tailscale/tsweb/tsweb.go

// Copyright (c) Tailscale Inc & AUTHORS
// SPDX-License-Identifier: BSD-3-Clause

// Package tsweb contains code used in various Tailscale webservers.
package tsweb

import (
	"bufio"
	"bytes"
	"cmp"
	"context"
	"errors"
	"expvar"
	"fmt"
	"net"
	"net/http"
	_ "net/http/pprof"
	"net/netip"
	"net/url"
	"os"
	"path/filepath"
	"regexp"
	"strconv"
	"strings"
	"sync"
	"time"

	"go4.org/mem"
	"tailscale.com/envknob"
	"tailscale.com/metrics"
	"tailscale.com/net/tsaddr"
	"tailscale.com/tsweb/varz"
	"tailscale.com/types/logger"
	"tailscale.com/util/vizerror"
)

// DevMode controls whether extra output in shown, for when the binary is being run in dev mode.
var DevMode bool

func DefaultCertDir(leafDir string) string {
	cacheDir, err := os.UserCacheDir()
	if err == nil {
		return filepath.Join(cacheDir, "tailscale", leafDir)
	}
	return ""
}

// IsProd443 reports whether addr is a Go listen address for port 443.
func IsProd443(addr string) bool {
	_, port, _ := net.SplitHostPort(addr)
	return port == "443" || port == "https"
}

// AllowDebugAccess reports whether r should be permitted to access
// various debug endpoints.
func AllowDebugAccess(r *http.Request) bool {
	if allowDebugAccessWithKey(r) {
		return true
	}
	if r.Header.Get("X-Forwarded-For") != "" {
		// TODO if/when needed. For now, conservative:
		return false
	}
	ipStr, _, err := net.SplitHostPort(r.RemoteAddr)
	if err != nil {
		return false
	}
	ip, err := netip.ParseAddr(ipStr)
	if err != nil {
		return false
	}
	if tsaddr.IsTailscaleIP(ip) || ip.IsLoopback() || ipStr == envknob.String("TS_ALLOW_DEBUG_IP") {
		return true
	}
	return false
}

func allowDebugAccessWithKey(r *http.Request) bool {
	if r.Method != "GET" {
		return false
	}
	urlKey := r.FormValue("debugkey")
	keyPath := envknob.String("TS_DEBUG_KEY_PATH")
	if urlKey != "" && keyPath != "" {
		slurp, err := os.ReadFile(keyPath)
		if err == nil && string(bytes.TrimSpace(slurp)) == urlKey {
			return true
		}
	}
	return false
}

// AcceptsEncoding reports whether r accepts the named encoding
// ("gzip", "br", etc).
func AcceptsEncoding(r *http.Request, enc string) bool {
	h := r.Header.Get("Accept-Encoding")
	if h == "" {
		return false
	}
	if !strings.Contains(h, enc) && !mem.ContainsFold(mem.S(h), mem.S(enc)) {
		return false
	}
	remain := h
	for len(remain) > 0 {
		var part string
		part, remain, _ = strings.Cut(remain, ",")
		part = strings.TrimSpace(part)
		part, _, _ = strings.Cut(part, ";")
		if part == enc {
			return true
		}
	}
	return false
}

// Protected wraps a provided debug handler, h, returning a Handler
// that enforces AllowDebugAccess and returns forbidden replies for
// unauthorized requests.
func Protected(h http.Handler) http.Handler {
	return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
		if !AllowDebugAccess(r) {
			msg := "debug access denied"
			if DevMode {
				ipStr, _, _ := net.SplitHostPort(r.RemoteAddr)
				msg += fmt.Sprintf("; to permit access, set TS_ALLOW_DEBUG_IP=%v", ipStr)
			}
			http.Error(w, msg, http.StatusForbidden)
			return
		}
		h.ServeHTTP(w, r)
	})
}

