caddy/modules/caddyhttp/reverseproxy/streaming.go

// Copyright 2015 Matthew Holt and The Caddy 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.

// Most of the code in this file was initially borrowed from the Go
// standard library and modified; It had this copyright notice:
// Copyright 2011 The Go Authors

package reverseproxy

import (
	"bufio"
	"context"
	"errors"
	"fmt"
	"io"
	weakrand "math/rand/v2"
	"mime"
	"net"
	"net/http"
	"sync"
	"sync/atomic"
	"time"
	"unsafe"

	"go.uber.org/zap"
	"go.uber.org/zap/zapcore"
	"golang.org/x/net/http/httpguts"

	"github.com/caddyserver/caddy/v2"
	"github.com/caddyserver/caddy/v2/modules/caddyhttp"
)

type extendedConnectReadWriteCloser struct {
	io.ReadCloser
	http.ResponseWriter
}

func (rwc extendedConnectReadWriteCloser) Write(p []byte) (n int, err error) {
	n, err = rwc.ResponseWriter.Write(p)
	if err != nil {
		return 0, err
	}

	//nolint:bodyclose
	err = http.NewResponseController(rwc.ResponseWriter).Flush()
	if err != nil {
		return 0, err
	}
	return n, nil
}

// SetWriteDeadline bounds writes to the extended CONNECT stream, so a peer
// that stopped reading cannot block the close message forever.
func (rwc extendedConnectReadWriteCloser) SetWriteDeadline(deadline time.Time) error {
	//nolint:bodyclose
	return http.NewResponseController(rwc.ResponseWriter).SetWriteDeadline(deadline)
}

// upstreamConnCtxKey is the context key for the *atomic.Pointer[net.Conn]
// recording the connection a websocket upgrade's round trip used.
const upstreamConnCtxKey caddy.CtxKey = "reverse_proxy.upstream_conn"

// upstreamWriteCloser returns where to write the close message to the
// upstream. The write must be bounded by a deadline, but for HTTP/1
// upgrades net/http hands the upstream socket back inside a body type that
// hides its deadline methods and writes straight through to the socket, so
// the close message goes to the socket recorded during the round trip
// instead. An HTTP/1 connection is dedicated to this request; HTTP/2 and
// HTTP/3 share theirs between streams, so their bodies are used as they are.
func upstreamWriteCloser(req *http.Request, res *http.Response, backConn io.ReadWriteCloser) io.ReadWriteCloser {
	if res.ProtoMajor != 1 {
		return backConn
	}
	recorded, ok := req.Context().Value(upstreamConnCtxKey).(*atomic.Pointer[net.Conn])
	if !ok {
		return backConn
	}
	if uc := recorded.Load(); uc != nil {
		return *uc
	}
	return backConn
}

func (h *Handler) handleUpgradeResponse(logger *zap.Logger, rw http.ResponseWriter, req *http.Request, res *http.Response, upstreamAddr string) {
	// Until the tunnel takes ownership, every setup failure must close the
	// backend body and any hijacked client connection. net/http no longer
	// owns the client socket after a successful hijack.
	var conn io.ReadWriteCloser
	tunnelOwnsConnections := false
	defer func() {
		if !tunnelOwnsConnections {
			res.Body.Close()
			if conn != nil {
				conn.Close()
			}
		}
	}()
	reqUpType := upgradeType(req.Header)
	resUpType := upgradeType(res.Header)

	// Taken from https://github.com/golang/go/commit/5c489514bc5e61ad9b5b07bd7d8ec65d66a0512a
	// We know reqUpType is ASCII, it's checked by the caller.
	if !asciiIsPrint(resUpType) {
		if c := logger.Check(zapcore.DebugLevel, "backend tried to switch to invalid protocol"); c != nil {
			c.Write(zap.String("backend_upgrade", resUpType))
		}
		return
	}
	if !asciiEqualFold(reqUpType, resUpType) {
		if c := logger.Check(zapcore.DebugLevel, "backend tried to switch to unexpected protocol via Upgrade header"); c != nil {
			c.Write(
				zap.String("backend_upgrade", resUpType),
				zap.String("requested_upgrade", reqUpType),
			)
		}
		return
	}

	backConn, ok := res.Body.(io.ReadWriteCloser)
	if !ok {
		logger.Error("internal error: 101 switching protocols response with non-writable body")
		return
	}

