ipn/ipnlocal: put a retry loop around Windows file deletes

oh, Windows.

Updates tailscale/corp#1626

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

.github/workflows/linux-race.yml

@@ -0,0 +1,48 @@
+name: Linux race
+
+on:
+  push:
+    branches:
+      - main
+  pull_request:
+    branches:
+      - '*'
+
+jobs:
+  build:
+    runs-on: ubuntu-latest
+
+    if: "!contains(github.event.head_commit.message, '[ci skip]')"
+
+    steps:
+
+    - name: Set up Go
+      uses: actions/setup-go@v1
+      with:
+        go-version: 1.16
+      id: go
+
+    - name: Check out code into the Go module directory
+      uses: actions/checkout@v1
+
+    - name: Basic build
+      run: go build ./cmd/...
+
+    - name: Run tests with -race flag on linux
+      run: go test -race ./...
+
+    - uses: k0kubun/action-slack@v2.0.0
+      with:
+        payload: |
+          {
+            "attachments": [{
+              "text": "${{ job.status }}: ${{ github.workflow }} <https://github.com/${{ github.repository }}/commit/${{ github.sha }}/checks|${{ env.COMMIT_DATE }} #${{ env.COMMIT_NUMBER_OF_DAY }}> " +
+                      "(<https://github.com/${{ github.repository }}/commit/${{ github.sha }}|" + "${{ github.sha }}".substring(0, 10) + ">) " +
+                      "of ${{ github.repository }}@" + "${{ github.ref }}".split('/').reverse()[0] + " by ${{ github.event.head_commit.committer.name }}",
+              "color": "danger"
+            }]
+          }
+      env:
+        SLACK_WEBHOOK_URL: ${{ secrets.SLACK_WEBHOOK_URL }}
+      if: failure() && github.event_name == 'push'
+

.github/workflows/windows-race.yml

@@ -0,0 +1,52 @@
+name: Windows race
+
+on:
+  push:
+    branches:
+      - main
+  pull_request:
+    branches:
+      - '*'
+
+jobs:
+  test:
+    runs-on: windows-latest
+
+    if: "!contains(github.event.head_commit.message, '[ci skip]')"
+
+    steps:
+
+    - name: Install Go
+      uses: actions/setup-go@v2
+      with:
+        go-version: 1.16.x
+
+    - name: Checkout code
+      uses: actions/checkout@v2
+
+    - name: Restore Cache
+      uses: actions/cache@v2
+      with:
+        path: ~/go/pkg/mod
+        key: ${{ runner.os }}-go-${{ hashFiles('**/go.sum') }}
+        restore-keys: |
+          ${{ runner.os }}-go-
+
+    - name: Test with -race flag
+      run: go test -race ./...
+
+    - uses: k0kubun/action-slack@v2.0.0
+      with:
+        payload: |
+          {
+            "attachments": [{
+              "text": "${{ job.status }}: ${{ github.workflow }} <https://github.com/${{ github.repository }}/commit/${{ github.sha }}/checks|${{ env.COMMIT_DATE }} #${{ env.COMMIT_NUMBER_OF_DAY }}> " +
+                      "(<https://github.com/${{ github.repository }}/commit/${{ github.sha }}|" + "${{ github.sha }}".substring(0, 10) + ">) " +
+                      "of ${{ github.repository }}@" + "${{ github.ref }}".split('/').reverse()[0] + " by ${{ github.event.head_commit.committer.name }}",
+              "color": "danger"
+            }]
+          }
+      env:
+        SLACK_WEBHOOK_URL: ${{ secrets.SLACK_WEBHOOK_URL }}
+      if: failure() && github.event_name == 'push'
+

api.md

@@ -367,10 +367,11 @@ Etag: "e0b2816b418b3f266309d94426ac7668ab3c1fa87798785bf82f1085cc2f6d9c"
 
 #### `POST /api/v2/tailnet/:tailnet/acl` - set ACL for a tailnet
 
-Sets the ACL for the given tailnet. HuJSON and JSON are both accepted inputs. An `If-Match` header can be set to avoid missed updates.
+Sets the ACL for the given domain.
+HuJSON and JSON are both accepted inputs.
+An `If-Match` header can be set to avoid missed updates.
 
-Returns error for invalid ACLs.
-Returns error if using an `If-Match` header and the ETag does not match.
+Returns the updated ACL in JSON or HuJSON according to the `Accept` header on success. Otherwise, errors are returned for incorrectly defined ACLs, ACLs with failing tests on attempted updates, and mismatched `If-Match` header and ETag.
 
 ##### Parameters
 
@@ -380,7 +381,17 @@ Returns error if using an `If-Match` header and the ETag does not match.
 `Accept` - Sets the return type of the updated ACL. Response is parsed `JSON` if `application/json` is explicitly named, otherwise HuJSON will be returned.
 
 ###### POST Body
-ACL JSON or HuJSON (see https://tailscale.com/kb/1018/acls)
+
+The POST body should be a JSON or [HuJSON](https://github.com/tailscale/hujson#hujson---human-json) formatted JSON object.
+An ACL policy may contain the following top-level properties:
+
+* `Groups` - Static groups of users which can be used for ACL rules.
+* `Hosts` - Hostname aliases to use in place of IP addresses or subnets.
+* `ACLs` - Access control lists.
+* `TagOwners` - Defines who is allowed to use which tags.
+* `Tests` - Run on ACL updates to check correct functionality of defined ACLs.
+
+See https://tailscale.com/kb/1018/acls for more information on those properties.
 
 ##### Example
 ```
@@ -411,7 +422,7 @@ curl 'https://api.tailscale.com/api/v2/tailnet/example.com/acl' \
 }'
 ```
 
-Response
+Response:
 ```
 // Example/default ACLs for unrestricted connections.
 {
@@ -436,9 +447,25 @@ Response
 }
 ```
 
+Failed test error response:
+```
+{
+    "message": "test(s) failed",
+    "data": [
+        {
+            "user": "user1@example.com",
+            "errors": [
+                "address \"user2@example.com:400\": want: Accept, got: Drop"
+            ]
+        }
+    ]
+}
+```
+
 <a name=tailnet-acl-preview-post></a>
 
 #### `POST /api/v2/tailnet/:tailnet/acl/preview` - preview rule matches on an ACL for a resource
+
 Determines what rules match for a user on an ACL without saving the ACL to the server.
 
