net/dns: crimes against unix and windows

This is preliminary work for using the directManager as
part of a wslManager on windows, where in addition to configuring
windows we'll use wsl.exe to edit the linux file system and modify the
system resolv.conf.

The pinholeFS is a little funky, but it's designed to work through
simple unix tools via wsl.exe without invoking bash. I would not have
thought it would stand on its own like this, but it turns out it's
useful for writing a test for the directManager.

Signed-off-by: David Crawshaw <crawshaw@tailscale.com>

.github/workflows/xe-experimental-vm-test.yml

@@ -0,0 +1,36 @@
+name: "integration-vms"
+
+on:
+  # # NOTE(Xe): uncomment this region when testing the test
+  # pull_request:
+  #   branches:
+  #     - 'main'
+  release:
+    types: [ created ]
+  schedule:
+    # At minute 0 past hour 6 and 18
+    # https://crontab.guru/#00_6,18_*_*_*
+    - cron: '00 6,18 * * *'
+
+jobs:
+  experimental-linux-vm-test:
+    # To set up a new runner, see tstest/integration/vms/runner.nix
+    runs-on: [ self-hosted, linux, vm_integration_test ]
+
+    if: "!contains(github.event.head_commit.message, '[ci skip]')"
+
+    steps:
+      - name: Checkout Code
+        uses: actions/checkout@v1
+
+      - name: Download VM Images
+        run: go test ./tstest/integration/vms -run-vm-tests -run=Download -timeout=60m
+        env:
+          XDG_CACHE_HOME: "/var/lib/ghrunner/cache"
+
+      - name: Run VM tests
+        run: go test ./tstest/integration/vms -v -run-vm-tests
+        env:
+          TMPDIR: "/tmp"
+          XDG_CACHE_HOME: "/var/lib/ghrunner/cache"
+

VERSION.txt

@@ -1 +1 @@
-1.9.0
+1.11.0

build_dist.sh

@@ -11,6 +11,36 @@
 
 set -eu
 
-eval $(./version/version.sh)
+IFS=".$IFS" read -r major minor patch <VERSION.txt
+git_hash=$(git rev-parse HEAD)
+if ! git diff-index --quiet HEAD; then
+	git_hash="${git_hash}-dirty"
+fi
+base_hash=$(git rev-list --max-count=1 HEAD -- VERSION.txt)
+change_count=$(git rev-list --count HEAD "^$base_hash")
+short_hash=$(echo "$git_hash" | cut -c1-9)
 
-exec go build -tags xversion -ldflags "-X tailscale.com/version.Long=${VERSION_LONG} -X tailscale.com/version.Short=${VERSION_SHORT} -X tailscale.com/version.GitCommit=${VERSION_GIT_HASH}" "$@"
+if expr "$minor" : "[0-9]*[13579]$" >/dev/null; then
+	patch="$change_count"
+	change_suffix=""
+elif [ "$change_count" != "0" ]; then
+	change_suffix="-$change_count"
+else
+	change_suffix=""
+fi
+
+long_suffix="$change_suffix-t$short_hash"
+SHORT="$major.$minor.$patch"
+LONG="${SHORT}$long_suffix"
+GIT_HASH="$git_hash"
+
+if [ "$1" = "shellvars" ]; then
+	cat <<EOF
+VERSION_SHORT="$SHORT"
+VERSION_LONG="$LONG"
+VERSION_GIT_HASH="$GIT_HASH"
+EOF
+	exit 0
+fi
+
+exec go build -ldflags "-X tailscale.com/version.Long=${LONG} -X tailscale.com/version.Short=${SHORT} -X tailscale.com/version.GitCommit=${GIT_HASH}" "$@"

client/tailscale/tailscale.go

@@ -256,3 +256,25 @@ func Logout(ctx context.Context) error {
 	_, err := send(ctx, "POST", "/localapi/v0/logout", http.StatusNoContent, nil)
 	return err
 }
+
+// SetDNS adds a DNS TXT record for the given domain name, containing
+// the provided TXT value. The intended use case is answering
+// LetsEncrypt/ACME dns-01 challenges.
+//
+// The control plane will only permit SetDNS requests with very
+// specific names and values. The name should be
+// "_acme-challenge." + your node's MagicDNS name. It's expected that
+// clients cache the certs from LetsEncrypt (or whichever CA is
+// providing them) and only request new ones as needed; the control plane
+// rate limits SetDNS requests.
+//
+// This is a low-level interface; it's expected that most Tailscale
+// users use a higher level interface to getting/using TLS
+// certificates.
+func SetDNS(ctx context.Context, name, value string) error {
+	v := url.Values{}
+	v.Set("name", name)
+	v.Set("value", value)
+	_, err := send(ctx, "POST", "/localapi/v0/set-dns?"+v.Encode(), 200, nil)
+	return err
+}

cmd/cloner/cloner.go

@@ -246,7 +246,9 @@ func gen(buf *bytes.Buffer, imports map[string]struct{}, name string, typ *types
 					writef("\t\tdst.%s[k] = append([]%s{}, src.%s[k]...)", fname, n, fname)
 					writef("\t}")
 				} else if containsPointers(ft.Elem()) {
-					writef("\t\t" + `panic("TODO map value pointers")`)
+					writef("\tfor k, v := range src.%s {", fname)
+					writef("\t\tdst.%s[k] = v.Clone()", fname)
+					writef("\t}")
 				} else {
 					writef("\tfor k, v := range src.%s {", fname)
 					writef("\t\tdst.%s[k] = v", fname)

cmd/derper/derper.go

@@ -12,8 +12,6 @@ import (
 	"errors"
 	"expvar"
 	"flag"
-	"fmt"
-	"html"
 	"io"
 	"io/ioutil"
 	"log"
@@ -35,7 +33,6 @@ import (
 	"tailscale.com/tsweb"
 	"tailscale.com/types/key"
 	"tailscale.com/types/wgkey"
-	"tailscale.com/version"
 )
 
 var (
@@ -49,6 +46,7 @@ var (
 	meshPSKFile   = flag.String("mesh-psk-file", defaultMeshPSKFile(), "if non-empty, path to file containing the mesh pre-shared key file. It should contain some hex string; whitespace is trimmed.")
 	meshWith      = flag.String("mesh-with", "", "optional comma-separated list of hostnames to mesh with; the server's own hostname can be in the list")
 	bootstrapDNS  = flag.String("bootstrap-dns-names", "", "optional comma-separated list of hostnames to make available at /bootstrap-dns")
+	verifyClients = flag.Bool("verify-clients", false, "verify clients to this DERP server through a local tailscaled instance.")
 )
 
 type config struct {
@@ -125,6 +123,7 @@ func main() {
 	letsEncrypt := tsweb.IsProd443(*addr)
 
 	s := derp.NewServer(key.Private(cfg.PrivateKey), log.Printf)
+	s.SetVerifyClient(*verifyClients)
 
 	if *meshPSKFile != "" {
 		b, err := ioutil.ReadFile(*meshPSKFile)
@@ -143,8 +142,7 @@ func main() {
 	}
 	expvar.Publish("derp", s.ExpVar())
 
-	// Create our own mux so we don't expose /debug/ stuff to the world.
-	mux := tsweb.NewMux(debugHandler(s))
+	mux := http.NewServeMux()
 	mux.Handle("/derp", derphttp.Handler(s))
 	go refreshBootstrapDNSLoop()
 	mux.HandleFunc("/bootstrap-dns", handleBootstrapDNS)
@@ -164,6 +162,18 @@ func main() {
 			io.WriteString(w, "<p>Debug info at <a href='/debug/'>/debug/</a>.</p>\n")
 		}
 	}))
+	debug := tsweb.Debugger(mux)
+	debug.KV("TLS hostname", *hostname)
+	debug.KV("Mesh key", s.HasMeshKey())
+	debug.Handle("check", "Consistency check", http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		err := s.ConsistencyCheck()
+		if err != nil {
+			http.Error(w, err.Error(), 500)
+		} else {
+			io.WriteString(w, "derp.Server ConsistencyCheck okay")
+		}
+	}))
+	debug.Handle("traffic", "Traffic check", http.HandlerFunc(s.ServeDebugTraffic))
 
 	if *runSTUN {
 		go serveSTUN()
@@ -217,39 +227,6 @@ func main() {
 	}
 }
 
-func debugHandler(s *derp.Server) http.Handler {
-	return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-		if r.RequestURI == "/debug/check" {
-			err := s.ConsistencyCheck()
-			if err != nil {
-				http.Error(w, err.Error(), 500)
-			} else {
-				io.WriteString(w, "derp.Server ConsistencyCheck okay")
-			}
-			return
-		}
-		f := func(format string, args ...interface{}) { fmt.Fprintf(w, format, args...) }
-		f(`<html><body>
-<h1>DERP debug</h1>
-<ul>
-`)
-		f("<li><b>Hostname:</b> %v</li>\n", html.EscapeString(*hostname))
-		f("<li><b>Uptime:</b> %v</li>\n", tsweb.Uptime())
-		f("<li><b>Mesh Key:</b> %v</li>\n", s.HasMeshKey())
-		f("<li><b>Version:</b> %v</li>\n", html.EscapeString(version.Long))
-
-		f(`<li><a href="/debug/vars">/debug/vars</a> (Go)</li>
-   <li><a href="/debug/varz">/debug/varz</a> (Prometheus)</li>
-   <li><a href="/debug/pprof/">/debug/pprof/</a></li>
-   <li><a href="/debug/pprof/goroutine?debug=1">/debug/pprof/goroutine</a> (collapsed)</li>
-   <li><a href="/debug/pprof/goroutine?debug=2">/debug/pprof/goroutine</a> (full)</li>
-   <li><a href="/debug/check">/debug/check</a> internal consistency check</li>
-<ul>
-</html>
-`)
-	})
-}
-
 func serveSTUN() {
 	pc, err := net.ListenPacket("udp", ":3478")
 	if err != nil {

cmd/microproxy/microproxy.go

@@ -55,9 +55,13 @@ func main() {
 		log.Fatalf("Couldn't parse URL %q: %v", *goVarsURL, err)
 	}
 
-	mux := tsweb.NewMux(http.HandlerFunc(debugHandler))
+	mux := http.NewServeMux()
+	tsweb.Debugger(mux) // registers /debug/*
 	mux.Handle("/metrics", tsweb.Protected(proxy))
-	mux.Handle("/varz", tsweb.Protected(tsweb.StdHandler(&goVarsHandler{*goVarsURL}, log.Printf)))
+	mux.Handle("/varz", tsweb.Protected(tsweb.StdHandler(&goVarsHandler{*goVarsURL}, tsweb.HandlerOptions{
+		Quiet200s: true,
+		Logf:      log.Printf,
+	})))
 
 	ch := &certHolder{
 		hostname: *hostname,
@@ -167,23 +171,3 @@ func (c *certHolder) loadLocked() error {
 	c.loaded = time.Now()
 	return nil
 }
-
-// debugHandler serves a page with links to tsweb-managed debug URLs
-// at /debug/.
-func debugHandler(w http.ResponseWriter, r *http.Request) {
-	f := func(format string, args ...interface{}) { fmt.Fprintf(w, format, args...) }
-	f(`<html><body>
-<h1>microproxy debug</h1>
-<ul>
-`)
-	f("<li><b>Hostname:</b> %v</li>\n", *hostname)
-	f("<li><b>Uptime:</b> %v</li>\n", tsweb.Uptime())
-	f(`<li><a href="/debug/vars">/debug/vars</a> (Go)</li>
-   <li><a href="/debug/varz">/debug/varz</a> (Prometheus)</li>
-   <li><a href="/debug/pprof/">/debug/pprof/</a></li>
-   <li><a href="/debug/pprof/goroutine?debug=1">/debug/pprof/goroutine</a> (collapsed)</li>
-   <li><a href="/debug/pprof/goroutine?debug=2">/debug/pprof/goroutine</a> (full)</li>
-<ul>
-</html>
-`)
-}

cmd/mkpkg/main.go

@@ -21,6 +21,9 @@ import (
 // into a map of filePathOnDisk -> filePathInPackage.
 func parseFiles(s string) (map[string]string, error) {
 	ret := map[string]string{}
+	if len(s) == 0 {
+		return ret, nil
+	}
 	for _, f := range strings.Split(s, ",") {
 		fs := strings.Split(f, ":")
 		if len(fs) != 2 {

cmd/tailscale/cli/auth-redirect.html

@@ -0,0 +1,57 @@
+<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>

cmd/tailscale/cli/file.go

@@ -74,7 +74,6 @@ func runCp(ctx context.Context, args []string) error {
 		return runCpTargets(ctx, args)
 	}
 	if len(args) < 2 {
-		//lint:ignore ST1005 no sorry need that colon at the end
 		return errors.New("usage: tailscale file cp <files...> <target>:")
 	}
 	files, target := args[:len(args)-1], args[len(args)-1]
@@ -97,14 +96,12 @@ func runCp(ctx context.Context, args []string) error {
 		return err
 	}
 
-	peerAPIBase, lastSeen, isOffline, err := discoverPeerAPIBase(ctx, ip)
+	peerAPIBase, isOffline, err := discoverPeerAPIBase(ctx, ip)
 	if err != nil {
 		return fmt.Errorf("can't send to %s: %v", target, err)
 	}
 	if isOffline {
 		fmt.Fprintf(os.Stderr, "# warning: %s is offline\n", target)
-	} else if !lastSeen.IsZero() && time.Since(lastSeen) > lastSeenOld {
-		fmt.Fprintf(os.Stderr, "# warning: %s last seen %v ago\n", target, time.Since(lastSeen).Round(time.Minute))
 	}
 
 	if len(files) > 1 {
@@ -182,14 +179,14 @@ func runCp(ctx context.Context, args []string) error {
 	return nil
 }
 
-func discoverPeerAPIBase(ctx context.Context, ipStr string) (base string, lastSeen time.Time, isOffline bool, err error) {
+func discoverPeerAPIBase(ctx context.Context, ipStr string) (base string, isOffline bool, err error) {
 	ip, err := netaddr.ParseIP(ipStr)
 	if err != nil {
-		return "", time.Time{}, false, err
+		return "", false, err
 	}
 	fts, err := tailscale.FileTargets(ctx)
 	if err != nil {
-		return "", time.Time{}, false, err
+		return "", false, err
 	}
 	for _, ft := range fts {
 		n := ft.Node
@@ -197,14 +194,11 @@ func discoverPeerAPIBase(ctx context.Context, ipStr string) (base string, lastSe
 			if a.IP() != ip {
 				continue
 			}
-			if n.LastSeen != nil {
-				lastSeen = *n.LastSeen
-			}
 			isOffline = n.Online != nil && !*n.Online
-			return ft.PeerAPIURL, lastSeen, isOffline, nil
+			return ft.PeerAPIURL, isOffline, nil
 		}
 	}
-	return "", time.Time{}, false, fileTargetErrorDetail(ctx, ip)
+	return "", false, fileTargetErrorDetail(ctx, ip)
 }
 
 // fileTargetErrorDetail returns a non-nil error saying why ip is an
@@ -274,8 +268,6 @@ func (r *slowReader) Read(p []byte) (n int, err error) {
 	return
 }
 
-const lastSeenOld = 20 * time.Minute
-
 func runCpTargets(ctx context.Context, args []string) error {
 	if len(args) > 0 {
 		return errors.New("invalid arguments with --targets")

cmd/tailscale/cli/up.go

@@ -230,7 +230,9 @@ func prefsFromUpArgs(upArgs upArgsT, warnf logger.Logf, st *ipnstate.Status, goo
 			warnf("netfilter=nodivert; add iptables calls to ts-* chains manually.")
 		case "off":
 			prefs.NetfilterMode = preftype.NetfilterOff
-			warnf("netfilter=off; configure iptables yourself.")
+			if defaultNetfilterMode() != "off" {
+				warnf("netfilter=off; configure iptables yourself.")
+			}
 		default:
 			return nil, fmt.Errorf("invalid value --netfilter-mode=%q", upArgs.netfilterMode)
 		}
@@ -266,7 +268,7 @@ func runUp(ctx context.Context, args []string) error {
 	}
 
 	if distro.Get() == distro.Synology {
-		notSupported := "not yet supported on Synology; see https://github.com/tailscale/tailscale/issues/451"
+		notSupported := "not supported on Synology; see https://github.com/tailscale/tailscale/issues/1995"
 		if upArgs.acceptRoutes {
 			return errors.New("--accept-routes is " + notSupported)
 		}

cmd/tailscale/cli/web.go

@@ -9,12 +9,15 @@ import (
 	"context"
 	_ "embed"
 	"encoding/json"
+	"encoding/xml"
 	"flag"
 	"fmt"
 	"html/template"
+	"io/ioutil"
 	"log"
 	"net/http"
 	"net/http/cgi"
+	"net/url"
 	"os/exec"
 	"runtime"
 	"strings"
@@ -24,6 +27,7 @@ import (
 	"tailscale.com/ipn"
 	"tailscale.com/tailcfg"
 	"tailscale.com/types/preftype"
+	"tailscale.com/util/groupmember"
 	"tailscale.com/version/distro"
 )
 
@@ -33,6 +37,9 @@ var webHTML string
 //go:embed web.css
 var webCSS string
 
+//go:embed auth-redirect.html
+var authenticationRedirectHTML string
+
 var tmpl *template.Template
 
 func init() {
@@ -82,23 +89,114 @@ func runWeb(ctx context.Context, args []string) error {
 	return http.ListenAndServe(webArgs.listen, http.HandlerFunc(webHandler))
 }
 
