lock. So racing is not possible. A rwsem microbenchmark was run for 5 seconds on a 2-socket 40-core 80-thread Skylake system with a v5.1 based kernel and 240 write_lock threads with 5us sleep critical section. Before the patch, the min/mean/max numbers of locking operations for the locking threads were 1/7,792/173,696. After the patch, the figures became 5,842/6,542/7,458. It can be seen that the rwsem became much more fair, though there was a drop of about 16% in the mean locking operations done which was a tradeoff of having better fairness. Making the waiter set the handoff bit right after the first wakeup can impact performance especially with a mixed reader/writer workload. With the same microbenchmark with short critical section and equal number of reader and writer threads (40/40), the reader/writer locking operation counts with the current patch were: 40 readers, Iterations Min/Mean/Max = 1,793/1,794/1,796 40 writers, Iterations Min/Mean/Max = 1,793/34,956/86,081 By making waiter set handoff bit immediately after wakeup: 40 readers, Iterations Min/Mean/Max = 43/44/46 40 writers, Iterations Min/Mean/Max = 43/1,263/3,191 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-8-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:06 +03:00
wait:
/* wait until we successfully acquire the lock */
set_current_state(state);
for (;;) {
if (rwsem_try_write_lock(sem, wstate)) {
/* rwsem_try_write_lock() implies ACQUIRE on success */
break;
}
locking/rwsem: Implement lock handoff to prevent lock starvation Because of writer lock stealing, it is possible that a constant stream of incoming writers will cause a waiting writer or reader to wait indefinitely leading to lock starvation. This patch implements a lock handoff mechanism to disable lock stealing and force lock handoff to the first waiter or waiters (for readers) in the queue after at least a 4ms waiting period unless it is a RT writer task which doesn't need to wait. The waiting period is used to avoid discouraging lock stealing too much to affect performance. The setting and clearing of the handoff bit is serialized by the wait_lock. So racing is not possible. A rwsem microbenchmark was run for 5 seconds on a 2-socket 40-core 80-thread Skylake system with a v5.1 based kernel and 240 write_lock threads with 5us sleep critical section. Before the patch, the min/mean/max numbers of locking operations for the locking threads were 1/7,792/173,696. After the patch, the figures became 5,842/6,542/7,458. It can be seen that the rwsem became much more fair, though there was a drop of about 16% in the mean locking operations done which was a tradeoff of having better fairness. Making the waiter set the handoff bit right after the first wakeup can impact performance especially with a mixed reader/writer workload. With the same microbenchmark with short critical section and equal number of reader and writer threads (40/40), the reader/writer locking operation counts with the current patch were: 40 readers, Iterations Min/Mean/Max = 1,793/1,794/1,796 40 writers, Iterations Min/Mean/Max = 1,793/34,956/86,081 By making waiter set handoff bit immediately after wakeup: 40 readers, Iterations Min/Mean/Max = 43/44/46 40 writers, Iterations Min/Mean/Max = 43/1,263/3,191 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-8-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:06 +03:00
raw_spin_unlock_irq(&sem->wait_lock);
locking/rwsem: Make handoff writer optimistically spin on owner When the handoff bit is set by a writer, no other tasks other than the setting writer itself is allowed to acquire the lock. If the to-be-handoff'ed writer goes to sleep, there will be a wakeup latency period where the lock is free, but no one can acquire it. That is less than ideal. To reduce that latency, the handoff writer will now optimistically spin on the owner if it happens to be a on-cpu writer. It will spin until it releases the lock and the to-be-handoff'ed writer can then acquire the lock immediately without any delay. Of course, if the owner is not a on-cpu writer, the to-be-handoff'ed writer will have to sleep anyway. The optimistic spinning code is also modified to not stop spinning when the handoff bit is set. This will prevent an occasional setting of handoff bit from causing a bunch of optimistic spinners from entering into the wait queue causing significant reduction in throughput. On a 1-socket 22-core 44-thread Skylake system, the AIM7 shared_memory workload was run with 7000 users. The throughput (jobs/min) of the following kernels were as follows: 1) 5.2-rc6 - 8,092,486 2) 5.2-rc6 + tip's rwsem patches - 7,567,568 3) 5.2-rc6 + tip's rwsem patches + this patch - 7,954,545 Using perf-record(1), the %cpu time used by rwsem_down_write_slowpath(), rwsem_down_write_failed() and their callees for the 3 kernels were 1.70%, 5.46% and 2.08% respectively. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: x86@kernel.org Cc: Ingo Molnar <mingo@redhat.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: "H. Peter Anvin" <hpa@zytor.com> Link: https://lkml.kernel.org/r/20190625143913.24154-1-longman@redhat.com
2019-06-25 17:39:13 +03:00
/*
* After setting the handoff bit and failing to acquire
* the lock, attempt to spin on owner to accelerate lock
* transfer. If the previous owner is a on-cpu writer and it
* has just released the lock, OWNER_NULL will be returned.
* In this case, we attempt to acquire the lock again
* without sleeping.
*/
if (wstate == WRITER_HANDOFF &&
locking/rwsem: Remove reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It also improves the chance that a reader can get the lock as writer optimistic spinning disproportionally favors writers much more than readers. Since commit d3681e269fff ("locking/rwsem: Wake up almost all readers in wait queue"), all the waiting readers are woken up so that they can all get the read lock and run in parallel. When the number of contending readers is large, allowing reader optimistic spinning will likely cause reader fragmentation where multiple smaller groups of readers can get the read lock in a sequential manner separated by writers. That reduces reader parallelism. One possible way to address that drawback is to limit the number of readers (preferably one) that can do optimistic spinning. These readers act as representatives of all the waiting readers in the wait queue as they will wake up all those waiting readers once they get the lock. Alternatively, as reader optimistic lock stealing has already enhanced fairness to readers, it may be easier to just remove reader optimistic spinning and simplifying the optimistic spinning code as a result. Performance measurements (locking throughput kops/s) using a locking microbenchmark with 50/50 reader/writer distribution and turbo-boost disabled was done on a 2-socket Cascade Lake system (48-core 96-thread) to see the impacts of these changes: 1) Vanilla - 5.10-rc3 kernel 2) Before - 5.10-rc3 kernel with previous patches in this series 2) limit-rspin - 5.10-rc3 kernel with limited reader spinning patch 3) no-rspin - 5.10-rc3 kernel with reader spinning disabled # of threads CS Load Vanilla Before limit-rspin no-rspin ------------ ------- ------- ------ ----------- -------- 2 1 5,185 5,662 5,214 5,077 4 1 5,107 4,983 5,188 4,760 8 1 4,782 4,564 4,720 4,628 16 1 4,680 4,053 4,567 3,402 32 1 4,299 1,115 1,118 1,098 64 1 3,218 983 1,001 957 96 1 1,938 944 957 930 2 20 2,008 2,128 2,264 1,665 4 20 1,390 1,033 1,046 1,101 8 20 1,472 1,155 1,098 1,213 16 20 1,332 1,077 1,089 1,122 32 20 967 914 917 980 64 20 787 874 891 858 96 20 730 836 847 844 2 100 372 356 360 355 4 100 492 425 434 392 8 100 533 537 529 538 16 100 548 572 568 598 32 100 499 520 527 537 64 100 466 517 526 512 96 100 406 497 506 509 The column "CS Load" represents the number of pause instructions issued in the locking critical section. A CS load of 1 is extremely short and is not likey in real situations. A load of 20 (moderate) and 100 (long) are more realistic. It can be seen that the previous patches in this series have reduced performance in general except in highly contended cases with moderate or long critical sections that performance improves a bit. This change is mostly caused by the "Prevent potential lock starvation" patch that reduce reader optimistic spinning and hence reduce reader fragmentation. The patch that further limit reader optimistic spinning doesn't seem to have too much impact on overall performance as shown in the benchmark data. The patch that disables reader optimistic spinning shows reduced performance at lightly loaded cases, but comparable or slightly better performance on with heavier contention. This patch just removes reader optimistic spinning for now. As readers are not going to do optimistic spinning anymore, we don't need to consider if the OSQ is empty or not when doing lock stealing. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-6-longman@redhat.com