 ##### Parameters
@@ -477,7 +504,7 @@ curl 'https://api.tailscale.com/api/v2/tailnet/example.com/acl?user=user1@exampl
 }'
 ```
 
-Response
+Response:
 ```
 {"matches":[{"users":["*"],"ports":["*:*"],"lineNumber":19}],"user":"user1@example.com"}
 ```

client/tailscale/apitype/apitype.go

@@ -0,0 +1,29 @@
+// Copyright (c) 2021 Tailscale Inc & AUTHORS All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// Package apitype contains types for the Tailscale local API.
+package apitype
+
+import "tailscale.com/tailcfg"
+
+// WhoIsResponse is the JSON type returned by tailscaled debug server's /whois?ip=$IP handler.
+type WhoIsResponse struct {
+	Node        *tailcfg.Node
+	UserProfile *tailcfg.UserProfile
+}
+
+// FileTarget is a node to which files can be sent, and the PeerAPI
+// URL base to do so via.
+type FileTarget struct {
+	Node *tailcfg.Node
+
+	// PeerAPI is the http://ip:port URL base of the node's peer API,
+	// without any path (not even a single slash).
+	PeerAPIURL string
+}
+
+type WaitingFile struct {
+	Name string
+	Size int64
+}

client/tailscale/tailscale.go

@@ -6,20 +6,29 @@
 package tailscale
 
 import (
+	"bytes"
 	"context"
 	"encoding/json"
+	"errors"
 	"fmt"
+	"io"
 	"io/ioutil"
 	"net"
 	"net/http"
 	"net/url"
 	"strconv"
+	"strings"
 
+	"tailscale.com/client/tailscale/apitype"
+	"tailscale.com/ipn"
 	"tailscale.com/ipn/ipnstate"
+	"tailscale.com/paths"
 	"tailscale.com/safesocket"
-	"tailscale.com/tailcfg"
 )
 
+// TailscaledSocket is the tailscaled Unix socket.
+var TailscaledSocket = paths.DefaultTailscaledSocket()
+
 // tsClient does HTTP requests to the local Tailscale daemon.
 var tsClient = &http.Client{
 	Transport: &http.Transport{
@@ -27,14 +36,16 @@ var tsClient = &http.Client{
 			if addr != "local-tailscaled.sock:80" {
 				return nil, fmt.Errorf("unexpected URL address %q", addr)
 			}
-			// On macOS, when dialing from non-sandboxed program to sandboxed GUI running
-			// a TCP server on a random port, find the random port. For HTTP connections,
-			// we don't send the token. It gets added in an HTTP Basic-Auth header.
-			if port, _, err := safesocket.LocalTCPPortAndToken(); err == nil {
-				var d net.Dialer
-				return d.DialContext(ctx, "tcp", "localhost:"+strconv.Itoa(port))
+			if TailscaledSocket == paths.DefaultTailscaledSocket() {
+				// On macOS, when dialing from non-sandboxed program to sandboxed GUI running
+				// a TCP server on a random port, find the random port. For HTTP connections,
+				// we don't send the token. It gets added in an HTTP Basic-Auth header.
+				if port, _, err := safesocket.LocalTCPPortAndToken(); err == nil {
+					var d net.Dialer
+					return d.DialContext(ctx, "tcp", "localhost:"+strconv.Itoa(port))
+				}
 			}
-			return safesocket.ConnectDefault()
+			return safesocket.Connect(TailscaledSocket, 41112)
 		},
 	},
 }
@@ -55,9 +66,22 @@ func DoLocalRequest(req *http.Request) (*http.Response, error) {
 	return tsClient.Do(req)
 }
 
-// WhoIs returns the owner of the remoteAddr, which must be an IP or IP:port.
-func WhoIs(ctx context.Context, remoteAddr string) (*tailcfg.WhoIsResponse, error) {
-	req, err := http.NewRequestWithContext(ctx, "GET", "http://local-tailscaled.sock/localapi/v0/whois?addr="+url.QueryEscape(remoteAddr), nil)
+type errorJSON struct {
+	Error string
+}
+
+// bestError returns either err, or if body contains a valid JSON
+// object of type errorJSON, its non-empty error body.
+func bestError(err error, body []byte) error {
+	var j errorJSON
+	if err := json.Unmarshal(body, &j); err == nil && j.Error != "" {
+		return errors.New(j.Error)
+	}
+	return err
+}
+
+func send(ctx context.Context, method, path string, wantStatus int, body io.Reader) ([]byte, error) {
+	req, err := http.NewRequestWithContext(ctx, method, "http://local-tailscaled.sock"+path, body)
 	if err != nil {
 		return nil, err
 	}
@@ -66,39 +90,49 @@ func WhoIs(ctx context.Context, remoteAddr string) (*tailcfg.WhoIsResponse, erro
 		return nil, err
 	}
 	defer res.Body.Close()
-	slurp, _ := ioutil.ReadAll(res.Body)
-	if res.StatusCode != 200 {
-		return nil, fmt.Errorf("HTTP %s: %s", res.Status, slurp)
+	slurp, err := ioutil.ReadAll(res.Body)
+	if err != nil {
+		return nil, err
 	}
-	r := new(tailcfg.WhoIsResponse)
-	if err := json.Unmarshal(slurp, r); err != nil {
-		if max := 200; len(slurp) > max {
-			slurp = slurp[:max]
+	if res.StatusCode != wantStatus {
+		err := fmt.Errorf("HTTP %s: %s (expected %v)", res.Status, slurp, wantStatus)
+		return nil, bestError(err, slurp)
+	}
+	return slurp, nil
+}
+
+func get200(ctx context.Context, path string) ([]byte, error) {
+	return send(ctx, "GET", path, 200, nil)
+}
+
+// WhoIs returns the owner of the remoteAddr, which must be an IP or IP:port.
+func WhoIs(ctx context.Context, remoteAddr string) (*apitype.WhoIsResponse, error) {
+	body, err := get200(ctx, "/localapi/v0/whois?addr="+url.QueryEscape(remoteAddr))
+	if err != nil {
+		return nil, err
+	}
+	r := new(apitype.WhoIsResponse)
+	if err := json.Unmarshal(body, r); err != nil {
+		if max := 200; len(body) > max {
+			body = append(body[:max], "..."...)
 		}
-		return nil, fmt.Errorf("failed to parse JSON WhoIsResponse from %q", slurp)
+		return nil, fmt.Errorf("failed to parse JSON WhoIsResponse from %q", body)
 	}
 	return r, nil
 }
 
 // Goroutines returns a dump of the Tailscale daemon's current goroutines.
 func Goroutines(ctx context.Context) ([]byte, error) {
-	req, err := http.NewRequestWithContext(ctx, "GET", "http://local-tailscaled.sock/localapi/v0/goroutines", nil)
+	return get200(ctx, "/localapi/v0/goroutines")
+}
+
+// BugReport logs and returns a log marker that can be shared by the user with support.
+func BugReport(ctx context.Context, note string) (string, error) {
+	body, err := send(ctx, "POST", "/localapi/v0/bugreport?note="+url.QueryEscape(note), 200, nil)
 	if err != nil {
-		return nil, err
+		return "", err
 	}
-	res, err := DoLocalRequest(req)
-	if err != nil {
-		return nil, err
-	}
-	defer res.Body.Close()
-	body, err := ioutil.ReadAll(res.Body)
-	if err != nil {
-		return nil, err
-	}
-	if res.StatusCode != 200 {
-		return nil, fmt.Errorf("HTTP %s: %s", res.Status, body)
-	}
-	return body, nil
+	return strings.TrimSpace(string(body)), nil
 }
 
 // Status returns the Tailscale daemon's status.
@@ -112,22 +146,113 @@ func StatusWithoutPeers(ctx context.Context) (*ipnstate.Status, error) {
 }
 
 func status(ctx context.Context, queryString string) (*ipnstate.Status, error) {
-	req, err := http.NewRequestWithContext(ctx, "GET", "http://local-tailscaled.sock/localapi/v0/status"+queryString, nil)
+	body, err := get200(ctx, "/localapi/v0/status"+queryString)
 	if err != nil {
 		return nil, err
 	}
-	res, err := DoLocalRequest(req)
-	if err != nil {
-		return nil, err
-	}
-	defer res.Body.Close()
-	if res.StatusCode != 200 {
-		body, _ := ioutil.ReadAll(res.Body)
-		return nil, fmt.Errorf("HTTP %s: %s", res.Status, body)
-	}
 	st := new(ipnstate.Status)
-	if err := json.NewDecoder(res.Body).Decode(st); err != nil {
+	if err := json.Unmarshal(body, st); err != nil {
 		return nil, err
 	}
 	return st, nil
 }
+
+func WaitingFiles(ctx context.Context) ([]apitype.WaitingFile, error) {
+	body, err := get200(ctx, "/localapi/v0/files/")
+	if err != nil {
+		return nil, err
+	}
+	var wfs []apitype.WaitingFile
+	if err := json.Unmarshal(body, &wfs); err != nil {
+		return nil, err
+	}
+	return wfs, nil
+}
+
+func DeleteWaitingFile(ctx context.Context, baseName string) error {
+	_, err := send(ctx, "DELETE", "/localapi/v0/files/"+url.PathEscape(baseName), http.StatusNoContent, nil)
+	return err
+}
+
+func GetWaitingFile(ctx context.Context, baseName string) (rc io.ReadCloser, size int64, err error) {
+	req, err := http.NewRequestWithContext(ctx, "GET", "http://local-tailscaled.sock/localapi/v0/files/"+url.PathEscape(baseName), nil)
+	if err != nil {
+		return nil, 0, err
+	}
+	res, err := DoLocalRequest(req)
+	if err != nil {
+		return nil, 0, err
+	}
+	if res.ContentLength == -1 {
+		res.Body.Close()
+		return nil, 0, fmt.Errorf("unexpected chunking")
+	}
+	if res.StatusCode != 200 {
+		body, _ := ioutil.ReadAll(res.Body)
+		res.Body.Close()
+		return nil, 0, fmt.Errorf("HTTP %s: %s", res.Status, body)
+	}
+	return res.Body, res.ContentLength, nil
+}
+
+func FileTargets(ctx context.Context) ([]apitype.FileTarget, error) {
+	body, err := get200(ctx, "/localapi/v0/file-targets")
+	if err != nil {
+		return nil, err
+	}
+	var fts []apitype.FileTarget
+	if err := json.Unmarshal(body, &fts); err != nil {
+		return nil, fmt.Errorf("invalid JSON: %w", err)
+	}
+	return fts, nil
+}
+
+func CheckIPForwarding(ctx context.Context) error {
+	body, err := get200(ctx, "/localapi/v0/check-ip-forwarding")
+	if err != nil {
+		return err
+	}
+	var jres struct {
+		Warning string
+	}
+	if err := json.Unmarshal(body, &jres); err != nil {
+		return fmt.Errorf("invalid JSON from check-ip-forwarding: %w", err)
+	}
+	if jres.Warning != "" {
+		return errors.New(jres.Warning)
+	}
+	return nil
+}
+
+func GetPrefs(ctx context.Context) (*ipn.Prefs, error) {
+	body, err := get200(ctx, "/localapi/v0/prefs")
+	if err != nil {
+		return nil, err
+	}
+	var p ipn.Prefs
+	if err := json.Unmarshal(body, &p); err != nil {
+		return nil, fmt.Errorf("invalid prefs JSON: %w", err)
+	}
+	return &p, nil
+}
+
+func EditPrefs(ctx context.Context, mp *ipn.MaskedPrefs) (*ipn.Prefs, error) {
+	mpj, err := json.Marshal(mp)
+	if err != nil {
+		return nil, err
+	}
+	body, err := send(ctx, "PATCH", "/localapi/v0/prefs", http.StatusOK, bytes.NewReader(mpj))
+	if err != nil {
+		return nil, err
+	}
+	var p ipn.Prefs
+	if err := json.Unmarshal(body, &p); err != nil {
+		return nil, fmt.Errorf("invalid prefs JSON: %w", err)
+	}
+	return &p, nil
+}
+
+func Logout(ctx context.Context) error {
+	_, err := send(ctx, "POST", "/localapi/v0/logout", http.StatusNoContent, nil)
+	return err
+}

cmd/hello/hello.go

@@ -18,7 +18,7 @@ import (
 	"strings"
 
 	"tailscale.com/client/tailscale"
-	"tailscale.com/tailcfg"
+	"tailscale.com/client/tailscale/apitype"
 )
 
 var (
@@ -107,7 +107,7 @@ type tmplData struct {
 	IP            string // "100.2.3.4"
 }
 
-func tailscaleIP(who *tailcfg.WhoIsResponse) string {
+func tailscaleIP(who *apitype.WhoIsResponse) string {
 	if who == nil {
 		return ""
 	}

cmd/tailscale/cli/bugreport.go

@@ -0,0 +1,38 @@
+// Copyright (c) 2021 Tailscale Inc & AUTHORS All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package cli
+
+import (
+	"context"
+	"errors"
+	"fmt"
+
+	"github.com/peterbourgon/ff/v2/ffcli"
+	"tailscale.com/client/tailscale"
+)
+
+var bugReportCmd = &ffcli.Command{
+	Name:       "bugreport",
+	Exec:       runBugReport,
+	ShortHelp:  "Print a shareable identifier to help diagnose issues",
+	ShortUsage: "bugreport [note]",
+}
+
+func runBugReport(ctx context.Context, args []string) error {
+	var note string
+	switch len(args) {
+	case 0:
+	case 1:
+		note = args[0]
+	default:
+		return errors.New("unknown argumets")
+	}
+	logMarker, err := tailscale.BugReport(ctx, note)
+	if err != nil {
+		return err
+	}
+	fmt.Println(logMarker)
+	return nil
+}

cmd/tailscale/cli/cli.go

@@ -15,29 +15,54 @@ import (
 	"os"
 	"os/signal"
 	"runtime"
+	"strconv"
 	"strings"
 	"syscall"
 	"text/tabwriter"
 
 	"github.com/peterbourgon/ff/v2/ffcli"
+	"tailscale.com/client/tailscale"
 	"tailscale.com/ipn"
 	"tailscale.com/paths"
 	"tailscale.com/safesocket"
+	"tailscale.com/syncs"
 )
 
 // ActLikeCLI reports whether a GUI application should act like the
 // CLI based on os.Args, GOOS, the context the process is running in
 // (pty, parent PID), etc.
 func ActLikeCLI() bool {
-	if len(os.Args) < 2 {
+	// This function is only used on macOS.
+	if runtime.GOOS != "darwin" {
 		return false
 	}
-	switch os.Args[1] {
-	case "up", "down", "status", "netcheck", "ping", "version",
-		"debug",
-		"-V", "--version", "-h", "--help":
+
+	// Escape hatch to let people force running the macOS
+	// GUI Tailscale binary as the CLI.
+	if v, _ := strconv.ParseBool(os.Getenv("TAILSCALE_BE_CLI")); v {
 		return true
 	}
+
+	// If our parent is launchd, we're definitely not
+	// being run as a CLI.
+	if os.Getppid() == 1 {
+		return false
+	}
+
+	// Looking at the environment of the GUI Tailscale app (ps eww
+	// $PID), empirically none of these environment variables are
+	// present. But all or some of these should be present with
+	// Terminal.all and bash or zsh.
+	for _, e := range []string{
+		"SHLVL",
+		"TERM",
+		"TERM_PROGRAM",
+		"PS1",
+	} {
+		if os.Getenv(e) != "" {
+			return true
+		}
+	}
 	return false
 }
 
@@ -63,6 +88,7 @@ change in the future.
 		Subcommands: []*ffcli.Command{
 			upCmd,
 			downCmd,
+			logoutCmd,
 			netcheckCmd,
 			ipCmd,
 			statusCmd,
@@ -70,6 +96,7 @@ change in the future.
 			versionCmd,
 			webCmd,
 			pushCmd,
+			bugReportCmd,
 		},
 		FlagSet:   rootfs,
 		Exec:      func(context.Context, []string) error { return flag.ErrHelp },
@@ -88,6 +115,8 @@ change in the future.
 		return err
 	}
 
+	tailscale.TailscaledSocket = rootArgs.socket
+
 	err := rootCmd.Run(context.Background())
 	if err == flag.ErrHelp {
 		return nil
@@ -104,6 +133,8 @@ var rootArgs struct {
 	socket string
 }
 
+var gotSignal syncs.AtomicBool
+
 func connect(ctx context.Context) (net.Conn, *ipn.BackendClient, context.Context, context.CancelFunc) {
 	c, err := safesocket.Connect(rootArgs.socket, 41112)
 	if err != nil {
@@ -121,7 +152,14 @@ func connect(ctx context.Context) (net.Conn, *ipn.BackendClient, context.Context
 	go func() {
 		interrupt := make(chan os.Signal, 1)
 		signal.Notify(interrupt, syscall.SIGINT, syscall.SIGTERM)
-		<-interrupt
+		select {
+		case <-interrupt:
+		case <-ctx.Done():
+			// Context canceled elsewhere.
+			signal.Reset(syscall.SIGINT, syscall.SIGTERM)
+			return
+		}
+		gotSignal.Set(true)
 		c.Close()
 		cancel()
 	}()
@@ -139,7 +177,9 @@ func pump(ctx context.Context, bc *ipn.BackendClient, conn net.Conn) {
 			if ctx.Err() != nil {
 				return
 			}
-			log.Printf("ReadMsg: %v\n", err)
+			if !gotSignal.Get() {
+				log.Printf("ReadMsg: %v\n", err)
+			}
 			break
 		}
 		bc.GotNotifyMsg(msg)

cmd/tailscale/cli/cli_test.go

@@ -0,0 +1,460 @@
+// Copyright (c) 2021 Tailscale Inc & AUTHORS All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package cli
+
+import (
+	"bytes"
+	"encoding/json"
+	"flag"
+	"fmt"
+	"strings"
+	"testing"
+
+	"inet.af/netaddr"
+	"tailscale.com/ipn"
+	"tailscale.com/ipn/ipnstate"
+	"tailscale.com/types/logger"
+	"tailscale.com/types/preftype"
+)
+
+// Test that checkForAccidentalSettingReverts's updateMaskedPrefsFromUpFlag can handle
+// all flags. This will panic if a new flag creeps in that's unhandled.
+func TestUpdateMaskedPrefsFromUpFlag(t *testing.T) {
+	mp := new(ipn.MaskedPrefs)
+	upFlagSet.VisitAll(func(f *flag.Flag) {
+		updateMaskedPrefsFromUpFlag(mp, f.Name)
+	})
+}
+
+func TestCheckForAccidentalSettingReverts(t *testing.T) {
+	f := func(flags ...string) map[string]bool {
+		m := make(map[string]bool)
+		for _, f := range flags {
+			m[f] = true
+		}
+		return m
+	}
+	tests := []struct {
+		name     string
+		flagSet  map[string]bool
+		curPrefs *ipn.Prefs
+		curUser  string // os.Getenv("USER") on the client side
+		mp       *ipn.MaskedPrefs
+		want     string
+	}{
+		{
+			name:    "bare_up_means_up",
+			flagSet: f(),
+			curPrefs: &ipn.Prefs{
+				ControlURL:  ipn.DefaultControlURL,
+				WantRunning: false,
+				Hostname:    "foo",
+			},
+			mp: &ipn.MaskedPrefs{
+				Prefs: ipn.Prefs{
+					WantRunning: true,
+				},
+				WantRunningSet: true,
+			},
+			want: "",
+		},
+		{
+			name:    "losing_hostname",
+			flagSet: f("accept-dns"),
+			curPrefs: &ipn.Prefs{
+				ControlURL:  ipn.DefaultControlURL,
+				WantRunning: false,
+				Hostname:    "foo",
+				CorpDNS:     true,
+			},
+			mp: &ipn.MaskedPrefs{
+				Prefs: ipn.Prefs{
+					ControlURL:  ipn.DefaultControlURL,
+					WantRunning: true,
+					CorpDNS:     true,
+				},
+				ControlURLSet:  true,
+				WantRunningSet: true,
+				CorpDNSSet:     true,
+			},
+			want: `'tailscale up' without --reset requires all preferences with changing values to be explicitly mentioned; --hostname is not specified but its default value of "" differs from current value "foo"`,
+		},
+		{
+			name:    "hostname_changing_explicitly",
+			flagSet: f("hostname"),
+			curPrefs: &ipn.Prefs{
+				ControlURL:  ipn.DefaultControlURL,
+				WantRunning: false,
+				Hostname:    "foo",
+			},
+			mp: &ipn.MaskedPrefs{
+				Prefs: ipn.Prefs{
+					ControlURL:  ipn.DefaultControlURL,
+					WantRunning: true,
+					Hostname:    "bar",
+				},
+				ControlURLSet:  true,
+				WantRunningSet: true,
+				HostnameSet:    true,
+			},
+			want: "",
+		},
+		{
+			name:    "hostname_changing_empty_explicitly",
+			flagSet: f("hostname"),
+			curPrefs: &ipn.Prefs{
+				ControlURL:  ipn.DefaultControlURL,
+				WantRunning: false,
+				Hostname:    "foo",
+			},
+			mp: &ipn.MaskedPrefs{
+				Prefs: ipn.Prefs{
+					ControlURL:  ipn.DefaultControlURL,
+					WantRunning: true,
+					Hostname:    "",
+				},
+				ControlURLSet:  true,
+				WantRunningSet: true,
+				HostnameSet:    true,
+			},
+			want: "",
+		},
+		{
+			name:    "empty_slice_equals_nil_slice",
+			flagSet: f("hostname"),
+			curPrefs: &ipn.Prefs{
+				ControlURL:      ipn.DefaultControlURL,
+				AdvertiseRoutes: []netaddr.IPPrefix{},
+			},
+			mp: &ipn.MaskedPrefs{
+				Prefs: ipn.Prefs{
+					ControlURL:      ipn.DefaultControlURL,
+					AdvertiseRoutes: nil,
+				},
+				ControlURLSet: true,
+			},
+			want: "",
+		},
+		{
+			// Issue 1725: "tailscale up --authkey=..." (or other non-empty flags) works from
+			// a fresh server's initial prefs.
+			name:     "up_with_default_prefs",
+			flagSet:  f("authkey"),
+			curPrefs: ipn.NewPrefs(),
+			mp: &ipn.MaskedPrefs{
+				Prefs:          *defaultPrefsFromUpArgs(t),
+				WantRunningSet: true,
+			},
+			want: "",
+		},
+		{
+			name:    "implicit_operator_change",
+			flagSet: f("hostname"),
+			curPrefs: &ipn.Prefs{
+				ControlURL:   ipn.DefaultControlURL,
+				OperatorUser: "alice",
+			},
+			curUser: "eve",
+			mp: &ipn.MaskedPrefs{
+				Prefs: ipn.Prefs{
+					ControlURL: ipn.DefaultControlURL,
+				},
+				ControlURLSet: true,
+			},
+			want: `'tailscale up' without --reset requires all preferences with changing values to be explicitly mentioned; --operator is not specified but its default value of "" differs from current value "alice"`,
+		},
+		{
+			name:    "implicit_operator_matches_shell_user",
+			flagSet: f("hostname"),
+			curPrefs: &ipn.Prefs{
+				ControlURL:   ipn.DefaultControlURL,
+				OperatorUser: "alice",
+			},
+			curUser: "alice",
+			mp: &ipn.MaskedPrefs{
+				Prefs: ipn.Prefs{
+					ControlURL: ipn.DefaultControlURL,
+				},
+				ControlURLSet: true,
+			},
+			want: "",
+		},
+		{
+			name:    "error_advertised_routes_exit_node_removed",
+			flagSet: f("advertise-routes"),
+			curPrefs: &ipn.Prefs{
+				ControlURL: ipn.DefaultControlURL,
+				AdvertiseRoutes: []netaddr.IPPrefix{
+					netaddr.MustParseIPPrefix("10.0.42.0/24"),
+					netaddr.MustParseIPPrefix("0.0.0.0/0"),
+					netaddr.MustParseIPPrefix("::/0"),
+				},
+			},
+			mp: &ipn.MaskedPrefs{
+				Prefs: ipn.Prefs{
+					ControlURL: ipn.DefaultControlURL,
+					AdvertiseRoutes: []netaddr.IPPrefix{
+						netaddr.MustParseIPPrefix("10.0.42.0/24"),
+					},
+				},
+				AdvertiseRoutesSet: true,
+			},
+			want: "'tailscale up' without --reset requires all preferences with changing values to be explicitly mentioned; --advertise-exit-node flag not mentioned but currently advertised routes are an exit node",
+		},
+		{
+			name:    "advertised_routes_exit_node_removed",
+			flagSet: f("advertise-routes", "advertise-exit-node"),
+			curPrefs: &ipn.Prefs{
+				ControlURL: ipn.DefaultControlURL,
+				AdvertiseRoutes: []netaddr.IPPrefix{
+					netaddr.MustParseIPPrefix("10.0.42.0/24"),
+					netaddr.MustParseIPPrefix("0.0.0.0/0"),
+					netaddr.MustParseIPPrefix("::/0"),
+				},
+			},
+			mp: &ipn.MaskedPrefs{
+				Prefs: ipn.Prefs{
+					ControlURL: ipn.DefaultControlURL,
+					AdvertiseRoutes: []netaddr.IPPrefix{
+						netaddr.MustParseIPPrefix("10.0.42.0/24"),
+					},
+				},
+				AdvertiseRoutesSet: true,
+			},
+			want: "",
+		},
+		{
+			name:    "advertised_routes_includes_the_0_routes", // but no --advertise-exit-node
+			flagSet: f("advertise-routes"),
+			curPrefs: &ipn.Prefs{
+				ControlURL: ipn.DefaultControlURL,
+				AdvertiseRoutes: []netaddr.IPPrefix{
+					netaddr.MustParseIPPrefix("10.0.42.0/24"),
+					netaddr.MustParseIPPrefix("0.0.0.0/0"),
+					netaddr.MustParseIPPrefix("::/0"),
+				},
+			},
+			mp: &ipn.MaskedPrefs{
+				Prefs: ipn.Prefs{
+					ControlURL: ipn.DefaultControlURL,
+					AdvertiseRoutes: []netaddr.IPPrefix{
+						netaddr.MustParseIPPrefix("11.1.43.0/24"),
+						netaddr.MustParseIPPrefix("0.0.0.0/0"),
+						netaddr.MustParseIPPrefix("::/0"),
+					},
+				},
+				AdvertiseRoutesSet: true,
+			},
+			want: "",
+		},
+		{
+			name:    "advertised_routes_includes_only_one_0_route", // and no --advertise-exit-node
+			flagSet: f("advertise-routes"),
+			curPrefs: &ipn.Prefs{
+				ControlURL: ipn.DefaultControlURL,
+				AdvertiseRoutes: []netaddr.IPPrefix{
+					netaddr.MustParseIPPrefix("10.0.42.0/24"),
+					netaddr.MustParseIPPrefix("0.0.0.0/0"),
+					netaddr.MustParseIPPrefix("::/0"),
+				},
+			},
+			mp: &ipn.MaskedPrefs{
+				Prefs: ipn.Prefs{
+					ControlURL: ipn.DefaultControlURL,
+					AdvertiseRoutes: []netaddr.IPPrefix{
+						netaddr.MustParseIPPrefix("11.1.43.0/24"),
+						netaddr.MustParseIPPrefix("0.0.0.0/0"),
+					},
+				},
+				AdvertiseRoutesSet: true,
+			},
+			want: "'tailscale up' without --reset requires all preferences with changing values to be explicitly mentioned; --advertise-exit-node flag not mentioned but currently advertised routes are an exit node",
+		},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			var got string
+			if err := checkForAccidentalSettingReverts(tt.flagSet, tt.curPrefs, tt.mp, tt.curUser); err != nil {
+				got = err.Error()
+			}
+			if got != tt.want {
+				t.Errorf("unexpected result\n got: %s\nwant: %s\n", got, tt.want)
+			}
+		})
+	}
+}
+
+func defaultPrefsFromUpArgs(t testing.TB) *ipn.Prefs {
+	upFlagSet.Parse(nil) // populates upArgs
+	if upFlagSet.Lookup("netfilter-mode") == nil && upArgs.netfilterMode == "" {
+		// This flag is not compiled on on-Linux platforms,
+		// but prefsFromUpArgs requires it be populated.
+		upArgs.netfilterMode = defaultNetfilterMode()
+	}
+	prefs, err := prefsFromUpArgs(upArgs, logger.Discard, new(ipnstate.Status), "linux")
+	if err != nil {
+		t.Fatalf("defaultPrefsFromUpArgs: %v", err)
+	}
+	prefs.WantRunning = true
+	return prefs
+}
+
+func TestPrefsFromUpArgs(t *testing.T) {
+	tests := []struct {
+		name     string
+		args     upArgsT
+		goos     string           // runtime.GOOS; empty means linux
+		st       *ipnstate.Status // or nil
+		want     *ipn.Prefs
+		wantErr  string
+		wantWarn string
+	}{
+		{
+			name: "zero",
+			goos: "windows",
+			args: upArgsT{},
+			want: &ipn.Prefs{
+				WantRunning:   true,
+				NoSNAT:        true,
+				NetfilterMode: preftype.NetfilterOn, // silly, but default from ipn.NewPref currently
+			},
+		},
+		{
+			name: "error_advertise_route_invalid_ip",
+			args: upArgsT{
+				advertiseRoutes: "foo",
+			},
+			wantErr: `"foo" is not a valid IP address or CIDR prefix`,
+		},
+		{
+			name: "error_advertise_route_unmasked_bits",
+			args: upArgsT{
+				advertiseRoutes: "1.2.3.4/16",
+			},
+			wantErr: `1.2.3.4/16 has non-address bits set; expected 1.2.0.0/16`,
+		},
+		{
+			name: "error_exit_node_bad_ip",
+			args: upArgsT{
+				exitNodeIP: "foo",
+			},
+			wantErr: `invalid IP address "foo" for --exit-node: unable to parse IP`,
+		},
+		{
+			name: "error_exit_node_allow_lan_without_exit_node",
+			args: upArgsT{
+				exitNodeAllowLANAccess: true,
+			},
+			wantErr: `--exit-node-allow-lan-access can only be used with --exit-node`,
+		},
+		{
+			name: "error_tag_prefix",
+			args: upArgsT{
+				advertiseTags: "foo",
+			},
+			wantErr: `tag: "foo": tags must start with 'tag:'`,
+		},
+		{
+			name: "error_long_hostname",
+			args: upArgsT{
+				hostname: strings.Repeat("a", 300),
+			},
+			wantErr: `hostname too long: 300 bytes (max 256)`,
+		},
+		{
+			name: "error_linux_netfilter_empty",
+			args: upArgsT{
+				netfilterMode: "",
+			},
+			wantErr: `invalid value --netfilter-mode=""`,
+		},
+		{
+			name: "error_linux_netfilter_bogus",
+			args: upArgsT{
+				netfilterMode: "bogus",
+			},
+			wantErr: `invalid value --netfilter-mode="bogus"`,
+		},
+		{
+			name: "error_exit_node_ip_is_self_ip",
+			args: upArgsT{
+				exitNodeIP: "100.105.106.107",
+			},
+			st: &ipnstate.Status{
+				TailscaleIPs: []netaddr.IP{netaddr.MustParseIP("100.105.106.107")},
+			},
+			wantErr: `cannot use 100.105.106.107 as the exit node as it is a local IP address to this machine, did you mean --advertise-exit-node?`,
+		},
+		{
+			name: "warn_linux_netfilter_nodivert",
+			goos: "linux",
+			args: upArgsT{
+				netfilterMode: "nodivert",
+			},
+			wantWarn: "netfilter=nodivert; add iptables calls to ts-* chains manually.",
+			want: &ipn.Prefs{
+				WantRunning:   true,
+				NetfilterMode: preftype.NetfilterNoDivert,
+				NoSNAT:        true,
+			},
+		},
+		{
+			name: "warn_linux_netfilter_off",
+			goos: "linux",
+			args: upArgsT{
+				netfilterMode: "off",
+			},
+			wantWarn: "netfilter=off; configure iptables yourself.",
+			want: &ipn.Prefs{
+				WantRunning:   true,
+				NetfilterMode: preftype.NetfilterOff,
+				NoSNAT:        true,
+			},
+		},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			var warnBuf bytes.Buffer
+			warnf := func(format string, a ...interface{}) {
+				fmt.Fprintf(&warnBuf, format, a...)
+			}
+			goos := tt.goos
+			if goos == "" {
+				goos = "linux"
+			}
+			st := tt.st
+			if st == nil {
+				st = new(ipnstate.Status)
+			}
+			got, err := prefsFromUpArgs(tt.args, warnf, st, goos)
+			gotErr := fmt.Sprint(err)
+			if tt.wantErr != "" {
+				if tt.wantErr != gotErr {
+					t.Errorf("wrong error.\n got error: %v\nwant error: %v\n", gotErr, tt.wantErr)
+				}
+				return
+			}
+			if err != nil {
+				t.Fatal(err)
+			}
+			if tt.want == nil {
+				t.Fatal("tt.want is nil")
+			}
+			if !got.Equals(tt.want) {
+				jgot, _ := json.MarshalIndent(got, "", "\t")
+				jwant, _ := json.MarshalIndent(tt.want, "", "\t")
+				if bytes.Equal(jgot, jwant) {
+					t.Logf("prefs differ only in non-JSON-visible ways (nil/non-nil zero-length arrays)")
+				}
+				t.Errorf("wrong prefs\n got: %s\nwant: %s\n\ngot: %s\nwant: %s\n",
+					got.Pretty(), tt.want.Pretty(),
+					jgot, jwant,
+				)
+
+			}
+		})
+	}
+
+}

cmd/tailscale/cli/debug.go

@@ -6,12 +6,21 @@ package cli
 
 import (
 	"context"
+	"encoding/json"
 	"errors"
 	"flag"
+	"fmt"
+	"io"
+	"log"
 	"os"
+	"runtime"
+	"strings"
 
 	"github.com/peterbourgon/ff/v2/ffcli"
 	"tailscale.com/client/tailscale"
+	"tailscale.com/ipn"
+	"tailscale.com/paths"
+	"tailscale.com/safesocket"
 )
 
 var debugCmd = &ffcli.Command{
@@ -20,24 +29,101 @@ var debugCmd = &ffcli.Command{
 	FlagSet: (func() *flag.FlagSet {
 		fs := flag.NewFlagSet("debug", flag.ExitOnError)
 		fs.BoolVar(&debugArgs.goroutines, "daemon-goroutines", false, "If true, dump the tailscaled daemon's goroutines")
+		fs.BoolVar(&debugArgs.ipn, "ipn", false, "If true, subscribe to IPN notifications")
+		fs.BoolVar(&debugArgs.prefs, "prefs", false, "If true, dump active prefs")
+		fs.BoolVar(&debugArgs.pretty, "pretty", false, "If true, pretty-print output (for --prefs)")
+		fs.BoolVar(&debugArgs.netMap, "netmap", true, "whether to include netmap in --ipn mode")
+		fs.BoolVar(&debugArgs.localCreds, "local-creds", false, "print how to connect to local tailscaled")
+		fs.StringVar(&debugArgs.file, "file", "", "get, delete:NAME, or NAME")
 		return fs
 	})(),
 }
 
 var debugArgs struct {
+	localCreds bool
 	goroutines bool
+	ipn        bool
+	netMap     bool
+	file       string
+	prefs      bool
+	pretty     bool
 }
 
 func runDebug(ctx context.Context, args []string) error {
 	if len(args) > 0 {
 		return errors.New("unknown arguments")
 	}
+	if debugArgs.localCreds {
+		port, token, err := safesocket.LocalTCPPortAndToken()
+		if err == nil {
+			fmt.Printf("curl -u:%s http://localhost:%d/localapi/v0/status\n", token, port)
+			return nil
+		}
+		if runtime.GOOS == "windows" {
+			fmt.Printf("curl http://localhost:41112/localapi/v0/status\n")
+			return nil
+		}
+		fmt.Printf("curl --unix-socket %s http://foo/localapi/v0/status\n", paths.DefaultTailscaledSocket())
+		return nil
+	}
+	if debugArgs.prefs {
+		prefs, err := tailscale.GetPrefs(ctx)
+		if err != nil {
+			return err
+		}
+		if debugArgs.pretty {
+			fmt.Println(prefs.Pretty())
+		} else {
+			j, _ := json.MarshalIndent(prefs, "", "\t")
+			fmt.Println(string(j))
+		}
+		return nil
+	}
 	if debugArgs.goroutines {
 		goroutines, err := tailscale.Goroutines(ctx)
 		if err != nil {
 			return err
 		}
 		os.Stdout.Write(goroutines)
+		return nil
+	}
+	if debugArgs.ipn {
+		c, bc, ctx, cancel := connect(ctx)
+		defer cancel()
+
+		bc.SetNotifyCallback(func(n ipn.Notify) {
+			if !debugArgs.netMap {
+				n.NetMap = nil
+			}
+			j, _ := json.MarshalIndent(n, "", "\t")
+			fmt.Printf("%s\n", j)
+		})
+		bc.RequestEngineStatus()
+		pump(ctx, bc, c)
+		return errors.New("exit")
+	}
+	if debugArgs.file != "" {
+		if debugArgs.file == "get" {
+			wfs, err := tailscale.WaitingFiles(ctx)
+			if err != nil {
+				log.Fatal(err)
+			}
+			e := json.NewEncoder(os.Stdout)
+			e.SetIndent("", "\t")
+			e.Encode(wfs)
+			return nil
+		}
+		delete := strings.HasPrefix(debugArgs.file, "delete:")
+		if delete {
+			return tailscale.DeleteWaitingFile(ctx, strings.TrimPrefix(debugArgs.file, "delete:"))
+		}
+		rc, size, err := tailscale.GetWaitingFile(ctx, debugArgs.file)
+		if err != nil {
+			return err
+		}
+		log.Printf("Size: %v\n", size)
+		io.Copy(os.Stdout, rc)
+		return nil
 	}
 	return nil
 }

cmd/tailscale/cli/down.go

@@ -8,7 +8,7 @@ import (
 	"context"
 	"fmt"
 	"log"
-	"time"
+	"os"
 
 	"github.com/peterbourgon/ff/v2/ffcli"
 	"tailscale.com/client/tailscale"
@@ -33,34 +33,14 @@ func runDown(ctx context.Context, args []string) error {
 		return fmt.Errorf("error fetching current status: %w", err)
 	}
 	if st.BackendState == "Stopped" {
-		log.Printf("already stopped")
+		fmt.Fprintf(os.Stderr, "Tailscale was already stopped.\n")
 		return nil
 	}
-	log.Printf("was in state %q", st.BackendState)
-
-	c, bc, ctx, cancel := connect(ctx)
-	defer cancel()
-
-	timer := time.AfterFunc(5*time.Second, func() {
-		log.Fatalf("timeout running stop")
+	_, err = tailscale.EditPrefs(ctx, &ipn.MaskedPrefs{
+		Prefs: ipn.Prefs{
+			WantRunning: false,
+		},
+		WantRunningSet: true,
 	})
-	defer timer.Stop()
-
-	bc.SetNotifyCallback(func(n ipn.Notify) {
-		if n.ErrMessage != nil {
-			log.Fatal(*n.ErrMessage)
-		}
-		if n.State != nil {
-			log.Printf("now in state %q", *n.State)
-			if *n.State == ipn.Stopped {
-				cancel()
-			}
-			return
-		}
-	})
-
-	bc.SetWantRunning(false)
-	pump(ctx, bc, c)
-
-	return nil
+	return err
 }

cmd/tailscale/cli/ip.go

@@ -11,13 +11,16 @@ import (
 	"fmt"
 
 	"github.com/peterbourgon/ff/v2/ffcli"
+	"inet.af/netaddr"
 	"tailscale.com/client/tailscale"
+	"tailscale.com/ipn/ipnstate"
 )
 
 var ipCmd = &ffcli.Command{
 	Name:       "ip",
-	ShortUsage: "ip [-4] [-6]",
-	ShortHelp:  "Show this machine's current Tailscale IP address(es)",
+	ShortUsage: "ip [-4] [-6] [peername]",
+	ShortHelp:  "Show current Tailscale IP address(es)",
+	LongHelp:   "Shows the Tailscale IP address of the current machine without an argument. With an argument, it shows the IP of a named peer.",
 	Exec:       runIP,
 	FlagSet: (func() *flag.FlagSet {
 		fs := flag.NewFlagSet("ip", flag.ExitOnError)
@@ -33,9 +36,14 @@ var ipArgs struct {
 }
 
 func runIP(ctx context.Context, args []string) error {
-	if len(args) > 0 {
+	if len(args) > 1 {
 		return errors.New("unknown arguments")
 	}
+	var of string
+	if len(args) == 1 {
+		of = args[0]
+	}
+
 	v4, v6 := ipArgs.want4, ipArgs.want6
 	if v4 && v6 {
 		return errors.New("tailscale up -4 and -6 are mutually exclusive")
@@ -47,11 +55,24 @@ func runIP(ctx context.Context, args []string) error {
 	if err != nil {
 		return err
 	}
-	if len(st.TailscaleIPs) == 0 {
+	ips := st.TailscaleIPs
+	if of != "" {
+		ip, err := tailscaleIPFromArg(ctx, of)
+		if err != nil {
+			return err
+		}
+		peer, ok := peerMatchingIP(st, ip)
+		if !ok {
+			return fmt.Errorf("no peer found with IP %v", ip)
+		}
+		ips = peer.TailscaleIPs
+	}
+	if len(ips) == 0 {
 		return fmt.Errorf("no current Tailscale IPs; state: %v", st.BackendState)
 	}
+
 	match := false
-	for _, ip := range st.TailscaleIPs {
+	for _, ip := range ips {
 		if ip.Is4() && v4 || ip.Is6() && v6 {
 			match = true
 			fmt.Println(ip)
@@ -67,3 +88,18 @@ func runIP(ctx context.Context, args []string) error {
 	}
 	return nil
 }
+
+func peerMatchingIP(st *ipnstate.Status, ipStr string) (ps *ipnstate.PeerStatus, ok bool) {
+	ip, err := netaddr.ParseIP(ipStr)
+	if err != nil {
+		return
+	}
+	for _, ps = range st.Peer {
+		for _, pip := range ps.TailscaleIPs {
+			if ip == pip {
+				return ps, true
+			}
+		}
+	}
+	return nil, false
+}

cmd/tailscale/cli/logout.go

@@ -0,0 +1,34 @@
+// Copyright (c) 2020 Tailscale Inc & AUTHORS All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package cli
+
+import (
+	"context"
+	"log"
+	"strings"
+
+	"github.com/peterbourgon/ff/v2/ffcli"
+	"tailscale.com/client/tailscale"
+)
+
+var logoutCmd = &ffcli.Command{
+	Name:       "logout",
+	ShortUsage: "logout [flags]",
+	ShortHelp:  "Disconnect from Tailscale and expire current node key",
+
+	LongHelp: strings.TrimSpace(`
+"tailscale logout" brings the network down and invalidates
+the current node key, forcing a future use of it to cause
+a reauthentication.
+`),
+	Exec: runLogout,
+}
+
+func runLogout(ctx context.Context, args []string) error {
+	if len(args) > 0 {
+		log.Fatalf("too many non-flag arguments: %q", args)
+	}
+	return tailscale.Logout(ctx)
+}

cmd/tailscale/cli/ping.go

@@ -154,7 +154,10 @@ func tailscaleIPFromArg(ctx context.Context, hostOrIP string) (ip string, err er
 	}
 	for _, ps := range st.Peer {
 		if hostOrIP == dnsOrQuoteHostname(st, ps) || hostOrIP == ps.DNSName {
-			return ps.TailAddr, nil
+			if len(ps.TailscaleIPs) == 0 {
+				return "", errors.New("node found but lacks an IP")
+			}
+			return ps.TailscaleIPs[0].String(), nil
 		}
 	}
 

cmd/tailscale/cli/push.go

@@ -13,16 +13,17 @@ import (
 	"io"
 	"log"
 	"mime"
-	"net"
 	"net/http"
 	"net/url"
 	"os"
+	"strconv"
 	"time"
 	"unicode/utf8"
 
 	"github.com/peterbourgon/ff/v2/ffcli"
-	"tailscale.com/ipn"
-	"tailscale.com/ipn/ipnstate"
+	"golang.org/x/time/rate"
+	"inet.af/netaddr"
+	"tailscale.com/client/tailscale"
 )
 
 var pushCmd = &ffcli.Command{
@@ -34,6 +35,7 @@ var pushCmd = &ffcli.Command{
 		fs := flag.NewFlagSet("push", flag.ExitOnError)
 		fs.StringVar(&pushArgs.name, "name", "", "alternate filename to use, especially useful when <file> is \"-\" (stdin)")
 		fs.BoolVar(&pushArgs.verbose, "verbose", false, "verbose output")
+		fs.BoolVar(&pushArgs.targets, "targets", false, "list possible push targets")
 		return fs
 	})(),
 }
@@ -41,11 +43,15 @@ var pushCmd = &ffcli.Command{
 var pushArgs struct {
 	name    string
 	verbose bool
+	targets bool
 }
 
 func runPush(ctx context.Context, args []string) error {
+	if pushArgs.targets {
+		return runPushTargets(ctx, args)
+	}
 	if len(args) != 2 || args[0] == "" {
-		return errors.New("usage: push <hostname-or-IP> <file>")
+		return errors.New("usage: push <hostname-or-IP> <file>\n       push --targets")
 	}
 	var ip string
 
@@ -55,13 +61,19 @@ func runPush(ctx context.Context, args []string) error {
 		return err
 	}
 
-	peerAPIPort, err := discoverPeerAPIPort(ctx, ip)
+	peerAPIBase, lastSeen, isOffline, err := discoverPeerAPIBase(ctx, ip)
 	if err != nil {
 		return err
 	}
+	if isOffline {
+		fmt.Fprintf(os.Stderr, "# warning: %s is offline\n", hostOrIP)
+	} else if !lastSeen.IsZero() && time.Since(lastSeen) > lastSeenOld {
+		fmt.Fprintf(os.Stderr, "# warning: %s last seen %v ago\n", hostOrIP, time.Since(lastSeen).Round(time.Minute))
+	}
 
 	var fileContents io.Reader
 	var name = pushArgs.name
+	var contentLength int64 = -1
 	if fileArg == "-" {
 		fileContents = os.Stdin
 		if name == "" {
@@ -76,17 +88,30 @@ func runPush(ctx context.Context, args []string) error {
 			return err
 		}
 		defer f.Close()
-		fileContents = f
+		fi, err := f.Stat()
+		if err != nil {
+			return err
+		}
+		if fi.IsDir() {
+			return errors.New("directories not supported")
+		}
+		contentLength = fi.Size()
+		fileContents = io.LimitReader(f, contentLength)
 		if name == "" {
 			name = fileArg
 		}
+
+		if slow, _ := strconv.ParseBool(os.Getenv("TS_DEBUG_SLOW_PUSH")); slow {
+			fileContents = &slowReader{r: fileContents}
+		}
 	}
 
-	dstURL := "http://" + net.JoinHostPort(ip, fmt.Sprint(peerAPIPort)) + "/v0/put/" + url.PathEscape(name)
+	dstURL := peerAPIBase + "/v0/put/" + url.PathEscape(name)
 	req, err := http.NewRequestWithContext(ctx, "PUT", dstURL, fileContents)
 	if err != nil {
 		return err
 	}
+	req.ContentLength = contentLength
 	if pushArgs.verbose {
 		log.Printf("sending to %v ...", dstURL)
 	}
@@ -102,51 +127,29 @@ func runPush(ctx context.Context, args []string) error {
 	return errors.New(res.Status)
 }
 
-func discoverPeerAPIPort(ctx context.Context, ip string) (port uint16, err error) {
-	c, bc, ctx, cancel := connect(ctx)
-	defer cancel()
-
-	prc := make(chan *ipnstate.PingResult, 2)
-	bc.SetNotifyCallback(func(n ipn.Notify) {
-		if n.ErrMessage != nil {
-			log.Fatal(*n.ErrMessage)
-		}
-		if pr := n.PingResult; pr != nil && pr.IP == ip {
-			prc <- pr
-		}
-	})
-	go pump(ctx, bc, c)
-
-	ticker := time.NewTicker(time.Second)
-	defer ticker.Stop()
-
-	discoPings := 0
-	timer := time.NewTimer(10 * time.Second)
-	defer timer.Stop()
-
-	sendPings := func() {
-		bc.Ping(ip, false)
-		bc.Ping(ip, true)
+func discoverPeerAPIBase(ctx context.Context, ipStr string) (base string, lastSeen time.Time, isOffline bool, err error) {
+	ip, err := netaddr.ParseIP(ipStr)
+	if err != nil {
+		return "", time.Time{}, false, err
 	}
-	sendPings()
-	for {
-		select {
-		case <-ticker.C:
-			sendPings()
-		case <-timer.C:
-			return 0, fmt.Errorf("timeout contacting %v; it offline?", ip)
-		case pr := <-prc:
-			if p := pr.PeerAPIPort; p != 0 {
-				return p, nil
+	fts, err := tailscale.FileTargets(ctx)
+	if err != nil {
+		return "", time.Time{}, false, err
+	}
+	for _, ft := range fts {
+		n := ft.Node
+		for _, a := range n.Addresses {
+			if a.IP != ip {
+				continue
 			}
-			discoPings++
-			if discoPings == 3 {
-				return 0, fmt.Errorf("%v is online, but seems to be running an old Tailscale version", ip)
+			if n.LastSeen != nil {
+				lastSeen = *n.LastSeen
 			}
-		case <-ctx.Done():
-			return 0, ctx.Err()
+			isOffline = n.Online != nil && !*n.Online
+			return ft.PeerAPIURL, lastSeen, isOffline, nil
 		}
 	}
+	return "", time.Time{}, false, errors.New("target seems to be running an old Tailscale version")
 }
 
 const maxSniff = 4 << 20
@@ -171,3 +174,54 @@ func pickStdinFilename() (name string, r io.Reader, err error) {
 	}
 	return "stdin" + ext(sniff), io.MultiReader(bytes.NewReader(sniff), os.Stdin), nil
 }
+
+type slowReader struct {
+	r  io.Reader
+	rl *rate.Limiter
+}
+
+func (r *slowReader) Read(p []byte) (n int, err error) {
+	const burst = 4 << 10
+	plen := len(p)
+	if plen > burst {
+		plen = burst
+	}
+	if r.rl == nil {
+		r.rl = rate.NewLimiter(rate.Limit(1<<10), burst)
+	}
+	n, err = r.r.Read(p[:plen])
+	r.rl.WaitN(context.Background(), n)
+	return
+}
+
+const lastSeenOld = 20 * time.Minute
+
+func runPushTargets(ctx context.Context, args []string) error {
+	if len(args) > 0 {
+		return errors.New("invalid arguments with --targets")
+	}
+	fts, err := tailscale.FileTargets(ctx)
+	if err != nil {
+		return err
+	}
+	for _, ft := range fts {
+		n := ft.Node
+		var detail string
+		if n.Online != nil {
+			if !*n.Online {
+				detail = "offline"
+			}
+		} else {
+			detail = "unknown-status"
+		}
+		if detail != "" && n.LastSeen != nil {
+			d := time.Since(*n.LastSeen)
+			detail += fmt.Sprintf("; last seen %v ago", d.Round(time.Minute))
+		}
+		if detail != "" {
+			detail = "\t" + detail
+		}
+		fmt.Printf("%s\t%s%s\n", n.Addresses[0].IP, n.ComputedName, detail)
+	}
+	return nil
+}

cmd/tailscale/cli/status.go

@@ -18,6 +18,7 @@ import (
 
 	"github.com/peterbourgon/ff/v2/ffcli"
 	"github.com/toqueteos/webbrowser"
+	"inet.af/netaddr"
 	"tailscale.com/client/tailscale"
 	"tailscale.com/ipn"
 	"tailscale.com/ipn/ipnstate"
@@ -106,9 +107,24 @@ func runStatus(ctx context.Context, args []string) error {
 		return err
 	}
 
-	if st.BackendState == ipn.Stopped.String() {
+	switch st.BackendState {
+	default:
+		fmt.Fprintf(os.Stderr, "unexpected state: %s\n", st.BackendState)
+		os.Exit(1)
+	case ipn.Stopped.String():
 		fmt.Println("Tailscale is stopped.")
 		os.Exit(1)
+	case ipn.NeedsLogin.String():
+		fmt.Println("Logged out.")
+		if st.AuthURL != "" {
+			fmt.Printf("\nLog in at: %s\n", st.AuthURL)
+		}
+		os.Exit(1)
+	case ipn.NeedsMachineAuth.String():
+		fmt.Println("Machine is not yet authorized by tailnet admin.")
+		os.Exit(1)
+	case ipn.Running.String():
+		// Run below.
 	}
 
 	var buf bytes.Buffer
@@ -116,7 +132,7 @@ func runStatus(ctx context.Context, args []string) error {
 	printPS := func(ps *ipnstate.PeerStatus) {
 		active := peerActive(ps)
 		f("%-15s %-20s %-12s %-7s ",
-			ps.TailAddr,
+			firstIPString(ps.TailscaleIPs),
 			dnsOrQuoteHostname(st, ps),
 			ownerLogin(st, ps),
 			ps.OS,
@@ -201,3 +217,10 @@ func ownerLogin(st *ipnstate.Status, ps *ipnstate.PeerStatus) string {
 	}
 	return u.LoginName
 }
+
+func firstIPString(v []netaddr.IP) string {
+	if len(v) == 0 {
+		return ""
+	}
+	return v[0].String()
+}

cmd/tailscale/cli/up.go

@@ -5,25 +5,27 @@
 package cli
 
 import (
-	"bytes"
 	"context"
 	"errors"
 	"flag"
 	"fmt"
-	"log"
 	"os"
-	"os/exec"
+	"reflect"
 	"runtime"
 	"sort"
 	"strconv"
 	"strings"
 	"sync"
 
+	"github.com/go-multierror/multierror"
 	"github.com/peterbourgon/ff/v2/ffcli"
 	"inet.af/netaddr"
 	"tailscale.com/client/tailscale"
 	"tailscale.com/ipn"
+	"tailscale.com/ipn/ipnstate"
+	"tailscale.com/safesocket"
 	"tailscale.com/tailcfg"
+	"tailscale.com/types/logger"
 	"tailscale.com/types/preftype"
 	"tailscale.com/version/distro"
 )
@@ -31,40 +33,56 @@ import (
 var upCmd = &ffcli.Command{
 	Name:       "up",
 	ShortUsage: "up [flags]",
-	ShortHelp:  "Connect to your Tailscale network",
+	ShortHelp:  "Connect to Tailscale, logging in if needed",
 
 	LongHelp: strings.TrimSpace(`
 "tailscale up" connects this machine to your Tailscale network,
 triggering authentication if necessary.
 
-The flags passed to this command are specific to this machine. If you don't
-specify any flags, options are reset to their default.
+With no flags, "tailscale up" brings the network online without
+changing any settings. (That is, it's the opposite of "tailscale
+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.
 `),
-	FlagSet: (func() *flag.FlagSet {
-		upf := flag.NewFlagSet("up", flag.ExitOnError)
-		upf.StringVar(&upArgs.server, "login-server", "https://login.tailscale.com", "base URL of control server")
-		upf.BoolVar(&upArgs.acceptRoutes, "accept-routes", false, "accept routes advertised by other Tailscale nodes")
-		upf.BoolVar(&upArgs.acceptDNS, "accept-dns", true, "accept DNS configuration from the admin panel")
-		upf.BoolVar(&upArgs.singleRoutes, "host-routes", true, "install host routes to other Tailscale nodes")
-		upf.StringVar(&upArgs.exitNodeIP, "exit-node", "", "Tailscale IP of the exit node for internet traffic")
-		upf.BoolVar(&upArgs.shieldsUp, "shields-up", false, "don't allow incoming connections")
-		upf.BoolVar(&upArgs.forceReauth, "force-reauth", false, "force reauthentication")
-		upf.StringVar(&upArgs.advertiseTags, "advertise-tags", "", "ACL tags to request (comma-separated, e.g. \"tag:eng,tag:montreal,tag:ssh\")")
-		upf.StringVar(&upArgs.authKey, "authkey", "", "node authorization key")
-		upf.StringVar(&upArgs.hostname, "hostname", "", "hostname to use instead of the one provided by the OS")
-		if runtime.GOOS == "linux" || isBSD(runtime.GOOS) {
-			upf.StringVar(&upArgs.advertiseRoutes, "advertise-routes", "", "routes to advertise to other nodes (comma-separated, e.g. \"10.0.0.0/8,192.168.0.0/24\")")
-			upf.BoolVar(&upArgs.advertiseDefaultRoute, "advertise-exit-node", false, "offer to be an exit node for internet traffic for the tailnet")
-		}
-		if runtime.GOOS == "linux" {
-			upf.BoolVar(&upArgs.snat, "snat-subnet-routes", true, "source NAT traffic to local routes advertised with --advertise-routes")
-			upf.StringVar(&upArgs.netfilterMode, "netfilter-mode", defaultNetfilterMode(), "netfilter mode (one of on, nodivert, off)")
-		}
-		return upf
-	})(),
-	Exec: runUp,
+	FlagSet: upFlagSet,
+	Exec:    runUp,
 }
 
+var upFlagSet = (func() *flag.FlagSet {
+	upf := flag.NewFlagSet("up", flag.ExitOnError)
+
+	upf.BoolVar(&upArgs.forceReauth, "force-reauth", false, "force reauthentication")
+	upf.BoolVar(&upArgs.reset, "reset", false, "reset unspecified settings to their default values")
+
+	upf.StringVar(&upArgs.server, "login-server", ipn.DefaultControlURL, "base URL of control server")
+	upf.BoolVar(&upArgs.acceptRoutes, "accept-routes", false, "accept routes advertised by other Tailscale nodes")
+	upf.BoolVar(&upArgs.acceptDNS, "accept-dns", true, "accept DNS configuration from the admin panel")
+	upf.BoolVar(&upArgs.singleRoutes, "host-routes", true, "install host routes to other Tailscale nodes")
+	upf.StringVar(&upArgs.exitNodeIP, "exit-node", "", "Tailscale IP of the exit node for internet traffic")
+	upf.BoolVar(&upArgs.exitNodeAllowLANAccess, "exit-node-allow-lan-access", false, "Allow direct access to the local network when routing traffic via an exit node")
+	upf.BoolVar(&upArgs.shieldsUp, "shields-up", false, "don't allow incoming connections")
+	upf.StringVar(&upArgs.advertiseTags, "advertise-tags", "", "comma-separated ACL tags to request; each must start with \"tag:\" (e.g. \"tag:eng,tag:montreal,tag:ssh\")")
+	upf.StringVar(&upArgs.authKey, "authkey", "", "node authorization key")
+	upf.StringVar(&upArgs.hostname, "hostname", "", "hostname to use instead of the one provided by the OS")
+	upf.StringVar(&upArgs.advertiseRoutes, "advertise-routes", "", "routes to advertise to other nodes (comma-separated, e.g. \"10.0.0.0/8,192.168.0.0/24\")")
+	upf.BoolVar(&upArgs.advertiseDefaultRoute, "advertise-exit-node", false, "offer to be an exit node for internet traffic for the tailnet")
+	if safesocket.PlatformUsesPeerCreds() {
+		upf.StringVar(&upArgs.opUser, "operator", "", "Unix username to allow to operate on tailscaled without sudo")
+	}
+	if runtime.GOOS == "linux" {
+		upf.BoolVar(&upArgs.snat, "snat-subnet-routes", true, "source NAT traffic to local routes advertised with --advertise-routes")
+		upf.StringVar(&upArgs.netfilterMode, "netfilter-mode", defaultNetfilterMode(), "netfilter mode (one of on, nodivert, off)")
+	}
+	if runtime.GOOS == "windows" {
+		upf.BoolVar(&upArgs.forceDaemon, "unattended", false, "run in \"Unattended Mode\" where Tailscale keeps running even after the current GUI user logs out (Windows-only)")
+	}
+	return upf
+})()
+
 func defaultNetfilterMode() string {
 	if distro.Get() == distro.Synology {
 		return "off"
@@ -72,85 +90,45 @@ func defaultNetfilterMode() string {
 	return "on"
 }
 
-var upArgs struct {
-	server                string
-	acceptRoutes          bool
-	acceptDNS             bool
-	singleRoutes          bool
-	exitNodeIP            string
-	shieldsUp             bool
-	forceReauth           bool
-	advertiseRoutes       string
-	advertiseDefaultRoute bool
-	advertiseTags         string
-	snat                  bool
-	netfilterMode         string
-	authKey               string
-	hostname              string
+type upArgsT struct {
+	reset                  bool
+	server                 string
+	acceptRoutes           bool
+	acceptDNS              bool
+	singleRoutes           bool
+	exitNodeIP             string
+	exitNodeAllowLANAccess bool
+	shieldsUp              bool
+	forceReauth            bool
+	forceDaemon            bool
+	advertiseRoutes        string
+	advertiseDefaultRoute  bool
+	advertiseTags          string
+	snat                   bool
+	netfilterMode          string
+	authKey                string
+	hostname               string
+	opUser                 string
 }
 
-func isBSD(s string) bool {
-	return s == "dragonfly" || s == "freebsd" || s == "netbsd" || s == "openbsd"
-}
+var upArgs upArgsT
 
 func warnf(format string, args ...interface{}) {
 	fmt.Printf("Warning: "+format+"\n", args...)
 }
 
-// checkIPForwarding prints warnings on linux if IP forwarding is not
-// enabled, or if we were unable to verify the state of IP forwarding.
-func checkIPForwarding() {
-	var key string
-
-	if runtime.GOOS == "linux" {
-		key = "net.ipv4.ip_forward"
-	} else if isBSD(runtime.GOOS) {
-		key = "net.inet.ip.forwarding"
-	} else {
-		return
-	}
-
-	bs, err := exec.Command("sysctl", "-n", key).Output()
-	if err != nil {
-		warnf("couldn't check %s (%v).\nSubnet routes won't work without IP forwarding.", key, err)
-		return
-	}
-	on, err := strconv.ParseBool(string(bytes.TrimSpace(bs)))
-	if err != nil {
-		warnf("couldn't parse %s (%v).\nSubnet routes won't work without IP forwarding.", key, err)
-		return
-	}
-	if !on {
-		warnf("%s is disabled. Subnet routes won't work.", key)
-	}
-}
-
 var (
 	ipv4default = netaddr.MustParseIPPrefix("0.0.0.0/0")
 	ipv6default = netaddr.MustParseIPPrefix("::/0")
 )
 
-func runUp(ctx context.Context, args []string) error {
-	if len(args) > 0 {
-		log.Fatalf("too many non-flag arguments: %q", args)
-	}
-
-	if distro.Get() == distro.Synology {
-		notSupported := "not yet supported on Synology; see https://github.com/tailscale/tailscale/issues/451"
-		if upArgs.advertiseRoutes != "" {
-			return errors.New("--advertise-routes is " + notSupported)
-		}
-		if upArgs.acceptRoutes {
-			return errors.New("--accept-routes is " + notSupported)
-		}
-		if upArgs.exitNodeIP != "" {
-			return errors.New("--exit-node is " + notSupported)
-		}
-		if upArgs.netfilterMode != "off" {
-			return errors.New("--netfilter-mode values besides \"off\" " + notSupported)
-		}
-	}
-
+// prefsFromUpArgs returns the ipn.Prefs for the provided args.
+//
+// Note that the parameters upArgs and warnf are named intentionally
+// to shadow the globals to prevent accidental misuse of them. This
+// function exists for testing and should have no side effects or
+// outside interactions (e.g. no making Tailscale local API calls).
+func prefsFromUpArgs(upArgs upArgsT, warnf logger.Logf, st *ipnstate.Status, goos string) (*ipn.Prefs, error) {
 	routeMap := map[netaddr.IPPrefix]bool{}
 	var default4, default6 bool
 	if upArgs.advertiseRoutes != "" {
@@ -158,10 +136,10 @@ func runUp(ctx context.Context, args []string) error {
 		for _, s := range advroutes {
 			ipp, err := netaddr.ParseIPPrefix(s)
 			if err != nil {
-				fatalf("%q is not a valid IP address or CIDR prefix", s)
+				return nil, fmt.Errorf("%q is not a valid IP address or CIDR prefix", s)
 			}
 			if ipp != ipp.Masked() {
-				fatalf("%s has non-address bits set; expected %s", ipp, ipp.Masked())
+				return nil, fmt.Errorf("%s has non-address bits set; expected %s", ipp, ipp.Masked())
 			}
 			if ipp == ipv4default {
 				default4 = true
@@ -171,21 +149,15 @@ func runUp(ctx context.Context, args []string) error {
 			routeMap[ipp] = true
 		}
 		if default4 && !default6 {
-			fatalf("%s advertised without its IPv6 counterpart, please also advertise %s", ipv4default, ipv6default)
+			return nil, fmt.Errorf("%s advertised without its IPv6 counterpart, please also advertise %s", ipv4default, ipv6default)
 		} else if default6 && !default4 {
-			fatalf("%s advertised without its IPv6 counterpart, please also advertise %s", ipv6default, ipv4default)
+			return nil, fmt.Errorf("%s advertised without its IPv6 counterpart, please also advertise %s", ipv6default, ipv4default)
 		}
 	}
 	if upArgs.advertiseDefaultRoute {
 		routeMap[netaddr.MustParseIPPrefix("0.0.0.0/0")] = true
 		routeMap[netaddr.MustParseIPPrefix("::/0")] = true
 	}
-	if len(routeMap) > 0 {
-		checkIPForwarding()
-		if isBSD(runtime.GOOS) {
-			warnf("Subnet routing and exit nodes only work with additional manual configuration on %v, and is not currently officially supported.", runtime.GOOS)
-		}
-	}
 	routes := make([]netaddr.IPPrefix, 0, len(routeMap))
 	for r := range routeMap {
 		routes = append(routes, r)
@@ -202,7 +174,17 @@ func runUp(ctx context.Context, args []string) error {
 		var err error
 		exitNodeIP, err = netaddr.ParseIP(upArgs.exitNodeIP)
 		if err != nil {
-			fatalf("invalid IP address %q for --exit-node: %v", upArgs.exitNodeIP, err)
+			return nil, fmt.Errorf("invalid IP address %q for --exit-node: %v", upArgs.exitNodeIP, err)
+		}
+	} else if upArgs.exitNodeAllowLANAccess {
+		return nil, fmt.Errorf("--exit-node-allow-lan-access can only be used with --exit-node")
+	}
+
+	if upArgs.exitNodeIP != "" {
+		for _, ip := range st.TailscaleIPs {
+			if exitNodeIP == ip {
+				return nil, fmt.Errorf("cannot use %s as the exit node as it is a local IP address to this machine, did you mean --advertise-exit-node?", upArgs.exitNodeIP)
+			}
 		}
 	}
 
@@ -212,13 +194,13 @@ func runUp(ctx context.Context, args []string) error {
 		for _, tag := range tags {
 			err := tailcfg.CheckTag(tag)
 			if err != nil {
-				fatalf("tag: %q: %s", tag, err)
+				return nil, fmt.Errorf("tag: %q: %s", tag, err)
 			}
 		}
 	}
 
 	if len(upArgs.hostname) > 256 {
-		fatalf("hostname too long: %d bytes (max 256)", len(upArgs.hostname))
+		return nil, fmt.Errorf("hostname too long: %d bytes (max 256)", len(upArgs.hostname))
 	}
 
 	prefs := ipn.NewPrefs()
@@ -226,6 +208,7 @@ func runUp(ctx context.Context, args []string) error {
 	prefs.WantRunning = true
 	prefs.RouteAll = upArgs.acceptRoutes
 	prefs.ExitNodeIP = exitNodeIP
+	prefs.ExitNodeAllowLANAccess = upArgs.exitNodeAllowLANAccess
 	prefs.CorpDNS = upArgs.acceptDNS
 	prefs.AllowSingleHosts = upArgs.singleRoutes
 	prefs.ShieldsUp = upArgs.shieldsUp
@@ -233,9 +216,10 @@ func runUp(ctx context.Context, args []string) error {
 	prefs.AdvertiseTags = tags
 	prefs.NoSNAT = !upArgs.snat
 	prefs.Hostname = upArgs.hostname
-	prefs.ForceDaemon = (runtime.GOOS == "windows")
+	prefs.ForceDaemon = upArgs.forceDaemon
+	prefs.OperatorUser = upArgs.opUser
 
-	if runtime.GOOS == "linux" {
+	if goos == "linux" {
 		switch upArgs.netfilterMode {
 		case "on":
 			prefs.NetfilterMode = preftype.NetfilterOn
@@ -246,98 +230,392 @@ func runUp(ctx context.Context, args []string) error {
 			prefs.NetfilterMode = preftype.NetfilterOff
 			warnf("netfilter=off; configure iptables yourself.")
 		default:
-			fatalf("invalid value --netfilter-mode: %q", upArgs.netfilterMode)
+			return nil, fmt.Errorf("invalid value --netfilter-mode=%q", upArgs.netfilterMode)
+		}
+	}
+	return prefs, nil
+}
+
+func runUp(ctx context.Context, args []string) error {
+	if len(args) > 0 {
+		fatalf("too many non-flag arguments: %q", args)
+	}
+
+	st, err := tailscale.Status(ctx)
+	if err != nil {
+		fatalf("can't fetch status from tailscaled: %v", err)
+	}
+	origAuthURL := st.AuthURL
+
+	// printAuthURL reports whether we should print out the
+	// provided auth URL from an IPN notify.
+	printAuthURL := func(url string) bool {
+		if upArgs.authKey != "" {
+			// Issue 1755: when using an authkey, don't
+			// show an authURL that might still be pending
+			// from a previous non-completed interactive
+			// login.
+			return false
+		}
+		if upArgs.forceReauth && url == origAuthURL {
+			return false
+		}
+		return true
+	}
+
+	if distro.Get() == distro.Synology {
+		notSupported := "not yet supported on Synology; see https://github.com/tailscale/tailscale/issues/451"
+		if upArgs.acceptRoutes {
+			return errors.New("--accept-routes is " + notSupported)
+		}
+		if upArgs.exitNodeIP != "" {
+			return errors.New("--exit-node is " + notSupported)
+		}
+		if upArgs.netfilterMode != "off" {
+			return errors.New("--netfilter-mode values besides \"off\" " + notSupported)
 		}
 	}
 
-	c, bc, ctx, cancel := connect(ctx)
+	prefs, err := prefsFromUpArgs(upArgs, warnf, st, runtime.GOOS)
+	if err != nil {
+		fatalf("%s", err)
+	}
+
+	if len(prefs.AdvertiseRoutes) > 0 {
+		if err := tailscale.CheckIPForwarding(context.Background()); err != nil {
+			warnf("%v", err)
+		}
+	}
+
+	curPrefs, err := tailscale.GetPrefs(ctx)
+	if err != nil {
+		return err
+	}
+
+	flagSet := map[string]bool{}
+	mp := new(ipn.MaskedPrefs)
+	mp.WantRunningSet = true
+	mp.Prefs = *prefs
+	upFlagSet.Visit(func(f *flag.Flag) {
+		updateMaskedPrefsFromUpFlag(mp, f.Name)
+		flagSet[f.Name] = true
+	})
+
+	if !upArgs.reset {
+		if err := checkForAccidentalSettingReverts(flagSet, curPrefs, mp, os.Getenv("USER")); err != nil {
+			fatalf("%s", err)
+		}
+	}
+
+	controlURLChanged := curPrefs.ControlURL != prefs.ControlURL
+	if controlURLChanged && st.BackendState == ipn.Running.String() && !upArgs.forceReauth {
+		fatalf("can't change --login-server without --force-reauth")
+	}
+
+	// If we're already running and none of the flags require a
+	// restart, we can just do an EditPrefs call and change the
+	// prefs at runtime (e.g. changing hostname, changinged
+	// advertised tags, routes, etc)
+	justEdit := st.BackendState == ipn.Running.String() &&
+		!upArgs.forceReauth &&
+		!upArgs.reset &&
+		upArgs.authKey == "" &&
+		!controlURLChanged
+	if justEdit {
+		_, err := tailscale.EditPrefs(ctx, mp)
+		return err
+	}
+
+	// simpleUp is whether we're running a simple "tailscale up"
+	// to transition to running from a previously-logged-in but
+	// down state, without changing any settings.
+	simpleUp := len(flagSet) == 0 && curPrefs.Persist != nil && curPrefs.Persist.LoginName != ""
+
+	// At this point we need to subscribe to the IPN bus to watch
+	// for state transitions and possible need to authenticate.
+	c, bc, pumpCtx, cancel := connect(ctx)
 	defer cancel()
 
-	if !prefs.ExitNodeIP.IsZero() {
-		st, err := tailscale.Status(ctx)
-		if err != nil {
-			fatalf("can't fetch status from tailscaled: %v", err)
-		}
-		for _, ip := range st.TailscaleIPs {
-			if prefs.ExitNodeIP == ip {
-				fatalf("cannot use %s as the exit node as it is a local IP address to this machine, did you mean --advertise-exit-node?", ip)
-			}
-		}
-	}
+	startingOrRunning := make(chan bool, 1) // gets value once starting or running
+	gotEngineUpdate := make(chan bool, 1)   // gets value upon an engine update
+	go pump(pumpCtx, bc, c)
 
-	var printed bool
+	printed := !simpleUp
 	var loginOnce sync.Once
 	startLoginInteractive := func() { loginOnce.Do(func() { bc.StartLoginInteractive() }) }
 
-	bc.SetPrefs(prefs)
-
-	opts := ipn.Options{
-		StateKey: ipn.GlobalDaemonStateKey,
-		AuthKey:  upArgs.authKey,
-		Notify: func(n ipn.Notify) {
-			if n.ErrMessage != nil {
-				msg := *n.ErrMessage
-				if msg == ipn.ErrMsgPermissionDenied {
-					switch runtime.GOOS {
-					case "windows":
-						msg += " (Tailscale service in use by other user?)"
-					default:
-						msg += " (try 'sudo tailscale up [...]')"
-					}
-				}
-				fatalf("backend error: %v\n", msg)
+	bc.SetNotifyCallback(func(n ipn.Notify) {
+		if n.Engine != nil {
+			select {
+			case gotEngineUpdate <- true:
+			default:
 			}
-			if s := n.State; s != nil {
-				switch *s {
-				case ipn.NeedsLogin:
-					printed = true
-					startLoginInteractive()
-				case ipn.NeedsMachineAuth:
-					printed = true
-					fmt.Fprintf(os.Stderr, "\nTo authorize your machine, visit (as admin):\n\n\t%s/admin/machines\n\n", upArgs.server)
-				case ipn.Starting, ipn.Running:
-					// Done full authentication process
-					if printed {
-						// Only need to print an update if we printed the "please click" message earlier.
-						fmt.Fprintf(os.Stderr, "Success.\n")
-					}
-					cancel()
+		}
+		if n.ErrMessage != nil {
+			msg := *n.ErrMessage
+			if msg == ipn.ErrMsgPermissionDenied {
+				switch runtime.GOOS {
+				case "windows":
+					msg += " (Tailscale service in use by other user?)"
+				default:
+					msg += " (try 'sudo tailscale up [...]')"
 				}
 			}
-			if url := n.BrowseToURL; url != nil {
-				fmt.Fprintf(os.Stderr, "\nTo authenticate, visit:\n\n\t%s\n\n", *url)
+			fatalf("backend error: %v\n", msg)
+		}
+		if s := n.State; s != nil {
+			switch *s {
+			case ipn.NeedsLogin:
+				printed = true
+				startLoginInteractive()
+			case ipn.NeedsMachineAuth:
+				printed = true
+				fmt.Fprintf(os.Stderr, "\nTo authorize your machine, visit (as admin):\n\n\t%s/admin/machines\n\n", upArgs.server)
+			case ipn.Starting, ipn.Running:
+				// Done full authentication process
+				if printed {
+					// Only need to print an update if we printed the "please click" message earlier.
+					fmt.Fprintf(os.Stderr, "Success.\n")
+				}
+				select {
+				case startingOrRunning <- true:
+				default:
+				}
+				cancel()
 			}
-		},
+		}
+		if url := n.BrowseToURL; url != nil && printAuthURL(*url) {
+			printed = true
+			fmt.Fprintf(os.Stderr, "\nTo authenticate, visit:\n\n\t%s\n\n", *url)
+		}
+	})
+	// Wait for backend client to be connected so we know
+	// we're subscribed to updates. Otherwise we can miss
+	// an update upon its transition to running. Do so by causing some traffic
+	// back to the bus that we then wait on.
+	bc.RequestEngineStatus()
+	select {
+	case <-gotEngineUpdate:
+	case <-pumpCtx.Done():
+		return pumpCtx.Err()
 	}
 
-	// On Windows, we still run in mostly the "legacy" way that
-	// predated the server's StateStore. That is, we send an empty
-	// StateKey and send the prefs directly. Although the Windows
-	// supports server mode, though, the transition to StateStore
-	// is only half complete. Only server mode uses it, and the
-	// Windows service (~tailscaled) is the one that computes the
-	// StateKey based on the connection identity. So for now, just
-	// do as the Windows GUI's always done:
-	if runtime.GOOS == "windows" {
-		// The Windows service will set this as needed based
-		// on our connection's identity.
-		opts.StateKey = ""
-		opts.Prefs = prefs
-	}
+	// Special case: bare "tailscale up" means to just start
+	// running, if there's ever been a login.
+	if simpleUp {
+		_, err := tailscale.EditPrefs(ctx, &ipn.MaskedPrefs{
+			Prefs: ipn.Prefs{
+				WantRunning: true,
+			},
+			WantRunningSet: true,
+		})
+		if err != nil {
+			return err
+		}
+	} else {
+		bc.SetPrefs(prefs)
 
-	// 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)
-	if upArgs.forceReauth {
-		printed = true
+		opts := ipn.Options{
+			StateKey: ipn.GlobalDaemonStateKey,
+			AuthKey:  upArgs.authKey,
+		}
+		// On Windows, we still run in mostly the "legacy" way that
+		// predated the server's StateStore. That is, we send an empty
+		// StateKey and send the prefs directly. Although the Windows
+		// supports server mode, though, the transition to StateStore
+		// is only half complete. Only server mode uses it, and the
+		// Windows service (~tailscaled) is the one that computes the
+		// StateKey based on the connection identity. So for now, just
+		// do as the Windows GUI's always done:
+		if runtime.GOOS == "windows" {
+			// The Windows service will set this as needed based
+			// on our connection's identity.
+			opts.StateKey = ""
+			opts.Prefs = prefs
+		}
+
+		bc.Start(opts)
 		startLoginInteractive()
 	}
-	pump(ctx, bc, c)
 
-	return nil
+	select {
+	case <-startingOrRunning:
+		return nil
+	case <-pumpCtx.Done():
+		select {
+		case <-startingOrRunning:
+			return nil
+		default:
+		}
+		return pumpCtx.Err()
+	}
+}
+
+var (
+	flagForPref = map[string]string{} // "ExitNodeIP" => "exit-node"
+	prefsOfFlag = map[string][]string{}
+)
+
+func init() {
+	addPrefFlagMapping("accept-dns", "CorpDNS")
+	addPrefFlagMapping("accept-routes", "RouteAll")
+	addPrefFlagMapping("advertise-routes", "AdvertiseRoutes")
+	addPrefFlagMapping("advertise-tags", "AdvertiseTags")
+	addPrefFlagMapping("host-routes", "AllowSingleHosts")
+	addPrefFlagMapping("hostname", "Hostname")
+	addPrefFlagMapping("login-server", "ControlURL")
+	addPrefFlagMapping("netfilter-mode", "NetfilterMode")
+	addPrefFlagMapping("shields-up", "ShieldsUp")
+	addPrefFlagMapping("snat-subnet-routes", "NoSNAT")
+	addPrefFlagMapping("exit-node", "ExitNodeIP", "ExitNodeIP")
+	addPrefFlagMapping("exit-node-allow-lan-access", "ExitNodeAllowLANAccess")
+	addPrefFlagMapping("unattended", "ForceDaemon")
+	addPrefFlagMapping("operator", "OperatorUser")
+}
+
+func addPrefFlagMapping(flagName string, prefNames ...string) {
+	prefsOfFlag[flagName] = prefNames
+	prefType := reflect.TypeOf(ipn.Prefs{})
+	for _, pref := range prefNames {
+		flagForPref[pref] = flagName
+
+		// Crash at runtime if there's a typo in the prefName.
+		if _, ok := prefType.FieldByName(pref); !ok {
+			panic(fmt.Sprintf("invalid ipn.Prefs field %q", pref))
+		}
+	}
+}
+
+func updateMaskedPrefsFromUpFlag(mp *ipn.MaskedPrefs, flagName string) {
+	if prefs, ok := prefsOfFlag[flagName]; ok {
+		for _, pref := range prefs {
+			reflect.ValueOf(mp).Elem().FieldByName(pref + "Set").SetBool(true)
+		}
+		return
+	}
+	switch flagName {
+	case "authkey", "force-reauth", "reset":
+		// Not pref-related flags.
+	case "advertise-exit-node":
+		// This pref is a shorthand for advertise-routes.
+	default:
+		panic(fmt.Sprintf("internal error: unhandled flag %q", flagName))
+	}
+}
+
+// checkForAccidentalSettingReverts checks for people running
+// "tailscale up" with a subset of the flags they originally ran it
+// with.
+//
+// For example, in Tailscale 1.6 and prior, a user might've advertised
+// a tag, but later tried to change just one other setting and forgot
+// to mention the tag later and silently wiped it out. We now
+// require --reset to change preferences to flag default values when
+// the flag is not mentioned on the command line.
+//
+// curPrefs is what's currently active on the server.
+//
+// mp is the mask of settings actually set, where mp.Prefs is the new
+// preferences to set, including any values set from implicit flags.
+func checkForAccidentalSettingReverts(flagSet map[string]bool, curPrefs *ipn.Prefs, mp *ipn.MaskedPrefs, curUser string) error {
+	if len(flagSet) == 0 {
+		// A bare "tailscale up" is a special case to just
+		// mean bringing the network up without any changes.
+		return nil
+	}
+	if curPrefs.ControlURL == "" {
+		// Don't validate things on initial "up" before a control URL has been set.
+		return nil
+	}
+	curWithExplicitEdits := curPrefs.Clone()
+	curWithExplicitEdits.ApplyEdits(mp)
+
+	prefType := reflect.TypeOf(ipn.Prefs{})
+
+	// Explicit values (current + explicit edit):
+	ev := reflect.ValueOf(curWithExplicitEdits).Elem()
+	// Implicit values (what we'd get if we replaced everything with flag defaults):
+	iv := reflect.ValueOf(&mp.Prefs).Elem()
+	var errs []error
+	var didExitNodeErr bool
+	for i := 0; i < prefType.NumField(); i++ {
+		prefName := prefType.Field(i).Name
+		if prefName == "Persist" {
+			continue
+		}
+		flagName, hasFlag := flagForPref[prefName]
+
+		// Special case for advertise-exit-node; which is a
+		// flag but doesn't have a corresponding pref.  The
+		// flag augments advertise-routes, so we have to infer
+		// the imaginary pref's current value from the routes.
+		if prefName == "AdvertiseRoutes" &&
+			hasExitNodeRoutes(curPrefs.AdvertiseRoutes) &&
+			!hasExitNodeRoutes(curWithExplicitEdits.AdvertiseRoutes) &&
+			!flagSet["advertise-exit-node"] {
+			errs = append(errs, errors.New("'tailscale up' without --reset requires all preferences with changing values to be explicitly mentioned; --advertise-exit-node flag not mentioned but currently advertised routes are an exit node"))
+		}
+
+		if hasFlag && flagSet[flagName] {
+			continue
+		}
+		// Get explicit value and implicit value
+		ex, im := ev.Field(i), iv.Field(i)
+		switch ex.Kind() {
+		case reflect.String, reflect.Slice:
+			if ex.Kind() == reflect.Slice && ex.Len() == 0 && im.Len() == 0 {
+				// Treat nil and non-nil empty slices as equivalent.
+				continue
+			}
+		}
+		exi, imi := ex.Interface(), im.Interface()
+
+		if reflect.DeepEqual(exi, imi) {
+			continue
+		}
+		if flagName == "operator" && imi == "" && exi == curUser {
+			// Don't require setting operator if the current user matches
+			// the configured operator.
+			continue
+		}
+		switch flagName {
+		case "":
+			errs = append(errs, fmt.Errorf("'tailscale up' without --reset requires all preferences with changing values to be explicitly mentioned; this command would change the value of flagless pref %q", prefName))
+		case "exit-node":
+			if !didExitNodeErr {
+				didExitNodeErr = true
+				errs = append(errs, errors.New("'tailscale up' without --reset requires all preferences with changing values to be explicitly mentioned; --exit-node is not specified but an exit node is currently configured"))
+			}
+		default:
+			errs = append(errs, fmt.Errorf("'tailscale up' without --reset requires all preferences with changing values to be explicitly mentioned; --%s is not specified but its default value of %v differs from current value %v",
+				flagName, fmtSettingVal(imi), fmtSettingVal(exi)))
+		}
+	}
+	return multierror.New(errs)
+}
+
+func fmtSettingVal(v interface{}) string {
+	switch v := v.(type) {
+	case bool:
+		return strconv.FormatBool(v)
+	case string, preftype.NetfilterMode:
+		return fmt.Sprintf("%q", v)
+	case []string:
+		return strings.Join(v, ",")
+	}
+	return fmt.Sprint(v)
+}
+
+func hasExitNodeRoutes(rr []netaddr.IPPrefix) bool {
+	var v4, v6 bool
+	for _, r := range rr {
+		if r.Bits == 0 {
+			if r.IP.Is4() {
+				v4 = true
+			} else if r.IP.Is6() {
+				v6 = true
+			}
+		}
+	}
+	return v4 && v6
 }

cmd/tailscale/cli/web.go

@@ -17,10 +17,12 @@ import (
 	"net/http/cgi"
 	"os/exec"
 	"runtime"
+	"strings"
 
 	"github.com/peterbourgon/ff/v2/ffcli"
 	"tailscale.com/client/tailscale"
 	"tailscale.com/ipn"
+	"tailscale.com/tailcfg"
 	"tailscale.com/types/preftype"
 	"tailscale.com/version/distro"
 )
@@ -28,9 +30,18 @@ import (
 //go:embed web.html
 var webHTML string
 
-var tmpl = template.Must(template.New("html").Parse(webHTML))
+//go:embed web.css
+var webCSS string
+
+var tmpl *template.Template
+
+func init() {
+	tmpl = template.Must(template.New("web.html").Parse(webHTML))
+	template.Must(tmpl.New("web.css").Parse(webCSS))
+}
 
 type tmplData struct {
+	Profile      tailcfg.UserProfile
 	SynologyUser string
 	Status       string
 	DeviceName   string
@@ -117,6 +128,67 @@ req.send(null);
 </body></html>
 `
 
+const authenticationRedirectHTML = `
+<html>
+<head>
+	<title>Redirecting...</title>
+	<style>
+		html,
+		body {
+			height: 100%;
+		}
+
+		html {
+			background-color: rgb(249, 247, 246);
+			font-family: ui-sans-serif, system-ui, -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, "Helvetica Neue", Arial, "Noto Sans", sans-serif, "Apple Color Emoji", "Segoe UI Emoji", "Segoe UI Symbol", "Noto Color Emoji";
+			line-height: 1.5;
+			-webkit-text-size-adjust: 100%;
+			-webkit-font-smoothing: antialiased;
+			-moz-osx-font-smoothing: grayscale;
+		}
+
+		body {
+			display: flex;
+			flex-direction: column;
+			align-items: center;
+			justify-content: center;
+		}
+
+		.spinner {
+			margin-bottom: 2rem;
+			border: 4px rgba(112, 110, 109, 0.5) solid;
+			border-left-color: transparent;
+			border-radius: 9999px;
+			width: 4rem;
+			height: 4rem;
+			-webkit-animation: spin 700ms linear infinite;
+      animation: spin 800ms linear infinite;
+		}
+
+		.label {
+			color: rgb(112, 110, 109);
+			padding-left: 0.4rem;
+		}
+
+		@-webkit-keyframes spin {
+			to {
+				transform: rotate(360deg);
+			}
+		}
+
+		@keyframes spin {
+			to {
+				transform: rotate(360deg);
+			}
+		}
+	</style>
+</head>
+<body>
+	<div class="spinner"></div>
+	<div class="label">Redirecting...</div>
+</body>
+`
+
 func webHandler(w http.ResponseWriter, r *http.Request) {
 	if synoTokenRedirect(w, r) {
 		return
@@ -128,6 +200,11 @@ func webHandler(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 
+	if r.URL.Path == "/redirect" || r.URL.Path == "/redirect/" {
+		w.Write([]byte(authenticationRedirectHTML))
+		return
+	}
+
 	if r.Method == "POST" {
 		type mi map[string]interface{}
 		w.Header().Set("Content-Type", "application/json")
@@ -143,12 +220,16 @@ func webHandler(w http.ResponseWriter, r *http.Request) {
 	st, err := tailscale.Status(r.Context())
 	if err != nil {
 		http.Error(w, err.Error(), 500)
+		return
 	}
 
+	profile := st.User[st.Self.UserID]
+	deviceName := strings.Split(st.Self.DNSName, ".")[0]
 	data := tmplData{
 		SynologyUser: user,
+		Profile:      profile,
 		Status:       st.BackendState,
-		DeviceName:   st.Self.DNSName,
+		DeviceName:   deviceName,
 	}
 	if len(st.TailscaleIPs) != 0 {
 		data.IP = st.TailscaleIPs[0].String()
@@ -165,7 +246,7 @@ func webHandler(w http.ResponseWriter, r *http.Request) {
 // TODO(crawshaw): some of this is very similar to the code in 'tailscale up', can we share anything?
 func tailscaleUp(ctx context.Context) (authURL string, retErr error) {
 	prefs := ipn.NewPrefs()
-	prefs.ControlURL = "https://login.tailscale.com"
+	prefs.ControlURL = ipn.DefaultControlURL
 	prefs.WantRunning = true
 	prefs.CorpDNS = true
 	prefs.AllowSingleHosts = true
@@ -178,30 +259,30 @@ func tailscaleUp(ctx context.Context) (authURL string, retErr error) {
 	c, bc, ctx, cancel := connect(ctx)
 	defer cancel()
 
+	bc.SetNotifyCallback(func(n ipn.Notify) {
+		if n.ErrMessage != nil {
+			msg := *n.ErrMessage
+			if msg == ipn.ErrMsgPermissionDenied {
+				switch runtime.GOOS {
+				case "windows":
+					msg += " (Tailscale service in use by other user?)"
+				default:
+					msg += " (try 'sudo tailscale up [...]')"
+				}
+			}
+			retErr = fmt.Errorf("backend error: %v", msg)
+			cancel()
+		} else if url := n.BrowseToURL; url != nil {
+			authURL = *url
+			cancel()
+		}
+	})
+
 	bc.SetPrefs(prefs)
 
-	opts := ipn.Options{
+	bc.Start(ipn.Options{
 		StateKey: ipn.GlobalDaemonStateKey,
-		Notify: func(n ipn.Notify) {
-			if n.ErrMessage != nil {
-				msg := *n.ErrMessage
-				if msg == ipn.ErrMsgPermissionDenied {
-					switch runtime.GOOS {
-					case "windows":
-						msg += " (Tailscale service in use by other user?)"
-					default:
-						msg += " (try 'sudo tailscale up [...]')"
-					}
-				}
-				retErr = fmt.Errorf("backend error: %v", msg)
-				cancel()
-			} else if url := n.BrowseToURL; url != nil {
-				authURL = *url
-				cancel()
-			}
-		},
-	}
-	bc.Start(opts)
+	})
 	bc.StartLoginInteractive()
 	pump(ctx, bc, c)
 

cmd/tailscale/cli/web.html

@@ -1,47 +1,150 @@
 <!doctype html>
-<html><title>Tailscale Client</title><body>
-<h1>Tailscale</h1>
-<div style="float:right;">{{.SynologyUser}}</div>
-<table>
-<tr><th>Status:</th><td>{{.Status}}</td></tr>
-<tr><th>Device Name:</th><td>{{.DeviceName}}</td></tr>
-<tr><th>Tailscale IP:</th><td>{{.IP}}</td></tr>
-</table>
+<html class="bg-gray-50">
 
-<p><input id="login" type="button" value="Log in…"></p>
+<head>
+	<meta charset="utf-8" />
+	<meta name="viewport" content="width=device-width, initial-scale=1" />
+	<link rel="shortcut icon"
+		href="data:image/png;base64,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" />
+	<title>Tailscale</title>
+	<style>{{template "web.css"}}</style>
+</head>
 
-<script>
-login.onclick = function() {
-	const urlParams = new URLSearchParams(window.location.search);
-	const token = urlParams.get("SynoToken");
+<body class="py-14">
+	<main class="container max-w-lg mx-auto py-6 px-8 bg-white rounded-md shadow-2xl" style="width: 95%">
+		<header class="flex justify-between items-center min-width-0 py-2 mb-8">
+			<svg width="26" height="26" viewBox="0 0 23 23" title="Tailscale" fill="none" xmlns="http://www.w3.org/2000/svg"
+				class="flex-shrink-0 mr-4">
+				<circle opacity="0.2" cx="3.4" cy="3.25" r="2.7" fill="currentColor"></circle>
+				<circle cx="3.4" cy="11.3" r="2.7" fill="currentColor"></circle>
+				<circle opacity="0.2" cx="3.4" cy="19.5" r="2.7" fill="currentColor"></circle>
+				<circle cx="11.5" cy="11.3" r="2.7" fill="currentColor"></circle>
+				<circle cx="11.5" cy="19.5" r="2.7" fill="currentColor"></circle>
+				<circle opacity="0.2" cx="11.5" cy="3.25" r="2.7" fill="currentColor"></circle>
+				<circle opacity="0.2" cx="19.5" cy="3.25" r="2.7" fill="currentColor"></circle>
+				<circle cx="19.5" cy="11.3" r="2.7" fill="currentColor"></circle>
+				<circle opacity="0.2" cx="19.5" cy="19.5" r="2.7" fill="currentColor"></circle>
+			</svg>
+			<div class="flex items-center justify-end space-x-2 w-2/3">
+				{{ with .Profile.LoginName }}
+				<div class="text-right truncate leading-4">
+					<h4 class="truncate">{{.}}</h4>
+					<a href="#" class="text-xs text-gray-500 hover:text-gray-700 js-loginButton">Switch account</a>
+				</div>
+				{{ end }}
+				<div class="relative flex-shrink-0 w-8 h-8 rounded-full overflow-hidden">
+					{{ with .Profile.ProfilePicURL }}
+					<div class="w-8 h-8 flex pointer-events-none rounded-full bg-gray-200"
+						style="background-image: url('{{.}}'); background-size: cover;"></div>
+					{{ else }}
+					<div class="w-8 h-8 flex pointer-events-none rounded-full border border-gray-400 border-dashed"></div>
+					{{ end }}
+				</div>
+			</div>
+		</header>
+		{{ if .IP }}
+		<div
+			class="border border-gray-200 bg-gray-0 rounded-lg p-2 pl-3 pr-3 mb-8 width-full flex items-center justify-between">
+			<div class="flex items-center min-width-0">
+				<svg class="flex-shrink-0 text-gray-600 mr-3 ml-1" xmlns="http://www.w3.org/2000/svg" width="20" height="20"
+					viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round"
+					stroke-linejoin="round">
+					<rect x="2" y="2" width="20" height="8" rx="2" ry="2"></rect>
+					<rect x="2" y="14" width="20" height="8" rx="2" ry="2"></rect>
+					<line x1="6" y1="6" x2="6.01" y2="6"></line>
+					<line x1="6" y1="18" x2="6.01" y2="18"></line>
+				</svg>
+				<h4 class="font-semibold truncate mr-2">{{.DeviceName}}</h4>
+			</div>
+			<h5>{{.IP}}</h5>
+		</div>
+		{{ end }}
+		{{ if or (eq .Status "NeedsLogin") (eq .Status "NoState") }}
+		{{ if .IP }}
+		<div class="mb-6">
+			<p class="text-gray-700">Your device's key has expired. Reauthenticate this device by logging in again, or <a
+					href="https://tailscale.com/kb/1028/key-expiry" class="link" target="_blank">learn more</a>.</p>
+		</div>
+		<a href="#" class="mb-4 js-loginButton" target="_blank">
+			<button class="button button-blue w-full">Reauthenticate</button>
+		</a>
+		{{ else }}
+		<div class="mb-6">
+			<h3 class="text-3xl font-semibold mb-3">Log in</h3>
+			<p class="text-gray-700">Get started by logging in to your Tailscale network. Or,&nbsp;learn&nbsp;more at <a
+					href="https://tailscale.com/" class="link" target="_blank">tailscale.com</a>.</p>
+		</div>
+		<a href="#" class="mb-4 js-loginButton" target="_blank">
+			<button class="button button-blue w-full">Log In</button>
+		</a>
+		{{ end }}
+		{{ else if eq .Status "NeedsMachineAuth" }}
+		<div class="mb-4">
+			This device is authorized, but needs approval from a network admin before it can connect to the network.
+		</div>
+		{{ else }}
+		<div class="mb-4">
+			<p>You are connected! Access this device over Tailscale using the device name or IP address above.</p>
+		</div>
+		<a href="#" class="mb-4 link font-medium js-loginButton" target="_blank">Reauthenticate</a>
+		{{ end }}
+	</main>
+	<script>
+		(function () {
+			let loginButtons = document.querySelectorAll(".js-loginButton");
+			let fetchingUrl = false;
 
-	var params = new URLSearchParams("up=true");
-	if (token) {
-		params.set("SynoToken", token)
-	}
+			function handleClick(e) {
+				e.preventDefault();
 
-	var req = new XMLHttpRequest();
-	const url = [location.protocol, '//', location.host, location.pathname, "?", params.toString()].join('');
-	req.overrideMimeType("application/json");
-	req.open("POST", url, true);
-	req.onload = function() {
-		var jsonResponse = JSON.parse(req.responseText);
-		const err = jsonResponse["error"];
-		if (err) {
-			document.body.innerText = err;
-			return
-		}
-		var url = jsonResponse["url"];
-		console.log("jsonResponse: ", jsonResponse);
-		if (url) {
-			document.location.href = url;
-		} else {
-			//location.reload();
-		}
-	};
-	req.send(null);
-}
-</script>
+				if (fetchingUrl) {
+					return;
+				}
 
+				fetchingUrl = true;
+				const urlParams = new URLSearchParams(window.location.search);
+				const token = urlParams.get("SynoToken");
+				const nextParams = new URLSearchParams({ up: true });
+				if (token) {
+					nextParams.set("SynoToken", token)
+				}
+				const nextUrl = new URL(window.location);
+				nextUrl.search = nextParams.toString()
+				const url = nextUrl.toString();
+
+				const tab = window.open("/redirect", "_blank");
+
+				fetch(url, {
+					method: "POST",
+					headers: {
+						"Accept": "application/json",
+						"Content-Type": "application/json",
+					}
+				}).then(res => res.json()).then(res => {
+					fetchingUrl = false;
+					const err = res["error"];
+					if (err) {
+						throw new Error(err);
+					}
+					const url = res["url"];
+					if (url) {
+						authUrl = url;
+						tab.location = url;
+						tab.focus();
+					} else {
+						location.reload();
+					}
+				}).catch(err => {
+					tab.close();
+					alert("Failed to log in: " + err.message);
+				});
+			}
+
+			Array.from(loginButtons).forEach(el => {
+				el.addEventListener("click", handleClick);
+			})
+		})();
+	</script>
 </body>
+
 </html>

cmd/tailscale/depaware.txt

@@ -2,6 +2,7 @@ tailscale.com/cmd/tailscale dependencies: (generated by github.com/tailscale/dep
 
    W 💣 github.com/alexbrainman/sspi                                 from github.com/alexbrainman/sspi/negotiate
    W 💣 github.com/alexbrainman/sspi/negotiate                       from tailscale.com/net/tshttpproxy
+        github.com/go-multierror/multierror                          from tailscale.com/cmd/tailscale/cli
         github.com/peterbourgon/ff/v2                                from github.com/peterbourgon/ff/v2/ffcli
         github.com/peterbourgon/ff/v2/ffcli                          from tailscale.com/cmd/tailscale/cli
         github.com/tcnksm/go-httpstat                                from tailscale.com/net/netcheck
@@ -14,12 +15,13 @@ tailscale.com/cmd/tailscale dependencies: (generated by github.com/tailscale/dep
         rsc.io/goversion/version                                     from tailscale.com/version
         tailscale.com/atomicfile                                     from tailscale.com/ipn
         tailscale.com/client/tailscale                               from tailscale.com/cmd/tailscale/cli
+        tailscale.com/client/tailscale/apitype                       from tailscale.com/client/tailscale
         tailscale.com/cmd/tailscale/cli                              from tailscale.com/cmd/tailscale
         tailscale.com/derp                                           from tailscale.com/derp/derphttp
         tailscale.com/derp/derphttp                                  from tailscale.com/net/netcheck
         tailscale.com/derp/derpmap                                   from tailscale.com/cmd/tailscale/cli
         tailscale.com/disco                                          from tailscale.com/derp
-        tailscale.com/ipn                                            from tailscale.com/cmd/tailscale/cli
+        tailscale.com/ipn                                            from tailscale.com/cmd/tailscale/cli+
         tailscale.com/ipn/ipnstate                                   from tailscale.com/cmd/tailscale/cli+
         tailscale.com/metrics                                        from tailscale.com/derp
         tailscale.com/net/dnscache                                   from tailscale.com/derp/derphttp

cmd/tailscaled/depaware.txt

@@ -22,6 +22,7 @@ tailscale.com/cmd/tailscaled dependencies: (generated by github.com/tailscale/de
    L    github.com/mdlayher/sdnotify                                 from tailscale.com/util/systemd
    W    github.com/pkg/errors                                        from github.com/github/certstore
      💣 github.com/tailscale/wireguard-go/conn                       from github.com/tailscale/wireguard-go/device+
+   W 💣 github.com/tailscale/wireguard-go/conn/winrio                from github.com/tailscale/wireguard-go/conn
      💣 github.com/tailscale/wireguard-go/device                     from tailscale.com/wgengine+
      💣 github.com/tailscale/wireguard-go/ipc                        from github.com/tailscale/wireguard-go/device
    W 💣 github.com/tailscale/wireguard-go/ipc/winpipe                from github.com/tailscale/wireguard-go/ipc
@@ -70,6 +71,7 @@ tailscale.com/cmd/tailscaled dependencies: (generated by github.com/tailscale/de
         inet.af/peercred                                             from tailscale.com/ipn/ipnserver
         rsc.io/goversion/version                                     from tailscale.com/version
         tailscale.com/atomicfile                                     from tailscale.com/ipn+
+        tailscale.com/client/tailscale/apitype                       from tailscale.com/ipn/ipnlocal+
         tailscale.com/control/controlclient                          from tailscale.com/ipn/ipnlocal+
         tailscale.com/derp                                           from tailscale.com/derp/derphttp+
         tailscale.com/derp/derphttp                                  from tailscale.com/net/netcheck+
@@ -91,6 +93,7 @@ tailscale.com/cmd/tailscaled dependencies: (generated by github.com/tailscale/de
         tailscale.com/logtail/filch                                  from tailscale.com/logpolicy
         tailscale.com/metrics                                        from tailscale.com/derp
         tailscale.com/net/dns                                        from tailscale.com/ipn/ipnlocal+
+        tailscale.com/net/dns/resolver                               from tailscale.com/wgengine+
         tailscale.com/net/dnscache                                   from tailscale.com/control/controlclient+
         tailscale.com/net/dnsfallback                                from tailscale.com/control/controlclient
         tailscale.com/net/flowtrack                                  from tailscale.com/wgengine/filter+
@@ -123,7 +126,6 @@ tailscale.com/cmd/tailscaled dependencies: (generated by github.com/tailscale/de
         tailscale.com/types/netmap                                   from tailscale.com/control/controlclient+
         tailscale.com/types/nettype                                  from tailscale.com/wgengine/magicsock
         tailscale.com/types/opt                                      from tailscale.com/control/controlclient+
-        tailscale.com/types/pad32                                    from tailscale.com/wgengine/magicsock
         tailscale.com/types/persist                                  from tailscale.com/control/controlclient+
         tailscale.com/types/preftype                                 from tailscale.com/ipn+
         tailscale.com/types/strbuilder                               from tailscale.com/net/packet
@@ -162,7 +164,7 @@ tailscale.com/cmd/tailscaled dependencies: (generated by github.com/tailscale/de
         golang.org/x/crypto/salsa20/salsa                            from golang.org/x/crypto/nacl/box+
         golang.org/x/net/bpf                                         from github.com/mdlayher/netlink+
         golang.org/x/net/dns/dnsmessage                              from net+
-        golang.org/x/net/http/httpguts                               from net/http
+        golang.org/x/net/http/httpguts                               from net/http+
         golang.org/x/net/http/httpproxy                              from net/http
         golang.org/x/net/http2/hpack                                 from net/http
         golang.org/x/net/idna                                        from golang.org/x/net/http/httpguts+
@@ -176,7 +178,8 @@ tailscale.com/cmd/tailscaled dependencies: (generated by github.com/tailscale/de
   LD    golang.org/x/sys/unix                                        from github.com/jsimonetti/rtnetlink/internal/unix+
    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
+   W    golang.org/x/sys/windows/svc                                 from tailscale.com/cmd/tailscaled+
+   W    golang.org/x/sys/windows/svc/mgr                             from tailscale.com/cmd/tailscaled
         golang.org/x/term                                            from tailscale.com/logpolicy
         golang.org/x/text/secure/bidirule                            from golang.org/x/net/idna
         golang.org/x/text/transform                                  from golang.org/x/text/secure/bidirule+
@@ -215,6 +218,7 @@ tailscale.com/cmd/tailscaled dependencies: (generated by github.com/tailscale/de
         debug/elf                                                    from rsc.io/goversion/version
         debug/macho                                                  from rsc.io/goversion/version
         debug/pe                                                     from rsc.io/goversion/version
+        embed                                                        from tailscale.com/net/dns
         encoding                                                     from encoding/json+
         encoding/asn1                                                from crypto/x509+
         encoding/base64                                              from encoding/json+
@@ -246,7 +250,8 @@ tailscale.com/cmd/tailscaled dependencies: (generated by github.com/tailscale/de
         net                                                          from crypto/tls+
         net/http                                                     from expvar+
         net/http/httptrace                                           from github.com/tcnksm/go-httpstat+
-        net/http/internal                                            from net/http
+        net/http/httputil                                            from tailscale.com/ipn/localapi
+        net/http/internal                                            from net/http+
         net/http/pprof                                               from tailscale.com/cmd/tailscaled
         net/textproto                                                from golang.org/x/net/http/httpguts+
         net/url                                                      from crypto/x509+

cmd/tailscaled/install_darwin.go

@@ -73,9 +73,18 @@ func uninstallSystemDaemonDarwin(args []string) (ret error) {
 		}
 	}
 
-	err = os.Remove(sysPlist)
-	if os.IsNotExist(err) {
-		err = nil
+	if err := os.Remove(sysPlist); err != nil {
+		if os.IsNotExist(err) {
+			err = nil
+		}
+		if ret == nil {
+			ret = err
+		}
+	}
+	if err := os.Remove(targetBin); err != nil {
+		if os.IsNotExist(err) {
+			err = nil
+		}
 		if ret == nil {
 			ret = err
 		}
@@ -93,6 +102,9 @@ func installSystemDaemonDarwin(args []string) (err error) {
 		}
 	}()
 
+	// Best effort:
+	uninstallSystemDaemonDarwin(nil)
+
 	// Copy ourselves to /usr/local/bin/tailscaled.
 	if err := os.MkdirAll(filepath.Dir(targetBin), 0755); err != nil {
 		return err
@@ -127,9 +139,6 @@ func installSystemDaemonDarwin(args []string) (err error) {
 		return err
 	}
 
-	// Best effort:
-	uninstallSystemDaemonDarwin(nil)
-
 	if err := ioutil.WriteFile(sysPlist, []byte(darwinLaunchdPlist), 0700); err != nil {
 		return err
 	}

cmd/tailscaled/install_windows.go

@@ -0,0 +1,119 @@
+// Copyright (c) 2021 Tailscale Inc & AUTHORS All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package main
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"os"
+	"time"
+
+	"golang.org/x/sys/windows"
+	"golang.org/x/sys/windows/svc"
+	"golang.org/x/sys/windows/svc/mgr"
+	"tailscale.com/logtail/backoff"
+	"tailscale.com/types/logger"
+)
+
+func init() {
+	installSystemDaemon = installSystemDaemonWindows
+	uninstallSystemDaemon = uninstallSystemDaemonWindows
+}
+
+func installSystemDaemonWindows(args []string) (err error) {
+	m, err := mgr.Connect()
+	if err != nil {
+		return fmt.Errorf("failed to connect to Windows service manager: %v", err)
+	}
+
+	service, err := m.OpenService(serviceName)
+	if err == nil {
+		service.Close()
+		return fmt.Errorf("service %q is already installed", serviceName)
+	}
+
+	// no such service; proceed to install the service.
+
+	exe, err := os.Executable()
+	if err != nil {
+		return err
+	}
+
+	c := mgr.Config{
+		ServiceType:  windows.SERVICE_WIN32_OWN_PROCESS,
+		StartType:    mgr.StartAutomatic,
+		ErrorControl: mgr.ErrorNormal,
+		DisplayName:  serviceName,
+		Description:  "Connects this computer to others on the Tailscale network.",
+	}
+
+	service, err = m.CreateService(serviceName, exe, c)
+	if err != nil {
+		return fmt.Errorf("failed to create %q service: %v", serviceName, err)
+	}
+	defer service.Close()
+
+	// Exponential backoff is often too aggressive, so use (mostly)
+	// squares instead.
+	ra := []mgr.RecoveryAction{
+		{mgr.ServiceRestart, 1 * time.Second},
+		{mgr.ServiceRestart, 2 * time.Second},
+		{mgr.ServiceRestart, 4 * time.Second},
+		{mgr.ServiceRestart, 9 * time.Second},
+		{mgr.ServiceRestart, 16 * time.Second},
+		{mgr.ServiceRestart, 25 * time.Second},
+		{mgr.ServiceRestart, 36 * time.Second},
+		{mgr.ServiceRestart, 49 * time.Second},
+		{mgr.ServiceRestart, 64 * time.Second},
+	}
+	const resetPeriodSecs = 60
+	err = service.SetRecoveryActions(ra, resetPeriodSecs)
+	if err != nil {
+		return fmt.Errorf("failed to set service recovery actions: %v", err)
+	}
+
+	return nil
+}
+
+func uninstallSystemDaemonWindows(args []string) (ret error) {
+	m, err := mgr.Connect()
+	if err != nil {
+		return fmt.Errorf("failed to connect to Windows service manager: %v", err)
+	}
+	defer m.Disconnect()
+
+	service, err := m.OpenService(serviceName)
+	if err != nil {
+		return fmt.Errorf("failed to open %q service: %v", serviceName, err)
+	}
+
+	st, err := service.Query()
+	if err != nil {
+		service.Close()
+		return fmt.Errorf("failed to query service state: %v", err)
+	}
+	if st.State != svc.Stopped {
+		service.Control(svc.Stop)
+	}
+	err = service.Delete()
+	service.Close()
+	if err != nil {
+		return fmt.Errorf("failed to delete service: %v", err)
+	}
+
+	bo := backoff.NewBackoff("uninstall", logger.Discard, 30*time.Second)
+	end := time.Now().Add(15 * time.Second)
+	for time.Until(end) > 0 {
+		service, err = m.OpenService(serviceName)
+		if err != nil {
+			// service is no longer openable; success!
+			break
+		}
+		service.Close()
+		bo.BackOff(context.Background(), errors.New("service not deleted"))
+	}
+	return nil
+}

cmd/tailscaled/tailscaled.go

@@ -29,9 +29,12 @@ import (
 	"time"
 
 	"github.com/go-multierror/multierror"
+	"inet.af/netaddr"
 	"tailscale.com/ipn/ipnserver"
 	"tailscale.com/logpolicy"
+	"tailscale.com/net/dns"
 	"tailscale.com/net/socks5"
+	"tailscale.com/net/tsaddr"
 	"tailscale.com/net/tstun"
 	"tailscale.com/paths"
 	"tailscale.com/types/flagtype"
@@ -192,6 +195,7 @@ func run() error {
 	logf = logger.RateLimitedFn(logf, 5*time.Second, 5, 100)
 
 	if args.cleanup {
+		dns.Cleanup(logf, args.tunname)
 		router.Cleanup(logf, args.tunname)
 		return nil
 	}
@@ -228,44 +232,40 @@ func run() error {
 	}
 
 	var ns *netstack.Impl
-	if useNetstack {
-		tunDev, magicConn, ok := e.(wgengine.InternalsGetter).GetInternals()
-		if !ok {
-			log.Fatalf("%T is not a wgengine.InternalsGetter", e)
-		}
-		ns, err = netstack.Create(logf, tunDev, e, magicConn)
-		if err != nil {
-			log.Fatalf("netstack.Create: %v", err)
-		}
-		if err := ns.Start(); err != nil {
-			log.Fatalf("failed to start netstack: %v", err)
-		}
+	if useNetstack || wrapNetstack {
+		onlySubnets := wrapNetstack && !useNetstack
+		ns = mustStartNetstack(logf, e, onlySubnets)
 	}
 
 	if socksListener != nil {
 		srv := &socks5.Server{
 			Logf: logger.WithPrefix(logf, "socks5: "),
 		}
-		if useNetstack {
-			srv.Dialer = func(ctx context.Context, network, addr string) (net.Conn, error) {
+		var (
+			mu  sync.Mutex // guards the following field
+			dns netstack.DNSMap
+		)
+		e.AddNetworkMapCallback(func(nm *netmap.NetworkMap) {
+			mu.Lock()
+			defer mu.Unlock()
+			dns = netstack.DNSMapFromNetworkMap(nm)
+		})
+		useNetstackForIP := func(ip netaddr.IP) bool {
+			// TODO(bradfitz): this isn't exactly right.
+			// We should also support subnets when the
+			// prefs are configured as such.
+			return tsaddr.IsTailscaleIP(ip)
+		}
+		srv.Dialer = func(ctx context.Context, network, addr string) (net.Conn, error) {
+			ipp, err := dns.Resolve(ctx, addr)
+			if err != nil {
+				return nil, err
+			}
+			if ns != nil && useNetstackForIP(ipp.IP) {
 				return ns.DialContextTCP(ctx, addr)
 			}
-		} else {
-			var mu sync.Mutex
-			var dns netstack.DNSMap
-			e.AddNetworkMapCallback(func(nm *netmap.NetworkMap) {
-				mu.Lock()
-				defer mu.Unlock()
-				dns = netstack.DNSMapFromNetworkMap(nm)
-			})
-			srv.Dialer = func(ctx context.Context, network, addr string) (net.Conn, error) {
-				ipp, err := dns.Resolve(ctx, addr)
-				if err != nil {
-					return nil, err
-				}
-				var d net.Dialer
-				return d.DialContext(ctx, network, ipp.String())
-			}
+			var d net.Dialer
+			return d.DialContext(ctx, network, ipp.String())
 		}
 		go func() {
 			log.Fatalf("SOCKS5 server exited: %v", srv.Serve(socksListener))
@@ -298,8 +298,7 @@ func run() error {
 		Port:               41112,
 		StatePath:          args.statepath,
 		AutostartStateKey:  globalStateKey,
-		LegacyConfigPath:   paths.LegacyConfigPath(),
-		SurviveDisconnects: true,
+		SurviveDisconnects: runtime.GOOS != "windows",
 		DebugMux:           debugMux,
 	}
 	err = ipnserver.Run(ctx, logf, pol.PublicID.String(), ipnserver.FixedEngine(e), opts)
@@ -312,16 +311,16 @@ func run() error {
 	return nil
 }
 
-func createEngine(logf logger.Logf, linkMon *monitor.Mon) (e wgengine.Engine, isUserspace bool, err error) {
+func createEngine(logf logger.Logf, linkMon *monitor.Mon) (e wgengine.Engine, useNetstack bool, err error) {
 	if args.tunname == "" {
 		return nil, false, errors.New("no --tun value specified")
 	}
 	var errs []error
 	for _, name := range strings.Split(args.tunname, ",") {
 		logf("wgengine.NewUserspaceEngine(tun %q) ...", name)
-		e, isUserspace, err = tryEngine(logf, linkMon, name)
+		e, useNetstack, err = tryEngine(logf, linkMon, name)
 		if err == nil {
-			return e, isUserspace, nil
+			return e, useNetstack, nil
 		}
 		logf("wgengine.NewUserspaceEngine(tun %q) error: %v", name, err)
 		errs = append(errs, err)
@@ -329,14 +328,36 @@ func createEngine(logf logger.Logf, linkMon *monitor.Mon) (e wgengine.Engine, is
 	return nil, false, multierror.New(errs)
 }
 
-func tryEngine(logf logger.Logf, linkMon *monitor.Mon, name string) (e wgengine.Engine, isUserspace bool, err error) {
+var wrapNetstack = shouldWrapNetstack()
+
+func shouldWrapNetstack() bool {
+	if e := os.Getenv("TS_DEBUG_WRAP_NETSTACK"); e != "" {
+		v, err := strconv.ParseBool(e)
+		if err != nil {
+			log.Fatalf("invalid TS_DEBUG_WRAP_NETSTACK value: %v", err)
+		}
+		return v
+	}
+	if distro.Get() == distro.Synology {
+		return true
+	}
+	switch runtime.GOOS {
+	case "windows", "darwin":
+		// Enable on Windows and tailscaled-on-macOS (this doesn't
+		// affect the GUI clients).
+		return true
+	}
+	return false
+}
+
+func tryEngine(logf logger.Logf, linkMon *monitor.Mon, name string) (e wgengine.Engine, useNetstack bool, err error) {
 	conf := wgengine.Config{
 		ListenPort:  args.port,
 		LinkMonitor: linkMon,
 	}
-	isUserspace = name == "userspace-networking"
-	if !isUserspace {
-		dev, err := tstun.New(logf, name)
+	useNetstack = name == "userspace-networking"
+	if !useNetstack {
+		dev, devName, err := tstun.New(logf, name)
 		if err != nil {
 			tstun.Diagnose(logf, name)
 			return nil, false, err
@@ -347,13 +368,21 @@ func tryEngine(logf logger.Logf, linkMon *monitor.Mon, name string) (e wgengine.
 			dev.Close()
 			return nil, false, err
 		}
+		d, err := dns.NewOSConfigurator(logf, devName)
+		if err != nil {
+			return nil, false, err
+		}
+		conf.DNS = d
 		conf.Router = r
+		if wrapNetstack {
+			conf.Router = netstack.NewSubnetRouterWrapper(conf.Router)
+		}
 	}
 	e, err = wgengine.NewUserspaceEngine(logf, conf)
 	if err != nil {
-		return nil, isUserspace, err
+		return nil, useNetstack, err
 	}
-	return e, isUserspace, nil
+	return e, useNetstack, nil
 }
 
 func newDebugMux() *http.ServeMux {
@@ -375,3 +404,18 @@ func runDebugServer(mux *http.ServeMux, addr string) {
 		log.Fatal(err)
 	}
 }
+
+func mustStartNetstack(logf logger.Logf, e wgengine.Engine, onlySubnets bool) *netstack.Impl {
+	tunDev, magicConn, ok := e.(wgengine.InternalsGetter).GetInternals()
+	if !ok {
+		log.Fatalf("%T is not a wgengine.InternalsGetter", e)
+	}
+	ns, err := netstack.Create(logf, tunDev, e, magicConn, onlySubnets)
+	if err != nil {
+		log.Fatalf("netstack.Create: %v", err)
+	}
+	if err := ns.Start(); err != nil {
+		log.Fatalf("failed to start netstack: %v", err)
+	}
+	return ns
+}

cmd/tailscaled/tailscaled_windows.go

@@ -30,11 +30,13 @@ import (
 	"golang.zx2c4.com/wireguard/windows/tunnel/winipcfg"
 	"tailscale.com/ipn/ipnserver"
 	"tailscale.com/logpolicy"
+	"tailscale.com/net/dns"
 	"tailscale.com/net/tstun"
 	"tailscale.com/tempfork/wireguard-windows/firewall"
 	"tailscale.com/types/logger"
 	"tailscale.com/version"
 	"tailscale.com/wgengine"
+	"tailscale.com/wgengine/netstack"
 	"tailscale.com/wgengine/router"
 )
 
@@ -157,86 +159,120 @@ func beFirewallKillswitch() bool {
 
 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) {
-		dev, err := tstun.New(logf, "Tailscale")
+	getEngineRaw := func() (wgengine.Engine, error) {
+		dev, devName, err := tstun.New(logf, "Tailscale")
 		if err != nil {
-			return nil, err
+			return nil, fmt.Errorf("TUN: %w", err)
 		}
 		r, err := router.New(logf, dev)
 		if err != nil {
 			dev.Close()
-			return nil, err
+			return nil, fmt.Errorf("Router: %w", err)
+		}
+		if wrapNetstack {
+			r = netstack.NewSubnetRouterWrapper(r)
+		}
+		d, err := dns.NewOSConfigurator(logf, devName)
+		if err != nil {
+			r.Close()
+			dev.Close()
+			return nil, fmt.Errorf("DNS: %w", err)
 		}
 		eng, err := wgengine.NewUserspaceEngine(logf, wgengine.Config{
 			Tun:        dev,
 			Router:     r,
+			DNS:        d,
 			ListenPort: 41641,
 		})
 		if err != nil {
 			r.Close()
 			dev.Close()
-			return nil, err
+			return nil, fmt.Errorf("Engine: %w", err)
+		}
+		onlySubnets := true
+		if wrapNetstack {
+			mustStartNetstack(logf, eng, onlySubnets)
 		}
 		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()
+	type engineOrError struct {
+		Engine wgengine.Engine
+		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, err := getEngineRaw()
+			d, dt := time.Since(t1).Round(ms), time.Since(t1).Round(ms)
 			if err != nil {
-				logf("wgengine.NewUserspaceEngine: (try %v) %v", try, err)
-				continue
+				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)
+				}
 			}
-			if try > 1 {
-				logf("wgengine.NewUserspaceEngine: ended up working on try %v", try)
+			timer := time.NewTimer(5 * time.Second)
+			engErrc <- engineOrError{eng, err}
+			if err == nil {
+				timer.Stop()
+				return
 			}
-			break
+			<-timer.C
 		}
-	}
-	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
+
+	// 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, error) {
+		if msg := os.Getenv("TS_DEBUG_WIN_FAIL"); msg != "" {
+			return nil, fmt.Errorf("pretending to be a service failure: %v", msg)
+		}
+		for {
+			res := <-engErrc
+			if res.Engine != nil {
+				return res.Engine, 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, fmt.Errorf("%w\n\nlogid: %v", res.Err, logid)
 		}
-	} else {
-		getEngine = ipnserver.FixedEngine(eng)
 	}
-	err = ipnserver.Run(ctx, logf, logid, getEngine, opts)
+	err := ipnserver.Run(ctx, logf, logid, getEngine, opts)
 	if err != nil {
 		logf("ipnserver.Run: %v", err)
 	}
 	return err
 }
+
+var (
+	kernel32           = windows.NewLazySystemDLL("kernel32.dll")
+	getTickCount64Proc = kernel32.NewProc("GetTickCount64")
+)
+
+func windowsUptime() time.Duration {
+	r, _, _ := getTickCount64Proc.Call()
+	return time.Duration(int64(r)) * time.Millisecond
+}

control/controlclient/auto.go

@@ -100,11 +100,22 @@ func (s Status) String() string {
 }
 
 type LoginGoal struct {
-	_            structs.Incomparable
-	wantLoggedIn bool                 // true if we *want* to be logged in
-	token        *tailcfg.Oauth2Token // oauth token to use when logging in
-	flags        LoginFlags           // flags to use when logging in
-	url          string               // auth url that needs to be visited
+	_               structs.Incomparable
+	wantLoggedIn    bool                 // true if we *want* to be logged in
+	token           *tailcfg.Oauth2Token // oauth token to use when logging in
+	flags           LoginFlags           // flags to use when logging in
+	url             string               // auth url that needs to be visited
+	loggedOutResult chan<- error
+}
+
+func (g *LoginGoal) sendLogoutError(err error) {
+	if g.loggedOutResult == nil {
+		return
+	}
+	select {
+	case g.loggedOutResult <- err:
+	default:
+	}
 }
 
 // Client connects to a tailcontrol server for a node.
@@ -363,6 +374,7 @@ func (c *Client) authRoutine() {
 
 		if !goal.wantLoggedIn {
 			err := c.direct.TryLogout(ctx)
+			goal.sendLogoutError(err)
 			if err != nil {
 				report(err, "TryLogout")
 				bo.BackOff(ctx, err)
@@ -402,9 +414,10 @@ func (c *Client) authRoutine() {
 				report(err, f)
 				bo.BackOff(ctx, err)
 				continue
-			} else if url != "" {
+			}
+			if url != "" {
 				if goal.url != "" {
-					err = fmt.Errorf("weird: server required a new url?")
+					err = fmt.Errorf("[unexpected] server required a new URL?")
 					report(err, "WaitLoginURL")
 				}
 
@@ -584,12 +597,16 @@ func (c *Client) mapRoutine() {
 }
 
 func (c *Client) AuthCantContinue() bool {
+	if c == nil {
+		return true
+	}
 	c.mu.Lock()
 	defer c.mu.Unlock()
 
 	return !c.loggedIn && (c.loginGoal == nil || c.loginGoal.url != "")
 }
 
+// SetStatusFunc sets fn as the callback to run on any status change.
 func (c *Client) SetStatusFunc(fn func(Status)) {
 	c.mu.Lock()
 	c.statusFunc = fn
@@ -681,19 +698,50 @@ func (c *Client) Login(t *tailcfg.Oauth2Token, flags LoginFlags) {
 	c.cancelAuth()
 }
 
-func (c *Client) Logout() {
-	c.logf("client.Logout()")
+func (c *Client) StartLogout() {
+	c.logf("client.StartLogout()")
 
 	c.mu.Lock()
 	c.loginGoal = &LoginGoal{
 		wantLoggedIn: false,
 	}
 	c.mu.Unlock()
-
 	c.cancelAuth()
 }
 
-func (c *Client) UpdateEndpoints(localPort uint16, endpoints []string) {
+func (c *Client) Logout(ctx context.Context) error {
+	c.logf("client.Logout()")
+
+	errc := make(chan error, 1)
+
+	c.mu.Lock()
+	c.loginGoal = &LoginGoal{
+		wantLoggedIn:    false,
+		loggedOutResult: errc,
+	}
+	c.mu.Unlock()
+	c.cancelAuth()
+
+	timer := time.NewTimer(10 * time.Second)
+	defer timer.Stop()
+	select {
+	case err := <-errc:
+		return err
+	case <-ctx.Done():
+		return ctx.Err()
+	case <-timer.C:
+		return context.DeadlineExceeded
+	}
+}
+
+// UpdateEndpoints sets the client's discovered endpoints and sends
+// them to the control server if they've changed.
+//
+// It does not retain the provided slice.
+//
+// The localPort field is unused except for integration tests in
+// another repo.
+func (c *Client) UpdateEndpoints(localPort uint16, endpoints []tailcfg.Endpoint) {
 	changed := c.direct.SetEndpoints(localPort, endpoints)
 	if changed {
 		c.sendNewMapRequest()

control/controlclient/direct.go

@@ -23,7 +23,6 @@ import (
 	"path/filepath"
 	"reflect"
 	"runtime"
-	"sort"
 	"strconv"
 	"strings"
 	"sync"
@@ -46,9 +45,9 @@ import (
 	"tailscale.com/types/opt"
 	"tailscale.com/types/persist"
 	"tailscale.com/types/wgkey"
+	"tailscale.com/util/dnsname"
 	"tailscale.com/util/systemd"
 	"tailscale.com/version"
-	"tailscale.com/wgengine/filter"
 	"tailscale.com/wgengine/monitor"
 )
 
@@ -63,9 +62,10 @@ type Direct struct {
 	logf                   logger.Logf
 	linkMon                *monitor.Mon // or nil
 	discoPubKey            tailcfg.DiscoKey
-	machinePrivKey         wgkey.Private
+	getMachinePrivKey      func() (wgkey.Private, error)
 	debugFlags             []string
 	keepSharerAndUserSplit bool
+	skipIPForwardingCheck  bool
 
 	mu           sync.Mutex // mutex guards the following fields
 	serverKey    wgkey.Key
@@ -75,29 +75,34 @@ type Direct struct {
 	expiry       *time.Time
 	// hostinfo is mutated in-place while mu is held.
 	hostinfo      *tailcfg.Hostinfo // always non-nil
-	endpoints     []string
+	endpoints     []tailcfg.Endpoint
 	everEndpoints bool   // whether we've ever had non-empty endpoints
 	localPort     uint16 // or zero to mean auto
 }
 
 type Options struct {
-	Persist           persist.Persist   // initial persistent data
-	MachinePrivateKey wgkey.Private     // the machine key to use
-	ServerURL         string            // URL of the tailcontrol server
-	AuthKey           string            // optional node auth key for auto registration
-	TimeNow           func() time.Time  // time.Now implementation used by Client
-	Hostinfo          *tailcfg.Hostinfo // non-nil passes ownership, nil means to use default using os.Hostname, etc
-	DiscoPublicKey    tailcfg.DiscoKey
-	NewDecompressor   func() (Decompressor, error)
-	KeepAlive         bool
-	Logf              logger.Logf
-	HTTPTestClient    *http.Client // optional HTTP client to use (for tests only)
-	DebugFlags        []string     // debug settings to send to control
-	LinkMonitor       *monitor.Mon // optional link monitor
+	Persist              persist.Persist               // initial persistent data
+	GetMachinePrivateKey func() (wgkey.Private, error) // returns the machine key to use
+	ServerURL            string                        // URL of the tailcontrol server
+	AuthKey              string                        // optional node auth key for auto registration
+	TimeNow              func() time.Time              // time.Now implementation used by Client
+	Hostinfo             *tailcfg.Hostinfo             // non-nil passes ownership, nil means to use default using os.Hostname, etc
+	DiscoPublicKey       tailcfg.DiscoKey
+	NewDecompressor      func() (Decompressor, error)
+	KeepAlive            bool
+	Logf                 logger.Logf
+	HTTPTestClient       *http.Client // optional HTTP client to use (for tests only)
+	DebugFlags           []string     // debug settings to send to control
+	LinkMonitor          *monitor.Mon // optional link monitor
 
 	// KeepSharerAndUserSplit controls whether the client
 	// understands Node.Sharer. If false, the Sharer is mapped to the User.
 	KeepSharerAndUserSplit bool
+
+	// SkipIPForwardingCheck declares that the host's IP
+	// forwarding works and should not be double-checked by the
+	// controlclient package.
+	SkipIPForwardingCheck bool
 }
 
 type Decompressor interface {
@@ -110,8 +115,8 @@ func NewDirect(opts Options) (*Direct, error) {
 	if opts.ServerURL == "" {
 		return nil, errors.New("controlclient.New: no server URL specified")
 	}
-	if opts.MachinePrivateKey.IsZero() {
-		return nil, errors.New("controlclient.New: no MachinePrivateKey specified")
+	if opts.GetMachinePrivateKey == nil {
+		return nil, errors.New("controlclient.New: no GetMachinePrivateKey specified")
 	}
 	opts.ServerURL = strings.TrimRight(opts.ServerURL, "/")
 	serverURL, err := url.Parse(opts.ServerURL)
@@ -138,7 +143,7 @@ func NewDirect(opts Options) (*Direct, error) {
 		tr := http.DefaultTransport.(*http.Transport).Clone()
 		tr.Proxy = tshttpproxy.ProxyFromEnvironment
 		tshttpproxy.SetTransportGetProxyConnectHeader(tr)
-		tr.TLSClientConfig = tlsdial.Config(serverURL.Host, tr.TLSClientConfig)
+		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.ForceAttemptHTTP2 = true
@@ -147,7 +152,7 @@ func NewDirect(opts Options) (*Direct, error) {
 
 	c := &Direct{
 		httpc:                  httpc,
-		machinePrivKey:         opts.MachinePrivateKey,
+		getMachinePrivKey:      opts.GetMachinePrivateKey,
 		serverURL:              opts.ServerURL,
 		timeNow:                opts.TimeNow,
 		logf:                   opts.Logf,
@@ -159,6 +164,7 @@ func NewDirect(opts Options) (*Direct, error) {
 		debugFlags:             opts.DebugFlags,
 		keepSharerAndUserSplit: opts.KeepSharerAndUserSplit,
 		linkMon:                opts.LinkMonitor,
+		skipIPForwardingCheck:  opts.SkipIPForwardingCheck,
 	}
 	if opts.Hostinfo == nil {
 		c.SetHostinfo(NewHostinfo())
@@ -172,6 +178,7 @@ var osVersion func() string // non-nil on some platforms
 
 func NewHostinfo() *tailcfg.Hostinfo {
 	hostname, _ := os.Hostname()
+	hostname = dnsname.FirstLabel(hostname)
 	var osv string
 	if osVersion != nil {
 		osv = osVersion()
@@ -254,40 +261,50 @@ const (
 func (c *Direct) TryLogout(ctx context.Context) error {
 	c.logf("direct.TryLogout()")
 
-	c.mu.Lock()
-	defer c.mu.Unlock()
+	mustRegen, newURL, err := c.doLogin(ctx, loginOpt{Logout: true})
+	c.logf("TryLogout control response: mustRegen=%v, newURL=%v, err=%v", mustRegen, newURL, err)
 
-	// TODO(crawshaw): Tell the server. This node key should be
-	// immediately invalidated.
-	//if !c.persist.PrivateNodeKey.IsZero() {
-	//}
+	c.mu.Lock()
 	c.persist = persist.Persist{}
-	return nil
+	c.mu.Unlock()
+
+	return err
 }
 
 func (c *Direct) TryLogin(ctx context.Context, t *tailcfg.Oauth2Token, flags LoginFlags) (url string, err error) {
 	c.logf("direct.TryLogin(token=%v, flags=%v)", t != nil, flags)
-	return c.doLoginOrRegen(ctx, t, flags, false, "")
+	return c.doLoginOrRegen(ctx, loginOpt{Token: t, Flags: flags})
 }
 
-func (c *Direct) WaitLoginURL(ctx context.Context, url string) (newUrl string, err error) {
+// WaitLoginURL sits in a long poll waiting for the user to authenticate at url.
+//
+// On success, newURL and err will both be nil.
+func (c *Direct) WaitLoginURL(ctx context.Context, url string) (newURL string, err error) {
 	c.logf("direct.WaitLoginURL")
-	return c.doLoginOrRegen(ctx, nil, LoginDefault, false, url)
+	return c.doLoginOrRegen(ctx, loginOpt{URL: url})
 }
 
-func (c *Direct) doLoginOrRegen(ctx context.Context, t *tailcfg.Oauth2Token, flags LoginFlags, regen bool, url string) (newUrl string, err error) {
-	mustregen, url, err := c.doLogin(ctx, t, flags, regen, url)
+func (c *Direct) doLoginOrRegen(ctx context.Context, opt loginOpt) (newURL string, err error) {
+	mustRegen, url, err := c.doLogin(ctx, opt)
 	if err != nil {
 		return url, err
 	}
-	if mustregen {
-		_, url, err = c.doLogin(ctx, t, flags, true, url)
+	if mustRegen {
+		opt.Regen = true
+		_, url, err = c.doLogin(ctx, opt)
 	}
-
 	return url, err
 }
 
-func (c *Direct) doLogin(ctx context.Context, t *tailcfg.Oauth2Token, flags LoginFlags, regen bool, url string) (mustregen bool, newurl string, err error) {
+type loginOpt struct {
+	Token  *tailcfg.Oauth2Token
+	Flags  LoginFlags
+	Regen  bool
+	URL    string
+	Logout bool
+}
+
+func (c *Direct) doLogin(ctx context.Context, opt loginOpt) (mustRegen bool, newURL string, err error) {
 	c.mu.Lock()
 	persist := c.persist
 	tryingNewKey := c.tryingNewKey
@@ -298,26 +315,35 @@ func (c *Direct) doLogin(ctx context.Context, t *tailcfg.Oauth2Token, flags Logi
 	expired := c.expiry != nil && !c.expiry.IsZero() && c.expiry.Before(c.timeNow())
 	c.mu.Unlock()
 
-	if c.machinePrivKey.IsZero() {
-		return false, "", errors.New("controlclient.Direct requires a machine private key")
+	machinePrivKey, err := c.getMachinePrivKey()
+	if err != nil {
+		return false, "", fmt.Errorf("getMachinePrivKey: %w", err)
+	}
+	if machinePrivKey.IsZero() {
+		return false, "", errors.New("getMachinePrivKey returned zero key")
 	}
 
-	if expired {
-		c.logf("Old key expired -> regen=true")
-		systemd.Status("key expired; run 'tailscale up' to authenticate")
-		regen = true
-	}
-	if (flags & LoginInteractive) != 0 {
-		c.logf("LoginInteractive -> regen=true")
-		regen = true
+	regen := opt.Regen
+	if opt.Logout {
+		c.logf("logging out...")
+	} else {
+		if expired {
+			c.logf("Old key expired -> regen=true")
+			systemd.Status("key expired; run 'tailscale up' to authenticate")
+			regen = true
+		}
+		if (opt.Flags & LoginInteractive) != 0 {
+			c.logf("LoginInteractive -> regen=true")
+			regen = true
+		}
 	}
 
-	c.logf("doLogin(regen=%v, hasUrl=%v)", regen, url != "")
+	c.logf("doLogin(regen=%v, hasUrl=%v)", regen, opt.URL != "")
 	if serverKey.IsZero() {
 		var err error
 		serverKey, err = loadServerKey(ctx, c.httpc, c.serverURL)
 		if err != nil {
-			return regen, url, err
+			return regen, opt.URL, err
 		}
 
 		c.mu.Lock()
@@ -326,17 +352,21 @@ func (c *Direct) doLogin(ctx context.Context, t *tailcfg.Oauth2Token, flags Logi
 	}
 
 	var oldNodeKey wgkey.Key
-	if url != "" {
-	} else if regen || persist.PrivateNodeKey.IsZero() {
+	switch {
+	case opt.Logout:
+		tryingNewKey = persist.PrivateNodeKey
+	case opt.URL != "":
+		// Nothing.
+	case regen || persist.PrivateNodeKey.IsZero():
 		c.logf("Generating a new nodekey.")
 		persist.OldPrivateNodeKey = persist.PrivateNodeKey
 		key, err := wgkey.NewPrivate()
 		if err != nil {
 			c.logf("login keygen: %v", err)
-			return regen, url, err
+			return regen, opt.URL, err
 		}
 		tryingNewKey = key
-	} else {
+	default:
 		// Try refreshing the current key first
 		tryingNewKey = persist.PrivateNodeKey
 	}
@@ -345,11 +375,14 @@ func (c *Direct) doLogin(ctx context.Context, t *tailcfg.Oauth2Token, flags Logi
 	}
 
 	if tryingNewKey.IsZero() {
+		if opt.Logout {
+			return false, "", errors.New("no nodekey to log out")
+		}
 		log.Fatalf("tryingNewKey is empty, give up")
 	}
 	if backendLogID == "" {
 		err = errors.New("hostinfo: BackendLogID missing")
-		return regen, url, err
+		return regen, opt.URL, err
 	}
 	now := time.Now().Round(time.Second)
 	request := tailcfg.RegisterRequest{
@@ -357,17 +390,20 @@ func (c *Direct) doLogin(ctx context.Context, t *tailcfg.Oauth2Token, flags Logi
 		OldNodeKey: tailcfg.NodeKey(oldNodeKey),
 		NodeKey:    tailcfg.NodeKey(tryingNewKey.Public()),
 		Hostinfo:   hostinfo,
-		Followup:   url,
+		Followup:   opt.URL,
 		Timestamp:  &now,
 	}
+	if opt.Logout {
+		request.Expiry = time.Unix(123, 0) // far in the past
+	}
 	c.logf("RegisterReq: onode=%v node=%v fup=%v",
 		request.OldNodeKey.ShortString(),
-		request.NodeKey.ShortString(), url != "")
-	request.Auth.Oauth2Token = t
+		request.NodeKey.ShortString(), opt.URL != "")
+	request.Auth.Oauth2Token = opt.Token
 	request.Auth.Provider = persist.Provider
 	request.Auth.LoginName = persist.LoginName
 	request.Auth.AuthKey = authKey
-	err = signRegisterRequest(&request, c.serverURL, c.serverKey, c.machinePrivKey.Public())
+	err = signRegisterRequest(&request, c.serverURL, c.serverKey, machinePrivKey.Public())
 	if err != nil {
 		// If signing failed, clear all related fields
 		request.SignatureType = tailcfg.SignatureNone
@@ -381,34 +417,44 @@ func (c *Direct) doLogin(ctx context.Context, t *tailcfg.Oauth2Token, flags Logi
 			c.logf("RegisterReq sign error: %v", err)
 		}
 	}
-	bodyData, err := encode(request, &serverKey, &c.machinePrivKey)
+	if debugRegister {
+		j, _ := json.MarshalIndent(request, "", "\t")
+		c.logf("RegisterRequest: %s", j)
+	}
+
+	bodyData, err := encode(request, &serverKey, &machinePrivKey)
 	if err != nil {
-		return regen, url, err
+		return regen, opt.URL, err
 	}
 	body := bytes.NewReader(bodyData)
 
-	u := fmt.Sprintf("%s/machine/%s", c.serverURL, c.machinePrivKey.Public().HexString())
+	u := fmt.Sprintf("%s/machine/%s", c.serverURL, machinePrivKey.Public().HexString())
