| lists.openwall.net | lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC | |
|
Open Source and information security mailing list archives
| ||
|
Date: Mon, 17 Jun 2013 12:05:33 -0700
From: Davidlohr Bueso <davidlohr.bueso@...com>
To: Tim Chen <tim.c.chen@...ux.intel.com>
Cc: Alex Shi <alex.shi@...el.com>, Ingo Molnar <mingo@...e.hu>,
Rik van Riel <riel@...hat.com>,
Peter Zijlstra <a.p.zijlstra@...llo.nl>,
Andrea Arcangeli <aarcange@...hat.com>,
Mel Gorman <mgorman@...e.de>, Andi Kleen <andi@...stfloor.org>,
Andrew Morton <akpm@...ux-foundation.org>,
Michel Lespinasse <walken@...gle.com>,
"Wilcox, Matthew R" <matthew.r.wilcox@...el.com>,
Dave Hansen <dave.hansen@...el.com>,
"linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
"linux-mm@...ck.org" <linux-mm@...ck.org>
Subject: Re: Performance regression from switching lock to rw-sem for
anon-vma tree
On Mon, 2013-06-17 at 11:45 -0700, Tim Chen wrote:
> On Mon, 2013-06-17 at 09:22 -0700, Davidlohr Bueso wrote:
> > On Sun, 2013-06-16 at 17:50 +0800, Alex Shi wrote:
> > > On 06/14/2013 07:43 AM, Davidlohr Bueso wrote:
> > > > I was hoping that the lack of spin on owner was the main difference with
> > > > rwsems and am/was in the middle of implementing it. Could you send your
> > > > patch so I can give it a try on my workloads?
> > > >
> > > > Note that there have been a few recent (3.10) changes to mutexes that
> > > > give a nice performance boost, specially on large systems, most
> > > > noticeably:
> > > >
> > > > commit 2bd2c92c (mutex: Make more scalable by doing less atomic
> > > > operations)
> > > >
> > > > commit 0dc8c730 (mutex: Queue mutex spinners with MCS lock to reduce
> > > > cacheline contention)
> > > >
> > > > It might be worth looking into doing something similar to commit
> > > > 0dc8c730, in addition to the optimistic spinning.
> > >
> > > It is a good tunning for large machine. I just following what the commit
> > > 0dc8c730 done, give a RFC patch here. I tried it on my NHM EP machine. seems no
> > > clear help on aim7. but maybe it is helpful on large machine. :)
> >
> > After a lot of benchmarking, I finally got the ideal results for aim7,
> > so far: this patch + optimistic spinning with preemption disabled. Just
> > like optimistic spinning, this patch by itself makes little to no
> > difference, yet combined is where we actually outperform 3.10-rc5. In
> > addition, I noticed extra throughput when disabling preemption in
> > try_optimistic_spin().
> >
> > With i_mmap as a rwsem and these changes I could see performance
> > benefits for alltests (+14.5%), custom (+17%), disk (+11%), high_systime
> > (+5%), shared (+15%) and short (+4%), most of them after around 500
> > users, for fewer users, it made little to no difference.
> >
>
> Thanks. Those are encouraging numbers. On my exim workload I didn't
> get a boost when I added in the preempt disable in optimistic spin and
> put Alex's changes in. Can you send me your combined patch to see if
> there may be something you did that I've missed. I have a tweak to
> Alex's patch below to simplify things a bit.
>
I'm using:
int rwsem_optimistic_spin(struct rw_semaphore *sem)
{
struct task_struct *owner;
/* sem->wait_lock should not be held when attempting optimistic spinning */
if (!rwsem_can_spin_on_owner(sem))
return 0;
preempt_disable();
for (;;) {
owner = ACCESS_ONCE(sem->owner);
if (owner && !rwsem_spin_on_owner(sem, owner))
break;
/* wait_lock will be acquired if write_lock is obtained */
if (rwsem_try_write_lock(sem->count, true, sem)) {
preempt_enable();
return 1;
}
/*
* When there's no owner, we might have preempted between the
* owner acquiring the lock and setting the owner field. If
* we're an RT task that will live-lock because we won't let
* the owner complete.
