We are going to split <linux/sched/debug.h> out of <linux/sched.h>, which
will have to be picked up from other headers and a couple of .c files.
Create a trivial placeholder <linux/sched/debug.h> file that just
maps to <linux/sched.h> to make this patch obviously correct and
bisectable.
Include the new header in the files that are going to need it.
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
We are going to split <linux/sched/wake_q.h> out of <linux/sched.h>, which
will have to be picked up from other headers and a couple of .c files.
Create a trivial placeholder <linux/sched/wake_q.h> file that just
maps to <linux/sched.h> to make this patch obviously correct and
bisectable.
Include the new header in the files that are going to need it.
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
This is a nasty interface and setting the state of a foreign task must
not be done. As of the following commit:
be628be095 ("bcache: Make gc wakeup sane, remove set_task_state()")
... everyone in the kernel calls set_task_state() with current, allowing
the helper to be removed.
However, as the comment indicates, it is still around for those archs
where computing current is more expensive than using a pointer, at least
in theory. An important arch that is affected is arm64, however this has
been addressed now [1] and performance is up to par making no difference
with either calls.
Of all the callers, if any, it's the locking bits that would care most
about this -- ie: we end up passing a tsk pointer to a lot of the lock
slowpath, and setting ->state on that. The following numbers are based
on two tests: a custom ad-hoc microbenchmark that just measures
latencies (for ~65 million calls) between get_task_state() vs
get_current_state().
Secondly for a higher overview, an unlink microbenchmark was used,
which pounds on a single file with open, close,unlink combos with
increasing thread counts (up to 4x ncpus). While the workload is quite
unrealistic, it does contend a lot on the inode mutex or now rwsem.
[1] https://lkml.kernel.org/r/1483468021-8237-1-git-send-email-mark.rutland@arm.com
== 1. x86-64 ==
Avg runtime set_task_state(): 601 msecs
Avg runtime set_current_state(): 552 msecs
vanilla dirty
Hmean unlink1-processes-2 36089.26 ( 0.00%) 38977.33 ( 8.00%)
Hmean unlink1-processes-5 28555.01 ( 0.00%) 29832.55 ( 4.28%)
Hmean unlink1-processes-8 37323.75 ( 0.00%) 44974.57 ( 20.50%)
Hmean unlink1-processes-12 43571.88 ( 0.00%) 44283.01 ( 1.63%)
Hmean unlink1-processes-21 34431.52 ( 0.00%) 38284.45 ( 11.19%)
Hmean unlink1-processes-30 34813.26 ( 0.00%) 37975.17 ( 9.08%)
Hmean unlink1-processes-48 37048.90 ( 0.00%) 39862.78 ( 7.59%)
Hmean unlink1-processes-79 35630.01 ( 0.00%) 36855.30 ( 3.44%)
Hmean unlink1-processes-110 36115.85 ( 0.00%) 39843.91 ( 10.32%)
Hmean unlink1-processes-141 32546.96 ( 0.00%) 35418.52 ( 8.82%)
Hmean unlink1-processes-172 34674.79 ( 0.00%) 36899.21 ( 6.42%)
Hmean unlink1-processes-203 37303.11 ( 0.00%) 36393.04 ( -2.44%)
Hmean unlink1-processes-224 35712.13 ( 0.00%) 36685.96 ( 2.73%)
== 2. ppc64le ==
Avg runtime set_task_state(): 938 msecs
Avg runtime set_current_state: 940 msecs
vanilla dirty
Hmean unlink1-processes-2 19269.19 ( 0.00%) 30704.50 ( 59.35%)
Hmean unlink1-processes-5 20106.15 ( 0.00%) 21804.15 ( 8.45%)
Hmean unlink1-processes-8 17496.97 ( 0.00%) 17243.28 ( -1.45%)
Hmean unlink1-processes-12 14224.15 ( 0.00%) 17240.21 ( 21.20%)
Hmean unlink1-processes-21 14155.66 ( 0.00%) 15681.23 ( 10.78%)
Hmean unlink1-processes-30 14450.70 ( 0.00%) 15995.83 ( 10.69%)
Hmean unlink1-processes-48 16945.57 ( 0.00%) 16370.42 ( -3.39%)
Hmean unlink1-processes-79 15788.39 ( 0.00%) 14639.27 ( -7.28%)
Hmean unlink1-processes-110 14268.48 ( 0.00%) 14377.40 ( 0.76%)
Hmean unlink1-processes-141 14023.65 ( 0.00%) 16271.69 ( 16.03%)
Hmean unlink1-processes-172 13417.62 ( 0.00%) 16067.55 ( 19.75%)
Hmean unlink1-processes-203 15293.08 ( 0.00%) 15440.40 ( 0.96%)
Hmean unlink1-processes-234 13719.32 ( 0.00%) 16190.74 ( 18.01%)
Hmean unlink1-processes-265 16400.97 ( 0.00%) 16115.22 ( -1.74%)
Hmean unlink1-processes-296 14388.60 ( 0.00%) 16216.13 ( 12.70%)
Hmean unlink1-processes-320 15771.85 ( 0.00%) 15905.96 ( 0.85%)
x86-64 (known to be fast for get_current()/this_cpu_read_stable() caching)
and ppc64 (with paca) show similar improvements in the unlink microbenches.
