reverseproxy.go

Documentation: net/http/httputil

     1  // Copyright 2011 The Go Authors. All rights reserved.
     2  // Use of this source code is governed by a BSD-style
     3  // license that can be found in the LICENSE file.
     4  
     5  // HTTP reverse proxy handler
     6  
     7  package httputil
     8  
     9  import (
    10  	"context"
    11  	"fmt"
    12  	"io"
    13  	"log"
    14  	"net"
    15  	"net/http"
    16  	"net/textproto"
    17  	"net/url"
    18  	"strings"
    19  	"sync"
    20  	"time"
    21  
    22  	"golang.org/x/net/http/httpguts"
    23  )
    24  
    25  // ReverseProxy is an HTTP Handler that takes an incoming request and
    26  // sends it to another server, proxying the response back to the
    27  // client.
    28  //
    29  // ReverseProxy by default sets the client IP as the value of the
    30  // X-Forwarded-For header.
    31  //
    32  // If an X-Forwarded-For header already exists, the client IP is
    33  // appended to the existing values. As a special case, if the header
    34  // exists in the Request.Header map but has a nil value (such as when
    35  // set by the Director func), the X-Forwarded-For header is
    36  // not modified.
    37  //
    38  // To prevent IP spoofing, be sure to delete any pre-existing
    39  // X-Forwarded-For header coming from the client or
    40  // an untrusted proxy.
    41  type ReverseProxy struct {
    42  	// Director must be a function which modifies
    43  	// the request into a new request to be sent
    44  	// using Transport. Its response is then copied
    45  	// back to the original client unmodified.
    46  	// Director must not access the provided Request
    47  	// after returning.
    48  	Director func(*http.Request)
    49  
    50  	// The transport used to perform proxy requests.
    51  	// If nil, http.DefaultTransport is used.
    52  	Transport http.RoundTripper
    53  
    54  	// FlushInterval specifies the flush interval
    55  	// to flush to the client while copying the
    56  	// response body.
    57  	// If zero, no periodic flushing is done.
    58  	// A negative value means to flush immediately
    59  	// after each write to the client.
    60  	// The FlushInterval is ignored when ReverseProxy
    61  	// recognizes a response as a streaming response, or
    62  	// if its ContentLength is -1; for such responses, writes
    63  	// are flushed to the client immediately.
    64  	FlushInterval time.Duration
    65  
    66  	// ErrorLog specifies an optional logger for errors
    67  	// that occur when attempting to proxy the request.
    68  	// If nil, logging is done via the log package's standard logger.
    69  	ErrorLog *log.Logger
    70  
    71  	// BufferPool optionally specifies a buffer pool to
    72  	// get byte slices for use by io.CopyBuffer when
    73  	// copying HTTP response bodies.
    74  	BufferPool BufferPool
    75  
    76  	// ModifyResponse is an optional function that modifies the
    77  	// Response from the backend. It is called if the backend
    78  	// returns a response at all, with any HTTP status code.
    79  	// If the backend is unreachable, the optional ErrorHandler is
    80  	// called without any call to ModifyResponse.
    81  	//
    82  	// If ModifyResponse returns an error, ErrorHandler is called
    83  	// with its error value. If ErrorHandler is nil, its default
    84  	// implementation is used.
    85  	ModifyResponse func(*http.Response) error
    86  
    87  	// ErrorHandler is an optional function that handles errors
    88  	// reaching the backend or errors from ModifyResponse.
    89  	//
    90  	// If nil, the default is to log the provided error and return
    91  	// a 502 Status Bad Gateway response.
    92  	ErrorHandler func(http.ResponseWriter, *http.Request, error)
    93  }
    94  
    95  // A BufferPool is an interface for getting and returning temporary
    96  // byte slices for use by io.CopyBuffer.
