Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1751664AbYLJLcd (ORCPT ); Wed, 10 Dec 2008 06:32:33 -0500 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S1751523AbYLJLcJ (ORCPT ); Wed, 10 Dec 2008 06:32:09 -0500 Received: from fgwmail5.fujitsu.co.jp ([192.51.44.35]:33459 "EHLO fgwmail5.fujitsu.co.jp" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1750805AbYLJLcH (ORCPT ); Wed, 10 Dec 2008 06:32:07 -0500 Message-ID: <31051.10.75.179.61.1228908723.squirrel@webmail-b.css.fujitsu.com> In-Reply-To: <20081210105000.GC25467@balbir.in.ibm.com> References: <20081210051947.GH7593@balbir.in.ibm.com> <20081210174906.7c1a1a50.kamezawa.hiroyu@jp.fujitsu.com> <20081210105000.GC25467@balbir.in.ibm.com> Date: Wed, 10 Dec 2008 20:32:03 +0900 (JST) Subject: Re: [RFC][RFT] memcg fix cgroup_mutex deadlock when cpusetreclaims memory From: "KAMEZAWA Hiroyuki" To: balbir@linux.vnet.ibm.com Cc: "KAMEZAWA Hiroyuki" , menage@google.com, "KAMEZAWA Hiroyuki" , "Daisuke Miyakawa" , "YAMAMOTO Takashi" , linux-kernel@vger.kernel.org, linux-mm@kvack.org User-Agent: SquirrelMail/1.4.3a X-Mailer: SquirrelMail/1.4.3a MIME-Version: 1.0 Content-Type: text/plain;charset=us-ascii Content-Transfer-Encoding: 8bit X-Priority: 3 (Normal) Importance: Normal Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 5283 Lines: 149 Balbir Singh said: > * KAMEZAWA Hiroyuki [2008-12-10 > 17:49:06]: > >> On Wed, 10 Dec 2008 10:49:47 +0530 >> Balbir Singh wrote: >> >> > Hi, >> > >> > Here is a proposed fix for the memory controller cgroup_mutex deadlock >> > reported. It is lightly tested and reviewed. I need help with review >> > and test. Is the reported deadlock reproducible after this patch? A >> > careful review of the cpuset impact will also be highly appreciated. >> > >> > From: Balbir Singh >> > >> > cpuset_migrate_mm() holds cgroup_mutex throughout the duration of >> > do_migrate_pages(). The issue with that is that >> > >> > 1. It can lead to deadlock with memcg, as do_migrate_pages() >> > enters reclaim >> > 2. It can lead to long latencies, preventing users from creating/ >> > destroying other cgroups anywhere else >> > >> > The patch holds callback_mutex through the duration of >> cpuset_migrate_mm() and >> > gives up cgroup_mutex while doing so. >> > >> > Signed-off-by: Balbir Singh >> > --- >> > >> > include/linux/cpuset.h | 13 ++++++++++++- >> > kernel/cpuset.c | 23 ++++++++++++----------- >> > 2 files changed, 24 insertions(+), 12 deletions(-) >> > >> > diff -puN kernel/cgroup.c~cpuset-remove-cgroup-mutex-from-update-path >> kernel/cgroup.c >> > diff -puN kernel/cpuset.c~cpuset-remove-cgroup-mutex-from-update-path >> kernel/cpuset.c >> > --- a/kernel/cpuset.c~cpuset-remove-cgroup-mutex-from-update-path >> > +++ a/kernel/cpuset.c >> > @@ -369,7 +369,7 @@ static void guarantee_online_mems(const >> > * task has been modifying its cpuset. >> > */ >> > >> > -void cpuset_update_task_memory_state(void) >> > +void __cpuset_update_task_memory_state(bool held) >> > { >> > int my_cpusets_mem_gen; >> > struct task_struct *tsk = current; >> > @@ -380,7 +380,8 @@ void cpuset_update_task_memory_state(voi >> > rcu_read_unlock(); >> > >> > if (my_cpusets_mem_gen != tsk->cpuset_mems_generation) { >> > - mutex_lock(&callback_mutex); >> > + if (!held) >> > + mutex_lock(&callback_mutex); >> > task_lock(tsk); >> > cs = task_cs(tsk); /* Maybe changed when task not locked */ >> > guarantee_online_mems(cs, &tsk->mems_allowed); >> > @@ -394,7 +395,8 @@ void cpuset_update_task_memory_state(voi >> > else >> > tsk->flags &= ~PF_SPREAD_SLAB; >> > task_unlock(tsk); >> > - mutex_unlock(&callback_mutex); >> > + if (!held) >> > + mutex_unlock(&callback_mutex); >> > mpol_rebind_task(tsk, &tsk->mems_allowed); >> > } >> > } >> > @@ -949,13 +951,15 @@ static int update_cpumask(struct cpuset >> > * so that the migration code can allocate pages on these nodes. >> > * >> > * Call holding cgroup_mutex, so current's cpuset won't change >> > - * during this call, as manage_mutex holds off any cpuset_attach() >> > + * during this call, as callback_mutex holds off any >> cpuset_attach() >> > * calls. Therefore we don't need to take task_lock around the >> > * call to guarantee_online_mems(), as we know no one is changing >> > * our task's cpuset. >> > * >> > * Hold callback_mutex around the two modifications of our tasks >> > - * mems_allowed to synchronize with cpuset_mems_allowed(). >> > + * mems_allowed to synchronize with cpuset_mems_allowed(). Give >> > + * up cgroup_mutex to avoid deadlocking with other subsystems >> > + * as we enter reclaim from do_migrate_pages(). >> > * >> > * While the mm_struct we are migrating is typically from some >> > * other task, the task_struct mems_allowed that we are hacking >> > @@ -976,17 +980,14 @@ static void cpuset_migrate_mm(struct mm_ >> > { >> > struct task_struct *tsk = current; >> > >> > - cpuset_update_task_memory_state(); >> > - >> > + cgroup_unlock(); >> > mutex_lock(&callback_mutex); >> > + cpuset_update_task_memory_state_locked(); >> > tsk->mems_allowed = *to; >> > - mutex_unlock(&callback_mutex); >> > - >> > do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL); >> > - >> > - mutex_lock(&callback_mutex); >> > guarantee_online_mems(task_cs(tsk),&tsk->mems_allowed); >> > mutex_unlock(&callback_mutex); >> > + cgroup_lock(); >> > } >> > >> >> Hmm...can't this happen ? >> >> Assume there is a task X and cgroup Z1 and Z2. Z1 and Z2 doesn't need to >> be in >> the same hierarchy. >> == >> CPU A attach task X to cgroup Z1 >> cgroup_lock() >> for_each_subsys_state() > > You mean for_each_subsys() right? > >> => attach(X,Z) >> => migrate_mm() >> => cgroup_unlock() >> migration >> >> CPU B attach task X to cgroup Z2 at the same time >> cgroup_lock() >> replace css_set. >> == >> >> Works on CPU B can't break for_each_subsys_state() in CPU A ? >> > > for_each_subsys is hierarchy aware, so if we try to add the same task > to different hierachies, it should not be a problem right? > Ah, maybe. But what happens when Z1 and Z2 is the same hierarchy ? Are there some locks ? -Kame -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/