cmd/tailscaled: run off internal state autonomously.

With this change, tailscaled can be restarted and reconnect
without interaction from `tailscale`, and `tailscale` is merely
there to provide login assistance and adjust preferences.

Signed-off-by: David Anderson <dave@natulte.net>

cmd/tailscale/tailscale.go

@@ -8,9 +8,7 @@ package main // import "tailscale.com/cmd/tailscale"
 
 import (
 	"context"
-	"encoding/json"
 	"fmt"
-	"io/ioutil"
 	"log"
 	"net"
 	"os"
@@ -19,12 +17,20 @@ import (
 
 	"github.com/apenwarr/fixconsole"
 	"github.com/pborman/getopt/v2"
-	"tailscale.com/atomicfile"
 	"tailscale.com/ipn"
 	"tailscale.com/logpolicy"
 	"tailscale.com/safesocket"
 )
 
+// globalStateKey is the ipn.StateKey that tailscaled loads on
+// startup.
+//
+// We have to support multiple state keys for other OSes (Windows in
+// particular), but right now Unix daemons run with a single
+// node-global state. To keep open the option of having per-user state
+// later, the global state key doesn't look like a username.
+const globalStateKey = "_daemon"
+
 // pump receives backend messages on conn and pushes them into bc.
 func pump(ctx context.Context, bc *ipn.BackendClient, conn net.Conn) {
 	defer log.Printf("Control connection done.\n")
@@ -45,34 +51,26 @@ func main() {
 		log.Printf("fixConsoleOutput: %v\n", err)
 	}
 
-	config := getopt.StringLong("config", 'f', "", "path to config file")
 	server := getopt.StringLong("server", 's', "https://login.tailscale.com", "URL to tailcontrol server")
 	nuroutes := getopt.BoolLong("no-single-routes", 'N', "disallow (non-subnet) routes to single nodes")
-	rroutes := getopt.BoolLong("remote-routes", 'R', "allow routing subnets to remote nodes")
+	routeall := getopt.BoolLong("remote-routes", 'R', "accept routes advertised by remote nodes")
-	droutes := getopt.BoolLong("default-routes", 'D', "allow default route on remote node")
+	nopf := getopt.BoolLong("no-packet-filter", 'F', "disable packet filter")
 	getopt.Parse()
-	if *config == "" {
+	pol := logpolicy.New("tailnode.log.tailscale.io", "tailscale")
-		logpolicy.New("tailnode.log.tailscale.io", "tailscale")
-		log.Fatal("no --config provided")
-	}
 	if len(getopt.Args()) > 0 {
 		log.Fatalf("too many non-flag arguments: %#v", getopt.Args()[0])
 	}
 
-	pol := logpolicy.New("tailnode.log.tailscale.io", *config)
 	defer pol.Close()
 
-	localCfg, err := loadConfig(*config)
-	if err != nil {
-		log.Fatal(err)
-	}
-
 	// TODO(apenwarr): fix different semantics between prefs and uflags
 	// TODO(apenwarr): allow setting/using CorpDNS
-	prefs := &localCfg
+	prefs := ipn.Prefs{
-	prefs.WantRunning = true
+		WantRunning:      true,
-	prefs.RouteAll = *rroutes || *droutes
+		RouteAll:         *routeall,
-	prefs.AllowSingleHosts = !*nuroutes
+		AllowSingleHosts: !*nuroutes,
+		UsePacketFilter:  !*nopf,
+	}
 
 	c, err := safesocket.Connect("", "Tailscale", "tailscaled", 41112)
 	if err != nil {
@@ -83,6 +81,7 @@ func main() {
 	}
 
 	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
 
 	go func() {
 		interrupt := make(chan os.Signal, 1)
@@ -92,11 +91,11 @@ func main() {
 	}()
 
 	bc := ipn.NewBackendClient(log.Printf, clientToServer)
+	bc.SetPrefs(prefs)
 	opts := ipn.Options{
-		Prefs:     prefs,
+		StateKey:  globalStateKey,
 		ServerURL: *server,
 		Notify: func(n ipn.Notify) {
-			log.Printf("Notify: %v\n", n)
 			if n.ErrMessage != nil {
 				log.Fatalf("backend error: %v\n", *n.ErrMessage)
 			}
@@ -108,41 +107,22 @@ func main() {
 					fmt.Fprintf(os.Stderr, "\nTo authorize your machine, visit (as admin):\n\n\t%s/admin/machines\n\n", *server)
 				case ipn.Starting, ipn.Running:
 					// Done full authentication process
+					fmt.Fprintf(os.Stderr, "\ntailscaled is authenticated, nothing more to do.\n\n")
 					cancel()
 				}
 			}
 			if url := n.BrowseToURL; url != nil {
 				fmt.Fprintf(os.Stderr, "\nTo authenticate, visit:\n\n\t%s\n\n", *url)
 			}
-			if p := n.Prefs; p != nil {
-				prefs = p
-				saveConfig(*config, *p)
-			}
 		},
 	}
+	// We still have to Start right now because it's the only way to
+	// set up notifications and whatnot. This causes a bunch of churn
+	// every time the CLI touches anything.
+	//
+	// TODO(danderson): redo the frontend/backend API to assume
+	// ephemeral frontends that read/modify/write state, once
+	// Windows/Mac state is moved into backend.
 	bc.Start(opts)
 	pump(ctx, bc, c)
 }
-
-func loadConfig(path string) (ipn.Prefs, error) {
-	b, err := ioutil.ReadFile(path)
-	if os.IsNotExist(err) {
-		log.Printf("config %s does not exist", path)
-		return ipn.NewPrefs(), nil
-	}
-	return ipn.PrefsFromBytes(b, false)
-}
-
-func saveConfig(path string, prefs ipn.Prefs) error {
-	if path == "" {
-		return nil
-	}
-	b, err := json.MarshalIndent(prefs, "", "\t")
-	if err != nil {
-		return fmt.Errorf("save config: %v", err)
-	}
-	if err := atomicfile.WriteFile(path, b, 0666); err != nil {
-		return fmt.Errorf("save config: %v", err)
-	}
-	return nil
-}