The small delta for ppc64 (2ms), does not represent the gains on the unlink
runs. In the case of x86, there was a decent amount of variation in the
latency runs, but always within a 20 to 50ms increase), ppc was more constant.
Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
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: dave@stgolabs.net
Cc: mark.rutland@arm.com
Link: http://lkml.kernel.org/r/1483479794-14013-5-git-send-email-dave@stgolabs.net
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Currently the wake_q data structure is defined by the WAKE_Q() macro.
This macro, however, looks like a function doing something as "wake" is
a verb. Even checkpatch.pl was confused as it reported warnings like
WARNING: Missing a blank line after declarations
#548: FILE: kernel/futex.c:3665:
+ int ret;
+ WAKE_Q(wake_q);
This patch renames the WAKE_Q() macro to DEFINE_WAKE_Q() which clarifies
what the macro is doing and eliminates the checkpatch.pl warnings.
Signed-off-by: Waiman Long <longman@redhat.com>
Acked-by: 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>
Link: http://lkml.kernel.org/r/1479401198-1765-1-git-send-email-longman@redhat.com
[ Resolved conflict and added missing rename. ]
Signed-off-by: Ingo Molnar <mingo@kernel.org>
When wanting to wakeup readers, __rwsem_mark_wakeup() currently
iterates the wait_list twice while looking to wakeup the first N
queued reader-tasks. While this can be quite inefficient, it was
there such that a awoken reader would be first and foremost
acknowledged by the lock counter.
Keeping the same logic, we can further benefit from the use of
wake_qs and avoid entirely the first wait_list iteration that sets
the counter as wake_up_process() isn't going to occur right away,
and therefore we maintain the counter->list order of going about
things.
Other than saving cycles with O(n) "scanning", this change also
nicely cleans up a good chunk of __rwsem_mark_wakeup(); both
visually and less tedious to read.
For example, the following improvements where seen on some will
it scale microbenchmarks, on a 48-core Haswell:
v4.7 v4.7-rwsem-v1
Hmean signal1-processes-8 5792691.42 ( 0.00%) 5771971.04 ( -0.36%)
Hmean signal1-processes-12 6081199.96 ( 0.00%) 6072174.38 ( -0.15%)
Hmean signal1-processes-21 3071137.71 ( 0.00%) 3041336.72 ( -0.97%)
Hmean signal1-processes-48 3712039.98 ( 0.00%) 3708113.59 ( -0.11%)
Hmean signal1-processes-79 4464573.45 ( 0.00%) 4682798.66 ( 4.89%)
Hmean signal1-processes-110 4486842.01 ( 0.00%) 4633781.71 ( 3.27%)
Hmean signal1-processes-141 4611816.83 ( 0.00%) 4692725.38 ( 1.75%)
Hmean signal1-processes-172 4638157.05 ( 0.00%) 4714387.86 ( 1.64%)
Hmean signal1-processes-203 4465077.80 ( 0.00%) 4690348.07 ( 5.05%)
Hmean signal1-processes-224 4410433.74 ( 0.00%) 4687534.43 ( 6.28%)
Stddev signal1-processes-8 6360.47 ( 0.00%) 8455.31 ( 32.94%)
Stddev signal1-processes-12 4004.98 ( 0.00%) 9156.13 (128.62%)
Stddev signal1-processes-21 3273.14 ( 0.00%) 5016.80 ( 53.27%)
Stddev signal1-processes-48 28420.25 ( 0.00%) 26576.22 ( -6.49%)
Stddev signal1-processes-79 22038.34 ( 0.00%) 18992.70 (-13.82%)
Stddev signal1-processes-110 23226.93 ( 0.00%) 17245.79 (-25.75%)
Stddev signal1-processes-141 6358.98 ( 0.00%) 7636.14 ( 20.08%)
Stddev signal1-processes-172 9523.70 ( 0.00%) 4824.75 (-49.34%)
Stddev signal1-processes-203 13915.33 ( 0.00%) 9326.33 (-32.98%)
Stddev signal1-processes-224 15573.94 ( 0.00%) 10613.82 (-31.85%)
Other runs that saw improvements include context_switch and pipe; and
as expected, this is particularly highlighted on larger thread counts
as it becomes more expensive to walk the list twice.
