cmd/tailscaled: unify the two Windows paths + separate IPN server path

tailscaled on Windows had two entirely separate start-up paths for running
as a service vs in the foreground. It's been causing problems for ages.
This unifies the two paths, making them be the same as the path used
for every other platform.

Also, it uses the new async LocalBackend support in ipnserver.Server
so the Server can start serving HTTP immediately, even if tun takes
awhile to come up.

Updates #6535

Change-Id: Icc8c4f96d4887b54a024d7ac15ad11096b5a58cf
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>

cmd/tailscaled/tailscaled.go

@@ -48,6 +48,7 @@ import (
 	"tailscale.com/paths"
 	"tailscale.com/safesocket"
 	"tailscale.com/smallzstd"
+	"tailscale.com/syncs"
 	"tailscale.com/tsweb"
 	"tailscale.com/types/flagtype"
 	"tailscale.com/types/logger"
@@ -110,6 +111,9 @@ func defaultPort() uint16 {
 			return uint16(p)
 		}
 	}
+	if envknob.GOOS() == "windows" {
+		return 41641
+	}
 	return 0
 }
 
@@ -320,11 +324,13 @@ func ipnServerOpts() (o serverOptions) {
 	return o
 }
 
-func run() error {
+var logPol *logpolicy.Policy
-	var err error
+var debugMux *http.ServeMux
 
+func run() error {
 	pol := logpolicy.New(logtail.CollectionNode)
 	pol.SetVerbosityLevel(args.verbose)
+	logPol = pol
 	defer func() {
 		// Finish uploading logs after closing everything else.
 		ctx, cancel := context.WithTimeout(context.Background(), time.Second)
@@ -368,23 +374,87 @@ func run() error {
 		log.Printf("error in synology migration: %v", err)
 	}
 
-	var debugMux *http.ServeMux
 	if args.debug != "" {
 		debugMux = newDebugMux()
 	}
 
+	logid := pol.PublicID.String()
+	return startIPNServer(context.Background(), logf, logid)
+}
+
+func startIPNServer(ctx context.Context, logf logger.Logf, logid string) error {
+	ln, _, err := safesocket.Listen(args.socketpath, safesocket.WindowsLocalPort)
+	if err != nil {
+		return fmt.Errorf("safesocket.Listen: %v", err)
+	}
+
+	ctx, cancel := context.WithCancel(ctx)
+	defer cancel()
+	// Exit gracefully by cancelling the ipnserver context in most common cases:
+	// interrupted from the TTY or killed by a service manager.
+	interrupt := make(chan os.Signal, 1)
+	signal.Notify(interrupt, syscall.SIGINT, syscall.SIGTERM)
+	// SIGPIPE sometimes gets generated when CLIs disconnect from
+	// tailscaled. The default action is to terminate the process, we
+	// want to keep running.
+	signal.Ignore(syscall.SIGPIPE)
+	go func() {
+		select {
+		case s := <-interrupt:
+			logf("tailscaled got signal %v; shutting down", s)
+			cancel()
+		case <-ctx.Done():
+			// continue
+		}
+	}()
+
+	srv := ipnserver.New(logf, logid)
+	if debugMux != nil {
+		debugMux.HandleFunc("/debug/ipn", srv.ServeHTMLStatus)
+	}
+	var lbErr syncs.AtomicValue[error]
+
+	go func() {
+		t0 := time.Now()
+		lb, err := getLocalBackend(ctx, logf, logid)
+		if err == nil {
+			logf("got LocalBackend in %v", time.Since(t0).Round(time.Millisecond))
+			srv.SetLocalBackend(lb)
+			return
+		}
+		lbErr.Store(err) // before the following cancel
+		cancel()         // make srv.Run below complete
+	}()
+
+	err = srv.Run(ctx, ln)
+
+	if err != nil && lbErr.Load() != nil {
+		return fmt.Errorf("getLocalBackend error: %v", lbErr.Load())
+	}
+
+	// Cancelation is not an error: it is the only way to stop ipnserver.
+	if err != nil && !errors.Is(err, context.Canceled) {
+		return fmt.Errorf("ipnserver.Run: %w", err)
+	}
+
+	return nil
+}
+
+func getLocalBackend(ctx context.Context, logf logger.Logf, logid string) (_ *ipnlocal.LocalBackend, retErr error) {
 	linkMon, err := monitor.New(logf)
 	if err != nil {
-		return fmt.Errorf("monitor.New: %w", err)
+		return nil, fmt.Errorf("monitor.New: %w", err)
+	}
+	if logPol != nil {
+		logPol.Logtail.SetLinkMonitor(linkMon)
 	}
-	pol.Logtail.SetLinkMonitor(linkMon)
 