2020-11-21 07:14:16 +03:00
rwsem_spin_on_owner(sem) == OWNER_NULL)
locking/rwsem: Make handoff writer optimistically spin on owner When the handoff bit is set by a writer, no other tasks other than the setting writer itself is allowed to acquire the lock. If the to-be-handoff'ed writer goes to sleep, there will be a wakeup latency period where the lock is free, but no one can acquire it. That is less than ideal. To reduce that latency, the handoff writer will now optimistically spin on the owner if it happens to be a on-cpu writer. It will spin until it releases the lock and the to-be-handoff'ed writer can then acquire the lock immediately without any delay. Of course, if the owner is not a on-cpu writer, the to-be-handoff'ed writer will have to sleep anyway. The optimistic spinning code is also modified to not stop spinning when the handoff bit is set. This will prevent an occasional setting of handoff bit from causing a bunch of optimistic spinners from entering into the wait queue causing significant reduction in throughput. On a 1-socket 22-core 44-thread Skylake system, the AIM7 shared_memory workload was run with 7000 users. The throughput (jobs/min) of the following kernels were as follows: 1) 5.2-rc6 - 8,092,486 2) 5.2-rc6 + tip's rwsem patches - 7,567,568 3) 5.2-rc6 + tip's rwsem patches + this patch - 7,954,545 Using perf-record(1), the %cpu time used by rwsem_down_write_slowpath(), rwsem_down_write_failed() and their callees for the 3 kernels were 1.70%, 5.46% and 2.08% respectively. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: x86@kernel.org Cc: Ingo Molnar <mingo@redhat.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: "H. Peter Anvin" <hpa@zytor.com> Link: https://lkml.kernel.org/r/20190625143913.24154-1-longman@redhat.com
2019-06-25 17:39:13 +03:00
goto trylock_again;
/* Block until there are no active lockers. */
locking/rwsem: Implement lock handoff to prevent lock starvation Because of writer lock stealing, it is possible that a constant stream of incoming writers will cause a waiting writer or reader to wait indefinitely leading to lock starvation. This patch implements a lock handoff mechanism to disable lock stealing and force lock handoff to the first waiter or waiters (for readers) in the queue after at least a 4ms waiting period unless it is a RT writer task which doesn't need to wait. The waiting period is used to avoid discouraging lock stealing too much to affect performance. The setting and clearing of the handoff bit is serialized by the wait_lock. So racing is not possible. A rwsem microbenchmark was run for 5 seconds on a 2-socket 40-core 80-thread Skylake system with a v5.1 based kernel and 240 write_lock threads with 5us sleep critical section. Before the patch, the min/mean/max numbers of locking operations for the locking threads were 1/7,792/173,696. After the patch, the figures became 5,842/6,542/7,458. It can be seen that the rwsem became much more fair, though there was a drop of about 16% in the mean locking operations done which was a tradeoff of having better fairness. Making the waiter set the handoff bit right after the first wakeup can impact performance especially with a mixed reader/writer workload. With the same microbenchmark with short critical section and equal number of reader and writer threads (40/40), the reader/writer locking operation counts with the current patch were: 40 readers, Iterations Min/Mean/Max = 1,793/1,794/1,796 40 writers, Iterations Min/Mean/Max = 1,793/34,956/86,081 By making waiter set handoff bit immediately after wakeup: 40 readers, Iterations Min/Mean/Max = 43/44/46 40 writers, Iterations Min/Mean/Max = 43/1,263/3,191 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-8-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:06 +03:00
for (;;) {
if (signal_pending_state(state, current))
goto out_nolock;
schedule();
lockevent_inc(rwsem_sleep_writer);
set_current_state(state);
locking/rwsem: Implement lock handoff to prevent lock starvation Because of writer lock stealing, it is possible that a constant stream of incoming writers will cause a waiting writer or reader to wait indefinitely leading to lock starvation. This patch implements a lock handoff mechanism to disable lock stealing and force lock handoff to the first waiter or waiters (for readers) in the queue after at least a 4ms waiting period unless it is a RT writer task which doesn't need to wait. The waiting period is used to avoid discouraging lock stealing too much to affect performance. The setting and clearing of the handoff bit is serialized by the wait_lock. So racing is not possible. A rwsem microbenchmark was run for 5 seconds on a 2-socket 40-core 80-thread Skylake system with a v5.1 based kernel and 240 write_lock threads with 5us sleep critical section. Before the patch, the min/mean/max numbers of locking operations for the locking threads were 1/7,792/173,696. After the patch, the figures became 5,842/6,542/7,458. It can be seen that the rwsem became much more fair, though there was a drop of about 16% in the mean locking operations done which was a tradeoff of having better fairness. Making the waiter set the handoff bit right after the first wakeup can impact performance especially with a mixed reader/writer workload. With the same microbenchmark with short critical section and equal number of reader and writer threads (40/40), the reader/writer locking operation counts with the current patch were: 40 readers, Iterations Min/Mean/Max = 1,793/1,794/1,796 40 writers, Iterations Min/Mean/Max = 1,793/34,956/86,081 By making waiter set handoff bit immediately after wakeup: 40 readers, Iterations Min/Mean/Max = 43/44/46 40 writers, Iterations Min/Mean/Max = 43/1,263/3,191 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-8-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:06 +03:00
/*
* If HANDOFF bit is set, unconditionally do
* a trylock.
*/
if (wstate == WRITER_HANDOFF)
break;
if ((wstate == WRITER_NOT_FIRST) &&
(rwsem_first_waiter(sem) == &waiter))
wstate = WRITER_FIRST;
count = atomic_long_read(&sem->count);
locking/rwsem: Implement lock handoff to prevent lock starvation Because of writer lock stealing, it is possible that a constant stream of incoming writers will cause a waiting writer or reader to wait indefinitely leading to lock starvation. This patch implements a lock handoff mechanism to disable lock stealing and force lock handoff to the first waiter or waiters (for readers) in the queue after at least a 4ms waiting period unless it is a RT writer task which doesn't need to wait. The waiting period is used to avoid discouraging lock stealing too much to affect performance. The setting and clearing of the handoff bit is serialized by the wait_lock. So racing is not possible. A rwsem microbenchmark was run for 5 seconds on a 2-socket 40-core 80-thread Skylake system with a v5.1 based kernel and 240 write_lock threads with 5us sleep critical section. Before the patch, the min/mean/max numbers of locking operations for the locking threads were 1/7,792/173,696. After the patch, the figures became 5,842/6,542/7,458. It can be seen that the rwsem became much more fair, though there was a drop of about 16% in the mean locking operations done which was a tradeoff of having better fairness. Making the waiter set the handoff bit right after the first wakeup can impact performance especially with a mixed reader/writer workload. With the same microbenchmark with short critical section and equal number of reader and writer threads (40/40), the reader/writer locking operation counts with the current patch were: 40 readers, Iterations Min/Mean/Max = 1,793/1,794/1,796 40 writers, Iterations Min/Mean/Max = 1,793/34,956/86,081 By making waiter set handoff bit immediately after wakeup: 40 readers, Iterations Min/Mean/Max = 43/44/46 40 writers, Iterations Min/Mean/Max = 43/1,263/3,191 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-8-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:06 +03:00
if (!(count & RWSEM_LOCK_MASK))
break;
/*
* The setting of the handoff bit is deferred
* until rwsem_try_write_lock() is called.