    97  type BufferPool interface {
    98  	Get() []byte
    99  	Put([]byte)
   100  }
   101  
   102  func singleJoiningSlash(a, b string) string {
   103  	aslash := strings.HasSuffix(a, "/")
   104  	bslash := strings.HasPrefix(b, "/")
   105  	switch {
   106  	case aslash && bslash:
   107  		return a + b[1:]
   108  	case !aslash && !bslash:
   109  		return a + "/" + b
   110  	}
   111  	return a + b
   112  }
   113  
   114  func joinURLPath(a, b *url.URL) (path, rawpath string) {
   115  	if a.RawPath == "" && b.RawPath == "" {
   116  		return singleJoiningSlash(a.Path, b.Path), ""
   117  	}
   118  	// Same as singleJoiningSlash, but uses EscapedPath to determine
   119  	// whether a slash should be added
   120  	apath := a.EscapedPath()
   121  	bpath := b.EscapedPath()
   122  
   123  	aslash := strings.HasSuffix(apath, "/")
   124  	bslash := strings.HasPrefix(bpath, "/")
   125  
   126  	switch {
   127  	case aslash && bslash:
   128  		return a.Path + b.Path[1:], apath + bpath[1:]
   129  	case !aslash && !bslash:
   130  		return a.Path + "/" + b.Path, apath + "/" + bpath
   131  	}
   132  	return a.Path + b.Path, apath + bpath
   133  }
   134  
   135  // NewSingleHostReverseProxy returns a new ReverseProxy that routes
   136  // URLs to the scheme, host, and base path provided in target. If the
   137  // target's path is "/base" and the incoming request was for "/dir",
   138  // the target request will be for /base/dir.
   139  // NewSingleHostReverseProxy does not rewrite the Host header.
   140  // To rewrite Host headers, use ReverseProxy directly with a custom
   141  // Director policy.
   142  func NewSingleHostReverseProxy(target *url.URL) *ReverseProxy {
   143  	targetQuery := target.RawQuery
   144  	director := func(req *http.Request) {
   145  		req.URL.Scheme = target.Scheme
   146  		req.URL.Host = target.Host
   147  		req.URL.Path, req.URL.RawPath = joinURLPath(target, req.URL)
   148  		if targetQuery == "" || req.URL.RawQuery == "" {
   149  			req.URL.RawQuery = targetQuery + req.URL.RawQuery
   150  		} else {
   151  			req.URL.RawQuery = targetQuery + "&" + req.URL.RawQuery
   152  		}
   153  		if _, ok := req.Header["User-Agent"]; !ok {
   154  			// explicitly disable User-Agent so it's not set to default value
   155  			req.Header.Set("User-Agent", "")
   156  		}
   157  	}
   158  	return &ReverseProxy{Director: director}
   159  }
   160  
   161  func copyHeader(dst, src http.Header) {
   162  	for k, vv := range src {
   163  		for _, v := range vv {
   164  			dst.Add(k, v)
   165  		}
   166  	}
   167  }
   168  
   169  // Hop-by-hop headers. These are removed when sent to the backend.
   170  // As of RFC 7230, hop-by-hop headers are required to appear in the
   171  // Connection header field. These are the headers defined by the
   172  // obsoleted RFC 2616 (section 13.5.1) and are used for backward
   173  // compatibility.
   174  var hopHeaders = []string{
   175  	"Connection",
   176  	"Proxy-Connection", // non-standard but still sent by libcurl and rejected by e.g. google
   177  	"Keep-Alive",
   178  	"Proxy-Authenticate",
   179  	"Proxy-Authorization",
   180  	"Te",      // canonicalized version of "TE"
   181  	"Trailer", // not Trailers per URL above; https://www.rfc-editor.org/errata_search.php?eid=4522
   182  	"Transfer-Encoding",
   183  	"Upgrade",
   184  }
   185  
   186  func (p *ReverseProxy) defaultErrorHandler(rw http.ResponseWriter, req *http.Request, err error) {
   187  	p.logf("http: proxy error: %v", err)
   188  	rw.WriteHeader(http.StatusBadGateway)
   189  }
   190  
   191  func (p *ReverseProxy) getErrorHandler() func(http.ResponseWriter, *http.Request, error) {
   192  	if p.ErrorHandler != nil {
   193  		return p.ErrorHandler
   194  	}
   195  	return p.defaultErrorHandler
   196  }
   197  
   198  // modifyResponse conditionally runs the optional ModifyResponse hook
   199  // and reports whether the request should proceed.