 	socksListener, httpProxyListener := mustStartProxyListeners(args.socksAddr, args.httpProxyAddr)
 
 	dialer := &tsdial.Dialer{Logf: logf} // mutated below (before used)
 	e, useNetstack, err := createEngine(logf, linkMon, dialer)
 	if err != nil {
-		return fmt.Errorf("createEngine: %w", err)
+		return nil, fmt.Errorf("createEngine: %w", err)
 	}
 	if _, ok := e.(wgengine.ResolvingEngine).GetResolver(); !ok {
 		panic("internal error: exit node resolver not wired up")
@@ -400,7 +470,7 @@ func run() error {
 
 	ns, err := newNetstack(logf, dialer, e)
 	if err != nil {
-		return fmt.Errorf("newNetstack: %w", err)
+		return nil, fmt.Errorf("newNetstack: %w", err)
 	}
 	ns.ProcessLocalIPs = useNetstack
 	ns.ProcessSubnets = useNetstack || shouldWrapNetstack()
@@ -434,36 +504,16 @@ func run() error {
 
 	e = wgengine.NewWatchdog(e)
 
-	ctx, cancel := context.WithCancel(context.Background())
-	// Exit gracefully by cancelling the ipnserver context in most common cases:
-	// interrupted from the TTY or killed by a service manager.
-	interrupt := make(chan os.Signal, 1)
-	signal.Notify(interrupt, syscall.SIGINT, syscall.SIGTERM)
-	// SIGPIPE sometimes gets generated when CLIs disconnect from
-	// tailscaled. The default action is to terminate the process, we
-	// want to keep running.
-	signal.Ignore(syscall.SIGPIPE)
-	go func() {
-		select {
-		case s := <-interrupt:
-			logf("tailscaled got signal %v; shutting down", s)
-			cancel()
-		case <-ctx.Done():
-			// continue
-		}
-	}()
-
 	opts := ipnServerOpts()
 
 	store, err := store.New(logf, statePathOrDefault())
 	if err != nil {
-		return fmt.Errorf("store.New: %w", err)
+		return nil, fmt.Errorf("store.New: %w", err)
 	}
-	logid := pol.PublicID.String()
 
 	lb, err := ipnlocal.NewLocalBackend(logf, logid, store, "", dialer, e, opts.LoginFlags)
 	if err != nil {
-		return fmt.Errorf("ipnlocal.NewLocalBackend: %w", err)
+		return nil, fmt.Errorf("ipnlocal.NewLocalBackend: %w", err)
 	}
 	lb.SetVarRoot(opts.VarRoot)
 	if root := lb.TailscaleVarRoot(); root != "" {
@@ -473,31 +523,11 @@ func run() error {
 		return smallzstd.NewDecoder(nil)
 	})
 	configureTaildrop(logf, lb)
-
-	srv := ipnserver.New(logf, logid)
-	srv.SetLocalBackend(lb)
 	ns.SetLocalBackend(lb)
 	if err := ns.Start(); err != nil {
 		log.Fatalf("failed to start netstack: %v", err)
 	}
-
+	return lb, nil
-	if debugMux != nil {
-		debugMux.HandleFunc("/debug/ipn", srv.ServeHTMLStatus)
-	}
-
-	ln, _, err := safesocket.Listen(args.socketpath, safesocket.WindowsLocalPort)
-	if err != nil {
-		return fmt.Errorf("safesocket.Listen: %v", err)
-	}
-	defer dialer.Close()
-
-	err = srv.Run(ctx, ln)
-	// Cancelation is not an error: it is the only way to stop ipnserver.
-	if err != nil && err != context.Canceled {
-		return fmt.Errorf("ipnserver.Run: %w", err)
-	}
-
-	return nil
 }
 
 func createEngine(logf logger.Logf, linkMon *monitor.Mon, dialer *tsdial.Dialer) (e wgengine.Engine, useNetstack bool, err error) {
@@ -533,6 +563,8 @@ func shouldWrapNetstack() bool {
 	return false
 }
 
