sched/deadline: Add more reschedule cases to prio_changed_dl()
authorValentin Schneider <vschneid@redhat.com>
Mon, 6 Feb 2023 14:06:12 +0000 (14:06 +0000)
committerPeter Zijlstra <peterz@infradead.org>
Sat, 11 Feb 2023 10:18:10 +0000 (11:18 +0100)
I've been tracking down an issue on a ~5.17ish kernel where:

  CPUx                           CPUy

  <DL task p0 owns an rtmutex M>
  <p0 depletes its runtime, gets throttled>
  <rq switches to the idle task>
 <DL task p1 blocks on M, boost/replenish p0>
 <No call to resched_curr() happens here>

  [idle task keeps running here until *something*
   accidentally sets TIF_NEED_RESCHED]

On that kernel, it is quite easy to trigger using rt-tests's deadline_test
[1] with the test running on isolated CPUs (this reduces the chance of
something unrelated setting TIF_NEED_RESCHED on the idle tasks, making the
issue even more obvious as the hung task detector chimes in).

I haven't been able to reproduce this using a mainline kernel, even if I
revert

  2972e3050e35 ("tracing: Make trace_marker{,_raw} stream-like")

which gets rid of the lock involved in the above test, *but* I cannot
convince myself the issue isn't there from looking at the code.

Make prio_changed_dl() issue a reschedule if the current task isn't a
deadline one. While at it, ensure a reschedule is emitted when a
queued-but-not-current task gets boosted with an earlier deadline that
current's.

[1]: https://git.kernel.org/pub/scm/utils/rt-tests/rt-tests.git

Signed-off-by: Valentin Schneider <vschneid@redhat.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Juri Lelli <juri.lelli@redhat.com>
Link: https://lore.kernel.org/r/20230206140612.701871-1-vschneid@redhat.com
kernel/sched/deadline.c

index 0d97d54276cc836367547d11faecdbc28230e069..71b24371a6f7762d231d6c4037dde80a273fe0bb 100644 (file)
@@ -2663,17 +2663,20 @@ static void switched_to_dl(struct rq *rq, struct task_struct *p)
 static void prio_changed_dl(struct rq *rq, struct task_struct *p,
                            int oldprio)
 {
-       if (task_on_rq_queued(p) || task_current(rq, p)) {
+       if (!task_on_rq_queued(p))
+               return;
+
 #ifdef CONFIG_SMP
-               /*
-                * This might be too much, but unfortunately
-                * we don't have the old deadline value, and
-                * we can't argue if the task is increasing
-                * or lowering its prio, so...
-                */
-               if (!rq->dl.overloaded)
-                       deadline_queue_pull_task(rq);
+       /*
+        * This might be too much, but unfortunately
+        * we don't have the old deadline value, and
+        * we can't argue if the task is increasing
+        * or lowering its prio, so...
+        */
+       if (!rq->dl.overloaded)
+               deadline_queue_pull_task(rq);
 
+       if (task_current(rq, p)) {
                /*
                 * If we now have a earlier deadline task than p,
                 * then reschedule, provided p is still on this
@@ -2681,15 +2684,24 @@ static void prio_changed_dl(struct rq *rq, struct task_struct *p,
                 */
                if (dl_time_before(rq->dl.earliest_dl.curr, p->dl.deadline))
                        resched_curr(rq);
-#else
+       } else {
                /*
-                * Again, we don't know if p has a earlier
-                * or later deadline, so let's blindly set a
-                * (maybe not needed) rescheduling point.
+                * Current may not be deadline in case p was throttled but we
+                * have just replenished it (e.g. rt_mutex_setprio()).
+                *
+                * Otherwise, if p was given an earlier deadline, reschedule.
                 */
-               resched_curr(rq);
-#endif /* CONFIG_SMP */
+               if (!dl_task(rq->curr) ||
+                   dl_time_before(p->dl.deadline, rq->curr->dl.deadline))
+                       resched_curr(rq);
        }
+#else
+       /*
+        * We don't know if p has a earlier or later deadline, so let's blindly
+        * set a (maybe not needed) rescheduling point.
+        */
+       resched_curr(rq);
+#endif
 }
 
 DEFINE_SCHED_CLASS(dl) = {