*/
if ((wstate == WRITER_FIRST) && (rt_task(current) ||
time_after(jiffies, waiter.timeout))) {
wstate = WRITER_HANDOFF;
lockevent_inc(rwsem_wlock_handoff);
break;
}
}
locking/rwsem: Make handoff writer optimistically spin on owner When the handoff bit is set by a writer, no other tasks other than the setting writer itself is allowed to acquire the lock. If the to-be-handoff'ed writer goes to sleep, there will be a wakeup latency period where the lock is free, but no one can acquire it. That is less than ideal. To reduce that latency, the handoff writer will now optimistically spin on the owner if it happens to be a on-cpu writer. It will spin until it releases the lock and the to-be-handoff'ed writer can then acquire the lock immediately without any delay. Of course, if the owner is not a on-cpu writer, the to-be-handoff'ed writer will have to sleep anyway. The optimistic spinning code is also modified to not stop spinning when the handoff bit is set. This will prevent an occasional setting of handoff bit from causing a bunch of optimistic spinners from entering into the wait queue causing significant reduction in throughput. On a 1-socket 22-core 44-thread Skylake system, the AIM7 shared_memory workload was run with 7000 users. The throughput (jobs/min) of the following kernels were as follows: 1) 5.2-rc6 - 8,092,486 2) 5.2-rc6 + tip's rwsem patches - 7,567,568 3) 5.2-rc6 + tip's rwsem patches + this patch - 7,954,545 Using perf-record(1), the %cpu time used by rwsem_down_write_slowpath(), rwsem_down_write_failed() and their callees for the 3 kernels were 1.70%, 5.46% and 2.08% respectively. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: x86@kernel.org Cc: Ingo Molnar <mingo@redhat.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: "H. Peter Anvin" <hpa@zytor.com> Link: https://lkml.kernel.org/r/20190625143913.24154-1-longman@redhat.com
2019-06-25 17:39:13 +03:00
trylock_again:
raw_spin_lock_irq(&sem->wait_lock);
}
__set_current_state(TASK_RUNNING);
list_del(&waiter.list);
raw_spin_unlock_irq(&sem->wait_lock);
lockevent_inc(rwsem_wlock);
return ret;
out_nolock:
__set_current_state(TASK_RUNNING);
raw_spin_lock_irq(&sem->wait_lock);
list_del(&waiter.list);
locking/rwsem: Implement lock handoff to prevent lock starvation Because of writer lock stealing, it is possible that a constant stream of incoming writers will cause a waiting writer or reader to wait indefinitely leading to lock starvation. This patch implements a lock handoff mechanism to disable lock stealing and force lock handoff to the first waiter or waiters (for readers) in the queue after at least a 4ms waiting period unless it is a RT writer task which doesn't need to wait. The waiting period is used to avoid discouraging lock stealing too much to affect performance. The setting and clearing of the handoff bit is serialized by the wait_lock. So racing is not possible. A rwsem microbenchmark was run for 5 seconds on a 2-socket 40-core 80-thread Skylake system with a v5.1 based kernel and 240 write_lock threads with 5us sleep critical section. Before the patch, the min/mean/max numbers of locking operations for the locking threads were 1/7,792/173,696. After the patch, the figures became 5,842/6,542/7,458. It can be seen that the rwsem became much more fair, though there was a drop of about 16% in the mean locking operations done which was a tradeoff of having better fairness. Making the waiter set the handoff bit right after the first wakeup can impact performance especially with a mixed reader/writer workload. With the same microbenchmark with short critical section and equal number of reader and writer threads (40/40), the reader/writer locking operation counts with the current patch were: 40 readers, Iterations Min/Mean/Max = 1,793/1,794/1,796 40 writers, Iterations Min/Mean/Max = 1,793/34,956/86,081 By making waiter set handoff bit immediately after wakeup: 40 readers, Iterations Min/Mean/Max = 43/44/46 40 writers, Iterations Min/Mean/Max = 43/1,263/3,191 Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-8-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:06 +03:00
if (unlikely(wstate == WRITER_HANDOFF))
atomic_long_add(-RWSEM_FLAG_HANDOFF, &sem->count);
if (list_empty(&sem->wait_list))
atomic_long_andnot(RWSEM_FLAG_WAITERS, &sem->count);
else
locking/rwsem: Code cleanup after files merging After merging all the relevant rwsem code into one single file, there are a number of optimizations and cleanups that can be done: 1) Remove all the EXPORT_SYMBOL() calls for functions that are not accessed elsewhere. 2) Remove all the __visible tags as none of the functions will be called from assembly code anymore. 3) Make all the internal functions static. 4) Remove some unneeded blank lines. 5) Remove the intermediate rwsem_down_{read|write}_failed*() functions and rename __rwsem_down_{read|write}_failed_common() to rwsem_down_{read|write}_slowpath(). 6) Remove "__" prefix of __rwsem_mark_wake(). 7) Use atomic_long_try_cmpxchg_acquire() as much as possible. 8) Remove the rwsem_rtrylock and rwsem_wtrylock lock events as they are not that useful. That enables the compiler to do better optimization and reduce code size. The text+data size of rwsem.o on an x86-64 machine with gcc8 was reduced from 10237 bytes to 5030 bytes with this change. Suggested-by: Peter Zijlstra <peterz@infradead.org> Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-6-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:04 +03:00
rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q);
raw_spin_unlock_irq(&sem->wait_lock);
wake_up_q(&wake_q);
lockevent_inc(rwsem_wlock_fail);
return ERR_PTR(-EINTR);
}
/*
* handle waking up a waiter on the semaphore
* - up_read/up_write has decremented the active part of count if we come here
*/
static struct rw_semaphore *rwsem_wake(struct rw_semaphore *sem)
{
unsigned long flags;
DEFINE_WAKE_Q(wake_q);
raw_spin_lock_irqsave(&sem->wait_lock, flags);
if (!list_empty(&sem->wait_list))
locking/rwsem: Code cleanup after files merging After merging all the relevant rwsem code into one single file, there are a number of optimizations and cleanups that can be done: 1) Remove all the EXPORT_SYMBOL() calls for functions that are not accessed elsewhere. 2) Remove all the __visible tags as none of the functions will be called from assembly code anymore. 3) Make all the internal functions static. 4) Remove some unneeded blank lines. 5) Remove the intermediate rwsem_down_{read|write}_failed*() functions and rename __rwsem_down_{read|write}_failed_common() to rwsem_down_{read|write}_slowpath(). 6) Remove "__" prefix of __rwsem_mark_wake(). 7) Use atomic_long_try_cmpxchg_acquire() as much as possible. 8) Remove the rwsem_rtrylock and rwsem_wtrylock lock events as they are not that useful. That enables the compiler to do better optimization and reduce code size. The text+data size of rwsem.o on an x86-64 machine with gcc8 was reduced from 10237 bytes to 5030 bytes with this change. Suggested-by: Peter Zijlstra <peterz@infradead.org> Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-6-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:04 +03:00
rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q);
raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
wake_up_q(&wake_q);
return sem;
}
/*
* downgrade a write lock into a read lock
* - caller incremented waiting part of count and discovered it still negative
* - just wake up any readers at the front of the queue
*/
locking/rwsem: Code cleanup after files merging After merging all the relevant rwsem code into one single file, there are a number of optimizations and cleanups that can be done: 1) Remove all the EXPORT_SYMBOL() calls for functions that are not accessed elsewhere. 2) Remove all the __visible tags as none of the functions will be called from assembly code anymore. 3) Make all the internal functions static. 4) Remove some unneeded blank lines. 5) Remove the intermediate rwsem_down_{read|write}_failed*() functions and rename __rwsem_down_{read|write}_failed_common() to rwsem_down_{read|write}_slowpath(). 6) Remove "__" prefix of __rwsem_mark_wake(). 7) Use atomic_long_try_cmpxchg_acquire() as much as possible. 8) Remove the rwsem_rtrylock and rwsem_wtrylock lock events as they are not that useful. That enables the compiler to do better optimization and reduce code size. The text+data size of rwsem.o on an x86-64 machine with gcc8 was reduced from 10237 bytes to 5030 bytes with this change. Suggested-by: Peter Zijlstra <peterz@infradead.org> Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-6-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:04 +03:00
static struct rw_semaphore *rwsem_downgrade_wake(struct rw_semaphore *sem)
{
unsigned long flags;
DEFINE_WAKE_Q(wake_q);
raw_spin_lock_irqsave(&sem->wait_lock, flags);
if (!list_empty(&sem->wait_list))
locking/rwsem: Code cleanup after files merging After merging all the relevant rwsem code into one single file, there are a number of optimizations and cleanups that can be done: 1) Remove all the EXPORT_SYMBOL() calls for functions that are not accessed elsewhere. 2) Remove all the __visible tags as none of the functions will be called from assembly code anymore. 3) Make all the internal functions static. 4) Remove some unneeded blank lines. 5) Remove the intermediate rwsem_down_{read|write}_failed*() functions and rename __rwsem_down_{read|write}_failed_common() to rwsem_down_{read|write}_slowpath(). 6) Remove "__" prefix of __rwsem_mark_wake(). 7) Use atomic_long_try_cmpxchg_acquire() as much as possible. 8) Remove the rwsem_rtrylock and rwsem_wtrylock lock events as they are not that useful. That enables the compiler to do better optimization and reduce code size. The text+data size of rwsem.o on an x86-64 machine with gcc8 was reduced from 10237 bytes to 5030 bytes with this change. Suggested-by: Peter Zijlstra <peterz@infradead.org> Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-6-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:04 +03:00
rwsem_mark_wake(sem, RWSEM_WAKE_READ_OWNED, &wake_q);
raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
wake_up_q(&wake_q);
return sem;
}
/*
* lock for reading
*/
static inline int __down_read_common(struct rw_semaphore *sem, int state)
{
long count;
if (!rwsem_read_trylock(sem, &count)) {
if (IS_ERR(rwsem_down_read_slowpath(sem, count, state)))
return -EINTR;
DEBUG_RWSEMS_WARN_ON(!is_rwsem_reader_owned(sem), sem);
}
return 0;
}
static inline void __down_read(struct rw_semaphore *sem)
{
__down_read_common(sem, TASK_UNINTERRUPTIBLE);
}
static inline int __down_read_interruptible(struct rw_semaphore *sem)
{
return __down_read_common(sem, TASK_INTERRUPTIBLE);
}
static inline int __down_read_killable(struct rw_semaphore *sem)
{
return __down_read_common(sem, TASK_KILLABLE);
}
static inline int __down_read_trylock(struct rw_semaphore *sem)
{
long tmp;
DEBUG_RWSEMS_WARN_ON(sem->magic != sem, sem);
/*
* Optimize for the case when the rwsem is not locked at all.
