Date: Fri, 6 Nov 2015 15:24:20 +0100 From: Hans Petter Selasky <hps@selasky.org> To: Mateusz Guzik <mjguzik@gmail.com>, Ian Lepore <ian@freebsd.org>, John Baldwin <jhb@freebsd.org>, Adrian Chadd <adrian.chadd@gmail.com>, freebsd-current <freebsd-current@freebsd.org>, Konstantin Belousov <kostikbel@gmail.com> Subject: Re: [PATCH] microoptimize by trying to avoid locking a locked mutex Message-ID: <563CB814.80905@selasky.org> In-Reply-To: <20151106005524.GA4877@dft-labs.eu> References: <20151104233218.GA27709@dft-labs.eu> <20151105192623.GB27709@dft-labs.eu> <CAJ-VmonnH4JJg0XqX1SoBXBa%2B9Xfmk%2BHFv58ETaQ9v1-uAAhdQ@mail.gmail.com> <1563180.x0Z3Ou4xid@ralph.baldwin.cx> <1446766522.91534.412.camel@freebsd.org> <20151106005524.GA4877@dft-labs.eu>
next in thread | previous in thread | raw e-mail | index | archive | help
On 11/06/15 01:55, Mateusz Guzik wrote:
> On Thu, Nov 05, 2015 at 04:35:22PM -0700, Ian Lepore wrote:
>> On Thu, 2015-11-05 at 14:19 -0800, John Baldwin wrote:
>>> On Thursday, November 05, 2015 01:45:19 PM Adrian Chadd wrote:
>>>> On 5 November 2015 at 11:26, Mateusz Guzik <mjguzik@gmail.com>
>>>> wrote:
>>>>> On Thu, Nov 05, 2015 at 11:04:13AM -0800, John Baldwin wrote:
>>>>>> On Thursday, November 05, 2015 04:26:28 PM Konstantin Belousov
>>>>>> wrote:
>>>>>>> On Thu, Nov 05, 2015 at 12:32:18AM +0100, Mateusz Guzik
>>>>>>> wrote:
>>>>>>>> mtx_lock will unconditionally try to grab the lock and if
>>>>>>>> that fails,
>>>>>>>> will call __mtx_lock_sleep which will immediately try to do
>>>>>>>> the same
>>>>>>>> atomic op again.
>>>>>>>>
>>>>>>>> So, the obvious microoptimization is to check the state in
>>>>>>>> __mtx_lock_sleep and avoid the operation if the lock is not
>>>>>>>> free.
>>>>>>>>
>>>>>>>> This gives me ~40% speedup in a microbenchmark of 40 find
>>>>>>>> processes
>>>>>>>> traversing tmpfs and contending on mount mtx (only used as
>>>>>>>> an easy
>>>>>>>> benchmark, I have WIP patches to get rid of it).
>>>>>>>>
>>>>>>>> Second part of the patch is optional and just checks the
>>>>>>>> state of the
>>>>>>>> lock prior to doing any atomic operations, but it gives a
>>>>>>>> very modest
>>>>>>>> speed up when applied on top of the __mtx_lock_sleep
>>>>>>>> change. As such,
>>>>>>>> I'm not going to defend this part.
>>>>>>> Shouldn't the same consideration applied to all spinning
>>>>>>> loops, i.e.
>>>>>>> also to the spin/thread mutexes, and to the spinning parts of
>>>>>>> sx and
>>>>>>> lockmgr ?
>>>>>>
>>>>>> I agree. I think both changes are good and worth doing in our
>>>>>> other
>>>>>> primitives.
>>>>>>
>>>>>
>>>>> I glanced over e.g. rw_rlock and it did not have the issue, now
>>>>> that I
>>>>> see _sx_xlock_hard it wuld indeed use fixing.
>>>>>
>>>>> Expect a patch in few h for all primitives I'll find. I'll stress
>>>>> test
>>>>> the kernel, but it is unlikely I'll do microbenchmarks for
>>>>> remaining
>>>>> primitives.
