time: TestTimerModifiedEarlier failures with "timer took too long" on plan9-arm

[bcmills]

#!watchflakes
post <- builder ~ `^plan9` && pkg == "time" && test == "TestTimerModifiedEarlier" && `timer took too long`

--- FAIL: TestTimerModifiedEarlier (10.08s)
    sleep_test.go:552: timer took too long (10.006295837s)
    sleep_test.go:561: 1 failures
FAIL
FAIL	time	13.241s

greplogs --dashboard -md -l -e 'FAIL: TestTimerModifiedEarlier'

2022-01-04T21:10:15-1c8f9d2/plan9-arm
2021-12-02T23:02:33-8835343/plan9-arm
2021-11-05T21:34:10-bb53fd7/plan9-arm

The failing test is the regression test for #47329.

CC @millerresearch @ianlancetaylor

[gopherbot]

Change https://golang.org/cl/375934 mentions this issue: time: skip TestTimerModifiedEarlier on plan9/arm

[bcmills]

I wonder if this may be related to the race fixed in https://go.dev/cl/470215.

[millerresearch]

I wonder if this may be related to the race fixed in https://go.dev/cl/470215.

It doesn't look related to me. A possible clue is that the delayed read of the reset-into-the-past timer always occurs after just over 10 seconds, ie when the other (10 second) deadline timer fires. Hypothesis: the select for some reason is not detecting that the timer has already expired, so the goroutine is descheduled until the deadine timer fires. The deadline timer reschedules the goroutine which then finds both timer channels ready, and inputs from the first one giving the observed error.

But what's special about plan9-arm builders here? I don't think the timer or scheduling logic is different from other platforms. But they are Raspberry Pi boards with small RAM and SD cards as swap device, so potentially subject to extremely long delays in wall-clock time.

[bcmills]

Hypothesis: the select for some reason is not detecting that the timer has already expired, so the goroutine is descheduled until the deadine timer fires.

That sounds very plausible to me.

It's been quite a while since we started skipping this test. Should we remove the Skip and see if it's still happening?

[millerresearch]

Should we remove the Skip and see if it's still happening?

Yes, it's worth a try.