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

[KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>]

On Mon, 23 Jan 2012 10:04:36 +0100
Michal Hocko <mhocko@suse.cz> wrote:

> On Fri 20-01-12 10:08:44, Ying Han wrote:
> > On Wed, Jan 18, 2012 at 6:17 PM, KAMEZAWA Hiroyuki
> > <kamezawa.hiroyu@jp.fujitsu.com> wrote:
> > > On Wed, 18 Jan 2012 13:37:59 +0100
> > > Michal Hocko <mhocko@suse.cz> wrote:
> > >
> > >> On Wed 18-01-12 09:06:56, KAMEZAWA Hiroyuki wrote:
> > >> > On Tue, 17 Jan 2012 16:26:35 +0100
> > >> > Michal Hocko <mhocko@suse.cz> 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@jp.fujitsu.com>
> > >> > > > 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 ?
> > >> > ==
> > >> > A  A update A  A  A  A  A  A  A  A  A  A  reference
> > >> >
> > >> > A  A CPU A A  A  A  A  A  A  A  A  A  A  A  A CPU B
> > >> > A  set value A  A  A  A  A  A  A  A rcu_read_lock()
> > >> > A  smp_wmb() A  A  A  A  A  A  A  A smp_rmb()
> > >> > A  A  A  A  A  A  A  A  A  A  A  A  A  A read_value
> > >> > A  A  A  A  A  A  A  A  A  A  A  A  A  A rcu_read_unlock()
> > >> > A  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
> > >
> > > ==
> > > A  A  A  A update A  A  A  A  A  A  A  A  A reference
> > > A  A  A  A CPU A A  A  A  A  A  A  A  A  A  CPU B
> > > A  A  A  A set value
> > > A  A  A  A synchronize_rcu() A  A  A  rcu_read_lock()
> > > A  A  A  A  A  A  A  A  A  A  A  A  A  A  A  A read_value <= find old value
> > > A  A  A  A  A  A  A  A  A  A  A  A  A  A  A  A rcu_read_unlock()
> > > A  A  A  A  A  A  A  A  A  A  A  A  A  A  A  A 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.
> 

Thank you. 

...Finally, I'd like to make this check to atomic_t rather than complicated
percpu counter. Hmm, do it now ?

Thanks,
-Kame


--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>