// Port80Handler is the handler to be given to
// autocert.Manager.HTTPHandler.  The inner handler is the mux
// returned by NewMux containing registered /debug handlers.
type Port80Handler struct {
	Main http.Handler
	// FQDN is used to redirect incoming requests to https://<FQDN>.
	// If it is not set, the hostname is calculated from the incoming
	// request.
	FQDN string
}

func (h Port80Handler) ServeHTTP(w http.ResponseWriter, r *http.Request) {
	path := r.RequestURI
	if path == "/debug" || strings.HasPrefix(path, "/debug") {
		h.Main.ServeHTTP(w, r)
		return
	}
	if r.Method != "GET" && r.Method != "HEAD" {
		http.Error(w, "Use HTTPS", http.StatusBadRequest)
		return
	}
	if path == "/" && AllowDebugAccess(r) {
		// Redirect authorized user to the debug handler.
		path = "/debug/"
	}
	host := cmp.Or(h.FQDN, r.Host)
	target := "https://" + host + path
	http.Redirect(w, r, target, http.StatusFound)
}

// ReturnHandler is like net/http.Handler, but the handler can return an
// error instead of writing to its ResponseWriter.
type ReturnHandler interface {
	// ServeHTTPReturn is like http.Handler.ServeHTTP, except that
	// it can choose to return an error instead of writing to its
	// http.ResponseWriter.
	//
	// If ServeHTTPReturn returns an error, it caller should handle
	// an error by serving an HTTP 500 response to the user. The
	// error details should not be sent to the client, as they may
	// contain sensitive information. If the error is an
	// HTTPError, though, callers should use the HTTP response
	// code and message as the response to the client.
	ServeHTTPReturn(http.ResponseWriter, *http.Request) error
}

// BucketedStatsOptions describes tsweb handler options surrounding
// the generation of metrics, grouped into buckets.
type BucketedStatsOptions struct {
	// Bucket returns which bucket the given request is in.
	// If nil, [NormalizedPath] is used to compute the bucket.
	Bucket func(req *http.Request) string

	// If non-nil, Started maintains a counter of all requests which
	// have begun processing.
	Started *metrics.LabelMap

	// If non-nil, Finished maintains a counter of all requests which
	// have finished processing with success (that is, the HTTP handler has
	// returned).
	Finished *metrics.LabelMap
}

// normalizePathRegex matches components in a HTTP request path
// that should be replaced.
//
// See: https://regex101.com/r/WIfpaR/3 for the explainer and test cases.
var normalizePathRegex = regexp.MustCompile("([a-fA-F0-9]{9,}|([^\\/])+\\.([^\\/]){2,}|((n|k|u|L|t|S)[a-zA-Z0-9]{5,}(CNTRL|Djz1H|LV5CY|mxgaY|jNy1b))|(([^\\/])+\\@passkey))")

// NormalizedPath returns the given path with the following modifications:
//
//   - any query parameters are removed
//   - any path component with a hex string of 9 or more characters is
//     replaced by an ellipsis
//   - any path component containing a period with at least two characters
//     after the period (i.e. an email or domain)
//   - any path component consisting of a common Tailscale Stable ID
//   - any path segment *@passkey.
func NormalizedPath(p string) string {
	// Fastpath: No hex sequences in there we might have to trim.
	// Avoids allocating.
	if normalizePathRegex.FindStringIndex(p) == nil {
		b, _, _ := strings.Cut(p, "?")
		return b
	}

	// If we got here, there's at least one hex sequences we need to
	// replace with an ellipsis.
	replaced := normalizePathRegex.ReplaceAllString(p, "…")
	b, _, _ := strings.Cut(replaced, "?")
	return b
}

func (o *BucketedStatsOptions) bucketForRequest(r *http.Request) string {
	if o.Bucket != nil {
		return o.Bucket(r)
	}

	return NormalizedPath(r.URL.Path)
}

type HandlerOptions struct {
	QuietLoggingIfSuccessful bool // if set, do not log successfully handled HTTP requests (200 and 304 status codes)
	Logf                     logger.Logf
	Now                      func() time.Time // if nil, defaults to time.Now

