fix(msg): interrupt control reads during dispatcher shutdown (#5561)

pkg/msg/handler.go

@@ -20,6 +20,7 @@ import (
 	"net"
 	"reflect"
 	"sync"
+	"time"
 
 	"github.com/fatedier/frp/pkg/proto/wire"
 )
@@ -117,9 +118,11 @@ type Dispatcher struct {
 	msgHandlers map[reflect.Type]func(Message)
 }
 
-// NewDispatcher uses Close, when rw implements io.Closer, to interrupt ReadMsg
-// after a write failure. Otherwise, the caller must arrange for ReadMsg to return
-// before Done can close.
+// NewDispatcher uses SetReadDeadline, when supported, and Close, when rw
+// implements io.Closer, to interrupt ReadMsg after a write failure. Without a
+// working read deadline, Close or the caller must arrange for ReadMsg to return.
+// These methods must return in bounded time and must not wait for Done: Done
+// closes only after ReadMsg, any synchronous handler, and stop have returned.
 func NewDispatcher(rw ReadWriter) *Dispatcher {
 	return &Dispatcher{
 		rw:          rw,
@@ -136,11 +139,14 @@ func (d *Dispatcher) Run() {
 	go d.readLoop()
 }
 
-// stop rejects new sends and interrupts ReadMsg when rw is closable. Done remains
+// stop rejects new sends and interrupts ReadMsg when supported by rw. Done remains
 // owned by readLoop so the control owner cannot clean up while a handler is running.
 func (d *Dispatcher) stop() {
 	d.stopOnce.Do(func() {
 		close(d.stopCh)
+		if conn, ok := d.rw.(interface{ SetReadDeadline(time.Time) error }); ok {
+			_ = conn.SetReadDeadline(time.Now())
+		}
 		if closer, ok := d.rw.(io.Closer); ok {
 			_ = closer.Close()
 		}

pkg/msg/handler_shutdown_test.go

@@ -31,18 +31,13 @@ func TestDispatcherEnqueueRejectsStoppedDispatcher(t *testing.T) {
 
 	// Model shutdown after Send's initial check: the stop and buffered send
 	// cases are both ready. A send selected here must still return EOF.
-	enqueued := false
-	for range 100 {
 	if err := d.enqueue(&ReqWorkConn{}); !errors.Is(err, io.EOF) {
 		t.Fatalf("enqueue racing with shutdown: %v, want EOF", err)
 	}
+	// This single call checks that a stopped dispatcher's enqueue returns EOF; it
+	// does not guarantee exercising the send case and post-enqueue check on every run.
 	if len(d.sendCh) > 0 {
-			enqueued = true
-			<-d.sendCh
-		}
-	}
-	if !enqueued {
-		t.Fatal("did not exercise the ready send case")
+		t.Log("selected the send case; post-enqueue check returned EOF")
 	}
 }
 
@@ -65,7 +60,7 @@ func TestDispatcherAcceptsReadWriterWithoutClose(t *testing.T) {
 		})
 
 		c.transport.arm()
-		if err := d.Send(&ReqWorkConn{}); err != nil {
+		if err := d.Send(&ReqWorkConn{}); err != nil && !errors.Is(err, io.EOF) {
 			t.Fatal(err)
 		}
 		if err := regressionReceiveWrite(t, c); !errors.Is(err, errRegressionWriteTimeout) {

pkg/msg/handler_write_error_test.go

@@ -207,7 +207,7 @@ func TestDispatcherWriteErrorClosesControlAndFreshSessionWorks(t *testing.T) {
 	d := NewDispatcher(c)
 	d.Run()
 	c.transport.arm()
-	if err := d.Send(&ReqWorkConn{}); err != nil {
+	if err := d.Send(&ReqWorkConn{}); err != nil && !errors.Is(err, io.EOF) {
 		t.Fatal(err)
 	}
 	if err := regressionReceiveWrite(t, c); !errors.Is(err, errRegressionWriteTimeout) {
@@ -215,11 +215,9 @@ func TestDispatcherWriteErrorClosesControlAndFreshSessionWorks(t *testing.T) {
 	}
 	regressionAwait(t, c.transport.closed, "failed control connection close")
 	regressionAwait(t, d.Done(), "failed dispatcher completion")
-	for range 100 {
 	if err := d.Send(&ReqWorkConn{}); !errors.Is(err, io.EOF) {
 		t.Fatalf("Send after completed shutdown: %v", err)
 	}
-	}
 	if c.closeCalls.Load() != 1 {
 		t.Fatalf("connection closed %d times", c.closeCalls.Load())
 	}
@@ -255,6 +253,7 @@ func TestDispatcherWriteErrorClosesControlAndFreshSessionWorks(t *testing.T) {
 }
 