cmd/tailscaled/tailscaled.go

@@ -23,6 +23,15 @@ import (
 	"tailscale.com/wgengine/magicsock"
 )
 
+// globalStateKey is the ipn.StateKey that tailscaled loads on
+// startup.
+//
+// We have to support multiple state keys for other OSes (Windows in
+// particular), but right now Unix daemons run with a single
+// node-global state. To keep open the option of having per-user state
+// later, the global state key doesn't look like a username.
+const globalStateKey = "_daemon"
+
 func main() {
 	fake := getopt.BoolLong("fake", 0, "fake tunnel+routing instead of tuntap")
 	debug := getopt.StringLong("debug", 0, "", "Address of debug server")
@@ -43,6 +52,10 @@ func main() {
 		log.Fatalf("too many non-flag arguments: %#v", getopt.Args()[0])
 	}
 
+	if *statepath == "" {
+		log.Fatalf("--state is required")
+	}
+
 	if *debug != "" {
 		go runDebugServer(*debug)
 	}
@@ -60,6 +73,7 @@ func main() {
 
 	opts := ipnserver.Options{
 		StatePath:          *statepath,
+		AutostartStateKey:  globalStateKey,
 		SurviveDisconnects: true,
 	}
 	err = ipnserver.Run(context.Background(), logf, pol.PublicID.String(), opts, e)

ipn/ipnserver/server.go

@@ -25,11 +25,30 @@ import (
 	"tailscale.com/logtail/backoff"
 	"tailscale.com/safesocket"
 	"tailscale.com/types/logger"
+	"tailscale.com/version"
 	"tailscale.com/wgengine"
 )
 
+// defaultLoginServer is the login URL used by an auto-starting
+// server.
+//
+// TODO(danderson): the reason this is hardcoded is that the server
+// URL is currently not stored in state, but passed in by the
+// frontend. This needs to be fixed.
+const defaultLoginServer = "https://login.tailscale.com"
+
+// Options is the configuration of the Tailscale node agent.
 type Options struct {
+	// StatePath is the path to the stored agent state.
 	StatePath string
+	// AutostartStateKey, if non-empty, immediately starts the agent
+	// using the given StateKey. If empty, the agent stays idle and
+	// waits for a frontend to start it.
+	AutostartStateKey ipn.StateKey
+	// SurviveDisconnects specifies how the server reacts to its
+	// frontend disconnecting. If true, the server keeps running on
+	// its existing state, and accepts new frontend connections. If
+	// false, the server dumps its state and becomes idle.
 	SurviveDisconnects bool
 }
 
@@ -57,6 +76,12 @@ func Run(rctx context.Context, logf logger.Logf, logid string, opts Options, e w
 	if err != nil {
 		return fmt.Errorf("safesocket.Listen: %v", err)
 	}
+	// Go listeners can't take a context, close it instead.
+	go func() {
+		<-rctx.Done()
+		listen.Close()
+	}()
+	logf("Listening on %v\n", listen.Addr())
 
 	var store ipn.StateStore
 	if opts.StatePath != "" {
@@ -86,13 +111,17 @@ func Run(rctx context.Context, logf logger.Logf, logid string, opts Options, e w
 
 	bs := ipn.NewBackendServer(logf, b, serverToClient)
 
-	logf("Listening on %v\n", listen.Addr())
+	if opts.AutostartStateKey != "" {
-
+		bs.GotCommand(&ipn.Command{
-	// Go listeners can't take a context, close it instead.
+			Version: version.LONG,
-	go func() {
+			Start: &ipn.StartArgs{
-		<-rctx.Done()
+				Opts: ipn.Options{
-		listen.Close()
+					ServerURL: defaultLoginServer,
-	}()
+					StateKey:  opts.AutostartStateKey,
+				},
+			},
+		})
+	}
 
 	var oldS net.Conn
 	//lint:ignore SA4006 ctx is never used, but has to be defined so

ipn/message.go

@@ -31,9 +31,12 @@ type FakeExpireAfterArgs struct {
 	Duration time.Duration
 }
 
-// A command message sent to the server. Exactly one of these must be non-nil.
+// Command is a command message that is JSON encoded and sent by a
+// frontend to a backend.
 type Command struct {
 	Version string
+
+	// Exactly one of the following must be non-nil.
 	Quit                  *NoArgs
 	Start                 *StartArgs
 	StartLoginInteractive *NoArgs