lkml.org 
[lkml]   [2016]   [Dec]   [14]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 2/2] mm, oom: do not enfore OOM killer for __GFP_NOFAIL automatically
On Mon 12-12-16 09:48:37, Michal Hocko wrote:
> On Sun 11-12-16 20:23:47, Tetsuo Handa wrote:
[...]
> > I believe that __GFP_NOFAIL should not imply invocation of the OOM killer.
> > Therefore, I want to change __GFP_NOFAIL not to invoke the OOM killer.
> > But since currently the OOM killer is not invoked unless either __GFP_FS or
> > __GFP_NOFAIL is specified, changing __GFP_NOFAIL not to invoke the OOM
> > killer introduces e.g. GFP_NOFS | __GFP_NOFAIL users a risk of livelocking
> > by not invoking the OOM killer. Although I can't prove that this change
> > never causes livelock, I don't want to provide an alternative flag like
> > __GFP_WANT_OOM_KILLER. Therefore, all existing __GFP_NOFAIL users must
> > agree with accepting the risk introduced by this change.
>
> I think you are seriously misled here. First of all, I have gone through
> GFP_NOFS | GFP_NOFAIL users and _none_ of them have added the nofail
> flag to enforce the OOM killer. Those users just want to express that an
> allocation failure is simply not acceptable. Most of them were simply
> conversions from the open-conded
> do { } while (! (page = page_alloc(GFP_NOFS));
> loops. Which _does_ not invoke the OOM killer. And that is the most
> importatnt point here. Why the above open coded (and as you say lockup
> prone) loop is OK while GFP_NOFAIL varian should behave any differently?
>
> > and confirm that all existing __GFP_NOFAIL users are willing to accept
> > the risk of livelocking by not invoking the OOM killer.
> >
> > Unless you do this procedure, I continue:
> >
> > Nacked-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
>
> I was hoping for some actual arguments but I am afraid this is circling
> in a loop. You are still handwaving with theoretical lockups without any
> actual proof they are real. While I am not saying the risk is not there
> I also say that there are other aspects to consider
> - lockups will happen only if there are no other GFP_FS requests
> which trigger the OOM which is quite unlikely in most
> situations
> - triggering oom for GFP_NOFS | GFP_NOFAIL has a non negligible
> risk of pre-mature OOM killer invocation for the same reason
> we do not trigger oom for GFP_NOFS. Even worse metadata heavy
> workloads are much harder to contain so this might be used as
> a DoS vector.
> - one of the primary point of GFP_NOFAIL existence is to prevent
> from open coding endless loops in the code because the page
> allocator can handle most situations more gracefully (e.g.
> grant access to memory reserves). Having a completely
> different OOM killer behavior is both confusing and encourages
> abuse. If we have users who definitely need to control the OOM
> behavior then we should add a gfp flag for them. But this
> needs a strong use case and consider whether there are other
> options to go around that.
>
> I can add the above to the changelog if you think this is helpful but I
> still maintain my position that your "this might cause lockups
> theoretically" is unfounded and not justified to block the patch. I will
> of course retract this patch if you can demonstrate the issue is real or
> that any of my argumentation in the changelog is not correct.

I was thinking about this some more and realized that there is a
different risk which this patch would introduce and have to be
considered. Heavy GFP_NOFS | __GFP_NOFAIL users might actually deplete
memory reserves. This was less of a problem with the current code
because we invoke the oom killer and so at least _some_ memory might be
freed. I will think about it some more but I guess I will just allow a
partial access in the no-oom case. I will post the patch 1 in the
meantime because I believe this is a reasonable cleanup.
--
Michal Hocko
SUSE Labs

\
 
 \ /
  Last update: 2016-12-14 11:35    [W:0.199 / U:0.152 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site