*/
tmp = RWSEM_UNLOCKED_VALUE;
do {
if (atomic_long_try_cmpxchg_acquire(&sem->count, &tmp,
tmp + RWSEM_READER_BIAS)) {
rwsem_set_reader_owned(sem);
return 1;
}
} while (!(tmp & RWSEM_READ_FAILED_MASK));
return 0;
}
/*
* lock for writing
*/
static inline int __down_write_common(struct rw_semaphore *sem, int state)
{
if (unlikely(!rwsem_write_trylock(sem))) {
if (IS_ERR(rwsem_down_write_slowpath(sem, state)))
return -EINTR;
locking/rwsem: Code cleanup after files merging After merging all the relevant rwsem code into one single file, there are a number of optimizations and cleanups that can be done: 1) Remove all the EXPORT_SYMBOL() calls for functions that are not accessed elsewhere. 2) Remove all the __visible tags as none of the functions will be called from assembly code anymore. 3) Make all the internal functions static. 4) Remove some unneeded blank lines. 5) Remove the intermediate rwsem_down_{read|write}_failed*() functions and rename __rwsem_down_{read|write}_failed_common() to rwsem_down_{read|write}_slowpath(). 6) Remove "__" prefix of __rwsem_mark_wake(). 7) Use atomic_long_try_cmpxchg_acquire() as much as possible. 8) Remove the rwsem_rtrylock and rwsem_wtrylock lock events as they are not that useful. That enables the compiler to do better optimization and reduce code size. The text+data size of rwsem.o on an x86-64 machine with gcc8 was reduced from 10237 bytes to 5030 bytes with this change. Suggested-by: Peter Zijlstra <peterz@infradead.org> Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-6-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:04 +03:00
}
return 0;
}
static inline void __down_write(struct rw_semaphore *sem)
{
__down_write_common(sem, TASK_UNINTERRUPTIBLE);
}
static inline int __down_write_killable(struct rw_semaphore *sem)
{
return __down_write_common(sem, TASK_KILLABLE);
}
static inline int __down_write_trylock(struct rw_semaphore *sem)
{
DEBUG_RWSEMS_WARN_ON(sem->magic != sem, sem);
return rwsem_write_trylock(sem);
}
/*
* unlock after reading
*/
static inline void __up_read(struct rw_semaphore *sem)
{
long tmp;
DEBUG_RWSEMS_WARN_ON(sem->magic != sem, sem);
DEBUG_RWSEMS_WARN_ON(!is_rwsem_reader_owned(sem), sem);
rwsem_clear_reader_owned(sem);
tmp = atomic_long_add_return_release(-RWSEM_READER_BIAS, &sem->count);
DEBUG_RWSEMS_WARN_ON(tmp < 0, sem);
locking/rwsem: Code cleanup after files merging After merging all the relevant rwsem code into one single file, there are a number of optimizations and cleanups that can be done: 1) Remove all the EXPORT_SYMBOL() calls for functions that are not accessed elsewhere. 2) Remove all the __visible tags as none of the functions will be called from assembly code anymore. 3) Make all the internal functions static. 4) Remove some unneeded blank lines. 5) Remove the intermediate rwsem_down_{read|write}_failed*() functions and rename __rwsem_down_{read|write}_failed_common() to rwsem_down_{read|write}_slowpath(). 6) Remove "__" prefix of __rwsem_mark_wake(). 7) Use atomic_long_try_cmpxchg_acquire() as much as possible. 8) Remove the rwsem_rtrylock and rwsem_wtrylock lock events as they are not that useful. That enables the compiler to do better optimization and reduce code size. The text+data size of rwsem.o on an x86-64 machine with gcc8 was reduced from 10237 bytes to 5030 bytes with this change. Suggested-by: Peter Zijlstra <peterz@infradead.org> Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-6-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:04 +03:00
if (unlikely((tmp & (RWSEM_LOCK_MASK|RWSEM_FLAG_WAITERS)) ==
locking/rwsem: Enable time-based spinning on reader-owned rwsem When the rwsem is owned by reader, writers stop optimistic spinning simply because there is no easy way to figure out if all the readers are actively running or not. However, there are scenarios where the readers are unlikely to sleep and optimistic spinning can help performance. This patch provides a simple mechanism for spinning on a reader-owned rwsem by a writer. It is a time threshold based spinning where the allowable spinning time can vary from 10us to 25us depending on the condition of the rwsem. When the time threshold is exceeded, the nonspinnable bits will be set in the owner field to indicate that no more optimistic spinning will be allowed on this rwsem until it becomes writer owned again. Not even readers is allowed to acquire the reader-locked rwsem by optimistic spinning for fairness. We also want a writer to acquire the lock after the readers hold the lock for a relatively long time. In order to give preference to writers under such a circumstance, the single RWSEM_NONSPINNABLE bit is now split into two - one for reader and one for writer. When optimistic spinning is disabled, both bits will be set. When the reader count drop down to 0, the writer nonspinnable bit will be cleared to allow writers to spin on the lock, but not the readers. When a writer acquires the lock, it will write its own task structure pointer into sem->owner and clear the reader nonspinnable bit in the process. The time taken for each iteration of the reader-owned rwsem spinning loop varies. Below are sample minimum elapsed times for 16 iterations of the loop. System Time for 16 Iterations ------ ---------------------- 1-socket Skylake ~800ns 4-socket Broadwell ~300ns 2-socket ThunderX2 (arm64) ~250ns When the lock cacheline is contended, we can see up to almost 10X increase in elapsed time. So 25us will be at most 500, 1300 and 1600 iterations for each of the above systems. With a locking microbenchmark running on 5.1 based kernel, the total locking rates (in kops/s) on a 8-socket IvyBridge-EX system with equal numbers of readers and writers before and after this patch were as follows: # of Threads Pre-patch Post-patch ------------ --------- ---------- 2 1,759 6,684 4 1,684 6,738 8 1,074 7,222 16 900 7,163 32 458 7,316 64 208 520 128 168 425 240 143 474 This patch gives a big boost in performance for mixed reader/writer workloads. With 32 locking threads, the rwsem lock event data were: rwsem_opt_fail=79850 rwsem_opt_nospin=5069 rwsem_opt_rlock=597484 rwsem_opt_wlock=957339 rwsem_sleep_reader=57782 rwsem_sleep_writer=55663 With 64 locking threads, the data looked like: rwsem_opt_fail=346723 rwsem_opt_nospin=6293 rwsem_opt_rlock=1127119 rwsem_opt_wlock=1400628 rwsem_sleep_reader=308201 rwsem_sleep_writer=72281 So a lot more threads acquired the lock in the slowpath and more threads went to sleep. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-15-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:13 +03:00
RWSEM_FLAG_WAITERS)) {
locking/rwsem: Remove reader optimistic spinning Reader optimistic spinning is helpful when the reader critical section is short and there aren't that many readers around. It also improves the chance that a reader can get the lock as writer optimistic spinning disproportionally favors writers much more than readers. Since commit d3681e269fff ("locking/rwsem: Wake up almost all readers in wait queue"), all the waiting readers are woken up so that they can all get the read lock and run in parallel. When the number of contending readers is large, allowing reader optimistic spinning will likely cause reader fragmentation where multiple smaller groups of readers can get the read lock in a sequential manner separated by writers. That reduces reader parallelism. One possible way to address that drawback is to limit the number of readers (preferably one) that can do optimistic spinning. These readers act as representatives of all the waiting readers in the wait queue as they will wake up all those waiting readers once they get the lock. Alternatively, as reader optimistic lock stealing has already enhanced fairness to readers, it may be easier to just remove reader optimistic spinning and simplifying the optimistic spinning code as a result. Performance measurements (locking throughput kops/s) using a locking microbenchmark with 50/50 reader/writer distribution and turbo-boost disabled was done on a 2-socket Cascade Lake system (48-core 96-thread) to see the impacts of these changes: 1) Vanilla - 5.10-rc3 kernel 2) Before - 5.10-rc3 kernel with previous patches in this series 2) limit-rspin - 5.10-rc3 kernel with limited reader spinning patch 3) no-rspin - 5.10-rc3 kernel with