>>>>
>>>> Is this stuff you're proposing still valid for non-x86 platforms?
>>>
>>> Yes. It just does a read before trying the atomic compare and swap
>>> and
>>> falls through to the hard case as if the atomic op failed if the
>>> result
>>> of the read would result in a compare failure.
>>>
>>
>> The atomic ops include barriers, the new pre-read of the variable
>> doesn't. Will that cause problems, especially for code inside a loop
>> where the compiler may decide to shuffle things around?
>>
>
> Lock value is volatile, so the compiler should not screw us up. I removed
> the store to the variable due to a different concern. In worst case we
> would have slept instead of spinning. (see below)
>
>> I suspect the performance gain will be biggest on the platforms where
>> atomic ops are expensive (I gather from various code comments that's
>> the case on x86).
>>
>
> I don't know about other architectures, on x86 atomic op performance on
> contended cachelines deteriorates drastically.
>
> =======================================================
>
> For the most port the patch below does 2 simple kinds of changes:
> 1. in macros for lock/unlock primitives the lock value is fetched and
> compared against relevant _UNLOCKED macro prior to the atomic op
> 2. in functions:
>
> before:
> while (!_mtx_obtain_lock(m, tid)) {
> v = m->mtx_lock;
> if (v != MTX_UNOWNED) {
> ....
> }
> .....
> }
>
> after:
> for (;;) {
> if (m->mtx_lock == MTX_UNOWNED && _mtx_obtain_lock(m, tid))
> break;
> v = m->mtx_lock;
> if (v != MTX_UNOWNED) {
> ....
> }
> .....
> }
>
> The original patch preloaded the 'v' variable. If the value was
> MTX_UNOWNED and _mtx_obtain_lock failed, we just lost the race. In
> order to spin waiting for the other thread to release the lock, we have
> to re-read the variable. Thus for simplicity it is no longer stored in
> 'v'. Note this is contrary to kib's earlier suggestion which would put
> such threads directly to sleep. I don't have proper measurements to have
> any strong opinion here, I went the aforementioned route to minimise
> changes.
>
> Note this is a trivial patch to improve the situation a little bit, I
> have no interest in trying to polish these primitives at this time.
>
> For overall results, on a machine with 32GB of RAM and the following:
> CPU: Intel(R) Xeon(R) CPU E5-2680 v2 @ 2.80GHz (2800.06-MHz K8-class
> CPU)
> FreeBSD/SMP: 2 package(s) x 10 core(s) x 2 SMT threads
>
> this reduced make -j 40 kernel in a 10.2 jail from ~2:15 to ~2:05.
>
> =======================================================
>
> diff --git a/sys/kern/kern_lock.c b/sys/kern/kern_lock.c
> index aa67180..198e93e 100644
> --- a/sys/kern/kern_lock.c
> +++ b/sys/kern/kern_lock.c
> @@ -787,8 +787,10 @@ __lockmgr_args(struct lock *lk, u_int flags, struct lock_object *ilk,
> break;
> }
>
> - while (!atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED,
> - tid)) {
> + for (;;) {
> + if (lk->lk_lock == LK_UNLOCKED &&
> + atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid))
> + break;
> #ifdef HWPMC_HOOKS
> PMC_SOFT_CALL( , , lock, failed);
> #endif
> @@ -1124,7 +1126,11 @@ __lockmgr_args(struct lock *lk, u_int flags, struct lock_object *ilk,
> __func__, iwmesg, file, line);
> }
>
> - while (!atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid)) {
> + for (;;) {
> + if (lk->lk_lock == LK_UNLOCKED &&
> + atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid))
> + break;
> +
> #ifdef HWPMC_HOOKS
> PMC_SOFT_CALL( , , lock, failed);
> #endif
> diff --git a/sys/kern/kern_mutex.c b/sys/kern/kern_mutex.c
> index bec8f6b..51e82a3 100644
> --- a/sys/kern/kern_mutex.c
> +++ b/sys/kern/kern_mutex.c
> @@ -419,7 +419,9 @@ __mtx_lock_sleep(volatile uintptr_t *c, uintptr_t tid, int opts,
> all_time -= lockstat_nsecs(&m->lock_object);
> #endif
>
> - while (!_mtx_obtain_lock(m, tid)) {
> + for (;;) {
> + if (m->mtx_lock == MTX_UNOWNED && _mtx_obtain_lock(m, tid))
> + break;
> #ifdef KDTRACE_HOOKS
> spin_cnt++;
> #endif
> @@ -602,8 +604,9 @@ _mtx_lock_spin_cookie(volatile uintptr_t *c, uintptr_t tid, int opts,
> #ifdef KDTRACE_HOOKS
> spin_time -= lockstat_nsecs(&m->lock_object);
> #endif
> - while (!_mtx_obtain_lock(m, tid)) {
> -
> + for (;;) {
> + if (m->mtx_lock == MTX_UNOWNED && _mtx_obtain_lock(m, tid))
> + break;
> /* Give interrupts a chance while we spin. */
> spinlock_exit();
> while (m->mtx_lock != MTX_UNOWNED) {
> @@ -675,7 +678,9 @@ retry:
> m->lock_object.lo_name, file, line));
> WITNESS_CHECKORDER(&m->lock_object,
> opts | LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL);
> - while (!_mtx_obtain_lock(m, tid)) {
> + for (;;) {
> + if (m->mtx_lock == MTX_UNOWNED && _mtx_obtain_lock(m, tid))
> + break;
> if (m->mtx_lock == tid) {
> m->mtx_recurse++;
> break;
> diff --git a/sys/kern/kern_rwlock.c b/sys/kern/kern_rwlock.c
> index 6541724..6a904d2 100644
> --- a/sys/kern/kern_rwlock.c
> +++ b/sys/kern/kern_rwlock.c
> @@ -771,7 +771,9 @@ __rw_wlock_hard(volatile uintptr_t *c, uintptr_t tid, const char *file,
> all_time -= lockstat_nsecs(&rw->lock_object);
> state = rw->rw_lock;
> #endif
> - while (!_rw_write_lock(rw, tid)) {
> + for (;;) {
> + if (rw->rw_lock == RW_UNLOCKED && _rw_write_lock(rw, tid))
> + break;
> #ifdef KDTRACE_HOOKS
> spin_cnt++;
> #endif
> diff --git a/sys/kern/kern_sx.c b/sys/kern/kern_sx.c
> index 96e117b..2a81c04 100644
> --- a/sys/kern/kern_sx.c
> +++ b/sys/kern/kern_sx.c
> @@ -544,7 +544,10 @@ _sx_xlock_hard(struct sx *sx, uintptr_t tid, int opts, const char *file,
> all_time -= lockstat_nsecs(&sx->lock_object);
> state = sx->sx_lock;
> #endif
> - while (!atomic_cmpset_acq_ptr(&sx->sx_lock, SX_LOCK_UNLOCKED, tid)) {
> + for (;;) {
> + if (sx->sx_lock == SX_LOCK_UNLOCKED &&
> + atomic_cmpset_acq_ptr(&sx->sx_lock, SX_LOCK_UNLOCKED, tid))
> + break;
> #ifdef KDTRACE_HOOKS
> spin_cnt++;
> #endif
> diff --git a/sys/sys/mutex.h b/sys/sys/mutex.h
> index a9ec072..0443922 100644
> --- a/sys/sys/mutex.h
> +++ b/sys/sys/mutex.h
