Re: [RFC 3/3] exit: Check for MMF_OOM_SKIP in exit_mmap

From: Michal Hocko <mhocko@suse.com>
To: Nico Pache <npache@redhat.com>
Cc: linux-kernel@vger.kernel.org, linux-mm@kvack.org,
	David Rientjes <rientjes@google.com>,
	Andrea Arcangeli <aarcange@redhat.com>
Subject: Re: [RFC 3/3] exit: Check for MMF_OOM_SKIP in exit_mmap
Date: Tue, 26 Apr 2022 08:59:29 +0200	[thread overview]
Message-ID: <YmeYUX7Nh47j3gKc@dhcp22.suse.cz> (raw)
In-Reply-To: <19303483-5700-fb6e-ba4a-398913370100@redhat.com>

On Mon 25-04-22 15:00:24, Nico Pache wrote:
> 
> 
> On 4/22/22 11:38, Michal Hocko wrote:
> > On Thu 21-04-22 15:05:33, Nico Pache wrote:
> >> The MMF_OOM_SKIP bit is used to indicate weather a mm_struct can not be
> >> invalided or has already been invalided. exit_mmap currently calls
> >> __oom_reap_task_mm unconditionally despite the fact that the oom reaper
> >> may have already called this.
> >>
> >> Add a check for the MMF_OOM_SKIP bit being set in exit_mmap to avoid
> >> unnessary calls to the invalidate code.
> > 
> > Why do we care about this?
> Is there no cost to the MMU/TLB invalidation? The MMU notifier contains a lock
> too so perhaps we can also avoids some unnecessary MMU notifier lock contention.

I am pretty sure that this area can be micro optimized but I do not
really see a strong reason for that. OOM victims are/should be really
rare so I do not think that any performance optimization would be really
visible.

If you want to improve the code then I think a much better plan would be
to get rid of the whole oom special case altogether. This might be much
closer than ever after Hugh's recent m{un}lock changes. I didn't have
time to think that through though. I believe Suren Baghdasaryan has been
looking into that as well.
-- 
Michal Hocko
SUSE Labs