cmd/tailscaled: move more of the Windows server setup code into tailscaled

Updates #1232

cmd/tailscaled/depaware.txt

@@ -2,8 +2,6 @@ tailscale.com/cmd/tailscaled dependencies: (generated by github.com/tailscale/de
 
    W 💣 github.com/alexbrainman/sspi                                 from github.com/alexbrainman/sspi/negotiate
    W 💣 github.com/alexbrainman/sspi/negotiate                       from tailscale.com/net/tshttpproxy
-        github.com/apenwarr/fixconsole                               from tailscale.com/cmd/tailscaled
-   W 💣 github.com/apenwarr/w32                                      from github.com/apenwarr/fixconsole
    L    github.com/coreos/go-iptables/iptables                       from tailscale.com/wgengine/router
   LW    github.com/go-multierror/multierror                          from tailscale.com/wgengine/router
    W 💣 github.com/go-ole/go-ole                                     from github.com/go-ole/go-ole/oleutil+
@@ -163,7 +161,7 @@ tailscale.com/cmd/tailscaled dependencies: (generated by github.com/tailscale/de
         golang.org/x/sync/singleflight                               from tailscale.com/net/dnscache
         golang.org/x/sys/cpu                                         from golang.org/x/crypto/blake2b+
   LD    golang.org/x/sys/unix                                        from github.com/jsimonetti/rtnetlink/internal/unix+
-   W    golang.org/x/sys/windows                                     from github.com/apenwarr/fixconsole+
+   W    golang.org/x/sys/windows                                     from github.com/tailscale/wireguard-go/conn+
    W    golang.org/x/sys/windows/registry                            from golang.zx2c4.com/wireguard/windows/tunnel/winipcfg+
    W    golang.org/x/sys/windows/svc                                 from tailscale.com/cmd/tailscaled
         golang.org/x/term                                            from tailscale.com/logpolicy

cmd/tailscaled/tailscaled.go

@@ -24,7 +24,6 @@ import (
 	"syscall"
 	"time"
 
-	"github.com/apenwarr/fixconsole"
 	"tailscale.com/ipn/ipnserver"
 	"tailscale.com/logpolicy"
 	"tailscale.com/paths"
@@ -88,11 +87,6 @@ func main() {
 	flag.StringVar(&args.socketpath, "socket", paths.DefaultTailscaledSocket(), "path of the service unix socket")
 	flag.BoolVar(&printVersion, "version", false, "print version information and exit")
 
-	err := fixconsole.FixConsoleIfNeeded()
-	if err != nil {
-		log.Fatalf("fixConsoleOutput: %v", err)
-	}
-
 	if len(os.Args) > 1 && os.Args[1] == "debug" {
 		if err := debugMode(os.Args[2:]); err != nil {
 			log.Fatal(err)
@@ -100,6 +94,10 @@ func main() {
 		return
 	}
 
+	if beWindowsSubprocess() {
+		return
+	}
+
 	flag.Parse()
 	if flag.NArg() > 0 {
 		log.Fatalf("tailscaled does not take non-flag arguments: %q", flag.Args())

cmd/tailscaled/tailscaled_notwindows.go

@@ -11,3 +11,5 @@ import "tailscale.com/logpolicy"
 func isWindowsService() bool { return false }
 
 func runWindowsService(pol *logpolicy.Policy) error { panic("unreachable") }
+
+func beWindowsSubprocess() bool { return false }

cmd/tailscaled/tailscaled_windows.go

@@ -4,14 +4,33 @@
 
 package main // import "tailscale.com/cmd/tailscaled"
 
+// TODO: check if administrator, like tswin does.
+//
+// TODO: try to load wintun.dll early at startup, before wireguard/tun
+//       does (which panics) and if we'd fail (e.g. due to access
+//       denied, even if administrator), use 'tasklist /m wintun.dll'
+//       to see if something else is currently using it and tell user.
+//
+// TODO: check if Tailscale service is already running, and fail early
+//       like tswin does.
+//
+// TODO: on failure, check if on a UNC drive and recommend copying it
+//       to C:\ to run it, like tswin does.
+
 import (
 	"context"
+	"fmt"
 	"log"
+	"os"
+	"time"
 
