VERSION.txt: this is v1.24.1

Change-Id: Iee1c07bdca08609117153d42a0cd46b034222524
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>

VERSION.txt

@@ -1 +1 @@
-1.23.0
+1.24.1

cmd/tailscale/cli/ssh.go

@@ -23,7 +23,6 @@ import (
 	"tailscale.com/client/tailscale"
 	"tailscale.com/envknob"
 	"tailscale.com/ipn/ipnstate"
-	"tailscale.com/version"
 )
 
 var sshCmd = &ffcli.Command{
@@ -76,36 +75,52 @@ func runSSH(ctx context.Context, args []string) error {
 		return err
 	}
 
-	argv := append([]string{
-		ssh,
+	argv := []string{ssh}
 
+	if envknob.Bool("TS_DEBUG_SSH_EXEC") {
+		argv = append(argv, "-vvv")
+	}
+	argv = append(argv,
 		// Only trust SSH hosts that we know about.
 		"-o", fmt.Sprintf("UserKnownHostsFile %q", knownHostsFile),
 		"-o", "UpdateHostKeys no",
 		"-o", "StrictHostKeyChecking yes",
+	)
 
-		"-o", fmt.Sprintf("ProxyCommand %q --socket=%q nc %%h %%p",
-			tailscaleBin,
-			rootArgs.socket,
-		),
+	// TODO(bradfitz): nc is currently broken on macOS:
+	// https://github.com/tailscale/tailscale/issues/4529
+	// So don't use it for now. MagicDNS is usually working on macOS anyway
+	// and they're not in userspace mode, so 'nc' isn't very useful.
+	if runtime.GOOS != "darwin" {
+		argv = append(argv,
+			"-o", fmt.Sprintf("ProxyCommand %q --socket=%q nc %%h %%p",
+				tailscaleBin,
+				rootArgs.socket,
+			))
+	}
 
-		// Explicitly rebuild the user@host argument rather than
-		// passing it through.  In general, the use of OpenSSH's ssh
-		// binary is a crutch for now.  We don't want to be
-		// Hyrum-locked into passing through all OpenSSH flags to the
-		// OpenSSH client forever. We try to make our flags and args
-		// be compatible, but only a subset. The "tailscale ssh"
-		// command should be a simple and portable one. If they want
-		// to use a different one, we'll later be making stock ssh
-		// work well by default too. (doing things like automatically
-		// setting known_hosts, etc)
-		username + "@" + hostForSSH,
-	}, argRest...)
+	// Explicitly rebuild the user@host argument rather than
+	// passing it through.  In general, the use of OpenSSH's ssh
+	// binary is a crutch for now.  We don't want to be
+	// Hyrum-locked into passing through all OpenSSH flags to the
+	// OpenSSH client forever. We try to make our flags and args
+	// be compatible, but only a subset. The "tailscale ssh"
+	// command should be a simple and portable one. If they want
+	// to use a different one, we'll later be making stock ssh
+	// work well by default too. (doing things like automatically
+	// setting known_hosts, etc)
+	argv = append(argv, username+"@"+hostForSSH)
 
-	if runtime.GOOS == "windows" || version.IsSandboxedMacOS() {
-		// Don't use syscall.Exec on Windows or in the macOS sandbox.
+	argv = append(argv, argRest...)
+
+	if envknob.Bool("TS_DEBUG_SSH_EXEC") {
+		log.Printf("Running: %q, %q ...", ssh, argv)
+	}
+
+	if runtime.GOOS == "windows" {
+		// Don't use syscall.Exec on Windows.
 		cmd := exec.Command(ssh, argv[1:]...)
-		cmd.Stderr = os.Stderr
+		cmd.Stdin = os.Stdin
 		cmd.Stdout = os.Stdout
 		cmd.Stderr = os.Stderr
 		var ee *exec.ExitError
@@ -116,9 +131,6 @@ func runSSH(ctx context.Context, args []string) error {
 		return err
 	}
 
-	if envknob.Bool("TS_DEBUG_SSH_EXEC") {
-		log.Printf("Running: %q, %q ...", ssh, argv)
-	}
 	if err := syscall.Exec(ssh, argv, os.Environ()); err != nil {
 		return err
 	}

cmd/tailscale/cli/up.go

@@ -52,7 +52,7 @@ down").
 If flags are specified, the flags must be the complete set of desired
 settings. An error is returned if any setting would be changed as a
 result of an unspecified flag's default value, unless the --reset flag
-is also used. (The flags --authkey, --force-reauth, and --qr are not
+is also used. (The flags --auth-key, --force-reauth, and --qr are not
 considered settings that need to be re-specified when modifying
 settings.)
 `),

cmd/tailscaled/tailscaled.go

@@ -332,6 +332,7 @@ func run() error {
 	socksListener, httpProxyListener := mustStartProxyListeners(args.socksAddr, args.httpProxyAddr)
 