	// write header first, response headers should not be counted in size
	// like the rest of handler chain.
	copyHeader(rw.Header(), res.Header)

	// Capture all h fields needed by the tunnel now, so that the Handler (h)
	// is not referenced after this function returns (for HTTP/1.1 hijacked
	// connections the tunnel runs in a detached goroutine).
	tunnel := h.tunnelTracker
	bufferSize := h.StreamBufferSize
	streamTimeout := time.Duration(h.StreamTimeout)

	var (
		brw      *bufio.ReadWriter
		detached bool
	)
	// websocket over http2 or http3 if extended connect is enabled,
	// assuming backend doesn't support this, the request will be
	// modified to http1.1 upgrade
	// TODO: once we can reliably detect backend support this, it can
	// be removed for those backends
	if body, ok := caddyhttp.GetVar(req.Context(), "extended_connect_websocket_body").(io.ReadCloser); ok {
		// websocket over extended connect can't be detached. rw and req.Body
		// are only valid while the handler goroutine is running
		req.Body = body
		rw.Header().Del("Upgrade")
		rw.Header().Del("Connection")
		rw.Header().Del("Sec-WebSocket-Accept")
		rw.WriteHeader(http.StatusOK)

		if c := logger.Check(zap.DebugLevel, "upgrading connection"); c != nil {
			c.Write(zap.Int("http_version", req.ProtoMajor))
		}

		//nolint:bodyclose
		flushErr := http.NewResponseController(rw).Flush()
		if flushErr != nil {
			if c := h.logger.Check(zap.ErrorLevel, "failed to flush extended_connect websocket response"); c != nil {
				c.Write(zap.Error(flushErr))
			}
			return
		}
		conn = extendedConnectReadWriteCloser{req.Body, rw}
		// bufio is not needed, use minimal buffer
		brw = bufio.NewReadWriter(bufio.NewReaderSize(conn, 1), bufio.NewWriterSize(conn, 1))
	} else {
		normalizeWebsocketHeaders(rw.Header())
		if h.StreamDetached {
			detached = caddyhttp.DetachResponseWriterAfterHijack(rw, req, true)
			if !detached {
				// Some outer writers may have accepted detachment before an
				// inner writer refused it. Restore attached byte accounting
				// and keep the handler running for the lifetime of the tunnel.
				caddyhttp.DetachResponseWriterAfterHijack(rw, req, false)
				if c := logger.Check(zap.DebugLevel, "detaching connection failed; keeping stream attached"); c != nil {
					c.Write(zap.String("tip", "check if your response writers have an Unwrap method or if already hijacked"))
				}
			}
		}
		rw.WriteHeader(res.StatusCode)

		if c := logger.Check(zap.DebugLevel, "upgrading connection"); c != nil {
			c.Write(zap.Int("http_version", req.ProtoMajor))
		}

		var hijackErr error
		//nolint:bodyclose
		conn, brw, hijackErr = http.NewResponseController(rw).Hijack()
		if errors.Is(hijackErr, http.ErrNotSupported) {
			if c := h.logger.Check(zap.ErrorLevel, "can't switch protocols using non-Hijacker ResponseWriter"); c != nil {
				c.Write(zap.String("type", fmt.Sprintf("%T", rw)))
			}
			return
		}

		if hijackErr != nil {
			if c := h.logger.Check(zap.ErrorLevel, "hijack failed on protocol switch"); c != nil {
				c.Write(zap.Error(hijackErr))
			}
			return
		}
	}

	if err := brw.Flush(); err != nil {
		if c := logger.Check(zapcore.DebugLevel, "response flush"); c != nil {
			c.Write(zap.Error(err))
		}
		return
	}

	// There may be buffered data in the *bufio.Reader
	// see: https://github.com/caddyserver/caddy/issues/6273
	if buffered := brw.Reader.Buffered(); buffered > 0 {
		data, _ := brw.Peek(buffered)
		_, err := backConn.Write(data)
		if err != nil {
			if c := logger.Check(zapcore.DebugLevel, "backConn write failed"); c != nil {
				c.Write(zap.Error(err))
			}
			return
		}
	}

