runtime: negative duration panic in timeHistogram.record

[mknyszek]

It's possible sometimes for timeHistogram to get a negative duration either due to a platform bug or surprising time semantics on a given platform.

Notably, this could happen when running under gVisor, where, for efficiency, gVisor provides its own VDSO for acquiring the current monotonic time, but falls back on a different clock if the system clock is wildly in error (https://github.com/google/gvisor/blob/master/vdso/vdso_time.cc#L149).

The result is a panic that can't be worked around. For resiliency's sake, timeHistogram should just accept negative durations and record them.

[gopherbot]

Change https://golang.org/cl/279468 mentions this issue: runtime: shift timeHistogram buckets and allow negative durations

[dmitshur]

One test failure where I saw this recently is openbsd-arm64-jsing at 740851bacafd8e47b9a6ce0cd8fa8e05506a7382.

[prattmic]

Ah, nice catch. There are several of these:

2020-12-18T00:16:17-740851b/openbsd-arm64-jsing
2020-12-07T17:19:41-7f9a2bc/openbsd-arm64-jsing
2020-12-03T21:23:16-b78b427/openbsd-arm64-jsing
2020-11-30T03:18:36-e5da18d/openbsd-arm64-jsing
2020-11-03T00:56:12-45205bc/netbsd-arm64-bsiegert