 	dialer := new(tsdial.Dialer) // mutated below (before used)
+	dialer.Logf = logf
 	e, useNetstack, err := createEngine(logf, linkMon, dialer)
 	if err != nil {
 		return fmt.Errorf("createEngine: %w", err)
@@ -394,6 +395,7 @@ func run() error {
 	// want to keep running.
 	signal.Ignore(syscall.SIGPIPE)
 	go func() {
+		defer dialer.Close()
 		select {
 		case s := <-interrupt:
 			logf("tailscaled got signal %v; shutting down", s)

control/controlclient/direct.go

@@ -38,9 +38,9 @@ import (
 	"tailscale.com/net/dnscache"
 	"tailscale.com/net/dnsfallback"
 	"tailscale.com/net/interfaces"
-	"tailscale.com/net/netns"
 	"tailscale.com/net/netutil"
 	"tailscale.com/net/tlsdial"
+	"tailscale.com/net/tsdial"
 	"tailscale.com/net/tshttpproxy"
 	"tailscale.com/tailcfg"
 	"tailscale.com/types/key"
@@ -57,7 +57,8 @@ import (
 // Direct is the client that connects to a tailcontrol server for a node.
 type Direct struct {
 	httpc                  *http.Client // HTTP client used to talk to tailcontrol
-	serverURL              string       // URL of the tailcontrol server
+	dialer                 *tsdial.Dialer
+	serverURL              string // URL of the tailcontrol server
 	timeNow                func() time.Time
 	lastPrintMap           time.Time
 	newDecompressor        func() (Decompressor, error)
@@ -106,6 +107,7 @@ type Options struct {
 	DebugFlags           []string         // debug settings to send to control
 	LinkMonitor          *monitor.Mon     // optional link monitor
 	PopBrowserURL        func(url string) // optional func to open browser
+	Dialer               *tsdial.Dialer   // non-nil
 
 	// KeepSharerAndUserSplit controls whether the client
 	// understands Node.Sharer. If false, the Sharer is mapped to the User.
@@ -170,13 +172,12 @@ func NewDirect(opts Options) (*Direct, error) {
 			UseLastGood:      true,
 			LookupIPFallback: dnsfallback.Lookup,
 		}
-		dialer := netns.NewDialer(opts.Logf)
 		tr := http.DefaultTransport.(*http.Transport).Clone()
 		tr.Proxy = tshttpproxy.ProxyFromEnvironment
 		tshttpproxy.SetTransportGetProxyConnectHeader(tr)
 		tr.TLSClientConfig = tlsdial.Config(serverURL.Hostname(), tr.TLSClientConfig)
-		tr.DialContext = dnscache.Dialer(dialer.DialContext, dnsCache)
-		tr.DialTLSContext = dnscache.TLSDialer(dialer.DialContext, dnsCache, tr.TLSClientConfig)
+		tr.DialContext = dnscache.Dialer(opts.Dialer.SystemDial, dnsCache)
+		tr.DialTLSContext = dnscache.TLSDialer(opts.Dialer.SystemDial, dnsCache, tr.TLSClientConfig)
 		tr.ForceAttemptHTTP2 = true
 		// Disable implicit gzip compression; the various
 		// handlers (register, map, set-dns, etc) do their own
@@ -202,6 +203,7 @@ func NewDirect(opts Options) (*Direct, error) {
 		skipIPForwardingCheck:  opts.SkipIPForwardingCheck,
 		pinger:                 opts.Pinger,
 		popBrowser:             opts.PopBrowserURL,
+		dialer:                 opts.Dialer,
 	}
 	if opts.Hostinfo == nil {
 		c.SetHostinfo(hostinfo.New())
@@ -376,7 +378,7 @@ func (c *Direct) doLogin(ctx context.Context, opt loginOpt) (mustRegen bool, new
 		if err != nil {
 			return regen, opt.URL, err
 		}
-		c.logf("control server key %s from %s", serverKey.ShortString(), c.serverURL)
+		c.logf("control server key from %s: ts2021=%s, legacy=%v", c.serverURL, keys.PublicKey.ShortString(), keys.LegacyPublicKey.ShortString())
 
 		c.mu.Lock()
 		c.serverKey = keys.LegacyPublicKey
@@ -1278,7 +1280,7 @@ func (c *Direct) getNoiseClient() (*noiseClient, error) {
 			return nil, err
 		}
 