 	"golang.org/x/sys/windows"
 	"golang.org/x/sys/windows/svc"
 	"tailscale.com/ipn/ipnserver"
 	"tailscale.com/logpolicy"
+	"tailscale.com/types/logger"
+	"tailscale.com/version"
+	"tailscale.com/wgengine"
 )
 
 const serviceName = "Tailscale IPN"
@@ -62,3 +81,100 @@ func (service *ipnService) Execute(args []string, r <-chan svc.ChangeRequest, ch
 	changes <- svc.Status{State: svc.StopPending}
 	return false, windows.NO_ERROR
 }
+
+func beWindowsSubprocess() bool {
+	if len(os.Args) != 3 || os.Args[1] != "/subproc" {
+		return false
+	}
+	logid := os.Args[2]
+
+	log.Printf("Program starting: v%v: %#v", version.Long, os.Args)
+	log.Printf("subproc mode: logid=%v", logid)
+
+	go func() {
+		b := make([]byte, 16)
+		for {
+			_, err := os.Stdin.Read(b)
+			if err != nil {
+				log.Fatalf("stdin err (parent process died): %v", err)
+			}
+		}
+	}()
+
+	err := startIPNServer(context.Background(), logid)
+	if err != nil {
+		log.Fatalf("ipnserver: %v", err)
+	}
+	return true
+}
+
+func startIPNServer(ctx context.Context, logid string) error {
+	var logf logger.Logf = log.Printf
+	var eng wgengine.Engine
+	var err error
+
+	getEngine := func() (wgengine.Engine, error) {
+		eng, err := wgengine.NewUserspaceEngine(logf, "Tailscale", 41641)
+		if err != nil {
+			return nil, err
+		}
+		return wgengine.NewWatchdog(eng), nil
+	}
+
+	if msg := os.Getenv("TS_DEBUG_WIN_FAIL"); msg != "" {
+		err = fmt.Errorf("pretending to be a service failure: %v", msg)
+	} else {
+		// We have a bunch of bug reports of wgengine.NewUserspaceEngine returning a few different errors,
+		// all intermittently. A few times I (Brad) have also seen sporadic failures that simply
+		// restarting fixed. So try a few times.
+		for try := 1; try <= 5; try++ {
+			if try > 1 {
+				// Only sleep a bit. Don't do some massive backoff because
+				// the frontend GUI has a 30 second timeout on connecting to us,
+				// but even 5 seconds is too long for them to get any results.
+				// 5 tries * 1 second each seems fine.
+				time.Sleep(time.Second)
+			}
+			eng, err = getEngine()
+			if err != nil {
+				logf("wgengine.NewUserspaceEngine: (try %v) %v", try, err)
+				continue
+			}
+			if try > 1 {
+				logf("wgengine.NewUserspaceEngine: ended up working on try %v", try)
+			}
+			break
+		}
+	}
+	if err != nil {
+		// Log the error, but don't fatalf. We want to
+		// propagate the error message to the UI frontend. So
+		// we continue and tell the ipnserver to return that
+		// in a Notify message.
+		logf("wgengine.NewUserspaceEngine: %v", err)
+	}
+	opts := ipnserver.Options{
+		Port:               41112,
+		SurviveDisconnects: false,
+		StatePath:          args.statepath,
+	}
+	if err != nil {
+		// Return nicer errors to users, annotated with logids, which helps
+		// when they file bugs.
+		rawGetEngine := getEngine // raw == without verbose logid-containing error
+		getEngine = func() (wgengine.Engine, error) {
+			eng, err := rawGetEngine()
+			if err != nil {
+				return nil, fmt.Errorf("wgengine.NewUserspaceEngine: %v\n\nlogid: %v", err, logid)
+			}
+			return eng, nil
+		}
+	} else {
+		getEngine = ipnserver.FixedEngine(eng)
+	}
+	err = ipnserver.Run(ctx, logf, logid, getEngine, opts)
+	if err != nil {
+		logf("ipnserver.Run: %v", err)
+	}
+	return err
+}

paths/paths.go

@@ -8,6 +8,7 @@ package paths
 
 import (
 	"os"
+	"path/filepath"
 	"runtime"
 )
 
@@ -42,5 +43,8 @@ func DefaultTailscaledStateFile() string {
 	if f := stateFileFunc; f != nil {
 		return f()
 	}
+	if runtime.GOOS == "windows" {
+		return filepath.Join(os.Getenv("LocalAppData"), "Tailscale", "server-state.conf")
+	}
 	return ""
 }