 func TestDispatcherWriteErrorWaitsForSynchronousHandler(t *testing.T) {
+	synctest.Test(t, func(t *testing.T) {
 		c := newRegressionControl(t)
 		regressionWarm(t, c)
 		defer c.transport.close()
@@ -270,25 +269,26 @@ func TestDispatcherWriteErrorWaitsForSynchronousHandler(t *testing.T) {
 		c.incoming <- &ReqWorkConn{}
 		regressionAwait(t, entered, "synchronous handler entry")
 		c.transport.arm()
-	if err := d.Send(&ReqWorkConn{}); err != nil {
+		if err := d.Send(&ReqWorkConn{}); err != nil && !errors.Is(err, io.EOF) {
 			t.Fatal(err)
 		}
 		if err := regressionReceiveWrite(t, c); !errors.Is(err, errRegressionWriteTimeout) {
 			t.Fatal(err)
 		}
 		regressionAwait(t, c.transport.closed, "write failure closes control during handler")
-	for range 100 {
 		if err := d.Send(&ReqWorkConn{}); !errors.Is(err, io.EOF) {
 			t.Fatalf("Send after stop while handler is active: %v", err)
 		}
-	}
+		// Let the writer finish stopping while the handler remains blocked.
+		synctest.Wait()
 		select {
 		case <-d.Done():
 			t.Fatal("Done closed before synchronous handler returned; owner cleanup could race proxy registration")
-	case <-time.After(100 * time.Millisecond):
+		default:
 		}
 		releaseOnce.Do(func() { close(release) })
 		regressionAwait(t, d.Done(), "completion after handler returned")
+	})
 }
 
 func TestDispatcherWriteErrorUnblocksSenderWithFullQueue(t *testing.T) {
@@ -335,8 +335,6 @@ func TestDispatcherWriteErrorUnblocksSenderWithFullQueue(t *testing.T) {
 }
 
 func TestDispatcherConcurrentReadWriteFailureClosesOnce(t *testing.T) {
-	for range 100 {
-		func() {
 	c := newRegressionControl(t)
 	regressionWarm(t, c)
 	defer c.transport.close()
@@ -382,6 +380,4 @@ func TestDispatcherConcurrentReadWriteFailureClosesOnce(t *testing.T) {
 	if c.closeCalls.Load() != 1 {
 		t.Fatalf("connection closed %d times", c.closeCalls.Load())
 	}
-		}()
-	}
 }