-		nc, err = newNoiseClient(k, serverNoiseKey, c.serverURL)
+		nc, err = newNoiseClient(k, serverNoiseKey, c.serverURL, c.dialer)
 		if err != nil {
 			return nil, err
 		}

control/controlclient/direct_test.go

@@ -14,6 +14,7 @@ import (
 	"inet.af/netaddr"
 	"tailscale.com/hostinfo"
 	"tailscale.com/ipn/ipnstate"
+	"tailscale.com/net/tsdial"
 	"tailscale.com/tailcfg"
 	"tailscale.com/types/key"
 )
@@ -30,6 +31,7 @@ func TestNewDirect(t *testing.T) {
 		GetMachinePrivateKey: func() (key.MachinePrivate, error) {
 			return k, nil
 		},
+		Dialer: new(tsdial.Dialer),
 	}
 	c, err := NewDirect(opts)
 	if err != nil {
@@ -106,6 +108,7 @@ func TestTsmpPing(t *testing.T) {
 		GetMachinePrivateKey: func() (key.MachinePrivate, error) {
 			return k, nil
 		},
+		Dialer: new(tsdial.Dialer),
 	}
 
 	c, err := NewDirect(opts)

control/controlclient/noise.go

@@ -18,6 +18,7 @@ import (
 	"golang.org/x/net/http2"
 	"tailscale.com/control/controlbase"
 	"tailscale.com/control/controlhttp"
+	"tailscale.com/net/tsdial"
 	"tailscale.com/tailcfg"
 	"tailscale.com/types/key"
 	"tailscale.com/util/mak"
@@ -46,6 +47,7 @@ func (c *noiseConn) Close() error {
 // the ts2021 protocol.
 type noiseClient struct {
 	*http.Client // HTTP client used to talk to tailcontrol
+	dialer       *tsdial.Dialer
 	privKey      key.MachinePrivate
 	serverPubKey key.MachinePublic
 	serverHost   string // the host:port part of serverURL
@@ -58,7 +60,7 @@ type noiseClient struct {
 
 // newNoiseClient returns a new noiseClient for the provided server and machine key.
 // serverURL is of the form https://<host>:<port> (no trailing slash).
-func newNoiseClient(priKey key.MachinePrivate, serverPubKey key.MachinePublic, serverURL string) (*noiseClient, error) {
+func newNoiseClient(priKey key.MachinePrivate, serverPubKey key.MachinePublic, serverURL string, dialer *tsdial.Dialer) (*noiseClient, error) {
 	u, err := url.Parse(serverURL)
 	if err != nil {
 		return nil, err
@@ -75,6 +77,7 @@ func newNoiseClient(priKey key.MachinePrivate, serverPubKey key.MachinePublic, s
 		serverPubKey: serverPubKey,
 		privKey:      priKey,
 		serverHost:   host,
+		dialer:       dialer,
 	}
 
 	// Create the HTTP/2 Transport using a net/http.Transport
@@ -151,7 +154,7 @@ func (nc *noiseClient) dial(_, _ string, _ *tls.Config) (net.Conn, error) {
 		// thousand version numbers before getting to this point.
 		panic("capability version is too high to fit in the wire protocol")
 	}
-	conn, err := controlhttp.Dial(ctx, nc.serverHost, nc.privKey, nc.serverPubKey, uint16(tailcfg.CurrentCapabilityVersion))
+	conn, err := controlhttp.Dial(ctx, nc.serverHost, nc.privKey, nc.serverPubKey, uint16(tailcfg.CurrentCapabilityVersion), nc.dialer.SystemDial)
 	if err != nil {
 		return nil, err
 	}

control/controlhttp/client.go

@@ -25,16 +25,15 @@ import (
 	"errors"
 	"fmt"
 	"io"
-	"log"
 	"net"
 	"net/http"
 	"net/http/httptrace"
 	"net/url"
+	"time"
 
 	"tailscale.com/control/controlbase"
 	"tailscale.com/net/dnscache"
 	"tailscale.com/net/dnsfallback"
-	"tailscale.com/net/netns"
 	"tailscale.com/net/netutil"
 	"tailscale.com/net/tlsdial"
 	"tailscale.com/net/tshttpproxy"
@@ -65,7 +64,7 @@ const (
 //
 // The provided ctx is only used for the initial connection, until
 // Dial returns. It does not affect the connection once established.
-func Dial(ctx context.Context, addr string, machineKey key.MachinePrivate, controlKey key.MachinePublic, protocolVersion uint16) (*controlbase.Conn, error) {
+func Dial(ctx context.Context, addr string, machineKey key.MachinePrivate, controlKey key.MachinePublic, protocolVersion uint16, dialer dnscache.DialContextFunc) (*controlbase.Conn, error) {
 	host, port, err := net.SplitHostPort(addr)
 	if err != nil {
 		return nil, err
@@ -79,6 +78,7 @@ func Dial(ctx context.Context, addr string, machineKey key.MachinePrivate, contr
 		controlKey: controlKey,
 		version:    protocolVersion,
 		proxyFunc:  tshttpproxy.ProxyFromEnvironment,
+		dialer:     dialer,
 	}
 	return a.dial()
 }
@@ -92,65 +92,115 @@ type dialParams struct {
 	controlKey key.MachinePublic
 	version    uint16
 	proxyFunc  func(*http.Request) (*url.URL, error) // or nil
+	dialer     dnscache.DialContextFunc
 
