Re: frequent lockups in 3.18rc4
From: Ingo Molnar
Date: Sat Dec 13 2014 - 02:43:48 EST
* Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
> On Fri, Dec 12, 2014 at 11:58 AM, David Lang <david@xxxxxxx> wrote:
> >
> > If the machine has NOHZ and has a cpu bound userspace task,
> > it could take quite a while before userspace would trigger a
> > reschedule (at least if I've understood the comments on this
> > thread properly)
>
> The thing is, we'd have to return to user space for that to
> happen. And when we do that, we check the "should we schedule"
> flag again. So races like this really shouldn't matter, but
> there could be something kind-of-similar that just ends up
> causing a wakeup to be delayed.
Furthermore there ought to be a scheduler tick active in that
case - which won't be as fast as an immediate reschedule, but
fast enough to beat the softlockup watchdog's threshold of 20
seconds or so.
That is why I think it would be interesting to examine how the
locked up state looks like: is the system truly locked up,
impossible to log in to, locks held but not released, etc., or is
the lockup transient?
> But it would need to be delayed for seconds (for the RCU
> threads) or for tens of seconds (for the watchdog) to matter.
>
> Which just seems unlikely. Even the "very high load" thing
> shouldn't really matter, since while that could delay one
> particular thread being scheduled, it shouldn't delay the next
> "should we schedule" test. In fact, high load would normally be
> extected to make the next "should we schedule" come faster.
>
> But this is where some load calculation overflow might screw
> things up, of course.
Also, the percpu watchdog threads are SCHED_FIFO:99, woken up
through percpu hrtimers, which are not easy to delay through high
SCHED_OTHER load.
Thanks,
Ingo
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/