From mboxrd@z Thu Jan 1 00:00:00 1970 From: Eric Dumazet Subject: Re: [PATCHv5 2/2] memory barrier: adding smp_mb__after_lock Date: Tue, 07 Jul 2009 16:42:01 +0200 Message-ID: <4A535EB9.2020406@gmail.com> References: <20090703081219.GE2902@jolsa.lab.eng.brq.redhat.com> <20090703081445.GG2902@jolsa.lab.eng.brq.redhat.com> <20090703090606.GA3902@elte.hu> <4A4DCD54.1080908@gmail.com> <20090703092438.GE3902@elte.hu> <20090703095659.GA4518@jolsa.lab.eng.brq.redhat.com> <20090703102530.GD32128@elte.hu> <20090703111848.GA10267@jolsa.lab.eng.brq.redhat.com> <20090707101816.GA6619@jolsa.lab.eng.brq.redhat.com> <20090707134601.GB6619@jolsa.lab.eng.brq.redhat.com> <20090707140135.GA5506@Krystal> Mime-Version: 1.0 Content-Type: text/plain; charset=ISO-8859-1 Content-Transfer-Encoding: QUOTED-PRINTABLE Cc: Jiri Olsa , Ingo Molnar , Eric Dumazet , Peter Zijlstra , netdev@vger.kernel.org, linux-kernel@vger.kernel.org, fbl@redhat.com, nhorman@redhat.com, davem@redhat.com, htejun@gmail.com, jarkao2@gmail.com, oleg@redhat.com, davidel@xmailserver.org To: Mathieu Desnoyers Return-path: Received: from gw1.cosmosbay.com ([212.99.114.194]:57645 "EHLO gw1.cosmosbay.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1756672AbZGGOmf (ORCPT ); Tue, 7 Jul 2009 10:42:35 -0400 In-Reply-To: <20090707140135.GA5506@Krystal> Sender: netdev-owner@vger.kernel.org List-ID: Mathieu Desnoyers a =E9crit : > * Jiri Olsa (jolsa@redhat.com) wrote: >> On Tue, Jul 07, 2009 at 12:18:16PM +0200, Jiri Olsa wrote: >>> On Fri, Jul 03, 2009 at 01:18:48PM +0200, Jiri Olsa wrote: >>>> On Fri, Jul 03, 2009 at 12:25:30PM +0200, Ingo Molnar wrote: >>>>> * Jiri Olsa wrote: >>>>> >>>>>> On Fri, Jul 03, 2009 at 11:24:38AM +0200, Ingo Molnar wrote: >>>>>>> * Eric Dumazet wrote: >>>>>>> >>>>>>>> Ingo Molnar a =E9crit : >>>>>>>>> * Jiri Olsa wrote: >>>>>>>>> >>>>>>>>>> +++ b/arch/x86/include/asm/spinlock.h >>>>>>>>>> @@ -302,4 +302,7 @@ static inline void __raw_write_unlock(ra= w_rwlock_t *rw) >>>>>>>>>> #define _raw_read_relax(lock) cpu_relax() >>>>>>>>>> #define _raw_write_relax(lock) cpu_relax() >>>>>>>>>> =20 >>>>>>>>>> +/* The {read|write|spin}_lock() on x86 are full memory barr= iers. */ >>>>>>>>>> +#define smp_mb__after_lock() do { } while (0) >>>>>>>>> Two small stylistic comments, please make this an inline func= tion: >>>>>>>>> >>>>>>>>> static inline void smp_mb__after_lock(void) { } >>>>>>>>> #define smp_mb__after_lock >>>>>>>>> >>>>>>>>> (untested) >>>>>>>>> >>>>>>>>>> +/* The lock does not imply full memory barrier. */ >>>>>>>>>> +#ifndef smp_mb__after_lock >>>>>>>>>> +#define smp_mb__after_lock() smp_mb() >>>>>>>>>> +#endif >>>>>>>>> ditto. >>>>>>>>> >>>>>>>>> Ingo >>>>>>>> This was following existing implementations of various smp_mb_= _??? helpers : >>>>>>>> >>>>>>>> # grep -4 smp_mb__before_clear_bit include/asm-generic/bitops.= h >>>>>>>> >>>>>>>> /* >>>>>>>> * clear_bit may not imply a memory barrier >>>>>>>> */ >>>>>>>> #ifndef smp_mb__before_clear_bit >>>>>>>> #define smp_mb__before_clear_bit() smp_mb() >>>>>>>> #define smp_mb__after_clear_bit() smp_mb() >>>>>>>> #endif >>>>>>> Did i mention that those should be fixed too? :-) >>>>>>> >>>>>>> Ingo >>>>>> ok, could I include it in the 2/2 or you prefer separate patch? >>>>> depends on whether it will regress ;-) >>>>> >>>>> If it regresses, it's better to have it separate. If it wont, it = can=20 >>>>> be included. If unsure, default to the more conservative option. >>>>> >>>>> Ingo >>>> >>>> how about this..