 	// For tests only
 	insecureTLS bool
 }
 
 func (a *dialParams) dial() (*controlbase.Conn, error) {
-	init, cont, err := controlbase.ClientDeferred(a.machineKey, a.controlKey, a.version)
-	if err != nil {
-		return nil, err
-	}
+	// Create one shared context used by both port 80 and port 443 dials.
+	// If port 80 is still in flight when 443 returns, this deferred cancel
+	// will stop the port 80 dial.
+	ctx, cancel := context.WithCancel(a.ctx)
+	defer cancel()
 
-	u := &url.URL{
+	// u80 and u443 are the URLs we'll try to hit over HTTP or HTTPS,
+	// respectively, in order to do the HTTP upgrade to a net.Conn over which
+	// we'll speak Noise.
+	u80 := &url.URL{
 		Scheme: "http",
 		Host:   net.JoinHostPort(a.host, a.httpPort),
 		Path:   serverUpgradePath,
 	}
-	conn, httpErr := a.tryURL(u, init)
-	if httpErr == nil {
-		ret, err := cont(a.ctx, conn)
-		if err != nil {
-			conn.Close()
-			return nil, err
+	u443 := &url.URL{
+		Scheme: "https",
+		Host:   net.JoinHostPort(a.host, a.httpsPort),
+		Path:   serverUpgradePath,
+	}
+	type tryURLRes struct {
+		u    *url.URL
+		conn net.Conn
+		cont controlbase.HandshakeContinuation
+		err  error
+	}
+	ch := make(chan tryURLRes) // must be unbuffered
+
+	try := func(u *url.URL) {
+		res := tryURLRes{u: u}
+		var init []byte
+		init, res.cont, res.err = controlbase.ClientDeferred(a.machineKey, a.controlKey, a.version)
+		if res.err == nil {
+			res.conn, res.err = a.tryURL(ctx, u, init)
+		}
+		select {
+		case ch <- res:
+		case <-ctx.Done():
+			if res.conn != nil {
+				res.conn.Close()
+			}
 		}
-		return ret, nil
 	}
 
-	// Connecting over plain HTTP failed, assume it's an HTTP proxy
-	// being difficult and see if we can get through over HTTPS.
-	u.Scheme = "https"
-	u.Host = net.JoinHostPort(a.host, a.httpsPort)
-	init, cont, err = controlbase.ClientDeferred(a.machineKey, a.controlKey, a.version)
-	if err != nil {
-		return nil, err
-	}
-	conn, tlsErr := a.tryURL(u, init)
-	if tlsErr == nil {
-		ret, err := cont(a.ctx, conn)
-		if err != nil {
-			conn.Close()
-			return nil, err
-		}
-		return ret, nil
-	}
+	// Start the plaintext HTTP attempt first.
+	go try(u80)
 
-	return nil, fmt.Errorf("all connection attempts failed (HTTP: %v, HTTPS: %v)", httpErr, tlsErr)
+	// In case outbound port 80 blocked or MITM'ed poorly, start a backup timer
+	// to dial port 443 if port 80 doesn't either succeed or fail quickly.
+	try443Timer := time.AfterFunc(500*time.Millisecond, func() { try(u443) })
+	defer try443Timer.Stop()
+
+	var err80, err443 error
+	for {
+		select {
+		case <-ctx.Done():
+			return nil, fmt.Errorf("connection attempts aborted by context: %w", ctx.Err())
+		case res := <-ch:
+			if res.err == nil {
+				ret, err := res.cont(ctx, res.conn)
+				if err != nil {
+					res.conn.Close()
+					return nil, err
+				}
+				return ret, nil
+			}
+			switch res.u {
+			case u80:
+				// Connecting over plain HTTP failed; assume it's an HTTP proxy
+				// being difficult and see if we can get through over HTTPS.
+				err80 = res.err
+				// Stop the fallback timer and run it immediately. We don't use
+				// Timer.Reset(0) here because on AfterFuncs, that can run it
+				// again.
+				if try443Timer.Stop() {
+					go try(u443)
+				} // else we lost the race and it started already which is what we want
+			case u443:
+				err443 = res.err
+			default:
+				panic("invalid")
+			}
+			if err80 != nil && err443 != nil {
+				return nil, fmt.Errorf("all connection attempts failed (HTTP: %v, HTTPS: %v)", err80, err443)
+			}
+		}
+	}
 }
 
