Re: [PATCH -rt] memcg: use migrate_disable()/migrate_enable( ) in memcg_check_events()

[Yong Zhang <yong.zhang0@gmail.com>]

On Wed, Nov 16, 2011 at 09:12:38AM -0500, Steven Rostedt wrote:
> On Wed, 2011-11-16 at 17:16 +0800, Yong Zhang wrote:
> > Looking at commit 4799401f [memcg: Fix race condition in
> > memcg_check_events() with this_cpu usage], we just want
> > to disable migration. So use the right API in -rt. This
> > will cure below warning.
> > 
> 
> >  mm/memcontrol.c |    4 ++--
> >  1 files changed, 2 insertions(+), 2 deletions(-)
> > 
> > diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> > index 6aff93c..afa1954 100644
> > --- a/mm/memcontrol.c
> > +++ b/mm/memcontrol.c
> > @@ -722,7 +722,7 @@ static void __mem_cgroup_target_update(struct mem_cgroup *memcg, int target)
> >   */
> >  static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
> >  {
> > -	preempt_disable();
> > +	migrate_disable();
> 
> No this won't work. Not even for -rt. If we disable migration but not
> preemption, then two tasks can take this path. And the checks in
> __memcg_event_check() will be corrupted because nothing is protecting
> the updates from two tasks going into the same path.

I assumed that we only care about migration, but obviously I'm wrong.

> 
> Perhaps a local_lock would work.

Yeah, will try tglx's patch later.

Thanks,
Yong