CC @neild @odeke-em @ianlancetaylor

I spent some time digging into this, somewhat out of curiosity but also to be sure it wasn't related to the fix for #42156 since the timing was suspect. It doesn't seem to be related to #42156 since http.Server.Shutdown is not involved.

It would be nice to simulate go test -count 100 -run TestNoSniffExpectRequestBody_h2 or so on one of these builders. It's my understanding gomote access is not given out much. If that is the case, would it be appropriate to debug by opening a DO NOT REVIEW/SUBMIT CL and have TryBot target one or more of these builders using TRY?

I'm pretty certain this has been fixed, but I haven't managed to bisect down to the exact point of the fix yet. It reproduces on these builders at the previously-failing commits with -count 100, but not at head.

The timer-related changes in 8fdc79e and b4b0144 before the first failure (7be9158) and in 89a6540 after the last observed failure (676ad56) seemed interesting but I wasn't able to induce any failures locally in the questionable commit range to dig deeper.

Bisection results point at this being fixed by e02ab89:

https://go.googlesource.com/go/+/e02ab89eb8994fa6f2dfa2924cdadb097633fcc1 (passes)
https://go.googlesource.com/go/+/a313eec3869c609c0da2402a4cac2d32113d599c (fails)

The test is flaky. The failures take ~10s to run, which points at the ExpectContinueTimeout in the test expiring for some reason. I'm not sure at the moment why the test is failing or why that CL seems to fix it.

\cc @ianlancetaylor

https://golang.org/cl/266304 shouldn't affect any correct code. It's a minor optimization which will cause the runtime poller to be woken up less often to run timers. It will only affect programs that set an I/O deadline or a time.Timer, and then change the deadline or the Timer to run earlier than when it was created. Programs that do that will see slightly different, hopefully more efficient, scheduling behavior.

It's really puzzling that it fails before CL 266304 and passes after it. If anything I would expect CL 266304 to wake up fewer goroutines. So why would waking up more goroutines cause a timeout to expire?

I do not know what is happening here.

It will only affect programs that set an I/O deadline or a time.Timer, and then change the deadline or the Timer to run earlier than when it was created.

It's worth nothing that is what the h2 bits do in the Expect: 100-continue case, using AfterFunc with a "huge duration" here:

And then using Reset to have the timer fire earlier:

Well, at least that explains why the CL has some effect, though I still can't explain why it has the effect that it does.

Actually, something else I wondered about when digging into this is the use of Stop and Reset together. Expanding the code around the Reset:

if s.timer.Stop() {
	s.timer.Reset(s.delay)
}

Since this timer was created with AfterFunc, is it questionable to use Stop and Reset in this way?

That usage is fine. The Stop method will return true if the function has not yet run. After calling Stop, it's fine to call Reset to set a new time.

This test stops passing completely when the change in CL 253259 is bundled into net/http.
I filed #42498 about that, it may or may not be related to this issue.

I can explain how the error occurs, but it doesn't explain why.
The test has a client POST some data with a 10s delay on sending the body via an Expect: 100-continue. The server sends back a 403 (Unauthorized) once it receives the headers. The client should receive the 403 and cancel the timer for the body write.

The body is being written after 10s but the timer should have been cancelled.

https://go-review.googlesource.com/c/net/+/269058/ should fix this as well

@bcmills Is this still an issue now that https://golang.org/cl/269058 is merged?

I don't see any occurrences of this test failure since October 30, 2020, so closing.