From: Michal Hocko <email@example.com>
To: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: firstname.lastname@example.org, email@example.com, firstname.lastname@example.org,
Subject: Re: [PATCH v2] mm/page_alloc: Wait for oom_lock before retrying.
Date: Tue, 18 Jul 2017 11:08:29 +0200 [thread overview]
Message-ID: <20170718090829.GA19133@dhcp22.suse.cz> (raw)
On Tue 18-07-17 06:42:31, Tetsuo Handa wrote:
> Michal Hocko wrote:
> > On Sun 16-07-17 19:59:51, Tetsuo Handa wrote:
> > > Since the whole memory reclaim path has never been designed to handle the
> > > scheduling priority inversions, those locations which are assuming that
> > > execution of some code path shall eventually complete without using
> > > synchronization mechanisms can get stuck (livelock) due to scheduling
> > > priority inversions, for CPU time is not guaranteed to be yielded to some
> > > thread doing such code path.
> > >
> > > mutex_trylock() in __alloc_pages_may_oom() (waiting for oom_lock) and
> > > schedule_timeout_killable(1) in out_of_memory() (already held oom_lock) is
> > > one of such locations, and it was demonstrated using artificial stressing
> > > that the system gets stuck effectively forever because SCHED_IDLE priority
> > > thread is unable to resume execution at schedule_timeout_killable(1) if
> > > a lot of !SCHED_IDLE priority threads are wasting CPU time .
> > >
> > > To solve this problem properly, complete redesign and rewrite of the whole
> > > memory reclaim path will be needed. But we are not going to think about
> > > reimplementing the the whole stack (at least for foreseeable future).
> > >
> > > Thus, this patch workarounds livelock by forcibly yielding enough CPU time
> > > to the thread holding oom_lock by using mutex_lock_killable() mechanism,
> > > so that the OOM killer/reaper can use CPU time yielded by this patch.
> > > Of course, this patch does not help if the cause of lack of CPU time is
> > > somewhere else (e.g. executing CPU intensive computation with very high
> > > scheduling priority), but that is not fault of this patch.
> > > This patch only manages not to lockup if the cause of lack of CPU time is
> > > direct reclaim storm wasting CPU time without making any progress while
> > > waiting for oom_lock.
> > I have to think about this some more. Hitting much more on the oom_lock
> > is a problem while __oom_reap_task_mm still depends on the oom_lock. With
> > http://email@example.com it
> > doesn't do anymore.
> I suggested preserving oom_lock serialization when setting MMF_OOM_SKIP in
> reply to that post (unless we use some trick for force calling
> get_page_from_freelist() after confirming that there is no !MMF_OOM_SKIP mm).
If that is necessary, which I believe it is not, it should be discussed
in that thread email.
> > Also this whole reasoning is little bit dubious to me. The whole reclaim
> > stack might still preempt the holder of the lock so you are addressin
> > only a very specific contention case where everybody hits the oom. I
> > suspect that a differently constructed testcase might result in the same
> > problem.
> I think that direct reclaim/compaction is primary source of CPU time
> consumption, for there will be nothing more to do other than
> get_page_from_freelist() and schedule_timeout_uninterruptible() if
> we are waiting for somebody else to make progress using the OOM killer.
> Thus, if we wait using mutex_lock_killable(), direct reclaim/compaction
> will not be called (i.e. the rest of whole reclaim stack will not preempt
> the holder of the oom_lock) after each allocating thread failed to acquire
> the oom_lock.
But you still assume that you are going to hit the oom path. That takes
some time because you have to go over all reclaim priorities, compaction
attempts and so on. Many allocation paths will eventually hit the oom
path but many will simply still try the reclaim until they get there.
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to firstname.lastname@example.org. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"email@example.com"> firstname.lastname@example.org </a>
prev parent reply other threads:[~2017-07-18 9:08 UTC|newest]
Thread overview: 7+ messages / expand[flat|nested] mbox.gz Atom feed top
2017-07-16 10:59 [PATCH v2] mm/page_alloc: Wait for oom_lock before retrying Tetsuo Handa
2017-07-17 8:56 ` Michal Hocko
2017-07-17 13:50 ` Tetsuo Handa
2017-07-17 15:15 ` Michal Hocko
2017-07-17 15:24 ` Michal Hocko
2017-07-17 21:42 ` Tetsuo Handa
2017-07-18 9:08 ` Michal Hocko [this message]
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).