Re: [PATCH] sched/dl: Fix race between dl_task_timer() and sched_setaffinity()

[Peter Zijlstra <peterz@infradead.org>]

On Tue, May 20, 2014 at 09:08:53AM +0400, Kirill Tkhai wrote:
> 
> 
> 20.05.2014, 04:00, "Peter Zijlstra" <peterz@infradead.org>:
> > On Mon, May 19, 2014 at 11:31:19PM +0400, Kirill Tkhai wrote:
> >
> >>  @@ -513,9 +513,17 @@ static enum hrtimer_restart dl_task_timer(struct hrtimer *timer)
> >>                                                        struct sched_dl_entity,
> >>                                                        dl_timer);
> >>           struct task_struct *p = dl_task_of(dl_se);
> >>  - struct rq *rq = task_rq(p);
> >>  + struct rq *rq;
> >>  +again:
> >>  + rq = task_rq(p);
> >>           raw_spin_lock(&rq->lock);
> >>
> >>  + if (unlikely(rq != task_rq(p))) {
> >>  + /* Task was moved, retrying. */
> >>  + raw_spin_unlock(&rq->lock);
> >>  + goto again;
> >>  + }
> >>  +
> >
> > That thing is called: rq = __task_rq_lock(p);
> 
> But p->pi_lock is not held. The problem is __task_rq_lock() has lockdep assert.
> Should we change it?

Ok, so now that I'm awake ;-)

So the trivial problem as described by your initial changelog isn't
right, because we cannot call sched_setaffinity() on deadline tasks, or
rather we can, but we can't actually change the affinity mask.

Now I suppose the problem can still actually happen when you change the
root domain and trigger a effective affinity change that way.

That said, no leave it as you proposed, adding a *task_rq_lock() variant
without lockdep assert in will only confuse things, as normally we
really should be also taking ->pi_lock.

The only reason we don't strictly need ->pi_lock now is because we're
guaranteed to have p->state == TASK_RUNNING here and are thus free of
ttwu races.