sched/rt: Plug rt_mutex_setprio() vs push_rt_task() race

John reported that push_rt_task() can end up invoking
find_lowest_rq(rq->curr) when curr is not an RT task (in this case a CFS
one), which causes mayhem down convert_prio().

This can happen when current gets demoted to e.g. CFS when releasing an
rt_mutex, and the local CPU gets hit with an rto_push_work irqwork before
getting the chance to reschedule. Exactly who triggers this work isn't
entirely clear to me - switched_from_rt() only invokes rt_queue_pull_task()
if there are no RT tasks on the local RQ, which means the local CPU can't
be in the rto_mask.

My current suspected sequence is something along the lines of the below,
with the demoted task being current.

  mark_wakeup_next_waiter()
    rt_mutex_adjust_prio()
      rt_mutex_setprio() // deboost originally-CFS task
	check_class_changed()
	  switched_from_rt() // Only rt_queue_pull_task() if !rq->rt.rt_nr_running
	  switched_to_fair() // Sets need_resched
      __balance_callbacks() // if pull_rt_task(), tell_cpu_to_push() can't select local CPU per the above
      raw_spin_rq_unlock(rq)

       // need_resched is set, so task_woken_rt() can't
       // invoke push_rt_tasks(). Best I can come up with is
       // local CPU has rt_nr_migratory >= 2 after the demotion, so stays
       // in the rto_mask, and then:

       <some other CPU running rto_push_irq_work_func() queues rto_push_work on this CPU>
	 push_rt_task()
	   // breakage follows here as rq->curr is CFS

Move an existing check to check rq->curr vs the next pushable task's
priority before getting anywhere near find_lowest_rq(). While at it, add an
explicit sched_class of rq->curr check prior to invoking
find_lowest_rq(rq->curr). Align the DL logic to also reschedule regardless
of next_task's migratability.

Fixes: a7c8155 ("sched: Fix migrate_disable() vs rt/dl balancing")
Reported-by: John Keeping <[email protected]>
Signed-off-by: Valentin Schneider <[email protected]>
Signed-off-by: Peter Zijlstra (Intel) <[email protected]>
Reviewed-by: Dietmar Eggemann <[email protected]>
Tested-by: John Keeping <[email protected]>
Link: https://lore.kernel.org/r/[email protected]

Author: Valentin Schneider

kernel/sched/deadline.c

@@ -2240,12 +2240,6 @@ static int push_dl_task(struct rq *rq)
 		return 0;
 
 retry:
-	if (is_migration_disabled(next_task))
-		return 0;
-
-	if (WARN_ON(next_task == rq->curr))
-		return 0;
-
 	/*
 	 * If next_task preempts rq->curr, and rq->curr
 	 * can move away, it makes sense to just reschedule
@@ -2258,6 +2252,12 @@ static int push_dl_task(struct rq *rq)
 		return 0;
 	}
 
+	if (is_migration_disabled(next_task))
+		return 0;
+
+	if (WARN_ON(next_task == rq->curr))
+		return 0;
+
 	/* We might release rq lock */
 	get_task_struct(next_task);
 

kernel/sched/rt.c

@@ -2026,13 +2026,35 @@ static int push_rt_task(struct rq *rq, bool pull)
 		return 0;
 
 retry:
+	/*
+	 * It's possible that the next_task slipped in of
+	 * higher priority than current. If that's the case
+	 * just reschedule current.
+	 */
+	if (unlikely(next_task->prio < rq->curr->prio)) {
+		resched_curr(rq);
+		return 0;
+	}
+
 	if (is_migration_disabled(next_task)) {
 		struct task_struct *push_task = NULL;
 		int cpu;
 
 		if (!pull || rq->push_busy)
 			return 0;
 
+		/*
+		 * Invoking find_lowest_rq() on anything but an RT task doesn't
+		 * make sense. Per the above priority check, curr has to
+		 * be of higher priority than next_task, so no need to
+		 * reschedule when bailing out.
+		 *
+		 * Note that the stoppers are masqueraded as SCHED_FIFO
+		 * (cf. sched_set_stop_task()), so we can't rely on rt_task().
+		 */
+		if (rq->curr->sched_class != &rt_sched_class)
+			return 0;
+
 		cpu = find_lowest_rq(rq->curr);
 		if (cpu == -1 || cpu == rq->cpu)
 			return 0;
@@ -2057,16 +2079,6 @@ static int push_rt_task(struct rq *rq, bool pull)
 	if (WARN_ON(next_task == rq->curr))
 		return 0;
 
-	/*
-	 * It's possible that the next_task slipped in of
-	 * higher priority than current. If that's the case
-	 * just reschedule current.
-	 */
-	if (unlikely(next_task->prio < rq->curr->prio)) {
-		resched_curr(rq);
-		return 0;
-	}
-
 	/* We might release rq lock */
 	get_task_struct(next_task);