Re: [patch 04/12] mm: oom_kill: remove unnecessary locking in exit_oom_victim()

From: Johannes Weiner <hannes@cmpxchg.org>
To: Michal Hocko <mhocko@suse.cz>
Cc: linux-mm@kvack.org, linux-fsdevel@vger.kernel.org,
	linux-kernel@vger.kernel.org,
	Linus Torvalds <torvalds@linux-foundation.org>,
	Andrew Morton <akpm@linux-foundation.org>,
	Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>,
	Huang Ying <ying.huang@intel.com>,
	Andrea Arcangeli <aarcange@redhat.com>,
	Dave Chinner <david@fromorbit.com>,
	"Theodore Ts'o" <tytso@mit.edu>
Subject: Re: [patch 04/12] mm: oom_kill: remove unnecessary locking in exit_oom_victim()
Date: Thu, 26 Mar 2015 11:10:50 -0400	[thread overview]
Message-ID: <20150326151050.GB23973@cmpxchg.org> (raw)
In-Reply-To: <20150326130106.GG15257@dhcp22.suse.cz>

On Thu, Mar 26, 2015 at 02:01:06PM +0100, Michal Hocko wrote:
> On Thu 26-03-15 13:53:48, Michal Hocko wrote:
> > On Wed 25-03-15 02:17:08, Johannes Weiner wrote:
> > > Disabling the OOM killer needs to exclude allocators from entering,
> > > not existing victims from exiting.
> > 
> > The idea was that exit_oom_victim doesn't miss a waiter.
> > 
> > exit_oom_victim is doing
> > 	atomic_dec_return(&oom_victims) && oom_killer_disabled)
> > 
> > so there is a full (implicit) memory barrier befor oom_killer_disabled
> > check. The other part is trickier. oom_killer_disable does:
> > 	oom_killer_disabled = true;
> >         up_write(&oom_sem);
> > 
> >         wait_event(oom_victims_wait, !atomic_read(&oom_victims));
> > 
> > up_write doesn't guarantee a full memory barrier AFAICS in
> > Documentation/memory-barriers.txt (although the generic and x86
> > implementations seem to implement it as a full barrier) but wait_event
> > implies the full memory barrier (prepare_to_wait_event does spin
> > lock&unlock) before checking the condition in the slow path. This should
> > be sufficient and docummented...
> > 
> > 	/*
> > 	 * We do not need to hold oom_sem here because oom_killer_disable
> > 	 * guarantees that oom_killer_disabled chage is visible before
> > 	 * the waiter is put into sleep (prepare_to_wait_event) so
> > 	 * we cannot miss a wake up.
> > 	 */
> > 
> > in unmark_oom_victim()
> 
> OK, I can see that the next patch removes oom_killer_disabled
> completely. The dependency won't be there and so the concerns about the
> memory barriers.
> 
> Is there any reason why the ordering is done this way? It would sound
> more logical to me.

I honestly didn't even think about the dependency between the lock and
this check.  They both looked unnecessary to me and I stopped putting
any more thought into it once I had convinced myself that they are.

The order was chosen because the waitqueue generalization seemed like
a bigger deal.  One is just an unnecessary lock, but this extra check
cost me quite some time debugging and seems like a much more harmful
piece of code to fix.  It's no problem to reorder the patches, though.