   200  func (p *ReverseProxy) modifyResponse(rw http.ResponseWriter, res *http.Response, req *http.Request) bool {
   201  	if p.ModifyResponse == nil {
   202  		return true
   203  	}
   204  	if err := p.ModifyResponse(res); err != nil {
   205  		res.Body.Close()
   206  		p.getErrorHandler()(rw, req, err)
   207  		return false
   208  	}
   209  	return true
   210  }
   211  
   212  func (p *ReverseProxy) ServeHTTP(rw http.ResponseWriter, req *http.Request) {
   213  	transport := p.Transport
   214  	if transport == nil {
   215  		transport = http.DefaultTransport
   216  	}
   217  
   218  	ctx := req.Context()
   219  	if cn, ok := rw.(http.CloseNotifier); ok {
   220  		var cancel context.CancelFunc
   221  		ctx, cancel = context.WithCancel(ctx)
   222  		defer cancel()
   223  		notifyChan := cn.CloseNotify()
   224  		go func() {
   225  			select {
   226  			case <-notifyChan:
   227  				cancel()
   228  			case <-ctx.Done():
   229  			}
   230  		}()
   231  	}
   232  
   233  	outreq := req.Clone(ctx)
   234  	if req.ContentLength == 0 {
   235  		outreq.Body = nil // Issue 16036: nil Body for http.Transport retries
   236  	}
   237  	if outreq.Header == nil {
   238  		outreq.Header = make(http.Header) // Issue 33142: historical behavior was to always allocate
   239  	}
   240  
   241  	p.Director(outreq)
   242  	outreq.Close = false
   243  
   244  	reqUpType := upgradeType(outreq.Header)
   245  	removeConnectionHeaders(outreq.Header)
   246  
   247  	// Remove hop-by-hop headers to the backend. Especially
   248  	// important is "Connection" because we want a persistent
   249  	// connection, regardless of what the client sent to us.
   250  	for _, h := range hopHeaders {
   251  		outreq.Header.Del(h)
   252  	}
   253  
   254  	// Issue 21096: tell backend applications that care about trailer support
   255  	// that we support trailers. (We do, but we don't go out of our way to
   256  	// advertise that unless the incoming client request thought it was worth
   257  	// mentioning.) Note that we look at req.Header, not outreq.Header, since
   258  	// the latter has passed through removeConnectionHeaders.
   259  	if httpguts.HeaderValuesContainsToken(req.Header["Te"], "trailers") {
   260  		outreq.Header.Set("Te", "trailers")
   261  	}
   262  
   263  	// After stripping all the hop-by-hop connection headers above, add back any
   264  	// necessary for protocol upgrades, such as for websockets.
   265  	if reqUpType != "" {
   266  		outreq.Header.Set("Connection", "Upgrade")
   267  		outreq.Header.Set("Upgrade", reqUpType)
   268  	}
   269  
   270  	if clientIP, _, err := net.SplitHostPort(req.RemoteAddr); err == nil {
   271  		// If we aren't the first proxy retain prior
   272  		// X-Forwarded-For information as a comma+space
   273  		// separated list and fold multiple headers into one.
   274  		prior, ok := outreq.Header["X-Forwarded-For"]
   275  		omit := ok && prior == nil // Issue 38079: nil now means don't populate the header
   276  		if len(prior) > 0 {
   277  			clientIP = strings.Join(prior, ", ") + ", " + clientIP
   278  		}
   279  		if !omit {
   280  			outreq.Header.Set("X-Forwarded-For", clientIP)
   281  		}
   282  	}
   283  
   284  	res, err := transport.RoundTrip(outreq)
   285  	if err != nil {
   286  		p.getErrorHandler()(rw, outreq, err)
   287  		return
   288  	}
   289  
   290  	// Deal with 101 Switching Protocols responses: (WebSocket, h2c, etc)
   291  	if res.StatusCode == http.StatusSwitchingProtocols {
   292  		if !p.modifyResponse(rw, res, outreq) {
   293  			return
   294  		}
   295  		p.handleUpgradeResponse(rw, outreq, res)
   296  		return
   297  	}
   298  
   299  	removeConnectionHeaders(res.Header)
   300  
   301  	for _, h := range hopHeaders {
   302  		res.Header.Del(h)
   303  	}
   304  
   305  	if !p.modifyResponse(rw, res, outreq) {
   306  		return
   307  	}
   308  
   309  	copyHeader(rw.Header(), res.Header)
   310  
   311  	// The "Trailer" header isn't included in the Transport's response,
   312  	// at least for *http.Transport. Build it up from Trailer.
