runtime: TestNetpollBreak, TestNonblockingPipe, TestSetNonblock failing on aix-ppc64 and flaky on netbsd-arm

[bcmills]

Now that the aix-ppc64 builder is back (#35032), it's showing multiple failures in the runtime tests (https://build.golang.org/log/289cf45ed86d9587769512a5b3df6e0d17be685e):

--- FAIL: TestNetpollBreak (129.61s)
    proc_test.go:1020: netpollBreak did not interrupt netpoll: slept for: 2m9.603103s
--- FAIL: TestNonblockingPipe (0.00s)
    nbpipe_test.go:30: fcntl(reader, F_GETFL) failed: invalid argument
    nbpipe_test.go:31: fcntl(reader, F_GETFD) failed: invalid argument
    nbpipe_test.go:32: fcntl(writer, F_GETFL) failed: invalid argument
    nbpipe_test.go:33: fcntl(writer, F_GETFD) failed: invalid argument
--- FAIL: TestSetNonblock (0.00s)
    nbpipe_test.go:86: fcntl(reader, F_GETFL) failed: invalid argument
    nbpipe_test.go:87: fcntl(writer, F_GETFL) failed: invalid argument
    nbpipe_test.go:91: fcntl(reader, F_GETFD) failed: invalid argument
    nbpipe_test.go:92: fcntl(writer, F_GETFD) failed: invalid argument
FAIL
FAIL	runtime	163.667s

From the failures, this seems to be related to CL 171824, although (because of #35032) we're missing a big chunk of test results preceding that change.

The failure mode for TestNetpollBreak has the same symptom as #35074, but it's not clear to me whether the underlying cause is related.

CC @ianlancetaylor @mknyszek @Helflym @trex58

[Helflym]

TestSetNonblock and TestNonblockingPipe are failing because syscall.Syscall isn't available on AIX. I've made the patch but I've a problem with my google account. I'll push it asap.
TestNetpollBreak is a bit more complicated. It seems to be related to goto retry at line 176. But I've to be sure we can remove it without triggering any problems.

[ianlancetaylor]

For TestNetpollBreak the timeout value should be greater than 0, so that goto should not be executed. And that case should only happen if the program receives a signal, which shouldn't be happening here. So it's not clear to me how that goto could be the problem.

[Helflym]

My bad, that was the wrong goto ! This one is problematic.

[ianlancetaylor]

Ah, yes, that does look somewhat wrong. Maybe we should at least return if some byte was read from rdwake.

[bcmills]

A similar failure on netbsd-arm-bsiegert: https://build.golang.org/log/d1b5ba3784cadac0bb5ea659876e53ec7cf0bac8

##### GOMAXPROCS=2 runtime -cpu=1,2,4 -quick
--- FAIL: TestNetpollBreak (10.02s)
    proc_test.go:1020: netpollBreak did not interrupt netpoll: slept for: 10.01633375s
FAIL
FAIL	runtime	196.708s

[bcmills]

Here's a different failure mode from TestNetpollBreak — an apparent deadlock on aix-ppc64:
https://build.golang.org/log/ffe49c47ca85f6dcf78d77483de389d3f6331aff

[ianlancetaylor]

I'm pretty sure the netbsd-arm-bsiegert flake was fixed by the ubiquitous https://golang.org/cl/204800.

[ianlancetaylor]

The AIX problems were fixed by https://golang.org/cl/204059 and https://golang.org/cl/204318.