tailscale/tsnet/tsnet.go

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

// Package tsnet provides Tailscale as a library.
package tsnet

import (
	"context"
	crand "crypto/rand"
	"crypto/tls"
	"encoding/hex"
	"errors"
	"fmt"
	"io"
	"log"
	"math"
	"net"
	"net/http"
	"net/netip"
	"os"
	"path/filepath"
	"runtime"
	"slices"
	"strconv"
	"strings"
	"sync"
	"time"

	"tailscale.com/client/local"
	"tailscale.com/control/controlclient"
	"tailscale.com/envknob"
	_ "tailscale.com/feature/c2n"
	_ "tailscale.com/feature/condregister/oauthkey"
	_ "tailscale.com/feature/condregister/portmapper"
	_ "tailscale.com/feature/condregister/useproxy"
	"tailscale.com/health"
	"tailscale.com/hostinfo"
	"tailscale.com/internal/client/tailscale"
	"tailscale.com/ipn"
	"tailscale.com/ipn/ipnauth"
	"tailscale.com/ipn/ipnlocal"
	"tailscale.com/ipn/ipnstate"
	"tailscale.com/ipn/localapi"
	"tailscale.com/ipn/store"
	"tailscale.com/ipn/store/mem"
	"tailscale.com/logpolicy"
	"tailscale.com/logtail"
	"tailscale.com/logtail/filch"
	"tailscale.com/net/memnet"
	"tailscale.com/net/netmon"
	"tailscale.com/net/proxymux"
	"tailscale.com/net/socks5"
	"tailscale.com/net/tsdial"
	"tailscale.com/tsd"
	"tailscale.com/types/bools"
	"tailscale.com/types/logger"
	"tailscale.com/types/logid"
	"tailscale.com/types/nettype"
	"tailscale.com/util/clientmetric"
	"tailscale.com/util/mak"
	"tailscale.com/util/set"
	"tailscale.com/util/testenv"
	"tailscale.com/wgengine"
	"tailscale.com/wgengine/netstack"
)

// Server is an embedded Tailscale server.
//
// Its exported fields may be changed until the first method call.
type Server struct {
	// Dir specifies the name of the directory to use for
	// state. If empty, a directory is selected automatically
	// under os.UserConfigDir (https://golang.org/pkg/os/#UserConfigDir).
	// based on the name of the binary.
	//
	// If you want to use multiple tsnet services in the same
	// binary, you will need to make sure that Dir is set uniquely
	// for each service. A good pattern for this is to have a
	// "base" directory (such as your mutable storage folder) and
	// then append the hostname on the end of it.
	Dir string

	// Store specifies the state store to use.
	//
	// If nil, a new FileStore is initialized at `Dir/tailscaled.state`.
	// See tailscale.com/ipn/store for supported stores.
	//
	// Logs will automatically be uploaded to log.tailscale.com,
	// where the configuration file for logging will be saved at
	// `Dir/tailscaled.log.conf`.
	Store ipn.StateStore

	// Hostname is the hostname to present to the control server.
	// If empty, the binary name is used.
	Hostname string

	// UserLogf, if non-nil, specifies the logger to use for logs generated by
	// the Server itself intended to be seen by the user such as the AuthURL for
	// login and status updates. If unset, log.Printf is used.
	UserLogf logger.Logf

	// Logf, if set is used for logs generated by the backend such as the
	// LocalBackend and MagicSock. It is verbose and intended for debugging.
	// If unset, logs are discarded.
	Logf logger.Logf

	// Ephemeral, if true, specifies that the instance should register
	// as an Ephemeral node (https://tailscale.com/s/ephemeral-nodes).
	Ephemeral bool

	// AuthKey, if non-empty, is the auth key to create the node
	// and will be preferred over the TS_AUTHKEY environment
	// variable. If the node is already created (from state
	// previously stored in Store), then this field is not
	// used.
	AuthKey string

	// ControlURL optionally specifies the coordination server URL.
	// If empty, the Tailscale default is used.
	ControlURL string

	// RunWebClient, if true, runs a client for managing this node over
	// its Tailscale interface on port 5252.
	RunWebClient bool

	// Port is the UDP port to listen on for WireGuard and peer-to-peer
	// traffic. If zero, a port is automatically selected. Leave this
	// field at zero unless you know what you are doing.
	Port uint16

	// AdvertiseTags specifies tags that should be applied to this node, for
	// purposes of ACL enforcement. These can be referenced from the ACL policy
	// document. Note that advertising a tag on the client doesn't guarantee
	// that the control server will allow the node to adopt that tag.
	AdvertiseTags []string

	getCertForTesting func(*tls.ClientHelloInfo) (*tls.Certificate, error)

	initOnce         sync.Once
	initErr          error
	lb               *ipnlocal.LocalBackend
	sys              *tsd.System
	netstack         *netstack.Impl
	netMon           *netmon.Monitor
	rootPath         string // the state directory
	hostname         string
	shutdownCtx      context.Context
	shutdownCancel   context.CancelFunc
	proxyCred        string        // SOCKS5 proxy auth for loopbackListener
	localAPICred     string        // basic auth password for loopbackListener
	loopbackListener net.Listener  // optional loopback for localapi and proxies
	localAPIListener net.Listener  // in-memory, used by localClient
	localClient      *local.Client // in-memory
	localAPIServer   *http.Server
	logbuffer        *filch.Filch
	logtail          *logtail.Logger
	logid            logid.PublicID

	mu                  sync.Mutex
	listeners           map[listenKey]*listener
	fallbackTCPHandlers set.HandleSet[FallbackTCPHandler]
	dialer              *tsdial.Dialer
	closed              bool
}

