Obviously I'm missing something but I really don't see how that is possible. I don't see any code path at all that leads to an exit status of 2.

@ianlancetaylor seems like it - could it be that it's the test binary that returns the exit status 2 somehow? Our usage seems to be simple, but if something weird happened there it could be possible that the exit code comes from somewhere other than the test we run via -test.run.

@aclements Can you let us know how often this test failure occurs? Or tell us how we can run your tool to find that out ourselves? Thanks.

Oh, I forgot he did check in his tool: x/build/cmd/fetchlogs

They're all technically checked in, but some are in personal repos because they're a mess (go get github.com/aclements/go-misc/greplogs; go get github.com/aclements/go-misc/findflakes)

$ fetchlogs
$ greplogs -dashboard -E 'pipe_test.*exit status 2' -md -l

2016-05-11T16:48:46-80423f1/darwin-amd64-10_9
2016-05-11T17:07:24-20e362d/darwin-amd64-10_9
2016-05-12T16:48:09-a8a2b38/darwin-amd64-10_9
2016-05-13T17:44:48-0cc710d/darwin-amd64-10_9
2016-05-14T13:49:51-d8b08c3/darwin-amd64-10_9
2016-05-14T13:49:57-6181db5/darwin-amd64-10_9
2016-05-14T17:27:44-5f83312/darwin-amd64-10_9
2016-05-14T17:30:16-2699da1/darwin-amd64-10_9
2016-05-14T23:06:17-aba7b3e/darwin-amd64-10_9
2016-05-15T06:24:02-b4bf066/darwin-amd64-10_9
2016-05-15T20:56:39-a101b85/darwin-amd64-10_9
2016-08-22T22:59:20-550caa1/darwin-amd64-10_8
2016-09-08T21:36:37-58de268/darwin-amd64-10_8
2016-09-10T17:04:21-3d562de/darwin-amd64-10_8
2017-01-24T19:56:23-98842ca/darwin-amd64-10_11
2017-03-31T14:05:12-4a11404/darwin-amd64-10_10
2017-05-10T14:47:41-41d0bbd/darwin-amd64-10_10
2017-07-10T20:09:00-6f83b75/darwin-amd64-10_12
2017-07-11T14:01:04-58ae050/darwin-amd64-10_12
2017-08-30T19:23:48-81b9d73/darwin-amd64-10_11
2017-10-18T19:57:28-9a84e52/darwin-amd64-10_12

$ greplogs -dashboard -E 'pipe_test.*exit status 2' -l | findflakes -paths
First observed 98842ca 24 Jan 19:56 2017 (3839 commits ago)
Last observed  9a84e52 18 Oct 19:57 2017 (656 commits ago)
36% chance failure is still happening
0.16% failure probability (6 of 3184 commits)
Likely culprits:
    0% 98842ca net/http: don't send body on redirects for 301, 302, 303 when GetBody is set
    0% 314180e net/http: fix a nit
    0% aad06da cmd/link: mark DWARF function symbols as reachable
    0% be9dcfe doc: mention testing.MainStart signature change
    0% a96e117 runtime: amd64, use 4-byte ops for memmove of 4 bytes
    0% 4cce27a cmd/compile: fix constant propagation through s390x MOVDNE instructions
    0% 1be957d misc/cgo/test: pass current environment to syscall.Exec
    0% ec654e2 misc/cgo/test: fix test when using GCC 7
    0% 256a605 cmd/compile: don't use nilcheck information until the next block
    0% e8d5989 cmd/compile: fix compilebench -alloc
Past failures:
  550caa1 22 Aug 22:59 2016 to 3d562de 10 Sep 17:04 2016
    0.65% failure probability (3 of 310 commits)
  80423f1 11 May 16:48 2016 to a101b85 15 May 20:56 2016
    25% failure probability (11 of 41 commits)

Thanks, @aclements .

I do now see a code path that can lead to an exit status of 2. The failure cases are when signal.Notify is not called. The program will get a SIGPIPE. The signal handler will catch the signal and return. The Write will see an EPIPE error, call os.epipecheck, then os.sigpipe (in runtime/signal_unix.go), then dieFromSignal. If dieFromSignal fails to actually die from the signal, it will call exit(2), which is what we are seeing.

Darwin is unusual in that the runtime.raise function used by dieFromSignal does not simply send the signal to the current thread, but instead sends it to the entire process using the kill system call. So Darwin is more likely to encounter an unexpected delay in signal delivery here, and that may be what is causing these very occasional failures. Also the Darwin osyield function just sleeps for 1 microsecond, rather than calling sched_yield which generally lets all other threads run before the current thread resumes.

(Darwin does actually have pthread_kill and sched_yield functions. I don't know why we don't use them.)

See https://golang.org/cl/37890 for some of why it doesn't use sched_yield. Apparently we're waiting on loop preemption (sigh).

Change https://golang.org/cl/82375 mentions this issue: runtime: sleep longer in dieFromSignal on Darwin