Re: [FALSE ALARM] Re: HPET (?) related hangs and breakage in 2.6.35,36

[Thomas Gleixner <tglx@linutronix.de>]

[-- Attachment #1: Type: TEXT/PLAIN, Size: 1782 bytes --]

On Wed, 10 Nov 2010, Andrew Lutomirski wrote:
> On Wed, Nov 10, 2010 at 3:52 PM, Thomas Gleixner <tglx@linutronix.de> wrote:
> > On Wed, 10 Nov 2010, Andrew Lutomirski wrote:
> >
> >> On Wed, Nov 10, 2010 at 1:50 PM, Borislav Petkov <bp@amd64.org> wrote:
> >> > On Wed, Nov 10, 2010 at 01:48:00PM -0500, Andrew Lutomirski wrote:
> >> >> > Clocksource: tsc unstable (delta = -34355296774 ns)
> >> >> > Switching: to clocksource hpet
> >> >>
> >> >> Please disregard -- this is a bug in nouveau (or drm) not hpet.  I'll
> >> >> send a bug report to the maintainers.
> >> >
> >> > Interesting! Joerg was complaining about similar symptoms with .36 today
> >> > too.
> >>
> >> Well, there is a clocksource sort-of-bug that could cause confusion:
> >> when something totally unrelated to clocksources goes out to lunch,
> >> the clocksource watchdog decides that the clocksource is unstable and
> >> complains, steering everyone toward filing the wrong bug.
> >
> > How should the clocksource watchdog code know that something went to
> > lunch? The fact that we need to monitor TSC at all is horrible enough,
> > adding further heuristics to detect extended lunch breaks would be
> > just a PITA.
> 
> Could the clocksource watchdog detect when it gets woken up (i.e. when
> the hardware timer fires) instead of when it gets scheduled?  It is
> internal to the timing code, after all...
> 
> Alternatively, maybe the watchdog could just compare the TSC timestamp
> to the current value according to some other clock (PIT?  whatever
> clockevent is in use?) instead of just using the time passed into the
> delayed work in the first place.

It compares to some other clock. i.e. HPET in your case, but the
extended lunch break was larger than the HPET wrap around time :(

Thanks,

	tglx