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

["Rafael J. Wysocki" <rjw@rjwysocki.net>]

On Thursday, September 11, 2014 04:45:09 PM Michal Hocko wrote:
> On Thu 11-09-14 16:52:50, Rafael J. Wysocki wrote:
> > On Thursday, September 11, 2014 04:28:22 PM Michal Hocko wrote:
> > > On Thu 11-09-14 16:32:50, Rafael J. Wysocki wrote:
> > > > On Thursday, September 11, 2014 04:10:51 PM Michal Hocko wrote:
> > > > > On Thu 11-09-14 16:26:56, Rafael J. Wysocki wrote:
> > > > > > On Thursday, September 11, 2014 04:04:48 PM Michal Hocko wrote:
> > > > > > > On Thu 11-09-14 16:17:56, Rafael J. Wysocki wrote:
> > > > > > > [...]
> > > > > > > > And I'm still wondering if the OOM killer may be made avoid killing frozen
> > > > > > > > tasks.
> > > > > > > 
> > > > > > > This is really tricky. OOM killer aims at the biggest memory hog. We
> > > > > > > shouldn't ignore it just because it hides into the fridge... So even
> > > > > > > if we "fix" oom killer to ignore frozen tasks (which is inherently
> > > > > > > racy btw.) then we have a potential problem of freezer abuse (e.g. in
> > > > > > > container environments). So I strongly believe that the OOM killer has
> > > > > > > to be able to kill a frozen tasks.
> > > > > > 
> > > > > > OK
> > > > > > 
> > > > > > Is the OOM killer the only place where TIF_MEMDIE is set?
> > > > > 
> > > > > Yes. To be precise, lowmemorykiller (staging android thingy), global OOM
> > > > > killer and memcg OOM killer. Any other users would be an abuse.
> > > > 
> > > > OK
> > > > 
> > > > So can we ensure that those things don't trigger during system suspend (or
> > > > equivalent) and then simply use the TIF_MEMDIE check in __refrigerator()?
> > > 
> > > That would require that no memory allocation triggers OOM killer during
> > > suspend. I don't think this will work out. OOM killer is the last resort
> > > action. We cannot simply give access to memory reserves just because the
> > > current context is in the middle of suspend.
> > 
> > But we can fail the allocation, can't we?
> 
> We already do that by oom_killer_disable after all tasks are frozen in
> freeze_processes. This is before all other device specific things are
> done so I guess we cannot start killing after any device is suspended,
> right? This should be sufficient.

Yes, it should.

> > > What is the worst thing that might happen when a task is killed in the
> > > middle of suspend? I thought that suspend would fail after several
> > > attempts to suspend all existing tasks.
> > 
> > The problem is what to do when we need to kill a frozen task.
> > 
> > In that case we need to thaw it and then it will die eventually.  Unfortunately,
> > it generally can do something undesirable before dying.  That may be accessing
> > a suspended device, for example.
> 
> OK, I have misunderstood you obviously. I thought we were discussing OOM
> while we are in the middle of freezing tasks. After they are frozen
> there is no OOM killer as per above.

Right.

-- 
I speak only for myself.
Rafael J. Wysocki, Intel Open Source Technology Center.