TUN-10621: Propagate max wait timeout

This PR addresses an issue where cloudflared prematurely closes the origin connection before the upstream-to-downstream goroutine finishes reading, causing intermittent connection drops when a client immediately closes the write-side of a connection.

When a client finishes writing data, it immediately closes its side of the connection. Under the current implementation in cloudflared's downstream-to-upstream goroutine does not wait for the second stream to complete. It unblocks the moment the first stream writes to the channel. Once this happens, the pipe returns control to `proxyTCPStream`, which prematurely closes the origin connection. Consequently, when the upstream-to-downstream goroutine attempts to read the remaining data from the origin connection, the connection is already gone, leading to unexpected failures.

We started propagating the `TimeoutAfterFirstClose` configuration/parameter. This allows the proxy to wait for a designated period, giving the second stream sufficient time to finish processing and read all remaining data before `proxyTCPStream` tears down the origin connection.
This commit is contained in:
Miguel da Costa Martins Marcelino
2026-06-29 10:57:12 +00:00
parent dba2d33a6b
commit 5c66bd68ab
9 changed files with 93 additions and 39 deletions
+1 -1
View File
@@ -37,7 +37,7 @@ func (ws *Websocket) ServeStream(options *StartOptions, conn io.ReadWriter) erro
}
defer func() { _ = wsConn.Close() }()
stream.Pipe(wsConn, conn, ws.log)
stream.Pipe(wsConn, conn, stream.DefaultTimeoutAfterFirstClose, ws.log)
return nil
}
+2 -2
View File
@@ -151,7 +151,7 @@ func wsEchoEndpoint(w ResponseWriter, r *http.Request) error {
}()
originConn := &echoPipe{reader: readPipe, writer: writePipe}
stream.Pipe(wsConn, originConn, &log)
stream.Pipe(wsConn, originConn, 0, &log)
cancel()
wsConn.Close()
return nil
@@ -190,7 +190,7 @@ func wsFlakyEndpoint(w ResponseWriter, r *http.Request) error {
rInt, _ := rand.Int(rand.Reader, big.NewInt(50))
closedAfter := time.Millisecond * time.Duration(rInt.Int64())
originConn := &flakyConn{closeAt: time.Now().Add(closedAfter)}
stream.Pipe(wsConn, originConn, &log)
stream.Pipe(wsConn, originConn, 0, &log)
cancel()
wsConn.Close()
return nil
+3 -3
View File
@@ -25,9 +25,9 @@ type OriginConnection interface {
type streamHandlerFunc func(originConn io.ReadWriter, remoteConn net.Conn, log *zerolog.Logger)
// DefaultStreamHandler is an implementation of streamHandlerFunc that
// performs a two way io.Copy between originConn and remoteConn.
// performs a two-way io.Copy between originConn and remoteConn.
func DefaultStreamHandler(originConn io.ReadWriter, remoteConn net.Conn, log *zerolog.Logger) {
stream.Pipe(originConn, remoteConn, log)
stream.Pipe(originConn, remoteConn, stream.DefaultTimeoutAfterFirstClose, log)
}
// tcpConnection is an OriginConnection that directly streams to raw TCP.
