Re: [PATCH] sched: fix tsk->pi_lock isn't held when do_set_cpus_allowed()

[Peter Zijlstra <peterz@infradead.org>]

On Tue, Aug 25, 2015 at 07:47:44PM +0800, Wanpeng Li wrote:
> On 8/25/15 6:32 PM, Peter Zijlstra wrote:

> >So Possibly, Maybe (I'm still to wrecked to say for sure), something
> >like this would work:
> >
> >	WARN_ON(debug_locks && (lockdep_is_held(&p->pi_lock) ||
> >				(p->on_rq && lockdep_is_held(&rq->lock))));
> >
> >Instead of those two separate lockdep asserts.
> >
> >Please consider carefully.

So the normal rules for changing task_struct::cpus_allowed are holding
both pi_lock and rq->lock, such that holding either stabilizes the mask.

This is so that wakeup can happen without rq->lock and load-balance
without pi_lock.

>From this we already get the relaxation that we can omit acquiring
rq->lock if the task is not on the rq, because in that case
load-balancing will not apply to it.

** these are the rules currently tested in do_set_cpus_allowed() **

Now, since __set_cpus_allowed_ptr() uses task_rq_lock() which
unconditionally acquires both locks, we could get away with holding just
rq->lock when on_rq for modification because that'd still exclude
__set_cpus_allowed_ptr(), it would also work against
__kthread_bind_mask() because that assumes !on_rq.

That said, this is all somewhat fragile.

> Commit (5e16bbc2f: sched: Streamline the task migration locking a little)
> won't hold the pi_lock in migrate_tasks() path any more, actually pi_lock
> was still not held when call select_fallback_rq() and it was held in
> __migrate_task() before  the commit. Then commit (25834c73f93: sched: Fix a
> race between __kthread_bind() and sched_setaffinity()) add a
> lockdep_assert_held() in do_set_cpus_allowed(), the bug is triggered. How
> about something like below:
> 
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -5186,6 +5186,15 @@ static void migrate_tasks(struct rq *dead_rq)
>                 BUG_ON(!next);
>                 next->sched_class->put_prev_task(rq, next);
> 
> +               raw_spin_unlock(&rq->lock);
> +               raw_spin_lock(&next->pi_lock);
> +               raw_spin_lock(&rq->lock);
> +               if (!(task_rq(next) == rq && task_on_rq_queued(next))) {
> +                       raw_spin_unlock(&rq->lock);
> +                       raw_spin_unlock(&next->pi_lock);
> +                       continue;
> +               }

Yeah, that's quite disgusting.. also you'll trip over the lockdep_pin if
you were to actually run this.

Now, I don't think dropping rq->lock is quite as disastrous as it
usually is because !cpu_active at this point, which means load-balance
will not interfere, but that too is somewhat fragile.


So we end up with a choice of two fragile.. let me ponder that a wee
bit more.