Re: [PATCH] sched: do not stop ticks when cpu is not idle

[Philippe Troin <phil@fifi.org>]

Thomas Gleixner <tglx@linutronix.de> writes:

> On Fri, 18 Jul 2008, eric miao wrote:
> > On Fri, Jul 18, 2008 at 9:52 PM, Thomas Gleixner <tglx@linutronix.de> wrote:
> > >> Thomas, Peter, Dmitry, do you concur with the analysis? (commit below)
> > >
> > > Yes. I did not understand the issue when Jack pointed it out to me,
> > > but with Erics explanation it's really clear. Thanks for tracking that
> > > down.
> > 
> > Actually, Jack did most of the analysis and came up with this quick
> > fix.
> > 
> > >
> > >> It looks a bit ugly to me in the middle of schedule() - is there no wait
> > >> to solve this within kernel/time/*.c ?
> > >
> > > Hmm, yes. I think the proper fix is to enable the tick stop mechanism
> > > in the idle loop and disable it before we go to schedule. That takes
> > > an additional parameter to tick_nohz_stop_sched_tick(), but we then
> > > gain a very clear section where the nohz mimic can be active.
> > >
> > > I'll whip up a patch.
> > 
> > Sounds great, thanks.
> 
> Hey, thanks for tracking that down. I was banging my head against the
> wall when I understood the problem.
> 
> I tried to pinpoint the occasional softlockup bug reports, but I
> probably stared too long into that code so I just saw what I expected
> to see.
> 
> Can you give the patch below a try please ?

Hi Thomas,

I've seen weird timer behavior on both i386 and x86_64 on SMP
machines.  By weird I mean:

 - time stops for a few hours, then resumes as if nothing happened;

 - time flows too fast or slow (4x faster to 2x slower depending on
   phase of the moon);

 - the last one I've seen (yesterday), was: 
      sleep(1) sleeps for 1 second, but
      select(0, NULL, NULL, NULL, 0.5) sleeps for nine seconds.

I have been trying to track this problem for a few weeks now, without
success.  Booting a CONFIG_NO_HZ-enabled kernel with "highres=off
nohz=off" does not make a difference.  However booting a kernel with
CONFIG_NO_HZ and CONFIG_HIGH_RES_TIMERS disabled seems to be working
(I cannot garantee that since I've been using that for 48h so far, but
sometimes the problem takes a few days to manifest itself).

After a cursory reading of your patch, it looks to me that the race
could happen on a kernel compiled with CONFIG_NO_HZ and
CONFIG_HIGH_RES_TIMERS and booted with "nohz=off highres=off".  Can
you confirm that?

If you need more details (dmesg, lspci, etc), I have posted some
details on LKML ( http://lkml.org/lkml/2008/7/9/330 ) and I have a bug
posted on the Fedora/RH bugzilla (
https://bugzilla.redhat.com/show_bug.cgi?id=451824 ).

Phil.