Re: [PATCH 1/2] clocksource: Avoid accidental unstable marking of clocksources

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

On Fri, Nov 12, 2021 at 10:43:15PM -0500, Waiman Long wrote:
> On 11/12/21 00:44, Feng Tang wrote:
> > On Thu, Nov 11, 2021 at 06:43:11AM -0800, Paul E. McKenney wrote:
> > > On Thu, Nov 11, 2021 at 12:57:03PM +0800, Feng Tang wrote:
> > > > On Wed, Nov 10, 2021 at 05:17:31PM -0500, Waiman Long wrote:
> > > > > Since commit db3a34e17433 ("clocksource: Retry clock read if long delays
> > > > > detected") and commit 2e27e793e280 ("clocksource: Reduce clocksource-skew
> > > > > threshold"), it is found that tsc clocksource fallback to hpet can
> > > > > sometimes happen on both Intel and AMD systems especially when they are
> > > > > running stressful benchmarking workloads. Of the 23 systems tested with
> > > > > a v5.14 kernel, 10 of them have switched to hpet clock source during
> > > > > the test run.
> > > > > 
> > > > > The result of falling back to hpet is a drastic reduction of performance
> > > > > when running benchmarks. For example, the fio performance tests can
> > > > > drop up to 70% whereas the iperf3 performance can drop up to 80%.
> > > > > 
> > > > > 4 hpet fallbacks happened during bootup. They were:
> > > > > 
> > > > >    [    8.749399] clocksource: timekeeping watchdog on CPU13: hpet read-back delay of 263750ns, attempt 4, marking unstable
> > > > >    [   12.044610] clocksource: timekeeping watchdog on CPU19: hpet read-back delay of 186166ns, attempt 4, marking unstable
> > > > >    [   17.336941] clocksource: timekeeping watchdog on CPU28: hpet read-back delay of 182291ns, attempt 4, marking unstable
> > > > >    [   17.518565] clocksource: timekeeping watchdog on CPU34: hpet read-back delay of 252196ns, attempt 4, marking unstable
> > > > > 
> > > > > Other fallbacks happen when the systems were running stressful
> > > > > benchmarks. For example:
> > > > > 
> > > > >    [ 2685.867873] clocksource: timekeeping watchdog on CPU117: hpet read-back delay of 57269ns, attempt 4, marking unstable
> > > > >    [46215.471228] clocksource: timekeeping watchdog on CPU8: hpet read-back delay of 61460ns, attempt 4, marking unstable
> > > > > 
> > > > > Commit 2e27e793e280 ("clocksource: Reduce clocksource-skew threshold"),
> > > > > changed the skew margin from 100us to 50us. I think this is too small
> > > > > and can easily be exceeded when running some stressful workloads on
> > > > > a thermally stressed system.  So it is switched back to 100us. On
> > > > > the other hand, it doesn't look like we need to increase the minimum
> > > > > uncertainty margin. So it is kept the same at 100us too.
> > > > > 
> > > > > Even a maximum skew margin of 100us may be too small in for some systems
> > > > > when booting up especially if those systems are under thermal stress. To
> > > > > eliminate the case that the large skew is due to the system being too
> > > > > busy slowing down the reading of both the watchdog and the clocksource,
> > > > > a final check is done by reading watchdog time again and comparing the
> > > > > consecutive watchdog timing read delay against WATCHDOG_MAX_SKEW/2. If
> > > > > that delay exceeds the limit, we assume that the system is just too
> > > > > busy. A warning will be printed to the console and the watchdog check
> > > > > is then skipped for this round. For example:
> > > > > 
> > > > >    [    8.789316] clocksource: timekeeping watchdog on CPU13: hpet consecutive read-back delay of 174541ns, system too busy
> > > > 
> > > > I think it may be better to add more details about the root cause, other
> > > > than that it looks good to me, as we tested similar patch on our test
> > > > platforms.
> > > > 
> > > > Reviewed-by: Feng Tang <feng.tang@intel.com>
> > > Thank you both!
> > > 
> > > I agree on the bit about root cause.  Would it make sense to compare the
> > > difference between HPET reads 1 and 2 (containing the read of the TSC)
> > > and the difference between HPET reads 2 and 3?  If the 2-1 difference was
> > > no more than (say) 8/7ths of the 3-2 difference, or the 2-1 difference
> > > was no more than (say) 20 microseconds more than the 3-2 difference,
> > > this could be considered a good-as-it-gets read, ending the retry loop.
> > > Then if the 3-1 difference was greater than the default (100 microseconds
> > > in current -rcu), that difference could be substituted for that particular
> > > clocksource watchdog check.  With a console message noting the unusually
> > > high overhead (but not a splat).
> > > 
> > > So if it took 75 microseconds for each HPET read and 1 microsecond for
> > > the TSC read, then 226 microseconds would be substituted for the default
> > > of 100 microseconds for that cycle's skew cutoff.  Unless the previous
> > > skew cutoff was larger, in which case the previous cutoff should be
> > > used instead.  Either way, the current cutoff is recorded for comparison
> > > for the next clocksource watchdog check.
> > > 
> > > If the 3-1 difference was greater than 62.5 milliseconds, a warning should
> > > probably be emitted anyway.
> > I can test the patch with our cases that could reproduce the problem.
> > 
> > > Or did you have something else in mind?
> > I'm not sure the detail in  Waiman's cases, and in our cases (stress-ng)
> > the delay between watchdog's (HPET here) read were not linear, that
> > from debug data, sometimes the 3-2 difference could be bigger or much
> > bigger than the 2-1 difference.
> > 
> > The reason could be the gap between 2 reads depends hugely on the system
> > pressure at that time that 3 HPET read happens. On our test box (a
> > 2-Socket Cascade Lake AP server), the 2-1 and 3-2 difference are stably
> > about 2.5 us,  while under the stress it could be bumped to from 6 us
> > to 2800 us.
> > 
> > So I think checking the 3-2 difference plus increasing the max retries
> > to 10 may be a simple way, if the watchdog read is found to be
> > abnormally long, we skip this round of check.
> 
> On one of the test system, I had measured that normal delay
> (hpet->tsc->hpet) was normally a bit over 2us. It was a bit more than 4us at
> bootup time. However, the same system under stress could have a delay of
> over 200us at bootup time. When I measured the consecutive hpet delay, it
> was about 180us. So hpet read did dominate the total clocksource read delay.

Thank you both for the data!

> I would not suggest increasing the max retries as it may still fail in most
> cases because the system stress will likely not be going away within a short
> time. So we are likely just wasting cpu times. I believe we should just skip
> it if it is the watchdog read that is causing most of the delay.

If anything, adding that extra read would cause me to -reduce- the number
of retries to avoid increasing the per-watchdog overhead.

Thoughts?

							Thanx, Paul