-func auth() (string, error) {
-	if distro.Get() == distro.Synology {
-		cmd := exec.Command("/usr/syno/synoman/webman/modules/authenticate.cgi")
-		out, err := cmd.CombinedOutput()
+// authorize returns the name of the user accessing the web UI after verifying
+// whether the user has access to the web UI. The function will write the
+// error to the provided http.ResponseWriter.
+// Note: This is different from a tailscale user, and is typically the local
+// user on the node.
+func authorize(w http.ResponseWriter, r *http.Request) (string, error) {
+	switch distro.Get() {
+	case distro.Synology:
+		user, err := synoAuthn()
 		if err != nil {
-			return "", fmt.Errorf("auth: %v: %s", err, out)
+			http.Error(w, err.Error(), http.StatusUnauthorized)
+			return "", err
 		}
-		return string(out), nil
+		if err := authorizeSynology(user); err != nil {
+			http.Error(w, err.Error(), http.StatusForbidden)
+			return "", err
+		}
+		return user, nil
+	case distro.QNAP:
+		user, resp, err := qnapAuthn(r)
+		if err != nil {
+			http.Error(w, err.Error(), http.StatusUnauthorized)
+			return "", err
+		}
+		if resp.IsAdmin == 0 {
+			http.Error(w, err.Error(), http.StatusForbidden)
+			return "", err
+		}
+		return user, nil
 	}
-
 	return "", nil
 }
 
-func synoTokenRedirect(w http.ResponseWriter, r *http.Request) bool {
-	if distro.Get() != distro.Synology {
-		return false
+// authorizeSynology checks whether the provided user has access to the web UI
+// by consulting the membership of the "administrators" group.
+func authorizeSynology(name string) error {
+	yes, err := groupmember.IsMemberOfGroup("administrators", name)
+	if err != nil {
+		return err
 	}
+	if !yes {
+		return fmt.Errorf("not a member of administrators group")
+	}
+	return nil
+}
+
+type qnapAuthResponse struct {
+	AuthPassed int    `xml:"authPassed"`
+	IsAdmin    int    `xml:"isAdmin"`
+	AuthSID    string `xml:"authSid"`
+	ErrorValue int    `xml:"errorValue"`
+}
+
+func qnapAuthn(r *http.Request) (string, *qnapAuthResponse, error) {
+	user, err := r.Cookie("NAS_USER")
+	if err != nil {
+		return "", nil, err
+	}
+	token, err := r.Cookie("qtoken")
+	if err != nil {
+		return "", nil, err
+	}
+	query := url.Values{
+		"qtoken": []string{token.Value},
+		"user":   []string{user.Value},
+	}
+	u := url.URL{
+		Scheme:   r.URL.Scheme,
+		Host:     r.URL.Host,
+		Path:     "/cgi-bin/authLogin.cgi",
+		RawQuery: query.Encode(),
+	}
+	resp, err := http.Get(u.String())
+	if err != nil {
+		return "", nil, err
+	}
+	defer resp.Body.Close()
+	out, err := ioutil.ReadAll(resp.Body)
+	if err != nil {
+		return "", nil, err
+	}
+	authResp := &qnapAuthResponse{}
+	if err := xml.Unmarshal(out, authResp); err != nil {
+		return "", nil, err
+	}
+	if authResp.AuthPassed == 0 {
+		return "", nil, fmt.Errorf("not authenticated")
+	}
+	return user.Value, authResp, nil
+}
+
+func synoAuthn() (string, error) {
+	cmd := exec.Command("/usr/syno/synoman/webman/modules/authenticate.cgi")
+	out, err := cmd.CombinedOutput()
+	if err != nil {
+		return "", fmt.Errorf("auth: %v: %s", err, out)
+	}
+	return strings.TrimSpace(string(out)), nil
+}
+
+func authRedirect(w http.ResponseWriter, r *http.Request) bool {
+	if distro.Get() == distro.Synology {
+		return synoTokenRedirect(w, r)
+	}
+	return false
+}
+
+func synoTokenRedirect(w http.ResponseWriter, r *http.Request) bool {
 	if r.Header.Get("X-Syno-Token") != "" {
 		return false
 	}
@@ -132,75 +230,13 @@ 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) {
+	if authRedirect(w, r) {
 		return
 	}
 
-	user, err := auth()
+	user, err := authorize(w, r)
 	if err != nil {
-		http.Error(w, err.Error(), http.StatusForbidden)
 		return
 	}
 
@@ -214,7 +250,8 @@ func webHandler(w http.ResponseWriter, r *http.Request) {
 		w.Header().Set("Content-Type", "application/json")
 		url, err := tailscaleUpForceReauth(r.Context())
 		if err != nil {
-			json.NewEncoder(w).Encode(mi{"error": err})
+			w.WriteHeader(http.StatusInternalServerError)
+			json.NewEncoder(w).Encode(mi{"error": err.Error()})
 			return
 		}
 		json.NewEncoder(w).Encode(mi{"url": url})
@@ -223,7 +260,7 @@ func webHandler(w http.ResponseWriter, r *http.Request) {
 
 	st, err := tailscale.Status(r.Context())
 	if err != nil {
-		http.Error(w, err.Error(), 500)
+		http.Error(w, err.Error(), http.StatusInternalServerError)
 		return
 	}
 
@@ -241,7 +278,7 @@ func webHandler(w http.ResponseWriter, r *http.Request) {
 
 	buf := new(bytes.Buffer)
 	if err := tmpl.Execute(buf, data); err != nil {
-		http.Error(w, err.Error(), 500)
+		http.Error(w, err.Error(), http.StatusInternalServerError)
 		return
 	}
 	w.Write(buf.Bytes())
@@ -320,6 +357,10 @@ func tailscaleUpForceReauth(ctx context.Context) (authURL string, retErr error)
 	})
 	bc.StartLoginInteractive()
 
+	<-pumpCtx.Done() // wait for authURL or complete failure
+	if authURL == "" && retErr == nil {
+		retErr = pumpCtx.Err()
+	}
 	if authURL == "" && retErr == nil {
 		return "", fmt.Errorf("login failed with no backend error message")
 	}

cmd/tailscale/depaware.txt

@@ -1,6 +1,7 @@
 tailscale.com/cmd/tailscale dependencies: (generated by github.com/tailscale/depaware)
 
-   W 💣 github.com/alexbrainman/sspi                                 from github.com/alexbrainman/sspi/negotiate
+   W 💣 github.com/alexbrainman/sspi                                 from github.com/alexbrainman/sspi/negotiate+
+   W    github.com/alexbrainman/sspi/internal/common                 from github.com/alexbrainman/sspi/negotiate
    W 💣 github.com/alexbrainman/sspi/negotiate                       from tailscale.com/net/tshttpproxy
         github.com/kballard/go-shellquote                            from tailscale.com/cmd/tailscale/cli
         github.com/peterbourgon/ff/v2                                from github.com/peterbourgon/ff/v2/ffcli
@@ -14,13 +15,14 @@ tailscale.com/cmd/tailscale dependencies: (generated by github.com/tailscale/dep
         inet.af/netaddr                                              from tailscale.com/cmd/tailscale/cli+
         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                               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/hostinfo                                       from tailscale.com/net/interfaces
         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
@@ -48,14 +50,14 @@ tailscale.com/cmd/tailscale dependencies: (generated by github.com/tailscale/dep
         tailscale.com/types/opt                                      from tailscale.com/net/netcheck+
         tailscale.com/types/persist                                  from tailscale.com/ipn
         tailscale.com/types/preftype                                 from tailscale.com/cmd/tailscale/cli+
-        tailscale.com/types/strbuilder                               from tailscale.com/net/packet
         tailscale.com/types/structs                                  from tailscale.com/ipn+
         tailscale.com/types/wgkey                                    from tailscale.com/types/netmap+
         tailscale.com/util/dnsname                                   from tailscale.com/cmd/tailscale/cli+
    W    tailscale.com/util/endian                                    from tailscale.com/net/netns
-   L    tailscale.com/util/lineread                                  from tailscale.com/net/interfaces
+        tailscale.com/util/groupmember                               from tailscale.com/cmd/tailscale/cli
+        tailscale.com/util/lineread                                  from tailscale.com/net/interfaces+
         tailscale.com/version                                        from tailscale.com/cmd/tailscale/cli+
-        tailscale.com/version/distro                                 from tailscale.com/cmd/tailscale/cli
+        tailscale.com/version/distro                                 from tailscale.com/cmd/tailscale/cli+
         tailscale.com/wgengine/filter                                from tailscale.com/types/netmap
         golang.org/x/crypto/blake2b                                  from golang.org/x/crypto/nacl/box
         golang.org/x/crypto/chacha20                                 from golang.org/x/crypto/chacha20poly1305
@@ -118,13 +120,14 @@ tailscale.com/cmd/tailscale dependencies: (generated by github.com/tailscale/dep
         debug/macho                                                  from rsc.io/goversion/version
         debug/pe                                                     from rsc.io/goversion/version
         embed                                                        from tailscale.com/cmd/tailscale/cli
-        encoding                                                     from encoding/json
+        encoding                                                     from encoding/json+
         encoding/asn1                                                from crypto/x509+
         encoding/base64                                              from encoding/json+
         encoding/binary                                              from compress/gzip+
         encoding/hex                                                 from crypto/x509+
         encoding/json                                                from expvar+
         encoding/pem                                                 from crypto/tls+
+        encoding/xml                                                 from tailscale.com/cmd/tailscale/cli
         errors                                                       from bufio+
         expvar                                                       from tailscale.com/derp+
         flag                                                         from github.com/peterbourgon/ff/v2+
@@ -156,6 +159,7 @@ tailscale.com/cmd/tailscale dependencies: (generated by github.com/tailscale/dep
         os                                                           from crypto/rand+
         os/exec                                                      from github.com/toqueteos/webbrowser+
         os/signal                                                    from tailscale.com/cmd/tailscale/cli
+        os/user                                                      from tailscale.com/util/groupmember
         path                                                         from debug/dwarf+
         path/filepath                                                from crypto/x509+
         reflect                                                      from crypto/x509+

cmd/tailscale/tailscale.go

@@ -4,7 +4,7 @@
 
 // The tailscale command is the Tailscale command-line client. It interacts
 // with the tailscaled node agent.
-package main // import "tailscale.com/cmd/tailscaled"
+package main // import "tailscale.com/cmd/tailscale"
 
 import (
 	"fmt"
@@ -12,10 +12,10 @@ import (
 	"path/filepath"
 	"strings"
 
-	"tailscale.com/cmd/tailscaled/cli"
+	"tailscale.com/cmd/tailscale/cli"
 )
 
-func tailscale_main() {
+func main() {
 	args := os.Args[1:]
 	if name, _ := os.Executable(); strings.HasSuffix(filepath.Base(name), ".cgi") {
 		args = []string{"web", "-cgi"}

cmd/tailscaled/depaware.txt

@@ -1,43 +1,47 @@
 tailscale.com/cmd/tailscaled dependencies: (generated by github.com/tailscale/depaware)
 
-   W 💣 github.com/alexbrainman/sspi                                 from github.com/alexbrainman/sspi/negotiate
+   W 💣 github.com/alexbrainman/sspi                                 from github.com/alexbrainman/sspi/negotiate+
+   W    github.com/alexbrainman/sspi/internal/common                 from github.com/alexbrainman/sspi/negotiate
    W 💣 github.com/alexbrainman/sspi/negotiate                       from tailscale.com/net/tshttpproxy
    L    github.com/coreos/go-iptables/iptables                       from tailscale.com/wgengine/router
-   W 💣 github.com/github/certstore                                  from tailscale.com/control/controlclient
         github.com/go-multierror/multierror                          from tailscale.com/wgengine/router+
    W 💣 github.com/go-ole/go-ole                                     from github.com/go-ole/go-ole/oleutil+
    W 💣 github.com/go-ole/go-ole/oleutil                             from tailscale.com/wgengine/winnet
    L 💣 github.com/godbus/dbus/v5                                    from tailscale.com/net/dns
+        github.com/golang/snappy                                     from github.com/klauspost/compress/zstd
         github.com/google/btree                                      from inet.af/netstack/tcpip/header+
    L    github.com/josharian/native                                  from github.com/mdlayher/netlink+
    L 💣 github.com/jsimonetti/rtnetlink                              from tailscale.com/wgengine/monitor
    L    github.com/jsimonetti/rtnetlink/internal/unix                from github.com/jsimonetti/rtnetlink
         github.com/klauspost/compress/fse                            from github.com/klauspost/compress/huff0
         github.com/klauspost/compress/huff0                          from github.com/klauspost/compress/zstd
-        github.com/klauspost/compress/snappy                         from github.com/klauspost/compress/zstd
         github.com/klauspost/compress/zstd                           from tailscale.com/smallzstd
         github.com/klauspost/compress/zstd/internal/xxhash           from github.com/klauspost/compress/zstd
-   L 💣 github.com/mdlayher/netlink                                  from github.com/jsimonetti/rtnetlink+
-   L 💣 github.com/mdlayher/netlink/nlenc                            from github.com/jsimonetti/rtnetlink+
+   L 💣 github.com/mdlayher/netlink                                  from tailscale.com/wgengine/monitor+
+   L 💣 github.com/mdlayher/netlink/nlenc                            from github.com/mdlayher/netlink+
    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
-        github.com/tailscale/wireguard-go/ratelimiter                from github.com/tailscale/wireguard-go/device
-        github.com/tailscale/wireguard-go/replay                     from github.com/tailscale/wireguard-go/device
-        github.com/tailscale/wireguard-go/rwcancel                   from github.com/tailscale/wireguard-go/device+
-        github.com/tailscale/wireguard-go/tai64n                     from github.com/tailscale/wireguard-go/device+
-     💣 github.com/tailscale/wireguard-go/tun                        from github.com/tailscale/wireguard-go/device+
-   W 💣 github.com/tailscale/wireguard-go/tun/wintun                 from github.com/tailscale/wireguard-go/tun+
+   L 💣 github.com/mdlayher/socket                                   from github.com/mdlayher/netlink
+   W    github.com/pkg/errors                                        from github.com/tailscale/certstore
+   W 💣 github.com/tailscale/certstore                               from tailscale.com/control/controlclient
         github.com/tcnksm/go-httpstat                                from tailscale.com/net/netcheck
      💣 go4.org/intern                                               from inet.af/netaddr
-     💣 go4.org/mem                                                  from tailscale.com/control/controlclient+
+     💣 go4.org/mem                                                  from tailscale.com/derp+
         go4.org/unsafe/assume-no-moving-gc                           from go4.org/intern
+     💣 golang.zx2c4.com/wireguard/conn                              from golang.zx2c4.com/wireguard/device+
+   W 💣 golang.zx2c4.com/wireguard/conn/winrio                       from golang.zx2c4.com/wireguard/conn
+     💣 golang.zx2c4.com/wireguard/device                            from tailscale.com/net/tstun+
+     💣 golang.zx2c4.com/wireguard/ipc                               from golang.zx2c4.com/wireguard/device
+   W 💣 golang.zx2c4.com/wireguard/ipc/winpipe                       from golang.zx2c4.com/wireguard/ipc
+        golang.zx2c4.com/wireguard/ratelimiter                       from golang.zx2c4.com/wireguard/device
+        golang.zx2c4.com/wireguard/replay                            from golang.zx2c4.com/wireguard/device
+        golang.zx2c4.com/wireguard/rwcancel                          from golang.zx2c4.com/wireguard/device+
+        golang.zx2c4.com/wireguard/tai64n                            from golang.zx2c4.com/wireguard/device+
+     💣 golang.zx2c4.com/wireguard/tun                               from golang.zx2c4.com/wireguard/device+
+   W 💣 golang.zx2c4.com/wireguard/tun/wintun                        from golang.zx2c4.com/wireguard/tun+
    W 💣 golang.zx2c4.com/wireguard/windows/tunnel/winipcfg           from tailscale.com/net/interfaces+
         inet.af/netaddr                                              from tailscale.com/control/controlclient+
+        inet.af/netstack/atomicbitops                                from inet.af/netstack/tcpip+
+     💣 inet.af/netstack/buffer                                      from inet.af/netstack/tcpip/stack
      💣 inet.af/netstack/gohacks                                     from inet.af/netstack/state/wire+
         inet.af/netstack/linewriter                                  from inet.af/netstack/log
         inet.af/netstack/log                                         from inet.af/netstack/state+
@@ -46,7 +50,7 @@ tailscale.com/cmd/tailscaled dependencies: (generated by github.com/tailscale/de
      💣 inet.af/netstack/state                                       from inet.af/netstack/tcpip+
         inet.af/netstack/state/wire                                  from inet.af/netstack/state
      💣 inet.af/netstack/sync                                        from inet.af/netstack/linewriter+
-     💣 inet.af/netstack/tcpip                                       from inet.af/netstack/tcpip/adapters/gonet+
+        inet.af/netstack/tcpip                                       from inet.af/netstack/tcpip/adapters/gonet+
         inet.af/netstack/tcpip/adapters/gonet                        from tailscale.com/wgengine/netstack
      💣 inet.af/netstack/tcpip/buffer                                from inet.af/netstack/tcpip/adapters/gonet+
         inet.af/netstack/tcpip/hash/jenkins                          from inet.af/netstack/tcpip/stack+
@@ -69,8 +73,10 @@ tailscale.com/cmd/tailscaled dependencies: (generated by github.com/tailscale/de
         inet.af/netstack/tcpip/transport/udp                         from inet.af/netstack/tcpip/adapters/gonet+
         inet.af/netstack/waiter                                      from inet.af/netstack/tcpip+
         inet.af/peercred                                             from tailscale.com/ipn/ipnserver
+   W 💣 inet.af/wf                                                   from tailscale.com/wf
         rsc.io/goversion/version                                     from tailscale.com/version
         tailscale.com/atomicfile                                     from tailscale.com/ipn+
+        tailscale.com/client/tailscale                               from tailscale.com/derp
         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+
@@ -78,6 +84,7 @@ tailscale.com/cmd/tailscaled dependencies: (generated by github.com/tailscale/de
         tailscale.com/derp/derpmap                                   from tailscale.com/cmd/tailscaled+
         tailscale.com/disco                                          from tailscale.com/derp+
         tailscale.com/health                                         from tailscale.com/control/controlclient+
+        tailscale.com/hostinfo                                       from tailscale.com/control/controlclient+
         tailscale.com/internal/deephash                              from tailscale.com/ipn/ipnlocal+
         tailscale.com/ipn                                            from tailscale.com/ipn/ipnserver+
         tailscale.com/ipn/ipnlocal                                   from tailscale.com/ipn/ipnserver+
@@ -111,11 +118,10 @@ tailscale.com/cmd/tailscaled dependencies: (generated by github.com/tailscale/de
         tailscale.com/net/tstun                                      from tailscale.com/cmd/tailscaled+
         tailscale.com/paths                                          from tailscale.com/cmd/tailscaled+
         tailscale.com/portlist                                       from tailscale.com/ipn/ipnlocal
-        tailscale.com/safesocket                                     from tailscale.com/ipn/ipnserver
+        tailscale.com/safesocket                                     from tailscale.com/ipn/ipnserver+
         tailscale.com/smallzstd                                      from tailscale.com/ipn/ipnserver+
         tailscale.com/syncs                                          from tailscale.com/net/interfaces+
         tailscale.com/tailcfg                                        from tailscale.com/control/controlclient+
-   W 💣 tailscale.com/tempfork/wireguard-windows/firewall            from tailscale.com/cmd/tailscaled
    W    tailscale.com/tsconst                                        from tailscale.com/net/interfaces
         tailscale.com/tstime                                         from tailscale.com/wgengine/magicsock
         tailscale.com/types/empty                                    from tailscale.com/control/controlclient+
@@ -128,13 +134,13 @@ tailscale.com/cmd/tailscaled dependencies: (generated by github.com/tailscale/de
         tailscale.com/types/opt                                      from tailscale.com/control/controlclient+
         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
         tailscale.com/types/structs                                  from tailscale.com/control/controlclient+
         tailscale.com/types/wgkey                                    from tailscale.com/control/controlclient+
    L    tailscale.com/util/cmpver                                    from tailscale.com/net/dns
         tailscale.com/util/dnsname                                   from tailscale.com/ipn/ipnstate+
   LW    tailscale.com/util/endian                                    from tailscale.com/net/netns+
-   L    tailscale.com/util/lineread                                  from tailscale.com/control/controlclient+
+        tailscale.com/util/groupmember                               from tailscale.com/ipn/ipnserver
+        tailscale.com/util/lineread                                  from tailscale.com/control/controlclient+
         tailscale.com/util/osshare                                   from tailscale.com/cmd/tailscaled+
         tailscale.com/util/pidowner                                  from tailscale.com/ipn/ipnserver
         tailscale.com/util/racebuild                                 from tailscale.com/logpolicy
@@ -143,6 +149,7 @@ tailscale.com/cmd/tailscaled dependencies: (generated by github.com/tailscale/de
         tailscale.com/util/winutil                                   from tailscale.com/logpolicy+
         tailscale.com/version                                        from tailscale.com/cmd/tailscaled+
         tailscale.com/version/distro                                 from tailscale.com/control/controlclient+
+   W    tailscale.com/wf                                             from tailscale.com/cmd/tailscaled
         tailscale.com/wgengine                                       from tailscale.com/cmd/tailscaled+
         tailscale.com/wgengine/filter                                from tailscale.com/control/controlclient+
         tailscale.com/wgengine/magicsock                             from tailscale.com/wgengine+
@@ -154,7 +161,7 @@ tailscale.com/cmd/tailscaled dependencies: (generated by github.com/tailscale/de
         tailscale.com/wgengine/wglog                                 from tailscale.com/wgengine
    W 💣 tailscale.com/wgengine/winnet                                from tailscale.com/wgengine/router
         golang.org/x/crypto/blake2b                                  from golang.org/x/crypto/nacl/box
-        golang.org/x/crypto/blake2s                                  from github.com/tailscale/wireguard-go/device+
+        golang.org/x/crypto/blake2s                                  from golang.zx2c4.com/wireguard/device+
         golang.org/x/crypto/chacha20                                 from golang.org/x/crypto/chacha20poly1305
         golang.org/x/crypto/chacha20poly1305                         from crypto/tls+
         golang.org/x/crypto/cryptobyte                               from crypto/ecdsa+
@@ -163,7 +170,7 @@ tailscale.com/cmd/tailscaled dependencies: (generated by github.com/tailscale/de
         golang.org/x/crypto/hkdf                                     from crypto/tls
         golang.org/x/crypto/nacl/box                                 from tailscale.com/control/controlclient+
         golang.org/x/crypto/nacl/secretbox                           from golang.org/x/crypto/nacl/box
-        golang.org/x/crypto/poly1305                                 from github.com/tailscale/wireguard-go/device+
+        golang.org/x/crypto/poly1305                                 from golang.org/x/crypto/chacha20poly1305+
         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+
@@ -171,15 +178,15 @@ tailscale.com/cmd/tailscaled dependencies: (generated by github.com/tailscale/de
         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+
-        golang.org/x/net/ipv4                                        from github.com/tailscale/wireguard-go/device
-        golang.org/x/net/ipv6                                        from github.com/tailscale/wireguard-go/device+
+        golang.org/x/net/ipv4                                        from golang.zx2c4.com/wireguard/device
+        golang.org/x/net/ipv6                                        from golang.zx2c4.com/wireguard/device+
         golang.org/x/net/proxy                                       from tailscale.com/net/netns
    D    golang.org/x/net/route                                       from net+
         golang.org/x/sync/errgroup                                   from tailscale.com/derp
         golang.org/x/sync/singleflight                               from tailscale.com/net/dnscache
         golang.org/x/sys/cpu                                         from golang.org/x/crypto/blake2b+
-  LD    golang.org/x/sys/unix                                        from github.com/jsimonetti/rtnetlink/internal/unix+
-   W    golang.org/x/sys/windows                                     from github.com/tailscale/wireguard-go/conn+
+  LD    golang.org/x/sys/unix                                        from github.com/mdlayher/netlink+
+   W    golang.org/x/sys/windows                                     from github.com/go-ole/go-ole+
    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/mgr                             from tailscale.com/cmd/tailscaled