 	req, err := http.NewRequest("POST", u, body)
 	if err != nil {
-		return regen, url, err
+		return regen, opt.URL, err
 	}
 	req = req.WithContext(ctx)
 
 	res, err := c.httpc.Do(req)
 	if err != nil {
-		return regen, url, fmt.Errorf("register request: %v", err)
+		return regen, opt.URL, fmt.Errorf("register request: %v", err)
 	}
 	if res.StatusCode != 200 {
 		msg, _ := ioutil.ReadAll(res.Body)
 		res.Body.Close()
-		return regen, url, fmt.Errorf("register request: http %d: %.200s",
+		return regen, opt.URL, fmt.Errorf("register request: http %d: %.200s",
 			res.StatusCode, strings.TrimSpace(string(msg)))
 	}
 	resp := tailcfg.RegisterResponse{}
-	if err := decode(res, &resp, &serverKey, &c.machinePrivKey); err != nil {
-		c.logf("error decoding RegisterResponse with server key %s and machine key %s: %v", serverKey, c.machinePrivKey.Public(), err)
-		return regen, url, fmt.Errorf("register request: %v", err)
+	if err := decode(res, &resp, &serverKey, &machinePrivKey); err != nil {
+		c.logf("error decoding RegisterResponse with server key %s and machine key %s: %v", serverKey, machinePrivKey.Public(), err)
+		return regen, opt.URL, fmt.Errorf("register request: %v", err)
 	}
+	if debugRegister {
+		j, _ := json.MarshalIndent(resp, "", "\t")
+		c.logf("RegisterResponse: %s", j)
+	}
+
 	// Log without PII:
 	c.logf("RegisterReq: got response; nodeKeyExpired=%v, machineAuthorized=%v; authURL=%v",
 		resp.NodeKeyExpired, resp.MachineAuthorized, resp.AuthURL != "")
@@ -460,7 +506,7 @@ func (c *Direct) doLogin(ctx context.Context, t *tailcfg.Oauth2Token, flags Logi
 	return false, resp.AuthURL, nil
 }
 
-func sameStrings(a, b []string) bool {
+func sameEndpoints(a, b []tailcfg.Endpoint) bool {
 	if len(a) != len(b) {
 		return false
 	}
@@ -476,15 +522,19 @@ func sameStrings(a, b []string) bool {
 // whether they've changed.
 //
 // It does not retain the provided slice.
-func (c *Direct) newEndpoints(localPort uint16, endpoints []string) (changed bool) {
+func (c *Direct) newEndpoints(localPort uint16, endpoints []tailcfg.Endpoint) (changed bool) {
 	c.mu.Lock()
 	defer c.mu.Unlock()
 
 	// Nothing new?
-	if c.localPort == localPort && sameStrings(c.endpoints, endpoints) {
+	if c.localPort == localPort && sameEndpoints(c.endpoints, endpoints) {
 		return false // unchanged
 	}
-	c.logf("client.newEndpoints(%v, %v)", localPort, endpoints)
+	var epStrs []string
+	for _, ep := range endpoints {
+		epStrs = append(epStrs, ep.Addr.String())
+	}
+	c.logf("client.newEndpoints(%v, %v)", localPort, epStrs)
 	c.localPort = localPort
 	c.endpoints = append(c.endpoints[:0], endpoints...)
 	if len(endpoints) > 0 {
@@ -496,7 +546,7 @@ func (c *Direct) newEndpoints(localPort uint16, endpoints []string) (changed boo
 // SetEndpoints updates the list of locally advertised endpoints.
 // It won't be replicated to the server until a *fresh* call to PollNetMap().
 // You don't need to restart PollNetMap if we return changed==false.
-func (c *Direct) SetEndpoints(localPort uint16, endpoints []string) (changed bool) {
+func (c *Direct) SetEndpoints(localPort uint16, endpoints []tailcfg.Endpoint) (changed bool) {
 	// (no log message on function entry, because it clutters the logs
 	//  if endpoints haven't changed. newEndpoints() will log it.)
 	return c.newEndpoints(localPort, endpoints)
@@ -520,6 +570,11 @@ func (c *Direct) SendLiteMapUpdate(ctx context.Context) error {
 	return c.sendMapRequest(ctx, 1, nil)
 }
 
+// If we go more than pollTimeout without hearing from the server,
+// end the long poll. We should be receiving a keep alive ping
+// every minute.
+const pollTimeout = 120 * time.Second
+
 // cb nil means to omit peers.
 func (c *Direct) sendMapRequest(ctx context.Context, maxPolls int, cb func(*netmap.NetworkMap)) error {
 	c.mu.Lock()
@@ -529,10 +584,23 @@ func (c *Direct) sendMapRequest(ctx context.Context, maxPolls int, cb func(*netm
 	hostinfo := c.hostinfo.Clone()
 	backendLogID := hostinfo.BackendLogID
 	localPort := c.localPort
-	ep := append([]string(nil), c.endpoints...)
+	var epStrs []string
+	var epTypes []tailcfg.EndpointType
+	for _, ep := range c.endpoints {
+		epStrs = append(epStrs, ep.Addr.String())
+		epTypes = append(epTypes, ep.Type)
+	}
 	everEndpoints := c.everEndpoints
 	c.mu.Unlock()
 
+	machinePrivKey, err := c.getMachinePrivKey()
+	if err != nil {
+		return fmt.Errorf("getMachinePrivKey: %w", err)
+	}
+	if machinePrivKey.IsZero() {
+		return errors.New("getMachinePrivKey returned zero key")
+	}
+
 	if persist.PrivateNodeKey.IsZero() {
 		return errors.New("privateNodeKey is zero")
 	}
@@ -541,7 +609,7 @@ func (c *Direct) sendMapRequest(ctx context.Context, maxPolls int, cb func(*netm
 	}
 
 	allowStream := maxPolls != 1
-	c.logf("[v1] PollNetMap: stream=%v :%v ep=%v", allowStream, localPort, ep)
+	c.logf("[v1] PollNetMap: stream=%v :%v ep=%v", allowStream, localPort, epStrs)
 
 	vlogf := logger.Discard
 	if Debug.NetMap {
@@ -551,18 +619,20 @@ func (c *Direct) sendMapRequest(ctx context.Context, maxPolls int, cb func(*netm
 	}
 
 	request := &tailcfg.MapRequest{
-		Version:    tailcfg.CurrentMapRequestVersion,
-		KeepAlive:  c.keepAlive,
-		NodeKey:    tailcfg.NodeKey(persist.PrivateNodeKey.Public()),
-		DiscoKey:   c.discoPubKey,
-		Endpoints:  ep,
-		Stream:     allowStream,
-		Hostinfo:   hostinfo,
-		DebugFlags: c.debugFlags,
-		OmitPeers:  cb == nil,
+		Version:       tailcfg.CurrentMapRequestVersion,
+		KeepAlive:     c.keepAlive,
+		NodeKey:       tailcfg.NodeKey(persist.PrivateNodeKey.Public()),
+		DiscoKey:      c.discoPubKey,
+		Endpoints:     epStrs,
+		EndpointTypes: epTypes,
+		Stream:        allowStream,
+		Hostinfo:      hostinfo,
+		DebugFlags:    c.debugFlags,
+		OmitPeers:     cb == nil,
 	}
 	var extraDebugFlags []string
-	if hostinfo != nil && c.linkMon != nil && ipForwardingBroken(hostinfo.RoutableIPs, c.linkMon.InterfaceState()) {
+	if hostinfo != nil && c.linkMon != nil && !c.skipIPForwardingCheck &&
+		ipForwardingBroken(hostinfo.RoutableIPs, c.linkMon.InterfaceState()) {
 		extraDebugFlags = append(extraDebugFlags, "warn-ip-forwarding-off")
 	}
 	if health.RouterHealth() != nil {
@@ -586,11 +656,11 @@ func (c *Direct) sendMapRequest(ctx context.Context, maxPolls int, cb func(*netm
 	// TODO(bradfitz): we skip this optimization in tests, though,
 	// because the e2e tests are currently hyperspecific about the
 	// ordering of things. The e2e tests need love.
-	if len(ep) == 0 && !everEndpoints && !inTest() {
+	if len(epStrs) == 0 && !everEndpoints && !inTest() {
 		request.ReadOnly = true
 	}
 
-	bodyData, err := encode(request, &serverKey, &c.machinePrivKey)
+	bodyData, err := encode(request, &serverKey, &machinePrivKey)
 	if err != nil {
 		vlogf("netmap: encode: %v", err)
 		return err
@@ -599,7 +669,7 @@ func (c *Direct) sendMapRequest(ctx context.Context, maxPolls int, cb func(*netm
 	ctx, cancel := context.WithCancel(ctx)
 	defer cancel()
 
-	machinePubKey := tailcfg.MachineKey(c.machinePrivKey.Public())
+	machinePubKey := tailcfg.MachineKey(machinePrivKey.Public())
 	t0 := time.Now()
 	u := fmt.Sprintf("%s/machine/%s/map", serverURL, machinePubKey.HexString())
 
@@ -629,10 +699,6 @@ func (c *Direct) sendMapRequest(ctx context.Context, maxPolls int, cb func(*netm
 		return nil
 	}
 
-	// If we go more than pollTimeout without hearing from the server,
-	// end the long poll. We should be receiving a keep alive ping
-	// every minute.
-	const pollTimeout = 120 * time.Second
 	timeout := time.NewTimer(pollTimeout)
 	timeoutReset := make(chan struct{})
 	pollDone := make(chan struct{})
@@ -662,10 +728,11 @@ func (c *Direct) sendMapRequest(ctx context.Context, maxPolls int, cb func(*netm
 		}
 	}()
 
-	var lastDERPMap *tailcfg.DERPMap
-	var lastUserProfile = map[tailcfg.UserID]tailcfg.UserProfile{}
-	var lastParsedPacketFilter []filter.Match
-	var collectServices bool
+	sess := newMapSession(persist.PrivateNodeKey)
+	sess.logf = c.logf
+	sess.vlogf = vlogf
+	sess.machinePubKey = machinePubKey
+	sess.keepSharerAndUserSplit = c.keepSharerAndUserSplit
 
 	// If allowStream, then the server will use an HTTP long poll to
 	// return incremental results. There is always one response right
@@ -674,7 +741,6 @@ func (c *Direct) sendMapRequest(ctx context.Context, maxPolls int, cb func(*netm
 	// the same format before just closing the connection.
 	// We can use this same read loop either way.
 	var msg []byte
-	var previousPeers []*tailcfg.Node // for delta-purposes
 	for i := 0; i < maxPolls || maxPolls < 0; i++ {
 		vlogf("netmap: starting size read after %v (poll %v)", time.Since(t0).Round(time.Millisecond), i)
 		var siz [4]byte
@@ -692,7 +758,7 @@ func (c *Direct) sendMapRequest(ctx context.Context, maxPolls int, cb func(*netm
 		vlogf("netmap: read body after %v", time.Since(t0).Round(time.Millisecond))
 
 		var resp tailcfg.MapResponse
-		if err := c.decodeMsg(msg, &resp); err != nil {
+		if err := c.decodeMsg(msg, &resp, &machinePrivKey); err != nil {
 			vlogf("netmap: decode error: %v")
 			return err
 		}
@@ -721,16 +787,6 @@ func (c *Direct) sendMapRequest(ctx context.Context, maxPolls int, cb func(*netm
 			continue
 		}
 
-		undeltaPeers(&resp, previousPeers)
-		previousPeers = cloneNodes(resp.Peers) // defensive/lazy clone, since this escapes to who knows where
-		for _, up := range resp.UserProfiles {
-			lastUserProfile[up.ID] = up
-		}
-
-		if resp.DERPMap != nil {
-			vlogf("netmap: new map contains DERP map")
-			lastDERPMap = resp.DERPMap
-		}
 		if resp.Debug != nil {
 			if resp.Debug.LogHeapPprof {
 				go logheap.LogHeap(resp.Debug.LogHeapURL)
@@ -740,18 +796,40 @@ func (c *Direct) sendMapRequest(ctx context.Context, maxPolls int, cb func(*netm
 			}
 			setControlAtomic(&controlUseDERPRoute, resp.Debug.DERPRoute)
 			setControlAtomic(&controlTrimWGConfig, resp.Debug.TrimWGConfig)
+			if sleep := time.Duration(resp.Debug.SleepSeconds * float64(time.Second)); sleep > 0 {
+				if err := sleepAsRequested(ctx, c.logf, timeoutReset, sleep); err != nil {
+					return err
+				}
+			}
 		}
+
+		nm := sess.netmapForResponse(&resp)
+		if nm.SelfNode == nil {
+			c.logf("MapResponse lacked node")
+			return errors.New("MapResponse lacked node")
+		}
+
 		// Temporarily (2020-06-29) support removing all but
 		// discovery-supporting nodes during development, for
 		// less noise.
 		if Debug.OnlyDisco {
-			filtered := resp.Peers[:0]
-			for _, p := range resp.Peers {
-				if !p.DiscoKey.IsZero() {
-					filtered = append(filtered, p)
+			anyOld, numDisco := false, 0
+			for _, p := range nm.Peers {
+				if p.DiscoKey.IsZero() {
+					anyOld = true
+				} else {
+					numDisco++
 				}
 			}
-			resp.Peers = filtered
+			if anyOld {
+				filtered := make([]*tailcfg.Node, 0, numDisco)
+				for _, p := range nm.Peers {
+					if !p.DiscoKey.IsZero() {
+						filtered = append(filtered, p)
+					}
+				}
+				nm.Peers = filtered
+			}
 		}
 		if Debug.StripEndpoints {
 			for _, p := range resp.Peers {
@@ -761,13 +839,8 @@ func (c *Direct) sendMapRequest(ctx context.Context, maxPolls int, cb func(*netm
 				p.Endpoints = []string{"127.9.9.9:456"}
 			}
 		}
-
-		if pf := resp.PacketFilter; pf != nil {
-			lastParsedPacketFilter = c.parsePacketFilter(pf)
-		}
-
-		if v, ok := resp.CollectServices.Get(); ok {
-			collectServices = v
+		if Debug.StripCaps {
+			nm.SelfNode.Capabilities = nil
 		}
 
 		// Get latest localPort. This might've changed if
@@ -775,67 +848,9 @@ func (c *Direct) sendMapRequest(ctx context.Context, maxPolls int, cb func(*netm
 		// the end-to-end test.
 		// TODO(bradfitz): remove the NetworkMap.LocalPort field entirely.
 		c.mu.Lock()
-		localPort = c.localPort
+		nm.LocalPort = c.localPort
 		c.mu.Unlock()
 
-		nm := &netmap.NetworkMap{
-			SelfNode:        resp.Node,
-			NodeKey:         tailcfg.NodeKey(persist.PrivateNodeKey.Public()),
-			PrivateKey:      persist.PrivateNodeKey,
-			MachineKey:      machinePubKey,
-			Expiry:          resp.Node.KeyExpiry,
-			Name:            resp.Node.Name,
-			Addresses:       resp.Node.Addresses,
-			Peers:           resp.Peers,
-			LocalPort:       localPort,
-			User:            resp.Node.User,
-			UserProfiles:    make(map[tailcfg.UserID]tailcfg.UserProfile),
-			Domain:          resp.Domain,
-			DNS:             resp.DNSConfig,
-			Hostinfo:        resp.Node.Hostinfo,
-			PacketFilter:    lastParsedPacketFilter,
-			CollectServices: collectServices,
-			DERPMap:         lastDERPMap,
-			Debug:           resp.Debug,
-		}
-		addUserProfile := func(userID tailcfg.UserID) {
-			if _, dup := nm.UserProfiles[userID]; dup {
-				// Already populated it from a previous peer.
-				return
-			}
-			if up, ok := lastUserProfile[userID]; ok {
-				nm.UserProfiles[userID] = up
-			}
-		}
-		addUserProfile(nm.User)
-		magicDNSSuffix := nm.MagicDNSSuffix()
-		nm.SelfNode.InitDisplayNames(magicDNSSuffix)
-		for _, peer := range resp.Peers {
-			peer.InitDisplayNames(magicDNSSuffix)
-			if !peer.Sharer.IsZero() {
-				if c.keepSharerAndUserSplit {
-					addUserProfile(peer.Sharer)
-				} else {
-					peer.User = peer.Sharer
-				}
-			}
-			addUserProfile(peer.User)
-		}
-		if resp.Node.MachineAuthorized {
-			nm.MachineStatus = tailcfg.MachineAuthorized
-		} else {
-			nm.MachineStatus = tailcfg.MachineUnauthorized
-		}
-		if len(resp.DNS) > 0 {
-			nm.DNS.Nameservers = resp.DNS
-		}
-		if len(resp.SearchPaths) > 0 {
-			nm.DNS.Domains = resp.SearchPaths
-		}
-		if Debug.ProxyDNS {
-			nm.DNS.Proxied = true
-		}
-
 		// Printing the netmap can be extremely verbose, but is very
 		// handy for debugging. Let's limit how often we do it.
 		// Code elsewhere prints netmap diffs every time, so this
@@ -871,16 +886,19 @@ func decode(res *http.Response, v interface{}, serverKey *wgkey.Key, mkey *wgkey
 	return decodeMsg(msg, v, serverKey, mkey)
 }
 
-var debugMap, _ = strconv.ParseBool(os.Getenv("TS_DEBUG_MAP"))
+var (
+	debugMap, _      = strconv.ParseBool(os.Getenv("TS_DEBUG_MAP"))
+	debugRegister, _ = strconv.ParseBool(os.Getenv("TS_DEBUG_REGISTER"))
+)
 
 var jsonEscapedZero = []byte(`\u0000`)
 
-func (c *Direct) decodeMsg(msg []byte, v interface{}) error {
+func (c *Direct) decodeMsg(msg []byte, v interface{}, machinePrivKey *wgkey.Private) error {
 	c.mu.Lock()
 	serverKey := c.serverKey
 	c.mu.Unlock()
 
-	decrypted, err := decryptMsg(msg, &serverKey, &c.machinePrivKey)
+	decrypted, err := decryptMsg(msg, &serverKey, machinePrivKey)
 	if err != nil {
 		return err
 	}
@@ -914,8 +932,8 @@ func (c *Direct) decodeMsg(msg []byte, v interface{}) error {
 
 }
 
-func decodeMsg(msg []byte, v interface{}, serverKey *wgkey.Key, mkey *wgkey.Private) error {
-	decrypted, err := decryptMsg(msg, serverKey, mkey)
+func decodeMsg(msg []byte, v interface{}, serverKey *wgkey.Key, machinePrivKey *wgkey.Private) error {
+	decrypted, err := decryptMsg(msg, serverKey, machinePrivKey)
 	if err != nil {
 		return err
 	}
@@ -998,6 +1016,7 @@ type debug struct {
 	OnlyDisco      bool
 	Disco          bool
 	StripEndpoints bool // strip endpoints from control (only use disco messages)
+	StripCaps      bool // strip all local node's control-provided capabilities
 }
 
 func initDebug() debug {
@@ -1006,6 +1025,7 @@ func initDebug() debug {
 		NetMap:         envBool("TS_DEBUG_NETMAP"),
 		ProxyDNS:       envBool("TS_DEBUG_PROXY_DNS"),
 		StripEndpoints: envBool("TS_DEBUG_STRIP_ENDPOINTS"),
+		StripCaps:      envBool("TS_DEBUG_STRIP_CAPS"),
 		OnlyDisco:      use == "only",
 		Disco:          use == "only" || use == "" || envBool("TS_DEBUG_USE_DISCO"),
 	}
@@ -1023,117 +1043,7 @@ func envBool(k string) bool {
 	return v
 }
 
-// undeltaPeers updates mapRes.Peers to be complete based on the provided previous peer list
-// and the PeersRemoved and PeersChanged fields in mapRes.
-// It then also nils out the delta fields.
-func undeltaPeers(mapRes *tailcfg.MapResponse, prev []*tailcfg.Node) {
-	if len(mapRes.Peers) > 0 {
-		// Not delta encoded.
-		if !nodesSorted(mapRes.Peers) {
-			log.Printf("netmap: undeltaPeers: MapResponse.Peers not sorted; sorting")
-			sortNodes(mapRes.Peers)
-		}
-		return
-	}
-
-	var removed map[tailcfg.NodeID]bool
-	if pr := mapRes.PeersRemoved; len(pr) > 0 {
-		removed = make(map[tailcfg.NodeID]bool, len(pr))
-		for _, id := range pr {
-			removed[id] = true
-		}
-	}
-	changed := mapRes.PeersChanged
-
-	if len(removed) == 0 && len(changed) == 0 {
-		// No changes fast path.
-		mapRes.Peers = prev
-		return
-	}
-
-	if !nodesSorted(changed) {
-		log.Printf("netmap: undeltaPeers: MapResponse.PeersChanged not sorted; sorting")
-		sortNodes(changed)
-	}
-	if !nodesSorted(prev) {
-		// Internal error (unrelated to the network) if we get here.
-		log.Printf("netmap: undeltaPeers: [unexpected] prev not sorted; sorting")
-		sortNodes(prev)
-	}
-
-	newFull := make([]*tailcfg.Node, 0, len(prev)-len(removed))
-	for len(prev) > 0 && len(changed) > 0 {
-		pID := prev[0].ID
-		cID := changed[0].ID
-		if removed[pID] {
-			prev = prev[1:]
-			continue
-		}
-		switch {
-		case pID < cID:
-			newFull = append(newFull, prev[0])
-			prev = prev[1:]
-		case pID == cID:
-			newFull = append(newFull, changed[0])
-			prev, changed = prev[1:], changed[1:]
-		case cID < pID:
-			newFull = append(newFull, changed[0])
-			changed = changed[1:]
-		}
-	}
-	newFull = append(newFull, changed...)
-	for _, n := range prev {
-		if !removed[n.ID] {
-			newFull = append(newFull, n)
-		}
-	}
-	sortNodes(newFull)
-
-	if mapRes.PeerSeenChange != nil {
-		peerByID := make(map[tailcfg.NodeID]*tailcfg.Node, len(newFull))
-		for _, n := range newFull {
-			peerByID[n.ID] = n
-		}
-		now := time.Now()
-		for nodeID, seen := range mapRes.PeerSeenChange {
-			if n, ok := peerByID[nodeID]; ok {
-				if seen {
-					n.LastSeen = &now
-				} else {
-					n.LastSeen = nil
-				}
-			}
-		}
-	}
-
-	mapRes.Peers = newFull
-	mapRes.PeersChanged = nil
-	mapRes.PeersRemoved = nil
-}
-
-func nodesSorted(v []*tailcfg.Node) bool {
-	for i, n := range v {
-		if i > 0 && n.ID <= v[i-1].ID {
-			return false
-		}
-	}
-	return true
-}
-
-func sortNodes(v []*tailcfg.Node) {
-	sort.Slice(v, func(i, j int) bool { return v[i].ID < v[j].ID })
-}
-
-func cloneNodes(v1 []*tailcfg.Node) []*tailcfg.Node {
-	if v1 == nil {
-		return nil
-	}
-	v2 := make([]*tailcfg.Node, len(v1))
-	for i, n := range v1 {
-		v2[i] = n.Clone()
-	}
-	return v2
-}
+var clockNow = time.Now
 
 // opt.Bool configs from control.
 var (
@@ -1166,6 +1076,11 @@ func TrimWGConfig() opt.Bool {
 // and will definitely not work for the routes provided.
 //
 // It should not return false positives.
+//
+// TODO(bradfitz): merge this code into LocalBackend.CheckIPForwarding
+// and change controlclient.Options.SkipIPForwardingCheck into a
+// func([]netaddr.IPPrefix) error signature instead. Then we only have
+// one copy of this code.
 func ipForwardingBroken(routes []netaddr.IPPrefix, state *interfaces.State) bool {
 	if len(routes) == 0 {
 		// Nothing to route, so no need to warn.
@@ -1272,3 +1187,34 @@ func answerPing(logf logger.Logf, c *http.Client, pr *tailcfg.PingRequest) {
 		logf("answerPing complete to %v (after %v)", pr.URL, d)
 	}
 }
+
+func sleepAsRequested(ctx context.Context, logf logger.Logf, timeoutReset chan<- struct{}, d time.Duration) error {
+	const maxSleep = 5 * time.Minute
+	if d > maxSleep {
+		logf("sleeping for %v, capped from server-requested %v ...", maxSleep, d)
+		d = maxSleep
+	} else {
+		logf("sleeping for server-requested %v ...", d)
+	}
+
+	ticker := time.NewTicker(pollTimeout / 2)
+	defer ticker.Stop()
+	timer := time.NewTimer(d)
+	defer timer.Stop()
+	for {
+		select {
+		case <-ctx.Done():
+			return ctx.Err()
+		case <-timer.C:
+			return nil
+		case <-ticker.C:
+			select {
+			case timeoutReset <- struct{}{}:
+			case <-timer.C:
+				return nil
+			case <-ctx.Done():
+				return ctx.Err()
+			}
+		}
+	}
+}

control/controlclient/direct_test.go

@@ -6,94 +6,13 @@ package controlclient
 
 import (
 	"encoding/json"
-	"fmt"
-	"reflect"
-	"strings"
 	"testing"
 
+	"inet.af/netaddr"
 	"tailscale.com/tailcfg"
 	"tailscale.com/types/wgkey"
 )
 
-func TestUndeltaPeers(t *testing.T) {
-	n := func(id tailcfg.NodeID, name string) *tailcfg.Node {
-		return &tailcfg.Node{ID: id, Name: name}
-	}
-	peers := func(nv ...*tailcfg.Node) []*tailcfg.Node { return nv }
-	tests := []struct {
-		name   string
-		mapRes *tailcfg.MapResponse
-		prev   []*tailcfg.Node
-		want   []*tailcfg.Node
-	}{
-		{
-			name: "full_peers",
-			mapRes: &tailcfg.MapResponse{
-				Peers: peers(n(1, "foo"), n(2, "bar")),
-			},
-			want: peers(n(1, "foo"), n(2, "bar")),
-		},
-		{
-			name: "full_peers_ignores_deltas",
-			mapRes: &tailcfg.MapResponse{
-				Peers:        peers(n(1, "foo"), n(2, "bar")),
-				PeersRemoved: []tailcfg.NodeID{2},
-			},
-			want: peers(n(1, "foo"), n(2, "bar")),
-		},
-		{
-			name: "add_and_update",
-			prev: peers(n(1, "foo"), n(2, "bar")),
-			mapRes: &tailcfg.MapResponse{
-				PeersChanged: peers(n(0, "zero"), n(2, "bar2"), n(3, "three")),
-			},
-			want: peers(n(0, "zero"), n(1, "foo"), n(2, "bar2"), n(3, "three")),
-		},
-		{
-			name: "remove",
-			prev: peers(n(1, "foo"), n(2, "bar")),
-			mapRes: &tailcfg.MapResponse{
-				PeersRemoved: []tailcfg.NodeID{1},
-			},
-			want: peers(n(2, "bar")),
-		},
-		{
-			name: "add_and_remove",
-			prev: peers(n(1, "foo"), n(2, "bar")),
-			mapRes: &tailcfg.MapResponse{
-				PeersChanged: peers(n(1, "foo2")),
-				PeersRemoved: []tailcfg.NodeID{2},
-			},
-			want: peers(n(1, "foo2")),
-		},
-		{
-			name:   "unchanged",
-			prev:   peers(n(1, "foo"), n(2, "bar")),
-			mapRes: &tailcfg.MapResponse{},
-			want:   peers(n(1, "foo"), n(2, "bar")),
-		},
-	}
-	for _, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
-			undeltaPeers(tt.mapRes, tt.prev)
-			if !reflect.DeepEqual(tt.mapRes.Peers, tt.want) {
-				t.Errorf("wrong results\n got: %s\nwant: %s", formatNodes(tt.mapRes.Peers), formatNodes(tt.want))
-			}
-		})
-	}
-}
-
-func formatNodes(nodes []*tailcfg.Node) string {
-	var sb strings.Builder
-	for i, n := range nodes {
-		if i > 0 {
-			sb.WriteString(", ")
-		}
-		fmt.Fprintf(&sb, "(%d, %q)", n.ID, n.Name)
-	}
-	return sb.String()
-}
-
 func TestNewDirect(t *testing.T) {
 	hi := NewHostinfo()
 	ni := tailcfg.NetInfo{LinkType: "wired"}
@@ -103,7 +22,13 @@ func TestNewDirect(t *testing.T) {
 	if err != nil {
 		t.Error(err)
 	}
-	opts := Options{ServerURL: "https://example.com", MachinePrivateKey: key, Hostinfo: hi}
+	opts := Options{
+		ServerURL: "https://example.com",
+		Hostinfo:  hi,
+		GetMachinePrivateKey: func() (wgkey.Private, error) {
+			return key, nil
+		},
+	}
 	c, err := NewDirect(opts)
 	if err != nil {
 		t.Fatal(err)
@@ -138,7 +63,7 @@ func TestNewDirect(t *testing.T) {
 		t.Errorf("c.SetHostinfo(hi) want true got %v", changed)
 	}
 
-	endpoints := []string{"1", "2", "3"}
+	endpoints := fakeEndpoints(1, 2, 3)
 	changed = c.newEndpoints(12, endpoints)
 	if !changed {
 		t.Errorf("c.newEndpoints(12) want true got %v", changed)
@@ -151,13 +76,22 @@ func TestNewDirect(t *testing.T) {
 	if !changed {
 		t.Errorf("c.newEndpoints(13) want true got %v", changed)
 	}
-	endpoints = []string{"4", "5", "6"}
+	endpoints = fakeEndpoints(4, 5, 6)
 	changed = c.newEndpoints(13, endpoints)
 	if !changed {
 		t.Errorf("c.newEndpoints(13) want true got %v", changed)
 	}
 }
 
+func fakeEndpoints(ports ...uint16) (ret []tailcfg.Endpoint) {
+	for _, port := range ports {
+		ret = append(ret, tailcfg.Endpoint{
+			Addr: netaddr.IPPort{Port: port},
+		})
+	}
+	return
+}
+
 func TestNewHostinfo(t *testing.T) {
 	hi := NewHostinfo()
 	if hi == nil {

control/controlclient/filter.go

@@ -1,20 +0,0 @@
-// Copyright (c) 2020 Tailscale Inc & AUTHORS All rights reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-
-package controlclient
-
-import (
-	"tailscale.com/tailcfg"
-	"tailscale.com/wgengine/filter"
-)
-
-// Parse a backward-compatible FilterRule used by control's wire
-// format, producing the most current filter format.
-func (c *Direct) parsePacketFilter(pf []tailcfg.FilterRule) []filter.Match {
-	mm, err := filter.MatchesFromFilterRules(pf)
-	if err != nil {
-		c.logf("parsePacketFilter: %s\n", err)
-	}
-	return mm
-}

control/controlclient/hostinfo_linux.go

@@ -11,6 +11,7 @@ import (
 	"fmt"
 	"io"
 	"io/ioutil"
+	"os"
 	"strings"
 	"syscall"
 
@@ -58,6 +59,9 @@ func osVersionLinux() string {
 	if inContainer() {
 		attrBuf.WriteString("; container")
 	}
+	if inKnative() {
+		attrBuf.WriteString("; env=kn")
+	}
 	attr := attrBuf.String()
 
 	id := m["ID"]
@@ -99,5 +103,21 @@ func inContainer() (ret bool) {
 		}
 		return nil
 	})
+	lineread.File("/proc/mounts", func(line []byte) error {
+		if mem.Contains(mem.B(line), mem.S("fuse.lxcfs")) {
+			ret = true
+			return io.EOF
+		}
+		return nil
+	})
 	return
 }
+
+func inKnative() bool {
+	// https://cloud.google.com/run/docs/reference/container-contract#env-vars
+	if os.Getenv("K_REVISION") != "" && os.Getenv("K_CONFIGURATION") != "" &&
+		os.Getenv("K_SERVICE") != "" && os.Getenv("PORT") != "" {
+		return true
+	}
+	return false
+}

control/controlclient/map.go

@@ -0,0 +1,282 @@
+// Copyright (c) 2021 Tailscale Inc & AUTHORS All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package controlclient
+
+import (
+	"log"
+	"sort"
+
+	"tailscale.com/tailcfg"
+	"tailscale.com/types/logger"
+	"tailscale.com/types/netmap"
+	"tailscale.com/types/wgkey"
+	"tailscale.com/wgengine/filter"
+)
+
+// mapSession holds the state over a long-polled "map" request to the
+// control plane.
+//
+// It accepts incremental tailcfg.MapResponse values to
+// netMapForResponse and returns fully inflated NetworkMaps, filling
+// in the omitted data implicit from prior MapResponse values from
+// within the same session (the same long-poll HTTP response to the
+// one MapRequest).
+type mapSession struct {
+	// Immutable fields.
+	privateNodeKey         wgkey.Private
+	logf                   logger.Logf
+	vlogf                  logger.Logf
+	machinePubKey          tailcfg.MachineKey
+	keepSharerAndUserSplit bool // see Options.KeepSharerAndUserSplit
+
+	// Fields storing state over the the coards of multiple MapResponses.
+	lastNode               *tailcfg.Node
+	lastDNSConfig          *tailcfg.DNSConfig
+	lastDERPMap            *tailcfg.DERPMap
+	lastUserProfile        map[tailcfg.UserID]tailcfg.UserProfile
+	lastParsedPacketFilter []filter.Match
+	collectServices        bool
+	previousPeers          []*tailcfg.Node // for delta-purposes
+	lastDomain             string
+
+	// netMapBuilding is non-nil during a netmapForResponse call,
+	// containing the value to be returned, once fully populated.
+	netMapBuilding *netmap.NetworkMap
+}
+
+func newMapSession(privateNodeKey wgkey.Private) *mapSession {
+	ms := &mapSession{
+		privateNodeKey:  privateNodeKey,
+		logf:            logger.Discard,
+		vlogf:           logger.Discard,
+		lastDNSConfig:   new(tailcfg.DNSConfig),
+		lastUserProfile: map[tailcfg.UserID]tailcfg.UserProfile{},
+	}
+	return ms
+}
+
+func (ms *mapSession) addUserProfile(userID tailcfg.UserID) {
+	nm := ms.netMapBuilding
+	if _, dup := nm.UserProfiles[userID]; dup {
+		// Already populated it from a previous peer.
+		return
+	}
+	if up, ok := ms.lastUserProfile[userID]; ok {
+		nm.UserProfiles[userID] = up
+	}
+}
+
+// netmapForResponse returns a fully populated NetworkMap from a full
+// or incremental MapResponse within the session, filling in omitted
+// information from prior MapResponse values.
+func (ms *mapSession) netmapForResponse(resp *tailcfg.MapResponse) *netmap.NetworkMap {
+	undeltaPeers(resp, ms.previousPeers)
+
+	ms.previousPeers = cloneNodes(resp.Peers) // defensive/lazy clone, since this escapes to who knows where
+	for _, up := range resp.UserProfiles {
+		ms.lastUserProfile[up.ID] = up
+	}
+
+	if resp.DERPMap != nil {
+		ms.vlogf("netmap: new map contains DERP map")
+		ms.lastDERPMap = resp.DERPMap
+	}
+
+	if pf := resp.PacketFilter; pf != nil {
+		var err error
+		ms.lastParsedPacketFilter, err = filter.MatchesFromFilterRules(pf)
+		if err != nil {
+			ms.logf("parsePacketFilter: %v", err)
+		}
+	}
+	if c := resp.DNSConfig; c != nil {
+		ms.lastDNSConfig = c
+	}
+
+	if v, ok := resp.CollectServices.Get(); ok {
+		ms.collectServices = v
+	}
+	if resp.Domain != "" {
+		ms.lastDomain = resp.Domain
+	}
+
+	nm := &netmap.NetworkMap{
+		NodeKey:         tailcfg.NodeKey(ms.privateNodeKey.Public()),
+		PrivateKey:      ms.privateNodeKey,
+		MachineKey:      ms.machinePubKey,
+		Peers:           resp.Peers,
+		UserProfiles:    make(map[tailcfg.UserID]tailcfg.UserProfile),
+		Domain:          ms.lastDomain,
+		DNS:             *ms.lastDNSConfig,
+		PacketFilter:    ms.lastParsedPacketFilter,
+		CollectServices: ms.collectServices,
+		DERPMap:         ms.lastDERPMap,
+		Debug:           resp.Debug,
+	}
+	ms.netMapBuilding = nm
+
+	if resp.Node != nil {
+		ms.lastNode = resp.Node
+	}
+	if node := ms.lastNode.Clone(); node != nil {
+		nm.SelfNode = node
+		nm.Expiry = node.KeyExpiry
+		nm.Name = node.Name
+		nm.Addresses = node.Addresses
+		nm.User = node.User
+		nm.Hostinfo = node.Hostinfo
+		if node.MachineAuthorized {
+			nm.MachineStatus = tailcfg.MachineAuthorized
+		} else {
+			nm.MachineStatus = tailcfg.MachineUnauthorized
+		}
+	}
+
+	ms.addUserProfile(nm.User)
+	magicDNSSuffix := nm.MagicDNSSuffix()
+	if nm.SelfNode != nil {
+		nm.SelfNode.InitDisplayNames(magicDNSSuffix)
+	}
+	for _, peer := range resp.Peers {
+		peer.InitDisplayNames(magicDNSSuffix)
+		if !peer.Sharer.IsZero() {
+			if ms.keepSharerAndUserSplit {
+				ms.addUserProfile(peer.Sharer)
+			} else {
+				peer.User = peer.Sharer
+			}
+		}
+		ms.addUserProfile(peer.User)
+	}
+	if len(resp.DNS) > 0 {
+		nm.DNS.Nameservers = resp.DNS
+	}
+	if len(resp.SearchPaths) > 0 {
+		nm.DNS.Domains = resp.SearchPaths
+	}
+	if Debug.ProxyDNS {
+		nm.DNS.Proxied = true
+	}
+	ms.netMapBuilding = nil
+	return nm
+}
+
+// undeltaPeers updates mapRes.Peers to be complete based on the
+// provided previous peer list and the PeersRemoved and PeersChanged
+// fields in mapRes, as well as the PeerSeenChange and OnlineChange
+// maps.
+//
+// It then also nils out the delta fields.
+func undeltaPeers(mapRes *tailcfg.MapResponse, prev []*tailcfg.Node) {
+	if len(mapRes.Peers) > 0 {
+		// Not delta encoded.
+		if !nodesSorted(mapRes.Peers) {
+			log.Printf("netmap: undeltaPeers: MapResponse.Peers not sorted; sorting")
+			sortNodes(mapRes.Peers)
+		}
+		return
+	}
+
+	var removed map[tailcfg.NodeID]bool
+	if pr := mapRes.PeersRemoved; len(pr) > 0 {
+		removed = make(map[tailcfg.NodeID]bool, len(pr))
+		for _, id := range pr {
+			removed[id] = true
+		}
+	}
+	changed := mapRes.PeersChanged
+
+	if !nodesSorted(changed) {
+		log.Printf("netmap: undeltaPeers: MapResponse.PeersChanged not sorted; sorting")
+		sortNodes(changed)
+	}
+	if !nodesSorted(prev) {
+		// Internal error (unrelated to the network) if we get here.
+		log.Printf("netmap: undeltaPeers: [unexpected] prev not sorted; sorting")
+		sortNodes(prev)
+	}
+
+	newFull := prev
+	if len(removed) > 0 || len(changed) > 0 {
+		newFull = make([]*tailcfg.Node, 0, len(prev)-len(removed))
+		for len(prev) > 0 && len(changed) > 0 {
+			pID := prev[0].ID
+			cID := changed[0].ID
+			if removed[pID] {
+				prev = prev[1:]
+				continue
+			}
+			switch {
+			case pID < cID:
+				newFull = append(newFull, prev[0])
+				prev = prev[1:]
+			case pID == cID:
+				newFull = append(newFull, changed[0])
+				prev, changed = prev[1:], changed[1:]
+			case cID < pID:
+				newFull = append(newFull, changed[0])
+				changed = changed[1:]
+			}
+		}
+		newFull = append(newFull, changed...)
+		for _, n := range prev {
+			if !removed[n.ID] {
+				newFull = append(newFull, n)
+			}
+		}
+		sortNodes(newFull)
+	}
+
+	if len(mapRes.PeerSeenChange) != 0 || len(mapRes.OnlineChange) != 0 {
+		peerByID := make(map[tailcfg.NodeID]*tailcfg.Node, len(newFull))
+		for _, n := range newFull {
+			peerByID[n.ID] = n
+		}
+		now := clockNow()
+		for nodeID, seen := range mapRes.PeerSeenChange {
+			if n, ok := peerByID[nodeID]; ok {
+				if seen {
+					n.LastSeen = &now
+				} else {
+					n.LastSeen = nil
+				}
+			}
+		}
+		for nodeID, online := range mapRes.OnlineChange {
+			if n, ok := peerByID[nodeID]; ok {
+				online := online
+				n.Online = &online
+			}
+		}
+	}
+
+	mapRes.Peers = newFull
+	mapRes.PeersChanged = nil
+	mapRes.PeersRemoved = nil
+}
+
+func nodesSorted(v []*tailcfg.Node) bool {
+	for i, n := range v {
+		if i > 0 && n.ID <= v[i-1].ID {
+			return false
+		}
+	}
+	return true
+}
+
+func sortNodes(v []*tailcfg.Node) {
+	sort.Slice(v, func(i, j int) bool { return v[i].ID < v[j].ID })
+}
+
+func cloneNodes(v1 []*tailcfg.Node) []*tailcfg.Node {
+	if v1 == nil {
+		return nil
+	}
+	v2 := make([]*tailcfg.Node, len(v1))
+	for i, n := range v1 {
+		v2[i] = n.Clone()
+	}
+	return v2
+}

control/controlclient/map_test.go

@@ -0,0 +1,311 @@
+// Copyright (c) 2021 Tailscale Inc & AUTHORS All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package controlclient
+
+import (
+	"encoding/json"
+	"fmt"
+	"reflect"
+	"strings"
+	"testing"
+	"time"
+
+	"tailscale.com/tailcfg"
+	"tailscale.com/types/netmap"
+	"tailscale.com/types/wgkey"
+)
+
+func TestUndeltaPeers(t *testing.T) {
+	defer func(old func() time.Time) { clockNow = old }(clockNow)
+
+	var curTime time.Time
+	clockNow = func() time.Time {
+		return curTime
+	}
+	online := func(v bool) func(*tailcfg.Node) {
+		return func(n *tailcfg.Node) {
+			n.Online = &v
+		}
+	}
+	seenAt := func(t time.Time) func(*tailcfg.Node) {
+		return func(n *tailcfg.Node) {
+			n.LastSeen = &t
+		}
+	}
+	n := func(id tailcfg.NodeID, name string, mod ...func(*tailcfg.Node)) *tailcfg.Node {
+		n := &tailcfg.Node{ID: id, Name: name}
+		for _, f := range mod {
+			f(n)
+		}
+		return n
+	}
+	peers := func(nv ...*tailcfg.Node) []*tailcfg.Node { return nv }
+	tests := []struct {
+		name    string
+		mapRes  *tailcfg.MapResponse
+		curTime time.Time
+		prev    []*tailcfg.Node
+		want    []*tailcfg.Node
+	}{
+		{
+			name: "full_peers",
+			mapRes: &tailcfg.MapResponse{
+				Peers: peers(n(1, "foo"), n(2, "bar")),
+			},
+			want: peers(n(1, "foo"), n(2, "bar")),
+		},
+		{
+			name: "full_peers_ignores_deltas",
+			mapRes: &tailcfg.MapResponse{
+				Peers:        peers(n(1, "foo"), n(2, "bar")),
+				PeersRemoved: []tailcfg.NodeID{2},
+			},
+			want: peers(n(1, "foo"), n(2, "bar")),
+		},
+		{
+			name: "add_and_update",
+			prev: peers(n(1, "foo"), n(2, "bar")),
+			mapRes: &tailcfg.MapResponse{
+				PeersChanged: peers(n(0, "zero"), n(2, "bar2"), n(3, "three")),
+			},
+			want: peers(n(0, "zero"), n(1, "foo"), n(2, "bar2"), n(3, "three")),
+		},
+		{
+			name: "remove",
+			prev: peers(n(1, "foo"), n(2, "bar")),
+			mapRes: &tailcfg.MapResponse{
+				PeersRemoved: []tailcfg.NodeID{1},
+			},
+			want: peers(n(2, "bar")),
+		},
+		{
+			name: "add_and_remove",
+			prev: peers(n(1, "foo"), n(2, "bar")),
+			mapRes: &tailcfg.MapResponse{
+				PeersChanged: peers(n(1, "foo2")),
+				PeersRemoved: []tailcfg.NodeID{2},
+			},
+			want: peers(n(1, "foo2")),
+		},
+		{
+			name:   "unchanged",
+			prev:   peers(n(1, "foo"), n(2, "bar")),
+			mapRes: &tailcfg.MapResponse{},
+			want:   peers(n(1, "foo"), n(2, "bar")),
+		},
+		{
+			name: "online_change",
+			prev: peers(n(1, "foo"), n(2, "bar")),
+			mapRes: &tailcfg.MapResponse{
+				OnlineChange: map[tailcfg.NodeID]bool{
+					1: true,
+				},
+			},
+			want: peers(
+				n(1, "foo", online(true)),
+				n(2, "bar"),
+			),
+		},
+		{
+			name: "online_change_offline",
+			prev: peers(n(1, "foo"), n(2, "bar")),
+			mapRes: &tailcfg.MapResponse{
+				OnlineChange: map[tailcfg.NodeID]bool{
+					1: false,
+					2: true,
+				},
+			},
+			want: peers(
+				n(1, "foo", online(false)),
+				n(2, "bar", online(true)),
+			),
+		},
+		{
+			name:    "peer_seen_at",
+			prev:    peers(n(1, "foo", seenAt(time.Unix(111, 0))), n(2, "bar")),
+			curTime: time.Unix(123, 0),
+			mapRes: &tailcfg.MapResponse{
+				PeerSeenChange: map[tailcfg.NodeID]bool{
+					1: false,
+					2: true,
+				},
+			},
+			want: peers(
+				n(1, "foo"),
+				n(2, "bar", seenAt(time.Unix(123, 0))),
+			),
+		},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			if !tt.curTime.IsZero() {
+				curTime = tt.curTime
+			}
+			undeltaPeers(tt.mapRes, tt.prev)
+			if !reflect.DeepEqual(tt.mapRes.Peers, tt.want) {
+				t.Errorf("wrong results\n got: %s\nwant: %s", formatNodes(tt.mapRes.Peers), formatNodes(tt.want))
+			}
+		})
+	}
+}
+
+func formatNodes(nodes []*tailcfg.Node) string {
+	var sb strings.Builder
+	for i, n := range nodes {
+		if i > 0 {
+			sb.WriteString(", ")
+		}
+		var extra string
+		if n.Online != nil {
+			extra += fmt.Sprintf(", online=%v", *n.Online)
+		}
+		if n.LastSeen != nil {
+			extra += fmt.Sprintf(", lastSeen=%v", n.LastSeen.Unix())
+		}
+		fmt.Fprintf(&sb, "(%d, %q%s)", n.ID, n.Name, extra)
+	}
+	return sb.String()
+}
+
+func newTestMapSession(t *testing.T) *mapSession {
+	k, err := wgkey.NewPrivate()
+	if err != nil {
+		t.Fatal(err)
+	}
+	return newMapSession(k)
+}
+
+func TestNetmapForResponse(t *testing.T) {
+	t.Run("implicit_packetfilter", func(t *testing.T) {
+		somePacketFilter := []tailcfg.FilterRule{
+			{
+				SrcIPs: []string{"*"},
+				DstPorts: []tailcfg.NetPortRange{
+					{IP: "10.2.3.4", Ports: tailcfg.PortRange{First: 22, Last: 22}},
+				},
+			},
+		}
+		ms := newTestMapSession(t)
+		nm1 := ms.netmapForResponse(&tailcfg.MapResponse{
+			Node:         new(tailcfg.Node),
+			PacketFilter: somePacketFilter,
+		})
+		if len(nm1.PacketFilter) == 0 {
+			t.Fatalf("zero length PacketFilter")
+		}
+		nm2 := ms.netmapForResponse(&tailcfg.MapResponse{
+			Node:         new(tailcfg.Node),
+			PacketFilter: nil, // testing that the server can omit this.
+		})
+		if len(nm1.PacketFilter) == 0 {
+			t.Fatalf("zero length PacketFilter in 2nd netmap")
+		}
+		if !reflect.DeepEqual(nm1.PacketFilter, nm2.PacketFilter) {
+			t.Error("packet filters differ")
+		}
+	})
+	t.Run("implicit_dnsconfig", func(t *testing.T) {
+		someDNSConfig := &tailcfg.DNSConfig{Domains: []string{"foo", "bar"}}
+		ms := newTestMapSession(t)
+		nm1 := ms.netmapForResponse(&tailcfg.MapResponse{
+			Node:      new(tailcfg.Node),
+			DNSConfig: someDNSConfig,
+		})
+		if !reflect.DeepEqual(nm1.DNS, *someDNSConfig) {
+			t.Fatalf("1st DNS wrong")
+		}
+		nm2 := ms.netmapForResponse(&tailcfg.MapResponse{
+			Node:      new(tailcfg.Node),
+			DNSConfig: nil, // implict
+		})
+		if !reflect.DeepEqual(nm2.DNS, *someDNSConfig) {
+			t.Fatalf("2nd DNS wrong")
+		}
+	})
+	t.Run("collect_services", func(t *testing.T) {
+		ms := newTestMapSession(t)
+		var nm *netmap.NetworkMap
+		wantCollect := func(v bool) {
+			t.Helper()
+			if nm.CollectServices != v {
+				t.Errorf("netmap.CollectServices = %v; want %v", nm.CollectServices, v)
+			}
+		}
+
+		nm = ms.netmapForResponse(&tailcfg.MapResponse{
+			Node: new(tailcfg.Node),
+		})
+		wantCollect(false)
+
+		nm = ms.netmapForResponse(&tailcfg.MapResponse{
+			Node:            new(tailcfg.Node),
+			CollectServices: "false",
+		})
+		wantCollect(false)
+
+		nm = ms.netmapForResponse(&tailcfg.MapResponse{
+			Node:            new(tailcfg.Node),
+			CollectServices: "true",
+		})
+		wantCollect(true)
+
+		nm = ms.netmapForResponse(&tailcfg.MapResponse{
+			Node:            new(tailcfg.Node),
+			CollectServices: "",
+		})
+		wantCollect(true)
+	})
+	t.Run("implicit_domain", func(t *testing.T) {
+		ms := newTestMapSession(t)
+		var nm *netmap.NetworkMap
+		want := func(v string) {
+			t.Helper()
+			if nm.Domain != v {
+				t.Errorf("netmap.Domain = %q; want %q", nm.Domain, v)
+			}
+		}
+		nm = ms.netmapForResponse(&tailcfg.MapResponse{
+			Node:   new(tailcfg.Node),
+			Domain: "foo.com",
+		})
+		want("foo.com")
+
+		nm = ms.netmapForResponse(&tailcfg.MapResponse{
+			Node: new(tailcfg.Node),
+		})
+		want("foo.com")
+	})
+	t.Run("implicit_node", func(t *testing.T) {
+		someNode := &tailcfg.Node{
+			Name: "foo",
+		}
+		wantNode := &tailcfg.Node{
+			Name:                 "foo",
+			ComputedName:         "foo",
+			ComputedNameWithHost: "foo",
+		}
+		ms := newTestMapSession(t)
+
+		nm1 := ms.netmapForResponse(&tailcfg.MapResponse{
+			Node: someNode,
+		})
+		if nm1.SelfNode == nil {
+			t.Fatal("nil Node in 1st netmap")
+		}
+		if !reflect.DeepEqual(nm1.SelfNode, wantNode) {
+			j, _ := json.Marshal(nm1.SelfNode)
+			t.Errorf("Node mismatch in 1st netmap; got: %s", j)
+		}
+
+		nm2 := ms.netmapForResponse(&tailcfg.MapResponse{})
+		if nm2.SelfNode == nil {
+			t.Fatal("nil Node in 1st netmap")
+		}
+		if !reflect.DeepEqual(nm2.SelfNode, wantNode) {
+			j, _ := json.Marshal(nm2.SelfNode)
+			t.Errorf("Node mismatch in 2nd netmap; got: %s", j)
+		}
+	})
+}

control/controlclient/sign_supported.go

@@ -51,6 +51,46 @@ var (
 	errBadRequest = errors.New("malformed request")
 )
 
+func isSupportedCertificate(cert *x509.Certificate) bool {
+	return cert.PublicKeyAlgorithm == x509.RSA
+}
+
+func isSubjectInChain(subject string, chain []*x509.Certificate) bool {
+	if len(chain) == 0 || chain[0] == nil {
+		return false
+	}
+
+	for _, c := range chain {
+		if c == nil {
+			continue
+		}
+		if c.Subject.String() == subject {
+			return true
+		}
+	}
+
+	return false
+}
+
+func selectIdentityFromSlice(subject string, ids []certstore.Identity) (certstore.Identity, []*x509.Certificate) {
+	for _, id := range ids {
+		chain, err := id.CertificateChain()
+		if err != nil {
+			continue
+		}
+
+		if !isSupportedCertificate(chain[0]) {
+			continue
+		}
+
+		if isSubjectInChain(subject, chain) {
+			return id, chain
+		}
+	}
+
+	return nil, nil
+}
+
 // findIdentity locates an identity from the Windows or Darwin certificate
 // store. It returns the first certificate with a matching Subject anywhere in
 // its certificate chain, so it is possible to search for the leaf certificate,
@@ -64,26 +104,7 @@ func findIdentity(subject string, st certstore.Store) (certstore.Identity, []*x5
 		return nil, nil, err
 	}
 
-	var selected certstore.Identity
-	var chain []*x509.Certificate
-
-	for _, id := range ids {
-		chain, err = id.CertificateChain()
-		if err != nil {
-			continue
-		}
-
-		if chain[0].PublicKeyAlgorithm != x509.RSA {
-			continue
-		}
-
-		for _, c := range chain {
-			if c.Subject.String() == subject {
-				selected = id
-				break
-			}
-		}
-	}
+	selected, chain := selectIdentityFromSlice(subject, ids)
 
 	for _, id := range ids {
 		if id != selected {

derp/derp_test.go

@@ -409,7 +409,7 @@ func TestSendFreeze(t *testing.T) {
 	for i := 0; i < cap(errCh); i++ {
 		err := <-errCh
 		if err != nil {
-			if errors.Is(err, io.EOF) {
+			if errors.Is(err, io.EOF) || errors.Is(err, net.ErrClosed) {
 				continue
 			}
 			t.Error(err)

derp/derpmap/derpmap.go

@@ -83,6 +83,9 @@ func Prod() *tailcfg.DERPMap {
 			10: derpRegion(10, "sea", "Seattle",
 				derpNode("a", "137.220.36.168", "2001:19f0:8001:2d9:5400:2ff:feef:bbb1"),
 			),
+			11: derpRegion(11, "sao", "São Paulo",
+				derpNode("a", "18.230.97.74", "2600:1f1e:ee4:5611:ec5c:1736:d43b:a454"),
+			),
 		},
 	}
 }

disco/disco_fuzzer.go

@@ -0,0 +1,18 @@
+// Copyright (c) 2021 Tailscale Inc & AUTHORS All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+// +build gofuzz
+
+package disco
+
+func Fuzz(data []byte) int {
+	m, _ := Parse(data)
+
+	newBytes := m.AppendMarshal(data)
+	parsedMarshall, _ := Parse(newBytes)
+
+	if m != parsedMarshall {
+		panic("Parsing error")
+	}
+	return 1
+}

go.mod

@@ -24,14 +24,14 @@ require (
 	github.com/peterbourgon/ff/v2 v2.0.0
 	github.com/pkg/errors v0.9.1 // indirect
 	github.com/tailscale/depaware v0.0.0-20201214215404-77d1e9757027
-	github.com/tailscale/wireguard-go v0.0.0-20210327173134-f6a42a1646a0
+	github.com/tailscale/wireguard-go v0.0.0-20210419202603-b32acd8f0292
 	github.com/tcnksm/go-httpstat v0.2.0
 	github.com/toqueteos/webbrowser v1.2.0
 	go4.org/mem v0.0.0-20201119185036-c04c5a6ff174
 	golang.org/x/crypto v0.0.0-20210317152858-513c2a44f670
 	golang.org/x/net v0.0.0-20210226172049-e18ecbb05110
 	golang.org/x/sync v0.0.0-20210220032951-036812b2e83c
-	golang.org/x/sys v0.0.0-20210317225723-c4fcb01b228e
+	golang.org/x/sys v0.0.0-20210403161142-5e06dd20ab57
 	golang.org/x/term v0.0.0-20210317153231-de623e64d2a6
 	golang.org/x/time v0.0.0-20210220033141-f8bda1e9f3ba
 	golang.org/x/tools v0.0.0-20201211185031-d93e913c1a58

go.sum

@@ -123,6 +123,24 @@ github.com/tailscale/wireguard-go v0.0.0-20210324165952-2963b66bc23a h1:tQ7Y0ALS
 github.com/tailscale/wireguard-go v0.0.0-20210324165952-2963b66bc23a/go.mod h1:6t0OVdJwFOKFnvaHaVMKG6GznWaHqkmiR2n3kH0t924=
 github.com/tailscale/wireguard-go v0.0.0-20210327173134-f6a42a1646a0 h1:7KFBvUmm3TW/K+bAN22D7M6xSSoY/39s+PajaNBGrLw=
 github.com/tailscale/wireguard-go v0.0.0-20210327173134-f6a42a1646a0/go.mod h1:6t0OVdJwFOKFnvaHaVMKG6GznWaHqkmiR2n3kH0t924=
+github.com/tailscale/wireguard-go v0.0.0-20210330185929-1689f2635004 h1:GNEPNdNHsYe5zhoR/0z2Pl/a9zXbr0IySmHV6PhCrzI=
+github.com/tailscale/wireguard-go v0.0.0-20210330185929-1689f2635004/go.mod h1:6t0OVdJwFOKFnvaHaVMKG6GznWaHqkmiR2n3kH0t924=
+github.com/tailscale/wireguard-go v0.0.0-20210330200845-4914b4a944c4 h1:7Y0H5NzrV3fwHeDrUXDFcTy8QNbAEDwr+qHyOfX4VyE=
+github.com/tailscale/wireguard-go v0.0.0-20210330200845-4914b4a944c4/go.mod h1:6t0OVdJwFOKFnvaHaVMKG6GznWaHqkmiR2n3kH0t924=
+github.com/tailscale/wireguard-go v0.0.0-20210401164443-2d6878b6b30d h1:zbDBqtYvc492gcRL5BB7AO5Aed+aVht2jbYg8SKoMYs=
+github.com/tailscale/wireguard-go v0.0.0-20210401164443-2d6878b6b30d/go.mod h1:6t0OVdJwFOKFnvaHaVMKG6GznWaHqkmiR2n3kH0t924=
+github.com/tailscale/wireguard-go v0.0.0-20210401172819-1aca620a8afb h1:6TGRROCOrjTKbt1ucBTZaDMBeScG6yVEXEjuabOiBzU=
+github.com/tailscale/wireguard-go v0.0.0-20210401172819-1aca620a8afb/go.mod h1:jy12FSeiDLRvS7VQvSoiaqH9WtpapbrC8YSzyZ7fUAk=
+github.com/tailscale/wireguard-go v0.0.0-20210401194826-bb7bc2f24083 h1:e3k65apTVs7NM6mhQ1c94XISLe+2gdizPfRdsImNL8Y=
+github.com/tailscale/wireguard-go v0.0.0-20210401194826-bb7bc2f24083/go.mod h1:jy12FSeiDLRvS7VQvSoiaqH9WtpapbrC8YSzyZ7fUAk=
+github.com/tailscale/wireguard-go v0.0.0-20210402173217-0a47c6e64d15 h1:13GZsTKbCmPGwDBurcSXT+ssYID2IfcX0MfsvhaaagY=
+github.com/tailscale/wireguard-go v0.0.0-20210402173217-0a47c6e64d15/go.mod h1:jy12FSeiDLRvS7VQvSoiaqH9WtpapbrC8YSzyZ7fUAk=
+github.com/tailscale/wireguard-go v0.0.0-20210402193818-fc309421dd43 h1:SRUknVD6AHsxfghv0By9SFjQ8dhn8K8gIFwxf3OEPyU=
+github.com/tailscale/wireguard-go v0.0.0-20210402193818-fc309421dd43/go.mod h1:g3WdWX37upLnDT8STKFWhvA34Gwrt4hIpnWR3HGufpM=
+github.com/tailscale/wireguard-go v0.0.0-20210403171604-17614717a9b5 h1:FegsXWjtyhCxpB8bBSL1kLzagtV+e7BaX07phMM8uQM=
+github.com/tailscale/wireguard-go v0.0.0-20210403171604-17614717a9b5/go.mod h1:ys4yUmhKncXy1jWP34qUHKipRjl322VVhxoh1Rkfo7c=
+github.com/tailscale/wireguard-go v0.0.0-20210419202603-b32acd8f0292 h1:rKgYi0k3TNqEz5f7sc6zNeufZcnxm1Efd6bb39cGGkY=
+github.com/tailscale/wireguard-go v0.0.0-20210419202603-b32acd8f0292/go.mod h1:ys4yUmhKncXy1jWP34qUHKipRjl322VVhxoh1Rkfo7c=
 github.com/tcnksm/go-httpstat v0.2.0 h1:rP7T5e5U2HfmOBmZzGgGZjBQ5/GluWUylujl0tJ04I0=
 github.com/tcnksm/go-httpstat v0.2.0/go.mod h1:s3JVJFtQxtBEBC9dwcdTTXS9xFnM3SXAZwPG41aurT8=
 github.com/toqueteos/webbrowser v1.2.0 h1:tVP/gpK69Fx+qMJKsLE7TD8LuGWPnEV71wBN9rrstGQ=
@@ -206,6 +224,10 @@ golang.org/x/sys v0.0.0-20210309040221-94ec62e08169/go.mod h1:h1NjWce9XRLGQEsW7w
 golang.org/x/sys v0.0.0-20210316164454-77fc1eacc6aa/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
 golang.org/x/sys v0.0.0-20210317225723-c4fcb01b228e h1:XNp2Flc/1eWQGk5BLzqTAN7fQIwIbfyVTuVxXxZh73M=
 golang.org/x/sys v0.0.0-20210317225723-c4fcb01b228e/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
+golang.org/x/sys v0.0.0-20210402192133-700132347e07 h1:4k6HsQjxj6hVMsI2Vf0yKlzt5lXxZsMW1q0zaq2k8zY=
+golang.org/x/sys v0.0.0-20210402192133-700132347e07/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
+golang.org/x/sys v0.0.0-20210403161142-5e06dd20ab57 h1:F5Gozwx4I1xtr/sr/8CFbb57iKi3297KFs0QDbGN60A=
+golang.org/x/sys v0.0.0-20210403161142-5e06dd20ab57/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
 golang.org/x/term v0.0.0-20201117132131-f5c789dd3221/go.mod h1:Nr5EML6q2oocZ2LXRh80K7BxOlk5/8JxuGnuhpl+muw=
 golang.org/x/term v0.0.0-20201126162022-7de9c90e9dd1/go.mod h1:bj7SfCRtBDWHUb9snDiAeCFNEtKQo2Wmx5Cou7ajbmo=
 golang.org/x/term v0.0.0-20210317153231-de623e64d2a6 h1:EC6+IGYTjPpRfv9a2b/6Puw0W+hLtAhkV1tPsXhutqs=

health/health.go

@@ -49,6 +49,9 @@ const (
 	// SysRouter is the name the wgengine/router subsystem.
 	SysRouter = Subsystem("router")
 
+	// SysDNS is the name of the net/dns subsystem.
+	SysDNS = Subsystem("dns")
+
 	// SysNetworkCategory is the name of the subsystem that sets
 	// the Windows network adapter's "category" (public, private, domain).
 	// If it's unhealthy, the Windows firewall rules won't match.
@@ -80,12 +83,18 @@ func RegisterWatcher(cb func(key Subsystem, err error)) (unregister func()) {
 	}
 }
 
-// SetRouter sets the state of the wgengine/router.Router.
+// SetRouterHealth sets the state of the wgengine/router.Router.
 func SetRouterHealth(err error) { set(SysRouter, err) }
 
 // RouterHealth returns the wgengine/router.Router error state.
 func RouterHealth() error { return get(SysRouter) }
 
+// SetDNSHealth sets the state of the net/dns.Manager
+func SetDNSHealth(err error) { set(SysDNS, err) }
+
+// DNSHealth returns the net/dns.Manager error state.
+func DNSHealth() error { return get(SysDNS) }
+
 // SetNetworkCategoryHealth sets the state of setting the network adaptor's category.
 // This only applies on Windows.
 func SetNetworkCategoryHealth(err error) { set(SysNetworkCategory, err) }

internal/deepprint/deepprint_test.go

@@ -9,7 +9,6 @@ import (
 	"testing"
 
 	"inet.af/netaddr"
-	"tailscale.com/net/dns"
 	"tailscale.com/wgengine/router"
 	"tailscale.com/wgengine/wgcfg"
 )
@@ -36,9 +35,8 @@ func TestDeepPrint(t *testing.T) {
 func getVal() []interface{} {
 	return []interface{}{
 		&wgcfg.Config{
-			Name:       "foo",
-			Addresses:  []netaddr.IPPrefix{{Bits: 5, IP: netaddr.IPFrom16([16]byte{3: 3})}},
-			ListenPort: 5,
+			Name:      "foo",
+			Addresses: []netaddr.IPPrefix{{Bits: 5, IP: netaddr.IPFrom16([16]byte{3: 3})}},
 			Peers: []wgcfg.Peer{
 				{
 					Endpoints: "foo:5",
@@ -46,9 +44,9 @@ func getVal() []interface{} {
 			},
 		},
 		&router.Config{
-			DNS: dns.Config{
-				Nameservers: []netaddr.IP{netaddr.IPv4(8, 8, 8, 8)},
-				Domains:     []string{"tailscale.net"},
+			Routes: []netaddr.IPPrefix{
+				netaddr.MustParseIPPrefix("1.2.3.0/24"),
+				netaddr.MustParseIPPrefix("1234::/64"),
 			},
 		},
 		map[string]string{

ipn/backend.go

@@ -5,7 +5,6 @@
 package ipn
 
 import (
-	"net/http"
 	"time"
 
 	"tailscale.com/ipn/ipnstate"
@@ -68,6 +67,18 @@ type Notify struct {
 	BackendLogID  *string            // public logtail id used by backend
 	PingResult    *ipnstate.PingResult
 
+	// FilesWaiting if non-nil means that files are buffered in
+	// the Tailscale daemon and ready for local transfer to the
+	// user's preferred storage location.
+	FilesWaiting *empty.Message `json:",omitempty"`
+
+	// IncomingFiles, if non-nil, specifies which files are in the
+	// process of being received. A nil IncomingFiles means this
+	// Notify should not update the state of file transfers. A non-nil
+	// but empty IncomingFiles means that no files are in the middle
+	// of being transferred.
+	IncomingFiles []PartialFile `json:",omitempty"`
+
 	// LocalTCPPort, if non-nil, informs the UI frontend which
 	// (non-zero) localhost TCP port it's listening on.
 	// This is currently only used by Tailscale when run in the
@@ -77,6 +88,24 @@ type Notify struct {
 	// type is mirrored in xcode/Shared/IPN.swift
 }
 
+// PartialFile represents an in-progress file transfer.
+type PartialFile struct {
+	Name         string    // e.g. "foo.jpg"
+	Started      time.Time // time transfer started
+	DeclaredSize int64     // or -1 if unknown
+	Received     int64     // bytes copied thus far
+
+	// PartialPath is set non-empty in "direct" file mode to the
+	// in-progress '*.partial' file's path when the peerapi isn't
+	// being used; see LocalBackend.SetDirectFileRoot.
+	PartialPath string `json:",omitempty"`
+
+	// Done is set in "direct" mode when the partial file has been
+	// closed and is ready for the caller to rename away the
+	// ".partial" suffix.
+	Done bool `json:",omitempty"`
+}
+
 // StateKey is an opaque identifier for a set of LocalBackend state
 // (preferences, private keys, etc.).
 //
@@ -112,19 +141,6 @@ type Options struct {
 	// AuthKey is an optional node auth key used to authorize a
 	// new node key without user interaction.
 	AuthKey string
-	// LegacyConfigPath optionally specifies the old-style relaynode
-	// relay.conf location. If both LegacyConfigPath and StateKey are
-	// specified and the requested state doesn't exist in the backend
-	// store, the backend migrates the config from LegacyConfigPath.
-	//
-	// TODO(danderson): remove some time after the transition to
-	// tailscaled is done.
-	LegacyConfigPath string
-	// Notify is called when backend events happen.
-	Notify func(Notify) `json:"-"`
-	// HTTPTestClient is an optional HTTP client to pass to controlclient
-	// (for tests only).
-	HTTPTestClient *http.Client
 }
 
 // Backend is the interface between Tailscale frontends
@@ -133,6 +149,9 @@ type Options struct {
 // (It has nothing to do with the interface between the backends
 // and the cloud control plane.)
 type Backend interface {
+	// SetNotifyCallback sets the callback to be called on updates
+	// from the backend to the client.
+	SetNotifyCallback(func(Notify))
 	// Start starts or restarts the backend, typically when a
 	// frontend client connects.
 	Start(Options) error
@@ -149,9 +168,6 @@ type Backend interface {
 	// WantRunning. This may cause the wireguard engine to
 	// reconfigure or stop.
 	SetPrefs(*Prefs)
-	// SetWantRunning is like SetPrefs but sets only the
-	// WantRunning field.
-	SetWantRunning(wantRunning bool)
 	// RequestEngineStatus polls for an update from the wireguard
 	// engine. Only needed if you want to display byte
 	// counts. Connection events are emitted automatically without

ipn/fake_test.go

@@ -5,7 +5,6 @@
 package ipn
 
 import (
-	"log"
 	"time"
 
 	"tailscale.com/ipn/ipnstate"
@@ -20,19 +19,29 @@ type FakeBackend struct {
 }
 
 func (b *FakeBackend) Start(opts Options) error {
-	b.serverURL = opts.Prefs.ControlURL
-	if opts.Notify == nil {
-		log.Fatalf("FakeBackend.Start: opts.Notify is nil\n")
+	b.serverURL = opts.Prefs.ControlURLOrDefault()
+	if b.notify == nil {
+		panic("FakeBackend.Start: SetNotifyCallback not called")
 	}
-	b.notify = opts.Notify
-	b.notify(Notify{Prefs: opts.Prefs})
 	nl := NeedsLogin
-	b.notify(Notify{State: &nl})
+	if b.notify != nil {
+		b.notify(Notify{Prefs: opts.Prefs})
+		b.notify(Notify{State: &nl})
+	}
 	return nil
 }
 
+func (b *FakeBackend) SetNotifyCallback(notify func(Notify)) {
+	if notify == nil {
+		panic("FakeBackend.SetNotifyCallback: notify is nil")
+	}
+	b.notify = notify
+}
+
 func (b *FakeBackend) newState(s State) {
-	b.notify(Notify{State: &s})
+	if b.notify != nil {
+		b.notify(Notify{State: &s})
+	}
 	if s == Running {
 		b.live = true
 	} else {
@@ -42,7 +51,9 @@ func (b *FakeBackend) newState(s State) {
 
 func (b *FakeBackend) StartLoginInteractive() {
 	u := b.serverURL + "/this/is/fake"
-	b.notify(Notify{BrowseToURL: &u})
+	if b.notify != nil {
+		b.notify(Notify{BrowseToURL: &u})
+	}
 	b.login()
 }
 
@@ -54,10 +65,14 @@ func (b *FakeBackend) login() {
 	b.newState(NeedsMachineAuth)
 	b.newState(Stopped)
 	// TODO(apenwarr): Fill in a more interesting netmap here.
-	b.notify(Notify{NetMap: &netmap.NetworkMap{}})
+	if b.notify != nil {
+		b.notify(Notify{NetMap: &netmap.NetworkMap{}})
+	}
 	b.newState(Starting)
 	// TODO(apenwarr): Fill in a more interesting status.
-	b.notify(Notify{Engine: &EngineStatus{}})
+	if b.notify != nil {
+		b.notify(Notify{Engine: &EngineStatus{}})
+	}
 	b.newState(Running)
 }
 
@@ -70,7 +85,9 @@ func (b *FakeBackend) SetPrefs(new *Prefs) {
 		panic("FakeBackend.SetPrefs got nil prefs")
 	}
 
-	b.notify(Notify{Prefs: new.Clone()})
+	if b.notify != nil {
+		b.notify(Notify{Prefs: new.Clone()})
+	}
 	if new.WantRunning && !b.live {
 		b.newState(Starting)
 		b.newState(Running)
@@ -79,18 +96,20 @@ func (b *FakeBackend) SetPrefs(new *Prefs) {
 	}
 }
 
-func (b *FakeBackend) SetWantRunning(v bool) {
-	b.SetPrefs(&Prefs{WantRunning: v})
-}
-
 func (b *FakeBackend) RequestEngineStatus() {
-	b.notify(Notify{Engine: &EngineStatus{}})
+	if b.notify != nil {
+		b.notify(Notify{Engine: &EngineStatus{}})
+	}
 }
 
 func (b *FakeBackend) FakeExpireAfter(x time.Duration) {
-	b.notify(Notify{NetMap: &netmap.NetworkMap{}})
+	if b.notify != nil {
+		b.notify(Notify{NetMap: &netmap.NetworkMap{}})
+	}
 }
 
 func (b *FakeBackend) Ping(ip string, useTSMP bool) {
-	b.notify(Notify{PingResult: &ipnstate.PingResult{}})
+	if b.notify != nil {
+		b.notify(Notify{PingResult: &ipnstate.PingResult{}})
+	}
 }

ipn/handle.go

@@ -15,25 +15,26 @@ import (
 )
 
 type Handle struct {
-	frontendLogID string
-	b             Backend
-	xnotify       func(Notify)
-	logf          logger.Logf
+	b    Backend
+	logf logger.Logf
 
 	// Mutex protects everything below
 	mu                sync.Mutex
+	xnotify           func(Notify)
+	frontendLogID     string
 	netmapCache       *netmap.NetworkMap
 	engineStatusCache EngineStatus
 	stateCache        State
 	prefsCache        *Prefs
 }
 
-func NewHandle(b Backend, logf logger.Logf, opts Options) (*Handle, error) {
+func NewHandle(b Backend, logf logger.Logf, notify func(Notify), opts Options) (*Handle, error) {
 	h := &Handle{
 		b:    b,
 		logf: logf,
 	}
 
+	h.SetNotifyCallback(notify)
 	err := h.Start(opts)
 	if err != nil {
 		return nil, err
@@ -42,18 +43,25 @@ func NewHandle(b Backend, logf logger.Logf, opts Options) (*Handle, error) {
 	return h, nil
 }
 
+func (h *Handle) SetNotifyCallback(notify func(Notify)) {
+	h.mu.Lock()
+	h.xnotify = notify
+	h.mu.Unlock()
+
+	h.b.SetNotifyCallback(h.notify)
+}
+
 func (h *Handle) Start(opts Options) error {
+	h.mu.Lock()
 	h.frontendLogID = opts.FrontendLogID
-	h.xnotify = opts.Notify
 	h.netmapCache = nil
 	h.engineStatusCache = EngineStatus{}
 	h.stateCache = NoState
 	if opts.Prefs != nil {
 		h.prefsCache = opts.Prefs.Clone()
 	}
-	xopts := opts
-	xopts.Notify = h.notify
-	return h.b.Start(xopts)
+	h.mu.Unlock()
+	return h.b.Start(opts)
 }
 
 func (h *Handle) Reset() {
@@ -148,7 +156,7 @@ func (h *Handle) Expiry() time.Time {
 }
 
 func (h *Handle) AdminPageURL() string {
-	return h.prefsCache.ControlURL + "/admin/machines"
+	return h.prefsCache.ControlURLOrDefault() + "/admin/machines"
 }
 
 func (h *Handle) StartLoginInteractive() {

ipn/ipnlocal/local_test.go

@@ -5,14 +5,20 @@
 package ipnlocal
 
 import (
+	"bytes"
+	"fmt"
+	"net/http"
 	"reflect"
+	"sync"
 	"testing"
 
 	"inet.af/netaddr"
+	"tailscale.com/ipn"
 	"tailscale.com/net/interfaces"
 	"tailscale.com/net/tsaddr"
 	"tailscale.com/tailcfg"
 	"tailscale.com/types/netmap"
+	"tailscale.com/wgengine"
 	"tailscale.com/wgengine/wgcfg"
 )
 
@@ -291,3 +297,176 @@ func TestPeerRoutes(t *testing.T) {
 	}
 
 }
+
+func TestPeerAPIBase(t *testing.T) {
+	tests := []struct {
+		name string
+		nm   *netmap.NetworkMap
+		peer *tailcfg.Node
+		want string
+	}{
+		{
+			name: "nil_netmap",
+			peer: new(tailcfg.Node),
+			want: "",
+		},
+		{
+			name: "nil_peer",
+			nm:   new(netmap.NetworkMap),
+			want: "",
+		},
+		{
+			name: "self_only_4_them_both",
+			nm: &netmap.NetworkMap{
+				Addresses: []netaddr.IPPrefix{
+					netaddr.MustParseIPPrefix("100.64.1.1/32"),
+				},
+			},
+			peer: &tailcfg.Node{
+				Addresses: []netaddr.IPPrefix{
+					netaddr.MustParseIPPrefix("100.64.1.2/32"),
+					netaddr.MustParseIPPrefix("fe70::2/128"),
+				},
+				Hostinfo: tailcfg.Hostinfo{
+					Services: []tailcfg.Service{
+						{Proto: "peerapi4", Port: 444},
+						{Proto: "peerapi6", Port: 666},
+					},
