Re: [RFC PATCH] mm, oom: allow oom reaper to race with exit_mmap

[Hugh Dickins]

On Thu, 20 Jul 2017, Michal Hocko wrote:
> On Wed 19-07-17 18:18:27, Hugh Dickins wrote:
> > 
> > But I haven't looked at the oom_kill or oom_reaper end of it at all,
> > perhaps you have an overriding argument on the placement from that end.
> 
> Well, the main problem here is that the oom_reaper tries to
> MADV_DONTNEED the oom victim and then hide it from the oom killer (by
> setting MMF_OOM_SKIP) to guarantee a forward progress. In order to do
> that it needs mmap_sem for read. Currently we try to avoid races with
> the eixt path by checking mm->mm_users and that can lead to premature
> MMF_OOM_SKIP and that in turn to additional oom victim(s) selection
> while the current one is still tearing the address space down.
> 
> One way around that is to allow final unmap race with the oom_reaper
> tear down.
> 
> I hope this clarify the motivation

Thanks, yes, if you have a good reason of that kind, then I agree that
it's appropriate to leave the down_write(mmap_sem) until reaching the
free_pgtables() stage.

Hugh