Michal Hocko wrote:
> On Sun 29-05-16 01:25:14, Tetsuo Handa wrote:
> > There has been three problems about SysRq-f (manual invocation of the OOM
> > killer) case. To make description simple, this patch assumes situation
> > where the OOM reaper is not called (because the OOM victim's mm is shared
> > by unkillable threads) or not available (due to kthread_run() failure or
> > CONFIG_MMU=n).
> >
> > First is that moom_callback() is not called by moom_work under OOM
> > livelock situation because it does not have a dedicated WQ like vmstat_wq.
> > This problem is not fixed yet.
>
> Why do you mention it in the changelog when it is not related to the
> patch then?
Just we won't forget about it.
>
> Btw. you can (ab)use oom_reaper for that purpose. The patch would be
> quite trivial.
How do you handle CONFIG_MMU=n case?
Are we going to provide oom_reaper for CONFIG_MMU=n case?
>
> > Second is that select_bad_process() chooses a thread group which already
> > has a TIF_MEMDIE thread. Since commit f44666b04605d1c7 ("mm,oom: speed up
> > select_bad_process() loop") changed oom_scan_process_group() to use
> > task->signal->oom_victims, non SysRq-f case will no longer select a
> > thread group which already has a TIF_MEMDIE thread.
>
> I am not sure the reference to the commit is really helpful. The
> behavior you are describing below was there before this commit, the only
> thing that has changed is the scope of the TIF_MEMDIE check.
Indeed. Traversing on all threads can always find a thread group which
already has a TIF_MEMDIE thread when there is already a TIF_MEMDIE thread.
>
> > But SysRq-f case will
> > select such thread group due to returning OOM_SCAN_OK. This patch makes
> > sure that oom_badness() is skipped by making oom_scan_process_group() to
> > return OOM_SCAN_CONTINUE for SysRq-f case.
>
> I am OK with this part. I was suggesting something similar except I
> wanted to skip over tasks which have fatal_signal_pending and that part
> got nacked by David AFAIR. Could you make this a separate patch, please?
I think it is better to change both part with this patch.
>
> > Third is that oom_kill_process() chooses a thread group which already
> > has a TIF_MEMDIE thread when the candidate select_bad_process() chose
> > has children because oom_badness() does not take TIF_MEMDIE into account.
> > This patch checks child->signal->oom_victims before calling oom_badness()
> > if oom_kill_process() was called by SysRq-f case. This resembles making
> > sure that oom_badness() is skipped by returning OOM_SCAN_CONTINUE.
>
> This makes sense to me as well but why should be limit this to sysrq case?
> Does it make any sense to select a child which already got killed for
> normal OOM killer? Anyway I think it would be better to split this into
> its own patch as well.
The reason is described in next paragraph.
Do we prefer immediately killing all children of the allocating task?
If yes, I think it should be a separate patch on top of this patch because
somebody might complain such behavior as a regression.
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>