cmd/tailscaled/main.go

@@ -1,16 +0,0 @@
-package main
-
-import (
-	"os"
-	"strings"
-)
-
-func main() {
-	if strings.HasSuffix(os.Args[0], "tailscaled") {
-		tailscaled_main()
-	} else if strings.HasSuffix(os.Args[0], "tailscale") {
-		tailscale_main()
-	} else {
-		panic(os.Args[0])
-	}
-}

cmd/tailscaled/tailscaled.go

@@ -101,7 +101,7 @@ var subCommands = map[string]*func([]string) error{
 	"debug":                   &debugModeFunc,
 }
 
-func tailscaled_main() {
+func main() {
 	// We aren't very performance sensitive, and the parts that are
 	// performance sensitive (wireguard) try hard not to do any memory
 	// allocations. So let's be aggressive about garbage collection,

cmd/tailscaled/tailscaled.openrc

@@ -0,0 +1,23 @@
+#!/sbin/openrc-run
+
+source /etc/default/tailscaled
+
+command="/usr/sbin/tailscaled"
+command_args="--state=/var/lib/tailscale/tailscaled.state --port=$PORT --socket=/var/run/tailscale/tailscaled.sock $FLAGS"
+command_background=true
+pidfile="/run/tailscaled.pid"
+start_stop_daemon_args="-1 /var/log/tailscaled.log -2 /var/log/tailscaled.log"
+
+depend() {
+    need net
+}
+
+start_pre() {
+    mkdir -p /var/run/tailscale
+    mkdir -p /var/lib/tailscale
+    $command --cleanup
+}
+
+stop_post() {
+    $command --cleanup
+}

cmd/tailscaled/tailscaled.service

@@ -2,7 +2,7 @@
 Description=Tailscale node agent
 Documentation=https://tailscale.com/kb/
 Wants=network-pre.target
-After=network-pre.target
+After=network-pre.target NetworkManager.service systemd-resolved.service
 
 [Service]
 EnvironmentFile=/etc/default/tailscaled

cmd/tailscaled/tailscaled_windows.go

@@ -21,7 +21,6 @@ import (
 	"context"
 	"fmt"
 	"log"
-	"net"
 	"os"
 	"time"
 
@@ -32,9 +31,9 @@ import (
 	"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/wf"
 	"tailscale.com/wgengine"
 	"tailscale.com/wgengine/netstack"
 	"tailscale.com/wgengine/router"
@@ -144,13 +143,13 @@ func beFirewallKillswitch() bool {
 
 	luid, err := winipcfg.LUIDFromGUID(&guid)
 	if err != nil {
-		log.Fatalf("no interface with GUID %q", guid)
+		log.Fatalf("no interface with GUID %q: %v", guid, err)
 	}
 
-	noProtection := false
-	var dnsIPs []net.IP // unused in called code.
 	start := time.Now()
-	firewall.EnableFirewall(uint64(luid), noProtection, dnsIPs)
+	if _, err := wf.New(uint64(luid)); err != nil {
+		log.Fatalf("filewall creation failed: %v", err)
+	}
 	log.Printf("killswitch enabled, took %s", time.Since(start))
 
 	// Block until the monitor goroutine shuts us down.

control/controlclient/auto.go

@@ -576,9 +576,12 @@ func (c *Auto) sendStatus(who string, err error, url string, nm *netmap.NetworkM
 	c.logf("[v1] sendStatus: %s: %v", who, state)
 
 	var p *persist.Persist
-	var fin *empty.Message
+	var loginFin, logoutFin *empty.Message
 	if state == StateAuthenticated {
-		fin = new(empty.Message)
+		loginFin = new(empty.Message)
+	}
+	if state == StateNotAuthenticated {
+		logoutFin = new(empty.Message)
 	}
 	if nm != nil && loggedIn && synced {
 		pp := c.direct.GetPersist()
@@ -589,12 +592,13 @@ func (c *Auto) sendStatus(who string, err error, url string, nm *netmap.NetworkM
 		nm = nil
 	}
 	new := Status{
-		LoginFinished: fin,
-		URL:           url,
-		Persist:       p,
-		NetMap:        nm,
-		Hostinfo:      hi,
-		State:         state,
+		LoginFinished:  loginFin,
+		LogoutFinished: logoutFin,
+		URL:            url,
+		Persist:        p,
+		NetMap:         nm,
+		Hostinfo:       hi,
+		State:          state,
 	}
 	if err != nil {
 		new.Err = err.Error()
@@ -712,3 +716,9 @@ func (c *Auto) TestOnlySetAuthKey(authkey string) {
 func (c *Auto) TestOnlyTimeNow() time.Time {
 	return c.timeNow()
 }
+
+// SetDNS sends the SetDNSRequest request to the control plane server,
+// requesting a DNS record be created or updated.
+func (c *Auto) SetDNS(ctx context.Context, req *tailcfg.SetDNSRequest) error {
+	return c.direct.SetDNS(ctx, req)
+}

control/controlclient/client.go

@@ -74,4 +74,7 @@ type Client interface {
 	// in a separate http request. It has nothing to do with the rest of
 	// the state machine.
 	UpdateEndpoints(localPort uint16, endpoints []tailcfg.Endpoint)
+	// SetDNS sends the SetDNSRequest request to the control plane server,
+	// requesting a DNS record be created or updated.
+	SetDNS(context.Context, *tailcfg.SetDNSRequest) error
 }

control/controlclient/controlclient_test.go

@@ -22,7 +22,7 @@ func fieldsOf(t reflect.Type) (fields []string) {
 
 func TestStatusEqual(t *testing.T) {
 	// Verify that the Equal method stays in sync with reality
-	equalHandles := []string{"LoginFinished", "Err", "URL", "NetMap", "State", "Persist", "Hostinfo"}
+	equalHandles := []string{"LoginFinished", "LogoutFinished", "Err", "URL", "NetMap", "State", "Persist", "Hostinfo"}
 	if have := fieldsOf(reflect.TypeOf(Status{})); !reflect.DeepEqual(have, equalHandles) {
 		t.Errorf("Status.Equal check might be out of sync\nfields: %q\nhandled: %q\n",
 			have, equalHandles)

control/controlclient/direct.go

@@ -32,6 +32,7 @@ import (
 	"golang.org/x/crypto/nacl/box"
 	"inet.af/netaddr"
 	"tailscale.com/health"
+	"tailscale.com/ipn/ipnstate"
 	"tailscale.com/log/logheap"
 	"tailscale.com/net/dnscache"
 	"tailscale.com/net/dnsfallback"
@@ -66,6 +67,7 @@ type Direct struct {
 	debugFlags             []string
 	keepSharerAndUserSplit bool
 	skipIPForwardingCheck  bool
+	pinger                 Pinger
 
 	mu           sync.Mutex // mutex guards the following fields
 	serverKey    wgkey.Key
@@ -78,6 +80,7 @@ type Direct struct {
 	endpoints     []tailcfg.Endpoint
 	everEndpoints bool   // whether we've ever had non-empty endpoints
 	localPort     uint16 // or zero to mean auto
+	lastPingURL   string // last PingRequest.URL received, for dup suppresion
 }
 
 type Options struct {
@@ -103,6 +106,18 @@ type Options struct {
 	// forwarding works and should not be double-checked by the
 	// controlclient package.
 	SkipIPForwardingCheck bool
+
+	// Pinger optionally specifies the Pinger to use to satisfy
+	// MapResponse.PingRequest queries from the control plane.
+	// If nil, PingRequest queries are not answered.
+	Pinger Pinger
+}
+
+// Pinger is a subset of the wgengine.Engine interface, containing just the Ping method.
+type Pinger interface {
+	// Ping is a request to start a discovery or TSMP ping with the peer handling
+	// the given IP and then call cb with its ping latency & method.
+	Ping(ip netaddr.IP, useTSMP bool, cb func(*ipnstate.PingResult))
 }
 
 type Decompressor interface {
@@ -165,6 +180,7 @@ func NewDirect(opts Options) (*Direct, error) {
 		keepSharerAndUserSplit: opts.KeepSharerAndUserSplit,
 		linkMon:                opts.LinkMonitor,
 		skipIPForwardingCheck:  opts.SkipIPForwardingCheck,
+		pinger:                 opts.Pinger,
 	}
 	if opts.Hostinfo == nil {
 		c.SetHostinfo(NewHostinfo())
@@ -760,7 +776,7 @@ func (c *Direct) sendMapRequest(ctx context.Context, maxPolls int, cb func(*netm
 			health.GotStreamedMapResponse()
 		}
 
-		if pr := resp.PingRequest; pr != nil {
+		if pr := resp.PingRequest; pr != nil && c.isUniquePingRequest(pr) {
 			go answerPing(c.logf, c.httpc, pr)
 		}
 
@@ -1155,6 +1171,23 @@ func ipForwardingBroken(routes []netaddr.IPPrefix, state *interfaces.State) bool
 	return false
 }
 
+// isUniquePingRequest reports whether pr contains a new PingRequest.URL
+// not already handled, noting its value when returning true.
+func (c *Direct) isUniquePingRequest(pr *tailcfg.PingRequest) bool {
+	if pr == nil || pr.URL == "" {
+		// Bogus.
+		return false
+	}
+
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	if pr.URL == c.lastPingURL {
+		return false
+	}
+	c.lastPingURL = pr.URL
+	return true
+}
+
 func answerPing(logf logger.Logf, c *http.Client, pr *tailcfg.PingRequest) {
 	if pr.URL == "" {
 		logf("invalid PingRequest with no URL")
@@ -1211,3 +1244,50 @@ func sleepAsRequested(ctx context.Context, logf logger.Logf, timeoutReset chan<-
 		}
 	}
 }
+
+// SetDNS sends the SetDNSRequest request to the control plane server,
+// requesting a DNS record be created or updated.
+func (c *Direct) SetDNS(ctx context.Context, req *tailcfg.SetDNSRequest) error {
+	c.mu.Lock()
+	serverKey := c.serverKey
+	c.mu.Unlock()
+
+	if serverKey.IsZero() {
+		return errors.New("zero serverKey")
+	}
+	machinePrivKey, err := c.getMachinePrivKey()
+	if err != nil {
+		return fmt.Errorf("getMachinePrivKey: %w", err)
+	}
+	if machinePrivKey.IsZero() {
+		return errors.New("getMachinePrivKey returned zero key")
+	}
+
+	bodyData, err := encode(req, &serverKey, &machinePrivKey)
+	if err != nil {
+		return err
+	}
+	body := bytes.NewReader(bodyData)
+
+	u := fmt.Sprintf("%s/machine/%s/set-dns", c.serverURL, machinePrivKey.Public().HexString())
+	hreq, err := http.NewRequestWithContext(ctx, "POST", u, body)
+	if err != nil {
+		return err
+	}
+	res, err := c.httpc.Do(hreq)
+	if err != nil {
+		return err
+	}
+	defer res.Body.Close()
+	if res.StatusCode != 200 {
+		msg, _ := ioutil.ReadAll(res.Body)
+		return fmt.Errorf("set-dns response: %v, %.200s", res.Status, strings.TrimSpace(string(msg)))
+	}
+	var setDNSRes struct{} // no fields yet
+	if err := decode(res, &setDNSRes, &serverKey, &machinePrivKey); err != nil {
+		c.logf("error decoding SetDNSResponse with server key %s and machine key %s: %v", serverKey, machinePrivKey.Public(), err)
+		return fmt.Errorf("set-dns-response: %v", err)
+	}
+
+	return nil
+}

control/controlclient/hostinfo_linux.go

@@ -9,13 +9,11 @@ package controlclient
 import (
 	"bytes"
 	"fmt"
-	"io"
 	"io/ioutil"
-	"os"
 	"strings"
 	"syscall"
 
-	"go4.org/mem"
+	"tailscale.com/hostinfo"
 	"tailscale.com/util/lineread"
 	"tailscale.com/version/distro"
 )
@@ -56,11 +54,11 @@ func osVersionLinux() string {
 		}
 		attrBuf.WriteByte(byte(b))
 	}
-	if inContainer() {
+	if hostinfo.InContainer() {
 		attrBuf.WriteString("; container")
 	}
-	if inKnative() {
-		attrBuf.WriteString("; env=kn")
+	if env := hostinfo.GetEnvType(); env != "" {
+		fmt.Fprintf(&attrBuf, "; env=%s", env)
 	}
 	attr := attrBuf.String()
 
