Re: [PATCH] mm,oom: Bring OOM notifier callbacks to outside of OOM killer.

[Paul E. McKenney]

On Wed, Jun 27, 2018 at 07:52:23PM +0900, Tetsuo Handa wrote:
> On 2018/06/27 8:50, Paul E. McKenney wrote:
> > On Wed, Jun 27, 2018 at 05:10:48AM +0900, Tetsuo Handa wrote:
> >> As far as I can see,
> >>
> >> -	atomic_set(&oom_callback_count, 1);
> >> +	atomic_inc(&oom_callback_count);
> >>
> >> should be sufficient.
> > 
> > I don't see how that helps.  For example, suppose that two tasks
> > invoked rcu_oom_notify() at about the same time.  Then they could
> > both see oom_callback_count equal to zero, both atomically increment
> > oom_callback_count, then both do the IPI invoking rcu_oom_notify_cpu()
> > on each online CPU.
> > 
> > So far, so good.  But rcu_oom_notify_cpu() enqueues a per-CPU RCU
> > callback, and enqueuing the same callback twice in quick succession
> > would fatally tangle RCU's callback lists.
> > 
> > What am I missing here?
> 
> You are pointing out that "number of rsp->call() is called" > "number of
> rcu_oom_callback() is called" can happen if concurrently called, aren't you?

Yes.  Reusing an rcu_head before invocation of the earlier use is
very bad indeed.  ;-)

> Then, you are not missing anything. You will need to use something equivalent
> to oom_lock even if you can convert rcu_oom_notify() to use shrinkers.

What should I look at to work out whether it makes sense to convert
rcu_oom_notify() to shrinkers, and if so, how to go about it?

Or are you simply asking me to serialize rcu_oom_notify()?  (Which is
of course not difficult, so please just let me know.)

							Thanx, Paul