Re: [Patch v4 1/2] freezer: check OOM kill while being frozen

[Michal Hocko <mhocko@suse.cz>]

On Wed 10-09-14 13:30:25, Cong Wang wrote:
> On Mon, Sep 8, 2014 at 4:29 PM, Cong Wang <xiyou.wangcong@gmail.com> wrote:
> > On Mon, Sep 8, 2014 at 4:42 PM, Rafael J. Wysocki <rjw@rjwysocki.net> wrote:
> >> On Monday, September 08, 2014 04:16:15 PM Cong Wang wrote:
> >>> On Mon, Sep 8, 2014 at 4:23 PM, Rafael J. Wysocki <rjw@rjwysocki.net> wrote:
> >>> >
> >>> > The reason why it matters for the suspend-time freezing is that we freeze tasks
> >>> > to take them out of the picture entirely until they are thawed.  Therefore we
> >>> > can't allow them to go back to the picture just for a while until they are
> >>> > killed.  Frozen tasks are not supposed to get back to the picture at all.
> >>> >
> >>>
> >>>
> >>> Ok, then checking TIF_MEMDIE is unsafe for PM freeze, we should
> >>> keep the cgroup_freezing() test to make sure freeze request is from
> >>> cgroup not PM. Question got answered. :)
> >>
> >> Do I think correctly that cgroups freezing and system suspend are
> >> mutually exclusive?  If not, then this still is problematic.
> >
> > Good point! Although rare, but it is possible we freeze a process both from
> > cgroup and PM. Hmm, this means we have to explicitly exclude PM rather
> > just checking cgroup freeze? Interesting, but I am not familiar with PM.
> >
> 
> I am wondering if the folllowing check makes any sense with regarding
> to rule out PM freeze:
> 
>         if ((!pm_nosig_freezing && !pm_freezing) &&
>             cgroup_freezing(current) && test_thread_flag(TIF_MEMDIE))
>                 return true;

Why is this needed in the first place? What if OOM happens in the middle
of task freezing during freeze_processes? Also killing a task at that
stage should be safe as no devices are suspended yet and OOM killer is
disabled after all tasks are frozen and allocations fail at that stage.
-- 
Michal Hocko
SUSE Labs