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 d3681e269f ("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
If the optimistic spinning queue is empty and the rwsem does not have
the handoff or write-lock bits set, it is actually not necessary to
call rwsem_optimistic_spin() to spin on it. Instead, it can steal the
lock directly as its reader bias is in the count already. If it is
the first reader in this state, it will try to wake up other readers
in the wait queue.
With this patch applied, the following were the lock event counts
after rebooting a 2-socket system and a "make -j96" kernel rebuild.
rwsem_opt_rlock=4437
rwsem_rlock=29
rwsem_rlock_steal=19
So lock stealing represents about 0.4% of all the read locks acquired
in the slow path.
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/20201121041416.12285-4-longman@redhat.com
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>
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>
This patch enables readers to optimistically spin on a
rwsem when it is owned by a writer instead of going to sleep
directly. The rwsem_can_spin_on_owner() function is extracted
out of rwsem_optimistic_spin() and is called directly by
rwsem_down_read_slowpath() and rwsem_down_write_slowpath().
With a locking microbenchmark running on 5.1 based kernel, the total
locking rates (in kops/s) on a 8-socket IvyBrige-EX system with equal
numbers of readers and writers before and after the patch were as
follows:
# of Threads Pre-patch Post-patch
------------ --------- ----------
4 1,674 1,684
8 1,062 1,074
16 924 900
32 300 458
64 195 208
128 164 168
240 149 143
The performance change wasn't significant in this case, but this change
is required by a follow-on patch.
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/20190520205918.22251-13-longman@redhat.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
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>
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>
Add lock event counting calls so that we can track the number of lock
events happening in the rwsem code.
With CONFIG_LOCK_EVENT_COUNTS on and booting a 4-socket 112-thread x86-64
system, the rwsem counts after system bootup were as follows:
rwsem_opt_fail=261
rwsem_opt_wlock=50636
rwsem_rlock=445
rwsem_rlock_fail=0
rwsem_rlock_fast=22
rwsem_rtrylock=810144
rwsem_sleep_reader=441
rwsem_sleep_writer=310
rwsem_wake_reader=355
rwsem_wake_writer=2335
rwsem_wlock=261
rwsem_wlock_fail=0
rwsem_wtrylock=20583
It can be seen that most of the lock acquisitions in the slowpath were
write-locks in the optimistic spinning code path with no sleeping at
all. For this system, over 97% of the locks are acquired via optimistic
spinning. It illustrates the importance of optimistic spinning in
improving the performance of rwsem.
Signed-off-by: Waiman Long <longman@redhat.com>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Acked-by: Davidlohr Bueso <dbueso@suse.de>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Davidlohr Bueso <dave@stgolabs.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
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>
Link: http://lkml.kernel.org/r/20190404174320.22416-11-longman@redhat.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
The percpu event counts used by qspinlock code can be useful for
other locking code as well. So a new set of lockevent_* counting APIs
is introduced with the lock event names extracted out into the new
lock_events_list.h header file for easier addition in the future.
The existing qstat_inc() calls are replaced by either lockevent_inc() or
lockevent_cond_inc() calls.
The qstat_hop() call is renamed to lockevent_pv_hop(). The "reset_counters"
debugfs file is also renamed to ".reset_counts".
Signed-off-by: Waiman Long <longman@redhat.com>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Acked-by: Davidlohr Bueso <dbueso@suse.de>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Davidlohr Bueso <dave@stgolabs.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
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>
Link: http://lkml.kernel.org/r/20190404174320.22416-8-longman@redhat.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Waiman Long