reader spinning disabled # of threads CS Load Vanilla Before limit-rspin no-rspin ------------ ------- ------- ------ ----------- -------- 2 1 5,185 5,662 5,214 5,077 4 1 5,107 4,983 5,188 4,760 8 1 4,782 4,564 4,720 4,628 16 1 4,680 4,053 4,567 3,402 32 1 4,299 1,115 1,118 1,098 64 1 3,218 983 1,001 957 96 1 1,938 944 957 930 2 20 2,008 2,128 2,264 1,665 4 20 1,390 1,033 1,046 1,101 8 20 1,472 1,155 1,098 1,213 16 20 1,332 1,077 1,089 1,122 32 20 967 914 917 980 64 20 787 874 891 858 96 20 730 836 847 844 2 100 372 356 360 355 4 100 492 425 434 392 8 100 533 537 529 538 16 100 548 572 568 598 32 100 499 520 527 537 64 100 466 517 526 512 96 100 406 497 506 509 The column "CS Load" represents the number of pause instructions issued in the locking critical section. A CS load of 1 is extremely short and is not likey in real situations. A load of 20 (moderate) and 100 (long) are more realistic. It can be seen that the previous patches in this series have reduced performance in general except in highly contended cases with moderate or long critical sections that performance improves a bit. This change is mostly caused by the "Prevent potential lock starvation" patch that reduce reader optimistic spinning and hence reduce reader fragmentation. The patch that further limit reader optimistic spinning doesn't seem to have too much impact on overall performance as shown in the benchmark data. The patch that disables reader optimistic spinning shows reduced performance at lightly loaded cases, but comparable or slightly better performance on with heavier contention. This patch just removes reader optimistic spinning for now. As readers are not going to do optimistic spinning anymore, we don't need to consider if the OSQ is empty or not when doing lock stealing. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Davidlohr Bueso <dbueso@suse.de> Link: https://lkml.kernel.org/r/20201121041416.12285-6-longman@redhat.com
2020-11-21 07:14:16 +03:00
clear_nonspinnable(sem);
rwsem_wake(sem);
locking/rwsem: Enable time-based spinning on reader-owned rwsem When the rwsem is owned by reader, writers stop optimistic spinning simply because there is no easy way to figure out if all the readers are actively running or not. However, there are scenarios where the readers are unlikely to sleep and optimistic spinning can help performance. This patch provides a simple mechanism for spinning on a reader-owned rwsem by a writer. It is a time threshold based spinning where the allowable spinning time can vary from 10us to 25us depending on the condition of the rwsem. When the time threshold is exceeded, the nonspinnable bits will be set in the owner field to indicate that no more optimistic spinning will be allowed on this rwsem until it becomes writer owned again. Not even readers is allowed to acquire the reader-locked rwsem by optimistic spinning for fairness. We also want a writer to acquire the lock after the readers hold the lock for a relatively long time. In order to give preference to writers under such a circumstance, the single RWSEM_NONSPINNABLE bit is now split into two - one for reader and one for writer. When optimistic spinning is disabled, both bits will be set. When the reader count drop down to 0, the writer nonspinnable bit will be cleared to allow writers to spin on the lock, but not the readers. When a writer acquires the lock, it will write its own task structure pointer into sem->owner and clear the reader nonspinnable bit in the process. The time taken for each iteration of the reader-owned rwsem spinning loop varies. Below are sample minimum elapsed times for 16 iterations of the loop. System Time for 16 Iterations ------ ---------------------- 1-socket Skylake ~800ns 4-socket Broadwell ~300ns 2-socket ThunderX2 (arm64) ~250ns When the lock cacheline is contended, we can see up to almost 10X increase in elapsed time. So 25us will be at most 500, 1300 and 1600 iterations for each of the above systems. With a locking microbenchmark running on 5.1 based kernel, the total locking rates (in kops/s) on a 8-socket IvyBridge-EX system with equal numbers of readers and writers before and after this patch were as follows: # of Threads Pre-patch Post-patch ------------ --------- ---------- 2 1,759 6,684 4 1,684 6,738 8 1,074 7,222 16 900 7,163 32 458 7,316 64 208 520 128 168 425 240 143 474 This patch gives a big boost in performance for mixed reader/writer workloads. With 32 locking threads, the rwsem lock event data were: rwsem_opt_fail=79850 rwsem_opt_nospin=5069 rwsem_opt_rlock=597484 rwsem_opt_wlock=957339 rwsem_sleep_reader=57782 rwsem_sleep_writer=55663 With 64 locking threads, the data looked like: rwsem_opt_fail=346723 rwsem_opt_nospin=6293 rwsem_opt_rlock=1127119 rwsem_opt_wlock=1400628 rwsem_sleep_reader=308201 rwsem_sleep_writer=72281 So a lot more threads acquired the lock in the slowpath and more threads went to sleep. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-15-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:13 +03:00
}
}
/*
* unlock after writing
*/
static inline void __up_write(struct rw_semaphore *sem)
{
locking/rwsem: Code cleanup after files merging After merging all the relevant rwsem code into one single file, there are a number of optimizations and cleanups that can be done: 1) Remove all the EXPORT_SYMBOL() calls for functions that are not accessed elsewhere. 2) Remove all the __visible tags as none of the functions will be called from assembly code anymore. 3) Make all the internal functions static. 4) Remove some unneeded blank lines. 5) Remove the intermediate rwsem_down_{read|write}_failed*() functions and rename __rwsem_down_{read|write}_failed_common() to rwsem_down_{read|write}_slowpath(). 6) Remove "__" prefix of __rwsem_mark_wake(). 7) Use atomic_long_try_cmpxchg_acquire() as much as possible. 8) Remove the rwsem_rtrylock and rwsem_wtrylock lock events as they are not that useful. That enables the compiler to do better optimization and reduce code size. The text+data size of rwsem.o on an x86-64 machine with gcc8 was reduced from 10237 bytes to 5030 bytes with this change. Suggested-by: Peter Zijlstra <peterz@infradead.org> Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-6-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:04 +03:00
long tmp;
DEBUG_RWSEMS_WARN_ON(sem->magic != sem, sem);
/*
* sem->owner may differ from current if the ownership is transferred
* to an anonymous writer by setting the RWSEM_NONSPINNABLE bits.
*/
DEBUG_RWSEMS_WARN_ON((rwsem_owner(sem) != current) &&
!rwsem_test_oflags(sem, RWSEM_NONSPINNABLE), sem);
rwsem_clear_owner(sem);
locking/rwsem: Code cleanup after files merging After merging all the relevant rwsem code into one single file, there are a number of optimizations and cleanups that can be done: 1) Remove all the EXPORT_SYMBOL() calls for functions that are not accessed elsewhere. 2) Remove all the __visible tags as none of the functions will be called from assembly code anymore. 3) Make all the internal functions static. 4) Remove some unneeded blank lines. 5) Remove the intermediate rwsem_down_{read|write}_failed*() functions and rename __rwsem_down_{read|write}_failed_common() to rwsem_down_{read|write}_slowpath(). 6) Remove "__" prefix of __rwsem_mark_wake(). 7) Use atomic_long_try_cmpxchg_acquire() as much as possible. 8) Remove the rwsem_rtrylock and rwsem_wtrylock lock events as they are not that useful. That enables the compiler to do better optimization and reduce code size. The text+data size of rwsem.o on an x86-64 machine with gcc8 was reduced from 10237 bytes to 5030 bytes with this change. Suggested-by: Peter Zijlstra <peterz@infradead.org> Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-6-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:04 +03:00
tmp = atomic_long_fetch_add_release(-RWSEM_WRITER_LOCKED, &sem->count);
if (unlikely(tmp & RWSEM_FLAG_WAITERS))
rwsem_wake(sem);
}
/*
* downgrade write lock to read lock
*/
static inline void __downgrade_write(struct rw_semaphore *sem)
{
long tmp;
/*
* When downgrading from exclusive to shared ownership,
* anything inside the write-locked region cannot leak
* into the read side. In contrast, anything in the
* read-locked region is ok to be re-ordered into the
* write side. As such, rely on RELEASE semantics.
