net/http: TestClientTimeoutCancel is flaky

[bcmills]

From the linux-ppc64le-power9osu (https://build.golang.org/log/bc146619a7844f5eadaff5632418865af83dacf4):

--- FAIL: TestClientTimeoutCancel (0.00s)
    client_test.go:1353: error = read tcp 127.0.0.1:42466->127.0.0.1:40975: use of closed network connection; want errRequestCanceled
FAIL
FAIL	net/http	7.264s

CC @laboger @bradfitz

[ceseo]

I'll take a look.

[ceseo]

I tried a standalone build in that system and can't reproduce. Can't reproduce in any of my POWER9 systems as well. Looks like it's one of those intermittent failures.

[laboger]

@ceseo I think you need to try and reproduce using GOMAXPROCS=2 since that is the number of processors on the build systems.

[bcmills]

2019-11-03T19:43:59-ab4a71f/windows-amd64-2008

[bcmills]

2019-11-11T12:40:04-f07059d/linux-amd64-jessie
2019-11-08T21:32:23-0bbcce9/windows-amd64-2012
2019-11-08T18:11:01-1fd3f8b/linux-mips-rtrk

[bcmills]

https://storage.googleapis.com/go-build-log/90770a22/windows-amd64-2016_87869979.log

[bcmills]

2020-04-21T00:38:14-12665b9/linux-amd64-longtest
2020-04-17T13:30:49-8ce21fa/linux-386-longtest
2020-03-25T19:14:38-fcb8f83/linux-amd64-longtest
2020-02-27T23:06:33-3061d89/freebsd-arm64-dmgk

[bcmills]

2020-08-24T06:01:43-328214e/linux-amd64-jessie
2020-08-17T21:05:19-ba97be4/linux-amd64-staticlockranking
2020-06-03T14:51:34-5aaeda1/darwin-amd64-nocgo

[bcmills]

Also https://storage.googleapis.com/go-build-log/d1511a2e/android-amd64-emu_dd85a88c.log (a TryBot).

[bcmills]

2020-10-28T13:46:11-a0a4439/windows-amd64-race
2020-10-22T13:49:19-b4c8b67/freebsd-386-11_2
2020-09-16T23:00:02-f5d59d0/linux-amd64-longtest
2020-09-16T03:02:13-eaa97fb/darwin-amd64-10_15

[bcmills]

CC @empijei @neild

[bcmills]

2021-03-19T17:34:25-1c59066/windows-amd64-longtest
2021-03-05T04:30:32-f0b6d37/windows-amd64-longtest
2021-02-22T18:51:18-b2bdadf/freebsd-amd64-11_2

[gopherbot]

Change https://golang.org/cl/303489 mentions this issue: net/http: fix request cancellation race

[neild]

This was really hard to track down.

The problem is that (*Transport).cancelRequest calls a per-request cancellation function, but a racing call to cancelRequest can return before that function has been called.