// FallbackTCPHandler describes the callback which
// conditionally handles an incoming TCP flow for the
// provided (src/port, dst/port) 4-tuple. These are registered
// as handlers of last resort, and are called only if no
// listener could handle the incoming flow.
//
// If the callback returns intercept=false, the flow is rejected.
//
// When intercept=true, the behavior depends on whether the returned handler
// is non-nil: if nil, the connection is rejected. If non-nil, handler takes
// over the TCP conn.
type FallbackTCPHandler func(src, dst netip.AddrPort) (handler func(net.Conn), intercept bool)

// Dial connects to the address on the tailnet.
// It will start the server if it has not been started yet.
func (s *Server) Dial(ctx context.Context, network, address string) (net.Conn, error) {
	if err := s.Start(); err != nil {
		return nil, err
	}
	if err := s.awaitRunning(ctx); err != nil {
		return nil, err
	}
	return s.dialer.UserDial(ctx, network, address)
}

// awaitRunning waits until the backend is in state Running.
// If the backend is in state Starting, it blocks until it reaches
// a terminal state (such as Stopped, NeedsMachineAuth)
// or the context expires.
func (s *Server) awaitRunning(ctx context.Context) error {
	st := s.lb.State()
	for {
		if err := ctx.Err(); err != nil {
			return err
		}
		switch st {
		case ipn.Running:
			return nil
		case ipn.NeedsLogin, ipn.Starting:
			// Even after LocalBackend.Start, the state machine is still briefly
			// in the "NeedsLogin" state. So treat that as also "Starting" and
			// wait for us to get out of that state.
			s.lb.WatchNotifications(ctx, ipn.NotifyInitialState, nil, func(n *ipn.Notify) (keepGoing bool) {
				if n.State != nil {
					st = *n.State
				}
				return st == ipn.NeedsLogin || st == ipn.Starting
			})
		default:
			return fmt.Errorf("tsnet: backend in state %v", st)
		}
	}
}

// HTTPClient returns an HTTP client that is configured to connect over Tailscale.
//
// This is useful if you need to have your tsnet services connect to other devices on
// your tailnet.
func (s *Server) HTTPClient() *http.Client {
	return &http.Client{
		Transport: &http.Transport{
			DialContext: s.Dial,
		},
	}
}

// LocalClient returns a LocalClient that speaks to s.
//
// It will start the server if it has not been started yet. If the server's
// already been started successfully, it doesn't return an error.
func (s *Server) LocalClient() (*local.Client, error) {
	if err := s.Start(); err != nil {
		return nil, err
	}
	return s.localClient, nil
}

// Loopback starts a routing server on a loopback address.
//
// The server has multiple functions.
//
// It can be used as a SOCKS5 proxy onto the tailnet.
// Authentication is required with the username "tsnet" and
// the value of proxyCred used as the password.
//
// The HTTP server also serves out the "LocalAPI" on /localapi.
// As the LocalAPI is powerful, access to endpoints requires BOTH passing a
// "Sec-Tailscale: localapi" HTTP header and passing localAPICred as basic auth.
//
// If you only need to use the LocalAPI from Go, then prefer LocalClient
// as it does not require communication via TCP.
func (s *Server) Loopback() (addr string, proxyCred, localAPICred string, err error) {
	if err := s.Start(); err != nil {
		return "", "", "", err
	}

	if s.loopbackListener == nil {
		var proxyCred [16]byte
		if _, err := crand.Read(proxyCred[:]); err != nil {
			return "", "", "", err
		}
		s.proxyCred = hex.EncodeToString(proxyCred[:])

		var cred [16]byte
		if _, err := crand.Read(cred[:]); err != nil {
			return "", "", "", err
		}
		s.localAPICred = hex.EncodeToString(cred[:])

		ln, err := net.Listen("tcp", "127.0.0.1:0")
		if err != nil {
			return "", "", "", err
		}
		s.loopbackListener = ln

		socksLn, httpLn := proxymux.SplitSOCKSAndHTTP(ln)

		// TODO: add HTTP proxy support. Probably requires factoring
		// out the CONNECT code from tailscaled/proxy.go that uses
		// httputil.ReverseProxy and adding auth support.
		go func() {
			lah := localapi.NewHandler(localapi.HandlerConfig{
				Actor:    ipnauth.Self,
				Backend:  s.lb,
				Logf:     s.logf,
				LogID:    s.logid,
				EventBus: s.sys.Bus.Get(),
			})
			lah.PermitWrite = true
			lah.PermitRead = true
			lah.RequiredPassword = s.localAPICred
			h := &localSecHandler{h: lah, cred: s.localAPICred}

			if err := http.Serve(httpLn, h); err != nil {
				s.logf("localapi tcp serve error: %v", err)
			}
		}()
		s5l := logger.WithPrefix(s.logf, "socks5: ")
		s5s := &socks5.Server{
			Logf:     s5l,
			Dialer:   s.dialer.UserDial,
			Username: "tsnet",
			Password: s.proxyCred,
		}
		go func() {
			s5l("SOCKS5 server exited: %v", s5s.Serve(socksLn))
		}()
	}

	lbAddr := s.loopbackListener.Addr()
	if lbAddr == nil {
		// https://github.com/tailscale/tailscale/issues/7488
		panic("loopbackListener has no Addr")
	}
	return lbAddr.String(), s.proxyCred, s.localAPICred, nil
}

type localSecHandler struct {
	h    http.Handler
	cred string
}

func (h *localSecHandler) ServeHTTP(w http.ResponseWriter, r *http.Request) {
	if r.Header.Get("Sec-Tailscale") != "localapi" {
		w.WriteHeader(403)
		io.WriteString(w, "missing 'Sec-Tailscale: localapi' header")
		return
	}
	h.h.ServeHTTP(w, r)
}

// Start connects the server to the tailnet.
// Optional: any calls to Dial/Listen will also call Start.
func (s *Server) Start() error {
	hostinfo.SetPackage("tsnet")
	s.initOnce.Do(s.doInit)
	return s.initErr
}

