net: TestConnClose failure with "connect: connection refused" on netbsd-arm-bsiegert

[bcmills]

--- FAIL: TestConnClose (0.00s)
    --- FAIL: TestConnClose/unix (0.37s)
        net_test.go:194: dial unix /var/gobuilder/buildlet/tmp/2224876548/sock: connect: connection refused
FAIL
FAIL	net	26.230s

greplogs --dashboard -md -l -e 'FAIL: TestConnClose .*(?:\n .*)* connect: connection refused'

2022-04-15T01:08:38-5a4f0b6/netbsd-arm-bsiegert

[bcmills]

@bsiegert, @coypoop: could this be a bug in the NetBSD implementation of Unix domain sockets?

This test is pretty simple — I don't see any way the test process could not be listening on the socket at that point.

[riastradh]

This is pretty weird. None of this code is obviously machine-dependent, and we haven't been able to reproduce this in thousands of trials (although I haven't tried a 32-bit arm system). Has this happened more than once?

In case this happens only under parallel tests that use the same file name sock -- I assume the temporary directory selection here is robust to collisions in the random directory name generation, right?

[bcmills]

Has this happened more than once?

Not that I can find.

In case this happens only under parallel tests that use the same file name sock -- I assume the temporary directory selection here is robust to collisions in the random directory name generation, right?

I sure hope so — I just reworked it for exactly that purpose in December! 😅
(See https://go.dev/cl/370694.)

But there may still be a bug remaining there somewhere — I haven't been able to track down the failure in #51441 yet.