=20 >>>> and similar change for smp_mb__before_clear_bit in a separate patc= h >>>> >>>> >>>> diff --git a/arch/x86/include/asm/spinlock.h b/arch/x86/include/as= m/spinlock.h >>>> index b7e5db8..4e77853 100644 >>>> --- a/arch/x86/include/asm/spinlock.h >>>> +++ b/arch/x86/include/asm/spinlock.h >>>> @@ -302,4 +302,8 @@ static inline void __raw_write_unlock(raw_rwlo= ck_t *rw) >>>> #define _raw_read_relax(lock) cpu_relax() >>>> #define _raw_write_relax(lock) cpu_relax() >>>> =20 >>>> +/* The {read|write|spin}_lock() on x86 are full memory barriers. = */ >>>> +static inline void smp_mb__after_lock(void) { } >>>> +#define ARCH_HAS_SMP_MB_AFTER_LOCK >>>> + >>>> #endif /* _ASM_X86_SPINLOCK_H */ >>>> diff --git a/include/linux/spinlock.h b/include/linux/spinlock.h >>>> index 252b245..4be57ab 100644 >>>> --- a/include/linux/spinlock.h >>>> +++ b/include/linux/spinlock.h >>>> @@ -132,6 +132,11 @@ do { \ >>>> #endif /*__raw_spin_is_contended*/ >>>> #endif >>>> =20 >>>> +/* The lock does not imply full memory barrier. */ >>>> +#ifndef ARCH_HAS_SMP_MB_AFTER_LOCK >>>> +static inline void smp_mb__after_lock(void) { smp_mb(); } >>>> +#endif >>>> + >>>> /** >>>> * spin_unlock_wait - wait until the spinlock gets unlocked >>>> * @lock: the spinlock in question. >>>> diff --git a/include/net/sock.h b/include/net/sock.h >>>> index 4eb8409..98afcd9 100644 >>>> --- a/include/net/sock.h >>>> +++ b/include/net/sock.h >>>> @@ -1271,6 +1271,9 @@ static inline int sk_has_allocations(const s= truct sock *sk) >>>> * in its cache, and so does the tp->rcv_nxt update on CPU2 side.= The CPU1 >>>> * could then endup calling schedule and sleep forever if there a= re no more >>>> * data on the socket. >>>> + * >>>> + * The sk_has_helper is always called right after a call to read_= lock, so we >>>> + * can use smp_mb__after_lock barrier. >>>> */ >>>> static inline int sk_has_sleeper(struct sock *sk) >>>> { >>>> @@ -1280,7 +1283,7 @@ static inline int sk_has_sleeper(struct sock= *sk) >>>> * >>>> * This memory barrier is paired in the sock_poll_wait. >>>> */ >>>> - smp_mb(); >>>> + smp_mb__after_lock(); >>>> return sk->sk_sleep && waitqueue_active(sk->sk_sleep); >>>> } >>>> =20 >>> any feedback on this?=20 >>> I'd send v6 if this way is acceptable.. >>> >>> thanks, >>> jirka >> also I checked the smp_mb__before_clear_bit/smp_mb__after_clear_bit = and >> it is used quite extensivelly. >> >> I'd prefer to send it in a separate patch, so we can move on with th= e=20 >> changes I've sent so far.. >> >=20 > As with any optimization (and this is one that adds a semantic that w= ill > just grow the memory barrier/locking rule complexity), it should come > with performance benchmarks showing real-life improvements. >=20 > Otherwise I'd recommend sticking to smp_mb() if this execution path i= s > not that critical, or to move to RCU if it's _that_ critical. >=20 > A valid argument would be if the data structures protected are so > complex that RCU is out of question but still the few cycles saved by > removing a memory barrier are really significant. And even then, the > proper solution would be more something like a > __read_lock()+smp_mb+smp_mb+__read_unlock(), so we get the performanc= e > improvements on architectures other than x86 as well. >=20 > So in all cases, I don't think the smp_mb__after_lock() is the > appropriate solution. RCU on this part is out of the question, as David already mentioned it. It would be a regression for short lived tcp/udp sessions, and some wor= kloads use them a lot... We gained about 20% performance between 2.6.26 and 2.6.31, carefuly rem= oving some atomic ops in network stack, adding RCU where it was sensible, but= this is a painful process, not something Jiri can use to fix bugs on legacy = RedHat kernels :) (We still are sorting out regressions) To solve problem pointed by Jiri, we have to insert an smp_mb() at this= point, (not mentioning the other change in select() logic of course) static void sock_def_readable(struct sock *sk, int len) { read_lock(&sk->sk_callback_lock); + smp_mb(); /* paired with opposite smp_mb() in sk poll logic */ if (sk->sk_sleep && waitqueue_active(sk->sk_sleep)) wake_up_interruptible_sync_poll(sk->sk_sleep, POLLIN | POLLRDNORM | POLLRDBAND); sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN); read_unlock(&sk->sk_callback_lock); } As about every incoming packet calls this path, we should be very caref= ul not slowing down stack if not necessary. On x86 this extra smp_mb() is not needed, since previous call to read_l= ock() already gives the full barrier for free.