runtime: netpoll failed on darwin

[bcmills]

#!watchflakes
post <- goos == "darwin" && log ~ `^fatal error: runtime: netpoll failed`

(Pulled out from #34988 (comment); CC @ianlancetaylor @golang/darwin)

[gopherbot]

Found new dashboard test flakes for:

#!watchflakes
post <- goos == "darwin" && log ~ `^fatal error: runtime: netpoll failed`

2023-04-04 16:58 darwin-amd64-nocgo go@3991f6c4 runtime/pprof.TestCPUProfileWithFork (log)

runtime: kevent on fd 3 failed with 60
fatal error: runtime: netpoll failed

runtime stack:
runtime.throw({0x11af512?, 0x0?})
	/tmp/buildlet/go/src/runtime/panic.go:1047 +0x5d fp=0x7ff7bfefec68 sp=0x7ff7bfefec38 pc=0x1034afd
runtime.netpoll(0x1054c74?)
	/tmp/buildlet/go/src/runtime/netpoll_kqueue.go:134 +0x38e fp=0x7ff7bfeff4f8 sp=0x7ff7bfefec68 pc=0x103160e
runtime.findRunnable()
	/tmp/buildlet/go/src/runtime/proc.go:2759 +0x2af fp=0x7ff7bfeff600 sp=0x7ff7bfeff4f8 pc=0x103c64f
...
	/tmp/buildlet/go/src/os/exec.go:109 +0x5a fp=0xc000173d28 sp=0xc000173ce0 pc=0x108a31a
os/exec.(*Cmd).Start(0xc00012a2c0)
	/tmp/buildlet/go/src/os/exec/exec.go:693 +0x5ee fp=0xc000173ea0 sp=0xc000173d28 pc=0x10e256e
os/exec.(*Cmd).Run(0x11435a0?)
	/tmp/buildlet/go/src/os/exec/exec.go:587 +0x1e fp=0xc000173ec0 sp=0xc000173ea0 pc=0x10e1f3e
os/exec.(*Cmd).CombinedOutput(0xc00012a2c0)
	/tmp/buildlet/go/src/os/exec/exec.go:1005 +0xa8 fp=0xc000173ee8 sp=0xc000173ec0 pc=0x10e3da8
runtime/pprof.TestCPUProfileWithFork(0xc000083380)
	/tmp/buildlet/go/src/runtime/pprof/pprof_test.go:668 +0x194 fp=0xc000173f70 sp=0xc000173ee8 pc=0x11530b4
testing.tRunner(0xc000083380, 0x11bbfb0)

— watchflakes

[ianlancetaylor]

> runtime: kevent on fd 3 failed with 60
> fatal error: runtime: netpoll failed

Error 60 is ETIMEDOUT. This shouldn't happen, and is not listed on the man page. When a call to kevent reaches the timeout, it is documented to return 0.

It would be easy for us to ignore a ETIMEDOUT error but it shouldn't be necessary.

[mknyszek]

In triage, we think this is either a bug in Darwin or a bug in the Darwin documentation. @prattmic thinks maybe an error is bubbling up from somewhere deeper, in which case maybe we just ignore it like @ianlancetaylor suggested.

[mknyszek]

We agree we will wait until it fails again to make a decision.