It was failing with multiple goroutines a few out of every thousand
runs (with errRequestCanceled) because it was using the same
*http.Request for all 5 RoundTrips, but the RoundTrips' goroutines
(notably the readLoop method) were all still running, sharing that
same pointer. Because the response has no body (which is what
TestZeroLengthPostAndResponse tests), the readLoop was marking the
connection as reusable early (before the caller read until the body's
EOF), but the Transport code was clearing the Request's cancelation
func *AFTER* the caller had already received it from RoundTrip. This
let the test continue looping and do the next request with the same
pointer, fetch a connection, and then between getConn and roundTrip
have an invariant violated: the Request's cancelation func was nil,
tripping this check:

        if !pc.t.replaceReqCanceler(req.Request, pc.cancelRequest) {
                pc.t.putIdleConn(pc)
                return nil, errRequestCanceled
        }

The solution is to clear the request cancelation func in the readLoop
goroutine in the no-body case before it's returned to the caller.

This now passes reliably:

$ go test -race -run=TestZeroLengthPostAndResponse -count=3000

I think we've only seen this recently because we now randomize scheduling
of goroutines in race mode (https://golang.org/cl/11795). This race
has existed for a long time but the window was hard to hit.

Change-Id: Idb91c582919f85aef5b9e5ef23706f1ba9126e9a
Reviewed-on: https://go-review.googlesource.com/13070
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>