// Up connects the server to the tailnet and waits until it is running.
// On success it returns the current status, including a Tailscale IP address.
func (s *Server) Up(ctx context.Context) (*ipnstate.Status, error) {
	lc, err := s.LocalClient() // calls Start
	if err != nil {
		return nil, fmt.Errorf("tsnet.Up: %w", err)
	}

	watcher, err := lc.WatchIPNBus(ctx, ipn.NotifyInitialState|ipn.NotifyNoPrivateKeys)
	if err != nil {
		return nil, fmt.Errorf("tsnet.Up: %w", err)
	}
	defer watcher.Close()

	for {
		n, err := watcher.Next()
		if err != nil {
			return nil, fmt.Errorf("tsnet.Up: %w", err)
		}
		if n.ErrMessage != nil {
			return nil, fmt.Errorf("tsnet.Up: backend: %s", *n.ErrMessage)
		}
		if s := n.State; s != nil {
			if *s == ipn.Running {
				status, err := lc.Status(ctx)
				if err != nil {
					return nil, fmt.Errorf("tsnet.Up: %w", err)
				}
				if len(status.TailscaleIPs) == 0 {
					return nil, errors.New("tsnet.Up: running, but no ip")
				}

				// Clear the persisted serve config state to prevent stale configuration
				// from code changes. This is a temporary workaround until we have a better
				// way to handle this. (2023-03-11)
				if err := lc.SetServeConfig(ctx, new(ipn.ServeConfig)); err != nil {
					return nil, fmt.Errorf("tsnet.Up: %w", err)
				}

				return status, nil
			}
			// TODO: in the future, return an error on ipn.NeedsLogin
			// and ipn.NeedsMachineAuth to improve the UX of trying
			// out the tsnet package.
			//
			// Unfortunately today, even when using an AuthKey we
			// briefly see these states. It would be nice to fix.
		}
	}
}

// Close stops the server.
//
// It must not be called before or concurrently with Start.
func (s *Server) Close() error {
	s.mu.Lock()
	defer s.mu.Unlock()
	if s.closed {
		return fmt.Errorf("tsnet: %w", net.ErrClosed)
	}
	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
	defer cancel()
	var wg sync.WaitGroup
	wg.Add(1)
	go func() {
		defer wg.Done()
		// Perform a best-effort final flush.
		if s.logtail != nil {
			s.logtail.Shutdown(ctx)
		}
		if s.logbuffer != nil {
			s.logbuffer.Close()
		}
	}()
	wg.Add(1)
	go func() {
		defer wg.Done()
		if s.localAPIServer != nil {
			s.localAPIServer.Shutdown(ctx)
		}
	}()

	if s.netstack != nil {
		s.netstack.Close()
		s.netstack = nil
	}
	if s.shutdownCancel != nil {
		s.shutdownCancel()
	}
	if s.lb != nil {
		s.lb.Shutdown()
	}
	if s.netMon != nil {
		s.netMon.Close()
	}
	if s.dialer != nil {
		s.dialer.Close()
	}
	if s.localAPIListener != nil {
		s.localAPIListener.Close()
	}
	if s.loopbackListener != nil {
		s.loopbackListener.Close()
	}

	for _, ln := range s.listeners {
		ln.closeLocked()
	}
	wg.Wait()
	s.sys.Bus.Get().Close()
	s.closed = true
	return nil
}

func (s *Server) doInit() {
	s.shutdownCtx, s.shutdownCancel = context.WithCancel(context.Background())
	if err := s.start(); err != nil {
		s.initErr = fmt.Errorf("tsnet: %w", err)
	}
}

// CertDomains returns the list of domains for which the server can
// provide TLS certificates. These are also the DNS names for the
// Server.
// If the server is not running, it returns nil.
func (s *Server) CertDomains() []string {
	nm := s.lb.NetMap()
	if nm == nil {
		return nil
	}
	return slices.Clone(nm.DNS.CertDomains)
}

// TailscaleIPs returns IPv4 and IPv6 addresses for this node. If the node
// has not yet joined a tailnet or is otherwise unaware of its own IP addresses,
// the returned ip4, ip6 will be !netip.IsValid().
func (s *Server) TailscaleIPs() (ip4, ip6 netip.Addr) {
	nm := s.lb.NetMap()
	if nm == nil {
		return
	}
	addrs := nm.GetAddresses()
	for _, addr := range addrs.All() {
		ip := addr.Addr()
		if ip.Is6() {
			ip6 = ip
		}
		if ip.Is4() {
			ip4 = ip
		}
	}

	return ip4, ip6
}

// LogtailWriter returns an [io.Writer] that writes to Tailscale's logging service and will be only visible to Tailscale's
// support team. Logs written there cannot be retrieved by the user. This method always returns a non-nil value.
func (s *Server) LogtailWriter() io.Writer {
	if s.logtail == nil {
		return io.Discard
	}

	return s.logtail
}

func (s *Server) getAuthKey() string {
	if v := s.AuthKey; v != "" {
		return v
	}
	if v := os.Getenv("TS_AUTHKEY"); v != "" {
		return v
	}
	return os.Getenv("TS_AUTH_KEY")
}

func (s *Server) start() (reterr error) {
	var closePool closeOnErrorPool
	defer closePool.closeAllIfError(&reterr)

	exe, err := os.Executable()
	if err != nil {
		switch runtime.GOOS {
		case "js", "wasip1":
			// These platforms don't implement os.Executable (at least as of Go
			// 1.21), but we don't really care much: it's only used as a default
			// directory and hostname when they're not supplied. But we can fall
			// back to "tsnet" as well.
			exe = "tsnet"
		case "ios":
			// When compiled as a framework (via TailscaleKit in libtailscale),
			// os.Executable() returns an error, so fall back to "tsnet" there
			// too.
			exe = "tsnet"
		default:
			return err
		}
	}
	prog := strings.TrimSuffix(strings.ToLower(filepath.Base(exe)), ".exe")

