Re: [PATCH v3 3/6] memcg: Simplify mem_cgroup_force_empty_list error handling

[Michal Hocko <mhocko-AlSwsSmVLrQ@public.gmane.org>]

On Tue 13-11-12 16:10:41, Johannes Weiner wrote:
> On Tue, Oct 30, 2012 at 11:35:59AM +0100, Michal Hocko wrote:
> > On Mon 29-10-12 15:00:22, Andrew Morton wrote:
> > > On Mon, 29 Oct 2012 17:58:45 +0400
> > > Glauber Costa <glommer-bzQdu9zFT3WakBO8gow8eQ@public.gmane.org> wrote:
> > > 
> > > > > + * move charges to its parent or the root cgroup if the group has no
> > > > > + * parent (aka use_hierarchy==0).
> > > > > + * Although this might fail (get_page_unless_zero, isolate_lru_page or
> > > > > + * mem_cgroup_move_account fails) the failure is always temporary and
> > > > > + * it signals a race with a page removal/uncharge or migration. In the
> > > > > + * first case the page is on the way out and it will vanish from the LRU
> > > > > + * on the next attempt and the call should be retried later.
> > > > > + * Isolation from the LRU fails only if page has been isolated from
> > > > > + * the LRU since we looked at it and that usually means either global
> > > > > + * reclaim or migration going on. The page will either get back to the
> > > > > + * LRU or vanish.
> > > > 
> > > > I just wonder for how long can it go in the worst case?
> > > 
> > > If the kernel is uniprocessor and the caller is SCHED_FIFO: ad infinitum!
> > 
> > You are right, if the rmdir (resp. echo > force_empty) at SCHED_FIFO
> > races with put_page (on a shared page) which gets preempted after
> > put_page_testzero and before __page_cache_release then we are screwed:
> > 
> > 						put_page(page)
> > 						  put_page_testzero
> > 						  <preempted and page still on LRU>
> > mem_cgroup_force_empty_list
> >   page = list_entry(list->prev, struct page, lru);
> >   mem_cgroup_move_parent(page)
> >     get_page_unless_zero <fails>
> >   cond_resched() <scheduled again>
> > 
> > The race window is really small but it is definitely possible. I am not
> > happy about this state and it should be probably mentioned in the
> > patch description but I do not see any way around (except for hacks like
> > sched_setscheduler for the current which is, ehm...) and still keep
> > do_not_fail contract here.
> > 
> > Can we consider this as a corner case (it is much easier to kill a
> > machine with SCHED_FIFO than this anyway) or the concern is really
> > strong and we should come with a solution before this can get merged?
> 
> Wouldn't the much bigger race window be reclaim having the page
> isolated and SCHED_FIFO preventing it from putback?

We wouldn't see the page on the LRU then, right?

> I also don't think this is a new class of problem, though.
> 
> Would it make sense to stick a wait_on_page_locked() in there just so
> that we don't busy spin on a page under migration/reclaim?

Hmm, this would also mean that get_page_unless_zero would fail as well
and so we would schedule in mem_cgroup_force_empty_list. It is true that
there might be no other runnable task so we can busy loop so yes this
would help. Care to cook the patch?

Thanks
-- 
Michal Hocko
SUSE Labs