	// Register both connections with the tunnel tracker. We also try to
	// gracefully close connections we recognize as websockets. We need to make
	// sure the client connection messages (i.e. to upstream) are masked, so we
	// need to know whether the connection is considered the server or the
	// client side of the proxy. Note that gracefulClose must not capture h,
	// since the tunnel may outlive the handler instance.
	gracefulClose := func(conn io.ReadWriteCloser, isClient bool) func(time.Time) error {
		if isWebsocket(req) {
			return func(deadline time.Time) error {
				return writeCloseControl(conn, isClient, deadline)
			}
		}
		return nil
	}
	deleteFrontConn := tunnel.registerConnection(conn, gracefulClose(conn, false), detached, upstreamAddr)
	deleteBackConn := tunnel.registerConnection(backConn, gracefulClose(upstreamWriteCloser(req, res, backConn), true), detached, upstreamAddr)
	if h.streamLogsSkipHandshake() {
		caddyhttp.SetVar(req.Context(), caddyhttp.LogSkipVar, true)
	}
	repl := req.Context().Value(caddy.ReplacerCtxKey).(*caddy.Replacer)
	repl.Set("http.reverse_proxy.upgraded", true)
	streamUUID, _ := repl.GetString("http.request.uuid")
	streamFields := makeStreamLogFields(streamUUID)
	streamLogger := h.streamLoggerForRequest(req)
	streamLevel := h.streamLogLevel
	finishMetrics := trackActiveStream(upstreamAddr)

	start := time.Now()

	tunnelOwnsConnections = true
	if !detached {
		handleUpgradeTunnel(
			streamLogger,
			streamLevel,
			conn,
			backConn,
			deleteFrontConn,
			deleteBackConn,
			bufferSize,
			streamTimeout,
			start,
			finishMetrics,
			streamFields,
		)
	} else {
		// start a new goroutine
		go handleUpgradeTunnel(
			streamLogger,
			streamLevel,
			conn,
			backConn,
			deleteFrontConn,
			deleteBackConn,
			bufferSize,
			streamTimeout,
			start,
			finishMetrics,
			streamFields,
		)
	}
}

// handleUpgradeTunnel returns when transfer is done.
func handleUpgradeTunnel(
	streamLogger *zap.Logger,
	streamLevel zapcore.Level,
	conn io.ReadWriteCloser,
	backConn io.ReadWriteCloser,
	deleteFrontConn func(),
	deleteBackConn func(),
	bufferSize int,
	streamTimeout time.Duration,
	start time.Time,
	finishMetrics func(result string, duration time.Duration, toBackend int64, fromBackend int64),
	streamFields []zap.Field,
) {
	defer deleteBackConn()
	defer deleteFrontConn()
	var (
		wg          sync.WaitGroup
		toBackend   int64
		fromBackend int64
		result      string
	)

	// when a stream timeout is encountered, no error will be read from errc
	// a buffer size of 2 will allow both the read and write goroutines to
	// send the error and exit
	// see: https://github.com/caddyserver/caddy/issues/7418
	errc := make(chan error, 2)
	spc := switchProtocolCopier{
		user:       conn,
		backend:    backConn,
		wg:         &wg,
		bufferSize: bufferSize,
		sent:       &toBackend,
		received:   &fromBackend,
	}
	wg.Add(2)

	var timeoutc <-chan time.Time
	if streamTimeout > 0 {
		timer := time.NewTimer(streamTimeout)
		defer timer.Stop()
		timeoutc = timer.C
	}