@@ -93,31 +91,3 @@ func osVersionLinux() string {
 	}
 	return fmt.Sprintf("Other%s", attr)
 }
-
-func inContainer() (ret bool) {
-	lineread.File("/proc/1/cgroup", func(line []byte) error {
-		if mem.Contains(mem.B(line), mem.S("/docker/")) ||
-			mem.Contains(mem.B(line), mem.S("/lxc/")) {
-			ret = true
-			return io.EOF // arbitrary non-nil error to stop loop
-		}
-		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/sign_supported.go

@@ -18,7 +18,7 @@ import (
 	"fmt"
 	"sync"
 
-	"github.com/github/certstore"
+	"github.com/tailscale/certstore"
 	"tailscale.com/tailcfg"
 	"tailscale.com/types/wgkey"
 	"tailscale.com/util/winutil"

control/controlclient/status.go

@@ -64,11 +64,12 @@ func (s State) String() string {
 }
 
 type Status struct {
-	_             structs.Incomparable
-	LoginFinished *empty.Message // nonempty when login finishes
-	Err           string
-	URL           string             // interactive URL to visit to finish logging in
-	NetMap        *netmap.NetworkMap // server-pushed configuration
+	_              structs.Incomparable
+	LoginFinished  *empty.Message // nonempty when login finishes
+	LogoutFinished *empty.Message // nonempty when logout finishes
+	Err            string
+	URL            string             // interactive URL to visit to finish logging in
+	NetMap         *netmap.NetworkMap // server-pushed configuration
 
 	// The internal state should not be exposed outside this
 	// package, but we have some automated tests elsewhere that need to
@@ -86,6 +87,7 @@ func (s *Status) Equal(s2 *Status) bool {
 	}
 	return s != nil && s2 != nil &&
 		(s.LoginFinished == nil) == (s2.LoginFinished == nil) &&
+		(s.LogoutFinished == nil) == (s2.LogoutFinished == nil) &&
 		s.Err == s2.Err &&
 		s.URL == s2.URL &&
 		reflect.DeepEqual(s.Persist, s2.Persist) &&

derp/derp_server.go

@@ -20,18 +20,24 @@ import (
 	"io"
 	"io/ioutil"
 	"log"
+	"math"
 	"math/big"
 	"math/rand"
+	"net/http"
 	"os"
+	"os/exec"
 	"runtime"
 	"strconv"
 	"strings"
 	"sync"
+	"sync/atomic"
 	"time"
 
 	"go4.org/mem"
 	"golang.org/x/crypto/nacl/box"
 	"golang.org/x/sync/errgroup"
+	"inet.af/netaddr"
+	"tailscale.com/client/tailscale"
 	"tailscale.com/disco"
 	"tailscale.com/metrics"
 	"tailscale.com/types/key"
@@ -120,6 +126,11 @@ type Server struct {
 	multiForwarderCreated    expvar.Int
 	multiForwarderDeleted    expvar.Int
 	removePktForwardOther    expvar.Int
+	avgQueueDuration         *uint64 // In milliseconds; accessed atomically
+
+	// verifyClients only accepts client connections to the DERP server if the clientKey is a
+	// known peer in the network, as specified by a running tailscaled's client's local api.
+	verifyClients bool
 
 	mu          sync.Mutex
 	closed      bool
@@ -138,6 +149,9 @@ type Server struct {
 	// because it includes intra-region forwarded packets as the
 	// src.
 	sentTo map[key.Public]map[key.Public]int64 // src => dst => dst's latest sclient.connNum
+
+	// maps from netaddr.IPPort to a client's public key
+	keyOfAddr map[netaddr.IPPort]key.Public
 }
 
 // PacketForwarder is something that can forward packets.
@@ -182,6 +196,8 @@ func NewServer(privateKey key.Private, logf logger.Logf) *Server {
 		memSys0:              ms.Sys,
 		watchers:             map[*sclient]bool{},
 		sentTo:               map[key.Public]map[key.Public]int64{},
+		avgQueueDuration:     new(uint64),
+		keyOfAddr:            map[netaddr.IPPort]key.Public{},
 	}
 	s.initMetacert()
 	s.packetsRecvDisco = s.packetsRecvByKind.Get("disco")
@@ -203,6 +219,13 @@ func (s *Server) SetMeshKey(v string) {
 	s.meshKey = v
 }
 
+// SetVerifyClients sets whether this DERP server verifies clients through tailscaled.
+//
+// It must be called before serving begins.
+func (s *Server) SetVerifyClient(v bool) {
+	s.verifyClients = v
+}
+
 // HasMeshKey reports whether the server is configured with a mesh key.
 func (s *Server) HasMeshKey() bool { return s.meshKey != "" }
 
@@ -339,6 +362,7 @@ func (s *Server) registerClient(c *sclient) {
 	if _, ok := s.clientsMesh[c.key]; !ok {
 		s.clientsMesh[c.key] = nil // just for varz of total users in cluster
 	}
+	s.keyOfAddr[c.remoteIPPort] = c.key
 	s.curClients.Add(1)
 	s.broadcastPeerStateChangeLocked(c.key, true)
 }
@@ -373,6 +397,8 @@ func (s *Server) unregisterClient(c *sclient) {
 		delete(s.watchers, c)
 	}
 
+	delete(s.keyOfAddr, c.remoteIPPort)
+
 	s.curClients.Add(-1)
 	if c.preferred {
 		s.curHomeClients.Add(-1)
@@ -446,20 +472,23 @@ func (s *Server) accept(nc Conn, brw *bufio.ReadWriter, remoteAddr string, connN
 	ctx, cancel := context.WithCancel(context.Background())
 	defer cancel()
 
+	remoteIPPort, _ := netaddr.ParseIPPort(remoteAddr)
+
 	c := &sclient{
-		connNum:     connNum,
-		s:           s,
-		key:         clientKey,
-		nc:          nc,
-		br:          br,
-		bw:          bw,
-		logf:        logger.WithPrefix(s.logf, fmt.Sprintf("derp client %v/%x: ", remoteAddr, clientKey)),
-		done:        ctx.Done(),
-		remoteAddr:  remoteAddr,
-		connectedAt: time.Now(),
-		sendQueue:   make(chan pkt, perClientSendQueueDepth),
-		peerGone:    make(chan key.Public),
-		canMesh:     clientInfo.MeshKey != "" && clientInfo.MeshKey == s.meshKey,
+		connNum:      connNum,
+		s:            s,
+		key:          clientKey,
+		nc:           nc,
+		br:           br,
+		bw:           bw,
+		logf:         logger.WithPrefix(s.logf, fmt.Sprintf("derp client %v/%x: ", remoteAddr, clientKey)),
+		done:         ctx.Done(),
+		remoteAddr:   remoteAddr,
+		remoteIPPort: remoteIPPort,
+		connectedAt:  time.Now(),
+		sendQueue:    make(chan pkt, perClientSendQueueDepth),
+		peerGone:     make(chan key.Public),
+		canMesh:      clientInfo.MeshKey != "" && clientInfo.MeshKey == s.meshKey,
 	}
 	if c.canMesh {
 		c.meshUpdate = make(chan struct{})
@@ -611,8 +640,9 @@ func (c *sclient) handleFrameForwardPacket(ft frameType, fl uint32) error {
 	}
 
 	return c.sendPkt(dst, pkt{
-		bs:  contents,
-		src: srcKey,
+		bs:         contents,
+		enqueuedAt: time.Now(),
+		src:        srcKey,
 	})
 }
 
@@ -665,8 +695,9 @@ func (c *sclient) handleFrameSendPacket(ft frameType, fl uint32) error {
 	}
 
 	p := pkt{
-		bs:  contents,
-		src: c.key,
+		bs:         contents,
+		enqueuedAt: time.Now(),
+		src:        c.key,
 	}
 	return c.sendPkt(dst, p)
 }
@@ -696,7 +727,7 @@ func (c *sclient) sendPkt(dst *sclient, p pkt) error {
 		}
 
 		select {
-		case <-dst.sendQueue:
+		case pkt := <-dst.sendQueue:
 			s.packetsDropped.Add(1)
 			s.packetsDroppedQueueHead.Add(1)
 			if verboseDropKeys[dstKey] {
@@ -705,6 +736,7 @@ func (c *sclient) sendPkt(dst *sclient, p pkt) error {
 				msg := fmt.Sprintf("tail drop %s -> %s", p.src.ShortString(), dstKey.ShortString())
 				c.s.limitedLogf(msg)
 			}
+			c.recordQueueTime(pkt.enqueuedAt)
 			if debug {
 				c.logf("dropping packet from client %x queue head", dstKey)
 			}
@@ -750,8 +782,17 @@ func (c *sclient) requestMeshUpdate() {
 }
 
 func (s *Server) verifyClient(clientKey key.Public, info *clientInfo) error {
-	// TODO(crawshaw): implement policy constraints on who can use the DERP server
-	// TODO(bradfitz): ... and at what rate.
+	if !s.verifyClients {
+		return nil
+	}
+	status, err := tailscale.Status(context.TODO())
+	if err != nil {
+		return fmt.Errorf("failed to query local tailscaled status: %w", err)
+	}
+	if _, exists := status.Peer[clientKey]; !exists {
+		return fmt.Errorf("client %v not in set of peers", clientKey)
+	}
+	// TODO(bradfitz): add policy for configurable bandwidth rate per client?
 	return nil
 }
 
@@ -885,18 +926,19 @@ func (s *Server) recvForwardPacket(br *bufio.Reader, frameLen uint32) (srcKey, d
 // (The "s" prefix is to more explicitly distinguish it from Client in derp_client.go)
 type sclient struct {
 	// Static after construction.
-	connNum    int64 // process-wide unique counter, incremented each Accept
-	s          *Server
-	nc         Conn
-	key        key.Public
-	info       clientInfo
-	logf       logger.Logf
-	done       <-chan struct{} // closed when connection closes
-	remoteAddr string          // usually ip:port from net.Conn.RemoteAddr().String()
-	sendQueue  chan pkt        // packets queued to this client; never closed
-	peerGone   chan key.Public // write request that a previous sender has disconnected (not used by mesh peers)
-	meshUpdate chan struct{}   // write request to write peerStateChange
-	canMesh    bool            // clientInfo had correct mesh token for inter-region routing
+	connNum      int64 // process-wide unique counter, incremented each Accept
+	s            *Server
+	nc           Conn
+	key          key.Public
+	info         clientInfo
+	logf         logger.Logf
+	done         <-chan struct{} // closed when connection closes
+	remoteAddr   string          // usually ip:port from net.Conn.RemoteAddr().String()
+	remoteIPPort netaddr.IPPort  // zero if remoteAddr is not ip:port.
+	sendQueue    chan pkt        // packets queued to this client; never closed
+	peerGone     chan key.Public // write request that a previous sender has disconnected (not used by mesh peers)
+	meshUpdate   chan struct{}   // write request to write peerStateChange
+	canMesh      bool            // clientInfo had correct mesh token for inter-region routing
 
 	// Owned by run, not thread-safe.
 	br          *bufio.Reader
@@ -927,11 +969,13 @@ type pkt struct {
 	// src is the who's the sender of the packet.
 	src key.Public
 
+	// enqueuedAt is when a packet was put onto a queue before it was sent,
+	// and is used for reporting metrics on the duration of packets in the queue.
+	enqueuedAt time.Time
+
 	// bs is the data packet bytes.
 	// The memory is owned by pkt.
 	bs []byte
-
-	// TODO(danderson): enqueue time, to measure queue latency?
 }
 
 func (c *sclient) setPreferred(v bool) {
@@ -959,6 +1003,25 @@ func (c *sclient) setPreferred(v bool) {
 	}
 }
 
+// expMovingAverage returns the new moving average given the previous average,
+// a new value, and an alpha decay factor.
+// https://en.wikipedia.org/wiki/Moving_average#Exponential_moving_average
+func expMovingAverage(prev, newValue, alpha float64) float64 {
+	return alpha*newValue + (1-alpha)*prev
+}
+
+// recordQueueTime updates the average queue duration metric after a packet has been sent.
+func (c *sclient) recordQueueTime(enqueuedAt time.Time) {
+	elapsed := float64(time.Since(enqueuedAt).Milliseconds())
+	for {
+		old := atomic.LoadUint64(c.s.avgQueueDuration)
+		newAvg := expMovingAverage(math.Float64frombits(old), elapsed, 0.1)
+		if atomic.CompareAndSwapUint64(c.s.avgQueueDuration, old, math.Float64bits(newAvg)) {
+			break
+		}
+	}
+}
+
 func (c *sclient) sendLoop(ctx context.Context) error {
 	defer func() {
 		// If the sender shuts down unilaterally due to an error, close so
@@ -1002,6 +1065,7 @@ func (c *sclient) sendLoop(ctx context.Context) error {
 			continue
 		case msg := <-c.sendQueue:
 			werr = c.sendPacket(msg.src, msg.bs)
+			c.recordQueueTime(msg.enqueuedAt)
 			continue
 		case <-keepAliveTick.C:
 			werr = c.sendKeepAlive()
@@ -1025,6 +1089,7 @@ func (c *sclient) sendLoop(ctx context.Context) error {
 			continue
 		case msg := <-c.sendQueue:
 			werr = c.sendPacket(msg.src, msg.bs)
+			c.recordQueueTime(msg.enqueuedAt)
 		case <-keepAliveTick.C:
 			werr = c.sendKeepAlive()
 		}
@@ -1290,6 +1355,9 @@ func (s *Server) ExpVar() expvar.Var {
 	m.Set("multiforwarder_created", &s.multiForwarderCreated)
 	m.Set("multiforwarder_deleted", &s.multiForwarderDeleted)
 	m.Set("packet_forwarder_delete_other_value", &s.removePktForwardOther)
+	m.Set("average_queue_duration_ms", expvar.Func(func() interface{} {
+		return math.Float64frombits(atomic.LoadUint64(s.avgQueueDuration))
+	}))
 	var expvarVersion expvar.String
 	expvarVersion.Set(version.Long)
 	m.Set("version", &expvarVersion)
@@ -1365,3 +1433,84 @@ func writePublicKey(bw *bufio.Writer, key *key.Public) error {
 	}
 	return nil
 }
+
+const minTimeBetweenLogs = 2 * time.Second
+
+// BytesSentRecv records the number of bytes that have been sent since the last traffic check
+// for a given process, as well as the public key of the process sending those bytes.
+type BytesSentRecv struct {
+	Sent uint64
+	Recv uint64
+	// Key is the public key of the client which sent/received these bytes.
+	Key key.Public
+}
+
+// parseSSOutput parses the output from the specific call to ss in ServeDebugTraffic.
+// Separated out for ease of testing.
+func parseSSOutput(raw string) map[netaddr.IPPort]BytesSentRecv {
+	newState := map[netaddr.IPPort]BytesSentRecv{}
+	// parse every 2 lines and get src and dst ips, and kv pairs
+	lines := strings.Split(raw, "\n")
+	for i := 0; i < len(lines); i += 2 {
+		ipInfo := strings.Fields(strings.TrimSpace(lines[i]))
+		if len(ipInfo) < 5 {
+			continue
+		}
+		src, err := netaddr.ParseIPPort(ipInfo[4])
+		if err != nil {
+			continue
+		}
+		stats := strings.Fields(strings.TrimSpace(lines[i+1]))
+		stat := BytesSentRecv{}
+		for _, s := range stats {
+			if strings.Contains(s, "bytes_sent") {
+				sent, err := strconv.Atoi(s[strings.Index(s, ":")+1:])
+				if err == nil {
+					stat.Sent = uint64(sent)
+				}
+			} else if strings.Contains(s, "bytes_received") {
+				recv, err := strconv.Atoi(s[strings.Index(s, ":")+1:])
+				if err == nil {
+					stat.Recv = uint64(recv)
+				}
+			}
+		}
+		newState[src] = stat
+	}
+	return newState
+}
+
+func (s *Server) ServeDebugTraffic(w http.ResponseWriter, r *http.Request) {
+	prevState := map[netaddr.IPPort]BytesSentRecv{}
+	enc := json.NewEncoder(w)
+	for r.Context().Err() == nil {
+		output, err := exec.Command("ss", "-i", "-H", "-t").Output()
+		if err != nil {
+			fmt.Fprintf(w, "ss failed: %v", err)
+			return
+		}
+		newState := parseSSOutput(string(output))
+		s.mu.Lock()
+		for k, next := range newState {
+			prev := prevState[k]
+			if prev.Sent < next.Sent || prev.Recv < next.Recv {
+				if pkey, ok := s.keyOfAddr[k]; ok {
+					next.Key = pkey
+					if err := enc.Encode(next); err != nil {
+						s.mu.Unlock()
+						return
+					}
+				}
+			}
+		}
+		s.mu.Unlock()
+		prevState = newState
+		if _, err := fmt.Fprintln(w); err != nil {
+			return
+		}
+		if f, ok := w.(http.Flusher); ok {
+			f.Flush()
+		}
+		time.Sleep(minTimeBetweenLogs)
+	}
+}

derp/derp_test.go

@@ -948,3 +948,14 @@ func waitConnect(t testing.TB, c *Client) {
 		t.Fatalf("client first Recv was unexpected type %T", v)
 	}
 }
+
+func TestParseSSOutput(t *testing.T) {
+	contents, err := ioutil.ReadFile("testdata/example_ss.txt")
+	if err != nil {
+		t.Errorf("ioutil.Readfile(example_ss.txt) failed: %v", err)
+	}
+	seen := parseSSOutput(string(contents))
+	if len(seen) == 0 {
+		t.Errorf("parseSSOutput expected non-empty map")
+	}
+}

