net/http: TestCancelRequestWhenSharingConnection can cause port exhaustion, leading to frequent test failures with "Only one usage of each socket address" on windows-amd64-longtest

[bcmills]

What version of Go are you using?

release-branch.go1.15

Does this issue reproduce with the latest release?

No.

What did you expect to see?

Tests on the current release branches (release-branch.go1.15 and release-branch.go1.16) consistently passing, especially for all first class ports.

What did you see instead?

Tests on the windows-amd64-longtest builder started failing intermittently in March, got better for a while in May, and are now failing in net/http on just about every run (https://build.golang.org/?repo=&branch=release-branch.go1.15#short):

The failure mode consists of many Only one usage of each socket address errors (https://build.golang.org/log/30a479e652cc31796c8d900ab0494ffff5765266), sometimes accompanied by a test deadlock or timeout (https://build.golang.org/log/3ff9e1b80ee03390ebb171468f0fb00ceb23795d).

These failures result in TryBot / SlowBot failures on the release branch for unrelated backports (https://golang.org/cl/332330). Since the builders are consistently broken, there is a substantial risk that these failures will cause some other regression on this branch to go unnoticed.

CC @neild @golang/release

[cagedmantis]

I believe the work to try and resolve this issue has been previously tracked in #45358.

[bcmills]

I believe the work to try and resolve this issue has been previously tracked in #45358.

Yeah, it's closely related.

I actually tried to find that issue before filing this one, but it didn't turn up in my search results — the problem described in the issue title seems to have been addressed, and the string in the current error message isn't mentioned in that issue at all. 😕

[dmitshur]

This flaky failure continues to be an issue but we're not going to block Go 1.15.14 on it. Moving to next milestone.

[cagedmantis]

As with the previous milestone: This flaky failure continues to be an issue but we're not going to block Go 1.15.15 on it. Moving to next milestone.

[gopherbot]

Change https://golang.org/cl/339593 mentions this issue: net/http: set socket linger time to 0 in TestCancelRequestWhenSharingConnection

[gopherbot]

Change https://golang.org/cl/339673 mentions this issue: [release-branch.go1.15] net/http: set socket linger time to 0 in TestCancelRequestWhenSharingConnection

[gopherbot]

Change https://golang.org/cl/339594 mentions this issue: net/http: speed up and deflake TestCancelRequestWhenSharingConnection