lkml.org 
[lkml]   [2022]   [Mar]   [30]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
SubjectRe: [PATCH v5] mm/oom_kill.c: futex: Close a race between do_exit and the oom_reaper
From


On 3/30/22 03:18, Michal Hocko wrote:
> Nico,
>
> On Wed 23-03-22 10:17:29, Michal Hocko wrote:
>> Let me skip over futex part which I need to digest and only focus on the
>> oom side of the things for clarification.
>>
>> On Tue 22-03-22 23:43:18, Thomas Gleixner wrote:
> [...]
>>> You can easily validate that by doing:
>>>
>>> wake_oom_reaper(task)
>>> task->reap_time = jiffies + HZ;
>>> queue_task(task);
>>> wakeup(reaper);
>>>
>>> and then:
>>>
>>> oom_reap_task(task)
>>> now = READ_ONCE(jiffies);
>>> if (time_before(now, task->reap_time)
>>> schedule_timeout_idle(task->reap_time - now);
>>>
>>> before trying to actually reap the mm.
>>>
>>> That will prevent the enforced race in most cases and allow the exiting
>>> and/or killed processes to cleanup themself. Not pretty, but it should
>>> reduce the chance of the reaper to win the race with the exiting and/or
>>> killed process significantly.
>>>
>>> It's not going to work when the problem is combined with a heavy VM
>>> overload situation which keeps a guest (or one/some it's vCPUs) away
>>> from being scheduled. See below for a discussion of guarantees.
>>>
>>> If it failed to do so when the sleep returns, then you still can reap
>>> it.
>>
>> Yes, this is certainly an option. Please note that the oom_reaper is not
>> the only way to trigger this. process_mrelease syscall performs the same
>> operation from the userspace. Arguably process_mrelease could be used
>> sanely/correctly because the userspace oom killer can do pro-cleanup
>> steps before going to final SIGKILL & process_mrelease. One way would be
>> to send SIGTERM in the first step and allow the victim to perform its
>> cleanup.
>
> are you working on another version of the fix/workaround based on the
> discussion so far?

We are indeed! Sorry for the delay we've been taking the time to do our due
diligence on some of the claims made. We are also spending time rewriting the
reproducer to include more test cases that Thomas brought up.

Ill summarize here, and reply to the original emails in more detail....

Firstly, we have implemented & tested the VMA skipping... it does fix our case.
Thomas brought up a few good points about the robust list head and the potential
waiters being in different VMAs; however, I think its a moot point, given that
the locks will only be reaped if allocated as ((private|anon)|| !shared).

If the locks are private|anon then there are not two processes sharing this lock
and OOMing will stop all the pthreads operating on this address space (no need
to wake the waiters).

If its a shared lock then the VMA will not be reaped and the dying process can
still recover the stuck waiter in the other process. We do however need to skip
the robust_list_head vma as that is in private|anon memory and can race between
the exit and the OOM as described in our previous commit logs.

We are trying to create a scenario in our reproducer that will break the VMA
skipping case, but have been unsuccessful so far.

>
> I guess the most reasonable way forward is to rework
> oom_reaper processing to be delayed. This can be either done by a
> delayed wake up or as Thomas suggests above by postponing the
> processing. I think the delayed wakeup would be _slightly_ easier to
> handle because the queue can contain many tasks to be reaped.
>
> More specifically something like delayed work but we cannot rely on
> the WQ here. I guess we do not have any delayed wait queue interface
> but the same trick with the timer should be applicable here as well.
> exit_mmap would then cancel the timer after __oom_reap_task_mm is done.
> Actually the timer could be canceled after mmu_notifier_release already
> but this shouldn't make much of a difference.

We have not tried implementing the delayed work yet, but we see value in that
solution IFF we are able to break the vma skipping case. In that case I believe
we combine the two approaches to further improve the robustness of the futex,
allowing for less 'best-effort' recoveries bailouts.

Cheers,
-- Nico

\
 
 \ /
  Last update: 2022-03-30 20:19    [W:0.254 / U:0.140 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site