	// If non-nil, StatusCodeCounters maintains counters
	// of status codes for handled responses.
	// The keys are "1xx", "2xx", "3xx", "4xx", and "5xx".
	StatusCodeCounters *expvar.Map
	// If non-nil, StatusCodeCountersFull maintains counters of status
	// codes for handled responses.
	// The keys are HTTP numeric response codes e.g. 200, 404, ...
	StatusCodeCountersFull *expvar.Map

	// If non-nil, BucketedStats computes and exposes statistics
	// for each bucket based on the contained parameters.
	BucketedStats *BucketedStatsOptions

	// OnError is called if the handler returned a HTTPError. This
	// is intended to be used to present pretty error pages if
	// the user agent is determined to be a browser.
	OnError ErrorHandlerFunc

	// OnCompletion is called when ServeHTTP is finished and gets
	// useful data that the implementor can use for metrics.
	OnCompletion OnCompletionFunc
}

// ErrorHandlerFunc is called to present a error response.
type ErrorHandlerFunc func(http.ResponseWriter, *http.Request, HTTPError)

// OnCompletionFunc is called when ServeHTTP is finished and gets
// useful data that the implementor can use for metrics.
type OnCompletionFunc func(*http.Request, AccessLogRecord)

// ReturnHandlerFunc is an adapter to allow the use of ordinary
// functions as ReturnHandlers. If f is a function with the
// appropriate signature, ReturnHandlerFunc(f) is a ReturnHandler that
// calls f.
type ReturnHandlerFunc func(http.ResponseWriter, *http.Request) error

// A Middleware is a function that wraps an http.Handler to extend or modify
// its behaviour.
//
// The implementation of the wrapper is responsible for delegating its input
// request to the underlying handler, if appropriate.
type Middleware func(h http.Handler) http.Handler

// ServeHTTPReturn calls f(w, r).
func (f ReturnHandlerFunc) ServeHTTPReturn(w http.ResponseWriter, r *http.Request) error {
	return f(w, r)
}

// StdHandler converts a ReturnHandler into a standard http.Handler.
// Handled requests are logged using opts.Logf, as are any errors.
// Errors are handled as specified by the Handler interface.
func StdHandler(h ReturnHandler, opts HandlerOptions) http.Handler {
	if opts.Now == nil {
		opts.Now = time.Now
	}
	if opts.Logf == nil {
		opts.Logf = logger.Discard
	}
	return retHandler{h, opts}
}

// retHandler is an http.Handler that wraps a Handler and handles errors.
type retHandler struct {
	rh   ReturnHandler
	opts HandlerOptions
}

// ServeHTTP implements the http.Handler interface.
func (h retHandler) ServeHTTP(w http.ResponseWriter, r *http.Request) {
	msg := AccessLogRecord{
		Time:       h.opts.Now(),
		RemoteAddr: r.RemoteAddr,
		Proto:      r.Proto,
		TLS:        r.TLS != nil,
		Host:       r.Host,
		Method:     r.Method,
		RequestURI: r.URL.RequestURI(),
		UserAgent:  r.UserAgent(),
		Referer:    r.Referer(),
		RequestID:  RequestIDFromContext(r.Context()),
	}

	var bucket string
	var startRecorded bool
	if bs := h.opts.BucketedStats; bs != nil {
		bucket = bs.bucketForRequest(r)
		if bs.Started != nil {
			switch v := bs.Started.Map.Get(bucket).(type) {
			case *expvar.Int:
				// If we've already seen this bucket for, count it immediately.
				// Otherwise, for newly seen paths, only count retroactively
				// (so started-finished doesn't go negative) so we don't fill
				// this LabelMap up with internet scanning spam.
				v.Add(1)
				startRecorded = true
			}
		}
	}

	lw := &loggingResponseWriter{ResponseWriter: w, logf: h.opts.Logf}

	// In case the handler panics, we want to recover and continue logging the
	// error before raising the panic again for the server to handle.
	var (
		didPanic bool
		panicRes any
	)
	defer func() {
		if didPanic {
			panic(panicRes)
		}
	}()
	runWithPanicProtection := func() (err error) {
		defer func() {
			if r := recover(); r != nil {
				didPanic = true
				panicRes = r
				// Even if r is an error, do not wrap it as an error here as
				// that would allow things like panic(vizerror.New("foo")) which
				// is really hard to define the behavior of.
				err = fmt.Errorf("panic: %v", r)
			}
		}()
		return h.rh.ServeHTTPReturn(lw, r)
	}
	err := runWithPanicProtection()

