On Sun 27-01-19 19:56:06, Tetsuo Handa wrote:
> On 2019/01/27 17:37, Michal Hocko wrote:
> > Thanks for the analysis and the patch. This should work, I believe but
> > I am not really thrilled to overload the meaning of the MMF_UNSTABLE.
> > The flag is meant to signal accessing address space is not stable and it
> > is not aimed to synchronize oom reaper with the oom path.
> >
> > Can we make use mark_oom_victim directly? I didn't get to think that
> > through right now so I might be missing something but this should
> > prevent repeating queueing as well.
>
> Yes, TIF_MEMDIE would work. But you are planning to remove TIF_MEMDIE. Also,
> TIF_MEMDIE can't avoid enqueuing many threads sharing mm_struct to the OOM
> reaper. There is no need to enqueue many threads sharing mm_struct because
> the OOM reaper acts on mm_struct rather than task_struct. Thus, enqueuing
> based on per mm_struct flag sounds better, but MMF_OOM_VICTIM cannot be
> set from wake_oom_reaper(victim) because victim's mm might be already inside
> exit_mmap() when wake_oom_reaper(victim) is called after task_unlock(victim).
>
> We could reintroduce MMF_OOM_KILLED in commit 855b018325737f76
> ("oom, oom_reaper: disable oom_reaper for oom_kill_allocating_task")
> if you don't like overloading the meaning of the MMF_UNSTABLE. But since
> MMF_UNSTABLE is available in Linux 4.9+ kernels (which covers all LTS stable
> versions with the OOM reaper support), we can temporarily use MMF_UNSTABLE
> for ease of backporting.
I agree that a per-mm state is more optimal but I would rather fix the
issue in a clear way first and only then think about an optimization on
top. Queueing based on mark_oom_victim (whatever that uses to guarantee
the victim is marked atomically and only once) makes sense from the
conceptual point of view and it makes a lot of sense to start from
there. MMF_UNSTABLE has a completely different purpose. So unless you
see a correctness issue with that then I would rather go that way.
--
Michal Hocko
SUSE Labs