	go spc.copyToBackend(errc)
	go spc.copyFromBackend(errc)
	// Wait for both directions to finish. A clean EOF in one direction is a
	// half-close of that direction only, not the end of the tunnel: the copier
	// propagates it to the destination with CloseWrite and reports nil, so the
	// other direction is left open and pending bytes can still drain. Anything
	// else ends the tunnel: a copy error, a failed CloseWrite, or errCopyDone
	// when the destination has no write half to close.
	//
	// net/http/httputil expresses the same wait as:
	//
	//	err := <-errc
	//	if err == nil {
	//		err = <-errc
	//	}
	//
	// which cannot be used verbatim here because this handler also selects on the
	// stream timeout, so the wait is repeated inside that select instead. Both
	// accept the same sequences; the loop only additionally lets the timeout win
	// at either wait.
	//
	// See https://github.com/caddyserver/caddy/issues/8026, and the same class
	// fixed upstream in net/http/httputil (https://go.dev/issue/35892).
	result = "closed"
copyLoop:
	for range 2 {
		select {
		case err := <-errc:
			if err != nil {
				result = classifyStreamResult(err)
				if c := streamLogger.Check(streamLevel, "streaming error"); c != nil && !errors.Is(err, errCopyDone) {
					c.Write(zap.Error(err))
				}
				break copyLoop
			}
		case t := <-timeoutc:
			result = "timeout"
			if c := streamLogger.Check(streamLevel, "stream timed out"); c != nil {
				c.Write(zap.Time("timeout", t))
			}
			break copyLoop
		}
	}

	// Close both ends to unblock the still-running copy goroutine,
	// then wait for it so byte counts are final before metrics/logging.
	conn.Close()
	backConn.Close()
	wg.Wait()

	finishMetrics(result, time.Since(start), toBackend, fromBackend)
	if c := streamLogger.Check(streamLevel, "connection closed"); c != nil {
		fields := append([]zap.Field{}, streamFields...)
		fields = append(fields,
			zap.Duration("duration", time.Since(start)),
			zap.Int64("bytes_to_backend", toBackend),
			zap.Int64("bytes_from_backend", fromBackend),
		)
		c.Write(fields...)
	}
}

func classifyStreamResult(err error) string {
	if err == nil ||
		errors.Is(err, errCopyDone) ||
		errors.Is(err, io.EOF) ||
		errors.Is(err, net.ErrClosed) ||
		errors.Is(err, context.Canceled) {
		return "closed"
	}
	return "error"
}

func makeStreamLogFields(streamUUID string) []zap.Field {
	fields := make([]zap.Field, 0, 1)
	if streamUUID != "" {
		fields = append(fields, zap.String("uuid", streamUUID))
	}
	return fields
}

// flushInterval returns the p.FlushInterval value, conditionally
// overriding its value for a specific request/response.
func (h Handler) flushInterval(req *http.Request, res *http.Response, incremental bool) time.Duration {
	// The upstream asked for the response to be forwarded incrementally.
	// See RFC 10036: https://www.rfc-editor.org/rfc/rfc10036.html
	if incremental || caddyhttp.IsIncremental(res.Header) {
		return -1 // negative means immediately
	}

	resCTHeader := res.Header.Get("Content-Type")
	resCT, _, err := mime.ParseMediaType(resCTHeader)

	// For Server-Sent Events responses, flush immediately.
	// The MIME type is defined in https://www.w3.org/TR/eventsource/#text-event-stream
	if err == nil && resCT == "text/event-stream" {
		return -1
	}

	// We might have the case of streaming for which Content-Length might be unset.
	if res.ContentLength == -1 {
		return -1
	}

	return time.Duration(h.FlushInterval)
}

func (h Handler) copyResponse(dst http.ResponseWriter, src io.Reader, flushInterval time.Duration, logger *zap.Logger) error {
	var w io.Writer = dst

	if flushInterval != 0 {
		var mlwLogger *zap.Logger
		if h.VerboseLogs {
			mlwLogger = logger.Named("max_latency_writer")
		} else {
			mlwLogger = zap.NewNop()
		}
		mlw := &maxLatencyWriter{
			dst: dst,
			//nolint:bodyclose
			flush:   http.NewResponseController(dst).Flush,
			latency: flushInterval,
			logger:  mlwLogger,
		}
		defer mlw.stop()

		// set up initial timer so headers get flushed even if body writes are delayed
		mlw.flushPending = true
		mlw.t = time.AfterFunc(flushInterval, mlw.delayedFlush)

		w = mlw
	}

	buf := streamingBufPool.Get().(*[]byte)
	defer streamingBufPool.Put(buf)

	var copyLogger *zap.Logger
	if h.VerboseLogs {
		copyLogger = logger
	} else {
		copyLogger = zap.NewNop()
	}