	s.hostname = s.Hostname
	if s.hostname == "" {
		s.hostname = prog
	}

	s.rootPath = s.Dir
	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")
		}
	}

	if s.rootPath == "" {
		confDir, err := os.UserConfigDir()
		if err != nil {
			return err
		}
		s.rootPath = filepath.Join(confDir, "tsnet-"+prog)
	}
	if err := os.MkdirAll(s.rootPath, 0700); err != nil {
		return err
	}
	if fi, err := os.Stat(s.rootPath); err != nil {
		return err
	} else if !fi.IsDir() {
		return fmt.Errorf("%v is not a directory", s.rootPath)
	}

	tsLogf := func(format string, a ...any) {
		if s.logtail != nil {
			s.logtail.Logf(format, a...)
		}
		if s.Logf == nil {
			return
		}
		s.Logf(format, a...)
	}

	sys := tsd.NewSystem()
	s.sys = sys
	if err := s.startLogger(&closePool, sys.HealthTracker.Get(), tsLogf); err != nil {
		return err
	}

	s.netMon, err = netmon.New(sys.Bus.Get(), tsLogf)
	if err != nil {
		return err
	}
	closePool.add(s.netMon)

	s.dialer = &tsdial.Dialer{Logf: tsLogf} // mutated below (before used)
	s.dialer.SetBus(sys.Bus.Get())
	eng, err := wgengine.NewUserspaceEngine(tsLogf, wgengine.Config{
		EventBus:      sys.Bus.Get(),
		ListenPort:    s.Port,
		NetMon:        s.netMon,
		Dialer:        s.dialer,
		SetSubsystem:  sys.Set,
		ControlKnobs:  sys.ControlKnobs(),
		HealthTracker: sys.HealthTracker.Get(),
		Metrics:       sys.UserMetricsRegistry(),
	})
	if err != nil {
		return err
	}
	closePool.add(s.dialer)
	sys.Set(eng)
	sys.HealthTracker.Get().SetMetricsRegistry(sys.UserMetricsRegistry())

	// TODO(oxtoacart): do we need to support Taildrive on tsnet, and if so, how?
	ns, err := netstack.Create(tsLogf, sys.Tun.Get(), eng, sys.MagicSock.Get(), s.dialer, sys.DNSManager.Get(), sys.ProxyMapper())
	if err != nil {
		return fmt.Errorf("netstack.Create: %w", err)
	}
	sys.Tun.Get().Start()
	sys.Set(ns)
	ns.ProcessLocalIPs = true
	ns.ProcessSubnets = true
	ns.GetTCPHandlerForFlow = s.getTCPHandlerForFlow
	ns.GetUDPHandlerForFlow = s.getUDPHandlerForFlow
	s.netstack = ns
	s.dialer.UseNetstackForIP = func(ip netip.Addr) bool {
		_, ok := eng.PeerForIP(ip)
		return ok
	}
	s.dialer.NetstackDialTCP = func(ctx context.Context, dst netip.AddrPort) (net.Conn, error) {
		// Note: don't just return ns.DialContextTCP or we'll return
		// *gonet.TCPConn(nil) instead of a nil interface which trips up
		// callers.
		v4, v6 := s.TailscaleIPs()
		src := bools.IfElse(dst.Addr().Is6(), v6, v4)
		tcpConn, err := ns.DialContextTCPWithBind(ctx, src, dst)
		if err != nil {
			return nil, err
		}
		return tcpConn, nil
	}
	s.dialer.NetstackDialUDP = func(ctx context.Context, dst netip.AddrPort) (net.Conn, error) {
		// Note: don't just return ns.DialContextUDP or we'll return
		// *gonet.UDPConn(nil) instead of a nil interface which trips up
		// callers.
		v4, v6 := s.TailscaleIPs()
		src := bools.IfElse(dst.Addr().Is6(), v6, v4)
		udpConn, err := ns.DialContextUDPWithBind(ctx, src, dst)
		if err != nil {
			return nil, err
		}
		return udpConn, nil
	}

	if s.Store == nil {
		stateFile := filepath.Join(s.rootPath, "tailscaled.state")
		s.logf("tsnet running state path %s", stateFile)
		s.Store, err = store.New(tsLogf, stateFile)
		if err != nil {
			return err
		}
	}
	sys.Set(s.Store)

	loginFlags := controlclient.LoginDefault
	if s.Ephemeral {
		loginFlags = controlclient.LoginEphemeral
	}
	lb, err := ipnlocal.NewLocalBackend(tsLogf, s.logid, sys, loginFlags|controlclient.LocalBackendStartKeyOSNeutral)
	if err != nil {
		return fmt.Errorf("NewLocalBackend: %v", err)
	}
	lb.SetTCPHandlerForFunnelFlow(s.getTCPHandlerForFunnelFlow)
	lb.SetVarRoot(s.rootPath)
	s.logf("tsnet starting with hostname %q, varRoot %q", s.hostname, s.rootPath)
	s.lb = lb
	if err := ns.Start(lb); err != nil {
		return fmt.Errorf("failed to start netstack: %w", err)
	}
	closePool.addFunc(func() { s.lb.Shutdown() })
	prefs := ipn.NewPrefs()
	prefs.Hostname = s.hostname
	prefs.WantRunning = true
	prefs.ControlURL = s.ControlURL
	prefs.RunWebClient = s.RunWebClient
	prefs.AdvertiseTags = s.AdvertiseTags
	authKey := s.getAuthKey()
	// Try to use an OAuth secret to generate an auth key if that functionality
	// is available.
	if f, ok := tailscale.HookResolveAuthKey.GetOk(); ok {
		authKey, err = f(s.shutdownCtx, s.getAuthKey(), prefs.AdvertiseTags)
		if err != nil {
			return fmt.Errorf("resolving auth key: %w", err)
		}
	}
	err = lb.Start(ipn.Options{
		UpdatePrefs: prefs,
		AuthKey:     authKey,
	})
	if err != nil {
		return fmt.Errorf("starting backend: %w", err)
	}
	st := lb.State()
	if st == ipn.NeedsLogin || envknob.Bool("TSNET_FORCE_LOGIN") {
		s.logf("LocalBackend state is %v; running StartLoginInteractive...", st)
		if err := s.lb.StartLoginInteractive(s.shutdownCtx); err != nil {
			return fmt.Errorf("StartLoginInteractive: %w", err)
		}
	} else if authKey != "" {
		s.logf("Authkey is set; but state is %v. Ignoring authkey. Re-run with TSNET_FORCE_LOGIN=1 to force use of authkey.", st)
	}
	go s.printAuthURLLoop()

