Tweak shutdown behaviour (again)

The previous behaviour was unsound, as it was prone to dropping
connections under (temporary) high load. The new behaviour requires that
users are well-behaved with respect to shutdown - the root listener
must be shut down before any of the child listeners are, otherwise
deadlocks may occur. This requirement seems reasonable.

bench_test.go

@@ -32,7 +32,6 @@ func BenchmarkCMuxConn(b *testing.B) {
 		}
 	}()
 
-	donec := make(chan struct{})
 	var wg sync.WaitGroup
 	wg.Add(b.N)
 
@@ -41,6 +40,6 @@ func BenchmarkCMuxConn(b *testing.B) {
 		c := &mockConn{
 			r: bytes.NewReader(benchHTTPPayload),
 		}
-		m.serve(c, donec, &wg)
+		m.serve(c, &wg)
 	}
 }

cmux.go

@@ -45,11 +45,16 @@ var ErrListenerClosed = errListenerClosed("mux: listener closed")
 
 // New instantiates a new connection multiplexer.
 func New(l net.Listener) CMux {
+	return NewSize(l, 1024)
+}
+
+// NewSize instantiates a new connection multiplexer which buffers up to size
+// connections in its child listeners.
+func NewSize(l net.Listener, size int) CMux {
 	return &cMux{
 		root:   l,
-		bufLen: 1024,
+		bufLen: size,
 		errh:   func(_ error) bool { return true },
-		donec:  make(chan struct{}),
 	}
 }
 
@@ -76,7 +81,6 @@ type cMux struct {
 	root   net.Listener
 	bufLen int
 	errh   ErrorHandler
-	donec  chan struct{}
 	sls    []matchersListener
 }
 
@@ -93,7 +97,6 @@ func (m *cMux) Serve() error {
 	var wg sync.WaitGroup
 
 	defer func() {
-		close(m.donec)
 		wg.Wait()
 
 		for _, sl := range m.sls {
@@ -111,11 +114,11 @@ func (m *cMux) Serve() error {
 		}
 
 		wg.Add(1)
-		go m.serve(c, m.donec, &wg)
+		go m.serve(c, &wg)
 	}
 }
 
-func (m *cMux) serve(c net.Conn, donec <-chan struct{}, wg *sync.WaitGroup) {
+func (m *cMux) serve(c net.Conn, wg *sync.WaitGroup) {
 	defer wg.Done()
 
 	muc := newMuxConn(c)
@@ -124,15 +127,7 @@ func (m *cMux) serve(c net.Conn, donec <-chan struct{}, wg *sync.WaitGroup) {
 			matched := s(muc.sniffer())
 			muc.reset()
 			if matched {
-				select {
-				case sl.l.connc <- muc:
-				default:
-					select {
-					case <-donec:
-						_ = c.Close()
-					default:
-					}
-				}
+				sl.l.connc <- muc
 				return
 			}
 		}

cmux_test.go

@@ -354,27 +354,47 @@ func TestClose(t *testing.T) {
 	}()
 	l := newChanListener()
 
-	c1, c2 := net.Pipe()
-
-	muxl := New(l)
+	muxl := NewSize(l, 0)
 	anyl := muxl.Match(Any())
 
+	defer func() {
+		// Listener is closed.
+		close(l.connCh)
+
+		// Root listener is closed now.
+		if _, err := anyl.Accept(); err != ErrListenerClosed {
+			t.Fatal(err)
+		}
+	}()
+
 	go safeServe(errCh, muxl)
 
+	c1, c2 := net.Pipe()
+
 	l.connCh <- c1
 
-	// First connection goes through.
+	// Simulate the child listener being temporarily blocked. The connection
+	// multiplexer must be unbuffered (size=0). Before the fix, this would cause
+	// connections to be lost.
+	time.Sleep(50 * time.Millisecond)
+
+	go func() {
+		// Simulate the child listener being temporarily blocked. The connection
+		// multiplexer must be unbuffered (size=0). Before the fix, this would cause
+		// connections to be lost.
+		time.Sleep(50 * time.Millisecond)
+
+		l.connCh <- c2
+	}()
+
+	// First connection goes through. Before the fix, c1 would be lost while we
+	// were temporarily blocked, so this consumed c2.
 	if _, err := anyl.Accept(); err != nil {
 		t.Fatal(err)
 	}
 
-	// Second connection is sent
-	l.connCh <- c2
-
-	// Listener is closed.
-	close(l.connCh)
-
-	// Second connection goes through.
+	// Second connection goes through. Before the fix, c2 would be consumed above, so
+	// this would block indefinitely.
 	if _, err := anyl.Accept(); err != nil {
 		t.Fatal(err)
 	}