[PATCH v6 28/29] x86/tsc: Restart NMI watchdog after refining tsc_khz

[Ricardo Neri]

On Tue, May 10, 2022 at 01:44:05PM +0200, Thomas Gleixner wrote:
> On Tue, May 10 2022 at 21:16, Nicholas Piggin wrote:
> > Excerpts from Ricardo Neri's message of May 6, 2022 10:00 am:
> >> +	/*
> >> +	 * If in use, the HPET hardlockup detector relies on tsc_khz.
> >> +	 * Reconfigure it to make use of the refined tsc_khz.
> >> +	 */
> >> +	lockup_detector_reconfigure();
> >
> > I don't know if the API is conceptually good.
> >
> > You change something that the lockup detector is currently using, 
> > *while* the detector is running asynchronously, and then reconfigure
> > it. What happens in the window? If this code is only used for small
> > adjustments maybe it does not really matter but in principle it's
> > a bad API to export.
> >
> > lockup_detector_reconfigure as an internal API is okay because it
> > reconfigures things while the watchdog is stopped [actually that
> > looks untrue for soft dog which uses watchdog_thresh in
> > is_softlockup(), but that should be fixed].
> >
> > You're the arch so you're allowed to stop the watchdog and configure
> > it, e.g., hardlockup_detector_perf_stop() is called in arch/.
> >
> > So you want to disable HPET watchdog if it was enabled, then update
> > wherever you're using tsc_khz, then re-enable.
> 
> The real question is whether making this refined tsc_khz value
> immediately effective matters at all. IMO, it does not because up to
> that point the watchdog was happily using the coarse calibrated value
> and the whole use TSC to assess whether the HPET fired mechanism is just
> a guestimate anyway. So what's the point of trying to guess 'more
> correct'.

In some of my test systems I observed that, the TSC value does not fall
within the expected error window the first time the HPET channel expires.

I inferred that the error computed using the coarser tsc_khz was wrong.
Recalculating the error window with refined tsc_khz would correct it.

However, restarting the timer has the side-effect of kicking the timer and,
therefore pushing the first HPET NMI further in the future.

Perhaps kicking HPET channel, not recomputing the error window, corrected
(masked?) the problem.

I will investigate further and rework or drop this patch as needed.

Thanks and BR,
Ricardo