runtime: TestNetpollDeadlock flakes across the board

[FiloSottile]

Saw it on a freebsd-amd64-12_0 TryBot, but it's been all over the place over the last few days.

--- FAIL: TestNetpollDeadlock (60.09s)
    crash_test.go:95: testprognet NetpollDeadlock exit status: exit status 2
    crash_test.go:450: output does not start with "done\n":
        dialing
        SIGQUIT: quit
        PC=0x45f1ae m=0 sigcode=65537

2019-11-04T17:19:36-1da575a/linux-amd64
2019-11-04T15:18:34-d3660e8/linux-amd64-clang
2019-11-04T10:13:30-1b2ff10/netbsd-amd64-8_0
2019-11-03T04:58:01-32190b0/freebsd-amd64-race
2019-11-03T04:57:50-c6621d9/freebsd-amd64-race
2019-11-03T01:44:46-d2c039f/linux-amd64-nocgo
2019-11-02T21:51:16-d16ec13/linux-ppc64le-power9osu
2019-11-02T21:51:14-a3ffb0d/freebsd-arm64-dmgk
2019-11-02T21:51:14-a3ffb0d/linux-arm64-packet
2019-11-02T19:25:54-afe50c1/linux-amd64
2019-11-02T00:36:59-48a0b97/openbsd-amd64-64
2019-11-01T22:04:05-130f3c0/darwin-amd64-10_14
2019-11-01T20:16:01-c9d89f6/darwin-amd64-race
2019-11-01T14:45:24-c0555a2/freebsd-amd64-12_0
2019-11-01T13:38:23-24e9c53/freebsd-amd64-race
2019-11-01T05:38:51-e96fd13/freebsd-386-11_2
2019-11-01T05:38:51-e96fd13/linux-amd64-race
2019-11-01T01:51:26-efd395f/freebsd-arm64-dmgk
2019-11-01T00:04:51-4a09a9b/linux-amd64-jessie
2019-10-31T20:45:40-d6f2f83/freebsd-386-12_0
2019-10-31T17:21:56-48c0cef/freebsd-386-11_2
2019-10-31T17:09:48-a9b37ae/linux-ppc64le-buildlet

/cc @ianlancetaylor

[FiloSottile]

Seems to have coincided with this rarer semi-related flake, which I noticed at https://storage.googleapis.com/go-build-log/b7ab0107/linux-amd64_99c6e062.log

runtime: netpollBreak write failed with 9
fatal error: runtime: netpollBreak write failed

2019-11-04T16:32:38-7dcd343/darwin-amd64-nocgo
2019-11-02T21:51:21-177a36a/dragonfly-amd64
2019-11-02T21:51:14-a3ffb0d/darwin-amd64-10_14
2019-11-02T21:51:07-40b7455/darwin-amd64-nocgo
2019-11-02T05:52:33-8de0bb7/netbsd-amd64-8_0
2019-11-01T21:41:41-9bde9b4/darwin-amd64-10_14
2019-11-01T05:38:51-e96fd13/darwin-amd64-10_14

[bradfitz]

Maybe https://go-review.googlesource.com/c/go/+/204800 fixes it.

[ianlancetaylor]

The netpollBreak write failed with 9 failures should be fixed by https://golang.org/cl/204937, which is awaiting review.

I'm less sure about the TestNetpollDeadlock failures. Let's see if they still happen now that CL 204800 is in.

[FiloSottile]

(FYI you did not assign a reviewer on CL 204937.)

[ianlancetaylor]

It wasn't in anyone's particular speciality. I left it for the default approach.

[gopherbot]

Change https://golang.org/cl/204937 mentions this issue: runtime: don't return from netpollGenericInit until init is complete

[danscales]

I reproduced this on linux-amd64 at 1da575a (the latest test failure above) by repeating this 100 times:

go test runtime -test.run TestNetpollDeadlock -test.count=100

(usually happened within about 20 times). Using this reproduction case, I showed that it was also fixed by Ian's commit d80ab3e ("runtime: wake netpoller when dropping P, don't sleep too long in sysmon"). The failure is reproducible in commit immediate before Ian's, and not reproducible with Ian's commit.

So, I think we can close this bug as well.