Re: sched_setscheduler() vs idle_balance() race

[Peter Zijlstra <peterz@infradead.org>]

On Thu, May 28, 2015 at 09:43:52AM +0200, Mike Galbraith wrote:
> Hi Peter,
> 
> I'm not seeing what prevents pull_task() from yanking a task out from
> under __sched_setscheduler().  A box sprinkling smoldering 3.0 kernel
> wreckage all over my bugzilla mbox isn't seeing it either ;-)
> 
> Scenario: rt task forks, wakes child to CPU foo, immediately tries to
> change child to fair class, calls switched_from_rt(), that leads to
> pull_rt_task() -> double_lock_balance() which momentarily drops child's
> rq->lock, letting some prick doing idle balancing over on CPU bar in to
> migrate the child.  Rt parent then calls switched_to_fair(), and box
> explodes when we use the passed rq as if the child still lived there.
> 
> I sent a patchlet to verify that the diagnosis is really really correct
> (can_migrate_task() says no if ->pi_lock is held), but I think it is,
> the 8x10 color glossy with circles and arrows clearly shows both tasks
> with their grubby mitts on that child at the same time, each thinking it
> has that child locked down tight.
> 
> Not seeing what should prevent that in mainline either, I'll just ask
> while I wait to (hopefully) hear "yup, all better".

The last patch to come close is 67dfa1b756f2 ("sched/deadline: Implement
cancel_dl_timer() to use in switched_from_dl()")

Which places the comment /* Possible rq-lock hole */ between
switched_from() and switched_to().

Which is exactly the hole you mean, right?

And that commit talks about how all that is 'safe' because all scheduler
operations take ->pi_lock, which is true, except for load-balancing,
which only uses rq->lock.

Furthermore, we call check_class_changed() _after_ we enqueue the task
on the new class, so balancing can indeed occur.

Lemme go stare at this; ideally we'd call check_class_changed() at
__setscheduler() time where the task is off all rqs, but I suspect
there's 'obvious' problems with that..