No change in wakeup ordering or semantics.
Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
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: Waiman.Long@hp.com
Cc: dave@stgolabs.net
Cc: jason.low2@hpe.com
Cc: wanpeng.li@hotmail.com
Link: http://lkml.kernel.org/r/1470384285-32163-4-git-send-email-dave@stgolabs.net
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Currently, it is not possible to determine for sure if a reader
owns a rwsem by looking at the content of the rwsem data structure.
This patch adds a new state RWSEM_READER_OWNED to the owner field
to indicate that readers currently own the lock. This enables us to
address the following 2 issues in the rwsem optimistic spinning code:
1) rwsem_can_spin_on_owner() will disallow optimistic spinning if
the owner field is NULL which can mean either the readers own
the lock or the owning writer hasn't set the owner field yet.
In the latter case, we miss the chance to do optimistic spinning.
2) While a writer is waiting in the OSQ and a reader takes the lock,
the writer will continue to spin when out of the OSQ in the main
rwsem_optimistic_spin() loop as the owner field is NULL wasting
CPU cycles if some of readers are sleeping.
Adding the new state will allow optimistic spinning to go forward as
long as the owner field is not RWSEM_READER_OWNED and the owner is
running, if set, but stop immediately when that state has been reached.
On a 4-socket Haswell machine running on a 4.6-rc1 based kernel, the
fio test with multithreaded randrw and randwrite tests on the same
file on a XFS partition on top of a NVDIMM were run, the aggregated
bandwidths before and after the patch were as follows:
Test BW before patch BW after patch % change
---- --------------- -------------- --------
randrw 988 MB/s 1192 MB/s +21%
randwrite 1513 MB/s 1623 MB/s +7.3%
The perf profile of the rwsem_down_write_failed() function in randrw
before and after the patch were:
19.95% 5.88% fio [kernel.vmlinux] [k] rwsem_down_write_failed
14.20% 1.52% fio [kernel.vmlinux] [k] rwsem_down_write_failed
The actual CPU cycles spend in rwsem_down_write_failed() dropped from
5.88% to 1.52% after the patch.
The xfstests was also run and no regression was observed.
Signed-off-by: Waiman Long <Waiman.Long@hpe.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Jason Low <jason.low2@hp.com>
Acked-by: Davidlohr Bueso <dave@stgolabs.net>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Douglas Hatch <doug.hatch@hpe.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: Peter Hurley <peter@hurleysoftware.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Scott J Norton <scott.norton@hpe.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/1463534783-38814-2-git-send-email-Waiman.Long@hpe.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Readers that are awoken will expect a nil ->task indicating
that a wakeup has occurred. Because of the way readers are
implemented, there's a small chance that the waiter will never
block in the slowpath (rwsem_down_read_failed), and therefore
requires some form of reference counting to avoid the following
scenario:
rwsem_down_read_failed() rwsem_wake()
get_task_struct();
spin_lock_irq(&wait_lock);
list_add_tail(&waiter.list)
spin_unlock_irq(&wait_lock);
raw_spin_lock_irqsave(&wait_lock)
__rwsem_do_wake()
while (1) {
set_task_state(TASK_UNINTERRUPTIBLE);
waiter->task = NULL
if (!waiter.task) // true
break;
schedule() // never reached
__set_task_state(TASK_RUNNING);
do_exit();
wake_up_process(tsk); // boom
... and therefore race with do_exit() when the caller returns.
