Re: [PATCH v4] sched/numa: fix unsafe get_task_struct() in task_numa_assign()

[Peter Zijlstra <peterz@infradead.org>]

On Mon, Nov 10, 2014 at 10:48:28AM -0500, Sasha Levin wrote:
> On 11/10/2014 05:03 AM, Peter Zijlstra wrote:
> > --- a/kernel/sched/fair.c
> > +++ b/kernel/sched/fair.c
> > @@ -1268,6 +1268,13 @@ static void task_numa_compare(struct tas
> >  	raw_spin_unlock_irq(&dst_rq->lock);
> >  
> >  	/*
> > +	 * Because we have preemption enabled we can get migrated around and
> > +	 * end try selecting ourselves (current == env->p) as a swap candidate.
> > +	 */
> > +	if (cur == env->p)
> > +		goto unlock;
> 
> This is too late though, because currently the lockup happens couple of lines
> above that at:
> 
>         raw_spin_lock_irq(&dst_rq->lock); <=== here
>         cur = dst_rq->curr;
> 
> Which got me a bit stuck trying to use your old patch since we can't access '->curr'
> without locking dst_rq, but locking dst_rq is causing a lockup.

confused... how can we lock up there. We should not be holding _any_
lock there.

That a different problem that the originally reported one.