@@ -38,7 +38,7 @@ type tcpConnection struct {
}
func (tc *tcpConnection) Stream(_ context.Context, tunnelConn io.ReadWriter, _ *zerolog.Logger) {
stream.Pipe(tunnelConn, tc, tc.logger)
stream.Pipe(tunnelConn, tc, stream.DefaultTimeoutAfterFirstClose, tc.logger)
}
func (tc *tcpConnection) Write(b []byte) (int, error) {
+2 -1
View File
@@ -38,6 +38,7 @@ func TestStreamTCPConnection(t *testing.T) {
tcpConn := tcpConnection{
Conn: cfdConn,
writeTimeout: 30 * time.Second,
logger: TestLogger,
}
eyeballConn, edgeConn := net.Pipe()
@@ -157,7 +158,7 @@ func TestSocksStreamWSOverTCPConnection(t *testing.T) {
require.NoError(t, err)
defer func() { _ = wsForwarderInConn.Close() }()
stream.Pipe(wsForwarderInConn, &wsEyeball{wsForwarderOutConn}, TestLogger)
stream.Pipe(wsForwarderInConn, &wsEyeball{wsForwarderOutConn}, stream.DefaultTimeoutAfterFirstClose, TestLogger)
return nil
})
+5 -5
View File
@@ -99,7 +99,6 @@ type rawTCPService struct {
name string
dialer net.Dialer
writeTimeout time.Duration
logger *zerolog.Logger
}
func (o *rawTCPService) String() string {
@@ -233,10 +232,12 @@ func (o *helloWorld) start(
if err != nil {
return errors.Wrap(err, "Cannot start Hello World Server")
}
go hello.StartHelloWorldServer(log, helloListener, shutdownC)
go func() {
_ = hello.StartHelloWorldServer(log, helloListener, shutdownC)
}()
o.server = helloListener
o.httpService.url = &url.URL{
o.url = &url.URL{
Scheme: "https",
Host: o.server.Addr().String(),
}
@@ -356,7 +357,7 @@ func newHTTPTransport(service OriginService, cfg OriginRequestConfig, log *zerol
IdleConnTimeout: cfg.KeepAliveTimeout.Duration,
TLSHandshakeTimeout: cfg.TLSTimeout.Duration,
ExpectContinueTimeout: 1 * time.Second,
TLSClientConfig: &tls.Config{RootCAs: originCertPool, InsecureSkipVerify: cfg.NoTLSVerify},
TLSClientConfig: &tls.Config{RootCAs: originCertPool, InsecureSkipVerify: cfg.NoTLSVerify}, //nolint: gosec
ForceAttemptHTTP2: cfg.Http2Origin,
}
if _, isHelloWorld := service.(*helloWorld); !isHelloWorld && cfg.OriginServerName != "" {
@@ -374,7 +375,6 @@ func newHTTPTransport(service OriginService, cfg OriginRequestConfig, log *zerol
// DialContext depends on which kind of origin is being used.
dialContext := dialer.DialContext
switch service := service.(type) {
// If this origin is a unix socket, enforce network type "unix".
case *unixSocketPath:
httpTransport.DialContext = func(ctx context.Context, _, _ string) (net.Conn, error) {
+4 -3
View File
@@ -256,7 +256,7 @@ func (p *Proxy) proxyHTTPRequest(
reader: tr.Body,
}
stream.Pipe(eyeballStream, rwc, logger)
stream.Pipe(eyeballStream, rwc, time.Second*0, logger)
return nil
}
@@ -311,7 +311,7 @@ func (p *Proxy) proxyStream(
// proxyTCPStream proxies private network type TCP connections as a stream towards an available origin.
//
// This is different than proxyStream because it's not leveraged ingress rule services and uses the
// This is different from proxyStream because it's not leveraged ingress rule services and uses the
// originDialer from OriginDialerService.
func (p *Proxy) proxyTCPStream(
tr *tracing.TracedContext,
@@ -344,7 +344,8 @@ func (p *Proxy) proxyTCPStream(
connectLatency.Observe(float64(time.Since(start).Milliseconds()))
logger.Debug().Msg("proxy stream acknowledged")
stream.Pipe(tunnelConn, originConn, logger)
stream.Pipe(tunnelConn, originConn, stream.DefaultTimeoutAfterFirstClose, logger)
return nil
}
+10 -8
View File
@@ -223,9 +223,9 @@ func testProxyWebsocket(proxy connection.OriginProxy) func(t *testing.T) {
}
if ctx.Err() == context.DeadlineExceeded {
t.Errorf("Test timed out")
readPipe.Close()
writePipe.Close()
responseWriter.Close()
_ = readPipe.Close()
_ = writePipe.Close()
_ = responseWriter.Close()
}
return nil
})
@@ -647,7 +647,7 @@ func TestConnections(t *testing.T) {
ingressServiceScheme: "tcp://",
originService: func(t *testing.T, ln net.Listener) {
// closing the listener created by the test.
ln.Close()
_ = ln.Close()
},
eyeballResponseWriter: newTCPRespWriter(replayer),
eyeballRequestBody: newTCPRequestBody([]byte("test2")),
@@ -756,6 +756,8 @@ func newTCPRequestBody(data []byte) *requestBody {
pr, pw := io.Pipe()
go func() {
_, _ = pw.Write(data)
// Close the write side once the payload has been sent.