+var tstunNew = tstun.New
+
 func tryEngine(logf logger.Logf, linkMon *monitor.Mon, dialer *tsdial.Dialer, name string) (e wgengine.Engine, useNetstack bool, err error) {
 	conf := wgengine.Config{
 		ListenPort:  args.port,
@@ -564,7 +596,7 @@ func tryEngine(logf logger.Logf, linkMon *monitor.Mon, dialer *tsdial.Dialer, na
 			}
 		}
 	} else {
-		dev, devName, err := tstun.New(logf, name)
+		dev, devName, err := tstunNew(logf, name)
 		if err != nil {
 			tstun.Diagnose(logf, name, err)
 			return nil, false, fmt.Errorf("tstun.New(%q): %w", name, err)

cmd/tailscaled/tailscaled_windows.go

@@ -26,12 +26,10 @@ import (
 	"errors"
 	"fmt"
 	"log"
-	"net"
 	"net/netip"
 	"os"
 	"os/exec"
 	"os/signal"
-	"path/filepath"
 	"sync"
 	"syscall"
 	"time"
@@ -40,29 +38,17 @@ import (
 	"golang.org/x/sys/windows"
 	"golang.org/x/sys/windows/svc"
 	"golang.org/x/sys/windows/svc/eventlog"
+	"golang.zx2c4.com/wireguard/tun"
 	"golang.zx2c4.com/wireguard/windows/tunnel/winipcfg"
-	"tailscale.com/control/controlclient"
 	"tailscale.com/envknob"
-	"tailscale.com/ipn"
-	"tailscale.com/ipn/ipnlocal"
-	"tailscale.com/ipn/ipnserver"
-	"tailscale.com/ipn/store"
 	"tailscale.com/logpolicy"
 	"tailscale.com/logtail/backoff"
 	"tailscale.com/net/dns"
-	"tailscale.com/net/dnsfallback"
-	"tailscale.com/net/tsdial"
 	"tailscale.com/net/tstun"
-	"tailscale.com/safesocket"
-	"tailscale.com/smallzstd"
 	"tailscale.com/types/logger"
 	"tailscale.com/util/winutil"
 	"tailscale.com/version"
 	"tailscale.com/wf"
-	"tailscale.com/wgengine"
-	"tailscale.com/wgengine/monitor"
-	"tailscale.com/wgengine/netstack"
-	"tailscale.com/wgengine/router"
 )
 
 func init() {
@@ -78,6 +64,37 @@ func init() {
 
 const serviceName = "Tailscale"
 
+func init() {
+	tstunNew = tstunNewWithWindowsRetries
+}
+
+// tstunNewOrRetry is a wrapper around tstun.New that retries on Windows for certain
+// errors.
+//
+// TODO(bradfitz): move this into tstun and/or just fix the problems so it doesn't
+// require a few tries to work.
+func tstunNewWithWindowsRetries(logf logger.Logf, tunName string) (_ tun.Device, devName string, _ error) {
+	bo := backoff.NewBackoff("tstunNew", logf, 10*time.Second)
+	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Minute)
+	defer cancel()
+	for {
+		dev, devName, err := tstun.New(logf, tunName)
+		if err == nil {
+			return dev, devName, err
+		}
+		if errors.Is(err, windows.ERROR_DEVICE_NOT_AVAILABLE) || windowsUptime() < 10*time.Minute {
+			// Wintun is not installing correctly. Dump the state of NetSetupSvc
+			// (which is a user-mode service that must be active for network devices
+			// to install) and its dependencies to the log.
+			winutil.LogSvcState(logf, "NetSetupSvc")
+		}
+		bo.BackOff(ctx, err)
+		if ctx.Err() != nil {
+			return nil, "", ctx.Err()
+		}
+	}
+}
+
 func isWindowsService() bool {
 	v, err := svc.IsWindowsService()
 	if err != nil {
@@ -131,6 +148,7 @@ func (service *ipnService) Execute(args []string, r <-chan svc.ChangeRequest, ch
 	}
 
 	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
 	doneCh := make(chan struct{})
 	go func() {
 		defer close(doneCh)
@@ -236,7 +254,7 @@ func beWindowsSubprocess() bool {
 		}
 	}()
 
-	err := startIPNServer(context.Background(), logid)
+	err := startIPNServer(context.Background(), log.Printf, logid)
 	if err != nil {
 		log.Fatalf("ipnserver: %v", err)
 	}
@@ -283,140 +301,6 @@ func beFirewallKillswitch() bool {
 	}
 }
 
