net: TestCloseWrite failures with "i/o timeout" on darwin/arm64

[bcmills]

#!watchflakes
post <- goos ~ `darwin|ios` && goarch == "arm64" && pkg == "net" && test == "TestCloseWrite"

greplogs --dashboard -md -l -e '(?m)FAIL: TestCloseWrite/.*\n(\s+.*\n)*.*i/o timeout'

2021-11-03T21:45:33-d3f5dd5/darwin-arm64-11_0-toothrot
2020-10-20T17:52:41-90c924f/ios-arm64-corellium
2020-09-28T21:30:01-af18bce/linux-mips64le-mengzhuo
2020-09-07T02:46:24-1a119ed/darwin-arm64-corellium
2020-08-26T16:15:32-6382809/darwin-arm64-corellium
2020-08-18T15:50:52-b58d297/darwin-arm64-corellium
2020-05-01T19:55:48-30f8074/darwin-arm64-corellium
2020-04-13T21:56:15-1b15c7f/darwin-arm64-corellium

See previously #34837.

[bcmills]

Still ongoing (but rare) on darwin/arm64.

greplogs -l -e '(?m)FAIL: TestCloseWrite/.*\n(\s+.*\n)*.*i/o timeout' --since=2021-11-04
2022-06-14T21:55:52-cb9bf93/darwin-arm64-11
2021-11-05T00:52:09-bd580a0/darwin-arm64-11_0-toothrot

(attn @neild, @ianlancetaylor)

[neild]

I wonder if this is a variation on #37795?

[neild]

The mips64le failure above has "too many open files" in the log, and is likely unrelated. I'll bet this is the same macOS bug as #37795. If it is, a similar workaround might work--add a small delay so that we don't close (or CloseWrite) a socket while reading from it.

[bcmills]

Seems worth a shot, at least! (We can always try it, and reopen the issue if it turns out not to be the right workaround.)

[gopherbot]

Change https://go.dev/cl/414534 mentions this issue: net: avoid darwin/arm64 platform bug in TestCloseWrite

[bcmills]

Sent https://go.dev/cl/414534 with a port of the workaround from #37795, prompted by another failure on the dashboard.

greplogs -l -e '(?m)FAIL: TestCloseWrite/.*\n(\s+.*\n)*.*i/o timeout' --since=2022-06-15
2022-06-26T00:21:33-666d736/darwin-arm64-12

[heschi]

Unless I'm misunderstanding something, this recurred on 6/29: 2022-06-29T22:28:30-b2cc0fe/darwin-arm64-11