+				},
+			},
+			want: "http://100.64.1.2:444",
+		},
+		{
+			name: "self_only_6_them_both",
+			nm: &netmap.NetworkMap{
+				Addresses: []netaddr.IPPrefix{
+					netaddr.MustParseIPPrefix("fe70::1/128"),
+				},
+			},
+			peer: &tailcfg.Node{
+				Addresses: []netaddr.IPPrefix{
+					netaddr.MustParseIPPrefix("100.64.1.2/32"),
+					netaddr.MustParseIPPrefix("fe70::2/128"),
+				},
+				Hostinfo: tailcfg.Hostinfo{
+					Services: []tailcfg.Service{
+						{Proto: "peerapi4", Port: 444},
+						{Proto: "peerapi6", Port: 666},
+					},
+				},
+			},
+			want: "http://[fe70::2]:666",
+		},
+		{
+			name: "self_both_them_only_4",
+			nm: &netmap.NetworkMap{
+				Addresses: []netaddr.IPPrefix{
+					netaddr.MustParseIPPrefix("100.64.1.1/32"),
+					netaddr.MustParseIPPrefix("fe70::1/128"),
+				},
+			},
+			peer: &tailcfg.Node{
+				Addresses: []netaddr.IPPrefix{
+					netaddr.MustParseIPPrefix("100.64.1.2/32"),
+					netaddr.MustParseIPPrefix("fe70::2/128"),
+				},
+				Hostinfo: tailcfg.Hostinfo{
+					Services: []tailcfg.Service{
+						{Proto: "peerapi4", Port: 444},
+					},
+				},
+			},
+			want: "http://100.64.1.2:444",
+		},
+		{
+			name: "self_both_them_only_6",
+			nm: &netmap.NetworkMap{
+				Addresses: []netaddr.IPPrefix{
+					netaddr.MustParseIPPrefix("100.64.1.1/32"),
+					netaddr.MustParseIPPrefix("fe70::1/128"),
+				},
+			},
+			peer: &tailcfg.Node{
+				Addresses: []netaddr.IPPrefix{
+					netaddr.MustParseIPPrefix("100.64.1.2/32"),
+					netaddr.MustParseIPPrefix("fe70::2/128"),
+				},
+				Hostinfo: tailcfg.Hostinfo{
+					Services: []tailcfg.Service{
+						{Proto: "peerapi6", Port: 666},
+					},
+				},
+			},
+			want: "http://[fe70::2]:666",
+		},
+		{
+			name: "self_both_them_no_peerapi_service",
+			nm: &netmap.NetworkMap{
+				Addresses: []netaddr.IPPrefix{
+					netaddr.MustParseIPPrefix("100.64.1.1/32"),
+					netaddr.MustParseIPPrefix("fe70::1/128"),
+				},
+			},
+			peer: &tailcfg.Node{
+				Addresses: []netaddr.IPPrefix{
+					netaddr.MustParseIPPrefix("100.64.1.2/32"),
+					netaddr.MustParseIPPrefix("fe70::2/128"),
+				},
+			},
+			want: "",
+		},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got := peerAPIBase(tt.nm, tt.peer)
+			if got != tt.want {
+				t.Errorf("got %q; want %q", got, tt.want)
+			}
+		})
+	}
+}
+
+type panicOnUseTransport struct{}
+
+func (panicOnUseTransport) RoundTrip(*http.Request) (*http.Response, error) {
+	panic("unexpected HTTP request")
+}
+
+var nl = []byte("\n")
+
+func TestStartsInNeedsLoginState(t *testing.T) {
+	var (
+		mu     sync.Mutex
+		logBuf bytes.Buffer
+	)
+	logf := func(format string, a ...interface{}) {
+		mu.Lock()
+		defer mu.Unlock()
+		fmt.Fprintf(&logBuf, format, a...)
+		if !bytes.HasSuffix(logBuf.Bytes(), nl) {
+			logBuf.Write(nl)
+		}
+	}
+	store := new(ipn.MemoryStore)
+	eng, err := wgengine.NewFakeUserspaceEngine(logf, 0)
+	if err != nil {
+		t.Fatalf("NewFakeUserspaceEngine: %v", err)
+	}
+	lb, err := NewLocalBackend(logf, "logid", store, eng)
+	if err != nil {
+		t.Fatalf("NewLocalBackend: %v", err)
+	}
+
+	lb.SetHTTPTestClient(&http.Client{
+		Transport: panicOnUseTransport{}, // validate we don't send HTTP requests
+	})
+
+	if err := lb.Start(ipn.Options{
+		StateKey: ipn.GlobalDaemonStateKey,
+	}); err != nil {
+		t.Fatalf("Start: %v", err)
+	}
+	if st := lb.State(); st != ipn.NeedsLogin {
+		t.Errorf("State = %v; want NeedsLogin", st)
+	}
+}

ipn/ipnlocal/loglines_test.go

@@ -82,7 +82,6 @@ func TestLocalLogLines(t *testing.T) {
 			LastHandshake: time.Now(),
 			NodeKey:       tailcfg.NodeKey(key.NewPrivate()),
 		}},
-		LocalAddrs: []string{"idk an address"},
 	}
 	lb.mu.Lock()
 	lb.parseWgStatusLocked(status)

ipn/ipnlocal/peerapi.go

@@ -20,19 +20,232 @@ import (
 	"runtime"
 	"strconv"
 	"strings"
+	"sync"
+	"time"
+	"unicode"
+	"unicode/utf8"
 
 	"inet.af/netaddr"
+	"tailscale.com/client/tailscale/apitype"
+	"tailscale.com/ipn"
+	"tailscale.com/logtail/backoff"
 	"tailscale.com/net/interfaces"
+	"tailscale.com/syncs"
 	"tailscale.com/tailcfg"
+	"tailscale.com/wgengine"
 )
 
 var initListenConfig func(*net.ListenConfig, netaddr.IP, *interfaces.State, string) error
 
 type peerAPIServer struct {
-	b        *LocalBackend
-	rootDir  string
-	tunName  string
-	selfNode *tailcfg.Node
+	b          *LocalBackend
+	rootDir    string
+	tunName    string
+	selfNode   *tailcfg.Node
+	knownEmpty syncs.AtomicBool
+
+	// directFileMode is whether we're writing files directly to a
+	// download directory (as *.partial files), rather than making
+	// the frontend retrieve it over localapi HTTP and write it
+	// somewhere itself. This is used on GUI macOS version.
+	directFileMode bool
+}
+
+const partialSuffix = ".partial"
+
+func validFilenameRune(r rune) bool {
+	switch r {
+	case '/':
+		return false
+	case '\\', ':', '*', '"', '<', '>', '|':
+		// Invalid stuff on Windows, but we reject them everywhere
+		// for now.
+		// TODO(bradfitz): figure out a better plan. We initially just
+		// wrote things to disk URL path-escaped, but that's gross
+		// when debugging, and just moves the problem to callers.
+		// So now we put the UTF-8 filenames on disk directly as
+		// sent.
+		return false
+	}
+	return unicode.IsPrint(r)
+}
+
+func (s *peerAPIServer) diskPath(baseName string) (fullPath string, ok bool) {
+	if !utf8.ValidString(baseName) {
+		return "", false
+	}
+	if strings.TrimSpace(baseName) != baseName {
+		return "", false
+	}
+	if len(baseName) > 255 {
+		return "", false
+	}
+	// TODO: validate unicode normalization form too? Varies by platform.
+	clean := path.Clean(baseName)
+	if clean != baseName ||
+		clean == "." || clean == ".." ||
+		strings.HasSuffix(clean, partialSuffix) {
+		return "", false
+	}
+	for _, r := range baseName {
+		if !validFilenameRune(r) {
+			return "", false
+		}
+	}
+	return filepath.Join(s.rootDir, baseName), true
+}
+
+// hasFilesWaiting reports whether any files are buffered in the
+// tailscaled daemon storage.
+func (s *peerAPIServer) hasFilesWaiting() bool {
+	if s.rootDir == "" || s.directFileMode {
+		return false
+	}
+	if s.knownEmpty.Get() {
+		// Optimization: this is usually empty, so avoid opening
+		// the directory and checking. We can't cache the actual
+		// has-files-or-not values as the macOS/iOS client might
+		// in the future use+delete the files directly. So only
+		// keep this negative cache.
+		return false
+	}
+	f, err := os.Open(s.rootDir)
+	if err != nil {
+		return false
+	}
+	defer f.Close()
+	for {
+		des, err := f.ReadDir(10)
+		for _, de := range des {
+			if strings.HasSuffix(de.Name(), partialSuffix) {
+				continue
+			}
+			if de.Type().IsRegular() {
+				return true
+			}
+		}
+		if err == io.EOF {
+			s.knownEmpty.Set(true)
+		}
+		if err != nil {
+			break
+		}
+	}
+	return false
+}
+
+func (s *peerAPIServer) WaitingFiles() (ret []apitype.WaitingFile, err error) {
+	if s.rootDir == "" {
+		return nil, errors.New("peerapi disabled; no storage configured")
+	}
+	if s.directFileMode {
+		return nil, nil
+	}
+	f, err := os.Open(s.rootDir)
+	if err != nil {
+		return nil, err
+	}
+	defer f.Close()
+	for {
+		des, err := f.ReadDir(10)
+		for _, de := range des {
+			name := de.Name()
+			if strings.HasSuffix(name, partialSuffix) {
+				continue
+			}
+			if de.Type().IsRegular() {
+				fi, err := de.Info()
+				if err != nil {
+					continue
+				}
+				ret = append(ret, apitype.WaitingFile{
+					Name: filepath.Base(name),
+					Size: fi.Size(),
+				})
+			}
+		}
+		if err == io.EOF {
+			break
+		}
+		if err != nil {
+			return nil, err
+		}
+	}
+	return ret, nil
+}
+
+func (s *peerAPIServer) DeleteFile(baseName string) error {
+	if s.rootDir == "" {
+		return errors.New("peerapi disabled; no storage configured")
+	}
+	if s.directFileMode {
+		return errors.New("deletes not allowed in direct mode")
+	}
+	path, ok := s.diskPath(baseName)
+	if !ok {
+		return errors.New("bad filename")
+	}
+	var bo *backoff.Backoff
+	logf := s.b.logf
+	t0 := time.Now()
+	for {
+		err := os.Remove(path)
+		if err != nil && !os.IsNotExist(err) {
+			if pe, ok := err.(*os.PathError); ok {
+				pe.Path = "redact"
+			}
+			// Put a retry loop around deletes on Windows. Windows
+			// file descriptor closes are effectively asynchronous,
+			// as a bunch of hooks run on/after close, and we can't
+			// necessarily delete the file for a while after close,
+			// as we need to wait for everybody to be done with
+			// it. (on Windows, unlike Unix, a file can't be deleted
+			// while open)
+			//
+			// TODO(bradfitz): we might instead want to just keep a
+			// map of logically deleted files and filter them out in
+			// WaitingFiles/OpenFile.  Then we can keep trying this
+			// delete in the background and/or in response to future
+			// WaitingFiles/OpenFile calls, and then remove from the
+			// logicallyDeleted map. But let's start with this retry
+			// loop.
+			if runtime.GOOS == "windows" {
+				if bo == nil {
+					bo = backoff.NewBackoff("delete-retry", logf, 1*time.Second)
+				}
+				if time.Since(t0) < 10*time.Second {
+					bo.BackOff(context.Background(), err)
+					continue
+				}
+			}
+			logf("peerapi: failed to DeleteFile: %v", err)
+			return err
+		}
+		return nil
+	}
+}
+
+func (s *peerAPIServer) OpenFile(baseName string) (rc io.ReadCloser, size int64, err error) {
+	if s.rootDir == "" {
+		return nil, 0, errors.New("peerapi disabled; no storage configured")
+	}
+	if s.directFileMode {
+		return nil, 0, errors.New("opens not allowed in direct mode")
+	}
+	path, ok := s.diskPath(baseName)
+	if !ok {
+		return nil, 0, errors.New("bad filename")
+	}
+	f, err := os.Open(path)
+	if err != nil {
+		return nil, 0, err
+	}
+	fi, err := f.Stat()
+	if err != nil {
+		f.Close()
+		return nil, 0, err
+	}
+	return f, fi.Size(), nil
 }
 
 func (s *peerAPIServer) listen(ip netaddr.IP, ifState *interfaces.State) (ln net.Listener, err error) {
@@ -51,6 +264,10 @@ func (s *peerAPIServer) listen(ip netaddr.IP, ifState *interfaces.State) (ln net
 		}
 	}
 
+	if wgengine.IsNetstack(s.b.e) {
+		ipStr = ""
+	}
+
 	tcp4or6 := "tcp4"
 	if ip.Is6() {
 		tcp4or6 = "tcp6"
@@ -79,22 +296,32 @@ func (s *peerAPIServer) listen(ip netaddr.IP, ifState *interfaces.State) (ln net
 }
 
 type peerAPIListener struct {
-	ps     *peerAPIServer
-	ip     netaddr.IP
-	ln     net.Listener
-	lb     *LocalBackend
+	ps *peerAPIServer
+	ip netaddr.IP
+	lb *LocalBackend
+
+	// ln is the Listener. It can be nil in netstack mode if there are more than
+	// 1 local addresses (e.g. both an IPv4 and IPv6). When it's nil, port
+	// and urlStr are still populated.
+	ln net.Listener
+
+	// urlStr is the base URL to access the peer API (http://ip:port/).
 	urlStr string
+	// port is just the port of urlStr.
+	port int
 }
 
-func (pln *peerAPIListener) Port() int {
-	ta, ok := pln.ln.Addr().(*net.TCPAddr)
-	if !ok {
-		return 0
+func (pln *peerAPIListener) Close() error {
+	if pln.ln != nil {
+		return pln.ln.Close()
 	}
-	return ta.Port
+	return nil
 }
 
 func (pln *peerAPIListener) serve() {
+	if pln.ln == nil {
+		return
+	}
 	defer pln.ln.Close()
 	logf := pln.lb.logf
 	for {
@@ -130,7 +357,6 @@ func (pln *peerAPIListener) serve() {
 			remoteAddr: ipp,
 			peerNode:   peerNode,
 			peerUser:   peerUser,
-			lb:         pln.lb,
 		}
 		httpServer := &http.Server{
 			Handler: h,
@@ -164,7 +390,6 @@ type peerAPIHandler struct {
 	isSelf     bool                // whether peerNode is owned by same user as this node
 	peerNode   *tailcfg.Node       // peerNode is who's making the request
 	peerUser   tailcfg.UserProfile // profile of peerNode
-	lb         *LocalBackend
 }
 
 func (h *peerAPIHandler) logf(format string, a ...interface{}) {
@@ -173,7 +398,7 @@ func (h *peerAPIHandler) logf(format string, a ...interface{}) {
 
 func (h *peerAPIHandler) ServeHTTP(w http.ResponseWriter, r *http.Request) {
 	if strings.HasPrefix(r.URL.Path, "/v0/put/") {
-		h.put(w, r)
+		h.handlePeerPut(w, r)
 		return
 	}
 	who := h.peerUser.DisplayName
@@ -189,26 +414,108 @@ This is my Tailscale device. Your device is %v.
 	}
 }
 
-func (h *peerAPIHandler) put(w http.ResponseWriter, r *http.Request) {
+type incomingFile struct {
+	name        string // "foo.jpg"
+	started     time.Time
+	size        int64     // or -1 if unknown; never 0
+	w           io.Writer // underlying writer
+	ph          *peerAPIHandler
+	partialPath string // non-empty in direct mode
+
+	mu         sync.Mutex
+	copied     int64
+	done       bool
+	lastNotify time.Time
+}
+
+func (f *incomingFile) markAndNotifyDone() {
+	f.mu.Lock()
+	f.done = true
+	f.mu.Unlock()
+	b := f.ph.ps.b
+	b.sendFileNotify()
+}
+
+func (f *incomingFile) Write(p []byte) (n int, err error) {
+	n, err = f.w.Write(p)
+
+	b := f.ph.ps.b
+	var needNotify bool
+	defer func() {
+		if needNotify {
+			b.sendFileNotify()
+		}
+	}()
+	if n > 0 {
+		f.mu.Lock()
+		defer f.mu.Unlock()
+		f.copied += int64(n)
+		now := time.Now()
+		if f.lastNotify.IsZero() || now.Sub(f.lastNotify) > time.Second {
+			f.lastNotify = now
+			needNotify = true
+		}
+	}
+	return n, err
+}
+
+func (f *incomingFile) PartialFile() ipn.PartialFile {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	return ipn.PartialFile{
+		Name:         f.name,
+		Started:      f.started,
+		DeclaredSize: f.size,
+		Received:     f.copied,
+		PartialPath:  f.partialPath,
+		Done:         f.done,
+	}
+}
+
+func (h *peerAPIHandler) handlePeerPut(w http.ResponseWriter, r *http.Request) {
 	if !h.isSelf {
 		http.Error(w, "not owner", http.StatusForbidden)
 		return
 	}
+	if !h.ps.b.hasCapFileSharing() {
+		http.Error(w, "file sharing not enabled by Tailscale admin", http.StatusForbidden)
+		return
+	}
 	if r.Method != "PUT" {
-		http.Error(w, "not method PUT", http.StatusMethodNotAllowed)
+		http.Error(w, "expected method PUT", http.StatusMethodNotAllowed)
 		return
 	}
 	if h.ps.rootDir == "" {
 		http.Error(w, "no rootdir", http.StatusInternalServerError)
 		return
 	}
-	name := path.Base(r.URL.Path)
-	if name == "." || name == "/" {
-		http.Error(w, "bad filename", http.StatusForbidden)
+	rawPath := r.URL.EscapedPath()
+	suffix := strings.TrimPrefix(rawPath, "/v0/put/")
+	if suffix == rawPath {
+		http.Error(w, "misconfigured internals", 500)
 		return
 	}
-	fileBase := strings.ReplaceAll(url.PathEscape(name), ":", "%3a")
-	dstFile := filepath.Join(h.ps.rootDir, fileBase)
+	if suffix == "" {
+		http.Error(w, "empty filename", 400)
+		return
+	}
+	if strings.Contains(suffix, "/") {
+		http.Error(w, "directories not supported", 400)
+		return
+	}
+	baseName, err := url.PathUnescape(suffix)
+	if err != nil {
+		http.Error(w, "bad path encoding", 400)
+		return
+	}
+	dstFile, ok := h.ps.diskPath(baseName)
+	if !ok {
+		http.Error(w, "bad filename", 400)
+		return
+	}
+	if h.ps.directFileMode {
+		dstFile += partialSuffix
+	}
 	f, err := os.Create(dstFile)
 	if err != nil {
 		h.logf("put Create error: %v", err)
@@ -221,23 +528,57 @@ func (h *peerAPIHandler) put(w http.ResponseWriter, r *http.Request) {
 			os.Remove(dstFile)
 		}
 	}()
-	n, err := io.Copy(f, r.Body)
-	if err != nil {
-		f.Close()
-		h.logf("put Copy error: %v", err)
-		http.Error(w, err.Error(), http.StatusInternalServerError)
-		return
+	var finalSize int64
+	var inFile *incomingFile
+	if r.ContentLength != 0 {
+		inFile = &incomingFile{
+			name:    baseName,
+			started: time.Now(),
+			size:    r.ContentLength,
+			w:       f,
+			ph:      h,
+		}
+		if h.ps.directFileMode {
+			inFile.partialPath = dstFile
+		}
+		h.ps.b.registerIncomingFile(inFile, true)
+		defer h.ps.b.registerIncomingFile(inFile, false)
+		n, err := io.Copy(inFile, r.Body)
+		if err != nil {
+			f.Close()
+			h.logf("put Copy error: %v", err)
+			http.Error(w, err.Error(), http.StatusInternalServerError)
+			return
+		}
+		finalSize = n
 	}
 	if err := f.Close(); err != nil {
 		h.logf("put Close error: %v", err)
 		http.Error(w, err.Error(), http.StatusInternalServerError)
 		return
 	}
+	if h.ps.directFileMode {
+		if inFile != nil { // non-zero length; TODO: notify even for zero length
+			inFile.markAndNotifyDone()
+		}
+	}
 
-	h.logf("put(%q): %d bytes from %v/%v", name, n, h.remoteAddr.IP, h.peerNode.ComputedName)
+	h.logf("put of %s from %v/%v", approxSize(finalSize), h.remoteAddr.IP, h.peerNode.ComputedName)
 
 	// TODO: set modtime
 	// TODO: some real response
 	success = true
 	io.WriteString(w, "{}\n")
+	h.ps.knownEmpty.Set(false)
+	h.ps.b.sendFileNotify()
+}
+
+func approxSize(n int64) string {
+	if n <= 1<<10 {
+		return "<=1KB"
+	}
+	if n <= 1<<20 {
+		return "<=1MB"
+	}
+	return fmt.Sprintf("~%dMB", n>>20)
 }

ipn/ipnlocal/peerapi_macios_ext.go

@@ -7,6 +7,7 @@
 package ipnlocal
 
 import (
+	"errors"
 	"fmt"
 	"log"
 	"net"
@@ -20,6 +21,7 @@ import (
 
 func init() {
 	initListenConfig = initListenConfigNetworkExtension
+	peerDialControlFunc = peerDialControlFuncNetworkExtension
 }
 
 // initListenConfigNetworkExtension configures nc for listening on IP
@@ -33,16 +35,7 @@ func initListenConfigNetworkExtension(nc *net.ListenConfig, ip netaddr.IP, st *i
 	nc.Control = func(network, address string, c syscall.RawConn) error {
 		var sockErr error
 		err := c.Control(func(fd uintptr) {
-
-			v6 := strings.Contains(address, "]:") || strings.HasSuffix(network, "6") // hacky test for v6
-			proto := unix.IPPROTO_IP
-			opt := unix.IP_BOUND_IF
-			if v6 {
-				proto = unix.IPPROTO_IPV6
-				opt = unix.IPV6_BOUND_IF
-			}
-
-			sockErr = unix.SetsockoptInt(int(fd), proto, opt, tunIf.Index)
+			sockErr = bindIf(fd, network, address, tunIf.Index)
 			log.Printf("peerapi: bind(%q, %q) on index %v = %v", network, address, tunIf.Index, sockErr)
 		})
 		if err != nil {
@@ -52,3 +45,40 @@ func initListenConfigNetworkExtension(nc *net.ListenConfig, ip netaddr.IP, st *i
 	}
 	return nil
 }
+
+func bindIf(fd uintptr, network, address string, ifIndex int) error {
+	v6 := strings.Contains(address, "]:") || strings.HasSuffix(network, "6") // hacky test for v6
+	proto := unix.IPPROTO_IP
+	opt := unix.IP_BOUND_IF
+	if v6 {
+		proto = unix.IPPROTO_IPV6
+		opt = unix.IPV6_BOUND_IF
+	}
+	return unix.SetsockoptInt(int(fd), proto, opt, ifIndex)
+}
+
+func peerDialControlFuncNetworkExtension(b *LocalBackend) func(network, address string, c syscall.RawConn) error {
+	b.mu.Lock()
+	defer b.mu.Unlock()
+	st := b.prevIfState
+	pas := b.peerAPIServer
+	index := -1
+	if st != nil && pas != nil && pas.tunName != "" {
+		if tunIf, ok := st.Interface[pas.tunName]; ok {
+			index = tunIf.Index
+		}
+	}
+	return func(network, address string, c syscall.RawConn) error {
+		if index == -1 {
+			return errors.New("failed to find TUN interface to bind to")
+		}
+		var sockErr error
+		err := c.Control(func(fd uintptr) {
+			sockErr = bindIf(fd, network, address, index)
+		})
+		if err != nil {
+			return err
+		}
+		return sockErr
+	}
+}

ipn/ipnlocal/peerapi_test.go

@@ -0,0 +1,476 @@
+// Copyright (c) 2021 Tailscale Inc & AUTHORS All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package ipnlocal
+
+import (
+	"bytes"
+	"fmt"
+	"io"
+	"io/fs"
+	"io/ioutil"
+	"math/rand"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"tailscale.com/tailcfg"
+	"tailscale.com/types/netmap"
+)
+
+type peerAPITestEnv struct {
+	ph     *peerAPIHandler
+	rr     *httptest.ResponseRecorder
+	logBuf bytes.Buffer
+}
+
+func (e *peerAPITestEnv) logf(format string, a ...interface{}) {
+	fmt.Fprintf(&e.logBuf, format, a...)
+}
+
+type check func(*testing.T, *peerAPITestEnv)
+
+func checks(vv ...check) []check { return vv }
+
+func httpStatus(wantStatus int) check {
+	return func(t *testing.T, e *peerAPITestEnv) {
+		if res := e.rr.Result(); res.StatusCode != wantStatus {
+			t.Errorf("HTTP response code = %v; want %v", res.Status, wantStatus)
+		}
+	}
+}
+
+func bodyContains(sub string) check {
+	return func(t *testing.T, e *peerAPITestEnv) {
+		if body := e.rr.Body.String(); !strings.Contains(body, sub) {
+			t.Errorf("HTTP response body does not contain %q; got: %s", sub, body)
+		}
+	}
+}
+
+func bodyNotContains(sub string) check {
+	return func(t *testing.T, e *peerAPITestEnv) {
+		if body := e.rr.Body.String(); strings.Contains(body, sub) {
+			t.Errorf("HTTP response body unexpectedly contains %q; got: %s", sub, body)
+		}
+	}
+}
+
+func fileHasSize(name string, size int) check {
+	return func(t *testing.T, e *peerAPITestEnv) {
+		root := e.ph.ps.rootDir
+		if root == "" {
+			t.Errorf("no rootdir; can't check whether %q has size %v", name, size)
+			return
+		}
+		path := filepath.Join(root, name)
+		if fi, err := os.Stat(path); err != nil {
+			t.Errorf("fileHasSize(%q, %v): %v", name, size, err)
+		} else if fi.Size() != int64(size) {
+			t.Errorf("file %q has size %v; want %v", name, fi.Size(), size)
+		}
+	}
+}
+
+func fileHasContents(name string, want string) check {
+	return func(t *testing.T, e *peerAPITestEnv) {
+		root := e.ph.ps.rootDir
+		if root == "" {
+			t.Errorf("no rootdir; can't check contents of %q", name)
+			return
+		}
+		path := filepath.Join(root, name)
+		got, err := ioutil.ReadFile(path)
+		if err != nil {
+			t.Errorf("fileHasContents: %v", err)
+			return
+		}
+		if string(got) != want {
+			t.Errorf("file contents = %q; want %q", got, want)
+		}
+	}
+}
+
+func hexAll(v string) string {
+	var sb strings.Builder
+	for i := 0; i < len(v); i++ {
+		fmt.Fprintf(&sb, "%%%02x", v[i])
+	}
+	return sb.String()
+}
+
+func TestHandlePeerPut(t *testing.T) {
+	tests := []struct {
+		name       string
+		isSelf     bool // the peer sending the request is owned by us
+		capSharing bool // self node has file sharing capabilty
+		omitRoot   bool // don't configure
+		req        *http.Request
+		checks     []check
+	}{
+		{
+			name:       "not_peer_api",
+			isSelf:     true,
+			capSharing: true,
+			req:        httptest.NewRequest("GET", "/", nil),
+			checks: checks(
+				httpStatus(200),
+				bodyContains("This is my Tailscale device."),
+				bodyContains("You are the owner of this node."),
+			),
+		},
+		{
+			name:       "not_peer_api_not_owner",
+			isSelf:     false,
+			capSharing: true,
+			req:        httptest.NewRequest("GET", "/", nil),
+			checks: checks(
+				httpStatus(200),
+				bodyContains("This is my Tailscale device."),
+				bodyNotContains("You are the owner of this node."),
+			),
+		},
+		{
+			name:       "reject_non_owner_put",
+			isSelf:     false,
+			capSharing: true,
+			req:        httptest.NewRequest("PUT", "/v0/put/foo", nil),
+			checks: checks(
+				httpStatus(http.StatusForbidden),
+				bodyContains("not owner"),
+			),
+		},
+		{
+			name:       "owner_without_cap",
+			isSelf:     true,
+			capSharing: false,
+			req:        httptest.NewRequest("PUT", "/v0/put/foo", nil),
+			checks: checks(
+				httpStatus(http.StatusForbidden),
+				bodyContains("file sharing not enabled by Tailscale admin"),
+			),
+		},
+		{
+			name:       "owner_with_cap_no_rootdir",
+			omitRoot:   true,
+			isSelf:     true,
+			capSharing: true,
+			req:        httptest.NewRequest("PUT", "/v0/put/foo", nil),
+			checks: checks(
+				httpStatus(http.StatusInternalServerError),
+				bodyContains("no rootdir"),
+			),
+		},
+		{
+			name:       "bad_method",
+			isSelf:     true,
+			capSharing: true,
+			req:        httptest.NewRequest("POST", "/v0/put/foo", nil),
+			checks: checks(
+				httpStatus(405),
+				bodyContains("expected method PUT"),
+			),
+		},
+		{
+			name:       "put_zero_length",
+			isSelf:     true,
+			capSharing: true,
+			req:        httptest.NewRequest("PUT", "/v0/put/foo", nil),
+			checks: checks(
+				httpStatus(200),
+				bodyContains("{}"),
+				fileHasSize("foo", 0),
+				fileHasContents("foo", ""),
+			),
+		},
+		{
+			name:       "put_non_zero_length_content_length",
+			isSelf:     true,
+			capSharing: true,
+			req:        httptest.NewRequest("PUT", "/v0/put/foo", strings.NewReader("contents")),
+			checks: checks(
+				httpStatus(200),
+				bodyContains("{}"),
+				fileHasSize("foo", len("contents")),
+				fileHasContents("foo", "contents"),
+			),
+		},
+		{
+			name:       "put_non_zero_length_chunked",
+			isSelf:     true,
+			capSharing: true,
+			req:        httptest.NewRequest("PUT", "/v0/put/foo", struct{ io.Reader }{strings.NewReader("contents")}),
+			checks: checks(
+				httpStatus(200),
+				bodyContains("{}"),
+				fileHasSize("foo", len("contents")),
+				fileHasContents("foo", "contents"),
+			),
+		},
+		{
+			name:       "bad_filename_partial",
+			isSelf:     true,
+			capSharing: true,
+			req:        httptest.NewRequest("PUT", "/v0/put/foo.partial", nil),
+			checks: checks(
+				httpStatus(400),
+				bodyContains("bad filename"),
+			),
+		},
+		{
+			name:       "bad_filename_dot",
+			isSelf:     true,
+			capSharing: true,
+			req:        httptest.NewRequest("PUT", "/v0/put/.", nil),
+			checks: checks(
+				httpStatus(400),
+				bodyContains("bad filename"),
+			),
+		},
+		{
+			name:       "bad_filename_empty",
+			isSelf:     true,
+			capSharing: true,
+			req:        httptest.NewRequest("PUT", "/v0/put/", nil),
+			checks: checks(
+				httpStatus(400),
+				bodyContains("empty filename"),
+			),
+		},
+		{
+			name:       "bad_filename_slash",
+			isSelf:     true,
+			capSharing: true,
+			req:        httptest.NewRequest("PUT", "/v0/put/foo/bar", nil),
+			checks: checks(
+				httpStatus(400),
+				bodyContains("directories not supported"),
+			),
+		},
+		{
+			name:       "bad_filename_encoded_dot",
+			isSelf:     true,
+			capSharing: true,
+			req:        httptest.NewRequest("PUT", "/v0/put/"+hexAll("."), nil),
+			checks: checks(
+				httpStatus(400),
+				bodyContains("bad filename"),
+			),
+		},
+		{
+			name:       "bad_filename_encoded_slash",
+			isSelf:     true,
+			capSharing: true,
+			req:        httptest.NewRequest("PUT", "/v0/put/"+hexAll("/"), nil),
+			checks: checks(
+				httpStatus(400),
+				bodyContains("bad filename"),
+			),
+		},
+		{
+			name:       "bad_filename_encoded_backslash",
+			isSelf:     true,
+			capSharing: true,
+			req:        httptest.NewRequest("PUT", "/v0/put/"+hexAll("\\"), nil),
+			checks: checks(
+				httpStatus(400),
+				bodyContains("bad filename"),
+			),
+		},
+		{
+			name:       "bad_filename_encoded_dotdot",
+			isSelf:     true,
+			capSharing: true,
+			req:        httptest.NewRequest("PUT", "/v0/put/"+hexAll(".."), nil),
+			checks: checks(
+				httpStatus(400),
+				bodyContains("bad filename"),
+			),
+		},
+		{
+			name:       "bad_filename_encoded_dotdot_out",
+			isSelf:     true,
+			capSharing: true,
+			req:        httptest.NewRequest("PUT", "/v0/put/"+hexAll("foo/../../../../../etc/passwd"), nil),
+			checks: checks(
+				httpStatus(400),
+				bodyContains("bad filename"),
+			),
+		},
+		{
+			name:       "put_spaces_and_caps",
+			isSelf:     true,
+			capSharing: true,
+			req:        httptest.NewRequest("PUT", "/v0/put/"+hexAll("Foo Bar.dat"), strings.NewReader("baz")),
+			checks: checks(
+				httpStatus(200),
+				bodyContains("{}"),
+				fileHasContents("Foo Bar.dat", "baz"),
+			),
+		},
+		{
+			name:       "put_unicode",
+			isSelf:     true,
+			capSharing: true,
+			req:        httptest.NewRequest("PUT", "/v0/put/"+hexAll("Томас и его друзья.mp3"), strings.NewReader("главный озорник")),
+			checks: checks(
+				httpStatus(200),
+				bodyContains("{}"),
+				fileHasContents("Томас и его друзья.mp3", "главный озорник"),
+			),
+		},
+		{
+			name:       "put_invalid_utf8",
+			isSelf:     true,
+			capSharing: true,
+			req:        httptest.NewRequest("PUT", "/v0/put/"+(hexAll("😜")[:3]), nil),
+			checks: checks(
+				httpStatus(400),
+				bodyContains("bad filename"),
+			),
+		},
+		{
+			name:       "put_invalid_null",
+			isSelf:     true,
+			capSharing: true,
+			req:        httptest.NewRequest("PUT", "/v0/put/%00", nil),
+			checks: checks(
+				httpStatus(400),
+				bodyContains("bad filename"),
+			),
+		},
+		{
+			name:       "put_invalid_non_printable",
+			isSelf:     true,
+			capSharing: true,
+			req:        httptest.NewRequest("PUT", "/v0/put/%01", nil),
+			checks: checks(
+				httpStatus(400),
+				bodyContains("bad filename"),
+			),
+		},
+		{
+			name:       "put_invalid_colon",
+			isSelf:     true,
+			capSharing: true,
+			req:        httptest.NewRequest("PUT", "/v0/put/"+hexAll("nul:"), nil),
+			checks: checks(
+				httpStatus(400),
+				bodyContains("bad filename"),
+			),
+		},
+		{
+			name:       "put_invalid_surrounding_whitespace",
+			isSelf:     true,
+			capSharing: true,
+			req:        httptest.NewRequest("PUT", "/v0/put/"+hexAll(" foo "), nil),
+			checks: checks(
+				httpStatus(400),
+				bodyContains("bad filename"),
+			),
+		},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			var caps []string
+			if tt.capSharing {
+				caps = append(caps, tailcfg.CapabilityFileSharing)
+			}
+			var e peerAPITestEnv
+			lb := &LocalBackend{
+				netMap: &netmap.NetworkMap{
+					SelfNode: &tailcfg.Node{
+						Capabilities: caps,
+					},
+				},
+				logf: e.logf,
+			}
+			e.ph = &peerAPIHandler{
+				isSelf: tt.isSelf,
+				peerNode: &tailcfg.Node{
+					ComputedName: "some-peer-name",
+				},
+				ps: &peerAPIServer{
+					b: lb,
+				},
+			}
+			var rootDir string
+			if !tt.omitRoot {
+				rootDir = t.TempDir()
+				e.ph.ps.rootDir = rootDir
+			}
+			e.rr = httptest.NewRecorder()
+			e.ph.ServeHTTP(e.rr, tt.req)
+			for _, f := range tt.checks {
+				f(t, &e)
+			}
+			if t.Failed() && rootDir != "" {
+				t.Logf("Contents of %s:", rootDir)
+				des, _ := fs.ReadDir(os.DirFS(rootDir), ".")
+				for _, de := range des {
+					fi, err := de.Info()
+					if err != nil {
+						t.Log(err)
+					} else {
+						t.Logf("  %v %5d %s", fi.Mode(), fi.Size(), de.Name())
+					}
+				}
+			}
+		})
+	}
+}
+
+// Windows likes to hold on to file descriptors for some indeterminate
+// amount of time after you close them and not let you delete them for
+// a bit. So test that we work around that sufficiently.
+func TestFileDeleteRace(t *testing.T) {
+	dir := t.TempDir()
+	ps := &peerAPIServer{
+		b: &LocalBackend{
+			logf: t.Logf,
+			netMap: &netmap.NetworkMap{
+				SelfNode: &tailcfg.Node{
+					Capabilities: []string{tailcfg.CapabilityFileSharing},
+				},
+			},
+		},
+		rootDir: dir,
+	}
+	ph := &peerAPIHandler{
+		isSelf: true,
+		peerNode: &tailcfg.Node{
+			ComputedName: "some-peer-name",
+		},
+		ps: ps,
+	}
+	buf := make([]byte, 2<<20)
+	for i := 0; i < 30; i++ {
+		rr := httptest.NewRecorder()
+		ph.ServeHTTP(rr, httptest.NewRequest("PUT", "/v0/put/foo.txt", bytes.NewReader(buf[:rand.Intn(len(buf))])))
+		if res := rr.Result(); res.StatusCode != 200 {
+			t.Fatal(res.Status)
+		}
+		wfs, err := ps.WaitingFiles()
+		if err != nil {
+			t.Fatal(err)
+		}
+		if len(wfs) != 1 {
+			t.Fatalf("waiting files = %d; want 1", len(wfs))
+		}
+
+		if err := ps.DeleteFile("foo.txt"); err != nil {
+			t.Fatal(err)
+		}
+		wfs, err = ps.WaitingFiles()
+		if err != nil {
+			t.Fatal(err)
+		}
+		if len(wfs) != 0 {
+			t.Fatalf("waiting files = %d; want 0", len(wfs))
+		}
+	}
+}

ipn/ipnserver/server.go

@@ -63,16 +63,6 @@ type Options struct {
 	// waits for a frontend to start it.
 	AutostartStateKey ipn.StateKey
 
-	// LegacyConfigPath optionally specifies the old-style relaynode
-	// relay.conf location. If both LegacyConfigPath and
-	// AutostartStateKey are specified and the requested state doesn't
-	// exist in the backend store, the backend migrates the config
-	// from LegacyConfigPath.
-	//
-	// TODO(danderson): remove some time after the transition to
-	// tailscaled is done.
-	LegacyConfigPath string
-
 	// 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
@@ -97,8 +87,9 @@ type Options struct {
 // server is an IPN backend and its set of 0 or more active connections
 // talking to an IPN backend.
 type server struct {
-	b    *ipnlocal.LocalBackend
-	logf logger.Logf
+	b            *ipnlocal.LocalBackend
+	logf         logger.Logf
+	backendLogID string
 	// resetOnZero is whether to call bs.Reset on transition from
 	// 1->0 connections.  That is, this is whether the backend is
 	// being run in "client mode" that requires an active GUI
@@ -296,7 +287,7 @@ func (s *server) serveConn(ctx context.Context, c net.Conn, logf logger.Logf) {
 	defer s.removeAndCloseConn(c)
 	logf("[v1] incoming control connection")
 
-	if isReadonlyConn(ci, logf) {
+	if isReadonlyConn(ci, s.b.OperatorUserID(), logf) {
 		ctx = ipn.ReadonlyContextOf(ctx)
 	}
 
@@ -322,7 +313,7 @@ func (s *server) serveConn(ctx context.Context, c net.Conn, logf logger.Logf) {
 	}
 }
 
-func isReadonlyConn(ci connIdentity, logf logger.Logf) bool {
+func isReadonlyConn(ci connIdentity, operatorUID string, logf logger.Logf) bool {
 	if runtime.GOOS == "windows" {
 		// Windows doesn't need/use this mechanism, at least yet. It
 		// has a different last-user-wins auth model.
@@ -351,6 +342,10 @@ func isReadonlyConn(ci connIdentity, logf logger.Logf) bool {
 		logf("connection from userid %v; connection from non-root user matching daemon has access", uid)
 		return rw
 	}
+	if operatorUID != "" && uid == operatorUID {
+		logf("connection from userid %v; is configured operator", uid)
+		return rw
+	}
 	var adminGroupID string
 	switch runtime.GOOS {
 	case "darwin":
@@ -444,7 +439,7 @@ func (s *server) localAPIPermissions(ci connIdentity) (read, write bool) {
 		return false, false
 	}
 	if ci.IsUnixSock {
-		return true, !isReadonlyConn(ci, logger.Discard)
+		return true, !isReadonlyConn(ci, s.b.OperatorUserID(), logger.Discard)
 	}
 	return false, false
 }
@@ -481,9 +476,7 @@ func (s *server) addConn(c net.Conn, isHTTP bool) (ci connIdentity, err error) {
 	defer func() {
 		if doReset {
 			s.logf("identity changed; resetting server")
-			s.bsMu.Lock()
-			s.bs.Reset(context.TODO())
-			s.bsMu.Unlock()
+			s.b.ResetForClientDisconnect()
 		}
 	}()
 
@@ -533,9 +526,7 @@ func (s *server) removeAndCloseConn(c net.Conn) {
 			s.logf("client disconnected; staying alive in server mode")
 		} else {
 			s.logf("client disconnected; stopping server")
-			s.bsMu.Lock()
-			s.bs.Reset(context.TODO())
-			s.bsMu.Unlock()
+			s.b.ResetForClientDisconnect()
 		}
 	}
 	c.Close()
@@ -601,6 +592,7 @@ func (s *server) writeToClients(b []byte) {
 // Run runs a Tailscale backend service.
 // The getEngine func is called repeatedly, once per connection, until it returns an engine successfully.
 func Run(ctx context.Context, logf logger.Logf, logid string, getEngine func() (wgengine.Engine, error), opts Options) error {
+	getEngine = getEngineUntilItWorksWrapper(getEngine)
 	runDone := make(chan struct{})
 	defer close(runDone)
 
@@ -610,8 +602,9 @@ func Run(ctx context.Context, logf logger.Logf, logid string, getEngine func() (
 	}
 
 	server := &server{
-		logf:        logf,
-		resetOnZero: !opts.SurviveDisconnects,
+		backendLogID: logid,
+		logf:         logf,
+		resetOnZero:  !opts.SurviveDisconnects,
 	}
 
 	// When the context is closed or when we return, whichever is first, close our listner
@@ -660,46 +653,6 @@ func Run(ctx context.Context, logf logger.Logf, logid string, getEngine func() (
 	eng, err := getEngine()
 	if err != nil {
 		logf("ipnserver: initial getEngine call: %v", err)
-
-		// Issue 1187: on Windows, in unattended mode,
-		// sometimes we try 5 times and fail to create the
-		// engine before the system's ready. Hack until the
-		// bug if fixed properly: if we're running in
-		// unattended mode on Windows, keep trying forever,
-		// waiting for the machine to be ready (networking to
-		// come up?) and then dial our own safesocket TCP
-		// listener to wake up the usual mechanism that lets
-		// us surface getEngine errors to UI clients. (We
-		// don't want to just call getEngine in a loop without
-		// the listener.Accept, as we do want to handle client
-		// connections so we can tell them about errors)
-
-		bootRaceWaitForEngine, bootRaceWaitForEngineCancel := context.WithTimeout(context.Background(), time.Minute)
-		if runtime.GOOS == "windows" && opts.AutostartStateKey != "" {
-			logf("ipnserver: in unattended mode, waiting for engine availability")
-			getEngine = getEngineUntilItWorksWrapper(getEngine)
-			// Wait for it to be ready.
-			go func() {
-				defer bootRaceWaitForEngineCancel()
-				t0 := time.Now()
-				for {
-					time.Sleep(10 * time.Second)
-					if _, err := getEngine(); err != nil {
-						logf("ipnserver: unattended mode engine load: %v", err)
-						continue
-					}
-					c, err := net.Dial("tcp", listen.Addr().String())
-					logf("ipnserver: engine created after %v; waking up Accept: Dial error: %v", time.Since(t0).Round(time.Second), err)
-					if err == nil {
-						c.Close()
-					}
-					break
-				}
-			}()
-		} else {
-			bootRaceWaitForEngineCancel()
-		}
-
 		for i := 1; ctx.Err() == nil; i++ {
 			c, err := listen.Accept()
 			if err != nil {
@@ -707,7 +660,6 @@ func Run(ctx context.Context, logf logger.Logf, logid string, getEngine func() (
 				bo.BackOff(ctx, err)
 				continue
 			}
-			<-bootRaceWaitForEngine.Done()
 			logf("ipnserver: try%d: trying getEngine again...", i)
 			eng, err = getEngine()
 			if err == nil {
@@ -752,10 +704,7 @@ func Run(ctx context.Context, logf logger.Logf, logid string, getEngine func() (
 		server.bs.GotCommand(context.TODO(), &ipn.Command{
 			Version: version.Long,
 			Start: &ipn.StartArgs{
-				Opts: ipn.Options{
-					StateKey:         opts.AutostartStateKey,
-					LegacyConfigPath: opts.LegacyConfigPath,
-				},
+				Opts: ipn.Options{StateKey: opts.AutostartStateKey},
 			},
 		})
 	}
@@ -982,7 +931,7 @@ func (psc *protoSwitchConn) Close() error {
 }
 
 func (s *server) localhostHandler(ci connIdentity) http.Handler {
-	lah := localapi.NewHandler(s.b)
+	lah := localapi.NewHandler(s.b, s.logf, s.backendLogID)
 	lah.PermitRead, lah.PermitWrite = s.localAPIPermissions(ci)
 
 	return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {

ipn/ipnstate/ipnstate.go

@@ -71,7 +71,8 @@ type PeerStatus struct {
 	OS        string // HostInfo.OS
 	UserID    tailcfg.UserID
 
-	TailAddr string // Tailscale IP
+	TailAddrDeprecated string       `json:"TailAddr"` // Tailscale IP
+	TailscaleIPs       []netaddr.IP // Tailscale IP(s) assigned to this node
 
 	// Endpoints:
 	Addrs   []string
@@ -87,7 +88,8 @@ type PeerStatus struct {
 	KeepAlive     bool
 	ExitNode      bool // true if this is the currently selected exit node.
 
-	PeerAPIURL []string
+	PeerAPIURL   []string
+	Capabilities []string `json:",omitempty"`
 
 	// ShareeNode indicates this node exists in the netmap because
 	// it's owned by a shared-to user and that node might connect
@@ -213,8 +215,11 @@ func (sb *StatusBuilder) AddPeer(peer key.Public, st *PeerStatus) {
 	if v := st.UserID; v != 0 {
 		e.UserID = v
 	}
-	if v := st.TailAddr; v != "" {
-		e.TailAddr = v
+	if v := st.TailAddrDeprecated; v != "" {
+		e.TailAddrDeprecated = v
+	}
+	if v := st.TailscaleIPs; v != nil {
+		e.TailscaleIPs = v
 	}
 	if v := st.OS; v != "" {
 		e.OS = st.OS
@@ -343,13 +348,17 @@ table tbody tr:nth-child(even) td { background-color: #f5f5f5; }
 			hostNameHTML = "<br>" + html.EscapeString(hostName)
 		}
 
+		var tailAddr string
+		if len(ps.TailscaleIPs) > 0 {
+			tailAddr = ps.TailscaleIPs[0].String()
+		}
 		f("<tr><td>%s</td><td class=acenter>%s</td>"+
 			"<td><b>%s</b>%s<div class=\"tailaddr\">%s</div></td><td class=\"acenter owner\">%s</td><td class=\"aright\">%v</td><td class=\"aright\">%v</td><td class=\"aright\">%v</td>",
 			ps.PublicKey.ShortString(),
 			osEmoji(ps.OS),
 			html.EscapeString(dnsName),
 			hostNameHTML,
-			ps.TailAddr,
+			tailAddr,
 			html.EscapeString(owner),
 			ps.RxBytes,
 			ps.TxBytes,
@@ -437,5 +446,9 @@ func sortKey(ps *PeerStatus) string {
 	if ps.HostName != "" {
 		return ps.HostName
 	}
-	return ps.TailAddr
+	// TODO(bradfitz): add PeerStatus.Less and avoid these allocs in a Less func.
+	if len(ps.TailscaleIPs) > 0 {
+		return ps.TailscaleIPs[0].String()
+	}
+	return string(ps.PublicKey[:])
 }

ipn/localapi/localapi.go

@@ -6,20 +6,40 @@
 package localapi
 
 import (
+	"crypto/rand"
+	"encoding/hex"
 	"encoding/json"
+	"errors"
+	"fmt"
 	"io"
+	"net"
 	"net/http"
+	"net/http/httputil"
+	"net/url"
+	"reflect"
 	"runtime"
 	"strconv"
+	"strings"
+	"sync"
+	"time"
 
 	"inet.af/netaddr"
+	"tailscale.com/client/tailscale/apitype"
+	"tailscale.com/ipn"
 	"tailscale.com/ipn/ipnlocal"
 	"tailscale.com/ipn/ipnstate"
 	"tailscale.com/tailcfg"
+	"tailscale.com/types/logger"
 )
 
-func NewHandler(b *ipnlocal.LocalBackend) *Handler {
-	return &Handler{b: b}
+func randHex(n int) string {
+	b := make([]byte, n)
+	rand.Read(b)
+	return hex.EncodeToString(b)
+}
+
+func NewHandler(b *ipnlocal.LocalBackend, logf logger.Logf, logID string) *Handler {
+	return &Handler{b: b, logf: logf, backendLogID: logID}
 }
 
 type Handler struct {
@@ -34,7 +54,9 @@ type Handler struct {
 	// PermitWrite is whether mutating HTTP handlers are allowed.
 	PermitWrite bool
 
-	b *ipnlocal.LocalBackend
+	b            *ipnlocal.LocalBackend
+	logf         logger.Logf
+	backendLogID string
 }
 
 func (h *Handler) ServeHTTP(w http.ResponseWriter, r *http.Request) {
@@ -53,6 +75,14 @@ func (h *Handler) ServeHTTP(w http.ResponseWriter, r *http.Request) {
 			return
 		}
 	}
+	if strings.HasPrefix(r.URL.Path, "/localapi/v0/files/") {
+		h.serveFiles(w, r)
+		return
+	}
+	if strings.HasPrefix(r.URL.Path, "/localapi/v0/file-put/") {
+		h.serveFilePut(w, r)
+		return
+	}
 	switch r.URL.Path {
 	case "/localapi/v0/whois":
 		h.serveWhoIs(w, r)
@@ -60,11 +90,38 @@ func (h *Handler) ServeHTTP(w http.ResponseWriter, r *http.Request) {
 		h.serveGoroutines(w, r)
 	case "/localapi/v0/status":
 		h.serveStatus(w, r)
-	default:
+	case "/localapi/v0/logout":
+		h.serveLogout(w, r)
+	case "/localapi/v0/prefs":
+		h.servePrefs(w, r)
+	case "/localapi/v0/check-ip-forwarding":
+		h.serveCheckIPForwarding(w, r)
+	case "/localapi/v0/bugreport":
+		h.serveBugReport(w, r)
+	case "/localapi/v0/file-targets":
+		h.serveFileTargets(w, r)
+	case "/":
 		io.WriteString(w, "tailscaled\n")
+	default:
+		http.Error(w, "404 not found", 404)
 	}
 }
 
+func (h *Handler) serveBugReport(w http.ResponseWriter, r *http.Request) {
+	if !h.PermitRead {
+		http.Error(w, "bugreport access denied", http.StatusForbidden)
+		return
+	}
+
+	logMarker := fmt.Sprintf("BUG-%v-%v-%v", h.backendLogID, time.Now().UTC().Format("20060102150405Z"), randHex(8))
+	h.logf("user bugreport: %s", logMarker)
+	if note := r.FormValue("note"); len(note) > 0 {
+		h.logf("user bugreport note: %s", note)
+	}
+	w.Header().Set("Content-Type", "text/plain")
+	fmt.Fprintln(w, logMarker)
+}
+
 func (h *Handler) serveWhoIs(w http.ResponseWriter, r *http.Request) {
 	if !h.PermitRead {
 		http.Error(w, "whois access denied", http.StatusForbidden)
@@ -88,7 +145,7 @@ func (h *Handler) serveWhoIs(w http.ResponseWriter, r *http.Request) {
 		http.Error(w, "no match for IP:port", 404)
 		return
 	}
-	res := &tailcfg.WhoIsResponse{
+	res := &apitype.WhoIsResponse{
 		Node:        n,
 		UserProfile: &u,
 	}
@@ -114,6 +171,23 @@ func (h *Handler) serveGoroutines(w http.ResponseWriter, r *http.Request) {
 	w.Write(buf)
 }
 
+func (h *Handler) serveCheckIPForwarding(w http.ResponseWriter, r *http.Request) {
+	if !h.PermitRead {
+		http.Error(w, "IP forwarding check access denied", http.StatusForbidden)
+		return
+	}
+	var warning string
+	if err := h.b.CheckIPForwarding(); err != nil {
+		warning = err.Error()
+	}
+	w.Header().Set("Content-Type", "application/json")
+	json.NewEncoder(w).Encode(struct {
+		Warning string
+	}{
+		Warning: warning,
+	})
+}
+
 func (h *Handler) serveStatus(w http.ResponseWriter, r *http.Request) {
 	if !h.PermitRead {
 		http.Error(w, "status access denied", http.StatusForbidden)
@@ -131,6 +205,203 @@ func (h *Handler) serveStatus(w http.ResponseWriter, r *http.Request) {
 	e.Encode(st)
 }
 
+func (h *Handler) serveLogout(w http.ResponseWriter, r *http.Request) {
+	if !h.PermitWrite {
+		http.Error(w, "logout access denied", http.StatusForbidden)
+		return
+	}
+	if r.Method != "POST" {
+		http.Error(w, "want POST", 400)
+		return
+	}
+	err := h.b.LogoutSync(r.Context())
+	if err == nil {
+		w.WriteHeader(http.StatusNoContent)
+		return
+	}
+	http.Error(w, err.Error(), 500)
+}
+
+func (h *Handler) servePrefs(w http.ResponseWriter, r *http.Request) {
+	if !h.PermitRead {
+		http.Error(w, "prefs access denied", http.StatusForbidden)
+		return
+	}
+	var prefs *ipn.Prefs
+	switch r.Method {
+	case "PATCH":
+		if !h.PermitWrite {
+			http.Error(w, "prefs write access denied", http.StatusForbidden)
+			return
+		}
+		mp := new(ipn.MaskedPrefs)
+		if err := json.NewDecoder(r.Body).Decode(mp); err != nil {
+			http.Error(w, err.Error(), 400)
+			return
+		}
+		var err error
+		prefs, err = h.b.EditPrefs(mp)
+		if err != nil {
+			http.Error(w, err.Error(), 400)
+			return
+		}
+	case "GET", "HEAD":
+		prefs = h.b.Prefs()
+	default:
+		http.Error(w, "unsupported method", http.StatusMethodNotAllowed)
+		return
+	}
+	w.Header().Set("Content-Type", "application/json")
+	e := json.NewEncoder(w)
+	e.SetIndent("", "\t")
+	e.Encode(prefs)
+}
+
+func (h *Handler) serveFiles(w http.ResponseWriter, r *http.Request) {
+	if !h.PermitWrite {
+		http.Error(w, "file access denied", http.StatusForbidden)
+		return
+	}
+	suffix := strings.TrimPrefix(r.URL.Path, "/localapi/v0/files/")
+	if suffix == "" {
+		if r.Method != "GET" {
+			http.Error(w, "want GET to list files", 400)
+			return
+		}
+		wfs, err := h.b.WaitingFiles()
+		if err != nil {
+			http.Error(w, err.Error(), 500)
+			return
+		}
+		w.Header().Set("Content-Type", "application/json")
+		json.NewEncoder(w).Encode(wfs)
+		return
+	}
+	name, err := url.PathUnescape(suffix)
+	if err != nil {
+		http.Error(w, "bad filename", 400)
+		return
+	}
+	if r.Method == "DELETE" {
+		if err := h.b.DeleteFile(name); err != nil {
+			http.Error(w, err.Error(), 500)
+			return
+		}
+		w.WriteHeader(http.StatusNoContent)
+		return
+	}
+	rc, size, err := h.b.OpenFile(name)
+	if err != nil {
+		http.Error(w, err.Error(), 500)
+		return
+	}
+	defer rc.Close()
+	w.Header().Set("Content-Length", fmt.Sprint(size))
+	io.Copy(w, rc)
+}
+
+func writeErrorJSON(w http.ResponseWriter, err error) {
+	if err == nil {
+		err = errors.New("unexpected nil error")
+	}
+	w.Header().Set("Content-Type", "application/json")
+	w.WriteHeader(500)
+	type E struct {
+		Error string `json:"error"`
+	}
+	json.NewEncoder(w).Encode(E{err.Error()})
+}
+
+func (h *Handler) serveFileTargets(w http.ResponseWriter, r *http.Request) {
+	if !h.PermitRead {
+		http.Error(w, "access denied", http.StatusForbidden)
+		return
+	}
+	if r.Method != "GET" {
+		http.Error(w, "want GET to list targets", 400)
+		return
+	}
+	fts, err := h.b.FileTargets()
+	if err != nil {
+		writeErrorJSON(w, err)
+		return
+	}
+	makeNonNil(&fts)
+	w.Header().Set("Content-Type", "application/json")
+	json.NewEncoder(w).Encode(fts)
+}
+
+func (h *Handler) serveFilePut(w http.ResponseWriter, r *http.Request) {
+	if !h.PermitWrite {
+		http.Error(w, "file access denied", http.StatusForbidden)
+		return
+	}
+	if r.Method != "PUT" {
+		http.Error(w, "want PUT to put file", 400)
+		return
+	}
+	fts, err := h.b.FileTargets()
+	if err != nil {
+		http.Error(w, err.Error(), 500)
+		return
+	}
+
+	upath := strings.TrimPrefix(r.URL.EscapedPath(), "/localapi/v0/file-put/")
+	slash := strings.Index(upath, "/")
+	if slash == -1 {
+		http.Error(w, "bogus URL", 400)
+		return
+	}
+	stableID, filenameEscaped := tailcfg.StableNodeID(upath[:slash]), upath[slash+1:]
+
+	var ft *apitype.FileTarget
+	for _, x := range fts {
+		if x.Node.StableID == stableID {
+			ft = x
+			break
+		}
+	}
+	if ft == nil {
+		http.Error(w, "node not found", 404)
+		return
+	}
+	dstURL, err := url.Parse(ft.PeerAPIURL)
+	if err != nil {
+		http.Error(w, "bogus peer URL", 500)
+		return
+	}
+	outReq, err := http.NewRequestWithContext(r.Context(), "PUT", "http://peer/v0/put/"+filenameEscaped, r.Body)
+	if err != nil {
+		http.Error(w, "bogus outreq", 500)
+		return
+	}
+	outReq.ContentLength = r.ContentLength
+
+	rp := httputil.NewSingleHostReverseProxy(dstURL)
+	rp.Transport = getDialPeerTransport(h.b)
+	rp.ServeHTTP(w, outReq)
+}
+
+var dialPeerTransportOnce struct {
+	sync.Once
+	v *http.Transport
+}
+
+func getDialPeerTransport(b *ipnlocal.LocalBackend) *http.Transport {
+	dialPeerTransportOnce.Do(func() {
+		t := http.DefaultTransport.(*http.Transport).Clone()
+		t.Dial = nil //lint:ignore SA1019 yes I know I'm setting it to nil defensively
+		dialer := net.Dialer{
+			Timeout:   30 * time.Second,
+			KeepAlive: 30 * time.Second,
+			Control:   b.PeerDialControlFunc(),
+		}
+		t.DialContext = dialer.DialContext
+		dialPeerTransportOnce.v = t
+	})
+	return dialPeerTransportOnce.v
+}
+
 func defBool(a string, def bool) bool {
 	if a == "" {
 		return def
@@ -141,3 +412,30 @@ func defBool(a string, def bool) bool {
 	}
 	return v
 }
+
+// makeNonNil takes a pointer to a Go data structure
+// (currently only a slice or a map) and makes sure it's non-nil for
+// JSON serialization. (In particular, JavaScript clients usually want
+// the field to be defined after they decode the JSON.)
+func makeNonNil(ptr interface{}) {
+	if ptr == nil {
+		panic("nil interface")
+	}
+	rv := reflect.ValueOf(ptr)
+	if rv.Kind() != reflect.Ptr {
+		panic(fmt.Sprintf("kind %v, not Ptr", rv.Kind()))
+	}
+	if rv.Pointer() == 0 {
+		panic("nil pointer")
+	}
+	rv = rv.Elem()
+	if rv.Pointer() != 0 {
+		return
+	}
+	switch rv.Type().Kind() {
+	case reflect.Slice:
+		rv.Set(reflect.MakeSlice(rv.Type(), 0, 0))
+	case reflect.Map:
+		rv.Set(reflect.MakeMap(rv.Type()))
+	}
+}

ipn/message.go

@@ -80,7 +80,6 @@ type Command struct {
 	Login                 *tailcfg.Oauth2Token
 	Logout                *NoArgs
 	SetPrefs              *SetPrefsArgs
-	SetWantRunning        *bool
 	RequestEngineStatus   *NoArgs
 	RequestStatus         *NoArgs
 	FakeExpireAfter       *FakeExpireAfterArgs
@@ -95,11 +94,13 @@ type BackendServer struct {
 }
 
 func NewBackendServer(logf logger.Logf, b Backend, sendNotifyMsg func(b []byte)) *BackendServer {
-	return &BackendServer{
+	bs := &BackendServer{
 		logf:          logf,
 		b:             b,
 		sendNotifyMsg: sendNotifyMsg,
 	}
+	b.SetNotifyCallback(bs.send)
+	return bs
 }
 
 func (bs *BackendServer) send(n Notify) {
@@ -142,11 +143,6 @@ func (bs *BackendServer) GotCommandMsg(ctx context.Context, b []byte) error {
 	return bs.GotCommand(ctx, cmd)
 }
 
-func (bs *BackendServer) GotFakeCommand(ctx context.Context, cmd *Command) error {
-	cmd.Version = version.Long
-	return bs.GotCommand(ctx, cmd)
-}
-
 // ErrMsgPermissionDenied is the Notify.ErrMessage value used an
 // operation was done from a user/context that didn't have permission.
 const ErrMsgPermissionDenied = "permission denied"
@@ -190,7 +186,6 @@ func (bs *BackendServer) GotCommand(ctx context.Context, cmd *Command) error {
 		return errors.New("Quit command received")
 	} else if c := cmd.Start; c != nil {
 		opts := c.Opts
-		opts.Notify = bs.send
 		return bs.b.Start(opts)
 	} else if c := cmd.StartLoginInteractive; c != nil {
 		bs.b.StartLoginInteractive()
@@ -204,9 +199,6 @@ func (bs *BackendServer) GotCommand(ctx context.Context, cmd *Command) error {
 	} else if c := cmd.SetPrefs; c != nil {
 		bs.b.SetPrefs(c.New)
 		return nil
-	} else if c := cmd.SetWantRunning; c != nil {
-		bs.b.SetWantRunning(*c)
-		return nil
 	} else if c := cmd.FakeExpireAfter; c != nil {
 		bs.b.FakeExpireAfter(c.Duration)
 		return nil
@@ -214,12 +206,6 @@ func (bs *BackendServer) GotCommand(ctx context.Context, cmd *Command) error {
 	return fmt.Errorf("BackendServer.Do: no command specified")
 }
 
-func (bs *BackendServer) Reset(ctx context.Context) error {
-	// Tell the backend we got a Logout command, which will cause it
-	// to forget all its authentication information.
-	return bs.GotFakeCommand(ctx, &Command{Logout: &NoArgs{}})
-}
-
 type BackendClient struct {
 	logf           logger.Logf
 	sendCommandMsg func(jsonb []byte)
@@ -247,7 +233,7 @@ func (bc *BackendClient) GotNotifyMsg(b []byte) {
 	}
 	n := Notify{}
 	if err := json.Unmarshal(b, &n); err != nil {
-		log.Fatalf("BackendClient.Notify: cannot decode message (length=%d)\n%#v", len(b), string(b))
+		log.Fatalf("BackendClient.Notify: cannot decode message (length=%d, %#q): %v", len(b), b, err)
 	}
 	if n.Version != version.Long && !bc.AllowVersionSkew {
 		vs := fmt.Sprintf("GotNotify: Version mismatch! frontend=%#v backend=%#v",
@@ -287,8 +273,6 @@ func (bc *BackendClient) Quit() error {
 }
 
 func (bc *BackendClient) Start(opts Options) error {
-	bc.notify = opts.Notify
-	opts.Notify = nil // server can't call our function pointer
 	bc.send(Command{Start: &StartArgs{Opts: opts}})
 	return nil // remote Start() errors must be handled remotely
 }
@@ -328,10 +312,6 @@ func (bc *BackendClient) Ping(ip string, useTSMP bool) {
 	}})
 }
 
-func (bc *BackendClient) SetWantRunning(v bool) {
-	bc.send(Command{SetWantRunning: &v})
-}
-
 // MaxMessageSize is the maximum message size, in bytes.
 const MaxMessageSize = 10 << 20
 

ipn/message_test.go

@@ -90,13 +90,11 @@ func TestClientServer(t *testing.T) {
 	bc = NewBackendClient(clogf, clientToServer)
 
 	ch := make(chan Notify, 256)
-	h, err := NewHandle(bc, clogf, Options{
+	notify := func(n Notify) { ch <- n }
+	h, err := NewHandle(bc, clogf, notify, Options{
 		Prefs: &Prefs{
 			ControlURL: "http://example.com/fake",
 		},
-		Notify: func(n Notify) {
-			ch <- n
-		},
 	})
 	if err != nil {
 		t.Fatalf("NewHandle error: %v\n", err)

ipn/policy/policy.go

@@ -6,12 +6,17 @@
 // shared between the node client & control server.
 package policy
 
-import "tailscale.com/tailcfg"
+import (
+	"tailscale.com/tailcfg"
+)
 
 // IsInterestingService reports whether service s on the given operating
 // system (a version.OS value) is an interesting enough port to report
 // to our peer nodes for discovery purposes.
 func IsInterestingService(s tailcfg.Service, os string) bool {
+	if s.Proto == "peerapi4" || s.Proto == "peerapi6" {
+		return true
+	}
 	if s.Proto != tailcfg.TCP {
 		return false
 	}

ipn/prefs.go

@@ -12,6 +12,7 @@ import (
 	"log"
 	"os"
 	"path/filepath"
+	"reflect"
 	"runtime"
 	"strings"
 
@@ -24,9 +25,18 @@ import (
 
 //go:generate go run tailscale.com/cmd/cloner -type=Prefs -output=prefs_clone.go
 
+// DefaultControlURL returns the URL base of the control plane
+// ("coordination server") for use when no explicit one is configured.
+// The default control plane is the hosted version run by Tailscale.com.
+const DefaultControlURL = "https://login.tailscale.com"
+
 // Prefs are the user modifiable settings of the Tailscale node agent.
 type Prefs struct {
 	// ControlURL is the URL of the control server to use.
+	//
+	// If empty, the default for new installs, DefaultControlURL
+	// is used. It's set non-empty once the daemon has been started
+	// for the first time.
 	ControlURL string
 
 	// RouteAll specifies whether to accept subnets advertised by
@@ -65,6 +75,10 @@ type Prefs struct {
 	ExitNodeID tailcfg.StableNodeID
 	ExitNodeIP netaddr.IP
 
+	// ExitNodeAllowLANAccess indicates whether locally accessible subnets should be
+	// routed directly or via the exit node.
+	ExitNodeAllowLANAccess bool
+
 	// CorpDNS specifies whether to install the Tailscale network's
 	// DNS configuration, if it exists.
 	CorpDNS bool
@@ -139,6 +153,10 @@ type Prefs struct {
 	// Tailscale, if at all.
 	NetfilterMode preftype.NetfilterMode
 
+	// OperatorUser is the local machine user name who is allowed to
+	// operate tailscaled without being root or using sudo.
+	OperatorUser string `json:",omitempty"`
+
 	// The Persist field is named 'Config' in the file for backward
 	// compatibility with earlier versions.
 	// TODO(apenwarr): We should move this out of here, it's not a pref.
@@ -147,6 +165,75 @@ type Prefs struct {
 	Persist *persist.Persist `json:"Config"`
 }
 
+// MaskedPrefs is a Prefs with an associated bitmask of which fields are set.
+type MaskedPrefs struct {
+	Prefs
+
+	ControlURLSet             bool `json:",omitempty"`
+	RouteAllSet               bool `json:",omitempty"`
+	AllowSingleHostsSet       bool `json:",omitempty"`
+	ExitNodeIDSet             bool `json:",omitempty"`
+	ExitNodeIPSet             bool `json:",omitempty"`
+	ExitNodeAllowLANAccessSet bool `json:",omitempty"`
+	CorpDNSSet                bool `json:",omitempty"`
+	WantRunningSet            bool `json:",omitempty"`
+	ShieldsUpSet              bool `json:",omitempty"`
+	AdvertiseTagsSet          bool `json:",omitempty"`
+	HostnameSet               bool `json:",omitempty"`
+	OSVersionSet              bool `json:",omitempty"`
+	DeviceModelSet            bool `json:",omitempty"`
+	NotepadURLsSet            bool `json:",omitempty"`
+	ForceDaemonSet            bool `json:",omitempty"`
+	AdvertiseRoutesSet        bool `json:",omitempty"`
+	NoSNATSet                 bool `json:",omitempty"`
+	NetfilterModeSet          bool `json:",omitempty"`
+	OperatorUserSet           bool `json:",omitempty"`
+}
+
+// ApplyEdits mutates p, assigning fields from m.Prefs for each MaskedPrefs
+// Set field that's true.
+func (p *Prefs) ApplyEdits(m *MaskedPrefs) {
+	if p == nil {
+		panic("can't edit nil Prefs")
+	}
+	pv := reflect.ValueOf(p).Elem()
+	mv := reflect.ValueOf(m).Elem()
+	mpv := reflect.ValueOf(&m.Prefs).Elem()
+	fields := mv.NumField()
+	for i := 1; i < fields; i++ {
+		if mv.Field(i).Bool() {
+			newFieldValue := mpv.Field(i - 1)
+			pv.Field(i - 1).Set(newFieldValue)
+		}
+	}
+}
+
+func (m *MaskedPrefs) Pretty() string {
+	if m == nil {
+		return "MaskedPrefs{<nil>}"
+	}
+	var sb strings.Builder
+	sb.WriteString("MaskedPrefs{")
+	mv := reflect.ValueOf(m).Elem()
+	mt := mv.Type()
+	mpv := reflect.ValueOf(&m.Prefs).Elem()
+	first := true
+	for i := 1; i < mt.NumField(); i++ {
+		name := mt.Field(i).Name
+		if mv.Field(i).Bool() {
+			if !first {
+				sb.WriteString(" ")
+			}
+			first = false
+			fmt.Fprintf(&sb, "%s=%#v",
+				strings.TrimSuffix(name, "Set"),
+				mpv.Field(i-1).Interface())
+		}
+	}
+	sb.WriteString("}")
+	return sb.String()
+}
+
 // IsEmpty reports whether p is nil or pointing to a Prefs zero value.
 func (p *Prefs) IsEmpty() bool { return p == nil || p.Equals(&Prefs{}) }
 
@@ -169,9 +256,9 @@ func (p *Prefs) pretty(goos string) string {
 		sb.WriteString("shields=true ")
 	}
 	if !p.ExitNodeIP.IsZero() {
-		fmt.Fprintf(&sb, "exit=%v ", p.ExitNodeIP)
+		fmt.Fprintf(&sb, "exit=%v lan=%t ", p.ExitNodeIP, p.ExitNodeAllowLANAccess)
 	} else if !p.ExitNodeID.IsZero() {
-		fmt.Fprintf(&sb, "exit=%v ", p.ExitNodeID)
+		fmt.Fprintf(&sb, "exit=%v lan=%t ", p.ExitNodeID, p.ExitNodeAllowLANAccess)
 	}
 	if len(p.AdvertiseRoutes) > 0 || goos == "linux" {
 		fmt.Fprintf(&sb, "routes=%v ", p.AdvertiseRoutes)
@@ -185,9 +272,15 @@ func (p *Prefs) pretty(goos string) string {
 	if goos == "linux" {
 		fmt.Fprintf(&sb, "nf=%v ", p.NetfilterMode)
 	}
-	if p.ControlURL != "" && p.ControlURL != "https://login.tailscale.com" {
+	if p.ControlURL != "" && p.ControlURL != DefaultControlURL {
 		fmt.Fprintf(&sb, "url=%q ", p.ControlURL)
 	}
+	if p.Hostname != "" {
+		fmt.Fprintf(&sb, "host=%q ", p.Hostname)
+	}
+	if p.OperatorUser != "" {
+		fmt.Fprintf(&sb, "op=%q ", p.OperatorUser)
+	}
 	if p.Persist != nil {
 		sb.WriteString(p.Persist.Pretty())
 	} else {
@@ -219,12 +312,14 @@ func (p *Prefs) Equals(p2 *Prefs) bool {
 		p.AllowSingleHosts == p2.AllowSingleHosts &&
 		p.ExitNodeID == p2.ExitNodeID &&
 		p.ExitNodeIP == p2.ExitNodeIP &&
+		p.ExitNodeAllowLANAccess == p2.ExitNodeAllowLANAccess &&
 		p.CorpDNS == p2.CorpDNS &&
 		p.WantRunning == p2.WantRunning &&
 		p.NotepadURLs == p2.NotepadURLs &&
 		p.ShieldsUp == p2.ShieldsUp &&
 		p.NoSNAT == p2.NoSNAT &&
 		p.NetfilterMode == p2.NetfilterMode &&
+		p.OperatorUser == p2.OperatorUser &&
 		p.Hostname == p2.Hostname &&
 		p.OSVersion == p2.OSVersion &&
 		p.DeviceModel == p2.DeviceModel &&
@@ -258,20 +353,36 @@ func compareStrings(a, b []string) bool {
 	return true
 }
 
+// NewPrefs returns the default preferences to use.
 func NewPrefs() *Prefs {