	// Run the localapi handler, to allow fetching LetsEncrypt certs.
	lah := localapi.NewHandler(localapi.HandlerConfig{
		Actor:    ipnauth.Self,
		Backend:  lb,
		Logf:     tsLogf,
		LogID:    s.logid,
		EventBus: sys.Bus.Get(),
	})
	lah.PermitWrite = true
	lah.PermitRead = true

	// Create an in-process listener.
	// nettest.Listen provides a in-memory pipe based implementation for net.Conn.
	lal := memnet.Listen("local-tailscaled.sock:80")
	s.localAPIListener = lal
	s.localClient = &local.Client{Dial: lal.Dial}
	s.localAPIServer = &http.Server{Handler: lah}
	s.lb.ConfigureWebClient(s.localClient)
	go func() {
		if err := s.localAPIServer.Serve(lal); err != nil && err != http.ErrServerClosed {
			s.logf("localapi serve error: %v", err)
		}
	}()
	closePool.add(s.localAPIListener)
	return nil
}

func (s *Server) startLogger(closePool *closeOnErrorPool, health *health.Tracker, tsLogf logger.Logf) error {
	if testenv.InTest() {
		return nil
	}
	cfgPath := filepath.Join(s.rootPath, "tailscaled.log.conf")
	lpc, err := logpolicy.ConfigFromFile(cfgPath)
	switch {
	case os.IsNotExist(err):
		lpc = logpolicy.NewConfig(logtail.CollectionNode)
		if err := lpc.Save(cfgPath); err != nil {
			return fmt.Errorf("logpolicy.Config.Save for %v: %w", cfgPath, err)
		}
	case err != nil:
		return fmt.Errorf("logpolicy.LoadConfig for %v: %w", cfgPath, err)
	}
	if err := lpc.Validate(logtail.CollectionNode); err != nil {
		return fmt.Errorf("logpolicy.Config.Validate for %v: %w", cfgPath, err)
	}
	s.logid = lpc.PublicID

	s.logbuffer, err = filch.New(filepath.Join(s.rootPath, "tailscaled"), filch.Options{ReplaceStderr: false})
	if err != nil {
		return fmt.Errorf("error creating filch: %w", err)
	}
	closePool.add(s.logbuffer)
	c := logtail.Config{
		Collection:   lpc.Collection,
		PrivateID:    lpc.PrivateID,
		Stderr:       io.Discard, // log everything to Buffer
		Buffer:       s.logbuffer,
		CompressLogs: true,
		Bus:          s.sys.Bus.Get(),
		HTTPC:        &http.Client{Transport: logpolicy.NewLogtailTransport(logtail.DefaultHost, s.netMon, health, tsLogf)},
		MetricsDelta: clientmetric.EncodeLogTailMetricsDelta,
	}
	s.logtail = logtail.NewLogger(c, tsLogf)
	closePool.addFunc(func() { s.logtail.Shutdown(context.Background()) })
	return nil
}

type closeOnErrorPool []func()

func (p *closeOnErrorPool) add(c io.Closer)   { *p = append(*p, func() { c.Close() }) }
func (p *closeOnErrorPool) addFunc(fn func()) { *p = append(*p, fn) }
func (p closeOnErrorPool) closeAllIfError(errp *error) {
	if *errp != nil {
		for _, closeFn := range p {
			closeFn()
		}
	}
}

func (s *Server) logf(format string, a ...any) {
	if s.logtail != nil {
		s.logtail.Logf(format, a...)
	}
	if s.UserLogf != nil {
		s.UserLogf(format, a...)
		return
	}
	log.Printf(format, a...)
}

// printAuthURLLoop loops once every few seconds while the server is still running and
// is in NeedsLogin state, printing out the auth URL.
func (s *Server) printAuthURLLoop() {
	for {
		if s.shutdownCtx.Err() != nil {
			return
		}
		if st := s.lb.State(); st != ipn.NeedsLogin && st != ipn.NoState {
			s.logf("AuthLoop: state is %v; done", st)
			return
		}
		st := s.lb.StatusWithoutPeers()
		if st.AuthURL != "" {
			s.logf("To start this tsnet server, restart with TS_AUTHKEY set, or go to: %s", st.AuthURL)
		}
		select {
		case <-time.After(5 * time.Second):
		case <-s.shutdownCtx.Done():
			return
		}
	}
}

// networkForFamily returns one of "tcp4", "tcp6", "udp4", or "udp6".
//
// netBase is "tcp" or "udp" (without any '4' or '6' suffix).
func networkForFamily(netBase string, is6 bool) string {
	switch netBase {
	case "tcp":
		if is6 {
			return "tcp6"
		}
		return "tcp4"
	case "udp":
		if is6 {
			return "udp6"
		}
		return "udp4"
	}
	panic("unexpected")
}

