net/http: TestCancelRequestWhenSharingConnection failures on longtest builders

[bcmills]

2020-12-02T14:20:12-0433845/linux-amd64-longtest
2020-12-01T21:42:10-7fca39a/windows-amd64-longtest
2020-12-01T20:42:53-20e2518/linux-amd64-longtest
2020-12-01T19:47:12-50b16f9/linux-amd64-longtest

Marking as release-blocker because this test was introduced Dec. 1 (in CL 257818).

CC @fraenkel @neild @bradfitz @golang/release

[bcmills]

There is also what appears to be a deadlock in this test:
2020-12-01T21:41:39-7430266/windows-amd64-longtest

[fraenkel]

This fix is incorrect and should be reverted.

[fraenkel]

Sorry, the fix is mostly correct. Confusing myself on the issue.
I need to see why the test fails sometimes which still means there is more to this.

[fraenkel]

CL 257818 did fix the case where:
Req 1 does a GET, the readLoop puts the persistconn back
Req 2 gets the pc, and has its context closed which drives cancelRequest -> pc.closech prior to Req 1 leaving roundtrip.

The new issue is similar but makes me believe there is a bigger issue afoot.
The first request starts to cancel (roundtrip), but puts the pc back into the pool (readLoop).
A second request picks up the pc.
The cancel completes, and the second request is cancelled.

[fraenkel]

There are multiple goroutines that fiddle with the cancelFn but most are blindly setting it rather than looking to see if its already been "cancelled".

[gopherbot]

Change https://golang.org/cl/274973 mentions this issue: net/http: add connections back that haven't been cancelled