	_, err := h.copyBuffer(w, src, *buf, copyLogger)
	return err
}

// errWritingDownstream wraps errors from writing the response to the client,
// to tell them apart from errors reading the response from the backend.
var errWritingDownstream = errors.New("writing")

// copyBuffer returns any write errors or non-EOF read errors, and the amount
// of bytes written.
func (h Handler) copyBuffer(dst io.Writer, src io.Reader, buf []byte, logger *zap.Logger) (int64, error) {
	if len(buf) == 0 {
		buf = make([]byte, defaultBufferSize)
	}
	var written int64
	for {
		logger.Debug("waiting to read from upstream")
		nr, rerr := src.Read(buf)
		logger := logger.With(zap.Int("read", nr))
		if c := logger.Check(zapcore.DebugLevel, "read from upstream"); c != nil {
			c.Write(zap.Error(rerr))
		}
		if rerr != nil && rerr != io.EOF && rerr != context.Canceled {
			// TODO: this could be useful to know (indeed, it revealed an error in our
			// fastcgi PoC earlier; but it's this single error report here that necessitates
			// a function separate from io.CopyBuffer, since io.CopyBuffer does not distinguish
			// between read or write errors; in a reverse proxy situation, write errors are not
			// something we need to report to the client, but read errors are a problem on our
			// end for sure. so we need to decide what we want.)
			// p.logf("copyBuffer: ReverseProxy read error during body copy: %v", rerr)
			if c := logger.Check(zapcore.ErrorLevel, "reading from backend"); c != nil {
				c.Write(zap.Error(rerr))
			}
		}
		if nr > 0 {
			logger.Debug("writing to downstream")
			nw, werr := dst.Write(buf[:nr])
			if nw > 0 {
				written += int64(nw)
			}
			if c := logger.Check(zapcore.DebugLevel, "wrote to downstream"); c != nil {
				c.Write(
					zap.Int("written", nw),
					zap.Int64("written_total", written),
					zap.Error(werr),
				)
			}
			if werr != nil {
				return written, fmt.Errorf("%w: %w", errWritingDownstream, werr)
			}
			if nr != nw {
				return written, io.ErrShortWrite
			}
		}
		if rerr != nil {
			if rerr == io.EOF {
				return written, nil
			}
			return written, fmt.Errorf("reading: %w", rerr)
		}
	}
}

// openConnection maps an open connection to an optional function for graceful
// close and records which upstream address the connection is proxying to.
// Also tracks whether the connection is detached, which means it should only be
// closed when the upstream is removed from the config, not on every reload.
type openConnection struct {
	conn          io.ReadWriteCloser
	gracefulClose func(deadline time.Time) error
	detached      bool
	upstream      string
}

// tunnelTracker tracks hijacked/upgraded connections for selective cleanup.
// This exists to detach the lifecycle of streaming connections from the proxy
// Handler and config, since we typically want them to survive past config reloads.
// It also allows for selective connection cleanup based on their attachment status.
type tunnelTracker struct {
	connections map[io.ReadWriteCloser]openConnection
	closeTimer  *time.Timer
	closeDelay  time.Duration
	stopped     bool
	mu          sync.Mutex
	logger      *zap.Logger
}

func newTunnelTracker(logger *zap.Logger, closeDelay time.Duration) *tunnelTracker {
	return &tunnelTracker{
		connections: make(map[io.ReadWriteCloser]openConnection),
		closeDelay:  closeDelay,
		logger:      logger,
	}
}

// registerConnection stores conn in the tracking map. The caller must invoke
// the returned del func when the connection is done.
func (ts *tunnelTracker) registerConnection(conn io.ReadWriteCloser, gracefulClose func(deadline time.Time) error, detached bool, upstream string) (del func()) {
	ts.mu.Lock()
	ts.connections[conn] = openConnection{conn, gracefulClose, detached, upstream}
	ts.mu.Unlock()
	return func() {
		ts.mu.Lock()
		delete(ts.connections, conn)
		empty := len(ts.connections) == 0 && ts.stopped
		if empty {
			if ts.closeTimer != nil {
				if ts.closeTimer.Stop() {
					ts.logger.Debug("stopped streaming connections close timer - all connections are already closed")
				}
				ts.closeTimer = nil
			}
		}
		ts.mu.Unlock()
		if empty {
			unregisterDetachedTunnelTrackers(ts)
		}
	}
}