*/
if (!owner && (need_resched() || rt_task(current)))
break;
/*
* The cpu_relax() call is a compiler barrier which forces
* everything in this loop to be re-loaded. We don't need
* memory barriers as we'll eventually observe the right
* values at the cost of a few extra spins.
*/
arch_mutex_cpu_relax();
}
preempt_enable();
return 0;
}
> >
> > >
> > >
> > > diff --git a/include/asm-generic/rwsem.h b/include/asm-generic/rwsem.h
> > > index bb1e2cd..240729a 100644
> > > --- a/include/asm-generic/rwsem.h
> > > +++ b/include/asm-generic/rwsem.h
> > > @@ -70,11 +70,11 @@ static inline void __down_write(struct rw_semaphore *sem)
> > >
> > > static inline int __down_write_trylock(struct rw_semaphore *sem)
> > > {
> > > - long tmp;
> > > + if (unlikely(&sem->count != RWSEM_UNLOCKED_VALUE))
> > > + return 0;
> > >
> > > - tmp = cmpxchg(&sem->count, RWSEM_UNLOCKED_VALUE,
> > > - RWSEM_ACTIVE_WRITE_BIAS);
> > > - return tmp == RWSEM_UNLOCKED_VALUE;
> > > + return cmpxchg(&sem->count, RWSEM_UNLOCKED_VALUE,
> > > + RWSEM_ACTIVE_WRITE_BIAS) == RWSEM_UNLOCKED_VALUE;
> > > }
> > >
> > > /*
> > > diff --git a/lib/rwsem.c b/lib/rwsem.c
> > > index 19c5fa9..9e54e20 100644
> > > --- a/lib/rwsem.c
> > > +++ b/lib/rwsem.c
> > > @@ -64,7 +64,7 @@ __rwsem_do_wake(struct rw_semaphore *sem, enum rwsem_wake_type wake_type)
> > > struct rwsem_waiter *waiter;
> > > struct task_struct *tsk;
> > > struct list_head *next;
> > > - long oldcount, woken, loop, adjustment;
> > > + long woken, loop, adjustment;
> > >
> > > waiter = list_entry(sem->wait_list.next, struct rwsem_waiter, list);
> > > if (waiter->type == RWSEM_WAITING_FOR_WRITE) {
> > > @@ -75,7 +75,7 @@ __rwsem_do_wake(struct rw_semaphore *sem, enum rwsem_wake_type wake_type)
> > > * will block as they will notice the queued writer.
> > > */
> > > wake_up_process(waiter->task);
> > > - goto out;
> > > + return sem;
> > > }
> > >
> > > /* Writers might steal the lock before we grant it to the next reader.
> > > @@ -85,15 +85,28 @@ __rwsem_do_wake(struct rw_semaphore *sem, enum rwsem_wake_type wake_type)
> > > adjustment = 0;
> > > if (wake_type != RWSEM_WAKE_READ_OWNED) {
> > > adjustment = RWSEM_ACTIVE_READ_BIAS;
> > > - try_reader_grant:
> > > - oldcount = rwsem_atomic_update(adjustment, sem) - adjustment;
> > > - if (unlikely(oldcount < RWSEM_WAITING_BIAS)) {
> > > - /* A writer stole the lock. Undo our reader grant. */
> > > + while (1) {
> > > + long oldcount;
> > > +
> > > + /* A writer stole the lock. */
> > > + if (unlikely(sem->count & RWSEM_ACTIVE_MASK))
> > > + return sem;
> > > +
> > > + if (unlikely(sem->count < RWSEM_WAITING_BIAS)) {
> > > + cpu_relax();
> > > + continue;
> > > + }
>
> The above two if statements could be cleaned up as a single check:
>
> if (unlikely(sem->count < RWSEM_WAITING_BIAS))
> return sem;
>
> This one statement is sufficient to check that we don't have a writer
> stolen the lock before we attempt to acquire the read lock by modifying
> sem->count.
We probably still want to keep the cpu relaxation if the statement
doesn't comply.
Thanks,
Davidlohr
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
Powered by blists - more mailing lists