_ = pw.Close()
}()
return &requestBody{
pr: pr,
@@ -801,8 +803,8 @@ func (p *pipedRequestBody) roundtrip(addr string) []byte {
if err != nil {
panic(err)
}
defer conn.Close()
defer resp.Body.Close()
defer func() { _ = conn.Close() }()
defer func() { _ = resp.Body.Close() }()
if resp.StatusCode != http.StatusSwitchingProtocols {
panic(fmt.Errorf("resp returned status code: %d", resp.StatusCode))
@@ -949,7 +951,7 @@ func runEchoTCPService(t *testing.T, l net.Listener) {
if err != nil {
panic(err)
}
defer conn.Close()
defer func() { _ = conn.Close() }()
for {
buf := make([]byte, 1024)
@@ -987,7 +989,7 @@ func runEchoWSService(t *testing.T, l net.Listener) {
t.Log(err)
return
}
defer conn.Close()
defer func() { _ = conn.Close() }()
for {
messageType, p, err := conn.ReadMessage()
+21 -11
View File
@@ -2,6 +2,7 @@ package stream
import (
"encoding/hex"
"errors"
"fmt"
"io"
"runtime/debug"
@@ -9,12 +10,17 @@ import (
"time"
"github.com/getsentry/sentry-go"
"github.com/pkg/errors"
"github.com/rs/zerolog"
"github.com/cloudflare/cloudflared/cfio"
)
// DefaultTimeoutAfterFirstClose controls the upper bound of how long we wait for the second stream to finish.
// We use bidirectional streams for our communication. Since the read and write sides can be closed independently,
// we must have a way to close the second stream once the first one finishes. We don't want to wait indefinitely,
// since we want to prevent misbehaving clients from blocking cloudflared.
const DefaultTimeoutAfterFirstClose = time.Second * 10
type Stream interface {
Reader
WriterCloser
@@ -37,7 +43,7 @@ type nopCloseWriterAdapter struct {
io.ReadWriter
}
func NopCloseWriterAdapter(stream io.ReadWriter) *nopCloseWriterAdapter {
func noopCloseWriter(stream io.ReadWriter) *nopCloseWriterAdapter {
return &nopCloseWriterAdapter{stream}
}
@@ -72,7 +78,7 @@ func (s *bidirectionalStreamStatus) wait(maxWaitForSecondStream time.Duration) e
select {
case <-timer.C:
return fmt.Errorf("timeout waiting for second stream to finish")
return fmt.Errorf("timeout waiting for second stream to finish %s", maxWaitForSecondStream)
case <-s.doneChan:
return nil
}
@@ -85,15 +91,19 @@ func (s *bidirectionalStreamStatus) isAnyDone() bool {
}
// Pipe copies copy data to & from provided io.ReadWriters.
func Pipe(tunnelConn, originConn io.ReadWriter, log *zerolog.Logger) {
_ = PipeBidirectional(NopCloseWriterAdapter(tunnelConn), NopCloseWriterAdapter(originConn), 0, log)
func Pipe(tunnelConn, originConn io.ReadWriter, timeoutAfterFirstClose time.Duration, log *zerolog.Logger) {
if err := PipeBidirectional(noopCloseWriter(tunnelConn), noopCloseWriter(originConn), timeoutAfterFirstClose, log); err != nil {
log.Warn().Err(err).Msg("Failed to pipe bidirectional stream")
}
}
// PipeBidirectional copies data to two unidirectional streams. It is a special case of Pipe where it receives a concept that allows for Read and Write side to be closed independently.
// The main difference is that when piping data from a reader to a writer, if EOF is read, then this implementation propagates the EOF signal to the destination/writer by closing the write side of the
// Bidirectional Stream.
// Finally, depending on once EOF is ready from one of the provided streams, the other direction of streaming data will have a configured time period to also finish, otherwise,
// the method will return immediately with a timeout error. It is however, the responsibility of the caller to close the associated streams in both ends in order to free all the resources/go-routines.