There is also a mismatch between the smp_mb() and its documentation,
in that the serialization is done between reading the task and the
nil store. Furthermore, in addition to having the overlapping of
loads and stores to waiter->task guaranteed to be ordered within
that CPU, both wake_up_process() originally and now wake_q_add()
already imply barriers upon successful calls, which serves the
comment.
Now, as an alternative to perhaps inverting the checks in the blocker
side (which has its own penalty in that schedule is unavoidable),
with lockless wakeups this situation is naturally addressed and we
can just use the refcount held by wake_q_add(), instead doing so
explicitly. Of course, we must guarantee that the nil store is done
as the _last_ operation in that the task must already be marked for
deletion to not fall into the race above. Spurious wakeups are also
handled transparently in that the task's reference is only removed
when wake_up_q() is actually called _after_ the nil store.
Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
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: Waiman.Long@hpe.com
Cc: dave@stgolabs.net
Cc: jason.low2@hp.com
Cc: peter@hurleysoftware.com
Link: http://lkml.kernel.org/r/1463165787-25937-3-git-send-email-dave@stgolabs.net
Signed-off-by: Ingo Molnar <mingo@kernel.org>
As wake_qs gain users, we can teach rwsems about them such that
waiters can be awoken without the wait_lock. This is for both
readers and writer, the former being the most ideal candidate
as we can batch the wakeups shortening the critical region that
much more -- ie writer task blocking a bunch of tasks waiting to
service page-faults (mmap_sem readers).
In general applying wake_qs to rwsem (xadd) is not difficult as
the wait_lock is intended to be released soon _anyways_, with
the exception of when a writer slowpath will proactively wakeup
any queued readers if it sees that the lock is owned by a reader,
in which we simply do the wakeups with the lock held (see comment
in __rwsem_down_write_failed_common()).
Similar to other locking primitives, delaying the waiter being
awoken does allow, at least in theory, the lock to be stolen in
the case of writers, however no harm was seen in this (in fact
lock stealing tends to be a _good_ thing in most workloads), and
this is a tiny window anyways.
Some page-fault (pft) and mmap_sem intensive benchmarks show some
pretty constant reduction in systime (by up to ~8 and ~10%) on a
2-socket, 12 core AMD box. In addition, on an 8-core Westmere doing
page allocations (page_test)
aim9:
4.6-rc6 4.6-rc6
rwsemv2
Min page_test 378167.89 ( 0.00%) 382613.33 ( 1.18%)
Min exec_test 499.00 ( 0.00%) 502.67 ( 0.74%)
Min fork_test 3395.47 ( 0.00%) 3537.64 ( 4.19%)
Hmean page_test 395433.06 ( 0.00%) 414693.68 ( 4.87%)
Hmean exec_test 499.67 ( 0.00%) 505.30 ( 1.13%)
Hmean fork_test 3504.22 ( 0.00%) 3594.95 ( 2.59%)
Stddev page_test 17426.57 ( 0.00%) 26649.92 (-52.93%)
Stddev exec_test 0.47 ( 0.00%) 1.41 (-199.05%)
Stddev fork_test 63.74 ( 0.00%) 32.59 ( 48.86%)
Max page_test 429873.33 ( 0.00%) 456960.00 ( 6.30%)
Max exec_test 500.33 ( 0.00%) 507.66 ( 1.47%)
Max fork_test 3653.33 ( 0.00%) 3650.90 ( -0.07%)
4.6-rc6 4.6-rc6
rwsemv2
User 1.12 0.04
System 0.23 0.04
Elapsed 727.27 721.98
Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
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: Waiman.Long@hpe.com
Cc: dave@stgolabs.net
Cc: jason.low2@hp.com
Cc: peter@hurleysoftware.com
Link: http://lkml.kernel.org/r/1463165787-25937-2-git-send-email-dave@stgolabs.net
Signed-off-by: Ingo Molnar <mingo@kernel.org>
In up_write()/up_read(), rwsem_wake() will be called whenever it
detects that some writers/readers are waiting. The rwsem_wake()
function will take the wait_lock and call __rwsem_do_wake() to do the
real wakeup. For a heavily contended rwsem, doing a spin_lock() on
wait_lock will cause further contention on the heavily contended rwsem
cacheline resulting in delay in the completion of the up_read/up_write
operations.