derp/testdata/example_ss.txt

@@ -0,0 +1,8 @@
+ESTAB 0      0       10.255.1.11:35238  34.210.105.16:https
+         cubic wscale:7,7 rto:236 rtt:34.14/3.432 ato:40 mss:1448 pmtu:1500 rcvmss:1448 advmss:1448 cwnd:8 ssthresh:6 bytes_sent:38056577 bytes_retrans:2918 bytes_acked:38053660 bytes_received:6973211 segs_out:165090 segs_in:124227 data_segs_out:78018 data_segs_in:71645 send 2.71Mbps lastsnd:1156 lastrcv:1120 lastack:1120 pacing_rate 3.26Mbps delivery_rate 2.35Mbps delivered:78017 app_limited busy:2586132ms retrans:0/6 dsack_dups:4 reordering:5 reord_seen:15 rcv_rtt:126355 rcv_space:65780 rcv_ssthresh:541928 minrtt:26.632
+ESTAB 0      80     100.79.58.14:ssh    100.95.73.104:58145
+         cubic wscale:6,7 rto:224 rtt:23.051/2.03 ato:172 mss:1228 pmtu:1280 rcvmss:1228 advmss:1228 cwnd:10 ssthresh:94 bytes_sent:1591815 bytes_retrans:944 bytes_acked:1590791 bytes_received:158925 segs_out:8070 segs_in:8858 data_segs_out:7452 data_segs_in:3789 send 4.26Mbps lastsnd:4 lastrcv:4 lastack:4 pacing_rate 8.52Mbps delivery_rate 10.9Mbps delivered:7451 app_limited busy:61656ms unacked:2 retrans:0/10 dsack_dups:10 rcv_rtt:174712 rcv_space:65025 rcv_ssthresh:64296 minrtt:16.186                                   
+ESTAB 0      374     10.255.1.11:43254 167.172.206.31:https
+         cubic wscale:7,7 rto:224 rtt:22.55/1.941 ato:40 mss:1448 pmtu:1500 rcvmss:1448 advmss:1448 cwnd:6 ssthresh:4 bytes_sent:14594668 bytes_retrans:173314 bytes_acked:14420981 bytes_received:4207111 segs_out:80566 segs_in:70310 data_segs_out:24317 data_segs_in:20365 send 3.08Mbps lastsnd:4 lastrcv:4 lastack:4 pacing_rate 3.7Mbps delivery_rate 3.05Mbps delivered:24111 app_limited busy:184820ms unacked:2 retrans:0/185 dsack_dups:1 reord_seen:3 rcv_rtt:651.262 rcv_space:226657 rcv_ssthresh:1557136 minrtt:10.18           
+ESTAB 0      0       10.255.1.11:33036    3.121.18.47:https
+         cubic wscale:7,7 rto:372 rtt:168.408/2.044 ato:40 mss:1448 pmtu:1500 rcvmss:1448 advmss:1448 cwnd:10 bytes_sent:27500 bytes_acked:27501 bytes_received:1386524 segs_out:10990 segs_in:11037 data_segs_out:303 data_segs_in:3414 send 688kbps lastsnd:125776 lastrcv:9640 lastack:22760 pacing_rate 1.38Mbps delivery_rate 482kbps delivered:304 app_limited busy:43024ms rcv_rtt:3345.12 rcv_space:62431 rcv_ssthresh:760472 minrtt:168.867

go.mod

@@ -3,48 +3,47 @@ module tailscale.com
 go 1.16
 
 require (
-	github.com/alexbrainman/sspi v0.0.0-20180613141037-e580b900e9f5
-	github.com/anmitsu/go-shlex v0.0.0-20161002113705-648efa622239 // indirect
-	github.com/coreos/go-iptables v0.4.5
-	github.com/flynn/go-shlex v0.0.0-20150515145356-3f9db97f8568 // indirect
-	github.com/frankban/quicktest v1.12.1
-	github.com/github/certstore v0.1.0
-	github.com/gliderlabs/ssh v0.2.2
+	github.com/alexbrainman/sspi v0.0.0-20210105120005-909beea2cc74
+	github.com/anmitsu/go-shlex v0.0.0-20200514113438-38f4b401e2be // indirect
+	github.com/aws/aws-sdk-go v1.38.52
+	github.com/coreos/go-iptables v0.6.0
+	github.com/frankban/quicktest v1.13.0
+	github.com/gliderlabs/ssh v0.3.2
 	github.com/go-multierror/multierror v1.0.2
-	github.com/go-ole/go-ole v1.2.4
-	github.com/godbus/dbus/v5 v5.0.3
-	github.com/google/go-cmp v0.5.5
-	github.com/goreleaser/nfpm v1.1.10
-	github.com/jsimonetti/rtnetlink v0.0.0-20210212075122-66c871082f2b
+	github.com/go-ole/go-ole v1.2.5
+	github.com/godbus/dbus/v5 v5.0.4
+	github.com/google/go-cmp v0.5.6
+	github.com/google/goexpect v0.0.0-20210430020637-ab937bf7fd6f
+	github.com/google/uuid v1.1.2
+	github.com/goreleaser/nfpm v1.10.3
+	github.com/jsimonetti/rtnetlink v0.0.0-20210525051524-4cc836578190
 	github.com/kballard/go-shellquote v0.0.0-20180428030007-95032a82bc51
-	github.com/klauspost/compress v1.10.10
+	github.com/klauspost/compress v1.12.2
 	github.com/kr/pty v1.1.8
-	github.com/mdlayher/netlink v1.3.2
-	github.com/mdlayher/sdnotify v0.0.0-20200625151349-e4a4f32afc4a
-	github.com/miekg/dns v1.1.30
-	github.com/pborman/getopt v0.0.0-20190409184431-ee0cd42419d3
+	github.com/mdlayher/netlink v1.4.1
+	github.com/mdlayher/sdnotify v0.0.0-20210228150836-ea3ec207d697
+	github.com/miekg/dns v1.1.42
+	github.com/pborman/getopt v1.1.0
 	github.com/peterbourgon/ff/v2 v2.0.0
-	github.com/pkg/errors v0.9.1 // indirect
+	github.com/pkg/sftp v1.13.0
+	github.com/tailscale/certstore v0.0.0-20210528134328-066c94b793d3
 	github.com/tailscale/depaware v0.0.0-20201214215404-77d1e9757027
-	github.com/tailscale/wireguard-go v0.0.0-20210510192616-d1aa5623121d
 	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 v0.0.0-20210510120150-4163338589ed
+	golang.org/x/crypto v0.0.0-20210513164829-c07d793c2f9a
+	golang.org/x/net v0.0.0-20210525063256-abc453219eb5
 	golang.org/x/sync v0.0.0-20210220032951-036812b2e83c
-	golang.org/x/sys v0.0.0-20210510120138-977fb7262007
-	golang.org/x/term v0.0.0-20210317153231-de623e64d2a6
-	golang.org/x/time v0.0.0-20210220033141-f8bda1e9f3ba
-	golang.org/x/tools v0.1.0
-	golang.zx2c4.com/wireguard/windows v0.1.2-0.20201113162609-9b85be97fdf8
-	gopkg.in/yaml.v2 v2.2.8 // indirect
-	honnef.co/go/tools v0.1.0
-	inet.af/netaddr v0.0.0-20210515010201-ad03edc7c841
-	inet.af/netstack v0.0.0-20210317161235-a1bf4e56ef22
-	inet.af/peercred v0.0.0-20210302202138-56e694897155
+	golang.org/x/sys v0.0.0-20210616094352-59db8d763f22
+	golang.org/x/term v0.0.0-20210503060354-a79de5458b56
+	golang.org/x/time v0.0.0-20210611083556-38a9dc6acbc6
+	golang.org/x/tools v0.1.2
+	golang.zx2c4.com/wireguard v0.0.0-20210525143454-64cb82f2b3f5
+	golang.zx2c4.com/wireguard/windows v0.3.15-0.20210525143335-94c0476d63e3
+	honnef.co/go/tools v0.1.4
+	inet.af/netaddr v0.0.0-20210602152128-50f8686885e3
+	inet.af/netstack v0.0.0-20210622165351-29b14ebc044e
+	inet.af/peercred v0.0.0-20210318190834-4259e17bb763
 	inet.af/wf v0.0.0-20210516214145-a5343001b756
 	rsc.io/goversion v1.2.0
 )
-
-replace github.com/github/certstore => github.com/cyolosecurity/certstore v0.0.0-20200922073901-ece7f1d353c2

hostinfo/hostinfo.go

@@ -0,0 +1,117 @@
+// 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 hostinfo answers questions about the host environment that Tailscale is
+// running on.
+//
+// TODO(bradfitz): move more of control/controlclient/hostinfo_* into this package.
+package hostinfo
+
+import (
+	"io"
+	"os"
+	"runtime"
+	"sync/atomic"
+
+	"go4.org/mem"
+	"tailscale.com/util/lineread"
+)
+
+// EnvType represents a known environment type.
+// The empty string, the default, means unknown.
+type EnvType string
+
+const (
+	KNative         = EnvType("kn")
+	AWSLambda       = EnvType("lm")
+	Heroku          = EnvType("hr")
+	AzureAppService = EnvType("az")
+)
+
+var envType atomic.Value // of EnvType
+
+func GetEnvType() EnvType {
+	if e, ok := envType.Load().(EnvType); ok {
+		return e
+	}
+	e := getEnvType()
+	envType.Store(e)
+	return e
+}
+
+func getEnvType() EnvType {
+	if inKnative() {
+		return KNative
+	}
+	if inAWSLambda() {
+		return AWSLambda
+	}
+	if inHerokuDyno() {
+		return Heroku
+	}
+	if inAzureAppService() {
+		return AzureAppService
+	}
+	return ""
+}
+
+// InContainer reports whether we're running in a container.
+func InContainer() bool {
+	if runtime.GOOS != "linux" {
+		return false
+	}
+	var ret bool
+	lineread.File("/proc/1/cgroup", func(line []byte) error {
+		if mem.Contains(mem.B(line), mem.S("/docker/")) ||
+			mem.Contains(mem.B(line), mem.S("/lxc/")) {
+			ret = true
+			return io.EOF // arbitrary non-nil error to stop loop
+		}
+		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 ret
+}
+
+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
+}
+
+func inAWSLambda() bool {
+	// https://docs.aws.amazon.com/lambda/latest/dg/configuration-envvars.html
+	if os.Getenv("AWS_LAMBDA_FUNCTION_NAME") != "" &&
+		os.Getenv("AWS_LAMBDA_FUNCTION_VERSION") != "" &&
+		os.Getenv("AWS_LAMBDA_INITIALIZATION_TYPE") != "" &&
+		os.Getenv("AWS_LAMBDA_RUNTIME_API") != "" {
+		return true
+	}
+	return false
+}
+
+func inHerokuDyno() bool {
+	// https://devcenter.heroku.com/articles/dynos#local-environment-variables
+	if os.Getenv("PORT") != "" && os.Getenv("DYNO") != "" {
+		return true
+	}
+	return false
+}
+
+func inAzureAppService() bool {
+	if os.Getenv("APPSVC_RUN_ZIP") != "" && os.Getenv("WEBSITE_STACK") != "" &&
+		os.Getenv("WEBSITE_AUTH_AUTO_AAD") != "" {
+		return true
+	}
+	return false
+}

internal/deephash/deephash.go

@@ -9,26 +9,28 @@ package deephash
 import (
 	"bufio"
 	"crypto/sha256"
+	"encoding/hex"
 	"fmt"
+	"hash"
 	"reflect"
-
-	"inet.af/netaddr"
-	"tailscale.com/tailcfg"
-	"tailscale.com/types/wgkey"
+	"strconv"
+	"sync"
 )
 
-func Hash(v ...interface{}) string {
+func calcHash(v interface{}) string {
 	h := sha256.New()
-	// 64 matches the chunk size in crypto/sha256/sha256.go
-	b := bufio.NewWriterSize(h, 64)
-	Print(b, v)
+	b := bufio.NewWriterSize(h, h.BlockSize())
+	scratch := make([]byte, 0, 128)
+	printTo(b, v, scratch)
 	b.Flush()
-	return fmt.Sprintf("%x", h.Sum(nil))
+	scratch = h.Sum(scratch[:0])
+	hex.Encode(scratch[:cap(scratch)], scratch[:sha256.Size])
+	return string(scratch[:sha256.Size*2])
 }
 
 // UpdateHash sets last to the hash of v and reports whether its value changed.
 func UpdateHash(last *string, v ...interface{}) (changed bool) {
-	sig := Hash(v)
+	sig := calcHash(v)
 	if *last != sig {
 		*last = sig
 		return true
@@ -36,81 +38,30 @@ func UpdateHash(last *string, v ...interface{}) (changed bool) {
 	return false
 }
 
-func Print(w *bufio.Writer, v ...interface{}) {
-	print(w, reflect.ValueOf(v), make(map[uintptr]bool))
+func printTo(w *bufio.Writer, v interface{}, scratch []byte) {
+	print(w, reflect.ValueOf(v), make(map[uintptr]bool), scratch)
 }
 
-var (
-	netaddrIPType       = reflect.TypeOf(netaddr.IP{})
-	netaddrIPPrefix     = reflect.TypeOf(netaddr.IPPrefix{})
-	wgkeyKeyType        = reflect.TypeOf(wgkey.Key{})
-	wgkeyPrivateType    = reflect.TypeOf(wgkey.Private{})
-	tailcfgDiscoKeyType = reflect.TypeOf(tailcfg.DiscoKey{})
-)
+var appenderToType = reflect.TypeOf((*appenderTo)(nil)).Elem()
 
-func print(w *bufio.Writer, v reflect.Value, visited map[uintptr]bool) {
+type appenderTo interface {
+	AppendTo([]byte) []byte
+}
+
+// print hashes v into w.
+// It reports whether it was able to do so without hitting a cycle.
+func print(w *bufio.Writer, v reflect.Value, visited map[uintptr]bool, scratch []byte) (acyclic bool) {
 	if !v.IsValid() {
-		return
+		return true
 	}
 
-	// Special case some common types.
 	if v.CanInterface() {
-		switch v.Type() {
-		case netaddrIPType:
-			var b []byte
-			var err error
-			if v.CanAddr() {
-				x := v.Addr().Interface().(*netaddr.IP)
-				b, err = x.MarshalText()
-			} else {
-				x := v.Interface().(netaddr.IP)
-				b, err = x.MarshalText()
-			}
-			if err == nil {
-				w.Write(b)
-				return
-			}
-		case netaddrIPPrefix:
-			var b []byte
-			var err error
-			if v.CanAddr() {
-				x := v.Addr().Interface().(*netaddr.IPPrefix)
-				b, err = x.MarshalText()
-			} else {
-				x := v.Interface().(netaddr.IPPrefix)
-				b, err = x.MarshalText()
-			}
-			if err == nil {
-				w.Write(b)
-				return
-			}
-		case wgkeyKeyType:
-			if v.CanAddr() {
-				x := v.Addr().Interface().(*wgkey.Key)
-				w.Write(x[:])
-			} else {
-				x := v.Interface().(wgkey.Key)
-				w.Write(x[:])
-			}
-			return
-		case wgkeyPrivateType:
-			if v.CanAddr() {
-				x := v.Addr().Interface().(*wgkey.Private)
-				w.Write(x[:])
-			} else {
-				x := v.Interface().(wgkey.Private)
-				w.Write(x[:])
-			}
-			return
-		case tailcfgDiscoKeyType:
-			if v.CanAddr() {
-				x := v.Addr().Interface().(*tailcfg.DiscoKey)
-				w.Write(x[:])
-			} else {
-				x := v.Interface().(tailcfg.DiscoKey)
-				w.Write(x[:])
-			}
-			return
+		// Use AppendTo methods, if available and cheap.
+		if v.CanAddr() && v.Type().Implements(appenderToType) {
+			a := v.Addr().Interface().(appenderTo)
+			scratch = a.AppendTo(scratch[:0])
+			w.Write(scratch)
+			return true
 		}
 	}
 
