Re: [PATCH v4 3/3] sched: optimize migration by forcing rmb() and updating to be called once

[Peter Zijlstra <peterz@infradead.org>]

On Tue, Nov 10, 2015 at 10:09:05AM +0900, Byungchul Park wrote:
> On Mon, Nov 09, 2015 at 02:29:14PM +0100, Peter Zijlstra wrote:
> > On Sat, Oct 24, 2015 at 01:16:21AM +0900, byungchul.park@lge.com wrote:
> > > +++ b/kernel/sched/core.c
> > > @@ -1264,6 +1264,8 @@ EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
> > >  
> > >  void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
> > >  {
> > > +	unsigned int prev_cpu = task_cpu(p);
> > > +
> > >  #ifdef CONFIG_SCHED_DEBUG
> > >  	/*
> > >  	 * We should never call set_task_cpu() on a blocked task,
> > > @@ -1289,15 +1291,14 @@ void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
> > >  #endif
> > >  
> > >  	trace_sched_migrate_task(p, new_cpu);
> > > +	__set_task_cpu(p, new_cpu);
> > >  
> > > -	if (task_cpu(p) != new_cpu) {
> > > +	if (prev_cpu != new_cpu) {
> > >  		if (p->sched_class->migrate_task_rq)
> > > -			p->sched_class->migrate_task_rq(p, new_cpu);
> > > +			p->sched_class->migrate_task_rq(p, prev_cpu);
> > >  		p->se.nr_migrations++;
> > >  		perf_event_task_migrate(p);
> > >  	}
> > > -
> > > -	__set_task_cpu(p, new_cpu);
> > >  }
> > 
> > I don't think this is safe, see the comment in __set_task_cpu(). We want
> > that to be last.
> 
> I am sorry but I don't understand what you said. I checked the comment in 
> __set_task_cpu().
> 
> 	/*
> 	 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
> 	 * successfuly executed on another CPU. We must ensure that updates of
> 	 * per-task data have been completed by this moment.
> 	 */
> 
> Of course, ->cpu should be set up to a new value for task_rq_lock() to be
> executed successfully on another CPU. Is this the case? Is there something
> i missed? I think it would be ok if task->pi_lock can work correctly within
> "if" statement in set_task_cpu(). Is there problem to do that?

So the problem is that as soon as that ->cpu store comes through, the
other rq->lock can happen, even though we might still hold a rq->lock
thinking we're serialized.

Take for instance move_queued_tasks(), it does:

	dequeue_task(rq, p, 0);
	p->on_rq = TASK_ON_RQ_MIGRATING;
	set_task_cpu(p, new_cpu) {
	  __set_task_cpu();

^^^ here holding rq->lock is insufficient and the below:

	  p->sched_class->migrate_task_rq()

would no longer be serialized by rq->lock.

	}
	raw_spin_unlock(&rq->lock);