fix race condition on concurrent use of Transport.Dial and Close (#4904)

transport.go

@@ -236,19 +236,13 @@ func (t *Transport) DialEarly(ctx context.Context, addr net.Addr, tlsConf *tls.C
 }
 
 func (t *Transport) dial(ctx context.Context, addr net.Addr, host string, tlsConf *tls.Config, conf *Config, use0RTT bool) (EarlyConnection, error) {
-	t.mutex.Lock()
-	if t.closeErr != nil {
-		t.mutex.Unlock()
-		return nil, t.closeErr
+	if err := t.init(t.isSingleUse); err != nil {
+		return nil, err
 	}
-	t.mutex.Unlock()
 	if err := validateConfig(conf); err != nil {
 		return nil, err
 	}
 	conf = populateConfig(conf)
-	if err := t.init(t.isSingleUse); err != nil {
-		return nil, err
-	}
 	tlsConf = tlsConf.Clone()
 	setTLSConfigServerName(tlsConf, addr, host)
 	return t.doDial(ctx,
@@ -283,6 +277,13 @@ func (t *Transport) doDial(
 
 	tracingID := nextConnTracingID()
 	ctx = context.WithValue(ctx, ConnectionTracingKey, tracingID)
+
+	t.mutex.Lock()
+	if t.closeErr != nil {
+		t.mutex.Unlock()
+		return nil, t.closeErr
+	}
+
 	var tracer *logging.ConnectionTracer
 	if config.Tracer != nil {
 		tracer = config.Tracer(ctx, protocol.PerspectiveClient, destConnID)
@@ -312,6 +313,7 @@ func (t *Transport) doDial(
 		version,
 	)
 	t.handlerMap.Add(srcConnID, conn)
+	t.mutex.Unlock()
 
 	// The error channel needs to be buffered, as the run loop will continue running
 	// after doDial returns (if the handshake is successful).
@@ -452,6 +454,9 @@ func (t *Transport) runSendQueue() {
 // If any listener was started, it will be closed as well.
 // It is invalid to start new listeners or connections after that.
 func (t *Transport) Close() error {
+	// avoid race condition if the transport is currently being initialized
+	t.init(false)
+
 	t.close(nil)
 	if t.createdConn {
 		if err := t.Conn.Close(); err != nil {

transport_test.go

@@ -120,8 +120,6 @@ func TestTransportPacketHandling(t *testing.T) {
 }
 
 func TestTransportAndListenerConcurrentClose(t *testing.T) {
-	// try 10 times to trigger race conditions
-	for i := 0; i < 10; i++ {
 	tr := &Transport{Conn: newUPDConnLocalhost(t)}
 	ln, err := tr.Listen(&tls.Config{}, nil)
 	require.NoError(t, err)
@@ -136,6 +134,26 @@ func TestTransportAndListenerConcurrentClose(t *testing.T) {
 		t.Fatal("timeout")
 	}
 }
+
+func TestTransportAndDialConcurrentClose(t *testing.T) {
+	server := newUPDConnLocalhost(t)
+
+	tr := &Transport{Conn: newUPDConnLocalhost(t)}
+	// close transport and dial concurrently
+	errChan := make(chan error, 1)
+	go func() { errChan <- tr.Close() }()
+	ctx, cancel := context.WithTimeout(context.Background(), 500*time.Millisecond)
+	defer cancel()
+	_, err := tr.Dial(ctx, server.LocalAddr(), &tls.Config{}, nil)
+	require.Error(t, err)
+	require.ErrorIs(t, err, ErrTransportClosed)
+	require.NotErrorIs(t, err, context.DeadlineExceeded)
+
+	select {
+	case <-errChan:
+	case <-time.After(time.Second):
+		t.Fatal("timeout")
+	}
 }
 
 func TestTransportErrFromConn(t *testing.T) {