   313  	announcedTrailers := len(res.Trailer)
   314  	if announcedTrailers > 0 {
   315  		trailerKeys := make([]string, 0, len(res.Trailer))
   316  		for k := range res.Trailer {
   317  			trailerKeys = append(trailerKeys, k)
   318  		}
   319  		rw.Header().Add("Trailer", strings.Join(trailerKeys, ", "))
   320  	}
   321  
   322  	rw.WriteHeader(res.StatusCode)
   323  
   324  	err = p.copyResponse(rw, res.Body, p.flushInterval(res))
   325  	if err != nil {
   326  		defer res.Body.Close()
   327  		// Since we're streaming the response, if we run into an error all we can do
   328  		// is abort the request. Issue 23643: ReverseProxy should use ErrAbortHandler
   329  		// on read error while copying body.
   330  		if !shouldPanicOnCopyError(req) {
   331  			p.logf("suppressing panic for copyResponse error in test; copy error: %v", err)
   332  			return
   333  		}
   334  		panic(http.ErrAbortHandler)
   335  	}
   336  	res.Body.Close() // close now, instead of defer, to populate res.Trailer
   337  
   338  	if len(res.Trailer) > 0 {
   339  		// Force chunking if we saw a response trailer.
   340  		// This prevents net/http from calculating the length for short
   341  		// bodies and adding a Content-Length.
   342  		if fl, ok := rw.(http.Flusher); ok {
   343  			fl.Flush()
   344  		}
   345  	}
   346  
   347  	if len(res.Trailer) == announcedTrailers {
   348  		copyHeader(rw.Header(), res.Trailer)
   349  		return
   350  	}
   351  
   352  	for k, vv := range res.Trailer {
   353  		k = http.TrailerPrefix + k
   354  		for _, v := range vv {
   355  			rw.Header().Add(k, v)
   356  		}
   357  	}
   358  }
   359  
   360  var inOurTests bool // whether we're in our own tests
   361  
   362  // shouldPanicOnCopyError reports whether the reverse proxy should
   363  // panic with http.ErrAbortHandler. This is the right thing to do by
   364  // default, but Go 1.10 and earlier did not, so existing unit tests
   365  // weren't expecting panics. Only panic in our own tests, or when
   366  // running under the HTTP server.
   367  func shouldPanicOnCopyError(req *http.Request) bool {
   368  	if inOurTests {
   369  		// Our tests know to handle this panic.
   370  		return true
   371  	}
   372  	if req.Context().Value(http.ServerContextKey) != nil {
   373  		// We seem to be running under an HTTP server, so
   374  		// it'll recover the panic.
   375  		return true
   376  	}
   377  	// Otherwise act like Go 1.10 and earlier to not break
   378  	// existing tests.
   379  	return false
   380  }
   381  
   382  // removeConnectionHeaders removes hop-by-hop headers listed in the "Connection" header of h.
   383  // See RFC 7230, section 6.1
   384  func removeConnectionHeaders(h http.Header) {
   385  	for _, f := range h["Connection"] {
   386  		for _, sf := range strings.Split(f, ",") {
   387  			if sf = textproto.TrimString(sf); sf != "" {
   388  				h.Del(sf)
   389  			}
   390  		}
   391  	}
   392  }
   393  
   394  // flushInterval returns the p.FlushInterval value, conditionally
   395  // overriding its value for a specific request/response.
   396  func (p *ReverseProxy) flushInterval(res *http.Response) time.Duration {
   397  	resCT := res.Header.Get("Content-Type")
   398  
   399  	// For Server-Sent Events responses, flush immediately.
   400  	// The MIME type is defined in https://www.w3.org/TR/eventsource/#text-event-stream
   401  	if resCT == "text/event-stream" {
   402  		return -1 // negative means immediately
   403  	}
   404  
   405  	// We might have the case of streaming for which Content-Length might be unset.
   406  	if res.ContentLength == -1 {
   407  		return -1
   408  	}
   409  
   410  	return p.FlushInterval
   411  }
   412  
   413  func (p *ReverseProxy) copyResponse(dst io.Writer, src io.Reader, flushInterval time.Duration) error {
   414  	if flushInterval != 0 {
   415  		if wf, ok := dst.(writeFlusher); ok {
   416  			mlw := &maxLatencyWriter{
   417  				dst:     wf,
   418  				latency: flushInterval,
   419  			}
   420  			defer mlw.stop()
   421  
   422  			// set up initial timer so headers get flushed even if body writes are delayed
   423  			mlw.flushPending = true
   424  			mlw.t = time.AfterFunc(flushInterval, mlw.delayedFlush)
   425  
   426  			dst = mlw
   427  		}
   428  	}
   429  
   430  	var buf []byte
   431  	if p.BufferPool != nil {
   432  		buf = p.BufferPool.Get()
   433  		defer p.BufferPool.Put(buf)
   434  	}
   435  	_, err := p.copyBuffer(dst, src, buf)
   436  	return err
   437  }
   438  
   439  // copyBuffer returns any write errors or non-EOF read errors, and the amount
   440  // of bytes written.
   441  func (p *ReverseProxy) copyBuffer(dst io.Writer, src io.Reader, buf []byte) (int64, error) {
   442  	if len(buf) == 0 {
   443  		buf = make([]byte, 32*1024)
   444  	}
   445  	var written int64
   446  	for {
   447  		nr, rerr := src.Read(buf)
   448  		if rerr != nil && rerr != io.EOF && rerr != context.Canceled {
   449  			p.logf("httputil: ReverseProxy read error during body copy: %v", rerr)
   450  		}
   451  		if nr > 0 {
   452  			nw, werr := dst.Write(buf[:nr])
   453  			if nw > 0 {
   454  				written += int64(nw)
   455  			}
   456  			if werr != nil {
   457  				return written, werr
   458  			}
   459  			if nr != nw {
   460  				return written, io.ErrShortWrite
   461  			}
   462  		}
   463  		if rerr != nil {
   464  			if rerr == io.EOF {
   465  				rerr = nil
   466  			}
   467  			return written, rerr
   468  		}
   469  	}
   470  }
   471  
   472  func (p *ReverseProxy) logf(format string, args ...interface{}) {
   473  	if p.ErrorLog != nil {
   474  		p.ErrorLog.Printf(format, args...)
   475  	} else {
   476  		log.Printf(format, args...)
   477  	}
   478  }
   479  
   480  type writeFlusher interface {
   481  	io.Writer
   482  	http.Flusher
   483  }
   484  
   485  type maxLatencyWriter struct {
   486  	dst     writeFlusher
   487  	latency time.Duration // non-zero; negative means to flush immediately
   488  
   489  	mu           sync.Mutex // protects t, flushPending, and dst.Flush
   490  	t            *time.Timer
   491  	flushPending bool
   492  }
   493  
   494  func (m *maxLatencyWriter) Write(p []byte) (n int, err error) {
   495  	m.mu.Lock()
   496  	defer m.mu.Unlock()
   497  	n, err = m.dst.Write(p)
   498  	if m.latency < 0 {
   499  		m.dst.Flush()
   500  		return
   501  	}
   502  	if m.flushPending {
   503  		return
   504  	}
   505  	if m.t == nil {
   506  		m.t = time.AfterFunc(m.latency, m.delayedFlush)
   507  	} else {
   508  		m.t.Reset(m.latency)
   509  	}
   510  	m.flushPending = true
   511  	return
   512  }
   513  
   514  func (m *maxLatencyWriter) delayedFlush() {
   515  	m.mu.Lock()
   516  	defer m.mu.Unlock()
   517  	if !m.flushPending { // if stop was called but AfterFunc already started this goroutine
   518  		return
   519  	}
   520  	m.dst.Flush()
   521  	m.flushPending = false
   522  }
   523  
   524  func (m *maxLatencyWriter) stop() {
   525  	m.mu.Lock()
   526  	defer m.mu.Unlock()
   527  	m.flushPending = false
   528  	if m.t != nil {
   529  		m.t.Stop()
   530  	}
   531  }
   532  
   533  func upgradeType(h http.Header) string {
   534  	if !httpguts.HeaderValuesContainsToken(h["Connection"], "Upgrade") {
   535  		return ""
   536  	}
   537  	return strings.ToLower(h.Get("Upgrade"))
   538  }
   539  
   540  func (p *ReverseProxy) handleUpgradeResponse(rw http.ResponseWriter, req *http.Request, res *http.Response) {
   541  	reqUpType := upgradeType(req.Header)
   542  	resUpType := upgradeType(res.Header)
   543  	if reqUpType != resUpType {
   544  		p.getErrorHandler()(rw, req, fmt.Errorf("backend tried to switch protocol %q when %q was requested", resUpType, reqUpType))
   545  		return
   546  	}
   547  
   548  	hj, ok := rw.(http.Hijacker)
   549  	if !ok {
   550  		p.getErrorHandler()(rw, req, fmt.Errorf("can't switch protocols using non-Hijacker ResponseWriter type %T", rw))
   551  		return
   552  	}
   553  	backConn, ok := res.Body.(io.ReadWriteCloser)
   554  	if !ok {
   555  		p.getErrorHandler()(rw, req, fmt.Errorf("internal error: 101 switching protocols response with non-writable body"))
   556  		return
   557  	}
   558  
   559  	backConnCloseCh := make(chan bool)
   560  	go func() {
   561  		// Ensure that the cancelation of a request closes the backend.
   562  		// See issue https://golang.org/issue/35559.
   563  		select {
   564  		case <-req.Context().Done():
   565  		case <-backConnCloseCh:
   566  		}
   567  		backConn.Close()
   568  	}()
   569  
   570  	defer close(backConnCloseCh)
   571  
   572  	conn, brw, err := hj.Hijack()
   573  	if err != nil {
   574  		p.getErrorHandler()(rw, req, fmt.Errorf("Hijack failed on protocol switch: %v", err))
   575  		return
   576  	}
   577  	defer conn.Close()
   578  
   579  	copyHeader(rw.Header(), res.Header)
   580  
   581  	res.Header = rw.Header()
   582  	res.Body = nil // so res.Write only writes the headers; we have res.Body in backConn above
   583  	if err := res.Write(brw); err != nil {
   584  		p.getErrorHandler()(rw, req, fmt.Errorf("response write: %v", err))
   585  		return
   586  	}
   587  	if err := brw.Flush(); err != nil {
   588  		p.getErrorHandler()(rw, req, fmt.Errorf("response flush: %v", err))
   589  		return
   590  	}
   591  	errc := make(chan error, 1)
   592  	spc := switchProtocolCopier{user: conn, backend: backConn}
   593  	go spc.copyToBackend(errc)
   594  	go spc.copyFromBackend(errc)
   595  	<-errc
   596  	return
   597  }
   598  
   599  // switchProtocolCopier exists so goroutines proxying data back and
   600  // forth have nice names in stacks.
   601  type switchProtocolCopier struct {
   602  	user, backend io.ReadWriter
   603  }
   604  
   605  func (c switchProtocolCopier) copyFromBackend(errc chan<- error) {
   606  	_, err := io.Copy(c.user, c.backend)
   607  	errc <- err
   608  }
   609  
   610  func (c switchProtocolCopier) copyToBackend(errc chan<- error) {
   611  	_, err := io.Copy(c.backend, c.user)
   612  	errc <- err
   613  }
   614
View as plain text