+	// Provide default values for options which might be missing
+	// from the json data for any reason. The json can still
+	// override them to false.
 	return &Prefs{
-		// Provide default values for options which might be missing
-		// from the json data for any reason. The json can still
-		// override them to false.
-		ControlURL:       "https://login.tailscale.com",
+		// ControlURL is explicitly not set to signal that
+		// it's not yet configured, which relaxes the CLI "up"
+		// safety net features. It will get set to DefaultControlURL
+		// on first up. Or, if not, DefaultControlURL will be used
+		// later anyway.
+		ControlURL: "",
+
 		RouteAll:         true,
 		AllowSingleHosts: true,
 		CorpDNS:          true,
-		WantRunning:      true,
+		WantRunning:      false,
 		NetfilterMode:    preftype.NetfilterOn,
 	}
 }
 
+// ControlURLOrDefault returns the coordination server's URL base.
+// If not configured, DefaultControlURL is returned instead.
+func (p *Prefs) ControlURLOrDefault() string {
+	if p.ControlURL != "" {
+		return p.ControlURL
+	}
+	return DefaultControlURL
+}
+
 // PrefsFromBytes deserializes Prefs from a JSON blob. If
 // enforceDefaults is true, Prefs.RouteAll and Prefs.AllowSingleHosts
 // are forced on.

ipn/prefs_clone.go

@@ -33,22 +33,24 @@ func (src *Prefs) Clone() *Prefs {
 // A compilation failure here means this code must be regenerated, with command:
 //   tailscale.com/cmd/cloner -type Prefs
 var _PrefsNeedsRegeneration = Prefs(struct {
-	ControlURL       string
-	RouteAll         bool
-	AllowSingleHosts bool
-	ExitNodeID       tailcfg.StableNodeID
-	ExitNodeIP       netaddr.IP
-	CorpDNS          bool
-	WantRunning      bool
-	ShieldsUp        bool
-	AdvertiseTags    []string
-	Hostname         string
-	OSVersion        string
-	DeviceModel      string
-	NotepadURLs      bool
-	ForceDaemon      bool
-	AdvertiseRoutes  []netaddr.IPPrefix
-	NoSNAT           bool
-	NetfilterMode    preftype.NetfilterMode
-	Persist          *persist.Persist
+	ControlURL             string
+	RouteAll               bool
+	AllowSingleHosts       bool
+	ExitNodeID             tailcfg.StableNodeID
+	ExitNodeIP             netaddr.IP
+	ExitNodeAllowLANAccess bool
+	CorpDNS                bool
+	WantRunning            bool
+	ShieldsUp              bool
+	AdvertiseTags          []string
+	Hostname               string
+	OSVersion              string
+	DeviceModel            string
+	NotepadURLs            bool
+	ForceDaemon            bool
+	AdvertiseRoutes        []netaddr.IPPrefix
+	NoSNAT                 bool
+	NetfilterMode          preftype.NetfilterMode
+	OperatorUser           string
+	Persist                *persist.Persist
 }{})

ipn/prefs_test.go

@@ -5,11 +5,13 @@
 package ipn
 
 import (
+	"encoding/json"
 	"errors"
 	"fmt"
 	"io/ioutil"
 	"os"
 	"reflect"
+	"strings"
 	"testing"
 	"time"
 
@@ -31,7 +33,28 @@ func fieldsOf(t reflect.Type) (fields []string) {
 func TestPrefsEqual(t *testing.T) {
 	tstest.PanicOnLog()
 
-	prefsHandles := []string{"ControlURL", "RouteAll", "AllowSingleHosts", "ExitNodeID", "ExitNodeIP", "CorpDNS", "WantRunning", "ShieldsUp", "AdvertiseTags", "Hostname", "OSVersion", "DeviceModel", "NotepadURLs", "ForceDaemon", "AdvertiseRoutes", "NoSNAT", "NetfilterMode", "Persist"}
+	prefsHandles := []string{
+		"ControlURL",
+		"RouteAll",
+		"AllowSingleHosts",
+		"ExitNodeID",
+		"ExitNodeIP",
+		"ExitNodeAllowLANAccess",
+		"CorpDNS",
+		"WantRunning",
+		"ShieldsUp",
+		"AdvertiseTags",
+		"Hostname",
+		"OSVersion",
+		"DeviceModel",
+		"NotepadURLs",
+		"ForceDaemon",
+		"AdvertiseRoutes",
+		"NoSNAT",
+		"NetfilterMode",
+		"OperatorUser",
+		"Persist",
+	}
 	if have := fieldsOf(reflect.TypeOf(Prefs{})); !reflect.DeepEqual(have, prefsHandles) {
 		t.Errorf("Prefs.Equal check might be out of sync\nfields: %q\nhandled: %q\n",
 			have, prefsHandles)
@@ -122,6 +145,17 @@ func TestPrefsEqual(t *testing.T) {
 			true,
 		},
 
+		{
+			&Prefs{},
+			&Prefs{ExitNodeAllowLANAccess: true},
+			false,
+		},
+		{
+			&Prefs{ExitNodeAllowLANAccess: true},
+			&Prefs{ExitNodeAllowLANAccess: true},
+			true,
+		},
+
 		{
 			&Prefs{CorpDNS: true},
 			&Prefs{CorpDNS: false},
@@ -382,14 +416,36 @@ func TestPrefsPretty(t *testing.T) {
 				ExitNodeIP: netaddr.MustParseIP("1.2.3.4"),
 			},
 			"linux",
-			`Prefs{ra=false mesh=false dns=false want=false exit=1.2.3.4 routes=[] nf=off Persist=nil}`,
+			`Prefs{ra=false mesh=false dns=false want=false exit=1.2.3.4 lan=false routes=[] nf=off Persist=nil}`,
 		},
 		{
 			Prefs{
 				ExitNodeID: tailcfg.StableNodeID("myNodeABC"),
 			},
 			"linux",
-			`Prefs{ra=false mesh=false dns=false want=false exit=myNodeABC routes=[] nf=off Persist=nil}`,
+			`Prefs{ra=false mesh=false dns=false want=false exit=myNodeABC lan=false routes=[] nf=off Persist=nil}`,
+		},
+		{
+			Prefs{
+				ExitNodeID:             tailcfg.StableNodeID("myNodeABC"),
+				ExitNodeAllowLANAccess: true,
+			},
+			"linux",
+			`Prefs{ra=false mesh=false dns=false want=false exit=myNodeABC lan=true routes=[] nf=off Persist=nil}`,
+		},
+		{
+			Prefs{
+				ExitNodeAllowLANAccess: true,
+			},
+			"linux",
+			`Prefs{ra=false mesh=false dns=false want=false routes=[] nf=off Persist=nil}`,
+		},
+		{
+			Prefs{
+				Hostname: "foo",
+			},
+			"linux",
+			`Prefs{ra=false mesh=false dns=false want=false routes=[] nf=off host="foo" Persist=nil}`,
 		},
 	}
 	for i, tt := range tests {
@@ -432,3 +488,146 @@ func TestLoadPrefsFileWithZeroInIt(t *testing.T) {
 	}
 	t.Fatalf("unexpected prefs=%#v, err=%v", p, err)
 }
+
+func TestMaskedPrefsFields(t *testing.T) {
+	have := map[string]bool{}
+	for _, f := range fieldsOf(reflect.TypeOf(Prefs{})) {
+		if f == "Persist" {
+			// This one can't be edited.
+			continue
+		}
+		have[f] = true
+	}
+	for _, f := range fieldsOf(reflect.TypeOf(MaskedPrefs{})) {
+		if f == "Prefs" {
+			continue
+		}
+		if !strings.HasSuffix(f, "Set") {
+			t.Errorf("unexpected non-/Set$/ field %q", f)
+			continue
+		}
+		bare := strings.TrimSuffix(f, "Set")
+		_, ok := have[bare]
+		if !ok {
+			t.Errorf("no corresponding Prefs.%s field for MaskedPrefs.%s", bare, f)
+			continue
+		}
+		delete(have, bare)
+	}
+	for f := range have {
+		t.Errorf("missing MaskedPrefs.%sSet for Prefs.%s", f, f)
+	}
+
+	// And also make sure they line up in the right order, which
+	// ApplyEdits assumes.
+	pt := reflect.TypeOf(Prefs{})
+	mt := reflect.TypeOf(MaskedPrefs{})
+	for i := 0; i < mt.NumField(); i++ {
+		name := mt.Field(i).Name
+		if i == 0 {
+			if name != "Prefs" {
+				t.Errorf("first field of MaskedPrefs should be Prefs")
+			}
+			continue
+		}
+		prefName := pt.Field(i - 1).Name
+		if prefName+"Set" != name {
+			t.Errorf("MaskedField[%d] = %s; want %sSet", i-1, name, prefName)
+		}
+	}
+}
+
+func TestPrefsApplyEdits(t *testing.T) {
+	tests := []struct {
+		name  string
+		prefs *Prefs
+		edit  *MaskedPrefs
+		want  *Prefs
+	}{
+		{
+			name: "no_change",
+			prefs: &Prefs{
+				Hostname: "foo",
+			},
+			edit: &MaskedPrefs{},
+			want: &Prefs{
+				Hostname: "foo",
+			},
+		},
+		{
+			name: "set1_decoy1",
+			prefs: &Prefs{
+				Hostname: "foo",
+			},
+			edit: &MaskedPrefs{
+				Prefs: Prefs{
+					Hostname:    "bar",
+					DeviceModel: "ignore-this", // not set
+				},
+				HostnameSet: true,
+			},
+			want: &Prefs{
+				Hostname: "bar",
+			},
+		},
+		{
+			name:  "set_several",
+			prefs: &Prefs{},
+			edit: &MaskedPrefs{
+				Prefs: Prefs{
+					Hostname:    "bar",
+					DeviceModel: "galaxybrain",
+				},
+				HostnameSet:    true,
+				DeviceModelSet: true,
+			},
+			want: &Prefs{
+				Hostname:    "bar",
+				DeviceModel: "galaxybrain",
+			},
+		},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got := tt.prefs.Clone()
+			got.ApplyEdits(tt.edit)
+			if !got.Equals(tt.want) {
+				gotj, _ := json.Marshal(got)
+				wantj, _ := json.Marshal(tt.want)
+				t.Errorf("fail.\n got: %s\nwant: %s\n", gotj, wantj)
+			}
+		})
+	}
+}
+
+func TestMaskedPrefsPretty(t *testing.T) {
+	tests := []struct {
+		m    *MaskedPrefs
+		want string
+	}{
+		{
+			m:    &MaskedPrefs{},
+			want: "MaskedPrefs{}",
+		},
+		{
+			m: &MaskedPrefs{
+				Prefs: Prefs{
+					Hostname:         "bar",
+					DeviceModel:      "galaxybrain",
+					AllowSingleHosts: true,
+					RouteAll:         false,
+				},
+				RouteAllSet:    true,
+				HostnameSet:    true,
+				DeviceModelSet: true,
+			},
+			want: `MaskedPrefs{RouteAll=false Hostname="bar" DeviceModel="galaxybrain"}`,
+		},
+	}
+	for i, tt := range tests {
+		got := tt.m.Pretty()
+		if got != tt.want {
+			t.Errorf("%d.\n got: %#q\nwant: %#q\n", i, got, tt.want)
+		}
+	}
+}

logtail/id.go

@@ -121,14 +121,17 @@ func (id PublicID) MarshalText() ([]byte, error) {
 }
 
 func (id *PublicID) UnmarshalText(s []byte) error {
-	b, err := hex.DecodeString(string(s))
-	if err != nil {
-		return fmt.Errorf("logtail.PublicID.UnmarshalText: %v", err)
+	if len(s) != len(id)*2 {
+		return fmt.Errorf("logtail.PublicID.UnmarshalText: invalid hex length: %d", len(s))
 	}
-	if len(b) != len(id) {
-		return fmt.Errorf("logtail.PublicID.UnmarshalText: invalid hex length: %d", len(b))
+	for i := range id {
+		a, ok1 := fromHexChar(s[i*2+0])
+		b, ok2 := fromHexChar(s[i*2+1])
+		if !ok1 || !ok2 {
+			return errors.New("invalid hex character")
+		}
+		id[i] = (a << 4) | b
 	}
-	copy(id[:], b)
 	return nil
 }
 

logtail/logtail_test.go

@@ -303,3 +303,26 @@ func TestParseAndRemoveLogLevel(t *testing.T) {
 		}
 	}
 }
+
+func TestPublicIDUnmarshalText(t *testing.T) {
+	const hexStr = "6c60a9e0e7af57170bb1347b2d477e4cbc27d4571a4923b21651456f931e3d55"
+	x := []byte(hexStr)
+
+	var id PublicID
+	if err := id.UnmarshalText(x); err != nil {
+		t.Fatal(err)
+	}
+	if id.String() != hexStr {
+		t.Errorf("String = %q; want %q", id.String(), hexStr)
+	}
+
+	n := int(testing.AllocsPerRun(1000, func() {
+		var id PublicID
+		if err := id.UnmarshalText(x); err != nil {
+			t.Fatal(err)
+		}
+	}))
+	if n != 0 {
+		t.Errorf("allocs = %v; want 0", n)
+	}
+}

net/dns/config.go

@@ -5,73 +5,119 @@
 package dns
 
 import (
-	"inet.af/netaddr"
+	"sort"
 
-	"tailscale.com/types/logger"
+	"inet.af/netaddr"
+	"tailscale.com/util/dnsname"
 )
 
-// Config is the set of parameters that uniquely determine
-// the state to which a manager should bring system DNS settings.
+// Config is a DNS configuration.
 type Config struct {
-	// Nameservers are the IP addresses of the nameservers to use.
-	Nameservers []netaddr.IP
-	// Domains are the search domains to use.
-	Domains []string
-	// PerDomain indicates whether it is preferred to use Nameservers
-	// only for DNS queries for subdomains of Domains.
-	// Note that Nameservers may still be applied to all queries
-	// if the manager does not support per-domain settings.
-	PerDomain bool
-	// Proxied indicates whether DNS requests are proxied through a dns.Resolver.
-	// This enables MagicDNS.
-	Proxied bool
+	// DefaultResolvers are the DNS resolvers to use for DNS names
+	// which aren't covered by more specific per-domain routes below.
+	// If empty, the OS's default resolvers (the ones that predate
+	// Tailscale altering the configuration) are used.
+	DefaultResolvers []netaddr.IPPort
+	// Routes maps a DNS suffix to the resolvers that should be used
+	// for queries that fall within that suffix.
+	// If a query doesn't match any entry in Routes, the
+	// DefaultResolvers are used.
+	Routes map[dnsname.FQDN][]netaddr.IPPort
+	// SearchDomains are DNS suffixes to try when expanding
+	// single-label queries.
+	SearchDomains []dnsname.FQDN
+	// Hosts maps DNS FQDNs to their IPs, which can be a mix of IPv4
+	// and IPv6.
+	// Queries matching entries in Hosts are resolved locally without
+	// recursing off-machine.
+	Hosts map[dnsname.FQDN][]netaddr.IP
+	// AuthoritativeSuffixes is a list of fully-qualified DNS suffixes
+	// for which the in-process Tailscale resolver is authoritative.
+	// Queries for names within AuthoritativeSuffixes can only be
+	// fulfilled by entries in Hosts. Queries with no match in Hosts
+	// return NXDOMAIN.
+	AuthoritativeSuffixes []dnsname.FQDN
 }
 
-// Equal determines whether its argument and receiver
-// represent equivalent DNS configurations (then DNS reconfig is a no-op).
-func (lhs Config) Equal(rhs Config) bool {
-	if lhs.Proxied != rhs.Proxied || lhs.PerDomain != rhs.PerDomain {
-		return false
-	}
+// needsAnyResolvers reports whether c requires a resolver to be set
+// at the OS level.
+func (c Config) needsOSResolver() bool {
+	return c.hasDefaultResolvers() || c.hasRoutes() || c.hasHosts()
+}
 
-	if len(lhs.Nameservers) != len(rhs.Nameservers) {
-		return false
-	}
+func (c Config) hasRoutes() bool {
+	return len(c.Routes) > 0
+}
 
-	if len(lhs.Domains) != len(rhs.Domains) {
-		return false
-	}
+// hasDefaultResolversOnly reports whether the only resolvers in c are
+// DefaultResolvers.
+func (c Config) hasDefaultResolversOnly() bool {
+	return c.hasDefaultResolvers() && !c.hasRoutes() && !c.hasHosts()
+}
 
-	// With how we perform resolution order shouldn't matter,
-	// but it is unlikely that we will encounter different orders.
-	for i, server := range lhs.Nameservers {
-		if rhs.Nameservers[i] != server {
-			return false
+func (c Config) hasDefaultResolvers() bool {
+	return len(c.DefaultResolvers) > 0
+}
+
+// singleResolverSet returns the resolvers used by c.Routes if all
+// routes use the same resolvers, or nil if multiple sets of resolvers
+// are specified.
+func (c Config) singleResolverSet() []netaddr.IPPort {
+	var first []netaddr.IPPort
+	for _, resolvers := range c.Routes {
+		if first == nil {
+			first = resolvers
+			continue
+		}
+		if !sameIPPorts(first, resolvers) {
+			return nil
 		}
 	}
+	return first
+}
 
-	// The order of domains, on the other hand, is significant.
-	for i, domain := range lhs.Domains {
-		if rhs.Domains[i] != domain {
+// hasHosts reports whether c requires resolution of MagicDNS hosts or
+// domains.
+func (c Config) hasHosts() bool {
+	return len(c.Hosts) > 0 || len(c.AuthoritativeSuffixes) > 0
+}
+
+// matchDomains returns the list of match suffixes needed by Routes,
+// AuthoritativeSuffixes. Hosts is not considered as we assume that
+// they're covered by AuthoritativeSuffixes for now.
+func (c Config) matchDomains() []dnsname.FQDN {
+	ret := make([]dnsname.FQDN, 0, len(c.Routes)+len(c.AuthoritativeSuffixes))
+	seen := map[dnsname.FQDN]bool{}
+	for _, suffix := range c.AuthoritativeSuffixes {
+		if seen[suffix] {
+			continue
+		}
+		ret = append(ret, suffix)
+		seen[suffix] = true
+	}
+	for suffix := range c.Routes {
+		if seen[suffix] {
+			continue
+		}
+		ret = append(ret, suffix)
+		seen[suffix] = true
+	}
+	sort.Slice(ret, func(i, j int) bool {
+		return ret[i].WithTrailingDot() < ret[j].WithTrailingDot()
+	})
+	return ret
+}
+
+func sameIPPorts(a, b []netaddr.IPPort) bool {
+	if len(a) != len(b) {
+		return false
+	}
+
+	for i := range a {
+		if a[i] != b[i] {
 			return false
 		}
 	}
 
 	return true
 }
-
-// ManagerConfig is the set of parameters from which
-// a manager implementation is chosen and initialized.
-type ManagerConfig struct {
-	// Logf is the logger for the manager to use.
-	// It is wrapped with a "dns: " prefix.
-	Logf logger.Logf
-	// InterfaceName is the name of the interface with which DNS settings should be associated.
-	InterfaceName string
-	// Cleanup indicates that the manager is created for cleanup only.
-	// A no-op manager will be instantiated if the system needs no cleanup.
-	Cleanup bool
-	// PerDomain indicates that a manager capable of per-domain configuration is preferred.
-	// Certain managers are per-domain only; they will not be considered if this is false.
-	PerDomain bool
-}

net/dns/debian_resolvconf.go

@@ -0,0 +1,184 @@
+// Copyright (c) 2020 Tailscale Inc & AUTHORS All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package dns
+
+import (
+	"bufio"
+	"bytes"
+	_ "embed"
+	"fmt"
+	"io/ioutil"
+	"os"
+	"os/exec"
+	"path/filepath"
+
+	"tailscale.com/atomicfile"
+	"tailscale.com/types/logger"
+)
+
+//go:embed resolvconf-workaround.sh
+var workaroundScript []byte
+
+// resolvconfConfigName is the name of the config submitted to
+// resolvconf.
+// The name starts with 'tun' in order to match the hardcoded
+// interface order in debian resolvconf, which will place this
+// configuration ahead of regular network links. In theory, this
+// doesn't matter because we then fix things up to ensure our config
+// is the only one in use, but in case that fails, this will make our
+// configuration slightly preferred.
+// The 'inet' suffix has no specific meaning, but conventionally
+// resolvconf implementations encourage adding a suffix roughly
+// indicating where the config came from, and "inet" is the "none of
+// the above" value (rather than, say, "ppp" or "dhcp").
+const resolvconfConfigName = "tun-tailscale.inet"
+
+// resolvconfLibcHookPath is the directory containing libc update
+// scripts, which are run by Debian resolvconf when /etc/resolv.conf
+// has been updated.
+const resolvconfLibcHookPath = "/etc/resolvconf/update-libc.d"
+
+// resolvconfHookPath is the name of the libc hook script we install
+// to force Tailscale's DNS config to take effect.
+var resolvconfHookPath = filepath.Join(resolvconfLibcHookPath, "tailscale")
+
+// resolvconfManager manages DNS configuration using the Debian
+// implementation of the `resolvconf` program, written by Thomas Hood.
+type resolvconfManager struct {
+	logf            logger.Logf
+	listRecordsPath string
+	interfacesDir   string
+	scriptInstalled bool // libc update script has been installed
+}
+
+func newDebianResolvconfManager(logf logger.Logf) (*resolvconfManager, error) {
+	ret := &resolvconfManager{
+		logf:            logf,
+		listRecordsPath: "/lib/resolvconf/list-records",
+		interfacesDir:   "/etc/resolvconf/run/interface", // panic fallback if nothing seems to work
+	}
+
+	if _, err := os.Stat(ret.listRecordsPath); os.IsNotExist(err) {
+		// This might be a Debian system from before the big /usr
+		// merge, try /usr instead.
+		ret.listRecordsPath = "/usr" + ret.listRecordsPath
+	}
+	// The runtime directory is currently (2020-04) canonically
+	// /etc/resolvconf/run, but the manpage is making noise about
+	// switching to /run/resolvconf and dropping the /etc path. So,
+	// let's probe the possible directories and use the first one
+	// that works.
+	for _, path := range []string{
+		"/etc/resolvconf/run/interface",
+		"/run/resolvconf/interface",
+		"/var/run/resolvconf/interface",
+	} {
+		if _, err := os.Stat(path); err == nil {
+			ret.interfacesDir = path
+			break
+		}
+	}
+	if ret.interfacesDir == "" {
+		// None of the paths seem to work, use the canonical location
+		// that the current manpage says to use.
+		ret.interfacesDir = "/etc/resolvconf/run/interfaces"
+	}
+
+	return ret, nil
+}
+
+func (m *resolvconfManager) deleteTailscaleConfig() error {
+	cmd := exec.Command("resolvconf", "-d", resolvconfConfigName)
+	out, err := cmd.CombinedOutput()
+	if err != nil {
+		return fmt.Errorf("running %s: %s", cmd, out)
+	}
+	return nil
+}
+
+func (m *resolvconfManager) SetDNS(config OSConfig) error {
+	if !m.scriptInstalled {
+		m.logf("injecting resolvconf workaround script")
+		if err := os.MkdirAll(resolvconfLibcHookPath, 0755); err != nil {
+			return err
+		}
+		if err := atomicfile.WriteFile(resolvconfHookPath, workaroundScript, 0755); err != nil {
+			return err
+		}
+		m.scriptInstalled = true
+	}
+
+	if config.IsZero() {
+		if err := m.deleteTailscaleConfig(); err != nil {
+			return err
+		}
+	} else {
+		stdin := new(bytes.Buffer)
+		writeResolvConf(stdin, config.Nameservers, config.SearchDomains) // dns_direct.go
+
+		// This resolvconf implementation doesn't support exclusive
+		// mode or interface priorities, so it will end up blending
+		// our configuration with other sources. However, this will
+		// get fixed up by the script we injected above.
+		cmd := exec.Command("resolvconf", "-a", resolvconfConfigName)
+		cmd.Stdin = stdin
+		out, err := cmd.CombinedOutput()
+		if err != nil {
+			return fmt.Errorf("running %s: %s", cmd, out)
+		}
+	}
+
+	return nil
+}
+
+func (m *resolvconfManager) SupportsSplitDNS() bool {
+	return false
+}
+
+func (m *resolvconfManager) GetBaseConfig() (OSConfig, error) {
+	var bs bytes.Buffer
+
+	cmd := exec.Command(m.listRecordsPath)
+	// list-records assumes it's being run with CWD set to the
+	// interfaces runtime dir, and returns nonsense otherwise.
+	cmd.Dir = m.interfacesDir
+	cmd.Stdout = &bs
+	if err := cmd.Run(); err != nil {
+		return OSConfig{}, err
+	}
+
+	var conf bytes.Buffer
+	sc := bufio.NewScanner(&bs)
+	for sc.Scan() {
+		if sc.Text() == resolvconfConfigName {
+			continue
+		}
+		bs, err := ioutil.ReadFile(filepath.Join(m.interfacesDir, sc.Text()))
+		if err != nil {
+			if os.IsNotExist(err) {
+				// Probably raced with a deletion, that's okay.
+				continue
+			}
+			return OSConfig{}, err
+		}
+		conf.Write(bs)
+		conf.WriteByte('\n')
+	}
+
+	return readResolv(&conf)
+}
+
+func (m *resolvconfManager) Close() error {
+	if err := m.deleteTailscaleConfig(); err != nil {
+		return err
+	}
+
+	if m.scriptInstalled {
+		m.logf("removing resolvconf workaround script")
+		os.Remove(resolvconfHookPath) // Best-effort
+	}
+
+	return nil
+}

net/dns/direct.go

@@ -2,14 +2,11 @@
 // Use of this source code is governed by a BSD-style
 // license that can be found in the LICENSE file.
 
-// +build linux freebsd openbsd
-
 package dns
 
 import (
 	"bufio"
 	"bytes"
-	"errors"
 	"fmt"
 	"io"
 	"io/ioutil"
@@ -20,16 +17,16 @@ import (
 
 	"inet.af/netaddr"
 	"tailscale.com/atomicfile"
+	"tailscale.com/util/dnsname"
 )
 
 const (
-	tsConf     = "/etc/resolv.tailscale.conf"
 	backupConf = "/etc/resolv.pre-tailscale-backup.conf"
 	resolvConf = "/etc/resolv.conf"
 )
 
 // writeResolvConf writes DNS configuration in resolv.conf format to the given writer.
-func writeResolvConf(w io.Writer, servers []netaddr.IP, domains []string) {
+func writeResolvConf(w io.Writer, servers []netaddr.IP, domains []dnsname.FQDN) {
 	io.WriteString(w, "# resolv.conf(5) file generated by tailscale\n")
 	io.WriteString(w, "# DO NOT EDIT THIS FILE BY HAND -- CHANGES WILL BE OVERWRITTEN\n\n")
 	for _, ns := range servers {
@@ -41,22 +38,14 @@ func writeResolvConf(w io.Writer, servers []netaddr.IP, domains []string) {
 		io.WriteString(w, "search")
 		for _, domain := range domains {
 			io.WriteString(w, " ")
-			io.WriteString(w, domain)
+			io.WriteString(w, domain.WithoutTrailingDot())
 		}
 		io.WriteString(w, "\n")
 	}
 }
 
-// readResolvConf reads DNS configuration from /etc/resolv.conf.
-func readResolvConf() (Config, error) {
-	var config Config
-
-	f, err := os.Open("/etc/resolv.conf")
-	if err != nil {
-		return config, err
-	}
-
-	scanner := bufio.NewScanner(f)
+func readResolv(r io.Reader) (config OSConfig, err error) {
+	scanner := bufio.NewScanner(r)
 	for scanner.Scan() {
 		line := strings.TrimSpace(scanner.Text())
 
@@ -65,7 +54,7 @@ func readResolvConf() (Config, error) {
 			nameserver = strings.TrimSpace(nameserver)
 			ip, err := netaddr.ParseIP(nameserver)
 			if err != nil {
-				return config, err
+				return OSConfig{}, err
 			}
 			config.Nameservers = append(config.Nameservers, ip)
 			continue
@@ -74,7 +63,11 @@ func readResolvConf() (Config, error) {
 		if strings.HasPrefix(line, "search") {
 			domain := strings.TrimPrefix(line, "search")
 			domain = strings.TrimSpace(domain)
-			config.Domains = append(config.Domains, domain)
+			fqdn, err := dnsname.ToFQDN(domain)
+			if err != nil {
+				return OSConfig{}, fmt.Errorf("parsing search domains %q: %w", line, err)
+			}
+			config.SearchDomains = append(config.SearchDomains, fqdn)
 			continue
 		}
 	}
@@ -82,6 +75,53 @@ func readResolvConf() (Config, error) {
 	return config, nil
 }
 
+func readResolvFile(path string) (OSConfig, error) {
+	var config OSConfig
+
+	f, err := os.Open(path)
+	if err != nil {
+		return config, err
+	}
+	defer f.Close()
+
+	return readResolv(f)
+}
+
+// readResolvConf reads DNS configuration from /etc/resolv.conf.
+func readResolvConf() (OSConfig, error) {
+	return readResolvFile(resolvConf)
+}
+
+// resolvOwner returns the apparent owner of the resolv.conf
+// configuration in bs - one of "resolvconf", "systemd-resolved" or
+// "NetworkManager", or "" if no known owner was found.
+func resolvOwner(bs []byte) string {
+	b := bytes.NewBuffer(bs)
+	for {
+		line, err := b.ReadString('\n')
+		if err != nil {
+			return ""
+		}
+		line = strings.TrimSpace(line)
+		if line == "" {
+			continue
+		}
+		if line[0] != '#' {
+			// First non-empty, non-comment line. Assume the owner
+			// isn't hiding further down.
+			return ""
+		}
+
+		if strings.Contains(line, "systemd-resolved") {
+			return "systemd-resolved"
+		} else if strings.Contains(line, "NetworkManager") {
+			return "NetworkManager"
+		} else if strings.Contains(line, "resolvconf") {
+			return "resolvconf"
+		}
+	}
+}
+
 // isResolvedRunning reports whether systemd-resolved is running on the system,
 // even if it is not managing the system DNS settings.
 func isResolvedRunning() bool {
@@ -110,75 +150,171 @@ func isResolvedRunning() bool {
 // or as cleanup if the program terminates unexpectedly.
 type directManager struct{}
 
-func newDirectManager(mconfig ManagerConfig) managerImpl {
-	return directManager{}
+func newDirectManager() (directManager, error) {
+	return directManager{}, nil
 }
 
-// Up implements managerImpl.
-func (m directManager) Up(config Config) error {
-	// Write the tsConf file.
-	buf := new(bytes.Buffer)
-	writeResolvConf(buf, config.Nameservers, config.Domains)
-	if err := atomicfile.WriteFile(tsConf, buf.Bytes(), 0644); err != nil {
-		return err
-	}
-
-	if linkPath, err := os.Readlink(resolvConf); err != nil {
-		// Remove any old backup that may exist.
-		os.Remove(backupConf)
-
-		// Backup the existing /etc/resolv.conf file.
-		contents, err := ioutil.ReadFile(resolvConf)
-		// If the original did not exist, still back up an empty file.
-		// The presence of a backup file is the way we know that Up ran.
-		if err != nil && !errors.Is(err, os.ErrNotExist) {
-			return err
-		}
-		if err := atomicfile.WriteFile(backupConf, contents, 0644); err != nil {
-			return err
-		}
-	} else if linkPath != tsConf {
-		// Backup the existing symlink.
-		os.Remove(backupConf)
-		if err := os.Symlink(linkPath, backupConf); err != nil {
-			return err
-		}
-	} else {
-		// Nothing to do, resolvConf already points to tsConf.
-		return nil
-	}
-
-	os.Remove(resolvConf)
-	if err := os.Symlink(tsConf, resolvConf); err != nil {
-		return err
-	}
-
-	if isResolvedRunning() {
-		exec.Command("systemctl", "restart", "systemd-resolved.service").Run() // Best-effort.
-	}
-
-	return nil
-}
-
-// Down implements managerImpl.
-func (m directManager) Down() error {
-	if _, err := os.Stat(backupConf); err != nil {
-		// If the backup file does not exist, then Up never ran successfully.
+// ownedByTailscale reports whether /etc/resolv.conf seems to be a
+// tailscale-managed file.
+func (m directManager) ownedByTailscale() (bool, error) {
+	st, err := os.Stat(resolvConf)
+	if err != nil {
 		if os.IsNotExist(err) {
+			return false, nil
+		}
+		return false, err
+	}
+	if !st.Mode().IsRegular() {
+		return false, nil
+	}
+	bs, err := ioutil.ReadFile(resolvConf)
+	if err != nil {
+		return false, err
+	}
+	if bytes.Contains(bs, []byte("generated by tailscale")) {
+		return true, nil
+	}
+	return false, nil
+}
+
+// backupConfig creates or updates a backup of /etc/resolv.conf, if
+// resolv.conf does not currently contain a Tailscale-managed config.
+func (m directManager) backupConfig() error {
+	if _, err := os.Stat(resolvConf); err != nil {
+		if os.IsNotExist(err) {
+			// No resolv.conf, nothing to back up. Also get rid of any
+			// existing backup file, to avoid restoring something old.
+			os.Remove(backupConf)
 			return nil
 		}
 		return err
 	}
 
-	if ln, err := os.Readlink(resolvConf); err != nil {
+	owned, err := m.ownedByTailscale()
+	if err != nil {
 		return err
-	} else if ln != tsConf {
-		return fmt.Errorf("resolv.conf is not a symlink to %s", tsConf)
 	}
+	if owned {
+		return nil
+	}
+
+	return os.Rename(resolvConf, backupConf)
+}
+
+func (m directManager) restoreBackup() error {
+	if _, err := os.Stat(backupConf); err != nil {
+		if os.IsNotExist(err) {
+			// No backup, nothing we can do.
+			return nil
+		}
+		return err
+	}
+	owned, err := m.ownedByTailscale()
+	if err != nil {
+		return err
+	}
+	if _, err := os.Stat(resolvConf); err != nil && !os.IsNotExist(err) {
+		return err
+	}
+	resolvConfExists := !os.IsNotExist(err)
+
+	if resolvConfExists && !owned {
+		// There's already a non-tailscale config in place, get rid of
+		// our backup.
+		os.Remove(backupConf)
+		return nil
+	}
+
+	// We own resolv.conf, and a backup exists.
+	if err := os.Rename(backupConf, resolvConf); err != nil {
+		return err
+	}
+
+	return nil
+}
+
+func (m directManager) SetDNS(config OSConfig) error {
+	if config.IsZero() {
+		if err := m.restoreBackup(); err != nil {
+			return err
+		}
+	} else {
+		if err := m.backupConfig(); err != nil {
+			return err
+		}
+
+		buf := new(bytes.Buffer)
+		writeResolvConf(buf, config.Nameservers, config.SearchDomains)
+		if err := atomicfile.WriteFile(resolvConf, buf.Bytes(), 0644); err != nil {
+			return err
+		}
+	}
+
+	// We might have taken over a configuration managed by resolved,
+	// in which case it will notice this on restart and gracefully
+	// start using our configuration. This shouldn't happen because we
+	// try to manage DNS through resolved when it's around, but as a
+	// best-effort fallback if we messed up the detection, try to
+	// restart resolved to make the system configuration consistent.
+	if isResolvedRunning() {
+		exec.Command("systemctl", "restart", "systemd-resolved.service").Run()
+	}
+
+	return nil
+}
+
+func (m directManager) SupportsSplitDNS() bool {
+	return false
+}
+
+func (m directManager) GetBaseConfig() (OSConfig, error) {
+	owned, err := m.ownedByTailscale()
+	if err != nil {
+		return OSConfig{}, err
+	}
+	fileToRead := resolvConf
+	if owned {
+		fileToRead = backupConf
+	}
+
+	return readResolvFile(fileToRead)
+}
+
+func (m directManager) Close() error {
+	// We used to keep a file for the tailscale config and symlinked
+	// to it, but then we stopped because /etc/resolv.conf being a
+	// symlink to surprising places breaks snaps and other sandboxing
+	// things. Clean it up if it's still there.
+	os.Remove("/etc/resolv.tailscale.conf")
+
+	if _, err := os.Stat(backupConf); err != nil {
+		if os.IsNotExist(err) {
+			// No backup, nothing we can do.
+			return nil
+		}
+		return err
+	}
+	owned, err := m.ownedByTailscale()
+	if err != nil {
+		return err
+	}
+	_, err = os.Stat(resolvConf)
+	if err != nil && !os.IsNotExist(err) {
+		return err
+	}
+	resolvConfExists := !os.IsNotExist(err)
+
+	if resolvConfExists && !owned {
+		// There's already a non-tailscale config in place, get rid of
+		// our backup.
+		os.Remove(backupConf)
+		return nil
+	}
+
+	// We own resolv.conf, and a backup exists.
 	if err := os.Rename(backupConf, resolvConf); err != nil {
 		return err
 	}
-	os.Remove(tsConf)
 
 	if isResolvedRunning() {
 		exec.Command("systemctl", "restart", "systemd-resolved.service").Run() // Best-effort.

net/dns/manager.go

@@ -5,9 +5,15 @@
 package dns
 
 import (
+	"strings"
 	"time"
 
+	"inet.af/netaddr"
+	"tailscale.com/net/dns/resolver"
+	"tailscale.com/net/tsaddr"
 	"tailscale.com/types/logger"
+	"tailscale.com/util/dnsname"
+	"tailscale.com/wgengine/monitor"
 )
 
 // We use file-ignore below instead of ignore because on some platforms,
@@ -23,78 +29,206 @@ import (
 // Such operations should be wrapped in a timeout context.
 const reconfigTimeout = time.Second
 
-type managerImpl interface {
-	// Up updates system DNS settings to match the given configuration.
-	Up(Config) error
-	// Down undoes the effects of Up.
-	// It is idempotent and performs no action if Up has never been called.
-	Down() error
-}
-
 // Manager manages system DNS settings.
 type Manager struct {
 	logf logger.Logf
 
-	impl managerImpl
+	resolver *resolver.Resolver
+	os       OSConfigurator
 
-	config  Config
-	mconfig ManagerConfig
+	config Config
 }
 
 // NewManagers created a new manager from the given config.
-func NewManager(mconfig ManagerConfig) *Manager {
-	mconfig.Logf = logger.WithPrefix(mconfig.Logf, "dns: ")
+func NewManager(logf logger.Logf, oscfg OSConfigurator, linkMon *monitor.Mon) *Manager {
+	logf = logger.WithPrefix(logf, "dns: ")
 	m := &Manager{
-		logf: mconfig.Logf,
-		impl: newManager(mconfig),
-
-		config:  Config{PerDomain: mconfig.PerDomain},
-		mconfig: mconfig,
+		logf:     logf,
+		resolver: resolver.New(logf, linkMon),
+		os:       oscfg,
 	}
-
-	m.logf("using %T", m.impl)
+	m.logf("using %T", m.os)
 	return m
 }
 
-func (m *Manager) Set(config Config) error {
-	if config.Equal(m.config) {
-		return nil
-	}
+func (m *Manager) Set(cfg Config) error {
+	m.logf("Set: %+v", cfg)
 
-	m.logf("Set: %+v", config)
-
-	if len(config.Nameservers) == 0 {
-		err := m.impl.Down()
-		// If we save the config, we will not retry next time. Only do this on success.
-		if err == nil {
-			m.config = config
-		}
+	rcfg, ocfg, err := m.compileConfig(cfg)
+	if err != nil {
 		return err
 	}
 
-	// Switching to and from per-domain mode may require a change of manager.
-	if config.PerDomain != m.config.PerDomain {
-		if err := m.impl.Down(); err != nil {
-			return err
-		}
-		m.mconfig.PerDomain = config.PerDomain
-		m.impl = newManager(m.mconfig)
-		m.logf("switched to %T", m.impl)
+	m.logf("Resolvercfg: %+v", rcfg)
+	m.logf("OScfg: %+v", ocfg)
+
+	if err := m.resolver.SetConfig(rcfg); err != nil {
+		return err
+	}
+	if err := m.os.SetDNS(ocfg); err != nil {
+		return err
 	}
 
-	err := m.impl.Up(config)
-	// If we save the config, we will not retry next time. Only do this on success.
-	if err == nil {
-		m.config = config
-	}
-
-	return err
+	return nil
 }
 
-func (m *Manager) Up() error {
-	return m.impl.Up(m.config)
+// compileConfig converts cfg into a quad-100 resolver configuration
+// and an OS-level configuration.
+func (m *Manager) compileConfig(cfg Config) (resolver.Config, OSConfig, error) {
+	// Deal with trivial configs first.
+	switch {
+	case !cfg.needsOSResolver():
+		// Set search domains, but nothing else. This also covers the
+		// case where cfg is entirely zero, in which case these
+		// configs clear all Tailscale DNS settings.
+		return resolver.Config{}, OSConfig{
+			SearchDomains: cfg.SearchDomains,
+		}, nil
+	case cfg.hasDefaultResolversOnly():
+		// Trivial CorpDNS configuration, just override the OS
+		// resolver.
+		return resolver.Config{}, OSConfig{
+			Nameservers:   toIPsOnly(cfg.DefaultResolvers),
+			SearchDomains: cfg.SearchDomains,
+		}, nil
+	case cfg.hasDefaultResolvers():
+		// Default resolvers plus other stuff always ends up proxying
+		// through quad-100.
+		rcfg := resolver.Config{
+			Routes: map[dnsname.FQDN][]netaddr.IPPort{
+				".": cfg.DefaultResolvers,
+			},
+			Hosts:        cfg.Hosts,
+			LocalDomains: cfg.AuthoritativeSuffixes,
+		}
+		for suffix, resolvers := range cfg.Routes {
+			rcfg.Routes[suffix] = resolvers
+		}
+		ocfg := OSConfig{
+			Nameservers:   []netaddr.IP{tsaddr.TailscaleServiceIP()},
+			SearchDomains: cfg.SearchDomains,
+		}
+		return rcfg, ocfg, nil
+	}
+
+	// From this point on, we're figuring out split DNS
+	// configurations. The possible cases don't return directly any
+	// more, because as a final step we have to handle the case where
+	// the OS can't do split DNS.
+	var rcfg resolver.Config
+	var ocfg OSConfig
+
+	if !cfg.hasHosts() && cfg.singleResolverSet() != nil && m.os.SupportsSplitDNS() {
+		// Split DNS configuration requested, where all split domains
+		// go to the same resolvers. We can let the OS do it.
+		return resolver.Config{}, OSConfig{
+			Nameservers:   toIPsOnly(cfg.singleResolverSet()),
+			SearchDomains: cfg.SearchDomains,
+			MatchDomains:  cfg.matchDomains(),
+		}, nil
+	}
+
+	// Split DNS configuration with either multiple upstream routes,
+	// or routes + MagicDNS, or just MagicDNS, or on an OS that cannot
+	// split-DNS. Install a split config pointing at quad-100.
+	rcfg = resolver.Config{
+		Hosts:        cfg.Hosts,
+		LocalDomains: cfg.AuthoritativeSuffixes,
+		Routes:       map[dnsname.FQDN][]netaddr.IPPort{},
+	}
+	for suffix, resolvers := range cfg.Routes {
+		rcfg.Routes[suffix] = resolvers
+	}
+	ocfg = OSConfig{
+		Nameservers:   []netaddr.IP{tsaddr.TailscaleServiceIP()},
+		SearchDomains: cfg.SearchDomains,
+	}
+
+	// If the OS can't do native split-dns, read out the underlying
+	// resolver config and blend it into our config.
+	if m.os.SupportsSplitDNS() {
+		ocfg.MatchDomains = cfg.matchDomains()
+	} else {
+		bcfg, err := m.os.GetBaseConfig()
+		if err != nil {
+			// Temporary hack to make OSes where split-DNS isn't fully
+			// implemented yet not completely crap out, but instead
+			// fall back to quad-9 as a hardcoded "backup resolver".
+			//
+			// This codepath currently only triggers when opted into
+			// the split-DNS feature server side, and when at least
+			// one search domain is something within tailscale.com, so
+			// we don't accidentally leak unstable user DNS queries to
+			// quad-9 if we accidentally go down this codepath.
+			canUseHack := false
+			for _, dom := range cfg.SearchDomains {
+				if strings.HasSuffix(dom.WithoutTrailingDot(), ".tailscale.com") {
+					canUseHack = true
+					break
+				}
+			}
+			if !canUseHack {
+				return resolver.Config{}, OSConfig{}, err
+			}
+			bcfg = OSConfig{
+				Nameservers: []netaddr.IP{netaddr.IPv4(9, 9, 9, 9)},
+			}
+		}
+		rcfg.Routes["."] = toIPPorts(bcfg.Nameservers)
+		ocfg.SearchDomains = append(ocfg.SearchDomains, bcfg.SearchDomains...)
+	}
+
+	return rcfg, ocfg, nil
+}
+
+// toIPsOnly returns only the IP portion of ipps.
+// TODO: this discards port information on the assumption that we're
+// always pointing at port 53.
+// https://github.com/tailscale/tailscale/issues/1666 tracks making
+// that not true, if we ever want to.
+func toIPsOnly(ipps []netaddr.IPPort) (ret []netaddr.IP) {
+	ret = make([]netaddr.IP, 0, len(ipps))
+	for _, ipp := range ipps {
+		ret = append(ret, ipp.IP)
+	}
+	return ret
+}
+
+func toIPPorts(ips []netaddr.IP) (ret []netaddr.IPPort) {
+	ret = make([]netaddr.IPPort, 0, len(ips))
+	for _, ip := range ips {
+		ret = append(ret, netaddr.IPPort{IP: ip, Port: 53})
+	}
+	return ret
+}
+
+func (m *Manager) EnqueueRequest(bs []byte, from netaddr.IPPort) error {
+	return m.resolver.EnqueueRequest(bs, from)
+}
+
+func (m *Manager) NextResponse() ([]byte, netaddr.IPPort, error) {
+	return m.resolver.NextResponse()
 }
 
 func (m *Manager) Down() error {
-	return m.impl.Down()
+	if err := m.os.Close(); err != nil {
+		return err
+	}
+	m.resolver.Close()
+	return nil
+}
+
+// Cleanup restores the system DNS configuration to its original state
+// in case the Tailscale daemon terminated without closing the router.
+// No other state needs to be instantiated before this runs.
+func Cleanup(logf logger.Logf, interfaceName string) {
+	oscfg, err := NewOSConfigurator(logf, interfaceName)
+	if err != nil {
+		logf("creating dns cleanup: %v", err)
+		return
+	}
+	dns := NewManager(logf, oscfg, nil)
+	if err := dns.Down(); err != nil {
+		logf("dns down: %v", err)
+	}
 }

net/dns/manager_default.go

@@ -6,9 +6,11 @@
 
 package dns
 
-func newManager(mconfig ManagerConfig) managerImpl {
+import "tailscale.com/types/logger"
+
+func NewOSConfigurator(logger.Logf, string) (OSConfigurator, error) {
 	// TODO(dmytro): on darwin, we should use a macOS-specific method such as scutil.
 	// This is currently not implemented. Editing /etc/resolv.conf does not work,
 	// as most applications use the system resolver, which disregards it.
-	return newNoopManager(mconfig)
+	return NewNoopManager()
 }

net/dns/manager_freebsd.go

@@ -4,11 +4,27 @@
 
 package dns
 
-func newManager(mconfig ManagerConfig) managerImpl {
-	switch {
-	case isResolvconfActive():
-		return newResolvconfManager(mconfig)
+import (
+	"fmt"
+	"io/ioutil"
+	"os"
+
+	"tailscale.com/types/logger"
+)
+
+func NewOSConfigurator(logf logger.Logf, _ string) (OSConfigurator, error) {
+	bs, err := ioutil.ReadFile("/etc/resolv.conf")
+	if os.IsNotExist(err) {
+		return newDirectManager()
+	}
+	if err != nil {
+		return nil, fmt.Errorf("reading /etc/resolv.conf: %w", err)
+	}
+
+	switch resolvOwner(bs) {
+	case "resolvconf":
+		return newResolvconfManager(logf)
 	default:
-		return newDirectManager(mconfig)
+		return newDirectManager()
 	}
 }

net/dns/manager_linux.go

@@ -4,24 +4,170 @@
 
 package dns
 
-func newManager(mconfig ManagerConfig) managerImpl {
-	switch {
-	// systemd-resolved should only activate per-domain.
-	case isResolvedActive() && mconfig.PerDomain:
-		if mconfig.Cleanup {
-			return newNoopManager(mconfig)
-		} else {
-			return newResolvedManager(mconfig)
+import (
+	"bytes"
+	"context"
+	"errors"
+	"fmt"
+	"io/ioutil"
+	"os"
+	"os/exec"
+	"time"
+
+	"github.com/godbus/dbus/v5"
+	"tailscale.com/types/logger"
+)
+
+type kv struct {
+	k, v string
+}
+
+func (kv kv) String() string {
+	return fmt.Sprintf("%s=%s", kv.k, kv.v)
+}
+
+func NewOSConfigurator(logf logger.Logf, interfaceName string) (ret OSConfigurator, err error) {
+	var debug []kv
+	dbg := func(k, v string) {
+		debug = append(debug, kv{k, v})
+	}
+	defer func() {
+		if ret != nil {
+			dbg("ret", fmt.Sprintf("%T", ret))
 		}
-	case isNMActive():
-		if mconfig.Cleanup {
-			return newNoopManager(mconfig)
-		} else {
-			return newNMManager(mconfig)
+		logf("dns: %v", debug)
+	}()
+
+	bs, err := ioutil.ReadFile("/etc/resolv.conf")
+	if os.IsNotExist(err) {
+		dbg("rc", "missing")
+		return newDirectManager()
+	}
+	if err != nil {
+		return nil, fmt.Errorf("reading /etc/resolv.conf: %w", err)
+	}
+
+	switch resolvOwner(bs) {
+	case "systemd-resolved":
+		dbg("rc", "resolved")
+		if err := dbusPing("org.freedesktop.resolve1", "/org/freedesktop/resolve1"); err != nil {
+			dbg("resolved", "no")
+			return newDirectManager()
 		}
-	case isResolvconfActive():
-		return newResolvconfManager(mconfig)
+		if err := dbusPing("org.freedesktop.NetworkManager", "/org/freedesktop/NetworkManager/DnsManager"); err != nil {
+			dbg("nm", "no")
+			return newResolvedManager(logf)
+		}
+		dbg("nm", "yes")
+		if err := nmIsUsingResolved(); err != nil {
+			dbg("nm-resolved", "no")
+			return newResolvedManager(logf)
+		}
+		dbg("nm-resolved", "yes")
+		return newNMManager(interfaceName)
+	case "resolvconf":
+		dbg("rc", "resolvconf")
+		if err := resolvconfSourceIsNM(bs); err == nil {
+			dbg("src-is-nm", "yes")
+			if err := dbusPing("org.freedesktop.NetworkManager", "/org/freedesktop/NetworkManager/DnsManager"); err == nil {
+				dbg("nm", "yes")
+				return newNMManager(interfaceName)
+			}
+			dbg("nm", "no")
+		}
+		dbg("src-is-nm", "no")
+		if _, err := exec.LookPath("resolvconf"); err != nil {
+			dbg("resolvconf", "no")
+			return newDirectManager()
+		}
+		dbg("resolvconf", "yes")
+		return newResolvconfManager(logf)
+	case "NetworkManager":
+		dbg("rc", "nm")
+		if err := dbusPing("org.freedesktop.NetworkManager", "/org/freedesktop/NetworkManager/DnsManager"); err != nil {
+			dbg("nm", "no")
+			return newDirectManager()
+		}
+		dbg("nm", "yes")
+		return newNMManager(interfaceName)
 	default:
-		return newDirectManager(mconfig)
+		dbg("rc", "unknown")
+		return newDirectManager()
 	}
 }
+
+func resolvconfSourceIsNM(resolvDotConf []byte) error {
+	b := bytes.NewBuffer(resolvDotConf)
+	cfg, err := readResolv(b)
+	if err != nil {
+		return fmt.Errorf("parsing /etc/resolv.conf: %w", err)
+	}
+
+	var (
+		paths = []string{
+			"/etc/resolvconf/run/interface/NetworkManager",
+			"/run/resolvconf/interface/NetworkManager",
+			"/var/run/resolvconf/interface/NetworkManager",
+			"/run/resolvconf/interfaces/NetworkManager",
+			"/var/run/resolvconf/interfaces/NetworkManager",
+		}
+		nmCfg OSConfig
+		found bool
+	)
+	for _, path := range paths {
+		nmCfg, err = readResolvFile(path)
+		if os.IsNotExist(err) {
+			continue
+		} else if err != nil {
+			return err
+		}
+		found = true
+		break
+	}
+	if !found {
+		return errors.New("NetworkManager resolvconf snippet not found")
+	}
+
+	if !nmCfg.Equal(cfg) {
+		return errors.New("NetworkManager config not applied by resolvconf")
+	}
+
+	return nil
+}
+
+func nmIsUsingResolved() error {
+	conn, err := dbus.SystemBus()
+	if err != nil {
+		// DBus probably not running.
+		return err
+	}
+
+	nm := conn.Object("org.freedesktop.NetworkManager", dbus.ObjectPath("/org/freedesktop/NetworkManager/DnsManager"))
+	v, err := nm.GetProperty("org.freedesktop.NetworkManager.DnsManager.Mode")
+	if err != nil {
+		return fmt.Errorf("getting NM mode: %w", err)
+	}
+	mode, ok := v.Value().(string)
+	if !ok {
+		return fmt.Errorf("unexpected type %T for NM DNS mode", v.Value())
+	}
+	if mode != "systemd-resolved" {
+		return errors.New("NetworkManager is not using systemd-resolved for DNS")
+	}
+	return nil
+}
+
+func dbusPing(name, objectPath string) error {
+	ctx, cancel := context.WithTimeout(context.Background(), time.Second)
+	defer cancel()
+
+	conn, err := dbus.SystemBus()
+	if err != nil {
+		// DBus probably not running.
+		return err
+	}
+
+	obj := conn.Object(name, dbus.ObjectPath(objectPath))
+	call := obj.CallWithContext(ctx, "org.freedesktop.DBus.Peer.Ping", 0)
+	return call.Err
+}

net/dns/manager_openbsd.go

@@ -4,6 +4,8 @@
 
 package dns
 
-func newManager(mconfig ManagerConfig) managerImpl {
-	return newDirectManager(mconfig)
+import "tailscale.com/types/logger"
+
+func NewOSConfigurator(logger.Logf, string) (OSConfigurator, error) {
+	return newDirectManager()
 }

net/dns/manager_test.go

@@ -0,0 +1,440 @@
+// Copyright (c) 2021 Tailscale Inc & AUTHORS All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package dns
+
+import (
+	"testing"
+
+	"github.com/google/go-cmp/cmp"
+	"github.com/google/go-cmp/cmp/cmpopts"
+	"inet.af/netaddr"
+	"tailscale.com/net/dns/resolver"
+	"tailscale.com/util/dnsname"
+)
+
+type fakeOSConfigurator struct {
+	SplitDNS   bool
+	BaseConfig OSConfig
+
+	OSConfig       OSConfig
+	ResolverConfig resolver.Config
+}
+
+func (c *fakeOSConfigurator) SetDNS(cfg OSConfig) error {
+	if !c.SplitDNS && len(cfg.MatchDomains) > 0 {
+		panic("split DNS config passed to non-split OSConfigurator")
+	}
+	c.OSConfig = cfg
+	return nil
+}
+
+func (c *fakeOSConfigurator) SetResolver(cfg resolver.Config) {
+	c.ResolverConfig = cfg
+}
+
+func (c *fakeOSConfigurator) SupportsSplitDNS() bool {
+	return c.SplitDNS
+}
+
+func (c *fakeOSConfigurator) GetBaseConfig() (OSConfig, error) {
+	return c.BaseConfig, nil
+}
+
+func (c *fakeOSConfigurator) Close() error { return nil }
+
+func TestManager(t *testing.T) {
+	// Note: these tests assume that it's safe to switch the
+	// OSConfigurator's split-dns support on and off between Set
+	// calls. Empirically this is currently true, because we reprobe
+	// the support every time we generate configs. It would be
+	// reasonable to make this unsupported as well, in which case
+	// these tests will need tweaking.
+	tests := []struct {
+		name  string
+		in    Config
+		split bool
+		bs    OSConfig
+		os    OSConfig
+		rs    resolver.Config
+	}{
+		{
+			name: "empty",
+		},
+		{
+			name: "search-only",
+			in: Config{
+				SearchDomains: fqdns("tailscale.com", "universe.tf"),
+			},
+			os: OSConfig{
+				SearchDomains: fqdns("tailscale.com", "universe.tf"),
+			},
+		},
+		{
+			name: "corp",
+			in: Config{
+				DefaultResolvers: mustIPPs("1.1.1.1:53", "9.9.9.9:53"),
+				SearchDomains:    fqdns("tailscale.com", "universe.tf"),
+			},
+			os: OSConfig{
+				Nameservers:   mustIPs("1.1.1.1", "9.9.9.9"),
+				SearchDomains: fqdns("tailscale.com", "universe.tf"),
+			},
+		},
+		{
+			name: "corp-split",
+			in: Config{
+				DefaultResolvers: mustIPPs("1.1.1.1:53", "9.9.9.9:53"),
+				SearchDomains:    fqdns("tailscale.com", "universe.tf"),
+			},
+			split: true,
+			os: OSConfig{
+				Nameservers:   mustIPs("1.1.1.1", "9.9.9.9"),
+				SearchDomains: fqdns("tailscale.com", "universe.tf"),
+			},
+		},
+		{
+			name: "corp-magic",
+			in: Config{
+				DefaultResolvers: mustIPPs("1.1.1.1:53", "9.9.9.9:53"),
+				SearchDomains:    fqdns("tailscale.com", "universe.tf"),
+				Hosts: hosts(
+					"dave.ts.com.", "1.2.3.4",
+					"bradfitz.ts.com.", "2.3.4.5"),
+				AuthoritativeSuffixes: fqdns("ts.com"),
+			},
+			os: OSConfig{
+				Nameservers:   mustIPs("100.100.100.100"),
+				SearchDomains: fqdns("tailscale.com", "universe.tf"),
+			},
+			rs: resolver.Config{
+				Routes: upstreams(".", "1.1.1.1:53", "9.9.9.9:53"),
+				Hosts: hosts(
+					"dave.ts.com.", "1.2.3.4",
+					"bradfitz.ts.com.", "2.3.4.5"),
+				LocalDomains: fqdns("ts.com."),
+			},
+		},
+		{
+			name: "corp-magic-split",
+			in: Config{
+				DefaultResolvers: mustIPPs("1.1.1.1:53", "9.9.9.9:53"),
+				SearchDomains:    fqdns("tailscale.com", "universe.tf"),
+				Hosts: hosts(
+					"dave.ts.com.", "1.2.3.4",
+					"bradfitz.ts.com.", "2.3.4.5"),
+				AuthoritativeSuffixes: fqdns("ts.com"),
+			},
+			split: true,
+			os: OSConfig{
+				Nameservers:   mustIPs("100.100.100.100"),
+				SearchDomains: fqdns("tailscale.com", "universe.tf"),
+			},
+			rs: resolver.Config{
+				Routes: upstreams(".", "1.1.1.1:53", "9.9.9.9:53"),
+				Hosts: hosts(
+					"dave.ts.com.", "1.2.3.4",
+					"bradfitz.ts.com.", "2.3.4.5"),
+				LocalDomains: fqdns("ts.com."),
+			},
+		},
+		{
+			name: "corp-routes",
+			in: Config{
+				DefaultResolvers: mustIPPs("1.1.1.1:53", "9.9.9.9:53"),
+				Routes:           upstreams("corp.com", "2.2.2.2:53"),
+				SearchDomains:    fqdns("tailscale.com", "universe.tf"),
+			},
+			os: OSConfig{
+				Nameservers:   mustIPs("100.100.100.100"),
+				SearchDomains: fqdns("tailscale.com", "universe.tf"),
+			},
+			rs: resolver.Config{
+				Routes: upstreams(
+					".", "1.1.1.1:53", "9.9.9.9:53",
+					"corp.com.", "2.2.2.2:53"),
+			},
+		},
+		{
+			name: "corp-routes-split",
+			in: Config{
+				DefaultResolvers: mustIPPs("1.1.1.1:53", "9.9.9.9:53"),
+				Routes:           upstreams("corp.com", "2.2.2.2:53"),
+				SearchDomains:    fqdns("tailscale.com", "universe.tf"),
+			},
+			split: true,
+			os: OSConfig{
+				Nameservers:   mustIPs("100.100.100.100"),
+				SearchDomains: fqdns("tailscale.com", "universe.tf"),
+			},
+			rs: resolver.Config{
+				Routes: upstreams(
+					".", "1.1.1.1:53", "9.9.9.9:53",
+					"corp.com.", "2.2.2.2:53"),
+			},
+		},
+		{
+			name: "routes",
+			in: Config{
+				Routes:        upstreams("corp.com", "2.2.2.2:53"),
+				SearchDomains: fqdns("tailscale.com", "universe.tf"),
+			},
+			bs: OSConfig{
+				Nameservers:   mustIPs("8.8.8.8"),
+				SearchDomains: fqdns("coffee.shop"),
+			},
+			os: OSConfig{
+				Nameservers:   mustIPs("100.100.100.100"),
+				SearchDomains: fqdns("tailscale.com", "universe.tf", "coffee.shop"),
+			},
+			rs: resolver.Config{
+				Routes: upstreams(
+					".", "8.8.8.8:53",
+					"corp.com.", "2.2.2.2:53"),
+			},
+		},
+		{
+			name: "routes-split",
+			in: Config{
+				Routes:        upstreams("corp.com", "2.2.2.2:53"),
+				SearchDomains: fqdns("tailscale.com", "universe.tf"),
+			},
+			split: true,
+			os: OSConfig{
+				Nameservers:   mustIPs("2.2.2.2"),
+				SearchDomains: fqdns("tailscale.com", "universe.tf"),
+				MatchDomains:  fqdns("corp.com"),
+			},
+		},
+		{
+			name: "routes-multi",
+			in: Config{
+				Routes: upstreams(
+					"corp.com", "2.2.2.2:53",
+					"bigco.net", "3.3.3.3:53"),
+				SearchDomains: fqdns("tailscale.com", "universe.tf"),
+			},
+			bs: OSConfig{
+				Nameservers:   mustIPs("8.8.8.8"),
+				SearchDomains: fqdns("coffee.shop"),
+			},
+			os: OSConfig{
+				Nameservers:   mustIPs("100.100.100.100"),
+				SearchDomains: fqdns("tailscale.com", "universe.tf", "coffee.shop"),
+			},
+			rs: resolver.Config{
+				Routes: upstreams(
+					".", "8.8.8.8:53",
+					"corp.com.", "2.2.2.2:53",
+					"bigco.net.", "3.3.3.3:53"),
+			},
+		},
+		{
+			name: "routes-multi-split",
+			in: Config{
+				Routes: upstreams(
+					"corp.com", "2.2.2.2:53",
+					"bigco.net", "3.3.3.3:53"),
+				SearchDomains: fqdns("tailscale.com", "universe.tf"),
+			},
+			split: true,
+			os: OSConfig{
+				Nameservers:   mustIPs("100.100.100.100"),
+				SearchDomains: fqdns("tailscale.com", "universe.tf"),
+				MatchDomains:  fqdns("bigco.net", "corp.com"),
+			},
+			rs: resolver.Config{
+				Routes: upstreams(
+					"corp.com.", "2.2.2.2:53",
+					"bigco.net.", "3.3.3.3:53"),
+			},
+		},
+		{
+			name: "magic",
+			in: Config{
+				Hosts: hosts(
+					"dave.ts.com.", "1.2.3.4",
+					"bradfitz.ts.com.", "2.3.4.5"),
+				AuthoritativeSuffixes: fqdns("ts.com"),
+				SearchDomains:         fqdns("tailscale.com", "universe.tf"),
+			},
+			bs: OSConfig{
+				Nameservers:   mustIPs("8.8.8.8"),
+				SearchDomains: fqdns("coffee.shop"),
+			},
+			os: OSConfig{
+				Nameservers:   mustIPs("100.100.100.100"),
+				SearchDomains: fqdns("tailscale.com", "universe.tf", "coffee.shop"),
+			},
+			rs: resolver.Config{
+				Routes: upstreams(".", "8.8.8.8:53"),
+				Hosts: hosts(
+					"dave.ts.com.", "1.2.3.4",
+					"bradfitz.ts.com.", "2.3.4.5"),
+				LocalDomains: fqdns("ts.com."),
+			},
+		},
+		{
+			name: "magic-split",
+			in: Config{
+				Hosts: hosts(
+					"dave.ts.com.", "1.2.3.4",
+					"bradfitz.ts.com.", "2.3.4.5"),
+				AuthoritativeSuffixes: fqdns("ts.com"),
+				SearchDomains:         fqdns("tailscale.com", "universe.tf"),
+			},
+			split: true,
+			os: OSConfig{
+				Nameservers:   mustIPs("100.100.100.100"),
+				SearchDomains: fqdns("tailscale.com", "universe.tf"),
+				MatchDomains:  fqdns("ts.com"),
+			},
+			rs: resolver.Config{
+				Hosts: hosts(
+					"dave.ts.com.", "1.2.3.4",
+					"bradfitz.ts.com.", "2.3.4.5"),
+				LocalDomains: fqdns("ts.com."),
+			},
+		},
+		{
+			name: "routes-magic",
+			in: Config{
+				Routes: upstreams("corp.com", "2.2.2.2:53"),
+				Hosts: hosts(
+					"dave.ts.com.", "1.2.3.4",
+					"bradfitz.ts.com.", "2.3.4.5"),
+				AuthoritativeSuffixes: fqdns("ts.com"),
+				SearchDomains:         fqdns("tailscale.com", "universe.tf"),
+			},
+			bs: OSConfig{
+				Nameservers:   mustIPs("8.8.8.8"),
+				SearchDomains: fqdns("coffee.shop"),
+			},
+			os: OSConfig{
+				Nameservers:   mustIPs("100.100.100.100"),
+				SearchDomains: fqdns("tailscale.com", "universe.tf", "coffee.shop"),
+			},
+			rs: resolver.Config{
+				Routes: upstreams(
+					"corp.com.", "2.2.2.2:53",
+					".", "8.8.8.8:53"),
+				Hosts: hosts(
+					"dave.ts.com.", "1.2.3.4",
+					"bradfitz.ts.com.", "2.3.4.5"),
+				LocalDomains: fqdns("ts.com."),
+			},
+		},
+		{
+			name: "routes-magic-split",
+			in: Config{
+				Routes: upstreams("corp.com", "2.2.2.2:53"),
+				Hosts: hosts(
+					"dave.ts.com.", "1.2.3.4",
+					"bradfitz.ts.com.", "2.3.4.5"),
+				AuthoritativeSuffixes: fqdns("ts.com"),
+				SearchDomains:         fqdns("tailscale.com", "universe.tf"),
+			},
+			split: true,
+			os: OSConfig{
+				Nameservers:   mustIPs("100.100.100.100"),
+				SearchDomains: fqdns("tailscale.com", "universe.tf"),
+				MatchDomains:  fqdns("corp.com", "ts.com"),
+			},
+			rs: resolver.Config{
+				Routes: upstreams("corp.com.", "2.2.2.2:53"),
+				Hosts: hosts(
+					"dave.ts.com.", "1.2.3.4",
+					"bradfitz.ts.com.", "2.3.4.5"),
+				LocalDomains: fqdns("ts.com."),
+			},
+		},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			f := fakeOSConfigurator{
+				SplitDNS:   test.split,
+				BaseConfig: test.bs,
+			}
+			m := NewManager(t.Logf, &f, nil)
+			m.resolver.TestOnlySetHook(f.SetResolver)
+
+			if err := m.Set(test.in); err != nil {
+				t.Fatalf("m.Set: %v", err)
+			}
+			tr := cmp.Transformer("ipStr", func(ip netaddr.IP) string { return ip.String() })
+			if diff := cmp.Diff(f.OSConfig, test.os, tr, cmpopts.EquateEmpty()); diff != "" {
+				t.Errorf("wrong OSConfig (-got+want)\n%s", diff)
+			}
+			if diff := cmp.Diff(f.ResolverConfig, test.rs, tr, cmpopts.EquateEmpty()); diff != "" {
+				t.Errorf("wrong resolver.Config (-got+want)\n%s", diff)
+			}
+		})
+	}
+}
+
+func mustIPs(strs ...string) (ret []netaddr.IP) {
+	for _, s := range strs {
+		ret = append(ret, netaddr.MustParseIP(s))
+	}
+	return ret
+}
+
+func mustIPPs(strs ...string) (ret []netaddr.IPPort) {
+	for _, s := range strs {
+		ret = append(ret, netaddr.MustParseIPPort(s))
+	}
+	return ret
+}
+
+func fqdns(strs ...string) (ret []dnsname.FQDN) {
+	for _, s := range strs {
+		fqdn, err := dnsname.ToFQDN(s)
+		if err != nil {
+			panic(err)
+		}
+		ret = append(ret, fqdn)
+	}
+	return ret
+}
+
+func hosts(strs ...string) (ret map[dnsname.FQDN][]netaddr.IP) {
+	var key dnsname.FQDN
+	ret = map[dnsname.FQDN][]netaddr.IP{}
+	for _, s := range strs {
+		if ip, err := netaddr.ParseIP(s); err == nil {
+			if key == "" {
+				panic("IP provided before name")
+			}
+			ret[key] = append(ret[key], ip)
+		} else {
+			fqdn, err := dnsname.ToFQDN(s)
+			if err != nil {
+				panic(err)
+			}
+			key = fqdn
+		}
+	}
+	return ret
+}
+
+func upstreams(strs ...string) (ret map[dnsname.FQDN][]netaddr.IPPort) {
+	var key dnsname.FQDN
+	ret = map[dnsname.FQDN][]netaddr.IPPort{}
+	for _, s := range strs {
+		if ipp, err := netaddr.ParseIPPort(s); err == nil {
+			if key == "" {
+				panic("IPPort provided before suffix")
+			}
+			ret[key] = append(ret[key], ipp)
+		} else {
+			fqdn, err := dnsname.ToFQDN(s)
+			if err != nil {
+				panic(err)
+			}
+			key = fqdn
+		}
+	}
+	return ret
+}

net/dns/manager_windows.go

@@ -5,31 +5,58 @@
 package dns
 
 import (
+	"errors"
 	"fmt"
 	"os/exec"
 	"strings"
 	"syscall"
 	"time"
 
+	"golang.org/x/sys/windows"
 	"golang.org/x/sys/windows/registry"
+	"golang.zx2c4.com/wireguard/windows/tunnel/winipcfg"
+	"inet.af/netaddr"
 	"tailscale.com/types/logger"
+	"tailscale.com/util/dnsname"
 )
 
 const (
 	ipv4RegBase = `SYSTEM\CurrentControlSet\Services\Tcpip\Parameters`
 	ipv6RegBase = `SYSTEM\CurrentControlSet\Services\Tcpip6\Parameters`
+
+	// the GUID is randomly generated. At present, Tailscale installs
+	// zero or one NRPT rules, so hardcoding a single GUID everywhere
+	// is fine.
+	nrptBase        = `SYSTEM\CurrentControlSet\services\Dnscache\Parameters\DnsPolicyConfig\{5abe529b-675b-4486-8459-25a634dacc23}`
+	nrptOverrideDNS = 0x8 // bitmask value for "use the provided override DNS resolvers"
+
+	versionKey = `SOFTWARE\Microsoft\Windows NT\CurrentVersion`
 )
 
 type windowsManager struct {
-	logf logger.Logf
-	guid string
+	logf      logger.Logf
+	guid      string
+	nrptWorks bool
 }
 
-func newManager(mconfig ManagerConfig) managerImpl {
-	return windowsManager{
-		logf: mconfig.Logf,
-		guid: mconfig.InterfaceName,
+func NewOSConfigurator(logf logger.Logf, interfaceName string) (OSConfigurator, error) {
+	ret := windowsManager{
+		logf:      logf,
+		guid:      interfaceName,
+		nrptWorks: !isWindows7(),
 	}
+
+	// Best-effort: if our NRPT rule exists, try to delete it. Unlike
+	// per-interface configuration, NRPT rules survive the unclean
+	// termination of the Tailscale process, and depending on the
+	// rule, it may prevent us from reaching login.tailscale.com to
+	// boot up. The bootstrap resolver logic will save us, but it
+	// slows down start-up a bunch.
+	if ret.nrptWorks {
+		ret.delKey(nrptBase)
+	}
+
+	return ret, nil
 }
 
 // keyOpenTimeout is how long we wait for a registry key to
@@ -38,37 +65,86 @@ func newManager(mconfig ManagerConfig) managerImpl {
 // can end up racing with that.
 const keyOpenTimeout = 20 * time.Second
 
-func setRegistryString(path, name, value string) error {
+func (m windowsManager) openKey(path string) (registry.Key, error) {
 	key, err := openKeyWait(registry.LOCAL_MACHINE, path, registry.SET_VALUE, keyOpenTimeout)
 	if err != nil {
-		return fmt.Errorf("opening %s: %w", path, err)
+		return 0, fmt.Errorf("opening %s: %w", path, err)
 	}
-	defer key.Close()