	var hErr HTTPError
	var hErrOK bool
	if errors.As(err, &hErr) {
		hErrOK = true
	} else if vizErr, ok := vizerror.As(err); ok {
		hErrOK = true
		hErr = HTTPError{Msg: vizErr.Error()}
	}

	if lw.code == 0 && err == nil && !lw.hijacked {
		// If the handler didn't write and didn't send a header, that still means 200.
		// (See https://play.golang.org/p/4P7nx_Tap7p)
		lw.code = 200
	}

	msg.Seconds = h.opts.Now().Sub(msg.Time).Seconds()
	msg.Code = lw.code
	msg.Bytes = lw.bytes

	switch {
	case lw.hijacked:
		// Connection no longer belongs to us, just log that we
		// switched protocols away from HTTP.
		if msg.Code == 0 {
			msg.Code = http.StatusSwitchingProtocols
		}
	case err != nil && r.Context().Err() == context.Canceled:
		msg.Code = 499 // nginx convention: Client Closed Request
		msg.Err = context.Canceled.Error()
	case hErrOK:
		// Handler asked us to send an error. Do so, if we haven't
		// already sent a response.
		msg.Err = hErr.Msg
		if hErr.Err != nil {
			if msg.Err == "" {
				msg.Err = hErr.Err.Error()
			} else {
				msg.Err = msg.Err + ": " + hErr.Err.Error()
			}
		}
		if lw.code != 0 {
			h.opts.Logf("[unexpected] handler returned HTTPError %v, but already sent a response with code %d", hErr, lw.code)
			break
		}
		msg.Code = hErr.Code
		if msg.Code == 0 {
			h.opts.Logf("[unexpected] HTTPError %v did not contain an HTTP status code, sending internal server error", hErr)
			msg.Code = http.StatusInternalServerError
		}
		if h.opts.OnError != nil {
			h.opts.OnError(lw, r, hErr)
		} else {
			// Default headers set by http.Error.
			lw.Header().Set("Content-Type", "text/plain; charset=utf-8")
			lw.Header().Set("X-Content-Type-Options", "nosniff")
			for k, vs := range hErr.Header {
				lw.Header()[k] = vs
			}
			lw.WriteHeader(msg.Code)
			fmt.Fprintln(lw, hErr.Msg)
			if msg.RequestID != "" {
				fmt.Fprintln(lw, msg.RequestID)
			}
		}
	case err != nil:
		const internalServerError = "internal server error"
		errorMessage := internalServerError
		if msg.RequestID != "" {
			errorMessage += "\n" + string(msg.RequestID)
		}
		// Handler returned a generic error. Serve an internal server
		// error, if necessary.
		msg.Err = err.Error()
		if lw.code == 0 {
			msg.Code = http.StatusInternalServerError
			http.Error(lw, errorMessage, msg.Code)
		}
	}

	if h.opts.OnCompletion != nil {
		h.opts.OnCompletion(r, msg)
	}

	if bs := h.opts.BucketedStats; bs != nil && bs.Finished != nil {
		// Only increment metrics for buckets that result in good HTTP statuses
		// or when we know the start was already counted.
		// Otherwise they get full of internet scanning noise. Only filtering 404
		// gets most of the way there but there are also plenty of URLs that are
		// almost right but result in 400s too. Seem easier to just only ignore
		// all 4xx and 5xx.
		if startRecorded {
			bs.Finished.Add(bucket, 1)
		} else if msg.Code < 400 {
			// This is the first non-error request for this bucket,
			// so count it now retroactively.
			bs.Started.Add(bucket, 1)
			bs.Finished.Add(bucket, 1)
		}
	}

