Re: [PATCH 8/9] sched: Fix migrate_disable() vs set_cpus_allowed_ptr()

From: Peter Zijlstra <peterz@infradead.org>
To: Valentin Schneider <valentin.schneider@arm.com>
Cc: tglx@linutronix.de, mingo@kernel.org,
	linux-kernel@vger.kernel.org, bigeasy@linutronix.de,
	qais.yousef@arm.com, swood@redhat.com, juri.lelli@redhat.com,
	vincent.guittot@linaro.org, dietmar.eggemann@arm.com,
	rostedt@goodmis.org, bsegall@google.com, mgorman@suse.de,
	bristot@redhat.com, vincent.donnefort@arm.com
Subject: Re: [PATCH 8/9] sched: Fix migrate_disable() vs set_cpus_allowed_ptr()
Date: Fri, 25 Sep 2020 11:56:15 +0200	[thread overview]
Message-ID: <20200925095615.GA2651@hirez.programming.kicks-ass.net> (raw)
In-Reply-To: <20200925090528.GV2628@hirez.programming.kicks-ass.net>

On Fri, Sep 25, 2020 at 11:05:28AM +0200, Peter Zijlstra wrote:
> On Thu, Sep 24, 2020 at 08:59:33PM +0100, Valentin Schneider wrote:
> > > @@ -2025,19 +2138,8 @@ static int __set_cpus_allowed_ptr(struct
> > >       if (cpumask_test_cpu(task_cpu(p), new_mask))
> > >               goto out;
> > 
> > I think this needs a cancellation of any potential pending migration
> > requests. Consider a task P0 running on CPU0:
> > 
> >    P0                     P1                               P2
> > 
> >    migrate_disable();
> >    <preempt>
> >                           set_cpus_allowed_ptr(P0, CPU1);
> >                           // waits for completion
> >                                                            set_cpus_allowed_ptr(P0, CPU0);
> >                                                            // Already good, no waiting for completion
> >    <resumes>
> >    migrate_enable();
> >    // task_cpu(p) allowed, no move_task()
> > 
> > AIUI in this scenario P1 would stay forever waiting.
> 

> The other approach is trying to handle that last condition in
> move_task(), but I'm quite sure that's going to be aweful too :/

Something like so perhaps?

--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -2039,6 +2039,10 @@ static int move_task(struct rq *rq, stru
 	if (WARN_ON_ONCE(!pending))
 		return -EINVAL;
 
+	/* Can the task run on the task's current CPU? If so, we're done */
+	if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask))
+		goto easy;
+
 	arg.done = &pending->done;
 
 	if (flags & SCA_MIGRATE_ENABLE) {
@@ -2063,6 +2067,7 @@ static int move_task(struct rq *rq, stru
 			if (task_on_rq_queued(p))
 				rq = move_queued_task(rq, rf, p, dest_cpu);
 
+easy:
 			p->migration_pending = NULL;
 			complete = true;
 		}
@@ -2151,10 +2156,6 @@ static int __set_cpus_allowed_ptr(struct
 			p->nr_cpus_allowed != 1);
 	}
 
-	/* Can the task run on the task's current CPU? If so, we're done */
-	if (cpumask_test_cpu(task_cpu(p), new_mask))
-		goto out;
-
 	return move_task(rq, &rf, p, dest_cpu, flags);
 
 out:
