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)
commit7ea98dfa44917a201e76d4fe96bf61d76e60f524
tree90a38374272d9e4b55482a703ec47f76f65002ef
parent829c1651e9c4a6f78398d3e67651cef9bb6b42cc
sched/deadline: Add more reschedule cases to prio_changed_dl()

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