-func startIPNServer(ctx context.Context, logid string) error {
-	var logf logger.Logf = log.Printf
-
-	linkMon, err := monitor.New(logf)
-	if err != nil {
-		return fmt.Errorf("monitor: %w", err)
-	}
-	dialer := &tsdial.Dialer{Logf: logf}
-
-	getEngineRaw := func() (wgengine.Engine, *netstack.Impl, error) {
-		dev, devName, err := tstun.New(logf, "Tailscale")
-		if err != nil {
-			if errors.Is(err, windows.ERROR_DEVICE_NOT_AVAILABLE) {
-				// Wintun is not installing correctly. Dump the state of NetSetupSvc
-				// (which is a user-mode service that must be active for network devices
-				// to install) and its dependencies to the log.
-				winutil.LogSvcState(logf, "NetSetupSvc")
-			}
-			return nil, nil, fmt.Errorf("TUN: %w", err)
-		}
-		r, err := router.New(logf, dev, nil)
-		if err != nil {
-			dev.Close()
-			return nil, nil, fmt.Errorf("router: %w", err)
-		}
-		if shouldWrapNetstack() {
-			r = netstack.NewSubnetRouterWrapper(r)
-		}
-		d, err := dns.NewOSConfigurator(logf, devName)
-		if err != nil {
-			r.Close()
-			dev.Close()
-			return nil, nil, fmt.Errorf("DNS: %w", err)
-		}
-		eng, err := wgengine.NewUserspaceEngine(logf, wgengine.Config{
-			Tun:         dev,
-			Router:      r,
-			DNS:         d,
-			ListenPort:  41641,
-			LinkMonitor: linkMon,
-			Dialer:      dialer,
-		})
-		if err != nil {
-			r.Close()
-			dev.Close()
-			return nil, nil, fmt.Errorf("engine: %w", err)
-		}
-		ns, err := newNetstack(logf, dialer, eng)
-		if err != nil {
-			return nil, nil, fmt.Errorf("newNetstack: %w", err)
-		}
-		ns.ProcessLocalIPs = false
-		ns.ProcessSubnets = shouldWrapNetstack()
-		if err := ns.Start(); err != nil {
-			return nil, nil, fmt.Errorf("failed to start netstack: %w", err)
-		}
-		return wgengine.NewWatchdog(eng), ns, nil
-	}
-
-	type engineOrError struct {
-		Engine   wgengine.Engine
-		Netstack *netstack.Impl
-		Err      error
-	}
-	engErrc := make(chan engineOrError)
-	t0 := time.Now()
-	go func() {
-		const ms = time.Millisecond
-		for try := 1; ; try++ {
-			logf("tailscaled: getting engine... (try %v)", try)
-			t1 := time.Now()
-			eng, ns, err := getEngineRaw()
-			d, dt := time.Since(t1).Round(ms), time.Since(t1).Round(ms)
-			if err != nil {
-				logf("tailscaled: engine fetch error (try %v) in %v (total %v, sysUptime %v): %v",
-					try, d, dt, windowsUptime().Round(time.Second), err)
-			} else {
-				if try > 1 {
-					logf("tailscaled: got engine on try %v in %v (total %v)", try, d, dt)
-				} else {
-					logf("tailscaled: got engine in %v", d)
-				}
-			}
-			timer := time.NewTimer(5 * time.Second)
-			engErrc <- engineOrError{eng, ns, err}
-			if err == nil {
-				timer.Stop()
-				return
-			}
-			<-timer.C
-		}
-	}()
-
-	// getEngine is called by ipnserver to get the engine. It's
-	// not called concurrently and is not called again once it
-	// successfully returns an engine.
-	getEngine := func() (wgengine.Engine, *netstack.Impl, error) {
-		if msg := envknob.String("TS_DEBUG_WIN_FAIL"); msg != "" {
-			return nil, nil, fmt.Errorf("pretending to be a service failure: %v", msg)
-		}
-		for {
-			res := <-engErrc
-			if res.Engine != nil {
-				return res.Engine, res.Netstack, nil
-			}
-			if time.Since(t0) < time.Minute || windowsUptime() < 10*time.Minute {
-				// Ignore errors during early boot. Windows 10 auto logs in the GUI
-				// way sooner than the networking stack components start up.
-				// So the network will fail for a bit (and require a few tries) while
-				// the GUI is still fine.
-				continue
-			}
-			// Return nicer errors to users, annotated with logids, which helps
-			// when they file bugs.
-			return nil, nil, fmt.Errorf("%w\n\nlogid: %v", res.Err, logid)
-		}
-	}
-	store, err := store.New(logf, statePathOrDefault())
-	if err != nil {
-		return fmt.Errorf("store: %w", err)
-	}
-
-	ln, _, err := safesocket.Listen(args.socketpath, safesocket.WindowsLocalPort)
-	if err != nil {
-		return fmt.Errorf("safesocket.Listen: %v", err)
-	}
-
-	err = ipnServerRunWithRetries(ctx, logf, ln, store, linkMon, dialer, logid, getEngine, ipnServerOpts())
