lkml.org 
[lkml]   [2018]   [Jun]   [15]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [patch] mm, oom: fix unnecessary killing of additional processes
On Fri, 15 Jun 2018, Michal Hocko wrote:

> > Signed-off-by: David Rientjes <rientjes@google.com>
>
> Nacked-by: Michal Hocko <mhocko@suse.com>
> as already explained elsewhere in this email thread.
>

I don't find this to be surprising, but I'm not sure that it actually
matters if you won't fix a regression that you introduced. Tetsuo
initially found this issue and presented a similar solution, so I think
his feedback on this is more important since it would fix a problem for
him as well.

> > ---
> > Note: I understand there is an objection based on timeout based delays.
> > This is currently the only possible way to avoid oom killing important
> > processes completely unnecessarily. If the oom reaper can someday free
> > all memory, including mlocked memory and those mm's with blockable mmu
> > notifiers, and is guaranteed to always be able to grab mm->mmap_sem,
> > this can be removed. I do not believe any such guarantee is possible
> > and consider the massive killing of additional processes unnecessarily
> > to be a regression introduced by the oom reaper and its very quick
> > setting of MMF_OOM_SKIP to allow additional processes to be oom killed.
>
> If you find oom reaper more harmful than useful I would be willing to
> ack a comman line option to disable it. Especially when you keep
> claiming that the lockups are not really happening in your environment.
>

There's no need to disable it, we simply need to ensure that it doesn't
set MMF_OOM_SKIP too early, which my patch does. We also need to avoid
setting MMF_OOM_SKIP in exit_mmap() until after all memory has been freed,
i.e. after free_pgtables().

I'd be happy to make the this timeout configurable, however, and default
it to perhaps one second as the blockable mmu notifier timeout in your own
code does. I find it somewhat sad that we'd need a sysctl for this, but
if that will appease you and it will help to move this into -mm then we
can do that.

> Other than that I've already pointed to a more robust solution. If you
> are reluctant to try it out I will do, but introducing a timeout is just
> papering over the real problem. Maybe we will not reach the state that
> _all_ the memory is reapable but we definitely should try to make as
> much as possible to be reapable and I do not see any fundamental
> problems in that direction.

You introduced the timeout already, I'm sure you realized yourself that
the oom reaper sets MMF_OOM_SKIP much too early. Trying to grab
mm->mmap_sem 10 times in a row with HZ/10 sleeps in between is a timeout.
If there are blockable mmu notifiers, your code puts the oom reaper to
sleep for HZ before setting MMF_OOM_SKIP, which is a timeout. This patch
moves the timeout to reaching exit_mmap() where we actually free all
memory possible and still allow for additional oom killing if there is a
very rare oom livelock.

You haven't provided any data that suggests oom livelocking isn't a very
rare event and that we need to respond immediately by randomly killing
more and more processes rather than wait a bounded period of time to allow
for forward progress to be made. I have constantly provided data showing
oom livelock in our fleet is extremely rare, less than 0.04% of the time.
Yet your solution is to kill many processes so this 0.04% is fast.

The reproducer on powerpc is very simple. Do an mmap() and mlock() the
length. Fork one 120MB process that does that and two 60MB processes that
do that in a 128MB memcg.

[ 402.064375] Killed process 17024 (a.out) total-vm:134080kB, anon-rss:122032kB, file-rss:1600kB
[ 402.107521] Killed process 17026 (a.out) total-vm:64448kB, anon-rss:44736kB, file-rss:1600kB

Completely reproducible and completely unnecessary. Killing two processes
pointlessly when the first oom kill would have been successful.

Killing processes is important, optimizing for 0.04% of cases of true oom
livelock by insisting everybody tolerate excessive oom killing is not. If
you have data to suggest the 0.04% is higher, please present it. I'd be
interested in any data you have that suggests its higher and has even
1/1,000,000th oom occurrence rate that I have shown.

It's inappropriate to merge code that oom kills many processes
unnecessarily when one happens to be mlocked or have blockable mmu
notifiers or when mm->mmap_sem can't be grabbed fast enough but forward
progress is actually being made. It's a regression, and it impacts real
users. Insisting that we fix the problem you introduced by making all mmu
notifiers unblockable and mlocked memory can always be reaped and
mm->mmap_sem can always be grabbed within a second is irresponsible.

\
 
 \ /
  Last update: 2018-06-16 01:16    [W:1.247 / U:0.176 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site