From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from psmtp.com (na3sys010amx191.postini.com [74.125.245.191]) by kanga.kvack.org (Postfix) with SMTP id E6CD26B0007 for ; Tue, 12 Feb 2013 12:56:57 -0500 (EST) Date: Tue, 12 Feb 2013 18:56:51 +0100 From: Michal Hocko Subject: Re: [PATCH v3 4/7] memcg: remove memcg from the reclaim iterators Message-ID: <20130212175651.GC17663@dhcp22.suse.cz> References: <20130211151649.GD19922@dhcp22.suse.cz> <20130211175619.GC13218@cmpxchg.org> <20130211192929.GB29000@dhcp22.suse.cz> <20130211195824.GB15951@cmpxchg.org> <20130211212756.GC29000@dhcp22.suse.cz> <20130211223943.GC15951@cmpxchg.org> <20130212095419.GB4863@dhcp22.suse.cz> <20130212151002.GD15951@cmpxchg.org> <20130212154330.GG4863@dhcp22.suse.cz> <20130212161051.GQ2666@linux.vnet.ibm.com> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <20130212161051.GQ2666@linux.vnet.ibm.com> Sender: owner-linux-mm@kvack.org List-ID: To: "Paul E. McKenney" Cc: Johannes Weiner , linux-mm@kvack.org, linux-kernel@vger.kernel.org, KAMEZAWA Hiroyuki , Ying Han , Tejun Heo , Glauber Costa , Li Zefan On Tue 12-02-13 08:10:51, Paul E. McKenney wrote: > On Tue, Feb 12, 2013 at 04:43:30PM +0100, Michal Hocko wrote: > > On Tue 12-02-13 10:10:02, Johannes Weiner wrote: > > > On Tue, Feb 12, 2013 at 10:54:19AM +0100, Michal Hocko wrote: > > > > On Mon 11-02-13 17:39:43, Johannes Weiner wrote: > > > > > On Mon, Feb 11, 2013 at 10:27:56PM +0100, Michal Hocko wrote: > > > > > > On Mon 11-02-13 14:58:24, Johannes Weiner wrote: > > > > > > > That way, if the dead count gives the go-ahead, you KNOW that the > > > > > > > position cache is valid, because it has been updated first. > > > > > > > > > > > > OK, you are right. We can live without css_tryget because dead_count is > > > > > > either OK which means that css would be alive at least this rcu period > > > > > > (and RCU walk would be safe as well) or it is incremented which means > > > > > > that we have started css_offline already and then css is dead already. > > > > > > So css_tryget can be dropped. > > > > > > > > > > Not quite :) > > > > > > > > > > The dead_count check is for completed destructions, > > > > > > > > Not quite :P. dead_count is incremented in css_offline callback which is > > > > called before the cgroup core releases its last reference and unlinks > > > > the group from the siblinks. css_tryget would already fail at this stage > > > > because CSS_DEACT_BIAS is in place at that time but this doesn't break > > > > RCU walk. So I think we are safe even without css_get. > > > > > > But you drop the RCU lock before you return. > > > > > > dead_count IS incremented for every destruction, but it's not reliable > > > for concurrent ones, is what I meant. Again, if there is a dead_count > > > mismatch, your pointer might be dangling, easy case. However, even if > > > there is no mismatch, you could still race with a destruction that has > > > marked the object dead, and then frees it once you drop the RCU lock, > > > so you need try_get() to check if the object is dead, or you could > > > return a pointer to freed or soon to be freed memory. > > > > Wait a moment. But what prevents from the following race? > > > > rcu_read_lock() > > mem_cgroup_css_offline(memcg) > > root->dead_count++ > > iter->last_dead_count = root->dead_count > > iter->last_visited = memcg > > // final > > css_put(memcg); > > // last_visited is still valid > > rcu_read_unlock() > > [...] > > // next iteration > > rcu_read_lock() > > iter->last_dead_count == root->dead_count > > // KABOOM > > > > The race window between dead_count++ and css_put is quite big but that > > is not important because that css_put can happen anytime before we start > > the next iteration and take rcu_read_lock. > > The usual approach is to make sure that there is a grace period (either > synchronize_rcu() or call_rcu()) between the time that the data is > made inaccessible to readers (this would be mem_cgroup_css_offline()?) > and the time it is freed (css_put(), correct?). Yes, that was my suggestion and I put it before dead_count is incremented down the mem_cgroup_css_offline road. Johannes still thinks we can do without it if we reduce the number of atomic_read(dead_count) which sounds like a way to go but I will rather think about it with a fresh head tomorrow. Anyway, thanks for jumping in. Earth is always a bit shaky when all the barriers and rcu mix together. -- Michal Hocko SUSE Labs -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@kvack.org. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: email@kvack.org