Remove a race condition in counting the number of connections per host,
which can cause a connCount underflow and a panic.
The race occurs when:
- A RoundTrip call attempts to use a HTTP/2 roundtripper (pconn.alt != nil)
and receives an isNoCachedConn error. The call removes the pconn from
the idle conn pool and decrements the connCount for its host.
- A second RoundTrip call on the same pconn succeeds,
and delivers the pconn to a third RoundTrip waiting for a conn.
- The third RoundTrip receives the pconn at the same moment its request
context is canceled. It places the pconn back into the idle conn pool.
At this time, the connCount is incorrect, because the conn returned to
the idle pool is not matched by an increment in the connCount.
Fix this by not adding HTTP/2 pconns back to the idle pool in
wantConn.cancel.
Fixes #61474
Change-Id: I104d6cf85a54d0382eebf3fcf5dda99c69a7c3f6
Reviewed-on: https://go-review.googlesource.com/c/go/+/703936
Auto-Submit: Damien Neil <dneil@google.com> Reviewed-by: Nicholas Husin <husin@google.com> Reviewed-by: Nicholas Husin <nsh@golang.org>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>