+	return key, nil
+}
 
-	err = key.SetStringValue(name, value)
-	if err != nil {
-		return fmt.Errorf("setting %s[%s]: %w", path, name, err)
+func (m windowsManager) ifPath(basePath string) string {
+	return fmt.Sprintf(`%s\Interfaces\%s`, basePath, m.guid)
+}
+
+func (m windowsManager) delKey(path string) error {
+	if err := registry.DeleteKey(registry.LOCAL_MACHINE, path); err != nil && err != registry.ErrNotExist {
+		return err
 	}
 	return nil
 }
 
-func (m windowsManager) setNameservers(basePath string, nameservers []string) error {
-	path := fmt.Sprintf(`%s\Interfaces\%s`, basePath, m.guid)
-	value := strings.Join(nameservers, ",")
-	return setRegistryString(path, "NameServer", value)
+func delValue(key registry.Key, name string) error {
+	if err := key.DeleteValue(name); err != nil && err != registry.ErrNotExist {
+		return err
+	}
+	return nil
 }
 
-func (m windowsManager) setDomains(basePath string, domains []string) error {
-	path := fmt.Sprintf(`%s\Interfaces\%s`, basePath, m.guid)
-	value := strings.Join(domains, ",")
-	return setRegistryString(path, "SearchList", value)
+// setSplitDNS configures an NRPT (Name Resolution Policy Table) rule
+// to resolve queries for domains using resolvers, rather than the
+// system's "primary" resolver.
+//
+// If no resolvers are provided, the Tailscale NRPT rule is deleted.
+func (m windowsManager) setSplitDNS(resolvers []netaddr.IP, domains []dnsname.FQDN) error {
+	if len(resolvers) == 0 {
+		return m.delKey(nrptBase)
+	}
+
+	servers := make([]string, 0, len(resolvers))
+	for _, resolver := range resolvers {
+		servers = append(servers, resolver.String())
+	}
+	doms := make([]string, 0, len(domains))
+	for _, domain := range domains {
+		// NRPT rules must have a leading dot, which is not usual for
+		// DNS search paths.
+		doms = append(doms, "."+domain.WithoutTrailingDot())
+	}
+
+	// CreateKey is actually open-or-create, which suits us fine.
+	key, _, err := registry.CreateKey(registry.LOCAL_MACHINE, nrptBase, registry.SET_VALUE)
+	if err != nil {
+		return fmt.Errorf("opening %s: %w", nrptBase, err)
+	}
+	defer key.Close()
+	if err := key.SetDWordValue("Version", 1); err != nil {
+		return err
+	}
+	if err := key.SetStringsValue("Name", doms); err != nil {
+		return err
+	}
+	if err := key.SetStringValue("GenericDNSServers", strings.Join(servers, "; ")); err != nil {
+		return err
+	}
+	if err := key.SetDWordValue("ConfigOptions", nrptOverrideDNS); err != nil {
+		return err
+	}
+
+	return nil
 }
 
-func (m windowsManager) Up(config Config) error {
+// setPrimaryDNS sets the given resolvers and domains as the Tailscale
+// interface's DNS configuration.
+// If resolvers is non-empty, those resolvers become the system's
+// "primary" resolvers.
+// domains can be set without resolvers, which just contributes new
+// paths to the global DNS search list.
+func (m windowsManager) setPrimaryDNS(resolvers []netaddr.IP, domains []dnsname.FQDN) error {
 	var ipsv4 []string
 	var ipsv6 []string
 
-	for _, ip := range config.Nameservers {
+	for _, ip := range resolvers {
 		if ip.Is4() {
 			ipsv4 = append(ipsv4, ip.String())
 		} else {
@@ -76,19 +152,111 @@ func (m windowsManager) Up(config Config) error {
 		}
 	}
 
-	if err := m.setNameservers(ipv4RegBase, ipsv4); err != nil {
+	domStrs := make([]string, 0, len(domains))
+	for _, dom := range domains {
+		domStrs = append(domStrs, dom.WithoutTrailingDot())
+	}
+
+	key4, err := m.openKey(m.ifPath(ipv4RegBase))
+	if err != nil {
 		return err
 	}
-	if err := m.setDomains(ipv4RegBase, config.Domains); err != nil {
+	defer key4.Close()
+
+	if len(ipsv4) == 0 {
+		if err := delValue(key4, "NameServer"); err != nil {
+			return err
+		}
+	} else if err := key4.SetStringValue("NameServer", strings.Join(ipsv4, ",")); err != nil {
 		return err
 	}
 
-	if err := m.setNameservers(ipv6RegBase, ipsv6); err != nil {
+	if len(domains) == 0 {
+		if err := delValue(key4, "SearchList"); err != nil {
+			return err
+		}
+	} else if err := key4.SetStringValue("SearchList", strings.Join(domStrs, ",")); err != nil {
 		return err
 	}
-	if err := m.setDomains(ipv6RegBase, config.Domains); err != nil {
+
+	key6, err := m.openKey(m.ifPath(ipv6RegBase))
+	if err != nil {
 		return err
 	}
+	defer key6.Close()
+
+	if len(ipsv6) == 0 {
+		if err := delValue(key6, "NameServer"); err != nil {
+			return err
+		}
+	} else if err := key6.SetStringValue("NameServer", strings.Join(ipsv6, ",")); err != nil {
+		return err
+	}
+
+	if len(domains) == 0 {
+		if err := delValue(key6, "SearchList"); err != nil {
+			return err
+		}
+	} else if err := key6.SetStringValue("SearchList", strings.Join(domStrs, ",")); err != nil {
+		return err
+	}
+
+	// Disable LLMNR on the Tailscale interface. We don't do
+	// multicast, and we certainly don't do LLMNR, so it's pointless
+	// to make Windows try it.
+	if err := key4.SetDWordValue("EnableMulticast", 0); err != nil {
+		return err
+	}
+	if err := key6.SetDWordValue("EnableMulticast", 0); err != nil {
+		return err
+	}
+
+	return nil
+}
+
+func (m windowsManager) SetDNS(cfg OSConfig) error {
+	// We can configure Windows DNS in one of two ways:
+	//
+	//  - In primary DNS mode, we set the NameServer and SearchList
+	//    registry keys on our interface. Because our interface metric
+	//    is very low, this turns us into the one and only "primary"
+	//    resolver for the OS, i.e. all queries flow to the
+	//    resolver(s) we specify.
+	//  - In split DNS mode, we set the Domain registry key on our
+	//    interface (which adds that domain to the global search list,
+	//    but does not contribute other DNS configuration from the
+	//    interface), and configure an NRPT (Name Resolution Policy
+	//    Table) rule to route queries for our suffixes to the
+	//    provided resolver.
+	//
+	// When switching modes, we delete all the configuration related
+	// to the other mode, so these two are an XOR.
+	//
+	// Windows actually supports much more advanced configurations as
+	// well, with arbitrary routing of hosts and suffixes to arbitrary
+	// resolvers. However, we use it in a "simple" split domain
+	// configuration only, routing one set of things to the "split"
+	// resolver and the rest to the primary.
+
+	if len(cfg.MatchDomains) == 0 {
+		if err := m.setSplitDNS(nil, nil); err != nil {
+			return err
+		}
+		if err := m.setPrimaryDNS(cfg.Nameservers, cfg.SearchDomains); err != nil {
+			return err
+		}
+	} else if !m.nrptWorks {
+		return errors.New("cannot set per-domain resolvers on Windows 7")
+	} else {
+		if err := m.setSplitDNS(cfg.Nameservers, cfg.MatchDomains); err != nil {
+			return err
+		}
+		// Still set search domains on the interface, since NRPT only
+		// handles query routing and not search domain expansion.
+		if err := m.setPrimaryDNS(nil, cfg.SearchDomains); err != nil {
+			return err
+		}
+	}
 
 	// Force DNS re-registration in Active Directory. What we actually
 	// care about is that this command invokes the undocumented hidden
@@ -97,22 +265,134 @@ func (m windowsManager) Up(config Config) error {
 	// effect.
 	//
 	// This command can take a few seconds to run, so run it async, best effort.
+	//
+	// After re-registering DNS, also flush the DNS cache to clear out
+	// any cached split-horizon queries that are no longer the correct
+	// answer.
 	go func() {
 		t0 := time.Now()
 		m.logf("running ipconfig /registerdns ...")
 		cmd := exec.Command("ipconfig", "/registerdns")
 		cmd.SysProcAttr = &syscall.SysProcAttr{HideWindow: true}
+		err := cmd.Run()
 		d := time.Since(t0).Round(time.Millisecond)
-		if err := cmd.Run(); err != nil {
+		if err != nil {
 			m.logf("error running ipconfig /registerdns after %v: %v", d, err)
 		} else {
 			m.logf("ran ipconfig /registerdns in %v", d)
 		}
+
+		t0 = time.Now()
+		m.logf("running ipconfig /registerdns ...")
+		cmd = exec.Command("ipconfig", "/flushdns")
+		cmd.SysProcAttr = &syscall.SysProcAttr{HideWindow: true}
+		err = cmd.Run()
+		d = time.Since(t0).Round(time.Millisecond)
+		if err != nil {
+			m.logf("error running ipconfig /flushdns after %v: %v", d, err)
+		} else {
+			m.logf("ran ipconfig /flushdns in %v", d)
+		}
 	}()
 
 	return nil
 }
 
-func (m windowsManager) Down() error {
-	return m.Up(Config{Nameservers: nil, Domains: nil})
+func (m windowsManager) SupportsSplitDNS() bool {
+	return m.nrptWorks
+}
+
+func (m windowsManager) Close() error {
+	return m.SetDNS(OSConfig{})
+}
+
+func (m windowsManager) GetBaseConfig() (OSConfig, error) {
+	resolvers, err := m.getBasePrimaryResolver()
+	if err != nil {
+		return OSConfig{}, err
+	}
+	return OSConfig{
+		Nameservers: resolvers,
+		// Don't return any search domains here, because even Windows
+		// 7 correctly handles blending search domains from multiple
+		// sources, and any search domains we add here will get tacked
+		// onto the Tailscale config unnecessarily.
+	}, nil
+}
+
+// getBasePrimaryResolver returns a guess of the non-Tailscale primary
+// resolver on the system.
+// It's used on Windows 7 to emulate split DNS by trying to figure out
+// what the "previous" primary resolver was. It might be wrong, or
+// incomplete.
+func (m windowsManager) getBasePrimaryResolver() (resolvers []netaddr.IP, err error) {
+	tsGUID, err := windows.GUIDFromString(m.guid)
+	if err != nil {
+		return nil, err
+	}
+	tsLUID, err := winipcfg.LUIDFromGUID(&tsGUID)
+	if err != nil {
+		return nil, err
+	}
+	ifrows, err := winipcfg.GetIPInterfaceTable(windows.AF_INET)
+	if err == windows.ERROR_NOT_FOUND {
+		// IPv4 seems disabled, try to get interface metrics from IPv6 instead.
+		ifrows, err = winipcfg.GetIPInterfaceTable(windows.AF_INET6)
+	}
+	if err != nil {
+		return nil, err
+	}
+
+	var (
+		primary winipcfg.LUID
+		best    = ^uint32(0)
+	)
+	for _, row := range ifrows {
+		if !row.Connected {
+			continue
+		}
+		if row.InterfaceLUID == tsLUID {
+			continue
+		}
+		if row.Metric < best {
+			primary = row.InterfaceLUID
+			best = row.Metric
+		}
+	}
+	if primary == 0 {
+		// No resolvers set outside of Tailscale.
+		return nil, nil
+	}
+
+	ips, err := primary.DNS()
+	if err != nil {
+		return nil, err
+	}
+	for _, stdip := range ips {
+		if ip, ok := netaddr.FromStdIP(stdip); ok {
+			resolvers = append(resolvers, ip)
+		}
+	}
+
+	return resolvers, nil
+}
+
+func isWindows7() bool {
+	key, err := registry.OpenKey(registry.LOCAL_MACHINE, versionKey, registry.READ)
+	if err != nil {
+		// Fail safe, assume Windows 7.
+		return true
+	}
+	ver, _, err := key.GetStringValue("CurrentVersion")
+	if err != nil {
+		return true
+	}
+	// Careful to not assume anything about version numbers beyond
+	// 6.3, Microsoft deprecated this registry key and locked its
+	// value to what it was in Windows 8.1. We can only use this to
+	// probe for versions before that. Good thing we only need Windows
+	// 7 (so far).
+	//
+	// And yes, Windows 7 is version 6.1. Don't ask.
+	return ver == "6.1"
 }

net/dns/map.go

@@ -1,160 +0,0 @@
-// Copyright (c) 2020 Tailscale Inc & AUTHORS All rights reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-
-package dns
-
-import (
-	"sort"
-	"strings"
-
-	"inet.af/netaddr"
-)
-
-// Map is all the data Resolver needs to resolve DNS queries within the Tailscale network.
-type Map struct {
-	// nameToIP is a mapping of Tailscale domain names to their IP addresses.
-	// For example, monitoring.tailscale.us -> 100.64.0.1.
-	nameToIP map[string]netaddr.IP
-	// ipToName is the inverse of nameToIP.
-	ipToName map[netaddr.IP]string
-	// names are the keys of nameToIP in sorted order.
-	names []string
-	// rootDomains are the domains whose subdomains should always
-	// be resolved locally to prevent leakage of sensitive names.
-	rootDomains []string // e.g. "user.provider.beta.tailscale.net."
-}
-
-// NewMap returns a new Map with name to address mapping given by nameToIP.
-//
-// rootDomains are the domains whose subdomains should always be
-// resolved locally to prevent leakage of sensitive names. They should
-// end in a period ("user-foo.tailscale.net.").
-func NewMap(initNameToIP map[string]netaddr.IP, rootDomains []string) *Map {
-	// TODO(dmytro): we have to allocate names and ipToName, but nameToIP can be avoided.
-	// It is here because control sends us names not in canonical form. Change this.
-	names := make([]string, 0, len(initNameToIP))
-	nameToIP := make(map[string]netaddr.IP, len(initNameToIP))
-	ipToName := make(map[netaddr.IP]string, len(initNameToIP))
-
-	for name, ip := range initNameToIP {
-		if len(name) == 0 {
-			// Nothing useful can be done with empty names.
-			continue
-		}
-		if name[len(name)-1] != '.' {
-			name += "."
-		}
-		names = append(names, name)
-		nameToIP[name] = ip
-		ipToName[ip] = name
-	}
-	sort.Strings(names)
-
-	return &Map{
-		nameToIP: nameToIP,
-		ipToName: ipToName,
-		names:    names,
-
-		rootDomains: rootDomains,
-	}
-}
-
-func printSingleNameIP(buf *strings.Builder, name string, ip netaddr.IP) {
-	buf.WriteString(name)
-	buf.WriteByte('\t')
-	buf.WriteString(ip.String())
-	buf.WriteByte('\n')
-}
-
-func (m *Map) Pretty() string {
-	buf := new(strings.Builder)
-	for _, name := range m.names {
-		printSingleNameIP(buf, name, m.nameToIP[name])
-	}
-	return buf.String()
-}
-
-func (m *Map) PrettyDiffFrom(old *Map) string {
-	var (
-		oldNameToIP map[string]netaddr.IP
-		newNameToIP map[string]netaddr.IP
-		oldNames    []string
-		newNames    []string
-	)
-	if old != nil {
-		oldNameToIP = old.nameToIP
-		oldNames = old.names
-	}
-	if m != nil {
-		newNameToIP = m.nameToIP
-		newNames = m.names
-	}
-
-	buf := new(strings.Builder)
-	space := func() bool {
-		return buf.Len() < (1 << 10)
-	}
-
-	for len(oldNames) > 0 && len(newNames) > 0 {
-		var name string
-
-		newName, oldName := newNames[0], oldNames[0]
-		switch {
-		case oldName < newName:
-			name = oldName
-			oldNames = oldNames[1:]
-		case oldName > newName:
-			name = newName
-			newNames = newNames[1:]
-		case oldNames[0] == newNames[0]:
-			name = oldNames[0]
-			oldNames = oldNames[1:]
-			newNames = newNames[1:]
-		}
-		if !space() {
-			continue
-		}
-
-		ipOld, inOld := oldNameToIP[name]
-		ipNew, inNew := newNameToIP[name]
-		switch {
-		case !inOld:
-			buf.WriteByte('+')
-			printSingleNameIP(buf, name, ipNew)
-		case !inNew:
-			buf.WriteByte('-')
-			printSingleNameIP(buf, name, ipOld)
-		case ipOld != ipNew:
-			buf.WriteByte('-')
-			printSingleNameIP(buf, name, ipOld)
-			buf.WriteByte('+')
-			printSingleNameIP(buf, name, ipNew)
-		}
-	}
-
-	for _, name := range oldNames {
-		if !space() {
-			break
-		}
-		if _, ok := newNameToIP[name]; !ok {
-			buf.WriteByte('-')
-			printSingleNameIP(buf, name, oldNameToIP[name])
-		}
-	}
-
-	for _, name := range newNames {
-		if !space() {
-			break
-		}
-		if _, ok := oldNameToIP[name]; !ok {
-			buf.WriteByte('+')
-			printSingleNameIP(buf, name, newNameToIP[name])
-		}
-	}
-	if !space() {
-		buf.WriteString("... [truncated]\n")
-	}
-
-	return buf.String()
-}

net/dns/map_test.go

@@ -1,156 +0,0 @@
-// Copyright (c) 2020 Tailscale Inc & AUTHORS All rights reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-
-package dns
-
-import (
-	"fmt"
-	"strings"
-	"testing"
-
-	"inet.af/netaddr"
-)
-
-func TestPretty(t *testing.T) {
-	tests := []struct {
-		name string
-		dmap *Map
-		want string
-	}{
-		{"empty", NewMap(nil, nil), ""},
-		{
-			"single",
-			NewMap(map[string]netaddr.IP{
-				"hello.ipn.dev.": netaddr.IPv4(100, 101, 102, 103),
-			}, nil),
-			"hello.ipn.dev.\t100.101.102.103\n",
-		},
-		{
-			"multiple",
-			NewMap(map[string]netaddr.IP{
-				"test1.domain.":     netaddr.IPv4(100, 101, 102, 103),
-				"test2.sub.domain.": netaddr.IPv4(100, 99, 9, 1),
-			}, nil),
-			"test1.domain.\t100.101.102.103\ntest2.sub.domain.\t100.99.9.1\n",
-		},
-	}
-
-	for _, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
-			got := tt.dmap.Pretty()
-			if tt.want != got {
-				t.Errorf("want %v; got %v", tt.want, got)
-			}
-		})
-	}
-}
-
-func TestPrettyDiffFrom(t *testing.T) {
-	tests := []struct {
-		name string
-		map1 *Map
-		map2 *Map
-		want string
-	}{
-		{
-			"from_empty",
-			nil,
-			NewMap(map[string]netaddr.IP{
-				"test1.ipn.dev.": netaddr.IPv4(100, 101, 102, 103),
-				"test2.ipn.dev.": netaddr.IPv4(100, 103, 102, 101),
-			}, nil),
-			"+test1.ipn.dev.\t100.101.102.103\n+test2.ipn.dev.\t100.103.102.101\n",
-		},
-		{
-			"equal",
-			NewMap(map[string]netaddr.IP{
-				"test1.ipn.dev.": netaddr.IPv4(100, 101, 102, 103),
-				"test2.ipn.dev.": netaddr.IPv4(100, 103, 102, 101),
-			}, nil),
-			NewMap(map[string]netaddr.IP{
-				"test2.ipn.dev.": netaddr.IPv4(100, 103, 102, 101),
-				"test1.ipn.dev.": netaddr.IPv4(100, 101, 102, 103),
-			}, nil),
-			"",
-		},
-		{
-			"changed_ip",
-			NewMap(map[string]netaddr.IP{
-				"test1.ipn.dev.": netaddr.IPv4(100, 101, 102, 103),
-				"test2.ipn.dev.": netaddr.IPv4(100, 103, 102, 101),
-			}, nil),
-			NewMap(map[string]netaddr.IP{
-				"test2.ipn.dev.": netaddr.IPv4(100, 104, 102, 101),
-				"test1.ipn.dev.": netaddr.IPv4(100, 101, 102, 103),
-			}, nil),
-			"-test2.ipn.dev.\t100.103.102.101\n+test2.ipn.dev.\t100.104.102.101\n",
-		},
-		{
-			"new_domain",
-			NewMap(map[string]netaddr.IP{
-				"test1.ipn.dev.": netaddr.IPv4(100, 101, 102, 103),
-				"test2.ipn.dev.": netaddr.IPv4(100, 103, 102, 101),
-			}, nil),
-			NewMap(map[string]netaddr.IP{
-				"test3.ipn.dev.": netaddr.IPv4(100, 105, 106, 107),
-				"test2.ipn.dev.": netaddr.IPv4(100, 103, 102, 101),
-				"test1.ipn.dev.": netaddr.IPv4(100, 101, 102, 103),
-			}, nil),
-			"+test3.ipn.dev.\t100.105.106.107\n",
-		},
-		{
-			"gone_domain",
-			NewMap(map[string]netaddr.IP{
-				"test1.ipn.dev.": netaddr.IPv4(100, 101, 102, 103),
-				"test2.ipn.dev.": netaddr.IPv4(100, 103, 102, 101),
-			}, nil),
-			NewMap(map[string]netaddr.IP{
-				"test1.ipn.dev.": netaddr.IPv4(100, 101, 102, 103),
-			}, nil),
-			"-test2.ipn.dev.\t100.103.102.101\n",
-		},
-		{
-			"mixed",
-			NewMap(map[string]netaddr.IP{
-				"test1.ipn.dev.": netaddr.IPv4(100, 101, 102, 103),
-				"test4.ipn.dev.": netaddr.IPv4(100, 107, 106, 105),
-				"test5.ipn.dev.": netaddr.IPv4(100, 64, 1, 1),
-				"test2.ipn.dev.": netaddr.IPv4(100, 103, 102, 101),
-			}, nil),
-			NewMap(map[string]netaddr.IP{
-				"test2.ipn.dev.": netaddr.IPv4(100, 104, 102, 101),
-				"test1.ipn.dev.": netaddr.IPv4(100, 100, 101, 102),
-				"test3.ipn.dev.": netaddr.IPv4(100, 64, 1, 1),
-			}, nil),
-			"-test1.ipn.dev.\t100.101.102.103\n+test1.ipn.dev.\t100.100.101.102\n" +
-				"-test2.ipn.dev.\t100.103.102.101\n+test2.ipn.dev.\t100.104.102.101\n" +
-				"+test3.ipn.dev.\t100.64.1.1\n-test4.ipn.dev.\t100.107.106.105\n-test5.ipn.dev.\t100.64.1.1\n",
-		},
-	}
-
-	for _, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
-			got := tt.map2.PrettyDiffFrom(tt.map1)
-			if tt.want != got {
-				t.Errorf("want %v; got %v", tt.want, got)
-			}
-		})
-	}
-
-	t.Run("truncated", func(t *testing.T) {
-		small := NewMap(nil, nil)
-		m := map[string]netaddr.IP{}
-		for i := 0; i < 5000; i++ {
-			m[fmt.Sprintf("host%d.ipn.dev.", i)] = netaddr.IPv4(100, 64, 1, 1)
-		}
-		veryBig := NewMap(m, nil)
-		diff := veryBig.PrettyDiffFrom(small)
-		if len(diff) > 3<<10 {
-			t.Errorf("pretty diff too large: %d bytes", len(diff))
-		}
-		if !strings.Contains(diff, "truncated") {
-			t.Errorf("big diff not truncated")
-		}
-	})
-}

net/dns/nm.go

@@ -4,70 +4,72 @@
 
 // +build linux
 
+//lint:file-ignore U1000 refactoring, temporarily unused code.
+
 package dns
 
 import (
-	"bufio"
-	"bytes"
 	"context"
 	"fmt"
-	"os"
-	"os/exec"
+	"sort"
+	"time"
 
 	"github.com/godbus/dbus/v5"
+	"inet.af/netaddr"
+	"tailscale.com/util/dnsname"
 	"tailscale.com/util/endian"
 )
 
-// isNMActive determines if NetworkManager is currently managing system DNS settings.
-func isNMActive() bool {
-	// This is somewhat tricky because NetworkManager supports a number
-	// of DNS configuration modes. In all cases, we expect it to be installed
-	// and /etc/resolv.conf to contain a mention of NetworkManager in the comments.
-	_, err := exec.LookPath("NetworkManager")
-	if err != nil {
-		return false
-	}
-
-	f, err := os.Open("/etc/resolv.conf")
-	if err != nil {
-		return false
-	}
-	defer f.Close()
-
-	scanner := bufio.NewScanner(f)
-	for scanner.Scan() {
-		line := scanner.Bytes()
-		// Look for the word "NetworkManager" until comments end.
-		if len(line) > 0 && line[0] != '#' {
-			return false
-		}
-		if bytes.Contains(line, []byte("NetworkManager")) {
-			return true
-		}
-	}
-	return false
-}
+const (
+	highestPriority = int32(-1 << 31)
+	mediumPriority  = int32(1)   // Highest priority that doesn't hard-override
+	lowerPriority   = int32(200) // lower than all builtin auto priorities
+)
 
 // nmManager uses the NetworkManager DBus API.
 type nmManager struct {
 	interfaceName string
+	manager       dbus.BusObject
+	dnsManager    dbus.BusObject
 }
 
-func newNMManager(mconfig ManagerConfig) managerImpl {
-	return nmManager{
-		interfaceName: mconfig.InterfaceName,
+func newNMManager(interfaceName string) (*nmManager, error) {
+	conn, err := dbus.SystemBus()
+	if err != nil {
+		return nil, err
 	}
+
+	return &nmManager{
+		interfaceName: interfaceName,
+		manager:       conn.Object("org.freedesktop.NetworkManager", dbus.ObjectPath("/org/freedesktop/NetworkManager")),
+		dnsManager:    conn.Object("org.freedesktop.NetworkManager", dbus.ObjectPath("/org/freedesktop/NetworkManager/DnsManager")),
+	}, nil
 }
 
 type nmConnectionSettings map[string]map[string]dbus.Variant
 
-// Up implements managerImpl.
-func (m nmManager) Up(config Config) error {
+func (m *nmManager) SetDNS(config OSConfig) error {
 	ctx, cancel := context.WithTimeout(context.Background(), reconfigTimeout)
 	defer cancel()
 
-	// conn is a shared connection whose lifecycle is managed by the dbus package.
-	// We should not interfere with that by closing it.
+	// NetworkManager only lets you set DNS settings on "active"
+	// connections, which requires an assigned IP address. This got
+	// configured before the DNS manager was invoked, but it might
+	// take a little time for the netlink notifications to propagate
+	// up. So, keep retrying for the duration of the reconfigTimeout.
+	var err error
+	for ctx.Err() == nil {
+		err = m.trySet(ctx, config)
+		if err == nil {
+			break
+		}
+		time.Sleep(10 * time.Millisecond)
+	}
+
+	return err
+}
+
+func (m *nmManager) trySet(ctx context.Context, config OSConfig) error {
 	conn, err := dbus.SystemBus()
 	if err != nil {
 		return fmt.Errorf("connecting to system bus: %w", err)
@@ -135,43 +137,79 @@ func (m nmManager) Up(config Config) error {
 		}
 	}
 
+	seen := map[dnsname.FQDN]bool{}
+	var search []string
+	for _, dom := range config.SearchDomains {
+		if seen[dom] {
+			continue
+		}
+		seen[dom] = true
+		search = append(search, dom.WithTrailingDot())
+	}
+	for _, dom := range config.MatchDomains {
+		if seen[dom] {
+			continue
+		}
+		seen[dom] = true
+		search = append(search, "~"+dom.WithTrailingDot())
+	}
+	if len(config.MatchDomains) == 0 {
+		// Non-split routing requested, add an all-domains match.
+		search = append(search, "~.")
+	}
+
+	general := settings["connection"]
+	general["llmnr"] = dbus.MakeVariant(0)
+	general["mdns"] = dbus.MakeVariant(0)
+
 	ipv4Map := settings["ipv4"]
 	ipv4Map["dns"] = dbus.MakeVariant(dnsv4)
-	ipv4Map["dns-search"] = dbus.MakeVariant(config.Domains)
+	ipv4Map["dns-search"] = dbus.MakeVariant(search)
 	// We should only request priority if we have nameservers to set.
 	if len(dnsv4) == 0 {
-		ipv4Map["dns-priority"] = dbus.MakeVariant(100)
+		ipv4Map["dns-priority"] = dbus.MakeVariant(lowerPriority)
+	} else if len(config.MatchDomains) > 0 {
+		// Set a fairly high priority, but don't override all other
+		// configs when in split-DNS mode.
+		ipv4Map["dns-priority"] = dbus.MakeVariant(mediumPriority)
 	} else {
-		// dns-priority = -1 ensures that we have priority
-		// over other interfaces, except those exploiting this same trick.
-		// Ref: https://bugs.launchpad.net/ubuntu/+source/network-manager/+bug/1211110/comments/92.
-		ipv4Map["dns-priority"] = dbus.MakeVariant(-1)
+		// Negative priority means only the settings from the most
+		// negative connection get used. The way this mixes with
+		// per-domain routing is unclear, but it _seems_ that the
+		// priority applies after routing has found possible
+		// candidates for a resolution.
+		ipv4Map["dns-priority"] = dbus.MakeVariant(highestPriority)
 	}
-	// In principle, we should not need set this to true,
-	// as our interface does not configure any automatic DNS settings (presumably via DHCP).
-	// All the same, better to be safe.
-	ipv4Map["ignore-auto-dns"] = dbus.MakeVariant(true)
 
 	ipv6Map := settings["ipv6"]
-	// This is a hack.
-	// Methods "disabled", "ignore", "link-local" (IPv6 default) prevent us from setting DNS.
-	// It seems that our only recourse is "manual" or "auto".
-	// "manual" requires addresses, so we use "auto", which will assign us a random IPv6 /64.
+	// In IPv6 settings, you're only allowed to provide additional
+	// static DNS settings in "auto" (SLAAC) or "manual" mode. In
+	// "manual" mode you also have to specify IP addresses, so we use
+	// "auto".
+	//
+	// NM actually documents that to set just DNS servers, you should
+	// use "auto" mode and then set ignore auto routes and DNS, which
+	// basically means "autoconfigure but ignore any autoconfiguration
+	// results you might get". As a safety, we also say that
+	// NetworkManager should never try to make us the default route
+	// (none of its business anyway, we handle our own default
+	// routing).
 	ipv6Map["method"] = dbus.MakeVariant("auto")
-	// Our IPv6 config is a fake, so it should never become the default route.
-	ipv6Map["never-default"] = dbus.MakeVariant(true)
-	// Moreover, we should ignore all autoconfigured routes (hopefully none), as they are bogus.
 	ipv6Map["ignore-auto-routes"] = dbus.MakeVariant(true)
-
-	// Finally, set the actual DNS config.
-	ipv6Map["dns"] = dbus.MakeVariant(dnsv6)
-	ipv6Map["dns-search"] = dbus.MakeVariant(config.Domains)
-	if len(dnsv6) == 0 {
-		ipv6Map["dns-priority"] = dbus.MakeVariant(100)
-	} else {
-		ipv6Map["dns-priority"] = dbus.MakeVariant(-1)
-	}
 	ipv6Map["ignore-auto-dns"] = dbus.MakeVariant(true)
+	ipv6Map["never-default"] = dbus.MakeVariant(true)
+
+	ipv6Map["dns"] = dbus.MakeVariant(dnsv6)
+	ipv6Map["dns-search"] = dbus.MakeVariant(search)
+	if len(dnsv6) == 0 {
+		ipv6Map["dns-priority"] = dbus.MakeVariant(lowerPriority)
+	} else if len(config.MatchDomains) > 0 {
+		// Set a fairly high priority, but don't override all other
+		// configs when in split-DNS mode.
+		ipv6Map["dns-priority"] = dbus.MakeVariant(mediumPriority)
+	} else {
+		ipv6Map["dns-priority"] = dbus.MakeVariant(highestPriority)
+	}
 
 	// deprecatedProperties are the properties in interface settings
 	// that are deprecated by NetworkManager.
@@ -188,18 +226,134 @@ func (m nmManager) Up(config Config) error {
 		delete(ipv6Map, property)
 	}
 
-	err = device.CallWithContext(
-		ctx, "org.freedesktop.NetworkManager.Device.Reapply", 0,
-		settings, version, uint32(0),
-	).Store()
-	if err != nil {
+	if call := device.CallWithContext(ctx, "org.freedesktop.NetworkManager.Device.Reapply", 0, settings, version, uint32(0)); call.Err != nil {
 		return fmt.Errorf("reapply: %w", err)
 	}
 
 	return nil
 }
 
-// Down implements managerImpl.
-func (m nmManager) Down() error {
-	return m.Up(Config{Nameservers: nil, Domains: nil})
+func (m *nmManager) SupportsSplitDNS() bool {
+	var mode string
+	v, err := m.dnsManager.GetProperty("org.freedesktop.NetworkManager.DnsManager.Mode")
+	if err != nil {
+		return false
+	}
+	mode, ok := v.Value().(string)
+	if !ok {
+		return false
+	}
+
+	// Per NM's documentation, it only does split-DNS when it's
+	// programming dnsmasq or systemd-resolved. All other modes are
+	// primary-only.
+	return mode == "dnsmasq" || mode == "systemd-resolved"
+}
+
+func (m *nmManager) GetBaseConfig() (OSConfig, error) {
+	conn, err := dbus.SystemBus()
+	if err != nil {
+		return OSConfig{}, err
+	}
+
+	nm := conn.Object("org.freedesktop.NetworkManager", dbus.ObjectPath("/org/freedesktop/NetworkManager/DnsManager"))
+	v, err := nm.GetProperty("org.freedesktop.NetworkManager.DnsManager.Configuration")
+	if err != nil {
+		return OSConfig{}, err
+	}
+	cfgs, ok := v.Value().([]map[string]dbus.Variant)
+	if !ok {
+		return OSConfig{}, fmt.Errorf("unexpected NM config type %T", v.Value())
+	}
+
+	if len(cfgs) == 0 {
+		return OSConfig{}, nil
+	}
+
+	type dnsPrio struct {
+		resolvers []netaddr.IP
+		domains   []string
+		priority  int32
+	}
+	order := make([]dnsPrio, 0, len(cfgs)-1)
+
+	for _, cfg := range cfgs {
+		if name, ok := cfg["interface"]; ok {
+			if s, ok := name.Value().(string); ok && s == m.interfaceName {
+				// Config for the taislcale interface, skip.
+				continue
+			}
+		}
+
+		var p dnsPrio
+
+		if v, ok := cfg["nameservers"]; ok {
+			if ips, ok := v.Value().([]string); ok {
+				for _, s := range ips {
+					ip, err := netaddr.ParseIP(s)
+					if err != nil {
+						// hmm, what do? Shouldn't really happen.
+						continue
+					}
+					p.resolvers = append(p.resolvers, ip)
+				}
+			}
+		}
+		if v, ok := cfg["domains"]; ok {
+			if domains, ok := v.Value().([]string); ok {
+				p.domains = domains
+			}
+		}
+		if v, ok := cfg["priority"]; ok {
+			if prio, ok := v.Value().(int32); ok {
+				p.priority = prio
+			}
+		}
+
+		order = append(order, p)
+	}
+
+	sort.Slice(order, func(i, j int) bool {
+		return order[i].priority < order[j].priority
+	})
+
+	var (
+		ret           OSConfig
+		seenResolvers = map[netaddr.IP]bool{}
+		seenSearch    = map[string]bool{}
+	)
+
+	for _, cfg := range order {
+		for _, resolver := range cfg.resolvers {
+			if seenResolvers[resolver] {
+				continue
+			}
+			ret.Nameservers = append(ret.Nameservers, resolver)
+			seenResolvers[resolver] = true
+		}
+		for _, dom := range cfg.domains {
+			if seenSearch[dom] {
+				continue
+			}
+			fqdn, err := dnsname.ToFQDN(dom)
+			if err != nil {
+				continue
+			}
+			ret.SearchDomains = append(ret.SearchDomains, fqdn)
+			seenSearch[dom] = true
+		}
+		if cfg.priority < 0 {
+			// exclusive configurations preempt all other
+			// configurations, so we're done.
+			break
+		}
+	}
+
+	return ret, nil
+}
+
+func (m *nmManager) Close() error {
+	// No need to do anything on close, NetworkManager will delete our
+	// settings when the tailscale interface goes away.
+	return nil
 }

net/dns/noop.go

@@ -6,12 +6,13 @@ package dns
 
 type noopManager struct{}
 
-// Up implements managerImpl.
-func (m noopManager) Up(Config) error { return nil }
-
-// Down implements managerImpl.
-func (m noopManager) Down() error { return nil }
-
-func newNoopManager(mconfig ManagerConfig) managerImpl {
-	return noopManager{}
+func (m noopManager) SetDNS(OSConfig) error  { return nil }
+func (m noopManager) SupportsSplitDNS() bool { return false }
+func (m noopManager) Close() error           { return nil }
+func (m noopManager) GetBaseConfig() (OSConfig, error) {
+	return OSConfig{}, ErrGetBaseConfigNotSupported
+}
+
+func NewNoopManager() (noopManager, error) {
+	return noopManager{}, nil
 }

net/dns/openresolv.go

@@ -0,0 +1,92 @@
+// Copyright (c) 2021 Tailscale Inc & AUTHORS All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package dns
+
+import (
+	"bytes"
+	"fmt"
+	"os/exec"
+	"strings"
+)
+
+// openresolvManager manages DNS configuration using the openresolv
+// implementation of the `resolvconf` program.
+type openresolvManager struct{}
+
+func newOpenresolvManager() (openresolvManager, error) {
+	return openresolvManager{}, nil
+}
+
+func (m openresolvManager) deleteTailscaleConfig() error {
+	cmd := exec.Command("resolvconf", "-f", "-d", "tailscale")
+	out, err := cmd.CombinedOutput()
+	if err != nil {
+		return fmt.Errorf("running %s: %s", cmd, out)
+	}
+	return nil
+}
+
+func (m openresolvManager) SetDNS(config OSConfig) error {
+	if config.IsZero() {
+		return m.deleteTailscaleConfig()
+	}
+
+	var stdin bytes.Buffer
+	writeResolvConf(&stdin, config.Nameservers, config.SearchDomains)
+
+	cmd := exec.Command("resolvconf", "-m", "0", "-x", "-a", "tailscale")
+	cmd.Stdin = &stdin
+	out, err := cmd.CombinedOutput()
+	if err != nil {
+		return fmt.Errorf("running %s: %s", cmd, out)
+	}
+	return nil
+}
+
+func (m openresolvManager) SupportsSplitDNS() bool {
+	return false
+}
+
+func (m openresolvManager) GetBaseConfig() (OSConfig, error) {
+	// List the names of all config snippets openresolv is aware
+	// of. Snippets get listed in priority order (most to least),
+	// which we'll exploit later.
+	bs, err := exec.Command("resolvconf", "-i").CombinedOutput()
+	if err != nil {
+		return OSConfig{}, err
+	}
+
+	// Remove the "tailscale" snippet from the list.
+	args := []string{"-l"}
+	for _, f := range strings.Split(strings.TrimSpace(string(bs)), " ") {
+		if f == "tailscale" {
+			continue
+		}
+		args = append(args, f)
+	}
+
+	// List all resolvconf snippets except our own, and parse that as
+	// a resolv.conf. This effectively generates a blended config of
+	// "everyone except tailscale", which is what would be in use if
+	// tailscale hadn't set exclusive mode.
+	//
+	// Note that this is not _entirely_ true. To be perfectly correct,
+	// we should be looking for other interfaces marked exclusive that
+	// predated tailscale, and stick to only those. However, in
+	// practice, openresolv uses are generally quite limited, and boil
+	// down to 1-2 DHCP leases, for which the correct outcome is a
+	// blended config like the one we produce here.
+	var buf bytes.Buffer
+	cmd := exec.Command("resolvconf", args...)
+	cmd.Stdout = &buf
+	if err := cmd.Run(); err != nil {
+		return OSConfig{}, err
+	}
+	return readResolv(&buf)
+}
+
+func (m openresolvManager) Close() error {
+	return m.deleteTailscaleConfig()
+}

net/dns/osconfig.go

@@ -0,0 +1,92 @@
+// Copyright (c) 2021 Tailscale Inc & AUTHORS All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package dns
+
+import (
+	"errors"
+
+	"inet.af/netaddr"
+	"tailscale.com/util/dnsname"
+)
+
+// An OSConfigurator applies DNS settings to the operating system.
+type OSConfigurator interface {
+	// SetDNS updates the OS's DNS configuration to match cfg.
+	// If cfg is the zero value, all Tailscale-related DNS
+	// configuration is removed.
+	// SetDNS must not be called after Close.
+	SetDNS(cfg OSConfig) error
+	// SupportsSplitDNS reports whether the configurator is capable of
+	// installing a resolver only for specific DNS suffixes. If false,
+	// the configurator can only set a global resolver.
+	SupportsSplitDNS() bool
+	// GetBaseConfig returns the OS's "base" configuration, i.e. the
+	// resolver settings the OS would use without Tailscale
+	// contributing any configuration.
+	// GetBaseConfig must return the tailscale-free base config even
+	// after SetDNS has been called to set a Tailscale configuration.
+	// Only works when SupportsSplitDNS=false.
+
+	// Implementations that don't support getting the base config must
+	// return ErrGetBaseConfigNotSupported.
+	GetBaseConfig() (OSConfig, error)
+	// Close removes Tailscale-related DNS configuration from the OS.
+	Close() error
+}
+
+// OSConfig is an OS DNS configuration.
+type OSConfig struct {
+	// Nameservers are the IP addresses of the nameservers to use.
+	Nameservers []netaddr.IP
+	// SearchDomains are the domain suffixes to use when expanding
+	// single-label name queries. SearchDomains is additive to
+	// whatever non-Tailscale search domains the OS has.
+	SearchDomains []dnsname.FQDN
+	// MatchDomains are the DNS suffixes for which Nameservers should
+	// be used. If empty, Nameservers is installed as the "primary" resolver.
+	// A non-empty MatchDomains requests a "split DNS" configuration
+	// from the OS, which will only work with OSConfigurators that
+	// report SupportsSplitDNS()=true.
+	MatchDomains []dnsname.FQDN
+}
+
+func (o OSConfig) IsZero() bool {
+	return len(o.Nameservers) == 0 && len(o.SearchDomains) == 0 && len(o.MatchDomains) == 0
+}
+
+func (a OSConfig) Equal(b OSConfig) bool {
+	if len(a.Nameservers) != len(b.Nameservers) {
+		return false
+	}
+	if len(a.SearchDomains) != len(b.SearchDomains) {
+		return false
+	}
+	if len(a.MatchDomains) != len(b.MatchDomains) {
+		return false
+	}
+
+	for i := range a.Nameservers {
+		if a.Nameservers[i] != b.Nameservers[i] {
+			return false
+		}
+	}
+	for i := range a.SearchDomains {
+		if a.SearchDomains[i] != b.SearchDomains[i] {
+			return false
+		}
+	}
+	for i := range a.MatchDomains {
+		if a.MatchDomains[i] != b.MatchDomains[i] {
+			return false
+		}
+	}
+
+	return true
+}
+
+// ErrGetBaseConfigNotSupported is the error
+// OSConfigurator.GetBaseConfig returns when the OSConfigurator
+// doesn't support reading the underlying configuration out of the OS.
+var ErrGetBaseConfigNotSupported = errors.New("getting OS base config is not supported")

net/dns/resolvconf-workaround.sh

@@ -0,0 +1,63 @@
+#!/bin/sh
+# Copyright (c) 2021 Tailscale Inc & AUTHORS All rights reserved.
+# Use of this source code is governed by a BSD-style
+# license that can be found in the LICENSE file.
+#
+# This script is a workaround for a vpn-unfriendly behavior of the
+# original resolvconf by Thomas Hood. Unlike the `openresolv`
+# implementation (whose binary is also called resolvconf,
+# confusingly), the original resolvconf lacks a way to specify
+# "exclusive mode" for a provider configuration. In practice, this
+# means that if Tailscale wants to install a DNS configuration, that
+# config will get "blended" with the configs from other sources,
+# rather than override those other sources.
+#
+# This script gets installed at /etc/resolvconf/update-libc.d, which
+# is a directory of hook scripts that get run after resolvconf's libc
+# helper has finished rewriting /etc/resolv.conf. It's meant to notify
+# consumers of resolv.conf of a new configuration.
+#
+# Instead, we use that hook mechanism to reach into resolvconf's
+# stuff, and rewrite the libc-generated resolv.conf to exclusively
+# contain Tailscale's configuration - effectively implementing
+# exclusive mode ourselves in post-production.
+
+set -e
+
+if [ -n "$TAILSCALE_RESOLVCONF_HOOK_LOOP" ]; then
+	# Hook script being invoked by itself, skip.
+	exit 0
+fi
+
+if [ ! -f tun-tailscale.inet ]; then
+	# Tailscale isn't trying to manage DNS, do nothing.
+	exit 0
+fi
+
+if ! grep resolvconf /etc/resolv.conf >/dev/null; then
+	# resolvconf isn't managing /etc/resolv.conf, do nothing.
+	exit 0
+fi
+
+# Write out a modified /etc/resolv.conf containing just our config.
+(
+	if [ -f /etc/resolvconf/resolv.conf.d/head ]; then
+		cat /etc/resolvconf/resolv.conf.d/head
+	fi
+	echo "# Tailscale workaround applied to set exclusive DNS configuration."
+	cat tun-tailscale.inet
+	if [ -f /etc/resolvconf/resolv.conf.d/base ]; then
+		# Keep options and sortlist, discard other base things since
+		# they're the things we're trying to override.
+		grep -e 'sortlist ' -e 'options ' /etc/resolvconf/resolv.conf.d/base || true
+	fi
+	if [ -f /etc/resolvconf/resolv.conf.d/tail ]; then
+		cat /etc/resolvconf/resolv.conf.d/tail
+	fi
+) >/etc/resolv.conf
+
+if [ -d /etc/resolvconf/update-libc.d ] ; then
+	# Re-notify libc watchers that we've changed resolv.conf again.
+	export TAILSCALE_RESOLVCONF_HOOK_LOOP=1
+	exec run-parts /etc/resolvconf/update-libc.d
+fi

net/dns/resolvconf.go

@@ -1,157 +1,27 @@
-// Copyright (c) 2020 Tailscale Inc & AUTHORS All rights reserved.
+// Copyright (c) 2021 Tailscale Inc & AUTHORS All rights reserved.
 // Use of this source code is governed by a BSD-style
 // license that can be found in the LICENSE file.
 
-// +build linux freebsd
-
 package dns
 
 import (
-	"bufio"
-	"bytes"
-	"fmt"
-	"os"
 	"os/exec"
+
+	"tailscale.com/types/logger"
 )
 
-// isResolvconfActive indicates whether the system appears to be using resolvconf.
-// If this is true, then directManager should be avoided:
-// resolvconf has exclusive ownership of /etc/resolv.conf.
-func isResolvconfActive() bool {
-	// Sanity-check first: if there is no resolvconf binary, then this is fruitless.
-	//
-	// However, this binary may be a shim like the one systemd-resolved provides.
-	// Such a shim may not behave as expected: in particular, systemd-resolved
-	// does not seem to respect the exclusive mode -x, saying:
-	//   -x            Send DNS traffic preferably over this interface
-	// whereas e.g. openresolv sends DNS traffix _exclusively_ over that interface,
-	// or not at all (in case of another exclusive-mode request later in time).
-	//
-	// Moreover, resolvconf may be installed but unused, in which case we should
-	// not use it either, lest we clobber existing configuration.
-	//
-	// To handle all the above correctly, we scan the comments in /etc/resolv.conf
-	// to ensure that it was generated by a resolvconf implementation.
-	_, err := exec.LookPath("resolvconf")
+func newResolvconfManager(logf logger.Logf) (OSConfigurator, error) {
+	_, err := exec.Command("resolvconf", "--version").CombinedOutput()
 	if err != nil {
-		return false
-	}
-
-	f, err := os.Open("/etc/resolv.conf")
-	if err != nil {
-		return false
-	}
-	defer f.Close()
-
-	scanner := bufio.NewScanner(f)
-	for scanner.Scan() {
-		line := scanner.Bytes()
-		// Look for the word "resolvconf" until comments end.
-		if len(line) > 0 && line[0] != '#' {
-			return false
-		}
-		if bytes.Contains(line, []byte("resolvconf")) {
-			return true
+		if exitErr, ok := err.(*exec.ExitError); ok && exitErr.ExitCode() == 99 {
+			// Debian resolvconf doesn't understand --version, and
+			// exits with a specific error code.
+			return newDebianResolvconfManager(logf)
 		}
 	}
-	return false
-}
-
-// resolvconfImpl enumerates supported implementations of the resolvconf CLI.
-type resolvconfImpl uint8
-
-const (
-	// resolvconfOpenresolv is the implementation packaged as "openresolv" on Ubuntu.
-	// It supports exclusive mode and interface metrics.
-	resolvconfOpenresolv resolvconfImpl = iota
-	// resolvconfLegacy is the implementation by Thomas Hood packaged as "resolvconf" on Ubuntu.
-	// It does not support exclusive mode or interface metrics.
-	resolvconfLegacy
-)
-
-func (impl resolvconfImpl) String() string {
-	switch impl {
-	case resolvconfOpenresolv:
-		return "openresolv"
-	case resolvconfLegacy:
-		return "legacy"
-	default:
-		return "unknown"
-	}
-}
-
-// getResolvconfImpl returns the implementation of resolvconf that appears to be in use.
-func getResolvconfImpl() resolvconfImpl {
-	err := exec.Command("resolvconf", "-v").Run()
-	if err != nil {
-		if exitErr, ok := err.(*exec.ExitError); ok {
-			// Thomas Hood's resolvconf has a minimal flag set
-			// and exits with code 99 when passed an unknown flag.
-			if exitErr.ExitCode() == 99 {
-				return resolvconfLegacy
-			}
-		}
-	}
-	return resolvconfOpenresolv
-}
-
-type resolvconfManager struct {
-	impl resolvconfImpl
-}
-
-func newResolvconfManager(mconfig ManagerConfig) managerImpl {
-	impl := getResolvconfImpl()
-	mconfig.Logf("resolvconf implementation is %s", impl)
-
-	return resolvconfManager{
-		impl: impl,
-	}
-}
-
-// resolvconfConfigName is the name of the config submitted to resolvconf.
-// It has this form to match the "tun*" rule in interface-order
-// when running resolvconfLegacy, hopefully placing our config first.
-const resolvconfConfigName = "tun-tailscale.inet"
-
-// Up implements managerImpl.
-func (m resolvconfManager) Up(config Config) error {
-	stdin := new(bytes.Buffer)
-	writeResolvConf(stdin, config.Nameservers, config.Domains) // dns_direct.go
-
-	var cmd *exec.Cmd
-	switch m.impl {
-	case resolvconfOpenresolv:
-		// Request maximal priority (metric 0) and exclusive mode.
-		cmd = exec.Command("resolvconf", "-m", "0", "-x", "-a", resolvconfConfigName)
-	case resolvconfLegacy:
-		// This does not quite give us the desired behavior (queries leak),
-		// but there is nothing else we can do without messing with other interfaces' settings.
-		cmd = exec.Command("resolvconf", "-a", resolvconfConfigName)
-	}
-	cmd.Stdin = stdin
-	out, err := cmd.CombinedOutput()
-	if err != nil {
-		return fmt.Errorf("running %s: %s", cmd, out)
-	}
-
-	return nil
-}
-
-// Down implements managerImpl.
-func (m resolvconfManager) Down() error {
-	var cmd *exec.Cmd
-	switch m.impl {
-	case resolvconfOpenresolv:
-		cmd = exec.Command("resolvconf", "-f", "-d", resolvconfConfigName)
-	case resolvconfLegacy:
-		// resolvconfLegacy lacks the -f flag.
-		// Instead, it succeeds even when the config does not exist.
-		cmd = exec.Command("resolvconf", "-d", resolvconfConfigName)
-	}
-	out, err := cmd.CombinedOutput()
-	if err != nil {
-		return fmt.Errorf("running %s: %s", cmd, out)
-	}
-
-	return nil
+	// If --version works, or we got some surprising error while
+	// probing, use openresolv. It's the more common implementation,
+	// so in cases where we can't figure things out, it's the least
+	// likely to misbehave.
+	return newOpenresolvManager()
 }

net/dns/resolved.go

@@ -4,18 +4,21 @@
 
 // +build linux
 
+//lint:file-ignore U1000 refactoring, temporarily unused code.
+
 package dns
 
 import (
 	"context"
 	"errors"
 	"fmt"
-	"os/exec"
 
 	"github.com/godbus/dbus/v5"
 	"golang.org/x/sys/unix"
 	"inet.af/netaddr"
 	"tailscale.com/net/interfaces"
+	"tailscale.com/types/logger"
+	"tailscale.com/util/dnsname"
 )
 
 // resolvedListenAddr is the listen address of the resolved stub resolver.
@@ -49,15 +52,20 @@ type resolvedLinkDomain struct {
 
 // isResolvedActive determines if resolved is currently managing system DNS settings.
 func isResolvedActive() bool {
-	// systemd-resolved is never installed without systemd.
-	_, err := exec.LookPath("systemctl")
+	ctx, cancel := context.WithTimeout(context.Background(), reconfigTimeout)
+	defer cancel()
+
+	conn, err := dbus.SystemBus()
 	if err != nil {
+		// Probably no DBus on the system, or we're not allowed to use
+		// it. Cannot control resolved.
 		return false
 	}
 
-	// is-active exits with code 3 if the service is not active.
-	err = exec.Command("systemctl", "is-active", "systemd-resolved").Run()
-	if err != nil {
+	rd := conn.Object("org.freedesktop.resolve1", dbus.ObjectPath("/org/freedesktop/resolve1"))
+	call := rd.CallWithContext(ctx, "org.freedesktop.DBus.Peer.Ping", 0)
+	if call.Err != nil {
+		// Can't talk to resolved.
 		return false
 	}
 
@@ -75,29 +83,28 @@ func isResolvedActive() bool {
 }
 
 // resolvedManager uses the systemd-resolved DBus API.
-type resolvedManager struct{}
+type resolvedManager struct {
+	logf     logger.Logf
+	resolved dbus.BusObject
+}
 
-func newResolvedManager(mconfig ManagerConfig) managerImpl {
-	return resolvedManager{}
+func newResolvedManager(logf logger.Logf) (*resolvedManager, error) {
+	conn, err := dbus.SystemBus()
+	if err != nil {
+		return nil, err
+	}
+
+	return &resolvedManager{
+		logf:     logf,
+		resolved: conn.Object("org.freedesktop.resolve1", dbus.ObjectPath("/org/freedesktop/resolve1")),
+	}, nil
 }
 
 // Up implements managerImpl.
-func (m resolvedManager) Up(config Config) error {
+func (m *resolvedManager) SetDNS(config OSConfig) error {
 	ctx, cancel := context.WithTimeout(context.Background(), reconfigTimeout)
 	defer cancel()
 
-	// conn is a shared connection whose lifecycle is managed by the dbus package.
-	// We should not interfere with that by closing it.
-	conn, err := dbus.SystemBus()
-	if err != nil {
-		return fmt.Errorf("connecting to system bus: %w", err)
-	}
-
-	resolved := conn.Object(
-		"org.freedesktop.resolve1",
-		dbus.ObjectPath("/org/freedesktop/resolve1"),
-	)
-
 	// In principle, we could persist this in the manager struct
 	// if we knew that interface indices are persistent. This does not seem to be the case.
 	_, iface, err := interfaces.Tailscale()
@@ -124,7 +131,7 @@ func (m resolvedManager) Up(config Config) error {
 		}
 	}
 
-	err = resolved.CallWithContext(
+	err = m.resolved.CallWithContext(
 		ctx, "org.freedesktop.resolve1.Manager.SetLinkDNS", 0,
 		iface.Index, linkNameservers,
 	).Store()
@@ -132,15 +139,40 @@ func (m resolvedManager) Up(config Config) error {
 		return fmt.Errorf("setLinkDNS: %w", err)
 	}
 
-	var linkDomains = make([]resolvedLinkDomain, len(config.Domains))
-	for i, domain := range config.Domains {
-		linkDomains[i] = resolvedLinkDomain{
-			Domain:      domain,
-			RoutingOnly: false,
+	linkDomains := make([]resolvedLinkDomain, 0, len(config.SearchDomains)+len(config.MatchDomains))
+	seenDomains := map[dnsname.FQDN]bool{}
+	for _, domain := range config.SearchDomains {
+		if seenDomains[domain] {
+			continue
 		}
+		seenDomains[domain] = true
+		linkDomains = append(linkDomains, resolvedLinkDomain{
+			Domain:      domain.WithTrailingDot(),
+			RoutingOnly: false,
+		})
+	}
+	for _, domain := range config.MatchDomains {
+		if seenDomains[domain] {
+			// Search domains act as both search and match in
+			// resolved, so it's correct to skip.
+			continue
+		}
+		seenDomains[domain] = true
+		linkDomains = append(linkDomains, resolvedLinkDomain{
+			Domain:      domain.WithTrailingDot(),
+			RoutingOnly: true,
+		})
+	}
+	if len(config.MatchDomains) == 0 && len(config.Nameservers) > 0 {
+		// Caller requested full DNS interception, install a
+		// routing-only root domain.
+		linkDomains = append(linkDomains, resolvedLinkDomain{
+			Domain:      ".",
+			RoutingOnly: true,
+		})
 	}
 
-	err = resolved.CallWithContext(
+	err = m.resolved.CallWithContext(
 		ctx, "org.freedesktop.resolve1.Manager.SetLinkDomains", 0,
 		iface.Index, linkDomains,
 	).Store()
@@ -148,26 +180,53 @@ func (m resolvedManager) Up(config Config) error {
 		return fmt.Errorf("setLinkDomains: %w", err)
 	}
 
+	if call := m.resolved.CallWithContext(ctx, "org.freedesktop.resolve1.Manager.SetLinkDefaultRoute", 0, iface.Index, len(config.MatchDomains) == 0); call.Err != nil {
+		return fmt.Errorf("setLinkDefaultRoute: %w", err)
+	}
+
+	// Some best-effort setting of things, but resolved should do the
+	// right thing if these fail (e.g. a really old resolved version
+	// or something).
+
+	// Disable LLMNR, we don't do multicast.
+	if call := m.resolved.CallWithContext(ctx, "org.freedesktop.resolve1.Manager.SetLinkLLMNR", 0, iface.Index, "no"); call.Err != nil {
+		m.logf("[v1] failed to disable LLMNR: %v", call.Err)
+	}
+
+	// Disable mdns.
+	if call := m.resolved.CallWithContext(ctx, "org.freedesktop.resolve1.Manager.SetLinkMulticastDNS", 0, iface.Index, "no"); call.Err != nil {
+		m.logf("[v1] failed to disable mdns: %v", call.Err)
+	}
+
+	// We don't support dnssec consistently right now, force it off to
+	// avoid partial failures when we split DNS internally.
+	if call := m.resolved.CallWithContext(ctx, "org.freedesktop.resolve1.Manager.SetLinkDNSSEC", 0, iface.Index, "no"); call.Err != nil {
+		m.logf("[v1] failed to disable DNSSEC: %v", call.Err)
+	}
+
+	if call := m.resolved.CallWithContext(ctx, "org.freedesktop.resolve1.Manager.SetLinkDNSOverTLS", 0, iface.Index, "no"); call.Err != nil {
+		m.logf("[v1] failed to disable DoT: %v", call.Err)
+	}
+
+	if call := m.resolved.CallWithContext(ctx, "org.freedesktop.resolve1.Manager.FlushCaches", 0); call.Err != nil {
+		m.logf("failed to flush resolved DNS cache: %v", call.Err)
+	}
+
 	return nil
 }
 
-// Down implements managerImpl.
-func (m resolvedManager) Down() error {
+func (m *resolvedManager) SupportsSplitDNS() bool {
+	return true
+}
+
+func (m *resolvedManager) GetBaseConfig() (OSConfig, error) {
+	return OSConfig{}, ErrGetBaseConfigNotSupported
+}
+
+func (m *resolvedManager) Close() error {
 	ctx, cancel := context.WithTimeout(context.Background(), reconfigTimeout)
 	defer cancel()
 
-	// conn is a shared connection whose lifecycle is managed by the dbus package.
-	// We should not interfere with that by closing it.
-	conn, err := dbus.SystemBus()
-	if err != nil {
-		return fmt.Errorf("connecting to system bus: %w", err)
-	}
-
-	resolved := conn.Object(
-		"org.freedesktop.resolve1",
-		dbus.ObjectPath("/org/freedesktop/resolve1"),
-	)
-
 	_, iface, err := interfaces.Tailscale()
 	if err != nil {
 		return fmt.Errorf("getting interface index: %w", err)
@@ -176,12 +235,8 @@ func (m resolvedManager) Down() error {
 		return errNotReady
 	}
 
-	err = resolved.CallWithContext(
-		ctx, "org.freedesktop.resolve1.Manager.RevertLink", 0,
-		iface.Index,
-	).Store()
-	if err != nil {
-		return fmt.Errorf("RevertLink: %w", err)
+	if call := m.resolved.CallWithContext(ctx, "org.freedesktop.resolve1.Manager.RevertLink", 0, iface.Index); call.Err != nil {
+		return fmt.Errorf("RevertLink: %w", call.Err)
 	}
 
 	return nil

net/dns/resolver/forwarder.go

@@ -2,7 +2,7 @@
 // Use of this source code is governed by a BSD-style
 // license that can be found in the LICENSE file.
 
-package dns
+package resolver
 
 import (
 	"bytes"
@@ -17,10 +17,11 @@ import (
 	"sync"
 	"time"
 
+	dns "golang.org/x/net/dns/dnsmessage"
 	"inet.af/netaddr"
 	"tailscale.com/logtail/backoff"
-	"tailscale.com/net/netns"
 	"tailscale.com/types/logger"
+	"tailscale.com/util/dnsname"
 )
 
 // headerBytes is the number of bytes in a DNS message header.
@@ -100,12 +101,17 @@ func getTxID(packet []byte) txid {
 	return (txid(hash) << 32) | txid(dnsid)
 }
 
+type route struct {
+	suffix    dnsname.FQDN
+	resolvers []netaddr.IPPort
+}
+
 // forwarder forwards DNS packets to a number of upstream nameservers.
 type forwarder struct {
 	logf logger.Logf
 
 	// responses is a channel by which responses are returned.
-	responses chan Packet
+	responses chan packet
 	// closed signals all goroutines to stop.
 	closed chan struct{}
 	// wg signals when all goroutines have stopped.
@@ -116,35 +122,32 @@ type forwarder struct {
 	conns []*fwdConn
 
 	mu sync.Mutex
-	// upstreams are the nameserver addresses that should be used for forwarding.
-	upstreams []net.Addr
-	// txMap maps DNS txids to active forwarding records.
-	txMap map[txid]forwardingRecord
+	// routes are per-suffix resolvers to use.
+	routes []route                   // most specific routes first
+	txMap  map[txid]forwardingRecord // txids to in-flight requests
 }
 
 func init() {
 	rand.Seed(time.Now().UnixNano())
 }
 
-func newForwarder(logf logger.Logf, responses chan Packet) *forwarder {
-	return &forwarder{
+func newForwarder(logf logger.Logf, responses chan packet) *forwarder {
+	ret := &forwarder{
 		logf:      logger.WithPrefix(logf, "forward: "),
 		responses: responses,
 		closed:    make(chan struct{}),
 		conns:     make([]*fwdConn, connCount),
 		txMap:     make(map[txid]forwardingRecord),
 	}
-}
 
-func (f *forwarder) Start() error {
-	f.wg.Add(connCount + 1)
-	for idx := range f.conns {
-		f.conns[idx] = newFwdConn(f.logf, idx)
-		go f.recv(f.conns[idx])
+	ret.wg.Add(connCount + 1)
+	for idx := range ret.conns {
+		ret.conns[idx] = newFwdConn(ret.logf, idx)
+		go ret.recv(ret.conns[idx])
 	}
-	go f.cleanMap()
+	go ret.cleanMap()
 
-	return nil
+	return ret
 }
 
 func (f *forwarder) Close() {
@@ -171,14 +174,14 @@ func (f *forwarder) rebindFromNetworkChange() {
 	}
 }
 
-func (f *forwarder) setUpstreams(upstreams []net.Addr) {
+func (f *forwarder) setRoutes(routes []route) {
 	f.mu.Lock()
-	f.upstreams = upstreams
+	f.routes = routes
 	f.mu.Unlock()
 }
 
 // send sends packet to dst. It is best effort.
-func (f *forwarder) send(packet []byte, dst net.Addr) {
+func (f *forwarder) send(packet []byte, dst netaddr.IPPort) {
 	connIdx := rand.Intn(connCount)
 	conn := f.conns[connIdx]
 	conn.send(packet, dst)
@@ -218,14 +221,11 @@ func (f *forwarder) recv(conn *fwdConn) {
 
 		f.mu.Unlock()
 
-		packet := Packet{
-			Payload: out,
-			Addr:    record.src,
-		}
+		pkt := packet{out, record.src}
 		select {
 		case <-f.closed:
 			return
-		case f.responses <- packet:
+		case f.responses <- pkt:
 			// continue
 		}
 	}
@@ -258,25 +258,39 @@ func (f *forwarder) cleanMap() {
 }
 
 // forward forwards the query to all upstream nameservers and returns the first response.
-func (f *forwarder) forward(query Packet) error {
-	txid := getTxID(query.Payload)
+func (f *forwarder) forward(query packet) error {
+	domain, err := nameFromQuery(query.bs)
+	if err != nil {
+		return err
+	}
+
+	txid := getTxID(query.bs)
 
 	f.mu.Lock()
-
-	upstreams := f.upstreams
-	if len(upstreams) == 0 {
-		f.mu.Unlock()
-		return errNoUpstreams
-	}
-	f.txMap[txid] = forwardingRecord{
-		src:       query.Addr,
-		createdAt: time.Now(),
-	}
-
+	routes := f.routes
 	f.mu.Unlock()
 
-	for _, upstream := range upstreams {
-		f.send(query.Payload, upstream)
+	var resolvers []netaddr.IPPort
+	for _, route := range routes {
+		if route.suffix != "." && !route.suffix.Contains(domain) {
+			continue
+		}
+		resolvers = route.resolvers
+		break
+	}
+	if len(resolvers) == 0 {
+		return errNoUpstreams
+	}
+
+	f.mu.Lock()
+	f.txMap[txid] = forwardingRecord{
+		src:       query.addr,
+		createdAt: time.Now(),
+	}
+	f.mu.Unlock()
+
+	for _, resolver := range resolvers {
+		f.send(query.bs, resolver)
 	}
 
 	return nil
@@ -312,7 +326,7 @@ func newFwdConn(logf logger.Logf, idx int) *fwdConn {
 
 // send sends packet to dst using c's connection.
 // It is best effort. It is UDP, after all. Failures are logged.
-func (c *fwdConn) send(packet []byte, dst net.Addr) {
+func (c *fwdConn) send(packet []byte, dst netaddr.IPPort) {
 	var b *backoff.Backoff // lazily initialized, since it is not needed in the common case
 	backOff := func(err error) {
 		if b == nil {
@@ -338,8 +352,9 @@ func (c *fwdConn) send(packet []byte, dst net.Addr) {
 		}
 		c.mu.Unlock()
 
+		a := dst.UDPAddr()
 		c.wg.Add(1)
-		_, err := conn.WriteTo(packet, dst)
+		_, err := conn.WriteTo(packet, a)
 		c.wg.Done()
 		if err == nil {
 			// Success
@@ -437,7 +452,7 @@ func (c *fwdConn) read(out []byte) int {
 func (c *fwdConn) reconnectLocked() {
 	c.closeConnLocked()
 	// Make a new connection.
-	conn, err := netns.Listener().ListenPacket(context.Background(), "udp", "")
+	conn, err := net.ListenPacket("udp", "")
 	if err != nil {
 		c.logf("ListenPacket failed: %v", err)
 	} else {
@@ -472,3 +487,24 @@ func (c *fwdConn) close() {
 	// Unblock any remaining readers.
 	c.change.Broadcast()
 }
+
+// nameFromQuery extracts the normalized query name from bs.
+func nameFromQuery(bs []byte) (dnsname.FQDN, error) {
+	var parser dns.Parser
+
+	hdr, err := parser.Start(bs)
+	if err != nil {
+		return "", err
+	}
+	if hdr.Response {
+		return "", errNotQuery
+	}
+
+	q, err := parser.Question()
+	if err != nil {
+		return "", err
+	}
+
+	n := q.Name.Data[:q.Name.Length]
+	return dnsname.ToFQDN(rawNameToLower(n))
+}

net/dns/resolver/neterr_darwin.go

@@ -2,7 +2,7 @@
 // Use of this source code is governed by a BSD-style
 // license that can be found in the LICENSE file.
 
-package dns
+package resolver
 
 import (
 	"errors"

net/dns/resolver/neterr_other.go

@@ -4,7 +4,7 @@
 
 // +build !darwin,!windows
 
-package dns
+package resolver
 
 func networkIsDown(err error) bool        { return false }
 func networkIsUnreachable(err error) bool { return false }

net/dns/resolver/neterr_windows.go

@@ -2,7 +2,7 @@
 // Use of this source code is governed by a BSD-style
 // license that can be found in the LICENSE file.
 
-package dns
+package resolver
 
 import (
 	"net"

net/dns/resolver/tsdns.go

@@ -2,14 +2,14 @@
 // Use of this source code is governed by a BSD-style
 // license that can be found in the LICENSE file.
 
-// Package dns provides a Resolver capable of resolving
-// domains on a Tailscale network.
-package dns
+// Package resolver implements a stub DNS resolver that can also serve
+// records out of an internal local zone.
+package resolver
 
 import (
 	"encoding/hex"
 	"errors"
-	"net"
+	"sort"
 	"strings"
 	"sync"
 	"time"
@@ -37,22 +37,33 @@ const defaultTTL = 600 * time.Second
 var ErrClosed = errors.New("closed")
 
 var (
-	errFullQueue      = errors.New("request queue full")
-	errMapNotSet      = errors.New("domain map not set")
-	errNotForwarding  = errors.New("forwarding disabled")