@@ -121,43 +72,45 @@ func print(w *bufio.Writer, v reflect.Value, visited map[uintptr]bool) {
 	case reflect.Ptr:
 		ptr := v.Pointer()
 		if visited[ptr] {
-			return
+			return false
 		}
 		visited[ptr] = true
-		print(w, v.Elem(), visited)
-		return
+		return print(w, v.Elem(), visited, scratch)
 	case reflect.Struct:
+		acyclic = true
 		w.WriteString("struct{\n")
 		for i, n := 0, v.NumField(); i < n; i++ {
 			fmt.Fprintf(w, " [%d]: ", i)
-			print(w, v.Field(i), visited)
+			if !print(w, v.Field(i), visited, scratch) {
+				acyclic = false
+			}
 			w.WriteString("\n")
 		}
 		w.WriteString("}\n")
+		return acyclic
 	case reflect.Slice, reflect.Array:
 		if v.Type().Elem().Kind() == reflect.Uint8 && v.CanInterface() {
 			fmt.Fprintf(w, "%q", v.Interface())
-			return
+			return true
 		}
 		fmt.Fprintf(w, "[%d]{\n", v.Len())
+		acyclic = true
 		for i, ln := 0, v.Len(); i < ln; i++ {
 			fmt.Fprintf(w, " [%d]: ", i)
-			print(w, v.Index(i), visited)
+			if !print(w, v.Index(i), visited, scratch) {
+				acyclic = false
+			}
 			w.WriteString("\n")
 		}
 		w.WriteString("}\n")
+		return acyclic
 	case reflect.Interface:
-		print(w, v.Elem(), visited)
+		return print(w, v.Elem(), visited, scratch)
 	case reflect.Map:
-		sm := newSortedMap(v)
-		fmt.Fprintf(w, "map[%d]{\n", len(sm.Key))
-		for i, k := range sm.Key {
-			print(w, k, visited)
-			w.WriteString(": ")
-			print(w, sm.Value[i], visited)
-			w.WriteString("\n")
+		if hashMapAcyclic(w, v, visited, scratch) {
+			return true
 		}
-		w.WriteString("}\n")
+		return hashMapFallback(w, v, visited, scratch)
 	case reflect.String:
 		w.WriteString(v.String())
 	case reflect.Bool:
@@ -165,10 +118,109 @@ func print(w *bufio.Writer, v reflect.Value, visited map[uintptr]bool) {
 	case reflect.Int, reflect.Int8, reflect.Int16, reflect.Int32, reflect.Int64:
 		fmt.Fprintf(w, "%v", v.Int())
 	case reflect.Uint, reflect.Uint8, reflect.Uint16, reflect.Uint32, reflect.Uint64, reflect.Uintptr:
-		fmt.Fprintf(w, "%v", v.Uint())
+		scratch = strconv.AppendUint(scratch[:0], v.Uint(), 10)
+		w.Write(scratch)
 	case reflect.Float32, reflect.Float64:
 		fmt.Fprintf(w, "%v", v.Float())
 	case reflect.Complex64, reflect.Complex128:
 		fmt.Fprintf(w, "%v", v.Complex())
 	}
+	return true
+}
+
+type mapHasher struct {
+	xbuf [sha256.Size]byte // XOR'ed accumulated buffer
+	ebuf [sha256.Size]byte // scratch buffer
+	s256 hash.Hash         // sha256 hash.Hash
+	bw   *bufio.Writer     // to hasher into ebuf
+	val  valueCache        // re-usable values for map iteration
+	iter *reflect.MapIter  // re-usable map iterator
+}
+
+func (mh *mapHasher) Reset() {
+	for i := range mh.xbuf {
+		mh.xbuf[i] = 0
+	}
+}
+
+func (mh *mapHasher) startEntry() {
+	for i := range mh.ebuf {
+		mh.ebuf[i] = 0
+	}
+	mh.bw.Flush()
+	mh.s256.Reset()
+}
+
+func (mh *mapHasher) endEntry() {
+	mh.bw.Flush()
+	for i, b := range mh.s256.Sum(mh.ebuf[:0]) {
+		mh.xbuf[i] ^= b
+	}
+}
+
+var mapHasherPool = &sync.Pool{
+	New: func() interface{} {
+		mh := new(mapHasher)
+		mh.s256 = sha256.New()
+		mh.bw = bufio.NewWriter(mh.s256)
+		mh.val = make(valueCache)
+		mh.iter = new(reflect.MapIter)
+		return mh
+	},
+}
+
+type valueCache map[reflect.Type]reflect.Value
+
+func (c valueCache) get(t reflect.Type) reflect.Value {
+	v, ok := c[t]
+	if !ok {
+		v = reflect.New(t).Elem()
+		c[t] = v
+	}
+	return v
+}
+
+// hashMapAcyclic is the faster sort-free version of map hashing. If
+// it detects a cycle it returns false and guarantees that nothing was
+// written to w.
+func hashMapAcyclic(w *bufio.Writer, v reflect.Value, visited map[uintptr]bool, scratch []byte) (acyclic bool) {
+	mh := mapHasherPool.Get().(*mapHasher)
+	defer mapHasherPool.Put(mh)
+	mh.Reset()
+	iter := mapIter(mh.iter, v)
+	defer mapIter(mh.iter, reflect.Value{}) // avoid pinning v from mh.iter when we return
+	k := mh.val.get(v.Type().Key())
+	e := mh.val.get(v.Type().Elem())
+	for iter.Next() {
+		key := iterKey(iter, k)
+		val := iterVal(iter, e)
+		mh.startEntry()
+		if !print(mh.bw, key, visited, scratch) {
+			return false
+		}
+		if !print(mh.bw, val, visited, scratch) {
+			return false
+		}
+		mh.endEntry()
+	}
+	w.Write(mh.xbuf[:])
+	return true
+}
+
+func hashMapFallback(w *bufio.Writer, v reflect.Value, visited map[uintptr]bool, scratch []byte) (acyclic bool) {
+	acyclic = true
+	sm := newSortedMap(v)
+	fmt.Fprintf(w, "map[%d]{\n", len(sm.Key))
+	for i, k := range sm.Key {
+		if !print(w, k, visited, scratch) {
+			acyclic = false
+		}
+		w.WriteString(": ")
+		if !print(w, sm.Value[i], visited, scratch) {
+			acyclic = false
+		}
+		w.WriteString("\n")
+	}
+	w.WriteString("}\n")
+	return acyclic
 }

internal/deephash/deephash_test.go

@@ -5,6 +5,10 @@
 package deephash
 
 import (
+	"bufio"
+	"bytes"
+	"fmt"
+	"reflect"
 	"testing"
 
 	"inet.af/netaddr"
@@ -14,15 +18,15 @@ import (
 	"tailscale.com/wgengine/wgcfg"
 )
 
-func TestDeepPrint(t *testing.T) {
+func TestDeepHash(t *testing.T) {
 	// v contains the types of values we care about for our current callers.
 	// Mostly we're just testing that we don't panic on handled types.
 	v := getVal()
 
-	hash1 := Hash(v)
+	hash1 := calcHash(v)
 	t.Logf("hash: %v", hash1)
 	for i := 0; i < 20; i++ {
-		hash2 := Hash(getVal())
+		hash2 := calcHash(getVal())
 		if hash1 != hash2 {
 			t.Error("second hash didn't match")
 		}
@@ -51,14 +55,23 @@ func getVal() []interface{} {
 		map[dnsname.FQDN][]netaddr.IP{
 			dnsname.FQDN("a."): {netaddr.MustParseIP("1.2.3.4"), netaddr.MustParseIP("4.3.2.1")},
 			dnsname.FQDN("b."): {netaddr.MustParseIP("8.8.8.8"), netaddr.MustParseIP("9.9.9.9")},
+			dnsname.FQDN("c."): {netaddr.MustParseIP("6.6.6.6"), netaddr.MustParseIP("7.7.7.7")},
+			dnsname.FQDN("d."): {netaddr.MustParseIP("6.7.6.6"), netaddr.MustParseIP("7.7.7.8")},
+			dnsname.FQDN("e."): {netaddr.MustParseIP("6.8.6.6"), netaddr.MustParseIP("7.7.7.9")},
+			dnsname.FQDN("f."): {netaddr.MustParseIP("6.9.6.6"), netaddr.MustParseIP("7.7.7.0")},
 		},
 		map[dnsname.FQDN][]netaddr.IPPort{
 			dnsname.FQDN("a."): {netaddr.MustParseIPPort("1.2.3.4:11"), netaddr.MustParseIPPort("4.3.2.1:22")},
 			dnsname.FQDN("b."): {netaddr.MustParseIPPort("8.8.8.8:11"), netaddr.MustParseIPPort("9.9.9.9:22")},
+			dnsname.FQDN("c."): {netaddr.MustParseIPPort("8.8.8.8:12"), netaddr.MustParseIPPort("9.9.9.9:23")},
+			dnsname.FQDN("d."): {netaddr.MustParseIPPort("8.8.8.8:13"), netaddr.MustParseIPPort("9.9.9.9:24")},
+			dnsname.FQDN("e."): {netaddr.MustParseIPPort("8.8.8.8:14"), netaddr.MustParseIPPort("9.9.9.9:25")},
 		},
 		map[tailcfg.DiscoKey]bool{
 			{1: 1}: true,
 			{1: 2}: false,
+			{2: 3}: true,
+			{3: 4}: false,
 		},
 	}
 }
@@ -67,6 +80,57 @@ func BenchmarkHash(b *testing.B) {
 	b.ReportAllocs()
 	v := getVal()
 	for i := 0; i < b.N; i++ {
-		Hash(v)
+		calcHash(v)
+	}
+}
+
+func TestHashMapAcyclic(t *testing.T) {
+	m := map[int]string{}
+	for i := 0; i < 100; i++ {
+		m[i] = fmt.Sprint(i)
+	}
+	got := map[string]bool{}
+
+	var buf bytes.Buffer
+	bw := bufio.NewWriter(&buf)
+
+	for i := 0; i < 20; i++ {
+		visited := map[uintptr]bool{}
+		scratch := make([]byte, 0, 64)
+		v := reflect.ValueOf(m)
+		buf.Reset()
+		bw.Reset(&buf)
+		if !hashMapAcyclic(bw, v, visited, scratch) {
+			t.Fatal("returned false")
+		}
+		if got[string(buf.Bytes())] {
+			continue
+		}
+		got[string(buf.Bytes())] = true
+	}
+	if len(got) != 1 {
+		t.Errorf("got %d results; want 1", len(got))
+	}
+}
+
+func BenchmarkHashMapAcyclic(b *testing.B) {
+	b.ReportAllocs()
+	m := map[int]string{}
+	for i := 0; i < 100; i++ {
+		m[i] = fmt.Sprint(i)
+	}
+
+	var buf bytes.Buffer
+	bw := bufio.NewWriter(&buf)
+	visited := map[uintptr]bool{}
+	scratch := make([]byte, 0, 64)
+	v := reflect.ValueOf(m)
+
+	for i := 0; i < b.N; i++ {
+		buf.Reset()
+		bw.Reset(&buf)
+		if !hashMapAcyclic(bw, v, visited, scratch) {
+			b.Fatal("returned false")
+		}
 	}
 }

internal/deephash/mapiter.go

@@ -0,0 +1,37 @@
+// 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 !tailscale_go
+
+package deephash
+
+import "reflect"
+
+// iterKey returns the current iter key.
+// scratch is a re-usable reflect.Value.
+// iterKey may store the iter key in scratch and return scratch,
+// or it may allocate and return a new reflect.Value.
+func iterKey(iter *reflect.MapIter, _ reflect.Value) reflect.Value {
+	return iter.Key()
+}
+
+// iterVal returns the current iter val.
+// scratch is a re-usable reflect.Value.
+// iterVal may store the iter val in scratch and return scratch,
+// or it may allocate and return a new reflect.Value.
+func iterVal(iter *reflect.MapIter, _ reflect.Value) reflect.Value {
+	return iter.Value()
+}
+
+// mapIter returns a map iterator for mapVal.
+// scratch is a re-usable reflect.MapIter.
+// mapIter may re-use scratch and return it,
+// or it may allocate and return a new *reflect.MapIter.
+// If mapVal is the zero reflect.Value, mapIter may return nil.
+func mapIter(_ *reflect.MapIter, mapVal reflect.Value) *reflect.MapIter {
+	if !mapVal.IsValid() {
+		return nil
+	}
+	return mapVal.MapRange()
+}

internal/deephash/mapiter_future.go

@@ -0,0 +1,42 @@
+// 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 tailscale_go
+
+package deephash
+
+import "reflect"
+
+// iterKey returns the current iter key.
+// scratch is a re-usable reflect.Value.
+// iterKey may store the iter key in scratch and return scratch,
+// or it may allocate and return a new reflect.Value.
+func iterKey(iter *reflect.MapIter, scratch reflect.Value) reflect.Value {
+	iter.SetKey(scratch)
+	return scratch
+}
+
+// iterVal returns the current iter val.
+// scratch is a re-usable reflect.Value.
+// iterVal may store the iter val in scratch and return scratch,
+// or it may allocate and return a new reflect.Value.
+func iterVal(iter *reflect.MapIter, scratch reflect.Value) reflect.Value {
+	iter.SetValue(scratch)
+	return scratch
+}
+
+// mapIter returns a map iterator for mapVal.
+// scratch is a re-usable reflect.MapIter.
+// mapIter may re-use scratch and return it,
+// or it may allocate and return a new *reflect.MapIter.
+// If mapVal is the zero reflect.Value, mapIter may return nil.
+func mapIter(scratch *reflect.MapIter, mapVal reflect.Value) *reflect.MapIter {
+	scratch.Reset(mapVal) // always Reset, to allow the caller to avoid pinning memory
+	if !mapVal.IsValid() {
+		// Returning scratch would also be OK.
+		// Do this for consistency with the non-optimized version.
+		return nil
+	}
+	return scratch
+}

ipn/ipnlocal/local.go

@@ -44,11 +44,13 @@ import (
 	"tailscale.com/types/logger"
 	"tailscale.com/types/netmap"
 	"tailscale.com/types/persist"
+	"tailscale.com/types/preftype"
 	"tailscale.com/types/wgkey"
 	"tailscale.com/util/dnsname"
 	"tailscale.com/util/osshare"
 	"tailscale.com/util/systemd"
 	"tailscale.com/version"
+	"tailscale.com/version/distro"
 	"tailscale.com/wgengine"
 	"tailscale.com/wgengine/filter"
 	"tailscale.com/wgengine/router"
@@ -242,7 +244,7 @@ func (b *LocalBackend) linkChange(major bool, ifst *interfaces.State) {
 	// need updating to tweak default routes.
 	b.updateFilter(b.netMap, b.prefs)
 
-	if runtime.GOOS == "windows" && b.netMap != nil && b.state == ipn.Running {
+	if peerAPIListenAsync && b.netMap != nil && b.state == ipn.Running {
 		want := len(b.netMap.Addresses)
 		b.logf("linkChange: peerAPIListeners too low; trying again")
 		if len(b.peerAPIListeners) < want {
@@ -323,6 +325,7 @@ func (b *LocalBackend) updateStatus(sb *ipnstate.StatusBuilder, extraLocked func
 		s.AuthURL = b.authURLSticky
 		if b.netMap != nil {
 			s.MagicDNSSuffix = b.netMap.MagicDNSSuffix()
+			s.CertDomains = append([]string(nil), b.netMap.DNS.CertDomains...)
 		}
 	})
 	sb.MutateSelfStatus(func(ss *ipnstate.PeerStatus) {
@@ -451,6 +454,13 @@ func (b *LocalBackend) setClientStatus(st controlclient.Status) {
 	// Lock b once and do only the things that require locking.
 	b.mu.Lock()
 
+	if st.LogoutFinished != nil {
+		// Since we're logged out now, our netmap cache is invalid.
+		// Since st.NetMap==nil means "netmap is unchanged", there is
+		// no other way to represent this change.
+		b.setNetMapLocked(nil)
+	}
+
 	prefs := b.prefs
 	stateKey := b.stateKey
 	netMap := b.netMap
@@ -648,6 +658,12 @@ func (b *LocalBackend) getNewControlClientFunc() clientGen {
 // startIsNoopLocked reports whether a Start call on this LocalBackend
 // with the provided Start Options would be a useless no-op.
 //
+// TODO(apenwarr): we shouldn't need this.
+//  The state machine is now nearly clean enough where it can accept a new
+//  connection while in any state, not just Running, and on any platform.
+//  We'd want to add a few more tests to state_test.go to ensure this continues
+//  to work as expected.
+//
 // b.mu must be held.
 func (b *LocalBackend) startIsNoopLocked(opts ipn.Options) bool {
 	// Options has 5 fields; check all of them:
@@ -701,6 +717,7 @@ func (b *LocalBackend) Start(opts ipn.Options) error {
 		b.send(ipn.Notify{
 			State:         &state,
 			NetMap:        nm,
+			Prefs:         b.prefs,
 			LoginFinished: new(empty.Message),
 		})
 		return nil
@@ -913,8 +930,8 @@ func (b *LocalBackend) updateFilter(netMap *netmap.NetworkMap, prefs *ipn.Prefs)
 			}
 		}
 	}
-	localNets := localNetsB.IPSet()
-	logNets := logNetsB.IPSet()
+	localNets, _ := localNetsB.IPSet()
+	logNets, _ := logNetsB.IPSet()
 
 	changed := deephash.UpdateHash(&b.filterHash, haveNetmap, addrs, packetFilter, localNets.Ranges(), logNets.Ranges(), shieldsUp)
 	if !changed {
@@ -971,7 +988,8 @@ func interfaceRoutes() (ips *netaddr.IPSet, hostIPs []netaddr.IP, err error) {
 		return nil, nil, err
 	}
 
-	return b.IPSet(), hostIPs, nil
+	ipSet, _ := b.IPSet()
+	return ipSet, hostIPs, nil
 }
 
 // shrinkDefaultRoute returns an IPSet representing the IPs in route,
@@ -1002,7 +1020,7 @@ func shrinkDefaultRoute(route netaddr.IPPrefix) (*netaddr.IPSet, error) {
 	for _, pfx := range removeFromDefaultRoute {
 		b.RemovePrefix(pfx)
 	}
-	return b.IPSet(), nil
+	return b.IPSet()
 }
 
 // dnsCIDRsEqual determines whether two CIDR lists are equal
@@ -1694,9 +1712,64 @@ func (b *LocalBackend) authReconfig() {
 
 	rcfg := b.routerConfig(cfg, uc)
 
-	var dcfg dns.Config
+	dcfg := dns.Config{
+		Routes: map[dnsname.FQDN][]netaddr.IPPort{},
+		Hosts:  map[dnsname.FQDN][]netaddr.IP{},
+	}
+
+	// Populate MagicDNS records. We do this unconditionally so that
+	// quad-100 can always respond to MagicDNS queries, even if the OS
+	// isn't configured to make MagicDNS resolution truly
+	// magic. Details in
+	// https://github.com/tailscale/tailscale/issues/1886.
+	set := func(name string, addrs []netaddr.IPPrefix) {
+		if len(addrs) == 0 || name == "" {
+			return
+		}
+		fqdn, err := dnsname.ToFQDN(name)
+		if err != nil {
+			return // TODO: propagate error?
+		}
+		var ips []netaddr.IP
+		for _, addr := range addrs {
+			// Remove IPv6 addresses for now, as we don't
+			// guarantee that the peer node actually can speak
+			// IPv6 correctly.
+			//
+			// https://github.com/tailscale/tailscale/issues/1152
+			// tracks adding the right capability reporting to
+			// enable AAAA in MagicDNS.
+			if addr.IP().Is6() {
+				continue
+			}
+			ips = append(ips, addr.IP())
+		}
+		dcfg.Hosts[fqdn] = ips
+	}
+	set(nm.Name, nm.Addresses)
+	for _, peer := range nm.Peers {
+		set(peer.Name, peer.Addresses)
+	}
+	for _, rec := range nm.DNS.ExtraRecords {
+		switch rec.Type {
+		case "", "A", "AAAA":
+			// Treat these all the same for now: infer from the value
+		default:
+			// TODO: more
+			continue
+		}
+		ip, err := netaddr.ParseIP(rec.Value)
+		if err != nil {
+			// Ignore.
+			continue
+		}
+		fqdn, err := dnsname.ToFQDN(rec.Name)
+		if err != nil {
+			continue
+		}
+		dcfg.Hosts[fqdn] = append(dcfg.Hosts[fqdn], ip)
+	}
 
