Re: [RFC] [PATCH 2/7 v2] memcg: add memory barrier for checking account move.

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

On Fri 20-01-12 10:08:44, Ying Han wrote:
> On Wed, Jan 18, 2012 at 6:17 PM, KAMEZAWA Hiroyuki
> <kamezawa.hiroyu-+CUm20s59erQFUHtdCDX3A@public.gmane.org> wrote:
> > On Wed, 18 Jan 2012 13:37:59 +0100
> > Michal Hocko <mhocko-AlSwsSmVLrQ@public.gmane.org> wrote:
> >
> >> On Wed 18-01-12 09:06:56, KAMEZAWA Hiroyuki wrote:
> >> > On Tue, 17 Jan 2012 16:26:35 +0100
> >> > Michal Hocko <mhocko-AlSwsSmVLrQ@public.gmane.org> wrote:
> >> >
> >> > > On Fri 13-01-12 17:33:47, KAMEZAWA Hiroyuki wrote:
> >> > > > I think this bugfix is needed before going ahead. thoughts?
> >> > > > ==
> >> > > > From 2cb491a41782b39aae9f6fe7255b9159ac6c1563 Mon Sep 17 00:00:00 2001
> >> > > > From: KAMEZAWA Hiroyuki <kamezawa.hiroyu-+CUm20s59erQFUHtdCDX3A@public.gmane.org>
> >> > > > Date: Fri, 13 Jan 2012 14:27:20 +0900
> >> > > > Subject: [PATCH 2/7] memcg: add memory barrier for checking account move.
> >> > > >
> >> > > > At starting move_account(), source memcg's per-cpu variable
> >> > > > MEM_CGROUP_ON_MOVE is set. The page status update
> >> > > > routine check it under rcu_read_lock(). But there is no memory
> >> > > > barrier. This patch adds one.
> >> > >
> >> > > OK this would help to enforce that the CPU would see the current value
> >> > > but what prevents us from the race with the value update without the
> >> > > lock? This is as racy as it was before AFAICS.
> >> > >
> >> >
> >> > Hm, do I misunderstand ?
> >> > ==
> >> >    update                     reference
> >> >
> >> >    CPU A                        CPU B
> >> >   set value                rcu_read_lock()
> >> >   smp_wmb()                smp_rmb()
> >> >                            read_value
> >> >                            rcu_read_unlock()
> >> >   synchronize_rcu().
> >> > ==
> >> > I expect
> >> > If synchronize_rcu() is called before rcu_read_lock() => move_lock_xxx will be held.
> >> > If synchronize_rcu() is called after rcu_read_lock() => update will be delayed.
> >>
> >> Ahh, OK I can see it now. Readers are not that important because it is
> >> actually the updater who is delayed until all preexisting rcu read
> >> sections are finished.
> >>
> >> In that case. Why do we need both barriers? spin_unlock is a full
> >> barrier so maybe we just need smp_rmb before we read value to make sure
> >> that we do not get stalled value when we start rcu_read section after
> >> synchronize_rcu?
> >>
> >
> > I doubt .... If no barrier, this case happens
> >
> > ==
> >        update                  reference
> >        CPU A                   CPU B
> >        set value
> >        synchronize_rcu()       rcu_read_lock()
> >                                read_value <= find old value
> >                                rcu_read_unlock()
> >                                do no lock
> > ==
> 
> Hi Kame,
> 
> Can you help to clarify a bit more on the example above? Why
> read_value got the old value after synchronize_rcu().

AFAIU it is because rcu_read_unlock doesn't force any memory barrier
and we synchronize only the updater (with synchronize_rcu), so nothing
guarantees that the value set on CPUA is visible to CPUB.

-- 
Michal Hocko
SUSE Labs
SUSE LINUX s.r.o.
Lihovarska 1060/12
190 00 Praha 9    
Czech Republic