	if !h.opts.QuietLoggingIfSuccessful || (msg.Code != http.StatusOK && msg.Code != http.StatusNotModified) {
		h.opts.Logf("%s", msg)
	}

	if h.opts.StatusCodeCounters != nil {
		h.opts.StatusCodeCounters.Add(responseCodeString(msg.Code/100), 1)
	}

	if h.opts.StatusCodeCountersFull != nil {
		h.opts.StatusCodeCountersFull.Add(responseCodeString(msg.Code), 1)
	}
}

func responseCodeString(code int) string {
	if v, ok := responseCodeCache.Load(code); ok {
		return v.(string)
	}

	var ret string
	if code < 10 {
		ret = fmt.Sprintf("%dxx", code)
	} else {
		ret = strconv.Itoa(code)
	}
	responseCodeCache.Store(code, ret)
	return ret
}

// responseCodeCache memoizes the string form of HTTP response codes,
// so that the hot request-handling codepath doesn't have to allocate
// in strconv/fmt for every request.
//
// Keys are either full HTTP response code ints (200, 404) or "family"
// ints representing entire families (e.g. 2 for 2xx codes). Values
// are the string form of that code/family.
var responseCodeCache sync.Map

// loggingResponseWriter wraps a ResponseWriter and record the HTTP
// response code that gets sent, if any.
type loggingResponseWriter struct {
	http.ResponseWriter
	code     int
	bytes    int
	hijacked bool
	logf     logger.Logf
}

// WriteHeader implements http.Handler.
func (l *loggingResponseWriter) WriteHeader(statusCode int) {
	if l.code != 0 {
		l.logf("[unexpected] HTTP handler set statusCode twice (%d and %d)", l.code, statusCode)
		return
	}
	l.code = statusCode
	l.ResponseWriter.WriteHeader(statusCode)
}

// Write implements http.Handler.
func (l *loggingResponseWriter) Write(bs []byte) (int, error) {
	if l.code == 0 {
		l.code = 200
	}
	n, err := l.ResponseWriter.Write(bs)
	l.bytes += n
	return n, err
}

// Hijack implements http.Hijacker. Note that hijacking can still fail
// because the wrapped ResponseWriter is not required to implement
// Hijacker, as this breaks HTTP/2.
func (l *loggingResponseWriter) Hijack() (net.Conn, *bufio.ReadWriter, error) {
	h, ok := l.ResponseWriter.(http.Hijacker)
	if !ok {
		return nil, nil, errors.New("ResponseWriter is not a Hijacker")
	}
	conn, buf, err := h.Hijack()
	if err == nil {
		l.hijacked = true
	}
	return conn, buf, err
}

func (l loggingResponseWriter) Flush() {
	f, _ := l.ResponseWriter.(http.Flusher)
	if f == nil {
		l.logf("[unexpected] tried to Flush a ResponseWriter that can't flush")
		return
	}
	f.Flush()
}

// HTTPError is an error with embedded HTTP response information.
//
// It is the error type to be (optionally) used by Handler.ServeHTTPReturn.
type HTTPError struct {
	Code   int         // HTTP response code to send to client; 0 means 500
	Msg    string      // Response body to send to client
	Err    error       // Detailed error to log on the server
	Header http.Header // Optional set of HTTP headers to set in the response
}

// Error implements the error interface.
func (e HTTPError) Error() string { return fmt.Sprintf("httperror{%d, %q, %v}", e.Code, e.Msg, e.Err) }
func (e HTTPError) Unwrap() error { return e.Err }

// Error returns an HTTPError containing the given information.
func Error(code int, msg string, err error) HTTPError {
	return HTTPError{Code: code, Msg: msg, Err: err}
}

// VarzHandler writes expvar values as Prometheus metrics.
// TODO: migrate all users to varz.Handler or promvarz.Handler and remove this.
func VarzHandler(w http.ResponseWriter, r *http.Request) {
	varz.Handler(w, r)
}