-func (a *dialParams) tryURL(u *url.URL, init []byte) (net.Conn, error) {
+// tryURL connects to u, and tries to upgrade it to a net.Conn.
+//
+// Only the provided ctx is used, not a.ctx.
+func (a *dialParams) tryURL(ctx context.Context, u *url.URL, init []byte) (net.Conn, error) {
 	dns := &dnscache.Resolver{
 		Forward:          dnscache.Get().Forward,
 		LookupIPFallback: dnsfallback.Lookup,
 		UseLastGood:      true,
 	}
-	dialer := netns.NewDialer(log.Printf)
 	tr := http.DefaultTransport.(*http.Transport).Clone()
 	defer tr.CloseIdleConnections()
 	tr.Proxy = a.proxyFunc
 	tshttpproxy.SetTransportGetProxyConnectHeader(tr)
-	tr.DialContext = dnscache.Dialer(dialer.DialContext, dns)
+	tr.DialContext = dnscache.Dialer(a.dialer, dns)
 	// Disable HTTP2, since h2 can't do protocol switching.
 	tr.TLSClientConfig.NextProtos = []string{}
 	tr.TLSNextProto = map[string]func(string, *tls.Conn) http.RoundTripper{}
@@ -159,7 +209,7 @@ func (a *dialParams) tryURL(u *url.URL, init []byte) (net.Conn, error) {
 		tr.TLSClientConfig.InsecureSkipVerify = true
 		tr.TLSClientConfig.VerifyConnection = nil
 	}
-	tr.DialTLSContext = dnscache.TLSDialer(dialer.DialContext, dns, tr.TLSClientConfig)
+	tr.DialTLSContext = dnscache.TLSDialer(a.dialer, dns, tr.TLSClientConfig)
 	tr.DisableCompression = true
 
 	// (mis)use httptrace to extract the underlying net.Conn from the
@@ -189,7 +239,7 @@ func (a *dialParams) tryURL(u *url.URL, init []byte) (net.Conn, error) {
 			connCh <- info.Conn
 		},
 	}
-	ctx := httptrace.WithClientTrace(a.ctx, &trace)
+	ctx = httptrace.WithClientTrace(ctx, &trace)
 	req := &http.Request{
 		Method: "POST",
 		URL:    u,

control/controlhttp/http_test.go

@@ -20,6 +20,7 @@ import (
 
 	"tailscale.com/control/controlbase"
 	"tailscale.com/net/socks5"
+	"tailscale.com/net/tsdial"
 	"tailscale.com/types/key"
 )
 
@@ -155,6 +156,7 @@ func testControlHTTP(t *testing.T, proxy proxy) {
 		controlKey:  server.Public(),
 		version:     testProtocolVersion,
 		insecureTLS: true,
+		dialer:      new(tsdial.Dialer).SystemDial,
 	}
 
 	if proxy != nil {

ipn/ipnlocal/local.go

@@ -1034,6 +1034,7 @@ func (b *LocalBackend) Start(opts ipn.Options) error {
 		LinkMonitor:          b.e.GetLinkMonitor(),
 		Pinger:               b.e,
 		PopBrowserURL:        b.tellClientToBrowseToURL,
+		Dialer:               b.Dialer(),
 
 		// Don't warn about broken Linux IP forwarding when
 		// netstack is being used.

ipn/ipnlocal/peerapi.go

@@ -563,6 +563,9 @@ func (h *peerAPIHandler) ServeHTTP(w http.ResponseWriter, r *http.Request) {
 	case "/v0/dnsfwd":
 		h.handleServeDNSFwd(w, r)
 		return
+	case "/v0/interfaces":
+		h.handleServeInterfaces(w, r)
+		return
 	}
 	who := h.peerUser.DisplayName
 	fmt.Fprintf(w, `<html>
@@ -577,6 +580,40 @@ This is my Tailscale device. Your device is %v.
 	}
 }
 
+func (h *peerAPIHandler) handleServeInterfaces(w http.ResponseWriter, r *http.Request) {
+	if !h.canDebug() {
+		http.Error(w, "denied; no debug access", http.StatusForbidden)
+		return
+	}
+	i, err := interfaces.GetList()
+	if err != nil {
+		http.Error(w, err.Error(), 500)
+	}
+
+	dr, err := interfaces.DefaultRoute()
+	if err != nil {
+		http.Error(w, err.Error(), 500)
+	}
+	w.Header().Set("Content-Type", "text/html; charset=utf-8")
+	fmt.Fprintln(w, "<h1>Interfaces</h1>")
+	fmt.Fprintf(w, "<h3>Default route is %q(%d)</h3>\n", dr.InterfaceName, dr.InterfaceIndex)
+
+	fmt.Fprintln(w, "<table>")
+	fmt.Fprint(w, "<tr>")
+	for _, v := range []any{"Index", "Name", "MTU", "Flags", "Addrs"} {
+		fmt.Fprintf(w, "<th>%v</th> ", v)
+	}
+	fmt.Fprint(w, "</tr>\n")
+	i.ForeachInterface(func(iface interfaces.Interface, ipps []netaddr.IPPrefix) {
+		fmt.Fprint(w, "<tr>")
+		for _, v := range []any{iface.Index, iface.Name, iface.MTU, iface.Flags, ipps} {
+			fmt.Fprintf(w, "<td>%v</td> ", v)
+		}
+		fmt.Fprint(w, "</tr>\n")
+	})
+	fmt.Fprintln(w, "</table>")
+}
+
 type incomingFile struct {
 	name        string // "foo.jpg"
 	started     time.Time

