Re: [PATCH v4 10/19] sched: Fix migrate_disable() vs set_cpus_allowed_ptr()

[Peter Zijlstra <peterz@infradead.org>]

On Thu, Nov 12, 2020 at 07:31:12PM +0000, Valentin Schneider wrote:

> I think what is happening here is:
> 
>   affine_move_task()
>       // task_running() case
>       stop_one_cpu()
>       wait_for_completion(&pending->done);
> 
> and this is !PREEMPT, so the stopper can very well hit:
> 
>   migration_cpu_stop()
>     // Task moved between unlocks and scheduling the stopper
>     task_rq(p) != rq &&
>     // task_running() case
>     dest_cpu >= 0
> 
>     => no complete_all(), ever :(

Damn...

> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> index 02076e6d3792..fad0a8e62aca 100644
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -1923,7 +1923,7 @@ static int migration_cpu_stop(void *data)
>  		else
>  			p->wake_cpu = dest_cpu;
>  
> -	} else if (dest_cpu < 0) {
> +	} else if (dest_cpu < 0 || pending) {
>  		/*
>  		 * This happens when we get migrated between migrate_enable()'s
>  		 * preempt_enable() and scheduling the stopper task. At that
> @@ -1933,6 +1933,17 @@ static int migration_cpu_stop(void *data)
>  		 * more likely.
>  		 */
>  
> +		/*
> +		 * The task moved before the stopper got to run. We're holding
> +		 * ->pi_lock, so the allowed mask is stable - if it got
> +		 * somewhere allowed, we're done.
> +		 */
> +		if (pending && cpumask_test_cpu(task_cpu(p), p->cpus_ptr)) {
> +			p->migration_pending = NULL;
> +			complete = true;
> +			goto out;
> +		}
> +
>  		/*
>  		 * When this was migrate_enable() but we no longer have an
>  		 * @pending, a concurrent SCA 'fixed' things and we should be

Agreed, this is very clearly a missing case and the proposed solution
seems straight forward enough; but I'm struggling to convince my sleep
deprived brain we're actually complete now.

I'll continue staring at it a little more. Could you make it into a
proper patch?