// CleanRedirectURL ensures that urlStr is a valid redirect URL to the
// current server, or one of allowedHosts. Returns the cleaned URL or
// a validation error.
func CleanRedirectURL(urlStr string, allowedHosts []string) (*url.URL, error) {
	if urlStr == "" {
		return &url.URL{}, nil
	}
	// In some places, we unfortunately query-escape the redirect URL
	// too many times, and end up needing to redirect to a URL that's
	// still escaped by one level. Try to unescape the input.
	unescaped, err := url.QueryUnescape(urlStr)
	if err == nil && unescaped != urlStr {
		urlStr = unescaped
	}

	// Go's URL parser and browser URL parsers disagree on the meaning
	// of malformed HTTP URLs. Given the input https:/evil.com, Go
	// parses it as hostname="", path="/evil.com". Browsers parse it
	// as hostname="evil.com", path="". This means that, using
	// malformed URLs, an attacker could trick us into approving of a
	// "local" redirect that in fact sends people elsewhere.
	//
	// This very blunt check enforces that we'll only process
	// redirects that are definitely well-formed URLs.
	//
	// Note that the check for just / also allows URLs of the form
	// "//foo.com/bar", which are scheme-relative redirects. These
	// must be handled with care below when determining whether a
	// redirect is relative to the current host. Notably,
	// url.URL.IsAbs reports // URLs as relative, whereas we want to
	// treat them as absolute redirects and verify the target host.
	if !hasSafeRedirectPrefix(urlStr) {
		return nil, fmt.Errorf("invalid redirect URL %q", urlStr)
	}

	url, err := url.Parse(urlStr)
	if err != nil {
		return nil, fmt.Errorf("invalid redirect URL %q: %w", urlStr, err)
	}
	// Redirects to self are always allowed. A self redirect must
	// start with url.Path, all prior URL sections must be empty.
	isSelfRedirect := url.Scheme == "" && url.Opaque == "" && url.User == nil && url.Host == ""
	if isSelfRedirect {
		return url, nil
	}
	for _, allowed := range allowedHosts {
		if strings.EqualFold(allowed, url.Hostname()) {
			return url, nil
		}
	}

	return nil, fmt.Errorf("disallowed target host %q in redirect URL %q", url.Hostname(), urlStr)
}

// hasSafeRedirectPrefix reports whether url starts with a slash, or
// one of the case-insensitive strings "http://" or "https://".
func hasSafeRedirectPrefix(url string) bool {
	if len(url) >= 1 && url[0] == '/' {
		return true
	}
	const http = "http://"
	if len(url) >= len(http) && strings.EqualFold(url[:len(http)], http) {
		return true
	}
	const https = "https://"
	if len(url) >= len(https) && strings.EqualFold(url[:len(https)], https) {
		return true
	}
	return false
}

// AddBrowserHeaders sets various HTTP security headers for browser-facing endpoints.
//
// The specific headers:
//   - require HTTPS access (HSTS)
//   - disallow iframe embedding
//   - mitigate MIME confusion attacks
//
// These headers are based on
// https://infosec.mozilla.org/guidelines/web_security
func AddBrowserHeaders(w http.ResponseWriter) {
	w.Header().Set("Strict-Transport-Security", "max-age=63072000; includeSubDomains")
	w.Header().Set("Content-Security-Policy", "default-src 'self'; frame-ancestors 'none'; form-action 'self'; base-uri 'self'; block-all-mixed-content; object-src 'none'")
	w.Header().Set("X-Frame-Options", "DENY")
	w.Header().Set("X-Content-Type-Options", "nosniff")
}

// BrowserHeaderHandler wraps the provided http.Handler with a call to
// AddBrowserHeaders.
func BrowserHeaderHandler(h http.Handler) http.Handler {
	return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
		AddBrowserHeaders(w)
		h.ServeHTTP(w, r)
	})
}

// BrowserHeaderHandlerFunc wraps the provided http.HandlerFunc with a call to
// AddBrowserHeaders.
func BrowserHeaderHandlerFunc(h http.HandlerFunc) http.HandlerFunc {
	return func(w http.ResponseWriter, r *http.Request) {
		AddBrowserHeaders(w)
		h.ServeHTTP(w, r)
	}
}