*/
DEBUG_RWSEMS_WARN_ON(rwsem_owner(sem) != current, sem);
tmp = atomic_long_fetch_add_release(
-RWSEM_WRITER_LOCKED+RWSEM_READER_BIAS, &sem->count);
rwsem_set_reader_owned(sem);
if (tmp & RWSEM_FLAG_WAITERS)
rwsem_downgrade_wake(sem);
}
locking/rwsem: Support optimistic spinning We have reached the point where our mutexes are quite fine tuned for a number of situations. This includes the use of heuristics and optimistic spinning, based on MCS locking techniques. Exclusive ownership of read-write semaphores are, conceptually, just about the same as mutexes, making them close cousins. To this end we need to make them both perform similarly, and right now, rwsems are simply not up to it. This was discovered by both reverting commit 4fc3f1d6 (mm/rmap, migration: Make rmap_walk_anon() and try_to_unmap_anon() more scalable) and similarly, converting some other mutexes (ie: i_mmap_mutex) to rwsems. This creates a situation where users have to choose between a rwsem and mutex taking into account this important performance difference. Specifically, biggest difference between both locks is when we fail to acquire a mutex in the fastpath, optimistic spinning comes in to play and we can avoid a large amount of unnecessary sleeping and overhead of moving tasks in and out of wait queue. Rwsems do not have such logic. This patch, based on the work from Tim Chen and I, adds support for write-side optimistic spinning when the lock is contended. It also includes support for the recently added cancelable MCS locking for adaptive spinning. Note that is is only applicable to the xadd method, and the spinlock rwsem variant remains intact. Allowing optimistic spinning before putting the writer on the wait queue reduces wait queue contention and provided greater chance for the rwsem to get acquired. With these changes, rwsem is on par with mutex. The performance benefits can be seen on a number of workloads. For instance, on a 8 socket, 80 core 64bit Westmere box, aim7 shows the following improvements in throughput: +--------------+---------------------+-----------------+ | Workload | throughput-increase | number of users | +--------------+---------------------+-----------------+ | alltests | 20% | >1000 | | custom | 27%, 60% | 10-100, >1000 | | high_systime | 36%, 30% | >100, >1000 | | shared | 58%, 29% | 10-100, >1000 | +--------------+---------------------+-----------------+ There was also improvement on smaller systems, such as a quad-core x86-64 laptop running a 30Gb PostgreSQL (pgbench) workload for up to +60% in throughput for over 50 clients. Additionally, benefits were also noticed in exim (mail server) workloads. Furthermore, no performance regression have been seen at all. Based-on-work-from: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Davidlohr Bueso <davidlohr@hp.com> [peterz: rej fixup due to comment patches, sched/rt.h header] Signed-off-by: Peter Zijlstra <peterz@infradead.org> Cc: Alex Shi <alex.shi@linaro.org> Cc: Andi Kleen <andi@firstfloor.org> Cc: Michel Lespinasse <walken@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Peter Hurley <peter@hurleysoftware.com> Cc: "Paul E.McKenney" <paulmck@linux.vnet.ibm.com> Cc: Jason Low <jason.low2@hp.com> Cc: Aswin Chandramouleeswaran <aswin@hp.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: "Scott J Norton" <scott.norton@hp.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Chris Mason <clm@fb.com> Cc: Josef Bacik <jbacik@fusionio.com> Link: http://lkml.kernel.org/r/1399055055.6275.15.camel@buesod1.americas.hpqcorp.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-05-02 22:24:15 +04:00
locking/rwsem: Add rtmutex based R/W semaphore implementation The RT specific R/W semaphore implementation used to restrict the number of readers to one, because a writer cannot block on multiple readers and inherit its priority or budget. The single reader restricting was painful in various ways: - Performance bottleneck for multi-threaded applications in the page fault path (mmap sem) - Progress blocker for drivers which are carefully crafted to avoid the potential reader/writer deadlock in mainline. The analysis of the writer code paths shows that properly written RT tasks should not take them. Syscalls like mmap(), file access which take mmap sem write locked have unbound latencies, which are completely unrelated to mmap sem. Other R/W sem users like graphics drivers are not suitable for RT tasks either. So there is little risk to hurt RT tasks when the RT rwsem implementation is done in the following way: - Allow concurrent readers - Make writers block until the last reader left the critical section. This blocking is not subject to priority/budget inheritance. - Readers blocked on a writer inherit their priority/budget in the normal way. There is a drawback with this scheme: R/W semaphores become writer unfair though the applications which have triggered writer starvation (mostly on mmap_sem) in the past are not really the typical workloads running on a RT system. So while it's unlikely to hit writer starvation, it's possible. If there are unexpected workloads on RT systems triggering it, the problem has to be revisited. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Link: https://lore.kernel.org/r/20210815211303.016885947@linutronix.de
2021-08-16 00:28:05 +03:00
#else /* !CONFIG_PREEMPT_RT */
#define RT_MUTEX_BUILD_MUTEX
locking/rwsem: Add rtmutex based R/W semaphore implementation The RT specific R/W semaphore implementation used to restrict the number of readers to one, because a writer cannot block on multiple readers and inherit its priority or budget. The single reader restricting was painful in various ways: - Performance bottleneck for multi-threaded applications in the page fault path (mmap sem) - Progress blocker for drivers which are carefully crafted to avoid the potential reader/writer deadlock in mainline. The analysis of the writer code paths shows that properly written RT tasks should not take them. Syscalls like mmap(), file access which take mmap sem write locked have unbound latencies, which are completely unrelated to mmap sem. Other R/W sem users like graphics drivers are not suitable for RT tasks either. So there is little risk to hurt RT tasks when the RT rwsem implementation is done in the following way: - Allow concurrent readers - Make writers block until the last reader left the critical section. This blocking is not subject to priority/budget inheritance. - Readers blocked on a writer inherit their priority/budget in the normal way. There is a drawback with this scheme: R/W semaphores become writer unfair though the applications which have triggered writer starvation (mostly on mmap_sem) in the past are not really the typical workloads running on a RT system. So while it's unlikely to hit writer starvation, it's possible. If there are unexpected workloads on RT systems triggering it, the problem has to be revisited. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Link: https://lore.kernel.org/r/20210815211303.016885947@linutronix.de
2021-08-16 00:28:05 +03:00
#include "rtmutex.c"
#define rwbase_set_and_save_current_state(state) \
set_current_state(state)
#define rwbase_restore_current_state() \
__set_current_state(TASK_RUNNING)
#define rwbase_rtmutex_lock_state(rtm, state) \
__rt_mutex_lock(rtm, state)
#define rwbase_rtmutex_slowlock_locked(rtm, state) \
__rt_mutex_slowlock_locked(rtm, NULL, state)
locking/rwsem: Add rtmutex based R/W semaphore implementation The RT specific R/W semaphore implementation used to restrict the number of readers to one, because a writer cannot block on multiple readers and inherit its priority or budget. The single reader restricting was painful in various ways: - Performance bottleneck for multi-threaded applications in the page fault path (mmap sem) - Progress blocker for drivers which are carefully crafted to avoid the potential reader/writer deadlock in mainline. The analysis of the writer code paths shows that properly written RT tasks should not take them. Syscalls like mmap(), file access which take mmap sem write locked have unbound latencies, which are completely unrelated to mmap sem. Other R/W sem users like graphics drivers are not suitable for RT tasks either. So there is little risk to hurt RT tasks when the RT rwsem implementation is done in the following way: - Allow concurrent readers - Make writers block until the last reader left the critical section. This blocking is not subject to priority/budget inheritance. - Readers blocked on a writer inherit their priority/budget in the normal way. There is a drawback with this scheme: R/W semaphores become writer unfair though the applications which have triggered writer starvation (mostly on mmap_sem) in the past are not really the typical workloads running on a RT system. So while it's unlikely to hit writer starvation, it's possible. If there are unexpected workloads on RT systems triggering it, the problem has to be revisited. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Link: https://lore.kernel.org/r/20210815211303.016885947@linutronix.de
2021-08-16 00:28:05 +03:00
#define rwbase_rtmutex_unlock(rtm) \
__rt_mutex_unlock(rtm)
#define rwbase_rtmutex_trylock(rtm) \
__rt_mutex_trylock(rtm)
#define rwbase_signal_pending_state(state, current) \
signal_pending_state(state, current)
#define rwbase_schedule() \
schedule()
#include "rwbase_rt.c"
void __init_rwsem(struct rw_semaphore *sem, const char *name,
