lkml.org 
[lkml]   [2013]   [Nov]   [30]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [merged] mm-memcg-handle-non-error-oom-situations-more-gracefully.patch removed from -mm tree
On Fri, 29 Nov 2013, Johannes Weiner wrote:

> > You said you have informed stable to not merge these patches until further
> > notice, I'd suggest simply avoid ever merging the whole series into a
> > stable kernel since the problem isn't serious enough. Marking changes
> > that do "goto nomem" seem fine to mark for stable, though.
>
> These are followup fixes for the series that is upstream but didn't go
> to stable. I truly have no idea what you are talking about.
>

I'm responding to your comments[*] that indicate you were going to
eventually be sending it to stable.

> > On the scale that we run memcg, we would see it daily in automated testing
> > primarily because we panic the machine for memcg oom conditions where
> > there are no killable processes. It would typically manifest by two
> > processes that are allocating memory in a memcg; one is oom killed, is
> > allowed to allocate, handles its SIGKILL, exits and frees its memory and
> > the second process which is oom disabled races with the uncharge and is
> > oom disabled so the machine panics.
>
> So why don't you implement proper synchronization instead of putting
> these random checks all over the map to make the race window just
> small enough to not matter most of the time?
>

The oom killer can be quite expensive, so we have found that is
advantageous after doing all that work that the memcg is still oom for
the charge order before needlessly killing a process. I am not suggesting
that we add synchronization to the uncharge path for such a race, but
merely a simple check as illustrated as due diligence. I think a simple
conditional in the oom killer to avoid needlessly killing a user job is
beneficial and avoids questions from customers who have a kernel log
showing an oom kill occurring in a memcg that is not oom. We could even
do the check in oom_kill_process() after dump_header() if you want to
reduce any chance of that to avoid getting bug reports about such cases.

> If you are really bothered by this race, then please have OOM kill
> invocations wait for any outstanding TIF_MEMDIE tasks in the same
> context.
>

The oom killer requires a tasklist scan, or an iteration over the set of
processes attached to the memcg for the memcg case, to find a victim. It
already defers if it finds eligible threads with TIF_MEMDIE set.

Thanks.

[*] http://marc.info/?l=linux-mm&m=138559524422298
http://marc.info/?l=linux-kernel&m=138539243412073


\
 
 \ /
  Last update: 2013-11-30 12:01    [W:0.092 / U:0.116 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site