// closeAttachedConnections closes all tracked attached connections.
func (ts *tunnelTracker) closeAttachedConnections() error {
	ts.mu.Lock()
	defer ts.mu.Unlock()
	// detached connections are only closed when the upstream is gone from the config
	return ts.closeConnectionsLocked(time.Now().Add(closeControlWriteTimeout), func(oc openConnection) bool {
		return !oc.detached
	})
}

// closeConnectionsLocked sends a close message to, then closes, every
// tracked connection that match selects. The caller must hold mu. Every
// graceful close shares one absolute deadline: the loop is serial and holds
// the lock, so per-connection timeouts would add up across stalled peers.
// All close messages are sent before any connection is closed, because
// closing one side of a tunnel makes its copy loops close the other side,
// which would lose a close message still waiting to be sent there.
// The first error is returned.
func (ts *tunnelTracker) closeConnectionsLocked(deadline time.Time, match func(openConnection) bool) error {
	var batch []openConnection
	for _, oc := range ts.connections {
		if match(oc) {
			batch = append(batch, oc)
		}
	}

	var err error
	for _, oc := range batch {
		if oc.gracefulClose != nil {
			if gracefulErr := oc.gracefulClose(deadline); gracefulErr != nil && err == nil {
				err = gracefulErr
			}
		}
	}
	for _, oc := range batch {
		// The other side of a tunnel may already have closed this one, which
		// is the expected outcome rather than an error worth logging.
		if closeErr := oc.conn.Close(); closeErr != nil && !errors.Is(closeErr, net.ErrClosed) && err == nil {
			err = closeErr
		}
	}
	return err
}

// cleanupAttachedConnections closes upgraded attached connections.
// Depending on closeDelay it does that either immediately or after a timer.
func (ts *tunnelTracker) cleanupAttachedConnections() error {
	ts.mu.Lock()
	// Mark the tracker stopped before upstream-removal notifications, even
	// when attached connections will be closed after a delay.
	ts.stopped = true
	if len(ts.connections) == 0 {
		ts.mu.Unlock()
		unregisterDetachedTunnelTrackers(ts)
		return nil
	}
	if ts.closeDelay == 0 {
		ts.mu.Unlock()
		return ts.closeAttachedConnections()
	}

	defer ts.mu.Unlock()
	delay := ts.closeDelay
	ts.closeTimer = time.AfterFunc(delay, func() {
		if c := ts.logger.Check(zapcore.DebugLevel, "closing streaming connections after delay"); c != nil {
			c.Write(zap.Duration("delay", delay))
		}
		err := ts.closeAttachedConnections()
		if err != nil {
			if c := ts.logger.Check(zapcore.ErrorLevel, "failed to close connections after delay"); c != nil {
				c.Write(
					zap.Error(err),
					zap.Duration("delay", delay),
				)
			}
		}
	})
	return nil
}

// writeCloseControl sends a best-effort Close control message to the given
// WebSocket connection. Thanks to @pascaldekloe who provided inspiration
// from his simple implementation of this I was able to learn from at:
// github.com/pascaldekloe/websocket. Further work for handling masking
// taken from github.com/gorilla/websocket.
//
// The write must finish by deadline: a peer that stopped reading leaves the
// tunnel's copy goroutine stuck writing, and the close message would wait
// behind it forever while the caller holds the tunnel tracker lock. The
// connection is closed right after, so the deadline needs no reset.
// Connections that cannot set a write deadline are skipped instead.
func writeCloseControl(conn io.Writer, isClient bool, deadline time.Time) error {
	dc, ok := conn.(interface{ SetWriteDeadline(time.Time) error })
	if !ok {
		return nil
	}
	if err := dc.SetWriteDeadline(deadline); err != nil {
		if errors.Is(err, http.ErrNotSupported) {
			return nil
		}
		return err
	}

	// Sources:
	// https://github.com/pascaldekloe/websocket/blob/32050af67a5d/websocket.go#L119
	// https://github.com/gorilla/websocket/blob/v1.5.0/conn.go#L413

