Re: [PATCH 0/2] clocksource: Avoid incorrect hpet fallback

["Paul E. McKenney" <paulmck@kernel.org>]

On Wed, Nov 10, 2021 at 06:25:14PM -0500, Waiman Long wrote:
> 
> On 11/10/21 17:32, Paul E. McKenney wrote:
> > On Wed, Nov 10, 2021 at 05:17:30PM -0500, Waiman Long wrote:
> > > It was found that when an x86 system was being stressed by running
> > > various different benchmark suites, the clocksource watchdog might
> > > occasionally mark TSC as unstable and fall back to hpet which will
> > > have a signficant impact on system performance.
> > > 
> > > The current watchdog clocksource skew threshold of 50us is found to be
> > > insufficient. So it is changed back to 100us before commit 2e27e793e280
> > > ("clocksource: Reduce clocksource-skew threshold") in patch 1. Patch 2
> > > adds a Kconfig option to allow kernel builder to control the actual
> > > threshold to be used.
> > > 
> > > Waiman Long (2):
> > >    clocksource: Avoid accidental unstable marking of clocksources
> > >    clocksource: Add a Kconfig option for WATCHDOG_MAX_SKEW
> > The ability to control the fine-grained threshold seems useful, but is
> > the TSC still marked unstable when this commit from -rcu is applied?
> > It has passed significant testing on other workloads.
> > 
> > 2a43fb0479aa ("clocksource: Forgive repeated long-latency watchdog clocksource reads")
> > 
> > If the patch below takes care of your situation, my thought is to
> > also take your second patch, which would allow people to set the
> > cutoff more loosely or more tightly, as their situation dictates.
> > 
> > Thoughts?
> 
> That is commit 14dbb29eda51 ("clocksource: Forgive repeated long-latency
> watchdog clocksource reads") in your linux-rcu git tree. From reading the
> patch, I believe it should be able to address the hpet fallback problem that
> Red Hat had encountered. Your patch said it was an out-of-tree patch. Are
> you planning to mainline it?

Yes, I expect to submit it into the next merge window (not the current
v5.16 merge window, but v5.17).  However, if your situation is urgent, and
if it works for you, I could submit it as a fix for an earlier regression.

> Patch 1 of this series contains some testing data that caused hpet fallback
> in our testing runs. In summary, a clock skew of 100us is found to be enough
> to avoid the problem with benchmark runs. However, we have some cases where
> TSC was marked unstable at bootup time with a skew of 200us or more which, I
> believe, was caused by the thermal stress that the system was experiencing
> after running stressful benchmarks for hours.

This sort of thing does show some value for allowing the threshold to
be adjusted.  I hope that it does not prove necessary to dynamically
adjust the threshold based on CPU clock frequency, but you never know.

> At the end, we have to revert your clocksource patches before shipping RHEL9
> beta last week.

Which has the disadvantage of leaving the initial clock-skew issues,
but I do understand that introducing one problem even while fixing
another one still counts as a regression.

							Thanx, Paul