pkg/msg/handler_yamux_test.go

@@ -0,0 +1,225 @@
+// Copyright 2026 The frp Authors
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package msg
+
+import (
+	"errors"
+	"io"
+	"net"
+	"sync"
+	"sync/atomic"
+	"testing"
+	"time"
+
+	fmux "github.com/fatedier/yamux"
+
+	"github.com/fatedier/frp/pkg/proto/wire"
+	netpkg "github.com/fatedier/frp/pkg/util/net"
+)
+
+// Delay one real TCP write without injecting an error. The stream's write
+// failure must come from yamux's ConnectionWriteTimeout.
+type delayedTCPConn struct {
+	net.Conn
+	armed   atomic.Bool
+	entered chan struct{}
+	release chan struct{}
+}
+
+func (c *delayedTCPConn) Write(p []byte) (int, error) {
+	if c.armed.Swap(false) {
+		close(c.entered)
+		<-c.release
+	}
+	return c.Conn.Write(p)
+}
+
+type observedYamuxConn struct {
+	net.Conn
+	writes      chan error
+	readEntered chan struct{}
+	readOnce    sync.Once
+	closed      chan struct{}
+	closeCalls  atomic.Int32
+}
+
+func (c *observedYamuxConn) Read(p []byte) (int, error) {
+	c.readOnce.Do(func() { close(c.readEntered) })
+	return c.Conn.Read(p)
+}
+
+func (c *observedYamuxConn) Write(p []byte) (int, error) {
+	n, err := c.Conn.Write(p)
+	c.writes <- err
+	return n, err
+}
+
+func (c *observedYamuxConn) Close() error {
+	c.closeCalls.Add(1)
+	err := c.Conn.Close()
+	close(c.closed)
+	return err
+}
+
+func newLoopbackYamuxPair(t *testing.T) (*fmux.Session, *fmux.Session, *delayedTCPConn) {
+	t.Helper()
+	listener, err := net.Listen("tcp", "127.0.0.1:0")
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer listener.Close()
+	clientTCP, err := net.Dial("tcp", listener.Addr().String())
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(func() { clientTCP.Close() })
+	serverTCP, err := listener.Accept()
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(func() { serverTCP.Close() })
+	gate := &delayedTCPConn{Conn: clientTCP, entered: make(chan struct{}), release: make(chan struct{})}
+	config := fmux.DefaultConfig()
+	config.ConnectionWriteTimeout = 80 * time.Millisecond
+	config.LogOutput = io.Discard
+	client, err := fmux.Client(gate, config)
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(func() { client.Close() })
+	server, err := fmux.Server(serverTCP, config)
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(func() { server.Close() })
+	return client, server, gate
+}
+
+func loopbackCryptoConn(t *testing.T, conn net.Conn, protocol string, role netpkg.AEADCryptoRole) *Conn {
+	t.Helper()
+	var rw io.ReadWriter
+	var err error
+	if protocol == wire.ProtocolV2 {
+		rw, err = netpkg.NewAEADCryptoReadWriter(
+			conn, regressionKey, role, wire.AEADAlgorithmAES256GCM, []byte("isolated-test-transcript"))
+	} else {
+		rw, err = netpkg.NewCryptoReadWriter(conn, regressionKey)
+	}
+	if err != nil {
+		t.Fatal(err)
+	}
+	return NewConn(conn, NewReadWriter(rw, protocol))
+}
+
+func loopbackControlPair(t *testing.T, client, server *fmux.Session, protocol string) (*Conn, *Conn, *observedYamuxConn) {
+	t.Helper()
+	local, err := client.OpenStream()
+	if err != nil {
+		t.Fatal(err)
+	}
+	remote, err := server.AcceptStream()
+	if err != nil {
+		t.Fatal(err)
+	}
+	observed := &observedYamuxConn{
+		Conn: local, writes: make(chan error, 16), readEntered: make(chan struct{}), closed: make(chan struct{}),
+	}
+	return loopbackCryptoConn(t, observed, protocol, netpkg.AEADCryptoRoleClient),
+		loopbackCryptoConn(t, remote, protocol, netpkg.AEADCryptoRoleServer), observed
+}
+
+func loopbackReadReq(t *testing.T, conn *Conn) {
+	t.Helper()
+	if err := conn.SetReadDeadline(time.Now().Add(time.Second)); err != nil {
+		t.Fatal(err)
+	}
+	defer conn.SetReadDeadline(time.Time{})
+	m, err := conn.ReadMsg()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, ok := m.(*ReqWorkConn); !ok {
+		t.Fatalf("decoded %T, want ReqWorkConn", m)
+	}
+}
+
+func TestDispatcherYamuxWriteTimeoutInterruptsRead(t *testing.T) {
+	for _, protocol := range []string{wire.ProtocolV1, wire.ProtocolV2} {
+		t.Run(protocol, func(t *testing.T) {
+			client, server, gate := newLoopbackYamuxPair(t)
+			var releaseOnce sync.Once
+			defer releaseOnce.Do(func() { close(gate.release) })
+			control, peer, observed := loopbackControlPair(t, client, server, protocol)
+			if err := control.WriteMsg(&ReqWorkConn{}); err != nil {
+				t.Fatal(err)
+			}
+			loopbackReadReq(t, peer)
+			// Warm-up has completed, including all crypto header writes.
+			for len(observed.writes) > 0 {
+				<-observed.writes
+			}
+			d := NewDispatcher(control)
+			readExited := make(chan struct{})
+			go func() {
+				defer close(readExited)
+				d.readLoop()
+			}()
+			regressionAwait(t, observed.readEntered, "control reader entry")
+			gate.armed.Store(true)
+			// Queue before starting the writer so this Send cannot race its failure.
+			if err := d.Send(&ReqWorkConn{}); err != nil {
+				t.Fatal(err)
+			}
+			go d.sendLoop()
+			regressionAwait(t, gate.entered, "TCP write delay barrier")
+			select {
+			case err := <-observed.writes:
+				if !errors.Is(err, fmux.ErrConnectionWriteTimeout) {
+					t.Fatalf("stream write error: %v, want yamux connection write timeout", err)
+				}
+				t.Logf("production yamux write failed: %v", err)
+			case <-time.After(time.Second):
+				t.Fatal("yamux write did not time out")
+			}
+			regressionAwait(t, d.stopCh, "write failure stops sends")
+			if err := d.Send(&ReqWorkConn{}); !errors.Is(err, io.EOF) {
+				t.Fatalf("Send after stop: %v, want EOF", err)
+			}
+			releaseOnce.Do(func() { close(gate.release) })
+			regressionAwait(t, observed.closed, "control Close returns")
+			// peer never closes its write half: local yamux Close alone cannot
+			// interrupt Read. Neither a peer FIN nor session cleanup may help.
+			regressionAwait(t, d.Done(), "Done without peer FIN")
+			regressionAwait(t, readExited, "control readLoop exit without peer FIN")
+			if observed.closeCalls.Load() != 1 {
+				t.Fatalf("control closed %d times, want once", observed.closeCalls.Load())
+			}
+			if client.IsClosed() || server.IsClosed() {
+				t.Fatal("yamux session unexpectedly closed")
+			}
+			fresh, freshPeer, freshObserved := loopbackControlPair(t, client, server, protocol)
+			for range 5 {
+				if err := fresh.WriteMsg(&ReqWorkConn{}); err != nil {
+					t.Fatal(err)
+				}
+				loopbackReadReq(t, freshPeer)
+				for len(freshObserved.writes) > 0 {
+					<-freshObserved.writes
+				}
+			}
+			t.Log("Done/readLoop exited without peer FIN; Close once; same yamux session decrypted five fresh messages")
+		})
+	}
+}