pc99: SCHED0011 sometimes times out

[lsf37]

In this run: https://github.com/seL4/util_libs/actions/runs/4920068732/jobs/8789972531?pr=156#step:4:3957 we're getting:

Tue, 09 May 2023 02:21:49 GMT  Test SCHED0010 passed
Tue, 09 May 2023 02:21:49 GMT  </testcase>
Tue, 09 May 2023 02:21:49 GMT  <testcase classname="sel4test" name="SCHED0011">
Tue, 09 May 2023 02:21:49 GMT  Running test SCHED0011 (Test scheduler accuracy)
Tue, 09 May 2023 02:36:11 GMT
Tue, 09 May 2023 02:36:11 GMT  [[Timeout]]
Tue, 09 May 2023 02:36:11 GMT  None
Tue, 09 May 2023 02:36:12 GMT  
Tue, 09 May 2023 02:36:12 GMT  console_run returned -1
Tue, 09 May 2023 02:36:12 GMT  Shutting down haswell4

Note the 15min passing between start of SCHED011 and the timeout.

The failing config was PC99_debug_MCS_clang_32

[wom-bat]

Seem,s to happen randomly. Today https://github.com/seL4/seL4/actions/runs/12040941622/job/33573351124#step:4:23728 on the gcc 32-bit build

[Indanz]

It's x86, you can get an NMI at any time which will take unknown amount of time to be handled by the system firmware. Still, 2 ms deviation seems like a lot.

SCHED0007 and SCHED0008 don't seem to clean up all spawned threads, not sure what effect that has on further tests.

[Indanz]

This might have been fixed with my sel4bench bugfixes recently, but either way haven't seen this for a while. Should we close this?

[lsf37]

I have also not seen this in a very long time. Would it be possible to run a stress test to confirm? (E.g. that one test a few hundred times -- that should exceed the frequency of previous occurrences by a lot).

[Indanz]

I have also not seen this in a very long time. Would it be possible to run a stress test to confirm? (E.g. that one test a few hundred times -- that should exceed the frequency of previous occurrences by a lot).

I'll give it a go. My stress tests usually run for at least an hour, or 24 hours, depending on what they test. Even to just filter out you need to run things for at least a thousand times usually.

That said, if it was a corner case bug or race in the timer driver, then repeating the same test probably won't trigger it again.

(Remind me if I forget to do this.)

[lsf37]

Agreed, it's not going to be certain, but it should be better than "we haven't seen this in a while" :-)