-	errNotImplemented = errors.New("query type not implemented")
-	errNotQuery       = errors.New("not a DNS query")
-	errNotOurName     = errors.New("not a Tailscale DNS name")
+	errFullQueue  = errors.New("request queue full")
+	errNotQuery   = errors.New("not a DNS query")
+	errNotOurName = errors.New("not a Tailscale DNS name")
 )
 
-// Packet represents a DNS payload together with the address of its origin.
-type Packet struct {
-	// Payload is the application layer DNS payload.
-	// Resolver assumes ownership of the request payload when it is enqueued
-	// and cedes ownership of the response payload when it is returned from NextResponse.
-	Payload []byte
-	// Addr is the source address for a request and the destination address for a response.
-	Addr netaddr.IPPort
+type packet struct {
+	bs   []byte
+	addr netaddr.IPPort // src for a request, dst for a response
+}
+
+// Config is a resolver configuration.
+// Given a Config, queries are resolved in the following order:
+// If the query is an exact match for an entry in LocalHosts, return that.
+// Else if the query suffix matches an entry in LocalDomains, return NXDOMAIN.
+// Else forward the query to the most specific matching entry in Routes.
+// Else return SERVFAIL.
+type Config struct {
+	// Routes is a map of DNS name suffix to the resolvers to use for
+	// queries within that suffix.
+	// Queries only match the most specific suffix.
+	// To register a "default route", add an entry for ".".
+	Routes map[dnsname.FQDN][]netaddr.IPPort
+	// LocalHosts is a map of FQDNs to corresponding IPs.
+	Hosts map[dnsname.FQDN][]netaddr.IP
+	// LocalDomains is a list of DNS name suffixes that should not be
+	// routed to upstream resolvers.
+	LocalDomains []dnsname.FQDN
 }
 
 // Resolver is a DNS resolver for nodes on the Tailscale network,
@@ -60,16 +71,17 @@ type Packet struct {
 // If it is asked to resolve a domain that is not of that form,
 // it delegates to upstream nameservers if any are set.
 type Resolver struct {
-	logf         logger.Logf
-	linkMon      *monitor.Mon // or nil
-	unregLinkMon func()       // or nil
+	logf               logger.Logf
+	linkMon            *monitor.Mon     // or nil
+	unregLinkMon       func()           // or nil
+	saveConfigForTests func(cfg Config) // used in tests to capture resolver config
 	// forwarder forwards requests to upstream nameservers.
 	forwarder *forwarder
 
 	// queue is a buffered channel holding DNS requests queued for resolution.
-	queue chan Packet
+	queue chan packet
 	// responses is an unbuffered channel to which responses are returned.
-	responses chan Packet
+	responses chan packet
 	// errors is an unbuffered channel to which errors are returned.
 	errors chan error
 	// closed signals all goroutines to stop.
@@ -78,56 +90,70 @@ type Resolver struct {
 	wg sync.WaitGroup
 
 	// mu guards the following fields from being updated while used.
-	mu sync.Mutex
-	// dnsMap is the map most recently received from the control server.
-	dnsMap *Map
+	mu           sync.Mutex
+	localDomains []dnsname.FQDN
+	hostToIP     map[dnsname.FQDN][]netaddr.IP
+	ipToHost     map[netaddr.IP]dnsname.FQDN
 }
 
-// ResolverConfig is the set of configuration options for a Resolver.
-type ResolverConfig struct {
-	// Logf is the logger to use throughout the Resolver.
-	Logf logger.Logf
-	// Forward determines whether the resolver will forward packets to
-	// nameservers set with SetUpstreams if the domain name is not of a Tailscale node.
-	Forward bool
-	// LinkMonitor optionally provides a link monitor to use to rebind
-	// connections on link changes.
-	// If nil, rebinds are not performend.
-	LinkMonitor *monitor.Mon
-}
-
-// NewResolver constructs a resolver associated with the given root domain.
-// The root domain must be in canonical form (with a trailing period).
-func NewResolver(config ResolverConfig) *Resolver {
+// New returns a new resolver.
+// linkMon optionally specifies a link monitor to use for socket rebinding.
+func New(logf logger.Logf, linkMon *monitor.Mon) *Resolver {
 	r := &Resolver{
-		logf:      logger.WithPrefix(config.Logf, "dns: "),
-		linkMon:   config.LinkMonitor,
-		queue:     make(chan Packet, queueSize),
-		responses: make(chan Packet),
+		logf:      logger.WithPrefix(logf, "dns: "),
+		linkMon:   linkMon,
+		queue:     make(chan packet, queueSize),
+		responses: make(chan packet),
 		errors:    make(chan error),
 		closed:    make(chan struct{}),
+		hostToIP:  map[dnsname.FQDN][]netaddr.IP{},
+		ipToHost:  map[netaddr.IP]dnsname.FQDN{},
 	}
-
-	if config.Forward {
-		r.forwarder = newForwarder(r.logf, r.responses)
-	}
+	r.forwarder = newForwarder(r.logf, r.responses)
 	if r.linkMon != nil {
 		r.unregLinkMon = r.linkMon.RegisterChangeCallback(r.onLinkMonitorChange)
 	}
 
-	return r
-}
-
-func (r *Resolver) Start() error {
-	if r.forwarder != nil {
-		if err := r.forwarder.Start(); err != nil {
-			return err
-		}
-	}
-
 	r.wg.Add(1)
 	go r.poll()
 
+	return r
+}
+
+func (r *Resolver) TestOnlySetHook(hook func(Config)) { r.saveConfigForTests = hook }
+
+func (r *Resolver) SetConfig(cfg Config) error {
+	if r.saveConfigForTests != nil {
+		r.saveConfigForTests(cfg)
+	}
+
+	routes := make([]route, 0, len(cfg.Routes))
+	reverse := make(map[netaddr.IP]dnsname.FQDN, len(cfg.Hosts))
+
+	for host, ips := range cfg.Hosts {
+		for _, ip := range ips {
+			reverse[ip] = host
+		}
+	}
+
+	for suffix, ips := range cfg.Routes {
+		routes = append(routes, route{
+			suffix:    suffix,
+			resolvers: ips,
+		})
+	}
+	// Sort from longest prefix to shortest.
+	sort.Slice(routes, func(i, j int) bool {
+		return routes[i].suffix.NumLabels() > routes[j].suffix.NumLabels()
+	})
+
+	r.forwarder.setRoutes(routes)
+
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	r.localDomains = cfg.LocalDomains
+	r.hostToIP = cfg.Hosts
+	r.ipToHost = reverse
 	return nil
 }
 
@@ -146,10 +172,7 @@ func (r *Resolver) Close() {
 		r.unregLinkMon()
 	}
 
-	if r.forwarder != nil {
-		r.forwarder.Close()
-	}
-
+	r.forwarder.Close()
 	r.wg.Wait()
 }
 
@@ -157,37 +180,17 @@ func (r *Resolver) onLinkMonitorChange(changed bool, state *interfaces.State) {
 	if !changed {
 		return
 	}
-	if r.forwarder != nil {
-		r.forwarder.rebindFromNetworkChange()
-	}
-}
-
-// SetMap sets the resolver's DNS map, taking ownership of it.
-func (r *Resolver) SetMap(m *Map) {
-	r.mu.Lock()
-	oldMap := r.dnsMap
-	r.dnsMap = m
-	r.mu.Unlock()
-	r.logf("map diff:\n%s", m.PrettyDiffFrom(oldMap))
-}
-
-// SetUpstreams sets the addresses of the resolver's
-// upstream nameservers, taking ownership of the argument.
-func (r *Resolver) SetUpstreams(upstreams []net.Addr) {
-	if r.forwarder != nil {
-		r.forwarder.setUpstreams(upstreams)
-	}
-	r.logf("set upstreams: %v", upstreams)
+	r.forwarder.rebindFromNetworkChange()
 }
 
 // EnqueueRequest places the given DNS request in the resolver's queue.
 // It takes ownership of the payload and does not block.
 // If the queue is full, the request will be dropped and an error will be returned.
-func (r *Resolver) EnqueueRequest(request Packet) error {
+func (r *Resolver) EnqueueRequest(bs []byte, from netaddr.IPPort) error {
 	select {
 	case <-r.closed:
 		return ErrClosed
-	case r.queue <- request:
+	case r.queue <- packet{bs, from}:
 		return nil
 	default:
 		return errFullQueue
@@ -196,73 +199,80 @@ func (r *Resolver) EnqueueRequest(request Packet) error {
 
 // NextResponse returns a DNS response to a previously enqueued request.
 // It blocks until a response is available and gives up ownership of the response payload.
-func (r *Resolver) NextResponse() (Packet, error) {
+func (r *Resolver) NextResponse() (packet []byte, to netaddr.IPPort, err error) {
 	select {
 	case <-r.closed:
-		return Packet{}, ErrClosed
+		return nil, netaddr.IPPort{}, ErrClosed
 	case resp := <-r.responses:
-		return resp, nil
+		return resp.bs, resp.addr, nil
 	case err := <-r.errors:
-		return Packet{}, err
+		return nil, netaddr.IPPort{}, err
 	}
 }
 
-// Resolve maps a given domain name to the IP address of the host that owns it,
-// if the IP address conforms to the DNS resource type given by tp (one of A, AAAA, ALL).
-// The domain name must be in canonical form (with a trailing period).
-func (r *Resolver) Resolve(domain string, tp dns.Type) (netaddr.IP, dns.RCode, error) {
+// resolveLocal returns an IP for the given domain, if domain is in
+// the local hosts map and has an IP corresponding to the requested
+// typ (A, AAAA, ALL).
+// Returns dns.RCodeRefused to indicate that the local map is not
+// authoritative for domain.
+func (r *Resolver) resolveLocal(domain dnsname.FQDN, typ dns.Type) (netaddr.IP, dns.RCode) {
+	// Reject .onion domains per RFC 7686.
+	if dnsname.HasSuffix(domain.WithoutTrailingDot(), ".onion") {
+		return netaddr.IP{}, dns.RCodeNameError
+	}
+
 	r.mu.Lock()
-	dnsMap := r.dnsMap
+	hosts := r.hostToIP
+	localDomains := r.localDomains
 	r.mu.Unlock()
 
-	if dnsMap == nil {
-		return netaddr.IP{}, dns.RCodeServerFailure, errMapNotSet
-	}
-
-	// Reject .onion domains per RFC 7686.
-	if dnsname.HasSuffix(domain, ".onion") {
-		return netaddr.IP{}, dns.RCodeNameError, nil
-	}
-
-	anyHasSuffix := false
-	for _, suffix := range dnsMap.rootDomains {
-		if dnsname.HasSuffix(domain, suffix) {
-			anyHasSuffix = true
-			break
-		}
-	}
-	addr, found := dnsMap.nameToIP[domain]
+	addrs, found := hosts[domain]
 	if !found {
-		if !anyHasSuffix {
-			return netaddr.IP{}, dns.RCodeRefused, nil
+		for _, suffix := range localDomains {
+			if suffix.Contains(domain) {
+				// We are authoritative for the queried domain.
+				return netaddr.IP{}, dns.RCodeNameError
+			}
 		}
-		return netaddr.IP{}, dns.RCodeNameError, nil
+		// Not authoritative, signal that forwarding is advisable.
+		return netaddr.IP{}, dns.RCodeRefused
 	}
 
 	// Refactoring note: this must happen after we check suffixes,
 	// otherwise we will respond with NOTIMP to requests that should be forwarded.
-	switch tp {
+	//
+	// DNS semantics subtlety: when a DNS name exists, but no records
+	// are available for the requested record type, we must return
+	// RCodeSuccess with no data, not NXDOMAIN.
+	switch typ {
 	case dns.TypeA:
-		if !addr.Is4() {
-			return netaddr.IP{}, dns.RCodeSuccess, nil
+		for _, ip := range addrs {
+			if ip.Is4() {
+				return ip, dns.RCodeSuccess
+			}
 		}
-		return addr, dns.RCodeSuccess, nil
+		return netaddr.IP{}, dns.RCodeSuccess
 	case dns.TypeAAAA:
-		if !addr.Is6() {
-			return netaddr.IP{}, dns.RCodeSuccess, nil
+		for _, ip := range addrs {
+			if ip.Is6() {
+				return ip, dns.RCodeSuccess
+			}
 		}
-		return addr, dns.RCodeSuccess, nil
+		return netaddr.IP{}, dns.RCodeSuccess
 	case dns.TypeALL:
 		// Answer with whatever we've got.
 		// It could be IPv4, IPv6, or a zero addr.
 		// TODO: Return all available resolutions (A and AAAA, if we have them).
-		return addr, dns.RCodeSuccess, nil
+		if len(addrs) == 0 {
+			return netaddr.IP{}, dns.RCodeSuccess
+		}
+		return addrs[0], dns.RCodeSuccess
 
 	// Leave some some record types explicitly unimplemented.
 	// These types relate to recursive resolution or special
-	// DNS sematics and might be implemented in the future.
+	// DNS semantics and might be implemented in the future.
 	case dns.TypeNS, dns.TypeSOA, dns.TypeAXFR, dns.TypeHINFO:
-		return netaddr.IP{}, dns.RCodeNotImplemented, errNotImplemented
+		return netaddr.IP{}, dns.RCodeNotImplemented
 
 	// For everything except for the few types above that are explictly not implemented, return no records.
 	// This is what other DNS systems do: always return NOERROR
@@ -271,51 +281,43 @@ func (r *Resolver) Resolve(domain string, tp dns.Type) (netaddr.IP, dns.RCode, e
 	//   dig -t TYPE9824 example.com
 	// and note that NOERROR is returned, despite that record type being made up.
 	default:
-		// no records exist of this type
-		return netaddr.IP{}, dns.RCodeSuccess, nil
+		// The name exists, but no records exist of the requested type.
+		return netaddr.IP{}, dns.RCodeSuccess
 	}
 }
 
-// ResolveReverse returns the unique domain name that maps to the given address.
-// The returned domain name is in canonical form (with a trailing period).
-func (r *Resolver) ResolveReverse(ip netaddr.IP) (string, dns.RCode, error) {
+// resolveReverse returns the unique domain name that maps to the given address.
+func (r *Resolver) resolveLocalReverse(ip netaddr.IP) (dnsname.FQDN, dns.RCode) {
 	r.mu.Lock()
-	dnsMap := r.dnsMap
+	ips := r.ipToHost
 	r.mu.Unlock()
 
-	if dnsMap == nil {
-		return "", dns.RCodeServerFailure, errMapNotSet
-	}
-	name, found := dnsMap.ipToName[ip]
+	name, found := ips[ip]
 	if !found {
-		return "", dns.RCodeNameError, nil
+		return "", dns.RCodeNameError
 	}
-	return name, dns.RCodeSuccess, nil
+	return name, dns.RCodeSuccess
 }
 
 func (r *Resolver) poll() {
 	defer r.wg.Done()
 
-	var packet Packet
+	var pkt packet
 	for {
 		select {
 		case <-r.closed:
 			return
-		case packet = <-r.queue:
+		case pkt = <-r.queue:
 			// continue
 		}
 
-		out, err := r.respond(packet.Payload)
+		out, err := r.respond(pkt.bs)
 
 		if err == errNotOurName {
-			if r.forwarder != nil {
-				err = r.forwarder.forward(packet)
-				if err == nil {
-					// forward will send response into r.responses, nothing to do.
-					continue
-				}
-			} else {
-				err = errNotForwarding
+			err = r.forwarder.forward(pkt)
+			if err == nil {
+				// forward will send response into r.responses, nothing to do.
+				continue
 			}
 		}
 
@@ -327,11 +329,11 @@ func (r *Resolver) poll() {
 				// continue
 			}
 		} else {
-			packet.Payload = out
+			pkt.bs = out
 			select {
 			case <-r.closed:
 				return
-			case r.responses <- packet:
+			case r.responses <- pkt:
 				// continue
 			}
 		}
@@ -342,12 +344,14 @@ type response struct {
 	Header   dns.Header
 	Question dns.Question
 	// Name is the response to a PTR query.
-	Name string
+	Name dnsname.FQDN
 	// IP is the response to an A, AAAA, or ALL query.
 	IP netaddr.IP
 }
 
 // parseQuery parses the query in given packet into a response struct.
+// if the parse is successful, resp.Name contains the normalized name being queried.
+// TODO: stuffing the query name in resp.Name temporarily is a hack. Clean it up.
 func parseQuery(query []byte, resp *response) error {
 	var parser dns.Parser
 	var err error
@@ -403,7 +407,7 @@ func marshalAAAARecord(name dns.Name, ip netaddr.IP, builder *dns.Builder) error
 
 // marshalPTRRecord serializes a PTR record into an active builder.
 // The caller may continue using the builder following the call.
-func marshalPTRRecord(queryName dns.Name, name string, builder *dns.Builder) error {
+func marshalPTRRecord(queryName dns.Name, name dnsname.FQDN, builder *dns.Builder) error {
 	var answer dns.PTRResource
 	var err error
 
@@ -413,7 +417,7 @@ func marshalPTRRecord(queryName dns.Name, name string, builder *dns.Builder) err
 		Class: dns.ClassINET,
 		TTL:   uint32(defaultTTL / time.Second),
 	}
-	answer.PTR, err = dns.NewName(name)
+	answer.PTR, err = dns.NewName(name.WithTrailingDot())
 	if err != nil {
 		return err
 	}
@@ -486,12 +490,13 @@ const (
 //   r._dns-sd._udp.<domain>.
 //  dr._dns-sd._udp.<domain>.
 //  lb._dns-sd._udp.<domain>.
-func hasRDNSBonjourPrefix(s string) bool {
+func hasRDNSBonjourPrefix(name dnsname.FQDN) bool {
 	// Even the shortest name containing a Bonjour prefix is long,
 	// so check length (cheap) and bail early if possible.
-	if len(s) < len("*._dns-sd._udp.0.0.0.0.in-addr.arpa.") {
+	if len(name) < len("*._dns-sd._udp.0.0.0.0.in-addr.arpa.") {
 		return false
 	}
+	s := name.WithTrailingDot()
 	dot := strings.IndexByte(s, '.')
 	if dot == -1 {
 		return false // shouldn't happen
@@ -526,9 +531,9 @@ func rawNameToLower(name []byte) string {
 //   4.3.2.1.in-addr.arpa
 // is transformed to
 //   1.2.3.4
-func rdnsNameToIPv4(name string) (ip netaddr.IP, ok bool) {
-	name = strings.TrimSuffix(name, rdnsv4Suffix)
-	ip, err := netaddr.ParseIP(string(name))
+func rdnsNameToIPv4(name dnsname.FQDN) (ip netaddr.IP, ok bool) {
+	s := strings.TrimSuffix(name.WithTrailingDot(), rdnsv4Suffix)
+	ip, err := netaddr.ParseIP(s)
 	if err != nil {
 		return netaddr.IP{}, false
 	}
@@ -545,21 +550,21 @@ func rdnsNameToIPv4(name string) (ip netaddr.IP, ok bool) {
 //   b.a.9.8.7.6.5.0.0.0.0.0.0.0.0.0.0.0.0.0.0.0.0.0.8.b.d.0.1.0.0.2.ip6.arpa.
 // is transformed to
 //   2001:db8::567:89ab
-func rdnsNameToIPv6(name string) (ip netaddr.IP, ok bool) {
+func rdnsNameToIPv6(name dnsname.FQDN) (ip netaddr.IP, ok bool) {
 	var b [32]byte
 	var ipb [16]byte
 
-	name = strings.TrimSuffix(name, rdnsv6Suffix)
+	s := strings.TrimSuffix(name.WithTrailingDot(), rdnsv6Suffix)
 	// 32 nibbles and 31 dots between them.
-	if len(name) != 63 {
+	if len(s) != 63 {
 		return netaddr.IP{}, false
 	}
 
 	// Dots and hex digits alternate.
 	prevDot := true
 	// i ranges over name backward; j ranges over b forward.
-	for i, j := len(name)-1, 0; i >= 0; i-- {
-		thisDot := (name[i] == '.')
+	for i, j := len(s)-1, 0; i >= 0; i-- {
+		thisDot := (s[i] == '.')
 		if prevDot == thisDot {
 			return netaddr.IP{}, false
 		}
@@ -568,7 +573,7 @@ func rdnsNameToIPv6(name string) (ip netaddr.IP, ok bool) {
 		if !thisDot {
 			// This is safe assuming alternation.
 			// We do not check that non-dots are hex digits: hex.Decode below will do that.
-			b[j] = name[i]
+			b[j] = s[i]
 			j++
 		}
 	}
@@ -583,7 +588,7 @@ func rdnsNameToIPv6(name string) (ip netaddr.IP, ok bool) {
 
 // respondReverse returns a DNS response to a PTR query.
 // It is assumed that resp.Question is populated by respond before this is called.
-func (r *Resolver) respondReverse(query []byte, name string, resp *response) ([]byte, error) {
+func (r *Resolver) respondReverse(query []byte, name dnsname.FQDN, resp *response) ([]byte, error) {
 	if hasRDNSBonjourPrefix(name) {
 		return nil, errNotOurName
 	}
@@ -591,9 +596,9 @@ func (r *Resolver) respondReverse(query []byte, name string, resp *response) ([]
 	var ip netaddr.IP
 	var ok bool
 	switch {
-	case strings.HasSuffix(name, rdnsv4Suffix):
+	case strings.HasSuffix(name.WithTrailingDot(), rdnsv4Suffix):
 		ip, ok = rdnsNameToIPv4(name)
-	case strings.HasSuffix(name, rdnsv6Suffix):
+	case strings.HasSuffix(name.WithTrailingDot(), rdnsv6Suffix):
 		ip, ok = rdnsNameToIPv6(name)
 	default:
 		return nil, errNotOurName
@@ -606,11 +611,7 @@ func (r *Resolver) respondReverse(query []byte, name string, resp *response) ([]
 		return nil, errNotOurName
 	}
 
-	var err error
-	resp.Name, resp.Header.RCode, err = r.ResolveReverse(ip)
-	if err != nil {
-		r.logf("resolving rdns: %v", ip, err)
-	}
+	resp.Name, resp.Header.RCode = r.resolveLocalReverse(ip)
 	if resp.Header.RCode == dns.RCodeNameError {
 		return nil, errNotOurName
 	}
@@ -638,7 +639,12 @@ func (r *Resolver) respond(query []byte) ([]byte, error) {
 		return marshalResponse(resp)
 	}
 	rawName := resp.Question.Name.Data[:resp.Question.Name.Length]
-	name := rawNameToLower(rawName)
+	name, err := dnsname.ToFQDN(rawNameToLower(rawName))
+	if err != nil {
+		// DNS packet unexpectedly contains an invalid FQDN.
+		resp.Header.RCode = dns.RCodeFormatError
+		return marshalResponse(resp)
+	}
 
 	// Always try to handle reverse lookups; delegate inside when not found.
 	// This way, queries for existent nodes do not leak,
@@ -647,16 +653,11 @@ func (r *Resolver) respond(query []byte) ([]byte, error) {
 		return r.respondReverse(query, name, resp)
 	}
 
-	resp.IP, resp.Header.RCode, err = r.Resolve(name, resp.Question.Type)
+	resp.IP, resp.Header.RCode = r.resolveLocal(name, resp.Question.Type)
 	// This return code is special: it requests forwarding.
 	if resp.Header.RCode == dns.RCodeRefused {
 		return nil, errNotOurName
 	}
 
-	// We will not return this error: it is the sender's fault.
-	if err != nil {
-		r.logf("resolving: %v", err)
-	}
-
 	return marshalResponse(resp)
 }

net/dns/resolver/tsdns_server_test.go

@@ -2,10 +2,10 @@
 // Use of this source code is governed by a BSD-style
 // license that can be found in the LICENSE file.
 
-package dns
+package resolver
 
 import (
-	"log"
+	"fmt"
 	"testing"
 
 	"github.com/miekg/dns"
@@ -16,8 +16,6 @@ import (
 // that depends on github.com/miekg/dns
 // from the rest, which only depends on dnsmessage.
 
-var dnsHandleFunc = dns.HandleFunc
-
 // resolveToIP returns a handler function which responds
 // to queries of type A it receives with an A record containing ipv4,
 // to queries of type AAAA with an AAAA record containing ipv6,
@@ -68,28 +66,38 @@ func resolveToIP(ipv4, ipv6 netaddr.IP, ns string) dns.HandlerFunc {
 	}
 }
 
-func resolveToNXDOMAIN(w dns.ResponseWriter, req *dns.Msg) {
+var resolveToNXDOMAIN = dns.HandlerFunc(func(w dns.ResponseWriter, req *dns.Msg) {
 	m := new(dns.Msg)
 	m.SetRcode(req, dns.RcodeNameError)
 	w.WriteMsg(m)
-}
-
-func serveDNS(tb testing.TB, addr string) (*dns.Server, chan error) {
-	server := &dns.Server{Addr: addr, Net: "udp"}
+})
 
+func serveDNS(tb testing.TB, addr string, records ...interface{}) *dns.Server {
+	if len(records)%2 != 0 {
+		panic("must have an even number of record values")
+	}
+	mux := dns.NewServeMux()
+	for i := 0; i < len(records); i += 2 {
+		name := records[i].(string)
+		handler := records[i+1].(dns.Handler)
+		mux.Handle(name, handler)
+	}
 	waitch := make(chan struct{})
-	server.NotifyStartedFunc = func() { close(waitch) }
+	server := &dns.Server{
+		Addr:              addr,
+		Net:               "udp",
+		Handler:           mux,
+		NotifyStartedFunc: func() { close(waitch) },
+		ReusePort:         true,
+	}
 
-	errch := make(chan error, 1)
 	go func() {
 		err := server.ListenAndServe()
 		if err != nil {
-			log.Printf("ListenAndServe(%q): %v", addr, err)
+			panic(fmt.Sprintf("ListenAndServe(%q): %v", addr, err))
 		}
-		errch <- err
-		close(errch)
 	}()
 
 	<-waitch
-	return server, errch
+	return server
 }

net/dns/resolver/tsdns_test.go

@@ -2,40 +2,35 @@
 // Use of this source code is governed by a BSD-style
 // license that can be found in the LICENSE file.
 
-package dns
+package resolver
 
 import (
 	"bytes"
 	"errors"
 	"net"
-	"sync"
 	"testing"
 
 	dns "golang.org/x/net/dns/dnsmessage"
 	"inet.af/netaddr"
 	"tailscale.com/tstest"
+	"tailscale.com/util/dnsname"
 )
 
-var testipv4 = netaddr.IPv4(1, 2, 3, 4)
-var testipv6 = netaddr.IPv6Raw([16]byte{
-	0x00, 0x01, 0x02, 0x03,
-	0x04, 0x05, 0x06, 0x07,
-	0x08, 0x09, 0x0a, 0x0b,
-	0x0c, 0x0d, 0x0e, 0x0f,
-})
+var testipv4 = netaddr.MustParseIP("1.2.3.4")
+var testipv6 = netaddr.MustParseIP("0001:0203:0405:0607:0809:0a0b:0c0d:0e0f")
 
-var dnsMap = NewMap(
-	map[string]netaddr.IP{
-		"test1.ipn.dev.": testipv4,
-		"test2.ipn.dev.": testipv6,
+var dnsCfg = Config{
+	Hosts: map[dnsname.FQDN][]netaddr.IP{
+		"test1.ipn.dev.": []netaddr.IP{testipv4},
+		"test2.ipn.dev.": []netaddr.IP{testipv6},
 	},
-	[]string{"ipn.dev."},
-)
+	LocalDomains: []dnsname.FQDN{"ipn.dev."},
+}
 
-func dnspacket(domain string, tp dns.Type) []byte {
+func dnspacket(domain dnsname.FQDN, tp dns.Type) []byte {
 	var dnsHeader dns.Header
 	question := dns.Question{
-		Name:  dns.MustNewName(domain),
+		Name:  dns.MustNewName(domain.WithTrailingDot()),
 		Type:  tp,
 		Class: dns.ClassINET,
 	}
@@ -50,7 +45,7 @@ func dnspacket(domain string, tp dns.Type) []byte {
 
 type dnsResponse struct {
 	ip    netaddr.IP
-	name  string
+	name  dnsname.FQDN
 	rcode dns.RCode
 }
 
@@ -100,7 +95,10 @@ func unpackResponse(payload []byte) (dnsResponse, error) {
 		if err != nil {
 			return response, err
 		}
-		response.name = res.NS.String()
+		response.name, err = dnsname.ToFQDN(res.NS.String())
+		if err != nil {
+			return response, err
+		}
 	default:
 		return response, errors.New("type not in {A, AAAA, NS}")
 	}
@@ -109,10 +107,9 @@ func unpackResponse(payload []byte) (dnsResponse, error) {
 }
 
 func syncRespond(r *Resolver, query []byte) ([]byte, error) {
-	request := Packet{Payload: query}
-	r.EnqueueRequest(request)
-	resp, err := r.NextResponse()
-	return resp.Payload, err
+	r.EnqueueRequest(query, netaddr.IPPort{})
+	payload, _, err := r.NextResponse()
+	return payload, err
 }
 
 func mustIP(str string) netaddr.IP {
@@ -126,7 +123,7 @@ func mustIP(str string) netaddr.IP {
 func TestRDNSNameToIPv4(t *testing.T) {
 	tests := []struct {
 		name   string
-		input  string
+		input  dnsname.FQDN
 		wantIP netaddr.IP
 		wantOK bool
 	}{
@@ -151,7 +148,7 @@ func TestRDNSNameToIPv4(t *testing.T) {
 func TestRDNSNameToIPv6(t *testing.T) {
 	tests := []struct {
 		name   string
-		input  string
+		input  dnsname.FQDN
 		wantIP netaddr.IP
 		wantOK bool
 	}{
@@ -193,18 +190,15 @@ func TestRDNSNameToIPv6(t *testing.T) {
 	}
 }
 
-func TestResolve(t *testing.T) {
-	r := NewResolver(ResolverConfig{Logf: t.Logf, Forward: false})
-	r.SetMap(dnsMap)
-
-	if err := r.Start(); err != nil {
-		t.Fatalf("start: %v", err)
-	}
+func TestResolveLocal(t *testing.T) {
+	r := New(t.Logf, nil)
 	defer r.Close()
 
+	r.SetConfig(dnsCfg)
+
 	tests := []struct {
 		name  string
-		qname string
+		qname dnsname.FQDN
 		qtype dns.Type
 		ip    netaddr.IP
 		code  dns.RCode
@@ -224,10 +218,7 @@ func TestResolve(t *testing.T) {
 
 	for _, tt := range tests {
 		t.Run(tt.name, func(t *testing.T) {
-			ip, code, err := r.Resolve(tt.qname, tt.qtype)
-			if err != nil {
-				t.Errorf("err = %v; want nil", err)
-			}
+			ip, code := r.resolveLocal(tt.qname, tt.qtype)
 			if code != tt.code {
 				t.Errorf("code = %v; want %v", code, tt.code)
 			}
@@ -239,19 +230,16 @@ func TestResolve(t *testing.T) {
 	}
 }
 
-func TestResolveReverse(t *testing.T) {
-	r := NewResolver(ResolverConfig{Logf: t.Logf, Forward: false})
-	r.SetMap(dnsMap)
-
-	if err := r.Start(); err != nil {
-		t.Fatalf("start: %v", err)
-	}
+func TestResolveLocalReverse(t *testing.T) {
+	r := New(t.Logf, nil)
 	defer r.Close()
 
+	r.SetConfig(dnsCfg)
+
 	tests := []struct {
 		name string
 		ip   netaddr.IP
-		want string
+		want dnsname.FQDN
 		code dns.RCode
 	}{
 		{"ipv4", testipv4, "test1.ipn.dev.", dns.RCodeSuccess},
@@ -261,10 +249,7 @@ func TestResolveReverse(t *testing.T) {
 
 	for _, tt := range tests {
 		t.Run(tt.name, func(t *testing.T) {
-			name, code, err := r.ResolveReverse(tt.ip)
-			if err != nil {
-				t.Errorf("err = %v; want nil", err)
-			}
+			name, code := r.resolveLocalReverse(tt.ip)
 			if code != tt.code {
 				t.Errorf("code = %v; want %v", code, tt.code)
 			}
@@ -291,45 +276,27 @@ func TestDelegate(t *testing.T) {
 		t.Skip("skipping test that requires localhost IPv6")
 	}
 
-	dnsHandleFunc("test.site.", resolveToIP(testipv4, testipv6, "dns.test.site."))
-	dnsHandleFunc("nxdomain.site.", resolveToNXDOMAIN)
+	v4server := serveDNS(t, "127.0.0.1:0",
+		"test.site.", resolveToIP(testipv4, testipv6, "dns.test.site."),
+		"nxdomain.site.", resolveToNXDOMAIN)
+	defer v4server.Shutdown()
+	v6server := serveDNS(t, "[::1]:0",
+		"test.site.", resolveToIP(testipv4, testipv6, "dns.test.site."),
+		"nxdomain.site.", resolveToNXDOMAIN)
+	defer v6server.Shutdown()
 
-	v4server, v4errch := serveDNS(t, "127.0.0.1:0")
-	v6server, v6errch := serveDNS(t, "[::1]:0")
-
-	defer func() {
-		if err := <-v4errch; err != nil {
-			t.Errorf("v4 server error: %v", err)
-		}
-		if err := <-v6errch; err != nil {
-			t.Errorf("v6 server error: %v", err)
-		}
-	}()
-	if v4server != nil {
-		defer v4server.Shutdown()
-	}
-	if v6server != nil {
-		defer v6server.Shutdown()
-	}
-
-	if v4server == nil || v6server == nil {
-		// There is an error in at least one of the channels
-		// and we cannot proceed; return to see it.
-		return
-	}
-
-	r := NewResolver(ResolverConfig{Logf: t.Logf, Forward: true})
-	r.SetMap(dnsMap)
-	r.SetUpstreams([]net.Addr{
-		v4server.PacketConn.LocalAddr(),
-		v6server.PacketConn.LocalAddr(),
-	})
-
-	if err := r.Start(); err != nil {
-		t.Fatalf("start: %v", err)
-	}
+	r := New(t.Logf, nil)
 	defer r.Close()
 
+	cfg := dnsCfg
+	cfg.Routes = map[dnsname.FQDN][]netaddr.IPPort{
+		".": {
+			netaddr.MustParseIPPort(v4server.PacketConn.LocalAddr().String()),
+			netaddr.MustParseIPPort(v6server.PacketConn.LocalAddr().String()),
+		},
+	}
+	r.SetConfig(cfg)
+
 	tests := []struct {
 		title    string
 		query    []byte
@@ -382,32 +349,87 @@ func TestDelegate(t *testing.T) {
 	}
 }
 
-func TestDelegateCollision(t *testing.T) {
-	dnsHandleFunc("test.site.", resolveToIP(testipv4, testipv6, "dns.test.site."))
+func TestDelegateSplitRoute(t *testing.T) {
+	test4 := netaddr.MustParseIP("2.3.4.5")
+	test6 := netaddr.MustParseIP("ff::1")
 
-	server, errch := serveDNS(t, "127.0.0.1:0")
-	defer func() {
-		if err := <-errch; err != nil {
-			t.Errorf("server error: %v", err)
-		}
-	}()
+	server1 := serveDNS(t, "127.0.0.1:0",
+		"test.site.", resolveToIP(testipv4, testipv6, "dns.test.site."))
+	defer server1.Shutdown()
+	server2 := serveDNS(t, "127.0.0.1:0",
+		"test.other.", resolveToIP(test4, test6, "dns.other."))
+	defer server2.Shutdown()
 
-	if server == nil {
-		return
-	}
-	defer server.Shutdown()
-
-	r := NewResolver(ResolverConfig{Logf: t.Logf, Forward: true})
-	r.SetMap(dnsMap)
-	r.SetUpstreams([]net.Addr{server.PacketConn.LocalAddr()})
-
-	if err := r.Start(); err != nil {
-		t.Fatalf("start: %v", err)
-	}
+	r := New(t.Logf, nil)
 	defer r.Close()
 
+	cfg := dnsCfg
+	cfg.Routes = map[dnsname.FQDN][]netaddr.IPPort{
+		".":      {netaddr.MustParseIPPort(server1.PacketConn.LocalAddr().String())},
+		"other.": {netaddr.MustParseIPPort(server2.PacketConn.LocalAddr().String())},
+	}
+	r.SetConfig(cfg)
+
+	tests := []struct {
+		title    string
+		query    []byte
+		response dnsResponse
+	}{
+		{
+			"general",
+			dnspacket("test.site.", dns.TypeA),
+			dnsResponse{ip: testipv4, rcode: dns.RCodeSuccess},
+		},
+		{
+			"override",
+			dnspacket("test.other.", dns.TypeA),
+			dnsResponse{ip: test4, rcode: dns.RCodeSuccess},
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.title, func(t *testing.T) {
+			payload, err := syncRespond(r, tt.query)
+			if err != nil {
+				t.Errorf("err = %v; want nil", err)
+				return
+			}
+			response, err := unpackResponse(payload)
+			if err != nil {
+				t.Errorf("extract: err = %v; want nil (in %x)", err, payload)
+				return
+			}
+			if response.rcode != tt.response.rcode {
+				t.Errorf("rcode = %v; want %v", response.rcode, tt.response.rcode)
+			}
+			if response.ip != tt.response.ip {
+				t.Errorf("ip = %v; want %v", response.ip, tt.response.ip)
+			}
+			if response.name != tt.response.name {
+				t.Errorf("name = %v; want %v", response.name, tt.response.name)
+			}
+		})
+	}
+}
+
+func TestDelegateCollision(t *testing.T) {
+	server := serveDNS(t, "127.0.0.1:0",
+		"test.site.", resolveToIP(testipv4, testipv6, "dns.test.site."))
+	defer server.Shutdown()
+
+	r := New(t.Logf, nil)
+	defer r.Close()
+
+	cfg := dnsCfg
+	cfg.Routes = map[dnsname.FQDN][]netaddr.IPPort{
+		".": {
+			netaddr.MustParseIPPort(server.PacketConn.LocalAddr().String()),
+		},
+	}
+	r.SetConfig(cfg)
+
 	packets := []struct {
-		qname string
+		qname dnsname.FQDN
 		qtype dns.Type
 		addr  netaddr.IPPort
 	}{
@@ -418,21 +440,20 @@ func TestDelegateCollision(t *testing.T) {
 	// packets will have the same dns txid.
 	for _, p := range packets {
 		payload := dnspacket(p.qname, p.qtype)
-		req := Packet{Payload: payload, Addr: p.addr}
-		err := r.EnqueueRequest(req)
+		err := r.EnqueueRequest(payload, p.addr)
 		if err != nil {
 			t.Error(err)
 		}
 	}
 
 	// Despite the txid collision, the answer(s) should still match the query.
-	resp, err := r.NextResponse()
+	resp, addr, err := r.NextResponse()
 	if err != nil {
 		t.Error(err)
 	}
 
 	var p dns.Parser
-	_, err = p.Start(resp.Payload)
+	_, err = p.Start(resp)
 	if err != nil {
 		t.Error(err)
 	}
@@ -456,72 +477,12 @@ func TestDelegateCollision(t *testing.T) {
 	}
 
 	for _, p := range packets {
-		if p.qtype == wantType && p.addr != resp.Addr {
-			t.Errorf("addr = %v; want %v", resp.Addr, p.addr)
+		if p.qtype == wantType && p.addr != addr {
+			t.Errorf("addr = %v; want %v", addr, p.addr)
 		}
 	}
 }
 
-func TestConcurrentSetMap(t *testing.T) {
-	r := NewResolver(ResolverConfig{Logf: t.Logf, Forward: false})
-
-	if err := r.Start(); err != nil {
-		t.Fatalf("start: %v", err)
-	}
-	defer r.Close()
-
-	// This is purely to ensure that Resolve does not race with SetMap.
-	var wg sync.WaitGroup
-	wg.Add(2)
-	go func() {
-		defer wg.Done()
-		r.SetMap(dnsMap)
-	}()
-	go func() {
-		defer wg.Done()
-		r.Resolve("test1.ipn.dev", dns.TypeA)
-	}()
-	wg.Wait()
-}
-
-func TestConcurrentSetUpstreams(t *testing.T) {
-	dnsHandleFunc("test.site.", resolveToIP(testipv4, testipv6, "dns.test.site."))
-
-	server, errch := serveDNS(t, "127.0.0.1:0")
-	defer func() {
-		if err := <-errch; err != nil {
-			t.Errorf("server error: %v", err)
-		}
-	}()
-
-	if server == nil {
-		return
-	}
-	defer server.Shutdown()
-
-	r := NewResolver(ResolverConfig{Logf: t.Logf, Forward: true})
-	r.SetMap(dnsMap)
-
-	if err := r.Start(); err != nil {
-		t.Fatalf("start: %v", err)
-	}
-	defer r.Close()
-
-	packet := dnspacket("test.site.", dns.TypeA)
-	// This is purely to ensure that delegation does not race with SetUpstreams.
-	var wg sync.WaitGroup
-	wg.Add(2)
-	go func() {
-		defer wg.Done()
-		r.SetUpstreams([]net.Addr{server.PacketConn.LocalAddr()})
-	}()
-	go func() {
-		defer wg.Done()
-		syncRespond(r, packet)
-	}()
-	wg.Wait()
-}
-
 var allResponse = []byte{
 	0x00, 0x00, // transaction id: 0
 	0x84, 0x00, // flags: response, authoritative, no error
@@ -670,14 +631,11 @@ var emptyResponse = []byte{
 }
 
 func TestFull(t *testing.T) {
-	r := NewResolver(ResolverConfig{Logf: t.Logf, Forward: false})
-	r.SetMap(dnsMap)
-
-	if err := r.Start(); err != nil {
-		t.Fatalf("start: %v", err)
-	}
+	r := New(t.Logf, nil)
 	defer r.Close()
 
+	r.SetConfig(dnsCfg)
+
 	// One full packet and one error packet
 	tests := []struct {
 		name     string
@@ -689,8 +647,8 @@ func TestFull(t *testing.T) {
 		{"ipv6", dnspacket("test2.ipn.dev.", dns.TypeAAAA), ipv6Response},
 		{"no-ipv6", dnspacket("test1.ipn.dev.", dns.TypeAAAA), emptyResponse},
 		{"upper", dnspacket("TEST1.IPN.DEV.", dns.TypeA), ipv4UppercaseResponse},
-		{"ptr", dnspacket("4.3.2.1.in-addr.arpa.", dns.TypePTR), ptrResponse},
-		{"ptr", dnspacket("f.0.e.0.d.0.c.0.b.0.a.0.9.0.8.0.7.0.6.0.5.0.4.0.3.0.2.0.1.0.0.0.ip6.arpa.",
+		{"ptr4", dnspacket("4.3.2.1.in-addr.arpa.", dns.TypePTR), ptrResponse},
+		{"ptr6", dnspacket("f.0.e.0.d.0.c.0.b.0.a.0.9.0.8.0.7.0.6.0.5.0.4.0.3.0.2.0.1.0.0.0.ip6.arpa.",
 			dns.TypePTR), ptrResponse6},
 		{"nxdomain", dnspacket("test3.ipn.dev.", dns.TypeA), nxdomainResponse},
 	}
@@ -709,13 +667,9 @@ func TestFull(t *testing.T) {
 }
 
 func TestAllocs(t *testing.T) {
-	r := NewResolver(ResolverConfig{Logf: t.Logf, Forward: false})
-	r.SetMap(dnsMap)
-
-	if err := r.Start(); err != nil {
-		t.Fatalf("start: %v", err)
-	}
+	r := New(t.Logf, nil)
 	defer r.Close()
+	r.SetConfig(dnsCfg)
 
 	// It is seemingly pointless to test allocs in the delegate path,
 	// as dialer.Dial -> Read -> Write alone comprise 12 allocs.
@@ -742,7 +696,7 @@ func TestAllocs(t *testing.T) {
 
 func TestTrimRDNSBonjourPrefix(t *testing.T) {
 	tests := []struct {
-		in   string
+		in   dnsname.FQDN
 		want bool
 	}{
 		{"b._dns-sd._udp.0.10.20.172.in-addr.arpa.", true},
@@ -752,7 +706,6 @@ func TestTrimRDNSBonjourPrefix(t *testing.T) {
 		{"lb._dns-sd._udp.0.10.20.172.in-addr.arpa.", true},
 		{"qq._dns-sd._udp.0.10.20.172.in-addr.arpa.", false},
 		{"0.10.20.172.in-addr.arpa.", false},
-		{"i-have-no-dot", false},
 	}
 
 	for _, test := range tests {
@@ -764,29 +717,20 @@ func TestTrimRDNSBonjourPrefix(t *testing.T) {
 }
 
 func BenchmarkFull(b *testing.B) {
-	dnsHandleFunc("test.site.", resolveToIP(testipv4, testipv6, "dns.test.site."))
-
-	server, errch := serveDNS(b, "127.0.0.1:0")
-	defer func() {
-		if err := <-errch; err != nil {
-			b.Errorf("server error: %v", err)
-		}
-	}()
-
-	if server == nil {
-		return
-	}
+	server := serveDNS(b, "127.0.0.1:0",
+		"test.site.", resolveToIP(testipv4, testipv6, "dns.test.site."))
 	defer server.Shutdown()
 
-	r := NewResolver(ResolverConfig{Logf: b.Logf, Forward: true})
-	r.SetMap(dnsMap)
-	r.SetUpstreams([]net.Addr{server.PacketConn.LocalAddr()})
-
-	if err := r.Start(); err != nil {
-		b.Fatalf("start: %v", err)
-	}
+	r := New(b.Logf, nil)
 	defer r.Close()
 
+	cfg := dnsCfg
+	cfg.Routes = map[dnsname.FQDN][]netaddr.IPPort{
+		".": {
+			netaddr.MustParseIPPort(server.PacketConn.LocalAddr().String()),
+		},
+	}
+
 	tests := []struct {
 		name    string
 		request []byte

net/interfaces/interfaces.go

@@ -500,7 +500,8 @@ func isPrivateIP(ip netaddr.IP) bool {
 }
 
 func isGlobalV6(ip netaddr.IP) bool {
-	return v6Global1.Contains(ip)
+	return v6Global1.Contains(ip) ||
+		(tsaddr.IsULA(ip) && !tsaddr.TailscaleULARange().Contains(ip))
 }
 
 func mustCIDR(s string) netaddr.IPPrefix {

net/interfaces/interfaces_default_route_test.go

@@ -2,16 +2,85 @@
 // Use of this source code is governed by a BSD-style
 // license that can be found in the LICENSE file.
 
-// +build linux darwin,!redo
+// +build linux,!redo
 
 package interfaces
 
-import "testing"
+import (
+	"fmt"
+	"io/ioutil"
+	"os"
+	"path/filepath"
+	"testing"
+)
 
 func TestDefaultRouteInterface(t *testing.T) {
+	// tests /proc/net/route on the local system, cannot make an assertion about
+	// the correct interface name, but good as a sanity check.
 	v, err := DefaultRouteInterface()
 	if err != nil {
 		t.Fatal(err)
 	}
 	t.Logf("got %q", v)
 }
+
+// test the specific /proc/net/route path as found on Google Cloud Run instances
+func TestGoogleCloudRunDefaultRouteInterface(t *testing.T) {
+	dir := t.TempDir()
+	savedProcNetRoutePath := procNetRoutePath
+	defer func() { procNetRoutePath = savedProcNetRoutePath }()
+	procNetRoutePath = filepath.Join(dir, "CloudRun")
+	buf := []byte("Iface\tDestination\tGateway\tFlags\tRefCnt\tUse\tMetric\tMask\tMTU\tWindow\tIRTT\n" +
+		"eth0\t8008FEA9\t00000000\t0001\t0\t0\t0\t01FFFFFF\t0\t0\t0\n" +
+		"eth1\t00000000\t00000000\t0001\t0\t0\t0\t00000000\t0\t0\t0\n")
+	err := ioutil.WriteFile(procNetRoutePath, buf, 0644)
+	if err != nil {
+		t.Fatal(err)
+	}
+	got, err := DefaultRouteInterface()
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	if got != "eth1" {
+		t.Fatalf("got %s, want eth1", got)
+	}
+}
+
+// we read chunks of /proc/net/route at a time, test that files longer than the chunk
+// size can be handled.
+func TestExtremelyLongProcNetRoute(t *testing.T) {
+	dir := t.TempDir()
+	savedProcNetRoutePath := procNetRoutePath
+	defer func() { procNetRoutePath = savedProcNetRoutePath }()
+	procNetRoutePath = filepath.Join(dir, "VeryLong")
+	f, err := os.Create(procNetRoutePath)
+	if err != nil {
+		t.Fatal(err)
+	}
+	_, err = f.Write([]byte("Iface\tDestination\tGateway\tFlags\tRefCnt\tUse\tMetric\tMask\tMTU\tWindow\tIRTT\n"))
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	for n := 0; n <= 1000; n++ {
+		line := fmt.Sprintf("eth%d\t8008FEA9\t00000000\t0001\t0\t0\t0\t01FFFFFF\t0\t0\t0\n", n)
+		_, err := f.Write([]byte(line))
+		if err != nil {
+			t.Fatal(err)
+		}
+	}
+	_, err = f.Write([]byte("tokenring1\t00000000\t00000000\t0001\t0\t0\t0\t00000000\t0\t0\t0\n"))
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	got, err := DefaultRouteInterface()
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	if got != "tokenring1" {
+		t.Fatalf("got %q, want tokenring1", got)
+	}
+}

net/interfaces/interfaces_linux.go

@@ -8,6 +8,7 @@ import (
 	"bufio"
 	"bytes"
 	"errors"
+	"fmt"
 	"io"
 	"log"
 	"os"
@@ -135,18 +136,20 @@ func DefaultRouteInterface() (string, error) {
 }
 
 var zeroRouteBytes = []byte("00000000")
+var procNetRoutePath = "/proc/net/route"
 
-func defaultRouteInterfaceProcNet() (string, error) {
-	f, err := os.Open("/proc/net/route")
+func defaultRouteInterfaceProcNetInternal(bufsize int) (string, error) {
+	f, err := os.Open(procNetRoutePath)
 	if err != nil {
 		return "", err
 	}
 	defer f.Close()
-	br := bufio.NewReaderSize(f, 128)
+
+	br := bufio.NewReaderSize(f, bufsize)
 	for {
 		line, err := br.ReadSlice('\n')
 		if err == io.EOF {
-			break
+			return "", fmt.Errorf("no default routes found: %w", err)
 		}
 		if err != nil {
 			return "", err
@@ -168,9 +171,28 @@ func defaultRouteInterfaceProcNet() (string, error) {
 			return ifc, nil // interface name
 		}
 	}
+}
 
-	return "", errors.New("no default routes found")
-
+// returns string interface name and an error.
+// io.EOF: full route table processed, no default route found.
+// other io error: something went wrong reading the route file.
+func defaultRouteInterfaceProcNet() (string, error) {
+	rc, err := defaultRouteInterfaceProcNetInternal(128)
+	if rc == "" && (errors.Is(err, io.EOF) || err == nil) {
+		// https://github.com/google/gvisor/issues/5732
+		// On a regular Linux kernel you can read the first 128 bytes of /proc/net/route,
+		// then come back later to read the next 128 bytes and so on.
+		//
+		// In Google Cloud Run, where /proc/net/route comes from gVisor, you have to
+		// read it all at once. If you read only the first few bytes then the second
+		// read returns 0 bytes no matter how much originally appeared to be in the file.
+		//
+		// At the time of this writing (Mar 2021) Google Cloud Run has eth0 and eth1
+		// with a 384 byte /proc/net/route. We allocate a large buffer to ensure we'll
+		// read it all in one call.
+		return defaultRouteInterfaceProcNetInternal(4096)
+	}
+	return rc, err
 }
 
 // defaultRouteInterfaceAndroidIPRoute tries to find the machine's default route interface name

net/interfaces/interfaces_test.go

@@ -7,6 +7,8 @@ package interfaces
 import (
 	"encoding/json"
 	"testing"
+
+	"inet.af/netaddr"
 )
 
 func TestGetState(t *testing.T) {
@@ -43,3 +45,24 @@ func TestLikelyHomeRouterIP(t *testing.T) {
 	}
 	t.Logf("myIP = %v; gw = %v", my, gw)
 }
+
+func TestIsGlobalV6(t *testing.T) {
+	tests := []struct {
+		name string
+		ip   string
+		want bool
+	}{
+		{"first ULA", "fc00::1", true},
+		{"Tailscale", "fd7a:115c:a1e0::1", false},
+		{"Cloud Run", "fddf:3978:feb1:d745::1", true},
+		{"zeros", "0000:0000:0000:0000:0000:0000:0000:0000", false},
+		{"Link Local", "fe80::1", false},
+		{"Global", "2602::1", true},
+	}
+
+	for _, test := range tests {
+		if got := isGlobalV6(netaddr.MustParseIP(test.ip)); got != test.want {
+			t.Errorf("isGlobalV6(%s) = %v, want %v", test.name, got, test.want)
+		}
+	}
+}

net/nettest/conn.go

@@ -5,21 +5,13 @@
 package nettest
 
 import (
-	"io"
+	"net"
 	"time"
 )
 
-// Conn is a bi-directional in-memory stream that looks like a TCP net.Conn.
+// Conn is a net.Conn that can additionally have its reads and writes blocked and unblocked.
 type Conn interface {
-	io.Reader
-	io.Writer
-	io.Closer
-
-	// The *Deadline methods follow the semantics of net.Conn.
-
-	SetDeadline(t time.Time) error
-	SetReadDeadline(t time.Time) error
-	SetWriteDeadline(t time.Time) error
+	net.Conn
 
 	// SetReadBlock blocks or unblocks the Read method of this Conn.
 	// It reports an error if the existing value matches the new value,
@@ -40,24 +32,37 @@ func NewConn(name string, maxBuf int) (Conn, Conn) {
 	return &connHalf{r: r, w: w}, &connHalf{r: w, w: r}
 }
 
+type connAddr string
+
+func (a connAddr) Network() string { return "mem" }
+func (a connAddr) String() string  { return string(a) }
+
 type connHalf struct {
 	r, w *Pipe
 }
 
+func (c *connHalf) LocalAddr() net.Addr {
+	return connAddr(c.r.name)
+}
+
+func (c *connHalf) RemoteAddr() net.Addr {
+	return connAddr(c.w.name)
+}
+
 func (c *connHalf) Read(b []byte) (n int, err error) {
 	return c.r.Read(b)
 }
 func (c *connHalf) Write(b []byte) (n int, err error) {
 	return c.w.Write(b)
 }
+
 func (c *connHalf) Close() error {
-	err1 := c.r.Close()
-	err2 := c.w.Close()
-	if err1 != nil {
-		return err1
+	if err := c.w.Close(); err != nil {
+		return err
 	}
-	return err2
+	return c.r.Close()
 }
+
 func (c *connHalf) SetDeadline(t time.Time) error {
 	err1 := c.SetReadDeadline(t)
 	err2 := c.SetWriteDeadline(t)
@@ -72,6 +77,7 @@ func (c *connHalf) SetReadDeadline(t time.Time) error {
 func (c *connHalf) SetWriteDeadline(t time.Time) error {
 	return c.w.SetWriteDeadline(t)
 }
+
 func (c *connHalf) SetReadBlock(b bool) error {
 	if b {
 		return c.r.Block()

net/nettest/conn_test.go

@@ -0,0 +1,22 @@
+// Copyright (c) 2021 Tailscale Inc & AUTHORS All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package nettest
+
+import (
+	"net"
+	"testing"
+
+	"golang.org/x/net/nettest"
+)
+
+func TestConn(t *testing.T) {
+	nettest.TestConn(t, func() (c1 net.Conn, c2 net.Conn, stop func(), err error) {
+		c1, c2 = NewConn("test", bufferSize)
+		return c1, c2, func() {
+			c1.Close()
+			c2.Close()
+		}, nil
+	})
+}

net/nettest/listener.go

@@ -0,0 +1,83 @@
+// Copyright (c) 2021 Tailscale Inc & AUTHORS All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package nettest
+
+import (
+	"context"
+	"net"
+	"strings"
+	"sync"
+)
+
+const (
+	bufferSize = 256 * 1024
+)
+
+// Listener is a net.Listener using using NewConn to create pairs of network
+// connections connected in memory using a buffered pipe. It also provides a
+// Dial method to establish new connections.
+type Listener struct {
+	addr      connAddr
+	ch        chan Conn
+	closeOnce sync.Once
+	closed    chan struct{}
+}
+
+// Listen returns a new Listener for the provided address.
+func Listen(addr string) *Listener {
+	return &Listener{
+		addr:   connAddr(addr),
+		ch:     make(chan Conn),
+		closed: make(chan struct{}),
+	}
+}
+
+// Addr implements net.Listener.Addr.
+func (l *Listener) Addr() net.Addr {
+	return l.addr
+}
+
+// Close closes the pipe listener.
+func (l *Listener) Close() error {
+	l.closeOnce.Do(func() {
+		close(l.closed)
+	})
+	return nil
+}
+
+// Accept blocks until a new connection is available or the listener is closed.
+func (l *Listener) Accept() (net.Conn, error) {
+	select {
+	case c := <-l.ch:
+		return c, nil
+	case <-l.closed:
+		return nil, net.ErrClosed
+	}
+}
+
+// Dial connects to the listener using the provided context.
+// The provided Context must be non-nil. If the context expires before the
+// connection is complete, an error is returned. Once successfully connected
+// any expiration of the context will not affect the connection.
+func (l *Listener) Dial(ctx context.Context, network, addr string) (net.Conn, error) {
+	if !strings.HasSuffix(network, "tcp") {
+		return nil, net.UnknownNetworkError(network)
+	}
+	if connAddr(addr) != l.addr {
+		return nil, &net.AddrError{
+			Err:  "invalid address",
+			Addr: addr,
+		}
+	}
+	c, s := NewConn(addr, bufferSize)
+	select {
+	case <-ctx.Done():
+		return nil, ctx.Err()
+	case <-l.closed:
+		return nil, net.ErrClosed
+	case l.ch <- s:
+		return c, nil
+	}
+}

net/nettest/listener_test.go

@@ -0,0 +1,34 @@
+// Copyright (c) 2021 Tailscale Inc & AUTHORS All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package nettest
+
+import (
+	"context"
+	"testing"
+)
+
+func TestListener(t *testing.T) {
+	l := Listen("srv.local")
+	defer l.Close()
+	go func() {
+		c, err := l.Accept()
+		if err != nil {
+			t.Error(err)
+			return
+		}
+		defer c.Close()
+	}()
+
+	if c, err := l.Dial(context.Background(), "tcp", "invalid"); err == nil {
+		c.Close()
+		t.Fatalf("dial to invalid address succeeded")
+	}
+	c, err := l.Dial(context.Background(), "tcp", "srv.local")
+	if err != nil {
+		t.Fatalf("dial failed: %v", err)
+		return
+	}
+	c.Close()
+}

net/nettest/pipe.go

@@ -5,11 +5,13 @@
 package nettest
 
 import (
+	"bytes"
 	"context"
-	"errors"
 	"fmt"
 	"io"
 	"log"
+	"net"
+	"os"
 	"sync"
 	"time"
 )
@@ -20,13 +22,12 @@ const debugPipe = false
 type Pipe struct {
 	name   string
 	maxBuf int
-	rCh    chan struct{}
-	wCh    chan struct{}
+	mu     sync.Mutex
+	cnd    *sync.Cond
 
-	mu               sync.Mutex
-	closed           bool
 	blocked          bool
-	buf              []byte
+	closed           bool
+	buf              bytes.Buffer
 	readTimeout      time.Time
 	writeTimeout     time.Time
 	cancelReadTimer  func()
@@ -35,21 +36,42 @@ type Pipe struct {
 
 // NewPipe creates a Pipe with a buffer size fixed at maxBuf.
 func NewPipe(name string, maxBuf int) *Pipe {
-	return &Pipe{
+	p := &Pipe{
 		name:   name,
 		maxBuf: maxBuf,
-		rCh:    make(chan struct{}, 1),
-		wCh:    make(chan struct{}, 1),
 	}
+	p.cnd = sync.NewCond(&p.mu)
+	return p
 }
 
-var (
-	ErrTimeout      = errors.New("timeout")
-	ErrReadTimeout  = fmt.Errorf("read %w", ErrTimeout)
-	ErrWriteTimeout = fmt.Errorf("write %w", ErrTimeout)
-)
+// readOrBlock attempts to read from the buffer, if the buffer is empty and
+// the connection hasn't been closed it will block until there is a change.
+func (p *Pipe) readOrBlock(b []byte) (int, error) {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	if !p.readTimeout.IsZero() && !time.Now().Before(p.readTimeout) {
+		return 0, os.ErrDeadlineExceeded
+	}
+	if p.blocked {
+		p.cnd.Wait()
+		return 0, nil
+	}
+
+	n, err := p.buf.Read(b)
+	// err will either be nil or io.EOF.
+	if err == io.EOF {
+		if p.closed {
+			return n, err
+		}
+		// Wait for something to change.
+		p.cnd.Wait()
+	}
+	return n, nil
+}
 
 // Read implements io.Reader.
+// Once the buffer is drained (i.e. after Close), subsequent calls will
+// return io.EOF.
 func (p *Pipe) Read(b []byte) (n int, err error) {
 	if debugPipe {
 		orig := b
@@ -57,35 +79,48 @@ func (p *Pipe) Read(b []byte) (n int, err error) {
 			log.Printf("Pipe(%q).Read( %q) n=%d, err=%v", p.name, string(orig[:n]), n, err)
 		}()
 	}
-	for {
-		p.mu.Lock()
-		closed := p.closed
-		timedout := !p.readTimeout.IsZero() && !time.Now().Before(p.readTimeout)
-		blocked := p.blocked
-		if !closed && !timedout && len(p.buf) > 0 {
-			n2 := copy(b, p.buf)
-			p.buf = p.buf[n2:]
-			b = b[n2:]
-			n += n2
+	for n == 0 {
+		n2, err := p.readOrBlock(b)
+		if err != nil {
+			return n2, err
 		}
-		p.mu.Unlock()
-
-		if closed {
-			return 0, fmt.Errorf("nettest.Pipe(%q): closed: %w", p.name, io.EOF)
-		}
-		if timedout {
-			return 0, fmt.Errorf("nettest.Pipe(%q): %w", p.name, ErrReadTimeout)
-		}
-		if blocked {
-			<-p.rCh
-			continue
-		}
-		if n > 0 {
-			p.signalWrite()
-			return n, nil
-		}
-		<-p.rCh
+		n += n2
 	}
+	p.cnd.Signal()
+	return n, nil
+}
+
+// writeOrBlock attempts to write to the buffer, if the buffer is full it will
+// block until there is a change.
+func (p *Pipe) writeOrBlock(b []byte) (int, error) {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	if p.closed {
+		return 0, net.ErrClosed
+	}
+	if !p.writeTimeout.IsZero() && !time.Now().Before(p.writeTimeout) {
+		return 0, os.ErrDeadlineExceeded
+	}
+	if p.blocked {
+		p.cnd.Wait()
+		return 0, nil
+	}
+
+	// Optimistically we want to write the entire slice.
+	n := len(b)
+	if limit := p.maxBuf - p.buf.Len(); limit < n {
+		// However, we don't have enough capacity to write everything.
+		n = limit
+	}
+	if n == 0 {
+		// Wait for something to change.
+		p.cnd.Wait()
+		return 0, nil
+	}
+
+	p.buf.Write(b[:n])
+	p.cnd.Signal()
+	return n, nil
 }
 
 // Write implements io.Writer.
@@ -96,47 +131,23 @@ func (p *Pipe) Write(b []byte) (n int, err error) {
 			log.Printf("Pipe(%q).Write(%q) n=%d, err=%v", p.name, string(orig), n, err)
 		}()
 	}
-	for {
-		p.mu.Lock()
-		closed := p.closed
-		timedout := !p.writeTimeout.IsZero() && !time.Now().Before(p.writeTimeout)
-		blocked := p.blocked
-		if !closed && !timedout {
-			n2 := len(b)
-			if limit := p.maxBuf - len(p.buf); limit < n2 {
-				n2 = limit
-			}
-			p.buf = append(p.buf, b[:n2]...)
-			b = b[n2:]
-			n += n2
+	for len(b) > 0 {
+		n2, err := p.writeOrBlock(b)
+		if err != nil {
+			return n + n2, err
 		}
-		p.mu.Unlock()
-
-		if closed {
-			return n, fmt.Errorf("nettest.Pipe(%q): closed: %w", p.name, io.EOF)
-		}
-		if timedout {
-			return n, fmt.Errorf("nettest.Pipe(%q): %w", p.name, ErrWriteTimeout)
-		}
-		if blocked {
-			<-p.wCh
-			continue
-		}
-		if n > 0 {
-			p.signalRead()
-		}
-		if len(b) == 0 {
-			return n, nil
-		}
-		<-p.wCh
+		n += n2
+		b = b[n2:]
 	}
+	return n, nil
 }
 
-// Close implements io.Closer.
+// Close closes the pipe.
 func (p *Pipe) Close() error {
 	p.mu.Lock()
-	closed := p.closed
+	defer p.mu.Unlock()
 	p.closed = true
+	p.blocked = false
 	if p.cancelWriteTimer != nil {
 		p.cancelWriteTimer()
 		p.cancelWriteTimer = nil
@@ -145,77 +156,65 @@ func (p *Pipe) Close() error {
 		p.cancelReadTimer()
 		p.cancelReadTimer = nil
 	}
-	p.mu.Unlock()
+	p.cnd.Broadcast()
 
-	if closed {
-		return fmt.Errorf("nettest.Pipe(%q).Close: already closed", p.name)
-	}
-
-	p.signalRead()
-	p.signalWrite()
 	return nil
 }
 
+func (p *Pipe) deadlineTimer(t time.Time) func() {
+	if t.IsZero() {
+		return nil
+	}
+	if t.Before(time.Now()) {
+		p.cnd.Broadcast()
+		return nil
+	}
+	ctx, cancel := context.WithDeadline(context.Background(), t)
+	go func() {
+		<-ctx.Done()
+		if ctx.Err() == context.DeadlineExceeded {
+			p.cnd.Broadcast()
+		}
+	}()
+	return cancel
+}
+
 // SetReadDeadline sets the deadline for future Read calls.
 func (p *Pipe) SetReadDeadline(t time.Time) error {
 	p.mu.Lock()
+	defer p.mu.Unlock()
 	p.readTimeout = t
+	// If we already have a deadline, cancel it and create a new one.
 	if p.cancelReadTimer != nil {
 		p.cancelReadTimer()
 		p.cancelReadTimer = nil
 	}
-	if d := time.Until(t); !t.IsZero() && d > 0 {
-		ctx, cancel := context.WithCancel(context.Background())
-		p.cancelReadTimer = cancel
-		go func() {
-			t := time.NewTimer(d)
-			defer t.Stop()
-			select {
-			case <-t.C:
-				p.signalRead()
-			case <-ctx.Done():
-			}
-		}()
-	}
-	p.mu.Unlock()
-
-	p.signalRead()
+	p.cancelReadTimer = p.deadlineTimer(t)
 	return nil
 }
 
 // SetWriteDeadline sets the deadline for future Write calls.
 func (p *Pipe) SetWriteDeadline(t time.Time) error {
 	p.mu.Lock()
+	defer p.mu.Unlock()
 	p.writeTimeout = t
+	// If we already have a deadline, cancel it and create a new one.
 	if p.cancelWriteTimer != nil {
 		p.cancelWriteTimer()
 		p.cancelWriteTimer = nil
 	}
-	if d := time.Until(t); !t.IsZero() && d > 0 {
-		ctx, cancel := context.WithCancel(context.Background())
-		p.cancelWriteTimer = cancel
-		go func() {
-			t := time.NewTimer(d)
-			defer t.Stop()
-			select {
-			case <-t.C:
-				p.signalWrite()
-			case <-ctx.Done():
-			}
-		}()
-	}
-	p.mu.Unlock()
-
-	p.signalWrite()
+	p.cancelWriteTimer = p.deadlineTimer(t)
 	return nil
 }
 
+// Block will cause all calls to Read and Write to block until they either
+// timeout, are unblocked or the pipe is closed.
 func (p *Pipe) Block() error {
 	p.mu.Lock()
+	defer p.mu.Unlock()
 	closed := p.closed
 	blocked := p.blocked
 	p.blocked = true
-	p.mu.Unlock()
 
 	if closed {
 		return fmt.Errorf("nettest.Pipe(%q).Block: closed", p.name)
@@ -223,17 +222,17 @@ func (p *Pipe) Block() error {
 	if blocked {
 		return fmt.Errorf("nettest.Pipe(%q).Block: already blocked", p.name)
 	}
-	p.signalRead()
-	p.signalWrite()
+	p.cnd.Broadcast()
 	return nil
 }
 
+// Unblock will cause all blocked Read/Write calls to continue execution.
 func (p *Pipe) Unblock() error {
 	p.mu.Lock()
+	defer p.mu.Unlock()
 	closed := p.closed
 	blocked := p.blocked
 	p.blocked = false
-	p.mu.Unlock()
 
 	if closed {
 		return fmt.Errorf("nettest.Pipe(%q).Block: closed", p.name)
@@ -241,21 +240,6 @@ func (p *Pipe) Unblock() error {
 	if !blocked {
 		return fmt.Errorf("nettest.Pipe(%q).Block: already unblocked", p.name)
 	}
-	p.signalRead()
-	p.signalWrite()
+	p.cnd.Broadcast()
 	return nil
 }
-
-func (p *Pipe) signalRead() {
-	select {
-	case p.rCh <- struct{}{}:
-	default:
-	}
-}
-
-func (p *Pipe) signalWrite() {
-	select {
-	case p.wCh <- struct{}{}:
-	default:
-	}
-}

net/nettest/pipe_test.go

@@ -7,6 +7,7 @@ package nettest
 import (
 	"errors"
 	"fmt"
+	"os"
 	"testing"
 	"time"
 )
@@ -35,7 +36,7 @@ func TestPipeTimeout(t *testing.T) {
 		p := NewPipe("p1", 1<<16)
 		p.SetWriteDeadline(time.Now().Add(-1 * time.Second))
 		n, err := p.Write([]byte{'h'})
-		if !errors.Is(err, ErrWriteTimeout) || !errors.Is(err, ErrTimeout) {
+		if !errors.Is(err, os.ErrDeadlineExceeded) {
 			t.Errorf("missing write timeout got err: %v", err)
 		}
 		if n != 0 {
@@ -49,7 +50,7 @@ func TestPipeTimeout(t *testing.T) {
 		p.SetReadDeadline(time.Now().Add(-1 * time.Second))
 		b := make([]byte, 1)
 		n, err := p.Read(b)
-		if !errors.Is(err, ErrReadTimeout) || !errors.Is(err, ErrTimeout) {
+		if !errors.Is(err, os.ErrDeadlineExceeded) {
 			t.Errorf("missing read timeout got err: %v", err)
 		}
 		if n != 0 {
@@ -65,7 +66,7 @@ func TestPipeTimeout(t *testing.T) {
 		if err := p.Block(); err != nil {
 			t.Fatal(err)
 		}
-		if _, err := p.Write([]byte{'h'}); !errors.Is(err, ErrWriteTimeout) {
+		if _, err := p.Write([]byte{'h'}); !errors.Is(err, os.ErrDeadlineExceeded) {
 			t.Fatalf("want write timeout got: %v", err)
 		}
 	})
@@ -80,7 +81,7 @@ func TestPipeTimeout(t *testing.T) {
 		if err := p.Block(); err != nil {
 			t.Fatal(err)
 		}
-		if _, err := p.Read(b); !errors.Is(err, ErrReadTimeout) {
+		if _, err := p.Read(b); !errors.Is(err, os.ErrDeadlineExceeded) {
 			t.Fatalf("want read timeout got: %v", err)
 		}
 	})

net/stun/stun_fuzzer.go

@@ -0,0 +1,13 @@
+// Copyright (c) 2021 Tailscale Inc & AUTHORS All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+// +build gofuzz
+
+package stun
+
+func FuzzStunParser(data []byte) int {
+	_, _, _, _ = ParseResponse(data)
+
+	_, _ = ParseBindingRequest(data)
+	return 1
+}

net/tsaddr/tsaddr.go

@@ -33,7 +33,9 @@ func CGNATRange() netaddr.IPPrefix {
 var (
 	cgnatRange   oncePrefix
 	ulaRange     oncePrefix
+	tsUlaRange   oncePrefix
 	ula4To6Range oncePrefix
+	ulaEph6Range oncePrefix
 )
 
 // TailscaleServiceIP returns the listen address of services
@@ -57,8 +59,8 @@ func IsTailscaleIP(ip netaddr.IP) bool {
 // TailscaleULARange returns the IPv6 Unique Local Address range that
 // is the superset range that Tailscale assigns out of.
 func TailscaleULARange() netaddr.IPPrefix {
-	ulaRange.Do(func() { mustPrefix(&ulaRange.v, "fd7a:115c:a1e0::/48") })
-	return ulaRange.v
+	tsUlaRange.Do(func() { mustPrefix(&tsUlaRange.v, "fd7a:115c:a1e0::/48") })
+	return tsUlaRange.v
 }
 
 // Tailscale4To6Range returns the subset of TailscaleULARange used for
@@ -71,6 +73,17 @@ func Tailscale4To6Range() netaddr.IPPrefix {
 	return ula4To6Range.v
 }
 
+// TailscaleEphemeral6Range returns the subset of TailscaleULARange
+// used for ephemeral IPv6-only Tailscale nodes.
+func TailscaleEphemeral6Range() netaddr.IPPrefix {
+	// This IP range has no significance, beyond being a subset of
+	// TailscaleULARange. The bits from /48 to /64 were picked at
+	// random, with the only criterion being to not be the conflict
+	// with the Tailscale4To6Range above.
+	ulaEph6Range.Do(func() { mustPrefix(&ulaEph6Range.v, "fd7a:115c:a1e0:efe3::/64") })
+	return ulaEph6Range.v
+}
+
 // Tailscale4To6Placeholder returns an IP address that can be used as
 // a source IP when one is required, but a netmap didn't provide
 // any. This address never gets allocated by the 4-to-6 algorithm in
@@ -95,6 +108,11 @@ func Tailscale4To6(ipv4 netaddr.IP) netaddr.IP {
 	return netaddr.IPFrom16(ret)
 }
 
+func IsULA(ip netaddr.IP) bool {
+	ulaRange.Do(func() { mustPrefix(&ulaRange.v, "fc00::/7") })
+	return ulaRange.v.Contains(ip)
+}
+
 func mustPrefix(v *netaddr.IPPrefix, prefix string) {
 	var err error
 	*v, err = netaddr.ParseIPPrefix(prefix)

net/tsaddr/tsaddr_test.go

@@ -42,3 +42,25 @@ func TestCGNATRange(t *testing.T) {
 		t.Errorf("got %q; want %q", got, want)
 	}
 }
+
+func TestIsUla(t *testing.T) {
+	tests := []struct {
+		name string
+		ip   string
+		want bool
+	}{
+		{"first ULA", "fc00::1", true},
+		{"not ULA", "fb00::1", false},
+		{"Tailscale", "fd7a:115c:a1e0::1", true},
+		{"Cloud Run", "fddf:3978:feb1:d745::1", true},
+		{"zeros", "0000:0000:0000:0000:0000:0000:0000:0000", false},
+		{"Link Local", "fe80::1", false},
+		{"Global", "2602::1", false},
+	}
+
+	for _, test := range tests {
+		if got := IsULA(netaddr.MustParseIP(test.ip)); got != test.want {
+			t.Errorf("IsULA(%s) = %v, want %v", test.name, got, test.want)
+		}
+	}
+}