-	if err != nil {
-		logf("ipnserver.Run: %v", err)
-	}
-	return err
-}
-
 func handleSessionChange(chgRequest svc.ChangeRequest) {
 	if chgRequest.Cmd != svc.SessionChange || chgRequest.EventType != windows.WTS_SESSION_UNLOCK {
 		return
@@ -563,127 +447,3 @@ func babysitProc(ctx context.Context, args []string, logf logger.Logf) {
 		}
 	}
 }
-
-// getEngineUntilItWorksWrapper returns a getEngine wrapper that does
-// not call getEngine concurrently and stops calling getEngine once
-// it's returned a working engine.
-func getEngineUntilItWorksWrapper(getEngine func() (wgengine.Engine, *netstack.Impl, error)) func() (wgengine.Engine, *netstack.Impl, error) {
-	var mu sync.Mutex
-	var engGood wgengine.Engine
-	var nsGood *netstack.Impl
-	return func() (wgengine.Engine, *netstack.Impl, error) {
-		mu.Lock()
-		defer mu.Unlock()
-		if engGood != nil {
-			return engGood, nsGood, nil
-		}
-		e, ns, err := getEngine()
-		if err != nil {
-			return nil, nil, err
-		}
-		engGood = e
-		nsGood = ns
-		return e, ns, nil
-	}
-}
-
-// listenerWithReadyConn is a net.Listener wrapper that has
-// one net.Conn ready to be accepted already.
-type listenerWithReadyConn struct {
-	net.Listener
-
-	mu sync.Mutex
-	c  net.Conn // if non-nil, ready to be Accepted
-}
-
-func (ln *listenerWithReadyConn) Accept() (net.Conn, error) {
-	ln.mu.Lock()
-	c := ln.c
-	ln.c = nil
-	ln.mu.Unlock()
-	if c != nil {
-		return c, nil
-	}
-	return ln.Listener.Accept()
-}
-
-// ipnServerRunWithRetries runs a Tailscale backend service.
-//
-// The getEngine func is called repeatedly, once per connection, until it
-// returns an engine successfully.
-//
-// This works around issues on Windows with the wgengine.Engine/wintun creation
-// failing or hanging. See https://github.com/tailscale/tailscale/issues/6522.
-func ipnServerRunWithRetries(ctx context.Context, logf logger.Logf, ln net.Listener, store ipn.StateStore, linkMon *monitor.Mon, dialer *tsdial.Dialer, logid string, getEngine func() (wgengine.Engine, *netstack.Impl, error), opts serverOptions) error {
-	getEngine = getEngineUntilItWorksWrapper(getEngine)
-	runDone := make(chan struct{})
-	defer close(runDone)
-
-	// When the context is closed or when we return, whichever is first, close our listener
-	// and all open connections.
-	go func() {
-		select {
-		case <-ctx.Done():
-		case <-runDone:
-		}
-		ln.Close()
-	}()
-	logf("Listening on %v", ln.Addr())
-
-	bo := backoff.NewBackoff("ipnserver", logf, 30*time.Second)
-	var unservedConn net.Conn // if non-nil, accepted, but hasn't served yet
-
-	eng, ns, err := getEngine()
-	if err != nil {
-		logf("ipnserver: initial getEngine call: %v", err)
-		for i := 1; ctx.Err() == nil; i++ {
-			c, err := ln.Accept()
-			if err != nil {
-				logf("%d: Accept: %v", i, err)
-				bo.BackOff(ctx, err)
-				continue
-			}
-			logf("ipnserver: try%d: trying getEngine again...", i)
-			eng, ns, err = getEngine()
-			if err == nil {
-				logf("%d: GetEngine worked; exiting failure loop", i)
-				unservedConn = c
-				break
-			}
-			logf("ipnserver%d: getEngine failed again: %v", i, err)
-			// TODO(bradfitz): queue this error up for the next IPN bus watcher call
-			// to get for the Windows GUI? We used to send it over the pre-HTTP
-			// protocol to the Windows GUI. Just close it.
-			c.Close()
-		}
-		if err := ctx.Err(); err != nil {
-			return err
-		}
-	}
-	if unservedConn != nil {
-		ln = &listenerWithReadyConn{
-			Listener: ln,
-			c:        unservedConn,
-		}
-	}
-
-	server := ipnserver.New(logf, logid)
-
-	lb, err := ipnlocal.NewLocalBackend(logf, logid, store, "", dialer, eng, opts.LoginFlags)
-	if err != nil {
-		return fmt.Errorf("ipnlocal.NewLocalBackend: %w", err)
-	}
-	lb.SetVarRoot(opts.VarRoot)
-	if root := lb.TailscaleVarRoot(); root != "" {
-		dnsfallback.SetCachePath(filepath.Join(root, "derpmap.cached.json"))
-	}
-	lb.SetDecompressor(func() (controlclient.Decompressor, error) {
-		return smallzstd.NewDecoder(nil)
-	})
-
-	server.SetLocalBackend(lb)
-	if ns != nil {
-		ns.SetLocalBackend(lb)
-	}
-	return server.Run(ctx, ln)
-}