	// For now, we're not using a reason. We might later, though.
	// The code handling the reason is left in
	var reason string // max 123 bytes (control frame payload limit is 125; status code takes 2)

	const closeMessage = 8
	const finalBit = 1 << 7 // Frame header byte 0 bits from Section 5.2 of RFC 6455
	const maskBit = 1 << 7  // Frame header byte 1 bits from Section 5.2 of RFC 6455
	const goingAwayUpper uint8 = 1001 >> 8
	const goingAwayLower uint8 = 1001 & 0xff

	b0 := byte(closeMessage) | finalBit
	b1 := byte(len(reason) + 2)
	if isClient {
		b1 |= maskBit
	}

	buf := make([]byte, 0, 127)
	buf = append(buf, b0, b1)
	msgLength := 4 + len(reason)

	// Both branches below append the "going away" code and reason
	appendMessage := func(buf []byte) []byte {
		buf = append(buf, goingAwayUpper, goingAwayLower)
		buf = append(buf, []byte(reason)...)
		return buf
	}

	// When we're the client, we need to mask the message as per
	// https://www.rfc-editor.org/rfc/rfc6455#section-5.3
	if isClient {
		key := newMaskKey()
		buf = append(buf, key[:]...)
		msgLength += len(key)
		buf = appendMessage(buf)
		maskBytes(key, 0, buf[2+len(key):])
	} else {
		buf = appendMessage(buf)
	}

	// simply best-effort, but return error for logging purposes
	// TODO: we might need to ensure we are the exclusive writer by this point (io.Copy is stopped)?
	_, err := conn.Write(buf[:msgLength])
	return err
}

// closeControlWriteTimeout is how long one batch of graceful closes waits
// for peers to accept their close messages before giving up on them. It is
// a variable so tests can shorten it.
var closeControlWriteTimeout = 5 * time.Second

// Copied from https://github.com/gorilla/websocket/blob/v1.5.0/mask.go
func maskBytes(key [4]byte, pos int, b []byte) int {
	// Mask one byte at a time for small buffers.
	if len(b) < 2*wordSize {
		for i := range b {
			b[i] ^= key[pos&3]
			pos++
		}
		return pos & 3
	}

	// Mask one byte at a time to word boundary.
	if n := int(uintptr(unsafe.Pointer(&b[0]))) % wordSize; n != 0 {
		n = wordSize - n
		for i := range b[:n] {
			b[i] ^= key[pos&3]
			pos++
		}
		b = b[n:]
	}

	// Create aligned word size key.
	var k [wordSize]byte
	for i := range k {
		k[i] = key[(pos+i)&3] // nolint:gosec // false positive, impossible to be out of bounds; see: https://github.com/securego/gosec/issues/1525
	}
	kw := *(*uintptr)(unsafe.Pointer(&k))

	// Mask one word at a time.
	n := (len(b) / wordSize) * wordSize
	for i := 0; i < n; i += wordSize {
		*(*uintptr)(unsafe.Add(unsafe.Pointer(&b[0]), i)) ^= kw
	}

	// Mask one byte at a time for remaining bytes.
	b = b[n:]
	for i := range b {
		b[i] ^= key[pos&3]
		pos++
	}

	return pos & 3
}

// Copied from https://github.com/gorilla/websocket/blob/v1.5.0/conn.go#L184
func newMaskKey() [4]byte {
	n := weakrand.Uint32()
	return [4]byte{byte(n), byte(n >> 8), byte(n >> 16), byte(n >> 24)}
}

// isWebsocket returns true if r looks to be an upgrade request for WebSockets.
// It is a fairly naive check.
func isWebsocket(r *http.Request) bool {
	return httpguts.HeaderValuesContainsToken(r.Header["Connection"], "upgrade") &&
		httpguts.HeaderValuesContainsToken(r.Header["Upgrade"], "websocket")
}