// listenerForDstAddr returns a listener for the provided network and
// destination IP/port. It matches from most specific to least specific.
// For example:
//
//   - ("tcp4", IP, port)
//   - ("tcp", IP, port)
//   - ("tcp4", "", port)
//   - ("tcp", "", port)
//
// The netBase is "tcp" or "udp" (without any '4' or '6' suffix).
//
// Listeners which do not specify an IP address will match for traffic
// for the local node (that is, a destination address of the IPv4 or
// IPv6 address of this node) only. To listen for traffic on other addresses
// such as those routed inbound via subnet routes, explicitly specify
// the listening address or use RegisterFallbackTCPHandler.
func (s *Server) listenerForDstAddr(netBase string, dst netip.AddrPort, funnel bool) (_ *listener, ok bool) {
	s.mu.Lock()
	defer s.mu.Unlock()

	// Search for a listener with the specified IP
	for _, net := range [2]string{
		networkForFamily(netBase, dst.Addr().Is6()),
		netBase,
	} {
		if ln, ok := s.listeners[listenKey{net, dst.Addr(), dst.Port(), funnel}]; ok {
			return ln, true
		}
	}

	// Search for a listener without an IP if the destination was
	// one of the native IPs of the node.
	if ip4, ip6 := s.TailscaleIPs(); dst.Addr() == ip4 || dst.Addr() == ip6 {
		for _, net := range [2]string{
			networkForFamily(netBase, dst.Addr().Is6()),
			netBase,
		} {
			if ln, ok := s.listeners[listenKey{net, netip.Addr{}, dst.Port(), funnel}]; ok {
				return ln, true
			}
		}
	}

	return nil, false
}

func (s *Server) getTCPHandlerForFunnelFlow(src netip.AddrPort, dstPort uint16) (handler func(net.Conn)) {
	ipv4, ipv6 := s.TailscaleIPs()
	var dst netip.AddrPort
	if src.Addr().Is4() {
		if !ipv4.IsValid() {
			return nil
		}
		dst = netip.AddrPortFrom(ipv4, dstPort)
	} else {
		if !ipv6.IsValid() {
			return nil
		}
		dst = netip.AddrPortFrom(ipv6, dstPort)
	}
	ln, ok := s.listenerForDstAddr("tcp", dst, true)
	if !ok {
		return nil
	}
	return ln.handle
}

func (s *Server) getTCPHandlerForFlow(src, dst netip.AddrPort) (handler func(net.Conn), intercept bool) {
	ln, ok := s.listenerForDstAddr("tcp", dst, false)
	if !ok {
		s.mu.Lock()
		defer s.mu.Unlock()
		for _, handler := range s.fallbackTCPHandlers {
			connHandler, intercept := handler(src, dst)
			if intercept {
				return connHandler, intercept
			}
		}
		return nil, true // don't handle, don't forward to localhost
	}
	return ln.handle, true
}

func (s *Server) getUDPHandlerForFlow(src, dst netip.AddrPort) (handler func(nettype.ConnPacketConn), intercept bool) {
	ln, ok := s.listenerForDstAddr("udp", dst, false)
	if !ok {
		return nil, true // don't handle, don't forward to localhost
	}
	return func(c nettype.ConnPacketConn) { ln.handle(c) }, true
}

// Listen announces only on the Tailscale network.
// It will start the server if it has not been started yet.
//
// Listeners which do not specify an IP address will match for traffic
// for the local node (that is, a destination address of the IPv4 or
// IPv6 address of this node) only. To listen for traffic on other addresses
// such as those routed inbound via subnet routes, explicitly specify
// the listening address or use RegisterFallbackTCPHandler.
func (s *Server) Listen(network, addr string) (net.Listener, error) {
	return s.listen(network, addr, listenOnTailnet)
}

// ListenPacket announces on the Tailscale network.
//
// The network must be "udp", "udp4" or "udp6". The addr must be of the form
// "ip:port" (or "[ip]:port") where ip is a valid IPv4 or IPv6 address
// corresponding to "udp4" or "udp6" respectively. IP must be specified.
//
// If s has not been started yet, it will be started.
func (s *Server) ListenPacket(network, addr string) (net.PacketConn, error) {
	ap, err := resolveListenAddr(network, addr)
	if err != nil {
		return nil, err
	}
	if !ap.Addr().IsValid() {
		return nil, fmt.Errorf("tsnet.ListenPacket(%q, %q): address must be a valid IP", network, addr)
	}
	if network == "udp" {
		if ap.Addr().Is4() {
			network = "udp4"
		} else {
			network = "udp6"
		}
	}
	if err := s.Start(); err != nil {
		return nil, err
	}
	return s.netstack.ListenPacket(network, ap.String())
}

// ListenTLS announces only on the Tailscale network.
// It returns a TLS listener wrapping the tsnet listener.
// It will start the server if it has not been started yet.
func (s *Server) ListenTLS(network, addr string) (net.Listener, error) {
	if network != "tcp" {
		return nil, fmt.Errorf("ListenTLS(%q, %q): only tcp is supported", network, addr)
	}
	ctx := context.Background()
	st, err := s.Up(ctx)
	if err != nil {
		return nil, err
	}
	if !st.CurrentTailnet.MagicDNSEnabled {
		return nil, errors.New("tsnet: you must enable MagicDNS in the DNS page of the admin panel to proceed. See https://tailscale.com/s/https")
	}
	if len(st.CertDomains) == 0 {
		return nil, errors.New("tsnet: you must enable HTTPS in the admin panel to proceed. See https://tailscale.com/s/https")
	}

	ln, err := s.listen(network, addr, listenOnTailnet)
	if err != nil {
		return nil, err
	}
	return tls.NewListener(ln, &tls.Config{
		GetCertificate: s.getCert,
	}), nil
}

// RegisterFallbackTCPHandler registers a callback which will be called
// to handle a TCP flow to this tsnet node, for which no listeners will handle.
//
// If multiple fallback handlers are registered, they will be called in an
// undefined order. See FallbackTCPHandler for details on handling a flow.
//
// The returned function can be used to deregister this callback.
func (s *Server) RegisterFallbackTCPHandler(cb FallbackTCPHandler) func() {
	s.mu.Lock()
	defer s.mu.Unlock()
	hnd := s.fallbackTCPHandlers.Add(cb)
	return func() {
		s.mu.Lock()
		defer s.mu.Unlock()
		delete(s.fallbackTCPHandlers, hnd)
	}
}