> @@ -185,7 +185,7 @@ void thread_lock_flags_(struct thread *, int, const char *, int);
> #define __mtx_lock(mp, tid, opts, file, line) do { \
> uintptr_t _tid = (uintptr_t)(tid); \
> \
> - if (!_mtx_obtain_lock((mp), _tid)) \
> + if (((mp)->mtx_lock != MTX_UNOWNED || !_mtx_obtain_lock((mp), _tid)))\
> _mtx_lock_sleep((mp), _tid, (opts), (file), (line)); \
> else \
> LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(adaptive__acquire, \
> @@ -203,7 +203,7 @@ void thread_lock_flags_(struct thread *, int, const char *, int);
> uintptr_t _tid = (uintptr_t)(tid); \
> \
> spinlock_enter(); \
> - if (!_mtx_obtain_lock((mp), _tid)) { \
> + if (((mp)->mtx_lock != MTX_UNOWNED || !_mtx_obtain_lock((mp), _tid))) {\
> if ((mp)->mtx_lock == _tid) \
> (mp)->mtx_recurse++; \
> else \
> @@ -232,7 +232,7 @@ void thread_lock_flags_(struct thread *, int, const char *, int);
> \
> if ((mp)->mtx_recurse == 0) \
> LOCKSTAT_PROFILE_RELEASE_LOCK(adaptive__release, mp); \
> - if (!_mtx_release_lock((mp), _tid)) \
> + if ((mp)->mtx_lock != _tid || !_mtx_release_lock((mp), _tid)) \
> _mtx_unlock_sleep((mp), (opts), (file), (line)); \
> } while (0)
>
> diff --git a/sys/sys/rwlock.h b/sys/sys/rwlock.h
> index f8947c5..6b4f505 100644
> --- a/sys/sys/rwlock.h
> +++ b/sys/sys/rwlock.h
> @@ -96,7 +96,7 @@
> #define __rw_wlock(rw, tid, file, line) do { \
> uintptr_t _tid = (uintptr_t)(tid); \
> \
> - if (!_rw_write_lock((rw), _tid)) \
> + if ((rw)->rw_lock != RW_UNLOCKED || !_rw_write_lock((rw), _tid))\
> _rw_wlock_hard((rw), _tid, (file), (line)); \
> else \
> LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, \
> @@ -112,7 +112,7 @@
> else { \
> LOCKSTAT_PROFILE_RELEASE_RWLOCK(rw__release, rw, \
> LOCKSTAT_WRITER); \
> - if (!_rw_write_unlock((rw), _tid)) \
> + if ((rw)->rw_lock != _tid || !_rw_write_unlock((rw), _tid))\
> _rw_wunlock_hard((rw), _tid, (file), (line)); \
> } \
> } while (0)
> diff --git a/sys/sys/sx.h b/sys/sys/sx.h
> index 96a664f..144cab5 100644
> --- a/sys/sys/sx.h
> +++ b/sys/sys/sx.h
> @@ -150,7 +150,8 @@ __sx_xlock(struct sx *sx, struct thread *td, int opts, const char *file,
> uintptr_t tid = (uintptr_t)td;
> int error = 0;
>
> - if (!atomic_cmpset_acq_ptr(&sx->sx_lock, SX_LOCK_UNLOCKED, tid))
> + if (sx->sx_lock != SX_LOCK_UNLOCKED ||
> + !atomic_cmpset_acq_ptr(&sx->sx_lock, SX_LOCK_UNLOCKED, tid))
> error = _sx_xlock_hard(sx, tid, opts, file, line);
> else
> LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
> @@ -168,7 +169,8 @@ __sx_xunlock(struct sx *sx, struct thread *td, const char *file, int line)
> if (sx->sx_recurse == 0)
> LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx,
> LOCKSTAT_WRITER);
> - if (!atomic_cmpset_rel_ptr(&sx->sx_lock, tid, SX_LOCK_UNLOCKED))
> + if (sx->sx_lock != tid ||
> + !atomic_cmpset_rel_ptr(&sx->sx_lock, tid, SX_LOCK_UNLOCKED))
> _sx_xunlock_hard(sx, tid, file, line);
> }
>
Just gave this patch a spin on my RPI2 and it seems to work fine.
--HPS
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?563CB814.80905>