Re: [PATCH 0/2] mm, oom: do not grant oom victims full memory reserves access

[Roman Gushchin <guro@fb.com>]

On Tue, Aug 01, 2017 at 02:29:05PM +0200, Michal Hocko wrote:
> On Tue 01-08-17 13:23:44, Roman Gushchin wrote:
> > On Tue, Aug 01, 2017 at 02:16:44PM +0200, Michal Hocko wrote:
> > > On Thu 27-07-17 11:03:55, Michal Hocko wrote:
> > > > Hi,
> > > > this is a part of a larger series I posted back in Oct last year [1]. I
> > > > have dropped patch 3 because it was incorrect and patch 4 is not
> > > > applicable without it.
> > > > 
> > > > The primary reason to apply patch 1 is to remove a risk of the complete
> > > > memory depletion by oom victims. While this is a theoretical risk right
> > > > now there is a demand for memcg aware oom killer which might kill all
> > > > processes inside a memcg which can be a lot of tasks. That would make
> > > > the risk quite real.
> > > > 
> > > > This issue is addressed by limiting access to memory reserves. We no
> > > > longer use TIF_MEMDIE to grant the access and use tsk_is_oom_victim
> > > > instead. See Patch 1 for more details. Patch 2 is a trivial follow up
> > > > cleanup.
> > > 
> > > Any comments, concerns? Can we merge it?
> > 
> > I've rebased the cgroup-aware OOM killer and ran some tests.
> > Everything works well.
> 
> Thanks for your testing. Can I assume your Tested-by?

Sure.

I wonder if we can get rid of TIF_MEMDIE completely,
if we will count OOM victims on per-oom-victim-signal-struct rather than
on per-thread basis? Say, assign oom_mm using cmpxchg, and call
exit_oom_victim() from __exit_signal()? __thaw_task() can be called from
mark_oom_victim() unconditionally.

Do you see any problems with this approach?