ipn/ipnserver/server.go

@@ -6,12 +6,14 @@ package ipnserver
 
 import (
 	"context"
+	"encoding/json"
 	"errors"
 	"fmt"
 	"io"
 	"net"
 	"net/http"
 	"os/user"
+	"strconv"
 	"strings"
 	"sync"
 	"sync/atomic"
@@ -25,6 +27,7 @@ import (
 	"tailscale.com/ipn/localapi"
 	"tailscale.com/types/logger"
 	"tailscale.com/util/mak"
+	"tailscale.com/util/set"
 	"tailscale.com/util/systemd"
 )
 
@@ -49,6 +52,7 @@ type Server struct {
 	mu            sync.Mutex
 	lastUserID    ipn.WindowsUserID // tracks last userid; on change, Reset state for paranoia
 	activeReqs    map[*http.Request]*ipnauth.ConnIdentity
+	backendWaiter set.HandleSet[context.CancelFunc] // values are wake-up funcs of lb waiters
 }
 
 func (s *Server) mustBackend() *ipnlocal.LocalBackend {
@@ -59,7 +63,64 @@ func (s *Server) mustBackend() *ipnlocal.LocalBackend {
 	return lb
 }
 
+func (s *Server) awaitBackend(ctx context.Context) (_ *ipnlocal.LocalBackend, ok bool) {
+	lb := s.lb.Load()
+	if lb != nil {
+		return lb, true
+	}
+	ctx, cancel := context.WithCancel(ctx)
+	defer cancel()
+
+	s.mu.Lock()
+	h := s.backendWaiter.Add(cancel)
+	s.mu.Unlock()
+	defer func() {
+		s.mu.Lock()
+		delete(s.backendWaiter, h)
+		s.mu.Unlock()
+	}()
+
+	// Try again, now that we've registered, in case there was a
+	// race.
+	lb = s.lb.Load()
+	if lb != nil {
+		return lb, true
+	}
+
+	<-ctx.Done()
+	lb = s.lb.Load()
+	return lb, lb != nil
+}
+
+// serveServerStatus serves the /server-status endpoint which reports whether
+// the LocalBackend is up yet.
+// This is primarily for the Windows GUI, because wintun can take awhile to
+// come up. See https://github.com/tailscale/tailscale/issues/6522.
+func (s *Server) serveServerStatus(w http.ResponseWriter, r *http.Request) {
+	ctx := r.Context()
+
+	w.Header().Set("Content-Type", "application/json")
+	var res struct {
+		Error string `json:"error,omitempty"`
+	}
+
+	lb := s.lb.Load()
+	if lb == nil {
+		w.WriteHeader(http.StatusServiceUnavailable)
+		if wait, _ := strconv.ParseBool(r.FormValue("wait")); wait {
+			w.(http.Flusher).Flush()
+			lb, _ = s.awaitBackend(ctx)
+		}
+	}
+
+	if lb == nil {
+		res.Error = "backend not ready"
+	}
+	json.NewEncoder(w).Encode(res)
+}
+
 func (s *Server) serveHTTP(w http.ResponseWriter, r *http.Request) {
+	ctx := r.Context()
 	if r.Method == "CONNECT" {
 		if envknob.GOOS() == "windows" {
 			// For the GUI client when using an exit node. See docs on handleProxyConnectConn.
@@ -70,11 +131,17 @@ func (s *Server) serveHTTP(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 
-	// TODO(bradfitz): add a status HTTP handler that returns whether there's a
+	// Check for this method before the awaitBackend call, as it reports whether
-	// LocalBackend yet, optionally blocking until there is one. See
+	// the backend is available.
-	// https://github.com/tailscale/tailscale/issues/6522
+	if r.Method == "GET" && r.URL.Path == "/server-status" {
-	lb := s.lb.Load()
+		s.serveServerStatus(w, r)
-	if lb == nil {
+		return
+	}
+
+	lb, ok := s.awaitBackend(ctx)
+	if !ok {
+		// Almost certainly because the context was canceled so the response
+		// here doesn't really matter. The client is gone.
 		http.Error(w, "no backend", http.StatusServiceUnavailable)
 		return
 	}
@@ -304,6 +371,13 @@ func (s *Server) SetLocalBackend(lb *ipnlocal.LocalBackend) {
 		panic("already set")
 	}
 	s.startBackendIfNeeded()
+
+	s.mu.Lock()
+	for _, wake := range s.backendWaiter {
+		wake() // they'll remove themselves when woken
+	}
+	s.mu.Unlock()
+
 	// TODO(bradfitz): send status update to GUI long poller waiter. See
 	// https://github.com/tailscale/tailscale/issues/6522
 }

