Re: [patch v3] mm, oom: fix unnecessary killing of additional processes

[David Rientjes <rientjes@google.com>]

On Fri, 20 Jul 2018, Tetsuo Handa wrote:

> > Absent oom_lock serialization, this is exactly working as intended.  You 
> > could argue that once the thread has reached exit_mmap() and begins oom 
> > reaping that it should be allowed to finish before the oom reaper declares 
> > MMF_OOM_SKIP.  That could certainly be helpful, I simply haven't 
> > encountered a usecase where it were needed.  Or, we could restart the oom 
> > expiration when MMF_UNSTABLE is set and deem that progress is being made 
> > so it give it some extra time.  In practice, again, we haven't seen this 
> > needed.  But either of those are very easy to add in as well.  Which would 
> > you prefer?
> 
> I don't think we need to introduce user-visible knob interface (even if it is in
> debugfs), for I think that my approach can solve your problem. Please try OOM lockup
> (CVE-2016-10723) mitigation patch ( https://marc.info/?l=linux-mm&m=153112243424285&w=4 )

The issue I am fixing has nothing to do with contention on oom_lock, it 
has to do with the inability of the oom reaper to free memory for one or 
more of several reasons: mlock, blockable mmus, ptes, mm->mmap_sem 
contention, and then the setting of MMF_OOM_SKIP to choose another victim 
before the original victim even reaches exit_mmap().  Thus, removing 
oom_lock from exit_mmap() will not fix this issue.

I agree that oom_lock can be removed from exit_mmap() and it would be 
helpful to do so, and may address a series of problems that we have yet to 
encounter, but this would not fix the almost immediate setting of 
MMF_OOM_SKIP that occurs with minimal memory freeing due to the oom 
reaper.