lkml.org 
[lkml]   [2016]   [Feb]   [18]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 6/6] mm,oom: wait for OOM victims when using oom_kill_allocating_task == 1
On Thu 18-02-16 19:45:45, Tetsuo Handa wrote:
> Michal Hocko wrote:
> > On Wed 17-02-16 19:36:36, Tetsuo Handa wrote:
> > > From 0b36864d4100ecbdcaa2fc2d1927c9e270f1b629 Mon Sep 17 00:00:00 2001
> > > From: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
> > > Date: Wed, 17 Feb 2016 16:37:59 +0900
> > > Subject: [PATCH 6/6] mm,oom: wait for OOM victims when using oom_kill_allocating_task == 1
> > >
> > > Currently, out_of_memory() does not wait for existing TIF_MEMDIE threads
> > > if /proc/sys/vm/oom_kill_allocating_task is set to 1. This can result in
> > > killing more OOM victims than needed. We can wait for the OOM reaper to
> > > reap memory used by existing TIF_MEMDIE threads if possible. If the OOM
> > > reaper is not available, the system will be kept OOM stalled until an
> > > OOM-unkillable thread does a GFP_FS allocation request and calls
> > > oom_kill_allocating_task == 0 path.
> > >
> > > This patch changes oom_kill_allocating_task == 1 case to call
> > > select_bad_process() in order to wait for existing TIF_MEMDIE threads.
> >
> > The primary motivation for oom_kill_allocating_task was to reduce the
> > overhead of select_bad_process. See fe071d7e8aae ("oom: add
> > oom_kill_allocating_task sysctl"). So this basically defeats the whole
> > purpose of the feature.
> >
>
> I didn't know that. But I think that printk()ing all candidates much more
> significantly degrades performance than scanning the tasklist.

I assume those who care do set oom_dump_tasks = 0.

> It would be
> nice if setting /proc/sys/vm/oom_dump_tasks = N (N > 1) shows only top N
> memory-hog processes.

You would need scanning of all tasks anyway and sorting etc... Not worth
bothering IMO.

[...]
> We have
>
> "Out of memory (oom_kill_allocating_task)"
> "Out of memory"
> "Memory cgroup out of memory"
>
> but we don't have
>
> "Memory cgroup out of memory (oom_kill_allocating_task)"
>
> I don't know whether we should use this condition for memcg OOM case.

memcg oom killer ignores follow oom_kill_allocating_task.

> > /*
> > * If any of p's children has a different mm and is eligible for kill,
> > * the one with the highest oom_badness() score is sacrificed for its
> > @@ -734,6 +737,7 @@ void oom_kill_process(struct oom_control *oc, struct task_struct *p,
> > }
> > read_unlock(&tasklist_lock);
> >
> > +kill:
> > p = find_lock_task_mm(victim);
> > if (!p) {
> > put_task_struct(victim);
> > @@ -888,6 +892,9 @@ bool out_of_memory(struct oom_control *oc)
> > if (sysctl_oom_kill_allocating_task && current->mm &&
> > !oom_unkillable_task(current, NULL, oc->nodemask) &&
> > current->signal->oom_score_adj != OOM_SCORE_ADJ_MIN) {
> > + if (test_thread_flag(TIF_MEMDIE))
> > + panic("Out of memory (oom_kill_allocating_task) not able to make a forward progress");
> > +
>
> If current thread got TIF_MEMDIE, current thread will not call out_of_memory()
> again because current thread will exit the allocation (unless __GFP_NOFAIL)
> due to use of ALLOC_NO_WATERMARKS.

exactly __GFP_NOFAIL has to be handled properly.

--
Michal Hocko
SUSE Labs

\
 
 \ /
  Last update: 2016-02-18 13:41    [W:1.052 / U:0.632 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site