tstest/integration/tailscaled_deps_test_darwin.go

@@ -36,6 +36,7 @@ import (
 	_ "tailscale.com/safesocket"
 	_ "tailscale.com/smallzstd"
 	_ "tailscale.com/ssh/tailssh"
+	_ "tailscale.com/syncs"
 	_ "tailscale.com/tailcfg"
 	_ "tailscale.com/tsweb"
 	_ "tailscale.com/types/flagtype"

tstest/integration/tailscaled_deps_test_freebsd.go

@@ -36,6 +36,7 @@ import (
 	_ "tailscale.com/safesocket"
 	_ "tailscale.com/smallzstd"
 	_ "tailscale.com/ssh/tailssh"
+	_ "tailscale.com/syncs"
 	_ "tailscale.com/tailcfg"
 	_ "tailscale.com/tsweb"
 	_ "tailscale.com/types/flagtype"

tstest/integration/tailscaled_deps_test_linux.go

@@ -36,6 +36,7 @@ import (
 	_ "tailscale.com/safesocket"
 	_ "tailscale.com/smallzstd"
 	_ "tailscale.com/ssh/tailssh"
+	_ "tailscale.com/syncs"
 	_ "tailscale.com/tailcfg"
 	_ "tailscale.com/tsweb"
 	_ "tailscale.com/types/flagtype"

tstest/integration/tailscaled_deps_test_openbsd.go

@@ -35,6 +35,7 @@ import (
 	_ "tailscale.com/paths"
 	_ "tailscale.com/safesocket"
 	_ "tailscale.com/smallzstd"
+	_ "tailscale.com/syncs"
 	_ "tailscale.com/tailcfg"
 	_ "tailscale.com/tsweb"
 	_ "tailscale.com/types/flagtype"

tstest/integration/tailscaled_deps_test_windows.go

@@ -16,6 +16,7 @@ import (
 	_ "golang.org/x/sys/windows/svc"
 	_ "golang.org/x/sys/windows/svc/eventlog"
 	_ "golang.org/x/sys/windows/svc/mgr"
+	_ "golang.zx2c4.com/wireguard/tun"
 	_ "golang.zx2c4.com/wireguard/windows/tunnel/winipcfg"
 	_ "tailscale.com/cmd/tailscaled/childproc"
 	_ "tailscale.com/control/controlclient"
@@ -41,6 +42,7 @@ import (
 	_ "tailscale.com/paths"
 	_ "tailscale.com/safesocket"
 	_ "tailscale.com/smallzstd"
+	_ "tailscale.com/syncs"
 	_ "tailscale.com/tailcfg"
 	_ "tailscale.com/tsweb"
 	_ "tailscale.com/types/flagtype"