Re: [RFC PATCH] clocksource: increase watchdog retries

On Mon, Nov 01, 2021 at 11:28:03AM +0100, Jiri Wiesner wrote:

On Thu, Oct 28, 2021 at 11:42:09AM -0700, Paul E. McKenney wrote:

quoted

On Thu, Oct 28, 2021 at 06:20:25PM +0200, Jiri Wiesner wrote:

quoted

On Wed, Oct 27, 2021 at 02:38:29PM -0700, Paul E. McKenney wrote:

quoted

I had something like this pending, but people came up with other workloads
that resulted in repeated delays.  In those cases, it does not make sense
to ever mark the affected clocksource unstable.  This led me to the patch
shown below, which splats after about 100 consecutive long-delay retries,
but which avoids marking the clocksource unstable.  This is queued on -rcu.

Does this work for you?

commit 9ec2a03bbf4bee3d9fbc02a402dee36efafc5a2d
Author: Paul E. McKenney [off-list ref]
Date:   Thu May 27 11:03:28 2021 -0700

    clocksource: Forgive repeated long-latency watchdog clocksource reads

Yes, it does. I have done 100 reboots of the testing machine (running
5.15-rc5 with the above patch applied) and TSC was stable every time. I
am going to start a longer test of 300 reboots for good measure and
report back next week. J.

Very good, and thank you for giving it a go!

Thank you for the fix! It resolves several strange results we got in our performance testing.

quoted

If it passes the upcoming tests

I have done 300 reboots of the testing machine. Again, TSC was stable every time.

quoted

may I have your Tested-by?

Absolutely:
Tested-by: Jiri Wiesner <redacted>

Applied, thank you!

							Thanx, Paul