locking/rwsem: Add rtmutex based R/W semaphore implementation The RT specific R/W semaphore implementation used to restrict the number of readers to one, because a writer cannot block on multiple readers and inherit its priority or budget. The single reader restricting was painful in various ways: - Performance bottleneck for multi-threaded applications in the page fault path (mmap sem) - Progress blocker for drivers which are carefully crafted to avoid the potential reader/writer deadlock in mainline. The analysis of the writer code paths shows that properly written RT tasks should not take them. Syscalls like mmap(), file access which take mmap sem write locked have unbound latencies, which are completely unrelated to mmap sem. Other R/W sem users like graphics drivers are not suitable for RT tasks either. So there is little risk to hurt RT tasks when the RT rwsem implementation is done in the following way: - Allow concurrent readers - Make writers block until the last reader left the critical section. This blocking is not subject to priority/budget inheritance. - Readers blocked on a writer inherit their priority/budget in the normal way. There is a drawback with this scheme: R/W semaphores become writer unfair though the applications which have triggered writer starvation (mostly on mmap_sem) in the past are not really the typical workloads running on a RT system. So while it's unlikely to hit writer starvation, it's possible. If there are unexpected workloads on RT systems triggering it, the problem has to be revisited. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Link: https://lore.kernel.org/r/20210815211303.016885947@linutronix.de
2021-08-16 00:28:05 +03:00
struct lock_class_key *key)
{
init_rwbase_rt(&(sem)->rwbase);
#ifdef CONFIG_DEBUG_LOCK_ALLOC
locking/rwsem: Add rtmutex based R/W semaphore implementation The RT specific R/W semaphore implementation used to restrict the number of readers to one, because a writer cannot block on multiple readers and inherit its priority or budget. The single reader restricting was painful in various ways: - Performance bottleneck for multi-threaded applications in the page fault path (mmap sem) - Progress blocker for drivers which are carefully crafted to avoid the potential reader/writer deadlock in mainline. The analysis of the writer code paths shows that properly written RT tasks should not take them. Syscalls like mmap(), file access which take mmap sem write locked have unbound latencies, which are completely unrelated to mmap sem. Other R/W sem users like graphics drivers are not suitable for RT tasks either. So there is little risk to hurt RT tasks when the RT rwsem implementation is done in the following way: - Allow concurrent readers - Make writers block until the last reader left the critical section. This blocking is not subject to priority/budget inheritance. - Readers blocked on a writer inherit their priority/budget in the normal way. There is a drawback with this scheme: R/W semaphores become writer unfair though the applications which have triggered writer starvation (mostly on mmap_sem) in the past are not really the typical workloads running on a RT system. So while it's unlikely to hit writer starvation, it's possible. If there are unexpected workloads on RT systems triggering it, the problem has to be revisited. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Link: https://lore.kernel.org/r/20210815211303.016885947@linutronix.de
2021-08-16 00:28:05 +03:00
debug_check_no_locks_freed((void *)sem, sizeof(*sem));
lockdep_init_map_wait(&sem->dep_map, name, key, 0, LD_WAIT_SLEEP);
#endif
}
EXPORT_SYMBOL(__init_rwsem);
locking/rwsem: Add rtmutex based R/W semaphore implementation The RT specific R/W semaphore implementation used to restrict the number of readers to one, because a writer cannot block on multiple readers and inherit its priority or budget. The single reader restricting was painful in various ways: - Performance bottleneck for multi-threaded applications in the page fault path (mmap sem) - Progress blocker for drivers which are carefully crafted to avoid the potential reader/writer deadlock in mainline. The analysis of the writer code paths shows that properly written RT tasks should not take them. Syscalls like mmap(), file access which take mmap sem write locked have unbound latencies, which are completely unrelated to mmap sem. Other R/W sem users like graphics drivers are not suitable for RT tasks either. So there is little risk to hurt RT tasks when the RT rwsem implementation is done in the following way: - Allow concurrent readers - Make writers block until the last reader left the critical section. This blocking is not subject to priority/budget inheritance. - Readers blocked on a writer inherit their priority/budget in the normal way. There is a drawback with this scheme: R/W semaphores become writer unfair though the applications which have triggered writer starvation (mostly on mmap_sem) in the past are not really the typical workloads running on a RT system. So while it's unlikely to hit writer starvation, it's possible. If there are unexpected workloads on RT systems triggering it, the problem has to be revisited. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Link: https://lore.kernel.org/r/20210815211303.016885947@linutronix.de
2021-08-16 00:28:05 +03:00
static inline void __down_read(struct rw_semaphore *sem)
{
rwbase_read_lock(&sem->rwbase, TASK_UNINTERRUPTIBLE);
}
static inline int __down_read_interruptible(struct rw_semaphore *sem)
{
return rwbase_read_lock(&sem->rwbase, TASK_INTERRUPTIBLE);
}
static inline int __down_read_killable(struct rw_semaphore *sem)
{
return rwbase_read_lock(&sem->rwbase, TASK_KILLABLE);
}
static inline int __down_read_trylock(struct rw_semaphore *sem)
{
return rwbase_read_trylock(&sem->rwbase);
}
static inline void __up_read(struct rw_semaphore *sem)
{
rwbase_read_unlock(&sem->rwbase, TASK_NORMAL);
}
static inline void __sched __down_write(struct rw_semaphore *sem)
{
rwbase_write_lock(&sem->rwbase, TASK_UNINTERRUPTIBLE);
}
static inline int __sched __down_write_killable(struct rw_semaphore *sem)
{
return rwbase_write_lock(&sem->rwbase, TASK_KILLABLE);
}
static inline int __down_write_trylock(struct rw_semaphore *sem)
{
return rwbase_write_trylock(&sem->rwbase);
}
static inline void __up_write(struct rw_semaphore *sem)
{
rwbase_write_unlock(&sem->rwbase);
}
static inline void __downgrade_write(struct rw_semaphore *sem)
{
rwbase_write_downgrade(&sem->rwbase);
}
/* Debug stubs for the common API */
#define DEBUG_RWSEMS_WARN_ON(c, sem)
static inline void __rwsem_set_reader_owned(struct rw_semaphore *sem,
struct task_struct *owner)
{
}
static inline bool is_rwsem_reader_owned(struct rw_semaphore *sem)
{
int count = atomic_read(&sem->rwbase.readers);
return count < 0 && count != READER_BIAS;
}
#endif /* CONFIG_PREEMPT_RT */
/*
* lock for reading
*/
void __sched down_read(struct rw_semaphore *sem)
{
might_sleep();
rwsem_acquire_read(&sem->dep_map, 0, 0, _RET_IP_);
LOCK_CONTENDED(sem, __down_read_trylock, __down_read);
}
EXPORT_SYMBOL(down_read);
int __sched down_read_interruptible(struct rw_semaphore *sem)
{
might_sleep();
rwsem_acquire_read(&sem->dep_map, 0, 0, _RET_IP_);
if (LOCK_CONTENDED_RETURN(sem, __down_read_trylock, __down_read_interruptible)) {
rwsem_release(&sem->dep_map, _RET_IP_);
return -EINTR;
}
return 0;
}
EXPORT_SYMBOL(down_read_interruptible);
int __sched down_read_killable(struct rw_semaphore *sem)
{
might_sleep();
rwsem_acquire_read(&sem->dep_map, 0, 0, _RET_IP_);
if (LOCK_CONTENDED_RETURN(sem, __down_read_trylock, __down_read_killable)) {
2019-09-19 19:09:40 +03:00
rwsem_release(&sem->dep_map, _RET_IP_);
return -EINTR;
}
return 0;
}
EXPORT_SYMBOL(down_read_killable);
/*
* trylock for reading -- returns 1 if successful, 0 if contention
*/
int down_read_trylock(struct rw_semaphore *sem)
{
int ret = __down_read_trylock(sem);
if (ret == 1)
rwsem_acquire_read(&sem->dep_map, 0, 1, _RET_IP_);
return ret;
}
EXPORT_SYMBOL(down_read_trylock);
/*
* lock for writing
*/
void __sched down_write(struct rw_semaphore *sem)
{
might_sleep();
rwsem_acquire(&sem->dep_map, 0, 0, _RET_IP_);
LOCK_CONTENDED(sem, __down_write_trylock, __down_write);
}
EXPORT_SYMBOL(down_write);
/*
* lock for writing
*/
int __sched down_write_killable(struct rw_semaphore *sem)
{
might_sleep();
rwsem_acquire(&sem->dep_map, 0, 0, _RET_IP_);
locking/rwsem: Code cleanup after files merging After merging all the relevant rwsem code into one single file, there are a number of optimizations and cleanups that can be done: 1) Remove all the EXPORT_SYMBOL() calls for functions that are not accessed elsewhere. 2) Remove all the __visible tags as none of the functions will be called from assembly code anymore. 3) Make all the internal functions static. 4) Remove some unneeded blank lines. 5) Remove the intermediate rwsem_down_{read|write}_failed*() functions and rename __rwsem_down_{read|write}_failed_common() to rwsem_down_{read|write}_slowpath(). 6) Remove "__" prefix of __rwsem_mark_wake(). 7) Use atomic_long_try_cmpxchg_acquire() as much as possible. 8) Remove the rwsem_rtrylock and rwsem_wtrylock lock events as they are not that useful. That enables the compiler to do better optimization and reduce code size. The text+data size of rwsem.o on an x86-64 machine with gcc8 was reduced from 10237 bytes to 5030 bytes with this change. Suggested-by: Peter Zijlstra <peterz@infradead.org> Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-6-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:04 +03:00
if (LOCK_CONTENDED_RETURN(sem, __down_write_trylock,