-	// If CorpDNS is false, dcfg remains the zero value.
 	if uc.CorpDNS {
 		addDefault := func(resolvers []tailcfg.DNSResolver) {
 			for _, resolver := range resolvers {
@@ -1710,9 +1783,6 @@ func (b *LocalBackend) authReconfig() {
 		}
 
 		addDefault(nm.DNS.Resolvers)
-		if len(nm.DNS.Routes) > 0 {
-			dcfg.Routes = map[dnsname.FQDN][]netaddr.IPPort{}
-		}
 		for suffix, resolvers := range nm.DNS.Routes {
 			fqdn, err := dnsname.ToFQDN(suffix)
 			if err != nil {
@@ -1734,36 +1804,9 @@ func (b *LocalBackend) authReconfig() {
 			}
 			dcfg.SearchDomains = append(dcfg.SearchDomains, fqdn)
 		}
-		set := func(name string, addrs []netaddr.IPPrefix) {
-			if len(addrs) == 0 || name == "" {
-				return
-			}
-			fqdn, err := dnsname.ToFQDN(name)
-			if err != nil {
-				return // TODO: propagate error?
-			}
-			var ips []netaddr.IP
-			for _, addr := range addrs {
-				// Remove IPv6 addresses for now, as we don't
-				// guarantee that the peer node actually can speak
-				// IPv6 correctly.
-				//
-				// https://github.com/tailscale/tailscale/issues/1152
-				// tracks adding the right capability reporting to
-				// enable AAAA in MagicDNS.
-				if addr.IP().Is6() {
-					continue
-				}
-				ips = append(ips, addr.IP())
-			}
-			dcfg.Hosts[fqdn] = ips
-		}
 		if nm.DNS.Proxied { // actually means "enable MagicDNS"
-			dcfg.AuthoritativeSuffixes = magicDNSRootDomains(nm)
-			dcfg.Hosts = map[dnsname.FQDN][]netaddr.IP{}
-			set(nm.Name, nm.Addresses)
-			for _, peer := range nm.Peers {
-				set(peer.Name, peer.Addresses)
+			for _, dom := range magicDNSRootDomains(nm) {
+				dcfg.Routes[dom] = nil // resolve internally with dcfg.Hosts
 			}
 		}
 
@@ -1787,7 +1830,7 @@ func (b *LocalBackend) authReconfig() {
 			//
 			// https://github.com/tailscale/tailscale/issues/1713
 			addDefault(nm.DNS.FallbackResolvers)
-		case len(dcfg.Routes) == 0 && len(dcfg.Hosts) == 0 && len(dcfg.AuthoritativeSuffixes) == 0:
+		case len(dcfg.Routes) == 0:
 			// No settings requiring split DNS, no problem.
 		case version.OS() == "android":
 			// We don't support split DNS at all on Android yet.
@@ -1795,7 +1838,7 @@ func (b *LocalBackend) authReconfig() {
 		}
 	}
 
-	err = b.e.Reconfig(cfg, rcfg, &dcfg)
+	err = b.e.Reconfig(cfg, rcfg, &dcfg, nm.Debug)
 	if err == wgengine.ErrNoChanges {
 		return
 	}
@@ -1815,8 +1858,9 @@ func parseResolver(cfg tailcfg.DNSResolver) (netaddr.IPPort, error) {
 // tailscaleVarRoot returns the root directory of Tailscale's writable
 // storage area. (e.g. "/var/lib/tailscale")
 func tailscaleVarRoot() string {
-	if runtime.GOOS == "ios" {
-		dir, _ := paths.IOSSharedDir.Load().(string)
+	switch runtime.GOOS {
+	case "ios", "android":
+		dir, _ := paths.AppSharedDir.Load().(string)
 		return dir
 	}
 	stateFile := paths.DefaultTailscaledStateFile()
@@ -1860,10 +1904,26 @@ func (b *LocalBackend) closePeerAPIListenersLocked() {
 	b.peerAPIListeners = nil
 }
 
+// peerAPIListenAsync is whether the operating system requires that we
+// retry listening on the peerAPI ip/port for whatever reason.
+//
+// On Windows, see Issue 1620.
+// On Android, see Issue 1960.
+const peerAPIListenAsync = runtime.GOOS == "windows" || runtime.GOOS == "android"
+
 func (b *LocalBackend) initPeerAPIListener() {
 	b.mu.Lock()
 	defer b.mu.Unlock()
 
+	if b.netMap == nil {
+		// We're called from authReconfig which checks that
+		// netMap is non-nil, but if a concurrent Logout,
+		// ResetForClientDisconnect, or Start happens when its
+		// mutex was released, the netMap could be
+		// nil'ed out (Issue 1996). Bail out early here if so.
+		return
+	}
+
 	if len(b.netMap.Addresses) == len(b.peerAPIListeners) {
 		allSame := true
 		for i, pln := range b.peerAPIListeners {
@@ -1914,13 +1974,12 @@ func (b *LocalBackend) initPeerAPIListener() {
 		if !skipListen {
 			ln, err = ps.listen(a.IP(), b.prevIfState)
 			if err != nil {
-				if runtime.GOOS == "windows" {
-					// Expected for now. See Issue 1620.
-					// But we fix it later in linkChange
+				if peerAPIListenAsync {
+					// Expected. But we fix it later in linkChange
 					// ("peerAPIListeners too low").
 					continue
 				}
-				b.logf("[unexpected] peerapi listen(%q) error: %v", a.IP, err)
+				b.logf("[unexpected] peerapi listen(%q) error: %v", a.IP(), err)
 				continue
 			}
 		}
@@ -2019,6 +2078,11 @@ func (b *LocalBackend) routerConfig(cfg *wgcfg.Config, prefs *ipn.Prefs) *router
 		Routes:           peerRoutes(cfg.Peers, 10_000),
 	}
 
+	if distro.Get() == distro.Synology {
+		// Issue 1995: we don't use iptables on Synology.
+		rs.NetfilterMode = preftype.NetfilterOff
+	}
+
 	// Sanity check: we expect the control server to program both a v4
 	// and a v6 default route, if default routing is on. Fill in
 	// blackhole routes appropriately if we're missing some. This is
@@ -2044,7 +2108,7 @@ func (b *LocalBackend) routerConfig(cfg *wgcfg.Config, prefs *ipn.Prefs) *router
 		if !default6 {
 			rs.Routes = append(rs.Routes, ipv6Default)
 		}
-		if runtime.GOOS == "linux" {
+		if runtime.GOOS == "linux" || runtime.GOOS == "darwin" {
 			// Only allow local lan access on linux machines for now.
 			ips, _, err := interfaceRoutes()
 			if err != nil {
@@ -2135,7 +2199,7 @@ func (b *LocalBackend) enterState(newState ipn.State) {
 		b.blockEngineUpdates(true)
 		fallthrough
 	case ipn.Stopped:
-		err := b.e.Reconfig(&wgcfg.Config{}, &router.Config{}, &dns.Config{})
+		err := b.e.Reconfig(&wgcfg.Config{}, &router.Config{}, &dns.Config{}, nil)
 		if err != nil {
 			b.logf("Reconfig(down): %v", err)
 		}
@@ -2255,7 +2319,7 @@ func (b *LocalBackend) stateMachine() {
 // a status update that predates the "I've shut down" update.
 func (b *LocalBackend) stopEngineAndWait() {
 	b.logf("stopEngineAndWait...")
-	b.e.Reconfig(&wgcfg.Config{}, &router.Config{}, &dns.Config{})
+	b.e.Reconfig(&wgcfg.Config{}, &router.Config{}, &dns.Config{}, nil)
 	b.requestEngineStatusAndWait()
 	b.logf("stopEngineAndWait: done.")
 }
@@ -2312,7 +2376,6 @@ func (b *LocalBackend) LogoutSync(ctx context.Context) error {
 func (b *LocalBackend) logout(ctx context.Context, sync bool) error {
 	b.mu.Lock()
 	cc := b.cc
-	b.setNetMapLocked(nil)
 	b.mu.Unlock()
 
 	b.EditPrefs(&ipn.MaskedPrefs{
@@ -2339,10 +2402,6 @@ func (b *LocalBackend) logout(ctx context.Context, sync bool) error {
 		cc.StartLogout()
 	}
 
-	b.mu.Lock()
-	b.setNetMapLocked(nil)
-	b.mu.Unlock()
-
 	b.stateMachine()
 	return err
 }
@@ -2544,6 +2603,42 @@ func (b *LocalBackend) FileTargets() ([]*apitype.FileTarget, error) {
 	return ret, nil
 }
 
+// SetDNS adds a DNS record for the given domain name & TXT record
+// value.
+//
+// It's meant for use with dns-01 ACME (LetsEncrypt) challenges.
+//
+// This is the low-level interface. Other layers will provide more
+// friendly options to get HTTPS certs.
+func (b *LocalBackend) SetDNS(ctx context.Context, name, value string) error {
+	req := &tailcfg.SetDNSRequest{
+		Version: 1,
+		Type:    "TXT",
+		Name:    name,
+		Value:   value,
+	}
+
+	b.mu.Lock()
+	cc := b.cc
+	if prefs := b.prefs; prefs != nil {
+		req.NodeKey = tailcfg.NodeKey(prefs.Persist.PrivateNodeKey.Public())
+	}
+	b.mu.Unlock()
+	if cc == nil {
+		return errors.New("not connected")
+	}
+	if req.NodeKey.IsZero() {
+		return errors.New("no nodekey")
+	}
+	if name == "" {
+		return errors.New("missing 'name'")
+	}
+	if value == "" {
+		return errors.New("missing 'value'")
+	}
+	return cc.SetDNS(ctx, req)
+}
+
 func (b *LocalBackend) registerIncomingFile(inf *incomingFile, active bool) {
 	b.mu.Lock()
 	defer b.mu.Unlock()
@@ -2616,7 +2711,6 @@ func (b *LocalBackend) CheckIPForwarding() error {
 		return nil
 	}
 	if isBSD(runtime.GOOS) {
-		//lint:ignore ST1005 output to users as is
 		return fmt.Errorf("Subnet routing and exit nodes only work with additional manual configuration on %v, and is not currently officially supported.", runtime.GOOS)
 	}
 
@@ -2633,16 +2727,13 @@ func (b *LocalBackend) CheckIPForwarding() error {
 	for _, key := range keys {
 		bs, err := exec.Command("sysctl", "-n", key).Output()
 		if err != nil {
-			//lint:ignore ST1005 output to users as is
 			return fmt.Errorf("couldn't check %s (%v).\nSubnet routes won't work without IP forwarding.", key, err)
 		}
 		on, err := strconv.ParseBool(string(bytes.TrimSpace(bs)))
 		if err != nil {
-			//lint:ignore ST1005 output to users as is
 			return fmt.Errorf("couldn't parse %s (%v).\nSubnet routes won't work without IP forwarding.", key, err)
 		}
 		if !on {
-			//lint:ignore ST1005 output to users as is
 			return fmt.Errorf("%s is disabled. Subnet routes won't work.", key)
 		}
 	}

ipn/ipnlocal/peerapi_macios_ext.go

@@ -2,21 +2,19 @@
 // Use of this source code is governed by a BSD-style
 // license that can be found in the LICENSE file.
 
-// +build darwin,redo ios,redo
+// +build darwin,ts_macext ios,ts_macext
 
 package ipnlocal
 
 import (
 	"errors"
 	"fmt"
-	"log"
 	"net"
-	"strings"
 	"syscall"
 
-	"golang.org/x/sys/unix"
 	"inet.af/netaddr"
 	"tailscale.com/net/interfaces"
+	"tailscale.com/net/netns"
 )
 
 func init() {
@@ -32,29 +30,7 @@ func initListenConfigNetworkExtension(nc *net.ListenConfig, ip netaddr.IP, st *i
 	if !ok {
 		return fmt.Errorf("no interface with name %q", tunIfName)
 	}
-	nc.Control = func(network, address string, c syscall.RawConn) error {
-		var sockErr error
-		err := c.Control(func(fd uintptr) {
-			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 {
-			return err
-		}
-		return sockErr
-	}
-	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)
+	return netns.SetListenConfigInterfaceIndex(nc, tunIf.Index)
 }
 
 func peerDialControlFuncNetworkExtension(b *LocalBackend) func(network, address string, c syscall.RawConn) error {
@@ -68,17 +44,12 @@ func peerDialControlFuncNetworkExtension(b *LocalBackend) func(network, address
 			index = tunIf.Index
 		}
 	}
+	var lc net.ListenConfig
+	netns.SetListenConfigInterfaceIndex(&lc, 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
+		return lc.Control(network, address, c)
 	}
 }

ipn/ipnlocal/state_test.go

@@ -140,6 +140,8 @@ func (cc *mockControl) send(err error, url string, loginFinished bool, nm *netma
 		}
 		if loginFinished {
 			s.LoginFinished = &empty.Message{}
+		} else if url == "" && err == nil && nm == nil {
+			s.LogoutFinished = &empty.Message{}
 		}
 		cc.statusFunc(s)
 	}
@@ -246,6 +248,10 @@ func (cc *mockControl) UpdateEndpoints(localPort uint16, endpoints []tailcfg.End
 	cc.called("UpdateEndpoints")
 }
 
+func (*mockControl) SetDNS(context.Context, *tailcfg.SetDNSRequest) error {
+	panic("unexpected SetDNS call")
+}
+
 // A very precise test of the sequence of function calls generated by
 // ipnlocal.Local into its controlclient instance, and the events it
 // produces upstream into the UI.
@@ -548,10 +554,7 @@ func TestStateMachine(t *testing.T) {
 		c.Assert(nn[0].State, qt.Not(qt.IsNil))
 		c.Assert(nn[0].LoginFinished, qt.Not(qt.IsNil))
 		c.Assert(nn[0].NetMap, qt.Not(qt.IsNil))
-		// BUG: Prefs should be sent too, or the UI could end up in
-		// a bad state. (iOS, the only current user of this feature,
-		// probably wouldn't notice because it happens to not display
-		// any prefs. Maybe exit nodes will look weird?)
+		c.Assert(nn[0].Prefs, qt.Not(qt.IsNil))
 	}
 
 	// undo the state hack above.
@@ -563,24 +566,25 @@ func TestStateMachine(t *testing.T) {
 	b.Logout()
 	{
 		nn := notifies.drain(2)
-		// BUG: now is not the time to unpause.
-		c.Assert([]string{"unpause", "StartLogout"}, qt.DeepEquals, cc.getCalls())
+		c.Assert([]string{"pause", "StartLogout"}, qt.DeepEquals, cc.getCalls())
 		c.Assert(nn[0].State, qt.Not(qt.IsNil))
 		c.Assert(nn[1].Prefs, qt.Not(qt.IsNil))
-		c.Assert(ipn.NeedsLogin, qt.Equals, *nn[0].State)
+		c.Assert(ipn.Stopped, qt.Equals, *nn[0].State)
 		c.Assert(nn[1].Prefs.LoggedOut, qt.IsTrue)
 		c.Assert(nn[1].Prefs.WantRunning, qt.IsFalse)
-		c.Assert(ipn.NeedsLogin, qt.Equals, b.State())
+		c.Assert(ipn.Stopped, qt.Equals, b.State())
 	}
 
 	// Let's make the logout succeed.
 	t.Logf("\n\nLogout (async) - succeed")
-	notifies.expect(0)
+	notifies.expect(1)
 	cc.setAuthBlocked(true)
 	cc.send(nil, "", false, nil)
 	{
-		notifies.drain(0)
-		c.Assert(cc.getCalls(), qt.HasLen, 0)
+		nn := notifies.drain(1)
+		c.Assert([]string{"unpause"}, qt.DeepEquals, cc.getCalls())
+		c.Assert(nn[0].State, qt.Not(qt.IsNil))
+		c.Assert(ipn.NeedsLogin, qt.Equals, *nn[0].State)
 		c.Assert(b.Prefs().LoggedOut, qt.IsTrue)
 		c.Assert(b.Prefs().WantRunning, qt.IsFalse)
 		c.Assert(ipn.NeedsLogin, qt.Equals, b.State())

ipn/ipnserver/server.go

@@ -24,7 +24,6 @@ import (
 	"strconv"
 	"strings"
 	"sync"
-	"sync/atomic"
 	"syscall"
 	"time"
 
@@ -41,9 +40,11 @@ import (
 	"tailscale.com/safesocket"
 	"tailscale.com/smallzstd"
 	"tailscale.com/types/logger"
+	"tailscale.com/util/groupmember"
 	"tailscale.com/util/pidowner"
 	"tailscale.com/util/systemd"
 	"tailscale.com/version"
+	"tailscale.com/version/distro"
 	"tailscale.com/wgengine"
 )
 
@@ -347,51 +348,32 @@ func isReadonlyConn(ci connIdentity, operatorUID string, logf logger.Logf) bool
 		logf("connection from userid %v; is configured operator", uid)
 		return rw
 	}
-	var adminGroupID string
-	switch runtime.GOOS {
-	case "darwin":
-		adminGroupID = darwinAdminGroupID()
-	default:
-		logf("connection from userid %v; read-only", uid)
+	if yes, err := isLocalAdmin(uid); err != nil {
+		logf("connection from userid %v; read-only; %v", uid, err)
 		return ro
-	}
-	if adminGroupID == "" {
-		logf("connection from userid %v; no system admin group found, read-only", uid)
-		return ro
-	}
-	u, err := user.LookupId(uid)
-	if err != nil {
-		logf("connection from userid %v; failed to look up user; read-only", uid)
-		return ro
-	}
-	gids, err := u.GroupIds()
-	if err != nil {
-		logf("connection from userid %v; failed to look up groups; read-only", uid)
-		return ro
-	}
-	for _, gid := range gids {
-		if gid == adminGroupID {
-			logf("connection from userid %v; is local admin, has access", uid)
-			return rw
-		}
+	} else if yes {
+		logf("connection from userid %v; is local admin, has access", uid)
+		return rw
 	}
 	logf("connection from userid %v; read-only", uid)
 	return ro
 }
 