ipn/prefs.go

@@ -428,9 +428,14 @@ func NewPrefs() *Prefs {
 }
 
 // ControlURLOrDefault returns the coordination server's URL base.
-// If not configured, DefaultControlURL is returned instead.
+//
+// If not configured, or if the configured value is a legacy name equivalent to
+// the default, then DefaultControlURL is returned instead.
 func (p *Prefs) ControlURLOrDefault() string {
 	if p.ControlURL != "" {
+		if p.ControlURL != DefaultControlURL && IsLoginServerSynonym(p.ControlURL) {
+			return DefaultControlURL
+		}
 		return p.ControlURL
 	}
 	return DefaultControlURL

ipn/prefs_test.go

@@ -810,3 +810,18 @@ func TestExitNodeIPOfArg(t *testing.T) {
 		})
 	}
 }
+
+func TestControlURLOrDefault(t *testing.T) {
+	var p Prefs
+	if got, want := p.ControlURLOrDefault(), DefaultControlURL; got != want {
+		t.Errorf("got %q; want %q", got, want)
+	}
+	p.ControlURL = "http://foo.bar"
+	if got, want := p.ControlURLOrDefault(), "http://foo.bar"; got != want {
+		t.Errorf("got %q; want %q", got, want)
+	}
+	p.ControlURL = "https://login.tailscale.com"
+	if got, want := p.ControlURLOrDefault(), DefaultControlURL; got != want {
+		t.Errorf("got %q; want %q", got, want)
+	}
+}

net/tsdial/tsdial.go

@@ -20,8 +20,12 @@ import (
 
 	"inet.af/netaddr"
 	"tailscale.com/net/dnscache"
+	"tailscale.com/net/interfaces"
 	"tailscale.com/net/netknob"
+	"tailscale.com/net/netns"
+	"tailscale.com/types/logger"
 	"tailscale.com/types/netmap"
+	"tailscale.com/util/mak"
 	"tailscale.com/wgengine/monitor"
 )
 
@@ -30,6 +34,7 @@ import (
 // (TUN, netstack), the OS network sandboxing style (macOS/iOS
 // Extension, none), user-selected route acceptance prefs, etc.
 type Dialer struct {
+	Logf logger.Logf
 	// UseNetstackForIP if non-nil is whether NetstackDialTCP (if
 	// it's non-nil) should be used to dial the provided IP.
 	UseNetstackForIP func(netaddr.IP) bool
@@ -46,12 +51,33 @@ type Dialer struct {
 	peerDialerOnce sync.Once
 	peerDialer     *net.Dialer
 
-	mu             sync.Mutex
-	dns            dnsMap
-	tunName        string // tun device name
-	linkMon        *monitor.Mon
-	exitDNSDoHBase string                 // non-empty if DoH-proxying exit node in use; base URL+path (without '?')
-	dnsCache       *dnscache.MessageCache // nil until first first non-empty SetExitDNSDoH
+	netnsDialerOnce sync.Once
+	netnsDialer     netns.Dialer
+
+	mu                sync.Mutex
+	closed            bool
+	dns               dnsMap
+	tunName           string // tun device name
+	linkMon           *monitor.Mon
+	linkMonUnregister func()
+	exitDNSDoHBase    string                 // non-empty if DoH-proxying exit node in use; base URL+path (without '?')
+	dnsCache          *dnscache.MessageCache // nil until first first non-empty SetExitDNSDoH
+	nextSysConnID     int
+	activeSysConns    map[int]net.Conn // active connections not yet closed
+}
+
+// sysConn wraps a net.Conn that was created using d.SystemDial.
+// It exists to track which connections are still open, and should be
+// closed on major link changes.
+type sysConn struct {
+	net.Conn
+	id int
+	d  *Dialer
+}
+
+func (c sysConn) Close() error {
+	c.d.closeSysConn(c.id)
+	return nil
 }
 
 // SetTUNName sets the name of the tun device in use ("tailscale0", "utun6",
@@ -91,10 +117,53 @@ func (d *Dialer) SetExitDNSDoH(doh string) {
 	}
 }
 
+func (d *Dialer) Close() error {
+	d.mu.Lock()
+	defer d.mu.Unlock()
+	d.closed = true
+	if d.linkMonUnregister != nil {
+		d.linkMonUnregister()
+		d.linkMonUnregister = nil
+	}
+	for _, c := range d.activeSysConns {
+		c.Close()
+	}
+	d.activeSysConns = nil
+	return nil
+}
+
 func (d *Dialer) SetLinkMonitor(mon *monitor.Mon) {
 	d.mu.Lock()
 	defer d.mu.Unlock()
+	if d.linkMonUnregister != nil {
+		go d.linkMonUnregister()
+		d.linkMonUnregister = nil
+	}
 	d.linkMon = mon
+	d.linkMonUnregister = d.linkMon.RegisterChangeCallback(d.linkChanged)
+}
+
+func (d *Dialer) linkChanged(major bool, state *interfaces.State) {
+	if !major {
+		return
+	}
+	d.mu.Lock()
+	defer d.mu.Unlock()
+	for id, c := range d.activeSysConns {
+		go c.Close()
+		delete(d.activeSysConns, id)
+	}
+}
+
+func (d *Dialer) closeSysConn(id int) {
+	d.mu.Lock()
+	defer d.mu.Unlock()
+	c, ok := d.activeSysConns[id]
+	if !ok {
+		return
+	}
+	delete(d.activeSysConns, id)
+	go c.Close() // ignore the error
 }
 
