[release-branch.go1.25] net/http: avoid connCount underflow race

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.

For #61474
Fixes #75539

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>
(cherry picked from commit 3203a5da290753e5c7aceb12f41f06b272356bd0)
Reviewed-on: https://go-review.googlesource.com/c/go/+/705376
Reviewed-by: Cherry Mui <cherryyz@google.com>

diff --git a/src/net/http/transport.go b/src/net/http/transport.go
index 07b3a9e1e72ba60122d83d48558d91fb464b47b7..2778db332e4319ebea85ae5448df4e4a6b411cc9 100644 (file)
--- a/src/net/http/transport.go
+++ b/src/net/http/transport.go
@@ -1372,7 +1372,10 @@ func (w *wantConn) cancel(t *Transport) {
        w.done = true
        w.mu.Unlock()
 
-       if pc != nil {
+       // HTTP/2 connections (pc.alt != nil) aren't removed from the idle pool on use,
+       // and should not be added back here. If the pconn isn't in the idle pool,
+       // it's because we removed it due to an error.
+       if pc != nil && pc.alt == nil {
                t.putOrCloseIdleConn(pc)
        }
 }

diff --git a/src/net/http/transport_test.go b/src/net/http/transport_test.go
index 9762f058867dc17a6ece731a6d7f913af3acf98b..28ad3eb6feb54fe20686ba1f649f385981a147e2 100644 (file)
--- a/src/net/http/transport_test.go
+++ b/src/net/http/transport_test.go
@@ -7559,3 +7559,35 @@ func TestTransportServerProtocols(t *testing.T) {
                })
        }
 }
+
+func TestIssue61474(t *testing.T) {
+       run(t, testIssue61474, []testMode{http2Mode})
+}
+func testIssue61474(t *testing.T, mode testMode) {
+       if testing.Short() {
+               return
+       }
+
+       // This test reliably exercises the condition causing #61474,
+       // but requires many iterations to do so.
+       // Keep the test around for now, but don't run it by default.
+       t.Skip("test is too large")
+
+       cst := newClientServerTest(t, mode, HandlerFunc(func(rw ResponseWriter, req *Request) {
+       }), func(tr *Transport) {
+               tr.MaxConnsPerHost = 1
+       })
+       var wg sync.WaitGroup
+       defer wg.Wait()
+       for range 100000 {
+               wg.Go(func() {
+                       ctx, cancel := context.WithTimeout(t.Context(), 1*time.Millisecond)
+                       defer cancel()
+                       req, _ := NewRequestWithContext(ctx, "GET", cst.ts.URL, nil)
+                       resp, err := cst.c.Do(req)
+                       if err == nil {
+                               resp.Body.Close()
+                       }
+               })
+       }
+}