// PipeBidirectional copies data between two unidirectional streams. It is a special case of Pipe that accepts streams
// whose read and write sides can be closed independently. The main difference is that when piping data from a reader
// to a writer, if EOF is read, this implementation propagates the EOF signal to the destination by closing the write
// side of the bidirectional stream.
// Finally, once EOF is received from one of the provided streams, the other direction has a configured grace period to
// finish; otherwise, the method returns a timeout error. It is, however, the responsibility of the caller to close
// the associated streams at both ends in order to free all resources and goroutines.
func PipeBidirectional(downstream, upstream Stream, maxWaitForSecondStream time.Duration, log *zerolog.Logger) error {
status := newBiStreamStatus()
@@ -101,7 +111,7 @@ func PipeBidirectional(downstream, upstream Stream, maxWaitForSecondStream time.
go unidirectionalStream(upstream, downstream, "downstream->upstream", status, log)
if err := status.wait(maxWaitForSecondStream); err != nil {
return errors.Wrap(err, "unable to wait for both streams while proxying")
return fmt.Errorf("unable to wait for both streams while proxying: %w", err)
}
return nil
+45 -5
View File
@@ -30,8 +30,8 @@ func TestPipeBidirectionalFinishOneSideTimeout(t *testing.T) {
func TestPipeBidirectionalClosingWriteBothSidesAlsoExists(t *testing.T) {
fun := func(upstream, downstream *mockedStream) {
downstream.CloseWrite()
upstream.CloseWrite()
_ = downstream.CloseWrite()
_ = upstream.CloseWrite()
downstream.writeToReader("abc")
upstream.writeToReader("abc")
@@ -42,7 +42,7 @@ func TestPipeBidirectionalClosingWriteBothSidesAlsoExists(t *testing.T) {
func TestPipeBidirectionalClosingWriteSingleSideAlsoExists(t *testing.T) {
fun := func(upstream, downstream *mockedStream) {
downstream.CloseWrite()
_ = downstream.CloseWrite()
downstream.writeToReader("abc")
upstream.writeToReader("abc")
@@ -51,6 +51,46 @@ func TestPipeBidirectionalClosingWriteSingleSideAlsoExists(t *testing.T) {
testPipeBidirectionalUnblocking(t, fun, time.Millisecond*200, true)
}
// TestPipeBidirectionalReturnsWhenBothSidesFinish verifies that
// PipeBidirectional returns as soon as both stream directions finish, without
// waiting for the full timeout grace period to expire. This guards against a
// regression where the second-stream wait would block for the whole timeout
// even when the result is already available.
func TestPipeBidirectionalReturnsWhenBothSidesFinish(t *testing.T) {
t.Parallel()
const (
timeout = time.Second * 5
maxWallTime = time.Millisecond * 500
)
logger := zerolog.Nop()
downstream := newMockedStream()
upstream := newMockedStream()
resultCh := make(chan error, 1)
go func() {
resultCh <- PipeBidirectional(downstream, upstream, timeout, &logger)
}()
// Close both reader sides so both stream directions reach EOF promptly.
downstream.closeReader()
upstream.closeReader()
start := time.Now()
select {
case err := <-resultCh:
elapsed := time.Since(start)
require.NoError(t, err)
require.Less(t, elapsed, maxWallTime,
"PipeBidirectional should return as soon as both streams finish, not after the full %s timeout (took %s)",
timeout, elapsed,
)
case <-time.After(timeout):
require.Fail(t, "PipeBidirectional did not return before the timeout expired")
}
}
func testPipeBidirectionalUnblocking(t *testing.T, afterFun func(*mockedStream, *mockedStream), timeout time.Duration, expectTimeout bool) {
logger := zerolog.Nop()
@@ -67,9 +107,9 @@ func testPipeBidirectionalUnblocking(t *testing.T, afterFun func(*mockedStream,
select {
case err := <-resultCh:
if expectTimeout {
require.NotNil(t, err)
require.Error(t, err)
} else {
require.Nil(t, err)
require.NoError(t, err)
}
case <-time.After(timeout * 2):