This patch makes the wait_lock taking and the call to __rwsem_do_wake()
optional if at least one spinning writer is present. The spinning
writer will be able to take the rwsem and call rwsem_wake() later
when it calls up_write(). With the presence of a spinning writer,
rwsem_wake() will now try to acquire the lock using trylock. If that
fails, it will just quit.
Suggested-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Waiman Long <Waiman.Long@hp.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Davidlohr Bueso <dave@stgolabs.net>
Acked-by: Jason Low <jason.low2@hp.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Douglas Hatch <doug.hatch@hp.com>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Scott J Norton <scott.norton@hp.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/1430428337-16802-2-git-send-email-Waiman.Long@hp.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Ming reported soft lockups occurring when running xfstest due to
the following tip:locking/core commit:
b3fd4f03ca ("locking/rwsem: Avoid deceiving lock spinners")
When doing optimistic spinning in rwsem, threads should stop
spinning when the lock owner is not running. While a thread is
spinning on owner, if the owner reschedules, owner->on_cpu
returns false and we stop spinning.
However, this commit essentially caused the check to get
ignored because when we break out of the spin loop due to
!on_cpu, we continue spinning if sem->owner != NULL.
This patch fixes this by making sure we stop spinning if the
owner is not running. Furthermore, just like with mutexes,
refactor the code such that we don't have separate checks for
owner_running(). This makes it more straightforward in terms of
why we exit the spin on owner loop and we would also avoid
needing to "guess" why we broke out of the loop to make this
more readable.
Reported-and-tested-by: Ming Lei <ming.lei@canonical.com>
Signed-off-by: Jason Low <jason.low2@hp.com>
Acked-by: Davidlohr Bueso <dave@stgolabs.net>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Dave Jones <davej@codemonkey.org.uk>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Michel Lespinasse <walken@google.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Sasha Levin <sasha.levin@oracle.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Tim Chen <tim.c.chen@linux.intel.com>
Link: http://lkml.kernel.org/r/1425714331.2475.388.camel@j-VirtualBox
Signed-off-by: Ingo Molnar <mingo@kernel.org>
37e9562453 ("locking/rwsem: Allow conservative optimistic
spinning when readers have lock") forced the default for
optimistic spinning to be disabled if the lock owner was
nil, which makes much sense for readers. However, while
it is not our priority, we can make some optimizations
for write-mostly workloads. We can bail the spinning step
and still be conservative if there are any active tasks,
otherwise there's really no reason not to spin, as the
semaphore is most likely unlocked.
This patch recovers most of a Unixbench 'execl' benchmark
throughput by sleeping less and making better average system
usage:
before:
CPU %user %nice %system %iowait %steal %idle
all 0.60 0.00 8.02 0.00 0.00 91.38
after:
CPU %user %nice %system %iowait %steal %idle
all 1.22 0.00 70.18 0.00 0.00 28.60
Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Jason Low <jason.low2@hp.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Michel Lespinasse <walken@google.com>
Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: Tim Chen <tim.c.chen@linux.intel.com>
Link: http://lkml.kernel.org/r/1422609267-15102-6-git-send-email-dave@stgolabs.net
Signed-off-by: Ingo Molnar <mingo@kernel.org>
When readers hold the semaphore, the ->owner is nil. As such,
and unlike mutexes, '!owner' does not necessarily imply that
the lock is free. This will cause writers to potentially spin
excessively as they've been mislead to thinking they have a
chance of acquiring the lock, instead of blocking.
This patch therefore enhances the counter check when the owner
is not set by the time we've broken out of the loop. Otherwise
we can return true as a new owner has the lock and thus we want
to continue spinning. While at it, we can make rwsem_spin_on_owner()
less ambiguos and return right away under need_resched conditions.
Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Jason Low <jason.low2@hp.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Michel Lespinasse <walken@google.com>
Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: Tim Chen <tim.c.chen@linux.intel.com>
Link: http://lkml.kernel.org/r/1422609267-15102-5-git-send-email-dave@stgolabs.net
Signed-off-by: Ingo Molnar <mingo@kernel.org>
In order to optimize the spinning step, we need to set the lock
owner as soon as the lock is acquired; after a successful counter
cmpxchg operation, that is. This is particularly useful as rwsems
need to set the owner to nil for readers, so there is a greater
chance of falling out of the spinning. Currently we only set the
owner much later in the game, in the more generic level -- latency
can be specially bad when waiting for a node->next pointer when
releasing the osq in up_write calls.
As such, update the owner inside rwsem_try_write_lock (when the
lock is obtained after blocking) and rwsem_try_write_lock_unqueued
(when the lock is obtained while spinning). This requires creating
a new internal rwsem.h header to share the owner related calls.
Also cleanup some headers for mutex and rwsem.
Suggested-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Jason Low <jason.low2@hp.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Michel Lespinasse <walken@google.com>
Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: Tim Chen <tim.c.chen@linux.intel.com>
Link: http://lkml.kernel.org/r/1422609267-15102-4-git-send-email-dave@stgolabs.net
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Commit 9b0fc9c09f ("rwsem: skip initial trylock in rwsem_down_write_failed")
checks for if there are known active lockers in order to avoid write trylocking
using expensive cmpxchg() when it likely wouldn't get the lock.
However, a subsequent patch was added such that we directly
check for sem->count == RWSEM_WAITING_BIAS right before trying
that cmpxchg().
Thus, commit 9b0fc9c09f now just adds overhead.
This patch modifies it so that we only do a check for if
count == RWSEM_WAITING_BIAS.
Also, add a comment on why we do an "extra check" of count
before the cmpxchg().
Signed-off-by: Jason Low <jason.low2@hp.com>
Acked-by: Davidlohr Bueso <dbueso@suse.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Aswin Chandramouleeswaran <aswin@hp.com>
Cc: Chegu Vinod <chegu_vinod@hp.com>
Cc: Peter Hurley <peter@hurleysoftware.com>
Cc: Tim Chen <tim.c.chen@linux.intel.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Link: http://lkml.kernel.org/r/1410913017.2447.22.camel@j-VirtualBox
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Commit 4fc828e24c ("locking/rwsem: Support optimistic spinning")
introduced a major performance regression for workloads such as
xfs_repair which mix read and write locking of the mmap_sem across
many threads. The result was xfs_repair ran 5x slower on 3.16-rc2
than on 3.15 and using 20x more system CPU time.
Perf profiles indicate in some workloads that significant time can
be spent spinning on !owner. This is because we don't set the lock
owner when readers(s) obtain the rwsem.
In this patch, we'll modify rwsem_can_spin_on_owner() such that we'll
return false if there is no lock owner. The rationale is that if we
just entered the slowpath, yet there is no lock owner, then there is
a possibility that a reader has the lock. To be conservative, we'll
avoid spinning in these situations.
This patch reduced the total run time of the xfs_repair workload from
about 4 minutes 24 seconds down to approximately 1 minute 26 seconds,
back to close to the same performance as on 3.15.
Retesting of AIM7, which were some of the workloads used to test the
original optimistic spinning code, confirmed that we still get big
performance gains with optimistic spinning, even with this additional
regression fix. Davidlohr found that while the 'custom' workload took
a performance hit of ~-14% to throughput for >300 users with this
additional patch, the overall gain with optimistic spinning is
still ~+45%. The 'disk' workload even improved by ~+15% at >1000 users.
Tested-by: Dave Chinner <dchinner@redhat.com>
Acked-by: Davidlohr Bueso <davidlohr@hp.com>
Signed-off-by: Jason Low <jason.low2@hp.com>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Cc: Tim Chen <tim.c.chen@linux.intel.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Link: http://lkml.kernel.org/r/1404532172.2572.30.camel@j-VirtualBox
Signed-off-by: Ingo Molnar <mingo@kernel.org>
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>