 func (d *Dialer) interfaceIndexLocked(ifName string) (index int, ok bool) {
@@ -197,6 +266,42 @@ func ipNetOfNetwork(n string) string {
 	return "ip"
 }
 
+// SystemDial connects to the provided network address without going over
+// Tailscale. It prefers going over the default interface and closes existing
+// connections if the default interface changes. It is used to connect to
+// Control and (in the future, as of 2022-04-27) DERPs..
+func (d *Dialer) SystemDial(ctx context.Context, network, addr string) (net.Conn, error) {
+	d.mu.Lock()
+	closed := d.closed
+	d.mu.Unlock()
+	if closed {
+		return nil, net.ErrClosed
+	}
+
+	d.netnsDialerOnce.Do(func() {
+		logf := d.Logf
+		if logf == nil {
+			logf = logger.Discard
+		}
+		d.netnsDialer = netns.NewDialer(logf)
+	})
+	c, err := d.netnsDialer.DialContext(ctx, network, addr)
+	if err != nil {
+		return nil, err
+	}
+	d.mu.Lock()
+	defer d.mu.Unlock()
+	id := d.nextSysConnID
+	d.nextSysConnID++
+	mak.Set(&d.activeSysConns, id, c)
+
+	return sysConn{
+		id:   id,
+		d:    d,
+		Conn: c,
+	}, nil
+}
+
 // UserDial connects to the provided network address as if a user were initiating the dial.
 // (e.g. from a SOCKS or HTTP outbound proxy)
 func (d *Dialer) UserDial(ctx context.Context, network, addr string) (net.Conn, error) {

net/tshttpproxy/tshttpproxy_synology.go

@@ -121,7 +121,7 @@ func parseSynologyConfig(r io.Reader) (*url.URL, error) {
 	return proxyURL, nil
 }
 
-// mtime stat's path and returns it's modification time. If path does not exist,
+// mtime stat's path and returns its modification time. If path does not exist,
 // it returns the unix epoch.
 func mtime(path string) time.Time {
 	fi, err := os.Stat(path)

safesocket/unixsocket.go

@@ -11,7 +11,6 @@ import (
 	"errors"
 	"fmt"
 	"io"
-	"io/ioutil"
 	"log"
 	"net"
 	"os"
@@ -19,7 +18,6 @@ import (
 	"path/filepath"
 	"runtime"
 	"strconv"
-	"strings"
 )
 
 // TODO(apenwarr): handle magic cookie auth
@@ -114,77 +112,30 @@ func socketPermissionsForOS() os.FileMode {
 	return 0600
 }
 
