xnet.TCPListener guard racy reading of closed bool (#200)

x/xnet/stack-go.go

@@ -290,9 +290,9 @@ func (u *udppktconn) Write(b []byte) (int, error) {
 }
 func (u *udppktconn) RemoteAddr() net.Addr { return nil }
 
+// tcplistener adapts [tcp.Listener] to [net.Listener].
 type tcplistener struct {
 	l         tcp.Listener
-	closed    bool
 	sleep     lneto.BackoffStrategy
 	localAddr net.Addr
 }
@@ -308,42 +308,26 @@ func (l *tcplistener) Addr() net.Addr {
 
 func (l *tcplistener) Shutdown() { l.Close() }
 
+// Accept blocks until the next connection is accepted and returned or until it is closed.
+// It ignores [lneto.ErrExhausted] which cause dropped connections.
 func (l *tcplistener) Accept() (net.Conn, error) {
-	if l.closed {
-		return nil, net.ErrClosed
-	}
 	var backoffs uint
 	for {
-		if l.closed {
-			return nil, net.ErrClosed
-		}
-		n := l.l.NumberOfReadyToAccept()
-		if n == 0 {
-			backoff(l.sleep, backoffs)
-			backoffs++
-			continue
-		}
-		backoffs = 0
 		c, _, err := l.l.TryAccept()
-		if err != nil {
-			return nil, err
+		if err == nil {
+			return tcpconn{
+				Conn:      c,
+				localAddr: l.localAddr,
+			}, nil
+		} else if err != lneto.ErrExhausted {
+			return nil, err // net.ErrClosed or failure.
 		}
-		cc := tcpconn{
-			Conn:      c,
-			localAddr: l.localAddr,
-		}
-		return cc, nil
+		backoff(l.sleep, backoffs)
+		backoffs++
 	}
 }
 
-func (l *tcplistener) Close() error {
-	if l.closed {
-		return net.ErrClosed
-	}
-	err := l.l.Close()
-	l.closed = true
-	return err
-}
+func (l *tcplistener) Close() error { return l.l.Close() }
 
 type tcpconn struct {
 	*tcp.Conn

x/xnet/xnet_listener_test.go

@@ -277,6 +277,67 @@ func TestListener_Close(t *testing.T) {
 	}
 }
 
+// TestTCPListener_CloseUnblocksAccept covers the net.Listener wrapper's Accept
+// poll loop being ended by a Close from another goroutine.
+func TestTCPListener_CloseUnblocksAccept(t *testing.T) {
+	const svPort uint16 = 80
+
+	pool, err := NewTCPPool(TCPPoolConfig{
+		PoolSize:           1,
+		QueueSize:          4,
+		TxBufSize:          512,
+		RxBufSize:          512,
+		EstablishedTimeout: 10e9,
+		ClosingTimeout:     10e9,
+		NewBackoff:         func() lneto.BackoffStrategy { return backoffYield },
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	var l tcplistener
+	// Sleep rather than yield between polls so Accept is genuinely parked in the
+	// loop when Close lands, instead of spinning a core for the whole test.
+	l.sleep = func(consecutiveBackoffs uint) time.Duration { return time.Millisecond }
+	l.localAddr = net.TCPAddrFromAddrPort(netip.AddrPortFrom(netip.AddrFrom4([4]byte{10, 0, 0, 1}), svPort))
+	err = l.l.Reset(svPort, pool)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	// No stack is driving this listener, so Accept can only ever block: nothing
+	// will become ready and the sole way out is the Close below.
+	accepted := make(chan error, 1)
+	go func() {
+		c, err := l.Accept()
+		if c != nil {
+			c.Close()
+		}
+		accepted <- err
+	}()
+	time.Sleep(20 * time.Millisecond) // Let Accept reach its poll loop.
+
+	if err := l.Close(); err != nil {
+		t.Fatal("Close while Accept is blocked:", err)
+	}
+	select {
+	case err := <-accepted:
+		if err != net.ErrClosed {
+			t.Fatalf("blocked Accept: want net.ErrClosed, got %v", err)
+		}
+	case <-time.After(5 * time.Second):
+		t.Fatal("Accept did not return after Close")
+	}
+
+	// A later Accept reports the close rather than blocking again.
+	if _, err := l.Accept(); err != net.ErrClosed {
+		t.Fatalf("Accept after Close: want net.ErrClosed, got %v", err)
+	}
+	if err := l.Close(); err != net.ErrClosed {
+		t.Fatalf("double Close: want net.ErrClosed, got %v", err)
+	}
+}
+
 func TestListener_ResetAfterClose(t *testing.T) {
 	const svPort uint16 = 80