// closeConnectionsForUpstream closes detached connections to the given
// upstream on stopped trackers. Attached connections keep their close delay.
// The deadline is shared so one upstream removal is bounded across every
// tracker it notifies.
func (ts *tunnelTracker) closeConnectionsForUpstream(addr string, deadline time.Time) error {
	ts.mu.Lock()
	defer ts.mu.Unlock()
	if !ts.stopped {
		return nil
	}
	return ts.closeConnectionsLocked(deadline, func(oc openConnection) bool {
		return oc.detached && oc.upstream == addr
	})
}

type maxLatencyWriter struct {
	dst     io.Writer
	flush   func() error
	latency time.Duration // non-zero; negative means to flush immediately

	mu           sync.Mutex // protects t, flushPending, and dst.Flush
	t            *time.Timer
	flushPending bool
	logger       *zap.Logger
}

func (m *maxLatencyWriter) Write(p []byte) (n int, err error) {
	m.mu.Lock()
	defer m.mu.Unlock()
	n, err = m.dst.Write(p)
	if c := m.logger.Check(zapcore.DebugLevel, "wrote bytes"); c != nil {
		c.Write(zap.Int("n", n), zap.Error(err))
	}
	if m.latency < 0 {
		m.logger.Debug("flushing immediately")
		//nolint:errcheck
		m.flush()
		return n, err
	}
	if m.flushPending {
		m.logger.Debug("delayed flush already pending")
		return n, err
	}
	if m.t == nil {
		m.t = time.AfterFunc(m.latency, m.delayedFlush)
	} else {
		m.t.Reset(m.latency)
	}
	if c := m.logger.Check(zapcore.DebugLevel, "timer set for delayed flush"); c != nil {
		c.Write(zap.Duration("duration", m.latency))
	}
	m.flushPending = true
	return n, err
}

func (m *maxLatencyWriter) delayedFlush() {
	m.mu.Lock()
	defer m.mu.Unlock()
	if !m.flushPending { // if stop was called but AfterFunc already started this goroutine
		m.logger.Debug("delayed flush is not pending")
		return
	}
	m.logger.Debug("delayed flush")
	//nolint:errcheck
	m.flush()
	m.flushPending = false
}

func (m *maxLatencyWriter) stop() {
	m.mu.Lock()
	defer m.mu.Unlock()
	m.flushPending = false
	if m.t != nil {
		m.t.Stop()
	}
}

// switchProtocolCopier exists so goroutines proxying data back and
// forth have nice names in stacks.
type switchProtocolCopier struct {
	user, backend io.ReadWriteCloser
	wg            *sync.WaitGroup
	bufferSize    int
	// sent and received accumulate byte counts for each direction.
	// They are written before wg.Done() and read after wg.Wait(), so no
	// additional synchronization is needed beyond the WaitGroup barrier.
	sent     *int64 // bytes copied to backend; must be non-nil
	received *int64 // bytes copied from backend; must be non-nil
}

// errCopyDone is reported by a direction that ended at a clean EOF on a
// connection whose write half cannot be closed, so the end of that direction
// could not be propagated to the peer.
var errCopyDone = errors.New("hijacked connection copy complete")

func (c switchProtocolCopier) copyFromBackend(errc chan<- error) {
	defer c.wg.Done()
	n, err := io.CopyBuffer(c.user, c.backend, c.buffer())
	*c.received = n
	if err != nil {
		errc <- err
		return
	}

	// backend conn has reached EOF so propagate close write to user conn
	if wc, ok := c.user.(interface{ CloseWrite() error }); ok {
		errc <- wc.CloseWrite()
		return
	}

	errc <- errCopyDone
}

func (c switchProtocolCopier) copyToBackend(errc chan<- error) {
	defer c.wg.Done()
	n, err := io.CopyBuffer(c.backend, c.user, c.buffer())
	*c.sent = n
	if err != nil {
		errc <- err
		return
	}

	// user conn has reached EOF so propagate close write to backend conn
	if wc, ok := c.backend.(interface{ CloseWrite() error }); ok {
		errc <- wc.CloseWrite()
		return
	}

	errc <- errCopyDone
}

func (c switchProtocolCopier) buffer() []byte {
	size := c.bufferSize
	if size <= 0 {
		size = defaultBufferSize
	}
	return make([]byte, size)
}

var streamingBufPool = sync.Pool{
	New: func() any {
		// The Pool's New function should generally only return pointer
		// types, since a pointer can be put into the return interface
		// value without an allocation
		// - (from the package docs)
		b := make([]byte, defaultBufferSize)
		return &b
	},
}

const (
	defaultBufferSize = 32 * 1024
	wordSize          = int(unsafe.Sizeof(uintptr(0)))
)