// getCert is the GetCertificate function used by ListenTLS.
//
// It calls GetCertificate on the localClient, passing in the ClientHelloInfo.
// For testing, if s.getCertForTesting is set, it will call that instead.
func (s *Server) getCert(hi *tls.ClientHelloInfo) (*tls.Certificate, error) {
	if s.getCertForTesting != nil {
		return s.getCertForTesting(hi)
	}
	lc, err := s.LocalClient()
	if err != nil {
		return nil, err
	}
	return lc.GetCertificate(hi)
}

// FunnelOption is an option passed to ListenFunnel to configure the listener.
type FunnelOption interface {
	funnelOption()
}

type funnelOnly struct{}

func (funnelOnly) funnelOption() {}

// FunnelOnly configures the listener to only respond to connections from Tailscale Funnel.
// The local tailnet will not be able to connect to the listener.
func FunnelOnly() FunnelOption { return funnelOnly{} }

type funnelTLSConfig struct{ conf *tls.Config }

func (f funnelTLSConfig) funnelOption() {}

// FunnelTLSConfig configures the TLS configuration for [Server.ListenFunnel]
//
// This is rarely needed but can permit requiring client certificates, specific
// ciphers suites, etc.
//
// The provided conf should at least be able to get a certificate, setting
// GetCertificate, Certificates or GetConfigForClient appropriately.
// The most common configuration is to set GetCertificate to
// Server.LocalClient's GetCertificate method.
//
// Unless [FunnelOnly] is also used, the configuration is also used for
// in-tailnet connections that don't arrive over Funnel.
func FunnelTLSConfig(conf *tls.Config) FunnelOption {
	return funnelTLSConfig{conf: conf}
}

// ListenFunnel announces on the public internet using Tailscale Funnel.
//
// It also by default listens on your local tailnet, so connections can
// come from either inside or outside your network. To restrict connections
// to be just from the internet, use the FunnelOnly option.
//
// Currently (2023-03-10), Funnel only supports TCP on ports 443, 8443, and 10000.
// The supported host name is limited to that configured for the tsnet.Server.
// As such, the standard way to create funnel is:
//
//	s.ListenFunnel("tcp", ":443")
//
// and the only other supported addrs currently are ":8443" and ":10000".
//
// It will start the server if it has not been started yet.
func (s *Server) ListenFunnel(network, addr string, opts ...FunnelOption) (net.Listener, error) {
	if network != "tcp" {
		return nil, fmt.Errorf("ListenFunnel(%q, %q): only tcp is supported", network, addr)
	}
	host, portStr, err := net.SplitHostPort(addr)
	if err != nil {
		return nil, err
	}
	if host != "" {
		return nil, fmt.Errorf("ListenFunnel(%q, %q): host must be empty", network, addr)
	}
	port, err := strconv.ParseUint(portStr, 10, 16)
	if err != nil {
		return nil, err
	}

	// Process, validate opts.
	lnOn := listenOnBoth
	var tlsConfig *tls.Config
	for _, opt := range opts {
		switch v := opt.(type) {
		case funnelTLSConfig:
			if v.conf == nil {
				return nil, errors.New("invalid nil FunnelTLSConfig")
			}
			tlsConfig = v.conf
		case funnelOnly:
			lnOn = listenOnFunnel
		default:
			return nil, fmt.Errorf("unknown opts FunnelOption type %T", v)
		}
	}
	if tlsConfig == nil {
		tlsConfig = &tls.Config{GetCertificate: s.getCert}
	}

	ctx := context.Background()
	st, err := s.Up(ctx)
	if err != nil {
		return nil, err
	}
	// TODO(sonia,tailscale/corp#10577): We may want to use the interactive enable
	// flow here instead of CheckFunnelAccess to allow the user to turn on Funnel
	// if not already on. Specifically when running from a terminal.
	// See cli.serveEnv.verifyFunnelEnabled.
	if err := ipn.CheckFunnelAccess(uint16(port), st.Self); err != nil {
		return nil, err
	}

	lc := s.localClient

	// May not have funnel enabled. Enable it.
	srvConfig, err := lc.GetServeConfig(ctx)
	if err != nil {
		return nil, err
	}
	if srvConfig == nil {
		srvConfig = &ipn.ServeConfig{}
	}
	if len(st.CertDomains) == 0 {
		return nil, errors.New("Funnel not available; HTTPS must be enabled. See https://tailscale.com/s/https")
	}
	domain := st.CertDomains[0]
	hp := ipn.HostPort(domain + ":" + portStr)
	if !srvConfig.AllowFunnel[hp] {
		mak.Set(&srvConfig.AllowFunnel, hp, true)
		srvConfig.AllowFunnel[hp] = true
		if err := lc.SetServeConfig(ctx, srvConfig); err != nil {
			return nil, err
		}
	}

	// Start a funnel listener.
	ln, err := s.listen(network, addr, lnOn)
	if err != nil {
		return nil, err
	}
	return tls.NewListener(ln, tlsConfig), nil
}

type listenOn string

const (
	listenOnTailnet = listenOn("listen-on-tailnet")
	listenOnFunnel  = listenOn("listen-on-funnel")
	listenOnBoth    = listenOn("listen-on-both")
)