-// connectMacOSAppSandbox connects to the Tailscale Network Extension,
-// which is necessarily running within the macOS App Sandbox.  Our
-// little dance to connect a regular user binary to the sandboxed
-// network extension is:
+// connectMacOSAppSandbox connects to the Tailscale Network Extension (macOS App
+// Store build) or App Extension (macsys standalone build), where the CLI itself
+// is either running within the macOS App Sandbox or built separately (e.g.
+// homebrew or go install). This little dance to connect a regular user binary
+// to the sandboxed network extension is:
 //
 //   * the sandboxed IPNExtension picks a random localhost:0 TCP port
 //     to listen on
 //   * it also picks a random hex string that acts as an auth token
-//   * it then creates a file named "sameuserproof-$PORT-$TOKEN" and leaves
-//     that file descriptor open forever.
-//
-// Then, we do different things depending on whether the user is
-// running cmd/tailscale that they built themselves (running as
-// themselves, outside the App Sandbox), or whether the user is
-// running the CLI via the GUI binary
-// (e.g. /Applications/Tailscale.app/Contents/MacOS/Tailscale <args>),
-// in which case we're running within the App Sandbox.
-//
-// If we're outside the App Sandbox:
-//
-//   * then we come along here, running as the same UID, but outside
-//     of the sandbox, and look for it. We can run lsof on our own processes,
-//     but other users on the system can't.
-//   * we parse out the localhost port number and the auth token
-//   * we connect to TCP localhost:$PORT
-//   * we send $TOKEN + "\n"
-//   * server verifies $TOKEN, sends "#IPN\n" if okay.
-//   * server is now protocol switched
-//   * we return the net.Conn and the caller speaks the normal protocol
-//
-// If we're inside the App Sandbox, then TS_MACOS_CLI_SHARED_DIR has
-// been set to our shared directory. We now have to find the most
-// recent "sameuserproof" file (there should only be 1, but previous
-// versions of the macOS app didn't clean them up).
+//   * the CLI looks on disk for that TCP port + auth token (see localTCPPortAndTokenDarwin)
+//   * we send it upon TCP connect to prove to the Tailscale daemon that
+//     we're a suitably privileged user to have access the files on disk
+//     which the Network/App Extension wrote.
 func connectMacOSAppSandbox() (net.Conn, error) {
-	// Are we running the Tailscale.app GUI binary as a CLI, running within the App Sandbox?
-	if d := os.Getenv("TS_MACOS_CLI_SHARED_DIR"); d != "" {
-		fis, err := ioutil.ReadDir(d)
-		if err != nil {
-			return nil, fmt.Errorf("reading TS_MACOS_CLI_SHARED_DIR: %w", err)
-		}
-		var best os.FileInfo
-		for _, fi := range fis {
-			if !strings.HasPrefix(fi.Name(), "sameuserproof-") || strings.Count(fi.Name(), "-") != 2 {
-				continue
-			}
-			if best == nil || fi.ModTime().After(best.ModTime()) {
-				best = fi
-			}
-		}
-		if best == nil {
-			return nil, fmt.Errorf("no sameuserproof token found in TS_MACOS_CLI_SHARED_DIR %q", d)
-		}
-		f := strings.SplitN(best.Name(), "-", 3)
-		portStr, token := f[1], f[2]
-		port, err := strconv.Atoi(portStr)
-		if err != nil {
-			return nil, fmt.Errorf("invalid port %q", portStr)
-		}
-		return connectMacTCP(port, token)
-	}
-
-	// Otherwise, assume we're running the cmd/tailscale binary from outside the
-	// App Sandbox.
 	port, token, err := LocalTCPPortAndToken()
 	if err != nil {
-		return nil, err
+		return nil, fmt.Errorf("failed to find local Tailscale daemon: %w", err)
 	}
 	return connectMacTCP(port, token)
 }
 
+// connectMacTCP creates an authenticated net.Conn to the local macOS Tailscale
+// daemon for used by the "IPN" JSON message bus protocol (Tailscale's original
+// local non-HTTP IPC protocol).
 func connectMacTCP(port int, token string) (net.Conn, error) {
 	c, err := net.Dial("tcp", "localhost:"+strconv.Itoa(port))
 	if err != nil {

tsnet/tsnet.go

@@ -105,6 +105,7 @@ func (s *Server) Close() error {
 	s.shutdownCancel()
 	s.lb.Shutdown()
 	s.linkMon.Close()
+	s.dialer.Close()
 	s.localAPIListener.Close()
 
 	s.mu.Lock()
@@ -137,8 +138,9 @@ func (s *Server) start() error {
 	}
 
 	s.rootPath = s.Dir
-	if s.Store != nil && !s.Ephemeral {
-		if _, ok := s.Store.(*mem.Store); !ok {
+	if s.Store != nil {
+		_, isMemStore := s.Store.(*mem.Store)
+		if isMemStore && !s.Ephemeral {
 			return fmt.Errorf("in-memory store is only supported for Ephemeral nodes")
 		}
 	}

wgengine/monitor/monitor_darwin.go

@@ -115,7 +115,8 @@ func addrType(addrs []route.Addr, rtaxType int) route.Addr {
 func (m *darwinRouteMon) IsInterestingInterface(iface string) bool {
 	baseName := strings.TrimRight(iface, "0123456789")
 	switch baseName {
-	case "llw", "awdl", "pdp_ip", "ipsec":
+	// TODO(maisem): figure out what this list should actually be.
+	case "llw", "awdl", "ipsec":
 		return false
 	}
 	return true

wgengine/netstack/netstack.go

@@ -651,6 +651,21 @@ func (ns *Impl) acceptTCP(r *tcp.ForwarderRequest) {
 	}
 	r.Complete(false)
 
+	// SetKeepAlive so that idle connections to peers that have forgotten about
+	// the connection or gone completely offline eventually time out.
+	// Applications might be setting this on a forwarded connection, but from
+	// userspace we can not see those, so the best we can do is to always
+	// perform them with conservative timing.
+	// TODO(tailscale/tailscale#4522): Netstack defaults match the Linux
+	// defaults, and results in a little over two hours before the socket would
+	// be closed due to keepalive. A shorter default might be better, or seeking
+	// a default from the host IP stack. This also might be a useful
+	// user-tunable, as in userspace mode this can have broad implications such
+	// as lingering connections to fork style daemons. On the other side of the
+	// fence, the long duration timers are low impact values for battery powered
+	// peers.
+	ep.SocketOptions().SetKeepAlive(true)
+
 	// The ForwarderRequest.CreateEndpoint above asynchronously
 	// starts the TCP handshake. Note that the gonet.TCPConn
 	// methods c.RemoteAddr() and c.LocalAddr() will return nil