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

[Michal Hocko]

On Wed 13-06-18 22:20:49, Tetsuo Handa wrote:
> On 2018/06/05 17:57, Michal Hocko wrote:
> >> For this reason, we see testing harnesses often oom killed immediately 
> >> after running a unittest that stresses reclaim or compaction by inducing a 
> >> system-wide oom condition.  The harness spawns the unittest which spawns 
> >> an antagonist memory hog that is intended to be oom killed.  When memory 
> >> is mlocked or there are a large number of threads faulting memory for the 
> >> antagonist, the unittest and the harness itself get oom killed because the 
> >> oom reaper sets MMF_OOM_SKIP; this ends up happening a lot on powerpc.  
> >> The memory hog has mm->mmap_sem readers queued ahead of a writer that is 
> >> doing mmap() so the oom reaper can't grab the sem quickly enough.
> > 
> > How come the writer doesn't back off. mmap paths should be taking an
> > exclusive mmap sem in killable sleep so it should back off. Or is the
> > holder of the lock deep inside mmap path doing something else and not
> > backing out with the exclusive lock held?
> > 
>  
> Here is an example where the writer doesn't back off.
> 
>   http://lkml.kernel.org/r/20180607150546.1c7db21f70221008e14b8bb8@xxxxxxxxxxxxxxxxxxxx
> 
> down_write_killable(&mm->mmap_sem) is nothing but increasing the possibility of
> successfully back off. There is no guarantee that the owner of that exclusive
> mmap sem will not be blocked by other unkillable waits.

but we are talking about mmap() path here. Sure there are other paths
which might need a back off while the lock is held and that should be
addressed if possible but this is not really related to what David wrote
above and I tried to understand.

-- 
Michal Hocko
SUSE Labs