// resolveListenAddr resolves a network and address into a netip.AddrPort. The
// returned netip.AddrPort.Addr will be the zero value if the address is empty.
// The port must be a valid port number. The caller is responsible for checking
// the network and address are valid.
//
// It resolves well-known port names and validates the address is a valid IP
// literal for the network.
func resolveListenAddr(network, addr string) (netip.AddrPort, error) {
	var zero netip.AddrPort
	host, portStr, err := net.SplitHostPort(addr)
	if err != nil {
		return zero, fmt.Errorf("tsnet: %w", err)
	}
	port, err := net.LookupPort(network, portStr)
	if err != nil || port < 0 || port > math.MaxUint16 {
		// LookupPort returns an error on out of range values so the bounds
		// checks on port should be unnecessary, but harmless. If they do
		// match, worst case this error message says "invalid port: <nil>".
		return zero, fmt.Errorf("invalid port: %w", err)
	}
	if host == "" {
		return netip.AddrPortFrom(netip.Addr{}, uint16(port)), nil
	}

	bindHostOrZero, err := netip.ParseAddr(host)
	if err != nil {
		return zero, fmt.Errorf("invalid Listen addr %q; host part must be empty or IP literal", host)
	}
	if strings.HasSuffix(network, "4") && !bindHostOrZero.Is4() {
		return zero, fmt.Errorf("invalid non-IPv4 addr %v for network %q", host, network)
	}
	if strings.HasSuffix(network, "6") && !bindHostOrZero.Is6() {
		return zero, fmt.Errorf("invalid non-IPv6 addr %v for network %q", host, network)
	}
	return netip.AddrPortFrom(bindHostOrZero, uint16(port)), nil
}

func (s *Server) listen(network, addr string, lnOn listenOn) (net.Listener, error) {
	switch network {
	case "", "tcp", "tcp4", "tcp6", "udp", "udp4", "udp6":
	default:
		return nil, errors.New("unsupported network type")
	}
	host, err := resolveListenAddr(network, addr)
	if err != nil {
		return nil, err
	}
	if err := s.Start(); err != nil {
		return nil, err
	}
	var keys []listenKey
	switch lnOn {
	case listenOnTailnet:
		keys = append(keys, listenKey{network, host.Addr(), host.Port(), false})
	case listenOnFunnel:
		keys = append(keys, listenKey{network, host.Addr(), host.Port(), true})
	case listenOnBoth:
		keys = append(keys, listenKey{network, host.Addr(), host.Port(), false})
		keys = append(keys, listenKey{network, host.Addr(), host.Port(), true})
	}

	ln := &listener{
		s:    s,
		keys: keys,
		addr: addr,

		closedc: make(chan struct{}),
		conn:    make(chan net.Conn),
	}
	s.mu.Lock()
	for _, key := range keys {
		if _, ok := s.listeners[key]; ok {
			s.mu.Unlock()
			return nil, fmt.Errorf("tsnet: listener already open for %s, %s", network, addr)
		}
	}
	if s.listeners == nil {
		s.listeners = make(map[listenKey]*listener)
	}
	for _, key := range keys {
		s.listeners[key] = ln
	}
	s.mu.Unlock()
	return ln, nil
}

// GetRootPath returns the root path of the tsnet server.
// This is where the state file and other data is stored.
func (s *Server) GetRootPath() string {
	return s.rootPath
}

// CapturePcap can be called by the application code compiled with tsnet to save a pcap
// of packets which the netstack within tsnet sees. This is expected to be useful during
// debugging, probably not useful for production.
//
// Packets will be written to the pcap until the process exits. The pcap needs a Lua dissector
// to be installed in WireShark in order to decode properly: wgengine/capture/ts-dissector.lua
// in this repository.
// https://tailscale.com/kb/1023/troubleshooting/#can-i-examine-network-traffic-inside-the-encrypted-tunnel
func (s *Server) CapturePcap(ctx context.Context, pcapFile string) error {
	stream, err := s.localClient.StreamDebugCapture(ctx)
	if err != nil {
		return err
	}

	f, err := os.OpenFile(pcapFile, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, 0644)
	if err != nil {
		stream.Close()
		return err
	}

	go func(stream io.ReadCloser, f *os.File) {
		defer stream.Close()
		defer f.Close()
		_, _ = io.Copy(f, stream)
	}(stream, f)

	return nil
}

// Sys returns a handle to the Tailscale subsystems of this node.
//
// This is not a stable API, nor are the APIs of the returned subsystems.
func (s *Server) Sys() *tsd.System {
	return s.sys
}

type listenKey struct {
	network string
	host    netip.Addr // or zero value for unspecified
	port    uint16
	funnel  bool
}

type listener struct {
	s       *Server
	keys    []listenKey
	addr    string
	conn    chan net.Conn // unbuffered, never closed
	closedc chan struct{} // closed on [listener.Close]
	closed  bool          // guarded by s.mu
}

func (ln *listener) Accept() (net.Conn, error) {
	select {
	case c := <-ln.conn:
		return c, nil
	case <-ln.closedc:
		return nil, fmt.Errorf("tsnet: %w", net.ErrClosed)
	}
}

func (ln *listener) Addr() net.Addr { return addr{ln} }

func (ln *listener) Close() error {
	ln.s.mu.Lock()
	defer ln.s.mu.Unlock()
	return ln.closeLocked()
}

// closeLocked closes the listener.
// It must be called with ln.s.mu held.
func (ln *listener) closeLocked() error {
	if ln.closed {
		return fmt.Errorf("tsnet: %w", net.ErrClosed)
	}
	for _, key := range ln.keys {
		if v, ok := ln.s.listeners[key]; ok && v == ln {
			delete(ln.s.listeners, key)
		}
	}
	close(ln.closedc)
	ln.closed = true
	return nil
}

func (ln *listener) handle(c net.Conn) {
	select {
	case ln.conn <- c:
		return
	case <-ln.closedc:
	case <-ln.s.shutdownCtx.Done():
	case <-time.After(time.Second):
		// TODO(bradfitz): this isn't ideal. Think about how
		// we how we want to do pushback.
	}
	c.Close()
}

// Server returns the tsnet Server associated with the listener.
func (ln *listener) Server() *Server { return ln.s }

type addr struct{ ln *listener }

func (a addr) Network() string { return a.ln.keys[0].network }
func (a addr) String() string  { return a.ln.addr }