__down_write_killable)) {
2019-09-19 19:09:40 +03:00
rwsem_release(&sem->dep_map, _RET_IP_);
return -EINTR;
}
return 0;
}
EXPORT_SYMBOL(down_write_killable);
/*
* trylock for writing -- returns 1 if successful, 0 if contention
*/
int down_write_trylock(struct rw_semaphore *sem)
{
int ret = __down_write_trylock(sem);
if (ret == 1)
rwsem_acquire(&sem->dep_map, 0, 1, _RET_IP_);
locking/rwsem: Support optimistic spinning We have reached the point where our mutexes are quite fine tuned for a number of situations. This includes the use of heuristics and optimistic spinning, based on MCS locking techniques. Exclusive ownership of read-write semaphores are, conceptually, just about the same as mutexes, making them close cousins. To this end we need to make them both perform similarly, and right now, rwsems are simply not up to it. This was discovered by both reverting commit 4fc3f1d6 (mm/rmap, migration: Make rmap_walk_anon() and try_to_unmap_anon() more scalable) and similarly, converting some other mutexes (ie: i_mmap_mutex) to rwsems. This creates a situation where users have to choose between a rwsem and mutex taking into account this important performance difference. Specifically, biggest difference between both locks is when we fail to acquire a mutex in the fastpath, optimistic spinning comes in to play and we can avoid a large amount of unnecessary sleeping and overhead of moving tasks in and out of wait queue. Rwsems do not have such logic. This patch, based on the work from Tim Chen and I, adds support for write-side optimistic spinning when the lock is contended. It also includes support for the recently added cancelable MCS locking for adaptive spinning. Note that is is only applicable to the xadd method, and the spinlock rwsem variant remains intact. Allowing optimistic spinning before putting the writer on the wait queue reduces wait queue contention and provided greater chance for the rwsem to get acquired. With these changes, rwsem is on par with mutex. The performance benefits can be seen on a number of workloads. For instance, on a 8 socket, 80 core 64bit Westmere box, aim7 shows the following improvements in throughput: +--------------+---------------------+-----------------+ | Workload | throughput-increase | number of users | +--------------+---------------------+-----------------+ | alltests | 20% | >1000 | | custom | 27%, 60% | 10-100, >1000 | | high_systime | 36%, 30% | >100, >1000 | | shared | 58%, 29% | 10-100, >1000 | +--------------+---------------------+-----------------+ There was also improvement on smaller systems, such as a quad-core x86-64 laptop running a 30Gb PostgreSQL (pgbench) workload for up to +60% in throughput for over 50 clients. Additionally, benefits were also noticed in exim (mail server) workloads. Furthermore, no performance regression have been seen at all. Based-on-work-from: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Davidlohr Bueso <davidlohr@hp.com> [peterz: rej fixup due to comment patches, sched/rt.h header] Signed-off-by: Peter Zijlstra <peterz@infradead.org> Cc: Alex Shi <alex.shi@linaro.org> Cc: Andi Kleen <andi@firstfloor.org> Cc: Michel Lespinasse <walken@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Peter Hurley <peter@hurleysoftware.com> Cc: "Paul E.McKenney" <paulmck@linux.vnet.ibm.com> Cc: Jason Low <jason.low2@hp.com> Cc: Aswin Chandramouleeswaran <aswin@hp.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: "Scott J Norton" <scott.norton@hp.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Chris Mason <clm@fb.com> Cc: Josef Bacik <jbacik@fusionio.com> Link: http://lkml.kernel.org/r/1399055055.6275.15.camel@buesod1.americas.hpqcorp.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-05-02 22:24:15 +04:00
return ret;
}
EXPORT_SYMBOL(down_write_trylock);
/*
* release a read lock
*/
void up_read(struct rw_semaphore *sem)
{
2019-09-19 19:09:40 +03:00
rwsem_release(&sem->dep_map, _RET_IP_);
__up_read(sem);
}
EXPORT_SYMBOL(up_read);
/*
* release a write lock
*/
void up_write(struct rw_semaphore *sem)
{
2019-09-19 19:09:40 +03:00
rwsem_release(&sem->dep_map, _RET_IP_);
__up_write(sem);
}
EXPORT_SYMBOL(up_write);
/*
* downgrade write lock to read lock
*/
void downgrade_write(struct rw_semaphore *sem)
{
lock_downgrade(&sem->dep_map, _RET_IP_);
__downgrade_write(sem);
}
EXPORT_SYMBOL(downgrade_write);
#ifdef CONFIG_DEBUG_LOCK_ALLOC
void down_read_nested(struct rw_semaphore *sem, int subclass)
{
might_sleep();
rwsem_acquire_read(&sem->dep_map, subclass, 0, _RET_IP_);
LOCK_CONTENDED(sem, __down_read_trylock, __down_read);
}
EXPORT_SYMBOL(down_read_nested);
int down_read_killable_nested(struct rw_semaphore *sem, int subclass)
{
might_sleep();
rwsem_acquire_read(&sem->dep_map, subclass, 0, _RET_IP_);
if (LOCK_CONTENDED_RETURN(sem, __down_read_trylock, __down_read_killable)) {
rwsem_release(&sem->dep_map, _RET_IP_);
return -EINTR;
}
return 0;
}
EXPORT_SYMBOL(down_read_killable_nested);
void _down_write_nest_lock(struct rw_semaphore *sem, struct lockdep_map *nest)
{
might_sleep();
rwsem_acquire_nest(&sem->dep_map, 0, 0, nest, _RET_IP_);
LOCK_CONTENDED(sem, __down_write_trylock, __down_write);
}
EXPORT_SYMBOL(_down_write_nest_lock);
void down_read_non_owner(struct rw_semaphore *sem)
{
might_sleep();
__down_read(sem);
locking/rwsem: Make owner store task pointer of last owning reader Currently, when a reader acquires a lock, it only sets the RWSEM_READER_OWNED bit in the owner field. The other bits are simply not used. When debugging hanging cases involving rwsems and readers, the owner value does not provide much useful information at all. This patch modifies the current behavior to always store the task_struct pointer of the last rwsem-acquiring reader in a reader-owned rwsem. This may be useful in debugging rwsem hanging cases especially if only one reader is involved. However, the task in the owner field may not the real owner or one of the real owners at all when the owner value is examined, for example, in a crash dump. So it is just an additional hint about the past history. If CONFIG_DEBUG_RWSEMS=y is enabled, the owner field will be checked at unlock time too to make sure the task pointer value is valid. That does have a slight performance cost and so is only enabled as part of that debug option. From the performance point of view, it is expected that the changes shouldn't have any noticeable performance impact. A rwsem microbenchmark (with 48 worker threads and 1:1 reader/writer ratio) was ran on a 2-socket 24-core 48-thread Haswell system. The locking rates on a 4.19-rc1 based kernel were as follows: 1) Unpatched kernel: 543.3 kops/s 2) Patched kernel: 549.2 kops/s 3) Patched kernel (CONFIG_DEBUG_RWSEMS on): 546.6 kops/s There was actually a slight increase in performance (1.1%) in this particular case. Maybe it was caused by the elimination of a branch or just a testing noise. Turning on the CONFIG_DEBUG_RWSEMS option also had less than the expected impact on performance. The least significant 2 bits of the owner value are now used to designate the rwsem is readers owned and the owners are anonymous. Signed-off-by: Waiman Long <longman@redhat.com> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Will Deacon <will.deacon@arm.com> Link: http://lkml.kernel.org/r/1536265114-10842-1-git-send-email-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-09-06 23:18:34 +03:00
__rwsem_set_reader_owned(sem, NULL);
}
EXPORT_SYMBOL(down_read_non_owner);
void down_write_nested(struct rw_semaphore *sem, int subclass)
{
might_sleep();
rwsem_acquire(&sem->dep_map, subclass, 0, _RET_IP_);
LOCK_CONTENDED(sem, __down_write_trylock, __down_write);
}
EXPORT_SYMBOL(down_write_nested);
int __sched down_write_killable_nested(struct rw_semaphore *sem, int subclass)
{
might_sleep();
rwsem_acquire(&sem->dep_map, subclass, 0, _RET_IP_);
locking/rwsem: Code cleanup after files merging After merging all the relevant rwsem code into one single file, there are a number of optimizations and cleanups that can be done: 1) Remove all the EXPORT_SYMBOL() calls for functions that are not accessed elsewhere. 2) Remove all the __visible tags as none of the functions will be called from assembly code anymore. 3) Make all the internal functions static. 4) Remove some unneeded blank lines. 5) Remove the intermediate rwsem_down_{read|write}_failed*() functions and rename __rwsem_down_{read|write}_failed_common() to rwsem_down_{read|write}_slowpath(). 6) Remove "__" prefix of __rwsem_mark_wake(). 7) Use atomic_long_try_cmpxchg_acquire() as much as possible. 8) Remove the rwsem_rtrylock and rwsem_wtrylock lock events as they are not that useful. That enables the compiler to do better optimization and reduce code size. The text+data size of rwsem.o on an x86-64 machine with gcc8 was reduced from 10237 bytes to 5030 bytes with this change. Suggested-by: Peter Zijlstra <peterz@infradead.org> Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Will Deacon <will.deacon@arm.com> Cc: huang ying <huang.ying.caritas@gmail.com> Link: https://lkml.kernel.org/r/20190520205918.22251-6-longman@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-05-20 23:59:04 +03:00
if (LOCK_CONTENDED_RETURN(sem, __down_write_trylock,
__down_write_killable)) {
2019-09-19 19:09:40 +03:00
rwsem_release(&sem->dep_map, _RET_IP_);
return -EINTR;
}
return 0;
}
EXPORT_SYMBOL(down_write_killable_nested);
void up_read_non_owner(struct rw_semaphore *sem)
{
DEBUG_RWSEMS_WARN_ON(!is_rwsem_reader_owned(sem), sem);
__up_read(sem);
}
EXPORT_SYMBOL(up_read_non_owner);
#endif