From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1753873AbYJ0Wej (ORCPT ); Mon, 27 Oct 2008 18:34:39 -0400 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S1751782AbYJ0Web (ORCPT ); Mon, 27 Oct 2008 18:34:31 -0400 Received: from e2.ny.us.ibm.com ([32.97.182.142]:44627 "EHLO e2.ny.us.ibm.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751328AbYJ0Wea (ORCPT ); Mon, 27 Oct 2008 18:34:30 -0400 Date: Mon, 27 Oct 2008 15:34:29 -0700 From: "Paul E. McKenney" To: Lai Jiangshan Cc: Ingo Molnar , Linux Kernel Mailing List Subject: Re: [PATCH] rcupdate: reduce sys's overhead when rcu_barrier()s called simultaneous Message-ID: <20081027223429.GJ6783@linux.vnet.ibm.com> Reply-To: paulmck@linux.vnet.ibm.com References: <49015F45.1070002@cn.fujitsu.com> <20081026215531.GA6236@linux.vnet.ibm.com> <4905567D.9080600@cn.fujitsu.com> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <4905567D.9080600@cn.fujitsu.com> User-Agent: Mutt/1.5.15+20070412 (2007-04-11) Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On Mon, Oct 27, 2008 at 01:49:49PM +0800, Lai Jiangshan wrote: > Paul E. McKenney wrote: > > On Fri, Oct 24, 2008 at 01:38:13PM +0800, Lai Jiangshan wrote: > >> rcu_barrier() queues rcu_head on all cpus, it will brings > >> large overhead for a large system which has a lots cpu. > >> this fix reduces sys's overhead when rcu_barrier()s called > >> simultaneous. > > > > Hello, Jiangshan, > > > > If we were to have problems with many concurrent rcu_barrier() > > calls stacking up, this patch looks like it would be a reasonable > > was of addressing those problems. > > > > But do we really have problems with this at the moment? > > > > Thanx, Paul > > Hi, Paul, > > Thanks, we do not have problems with this at the moment. > > I suddenly had an association of ideas to synchronize_srcu(), > so I made use of the ideas in synchronize_srcu() and this patch was made. Absolutely nothing wrong with that! ;-) We should make sure that this patch is kept somewhere so that if problems arise, it can be brought to bear quickly and easily. One approach would be to keep it in a git tree, but that would likely incur merge overhead. Another approach would be to have an RCU web page maintaining patches such as this one. Thoughts? Thanx, Paul > Thanx, Lai. > > > > > >> Signed-off-by: Lai Jiangshan > >> --- > >> diff --git a/kernel/rcupdate.c b/kernel/rcupdate.c > >> index ad63af8..734850b 100644 > >> --- a/kernel/rcupdate.c > >> +++ b/kernel/rcupdate.c > >> @@ -53,6 +53,9 @@ enum rcu_barrier { > >> > >> static DEFINE_PER_CPU(struct rcu_head, rcu_barrier_head) = {NULL}; > >> static atomic_t rcu_barrier_cpu_count; > >> +static unsigned long rcu_barrier_completed; > >> +static unsigned long rcu_barrier_completed_bh; > >> +static unsigned long rcu_barrier_completed_sched; > >> static DEFINE_MUTEX(rcu_barrier_mutex); > >> static struct completion rcu_barrier_completion; > >> > >> @@ -60,7 +63,7 @@ static struct completion rcu_barrier_completion; > >> * Awaken the corresponding synchronize_rcu() instance now that a > >> * grace period has elapsed. > >> */ > >> -void wakeme_after_rcu(struct rcu_head *head) > >> +void wakeme_after_rcu(struct rcu_head *head) > >> { > >> struct rcu_synchronize *rcu; > >> > >> @@ -113,11 +116,20 @@ static void rcu_barrier_func(void *type) > >> * Orchestrate the specified type of RCU barrier, waiting for all > >> * RCU callbacks of the specified type to complete. > >> */ > >> -static void _rcu_barrier(enum rcu_barrier type) > >> +static void _rcu_barrier(enum rcu_barrier type, unsigned long *completed) > >> { > >> + unsigned long batch = ACCESS_ONCE(*completed); > >> BUG_ON(in_interrupt()); > >> /* Take cpucontrol mutex to protect against CPU hotplug */ > >> mutex_lock(&rcu_barrier_mutex); > >> + > >> + BUG_ON(*completed & 1); > >> + if ((*completed - batch) >= 2) { > >> + mutex_unlock(&rcu_barrier_mutex); > >> + return; > >> + } > >> + (*completed)++; > >> + > >> init_completion(&rcu_barrier_completion); > >> /* > >> * Initialize rcu_barrier_cpu_count to 1, then invoke > >> @@ -133,6 +145,7 @@ static void _rcu_barrier(enum rcu_barrier type) > >> if (atomic_dec_and_test(&rcu_barrier_cpu_count)) > >> complete(&rcu_barrier_completion); > >> wait_for_completion(&rcu_barrier_completion); > >> + (*completed)++; > >> mutex_unlock(&rcu_barrier_mutex); > >> } > >> > >> @@ -141,7 +154,7 @@ static void _rcu_barrier(enum rcu_barrier type) > >> */ > >> void rcu_barrier(void) > >> { > >> - _rcu_barrier(RCU_BARRIER_STD); > >> + _rcu_barrier(RCU_BARRIER_STD, &rcu_barrier_completed); > >> } > >> EXPORT_SYMBOL_GPL(rcu_barrier); > >> > >> @@ -150,7 +163,7 @@ EXPORT_SYMBOL_GPL(rcu_barrier); > >> */ > >> void rcu_barrier_bh(void) > >> { > >> - _rcu_barrier(RCU_BARRIER_BH); > >> + _rcu_barrier(RCU_BARRIER_BH, &rcu_barrier_completed_bh); > >> } > >> EXPORT_SYMBOL_GPL(rcu_barrier_bh); > >> > >> @@ -159,7 +172,7 @@ EXPORT_SYMBOL_GPL(rcu_barrier_bh); > >> */ > >> void rcu_barrier_sched(void) > >> { > >> - _rcu_barrier(RCU_BARRIER_SCHED); > >> + _rcu_barrier(RCU_BARRIER_SCHED, &rcu_barrier_completed_sched); > >> } > >> EXPORT_SYMBOL_GPL(rcu_barrier_sched); > >> > >> > >> > > > > > > > >