-var darwinAdminGroupIDCache atomic.Value // of string
-
-func darwinAdminGroupID() string {
-	s, _ := darwinAdminGroupIDCache.Load().(string)
-	if s != "" {
-		return s
-	}
-	g, err := user.LookupGroup("admin")
+func isLocalAdmin(uid string) (bool, error) {
+	u, err := user.LookupId(uid)
 	if err != nil {
-		return ""
+		return false, err
 	}
-	darwinAdminGroupIDCache.Store(g.Gid)
-	return g.Gid
+	var adminGroup string
+	switch {
+	case runtime.GOOS == "darwin":
+		adminGroup = "admin"
+	case distro.Get() == distro.QNAP:
+		adminGroup = "administrators"
+	default:
+		return false, fmt.Errorf("no system admin group found")
+	}
+	return groupmember.IsMemberOfGroup(adminGroup, u.Username)
 }
 
 // inUseOtherUserError is the error type for when the server is in use
@@ -415,12 +397,10 @@ func (s *server) checkConnIdentityLocked(ci connIdentity) error {
 			break
 		}
 		if ci.UserID != active.UserID {
-			//lint:ignore ST1005 we want to capitalize Tailscale here
 			return inUseOtherUserError{fmt.Errorf("Tailscale already in use by %s, pid %d", active.User.Username, active.Pid)}
 		}
 	}
 	if su := s.serverModeUser; su != nil && ci.UserID != su.Uid {
-		//lint:ignore ST1005 we want to capitalize Tailscale here
 		return inUseOtherUserError{fmt.Errorf("Tailscale already in use by %s", su.Username)}
 	}
 	return nil

ipn/ipnstate/ipnstate.go

@@ -45,6 +45,13 @@ type Status struct {
 	// has MagicDNS enabled.
 	MagicDNSSuffix string
 
+	// CertDomains are the set of DNS names for which the control
+	// plane server will assist with provisioning TLS
+	// certificates. See SetDNSRequest for dns-01 ACME challenges
+	// for e.g. LetsEncrypt. These names are FQDNs without
+	// trailing periods, and without any "_acme-challenge." prefix.
+	CertDomains []string
+
 	Peer map[key.Public]*PeerStatus
 	User map[tailcfg.UserID]tailcfg.UserProfile
 }

ipn/localapi/localapi.go

@@ -100,6 +100,8 @@ func (h *Handler) ServeHTTP(w http.ResponseWriter, r *http.Request) {
 		h.serveBugReport(w, r)
 	case "/localapi/v0/file-targets":
 		h.serveFileTargets(w, r)
+	case "/localapi/v0/set-dns":
+		h.serveSetDNS(w, r)
 	case "/":
 		io.WriteString(w, "tailscaled\n")
 	default:
@@ -382,6 +384,25 @@ func (h *Handler) serveFilePut(w http.ResponseWriter, r *http.Request) {
 	rp.ServeHTTP(w, outReq)
 }
 
+func (h *Handler) serveSetDNS(w http.ResponseWriter, r *http.Request) {
+	if !h.PermitWrite {
+		http.Error(w, "access denied", http.StatusForbidden)
+		return
+	}
+	if r.Method != "POST" {
+		http.Error(w, "want POST", 400)
+		return
+	}
+	ctx := r.Context()
+	err := h.b.SetDNS(ctx, r.FormValue("name"), r.FormValue("value"))
+	if err != nil {
+		writeErrorJSON(w, err)
+		return
+	}
+	w.Header().Set("Content-Type", "application/json")
+	json.NewEncoder(w).Encode(struct{}{})
+}
+
 var dialPeerTransportOnce struct {
 	sync.Once
 	v *http.Transport
@@ -390,7 +411,7 @@ var dialPeerTransportOnce struct {
 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
+		t.Dial = nil
 		dialer := net.Dialer{
 			Timeout:   30 * time.Second,
 			KeepAlive: 30 * time.Second,

ipn/message.go

@@ -104,7 +104,9 @@ func NewBackendServer(logf logger.Logf, b Backend, sendNotifyMsg func(Notify)) *
 		b:             b,
 		sendNotifyMsg: sendNotifyMsg,
 	}
-	if sendNotifyMsg != nil {
+	// b may be nil if the BackendServer is being created just to
+	// encapsulate and send an error message.
+	if sendNotifyMsg != nil && b != nil {
 		b.SetNotifyCallback(bs.send)
 	}
 	return bs

ipn/message_test.go

@@ -187,3 +187,17 @@ func TestClientServer(t *testing.T) {
 	})
 	flushUntil(Running)
 }
+
+func TestNilBackend(t *testing.T) {
+	var called *Notify
+	bs := NewBackendServer(t.Logf, nil, func(n Notify) {
+		called = &n
+	})
+	bs.SendErrorMessage("Danger, Will Robinson!")
+	if called == nil {
+		t.Errorf("expect callback to be called, wasn't")
+	}
+	if called.ErrMessage == nil || *called.ErrMessage != "Danger, Will Robinson!" {
+		t.Errorf("callback got wrong error: %v", called.ErrMessage)
+	}
+}

logtail/filch/filch.go

@@ -15,7 +15,6 @@ import (
 	"sync"
 )
 
-//lint:ignore U1000 work around false positive: https://github.com/dominikh/go-tools/issues/983
 var stderrFD = 2 // a variable for testing
 
 type Options struct {

logtail/logtail.go

@@ -15,6 +15,7 @@ import (
 	"net/http"
 	"os"
 	"strconv"
+	"sync/atomic"
 	"time"
 
 	"tailscale.com/logtail/backoff"
@@ -72,7 +73,7 @@ func NewLogger(cfg Config, logf tslogger.Logf) *Logger {
 	}
 	l := &Logger{
 		stderr:         cfg.Stderr,
-		stderrLevel:    cfg.StderrLevel,
+		stderrLevel:    int64(cfg.StderrLevel),
 		httpc:          cfg.HTTPC,
 		url:            cfg.BaseURL + "/c/" + cfg.Collection + "/" + cfg.PrivateID.String(),
 		lowMem:         cfg.LowMemory,
@@ -103,7 +104,7 @@ func NewLogger(cfg Config, logf tslogger.Logf) *Logger {
 // logging facilities and uploading to a log server.
 type Logger struct {
 	stderr         io.Writer
-	stderrLevel    int
+	stderrLevel    int64 // accessed atomically
 	httpc          *http.Client
 	url            string
 	lowMem         bool
@@ -125,10 +126,8 @@ type Logger struct {
 // SetVerbosityLevel controls the verbosity level that should be
 // written to stderr. 0 is the default (not verbose). Levels 1 or higher
 // are increasingly verbose.
-//
-// It should not be changed concurrently with log writes.
 func (l *Logger) SetVerbosityLevel(level int) {
-	l.stderrLevel = level
+	atomic.StoreInt64(&l.stderrLevel, int64(level))
 }
 
 // SetLinkMonitor sets the optional the link monitor.
@@ -514,7 +513,7 @@ func (l *Logger) Write(buf []byte) (int, error) {
 		return 0, nil
 	}
 	level, buf := parseAndRemoveLogLevel(buf)
-	if l.stderr != nil && l.stderr != ioutil.Discard && level <= l.stderrLevel {
+	if l.stderr != nil && l.stderr != ioutil.Discard && int64(level) <= atomic.LoadInt64(&l.stderrLevel) {
 		if buf[len(buf)-1] == '\n' {
 			l.stderr.Write(buf)
 		} else {

net/dns/config.go

@@ -22,27 +22,26 @@ type Config struct {
 	// for queries that fall within that suffix.
 	// If a query doesn't match any entry in Routes, the
 	// DefaultResolvers are used.
+	// A Routes entry with no resolvers means the route should be
+	// authoritatively answered using the contents of Hosts.
 	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.
+	// Queries matching entries in Hosts are resolved locally by
+	// 100.100.100.100 without leaving the machine.
+	// Adding an entry to Hosts merely creates the record. If you want
+	// it to resolve, you also need to add appropriate routes to
+	// Routes.
 	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
 }
 
 // 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()
+	return c.hasDefaultResolvers() || c.hasRoutes()
 }
 
 func (c Config) hasRoutes() bool {
@@ -52,7 +51,7 @@ func (c Config) hasRoutes() bool {
 // hasDefaultResolversOnly reports whether the only resolvers in c are
 // DefaultResolvers.
 func (c Config) hasDefaultResolversOnly() bool {
-	return c.hasDefaultResolvers() && !c.hasRoutes() && !c.hasHosts()
+	return c.hasDefaultResolvers() && !c.hasRoutes()
 }
 
 func (c Config) hasDefaultResolvers() bool {
@@ -63,44 +62,28 @@ func (c Config) hasDefaultResolvers() bool {
 // routes use the same resolvers, or nil if multiple sets of resolvers
 // are specified.
 func (c Config) singleResolverSet() []netaddr.IPPort {
-	var first []netaddr.IPPort
+	var (
+		prev            []netaddr.IPPort
+		prevInitialized bool
+	)
 	for _, resolvers := range c.Routes {
-		if first == nil {
-			first = resolvers
+		if !prevInitialized {
+			prev = resolvers
+			prevInitialized = true
 			continue
 		}
-		if !sameIPPorts(first, resolvers) {
+		if !sameIPPorts(prev, resolvers) {
 			return nil
 		}
 	}
-	return first
+	return prev
 }
 
-// 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.
+// matchDomains returns the list of match suffixes needed by Routes.
 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
-	}
+	ret := make([]dnsname.FQDN, 0, len(c.Routes))
 	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()

net/dns/direct.go

@@ -2,13 +2,12 @@
 // 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"
+	"crypto/rand"
 	"fmt"
 	"io"
 	"io/ioutil"
@@ -18,7 +17,6 @@ import (
 	"strings"
 
 	"inet.af/netaddr"
-	"tailscale.com/atomicfile"
 	"tailscale.com/util/dnsname"
 )
 
@@ -77,21 +75,17 @@ func readResolv(r io.Reader) (config OSConfig, err error) {
 	return config, nil
 }
 
-func readResolvFile(path string) (OSConfig, error) {
-	var config OSConfig
-
-	f, err := os.Open(path)
+func (m directManager) readResolvFile(path string) (OSConfig, error) {
+	b, err := m.fs.ReadFile(path)
 	if err != nil {
-		return config, err
+		return OSConfig{}, err
 	}
-	defer f.Close()
-
-	return readResolv(f)
+	return readResolv(bytes.NewReader(b))
 }
 
 // readResolvConf reads DNS configuration from /etc/resolv.conf.
-func readResolvConf() (OSConfig, error) {
-	return readResolvFile(resolvConf)
+func (m directManager) readResolvConf() (OSConfig, error) {
+	return m.readResolvFile(resolvConf)
 }
 
 // resolvOwner returns the apparent owner of the resolv.conf
@@ -143,33 +137,39 @@ func isResolvedRunning() bool {
 	return err == nil
 }
 
-// directManager is a managerImpl which replaces /etc/resolv.conf with a file
+// directManager is an OSConfigurator which replaces /etc/resolv.conf with a file
 // generated from the given configuration, creating a backup of its old state.
 //
 // This way of configuring DNS is precarious, since it does not react
 // to the disappearance of the Tailscale interface.
 // The caller must call Down before program shutdown
 // or as cleanup if the program terminates unexpectedly.
-type directManager struct{}
+type directManager struct {
+	fs pinholeFS
+}
 
-func newDirectManager() (directManager, error) {
-	return directManager{}, nil
+func newDirectManager() directManager {
+	return directManager{fs: directFS{}}
+}
+
+func newDirectManagerOnFS(fs pinholeFS) directManager {
+	return directManager{fs: fs}
 }
 
 // ownedByTailscale reports whether /etc/resolv.conf seems to be a
 // tailscale-managed file.
 func (m directManager) ownedByTailscale() (bool, error) {
-	st, err := os.Stat(resolvConf)
+	isRegular, err := m.fs.Stat(resolvConf)
 	if err != nil {
 		if os.IsNotExist(err) {
 			return false, nil
 		}
 		return false, err
 	}
-	if !st.Mode().IsRegular() {
+	if !isRegular {
 		return false, nil
 	}
-	bs, err := ioutil.ReadFile(resolvConf)
+	bs, err := m.fs.ReadFile(resolvConf)
 	if err != nil {
 		return false, err
 	}
@@ -182,11 +182,11 @@ func (m directManager) ownedByTailscale() (bool, error) {
 // 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 _, err := m.fs.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)
+			m.fs.Remove(backupConf)
 			return nil
 		}
 		return err
@@ -200,11 +200,11 @@ func (m directManager) backupConfig() error {
 		return nil
 	}
 
-	return os.Rename(resolvConf, backupConf)
+	return m.fs.Rename(resolvConf, backupConf)
 }
 
 func (m directManager) restoreBackup() error {
-	if _, err := os.Stat(backupConf); err != nil {
+	if _, err := m.fs.Stat(backupConf); err != nil {
 		if os.IsNotExist(err) {
 			// No backup, nothing we can do.
 			return nil
@@ -215,7 +215,7 @@ func (m directManager) restoreBackup() error {
 	if err != nil {
 		return err
 	}
-	if _, err := os.Stat(resolvConf); err != nil && !os.IsNotExist(err) {
+	if _, err := m.fs.Stat(resolvConf); err != nil && !os.IsNotExist(err) {
 		return err
 	}
 	resolvConfExists := !os.IsNotExist(err)
@@ -223,12 +223,12 @@ func (m directManager) restoreBackup() error {
 	if resolvConfExists && !owned {
 		// There's already a non-tailscale config in place, get rid of
 		// our backup.
-		os.Remove(backupConf)
+		m.fs.Remove(backupConf)
 		return nil
 	}
 
 	// We own resolv.conf, and a backup exists.
-	if err := os.Rename(backupConf, resolvConf); err != nil {
+	if err := m.fs.Rename(backupConf, resolvConf); err != nil {
 		return err
 	}
 
@@ -247,7 +247,7 @@ func (m directManager) SetDNS(config OSConfig) error {
 
 		buf := new(bytes.Buffer)
 		writeResolvConf(buf, config.Nameservers, config.SearchDomains)
-		if err := atomicfile.WriteFile(resolvConf, buf.Bytes(), 0644); err != nil {
+		if err := atomicWriteFile(m.fs, resolvConf, buf.Bytes(), 0444); err != nil {
 			return err
 		}
 	}
@@ -279,7 +279,7 @@ func (m directManager) GetBaseConfig() (OSConfig, error) {
 		fileToRead = backupConf
 	}
 
-	return readResolvFile(fileToRead)
+	return m.readResolvFile(fileToRead)
 }
 
 func (m directManager) Close() error {
@@ -287,9 +287,9 @@ func (m directManager) Close() error {
 	// 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")
+	m.fs.Remove("/etc/resolv.tailscale.conf")
 
-	if _, err := os.Stat(backupConf); err != nil {
+	if _, err := m.fs.Stat(backupConf); err != nil {
 		if os.IsNotExist(err) {
 			// No backup, nothing we can do.
 			return nil
@@ -300,7 +300,7 @@ func (m directManager) Close() error {
 	if err != nil {
 		return err
 	}
-	_, err = os.Stat(resolvConf)
+	_, err = m.fs.Stat(resolvConf)
 	if err != nil && !os.IsNotExist(err) {
 		return err
 	}
@@ -309,12 +309,12 @@ func (m directManager) Close() error {
 	if resolvConfExists && !owned {
 		// There's already a non-tailscale config in place, get rid of
 		// our backup.
-		os.Remove(backupConf)
+		m.fs.Remove(backupConf)
 		return nil
 	}
 
 	// We own resolv.conf, and a backup exists.
-	if err := os.Rename(backupConf, resolvConf); err != nil {
+	if err := m.fs.Rename(backupConf, resolvConf); err != nil {
 		return err
 	}
 
@@ -324,3 +324,65 @@ func (m directManager) Close() error {
 
 	return nil
 }
+
+func atomicWriteFile(fs pinholeFS, filename string, data []byte, perm os.FileMode) error {
+	var randBytes [12]byte
+	if _, err := rand.Read(randBytes[:]); err != nil {
+		return fmt.Errorf("atomicWriteFile: %w", err)
+	}
+
+	tmpName := fmt.Sprintf("%s.%x.tmp", filename, randBytes[:])
+	if runtime.GOOS != "windows" {
+		// TODO: explain why the rename has to be the last command
+		defer fs.Remove(tmpName)
+	}
+
+	if err := fs.WriteFile(tmpName, data, perm); err != nil {
+		return fmt.Errorf("atomicWriteFile: %w", err)
+	}
+	if err := fs.WriteFile(tmpName+".2", data, perm); err != nil {
+		return fmt.Errorf("atomicWriteFile: %w", err)
+	}
+	return fs.Rename(tmpName, filename)
+}
+
+// pinholeFS is a high-level file system abstraction designed just for use
+// by directManager, with the goal that it is easy to implement over wsl.exe.
+type pinholeFS interface {
+	Stat(name string) (isRegular bool, err error)
+	Rename(oldName, newName string) error
+	Remove(name string) error
+	ReadFile(name string) ([]byte, error)
+	WriteFile(name string, contents []byte, perm os.FileMode) error
+	Symlink(oldName, newName string) error
+}
+
+// directFS is a pinholeFS implemented directly on the OS.
+type directFS struct {
+	prefix string // file path prefix; used for testing
+}
+
+func (fs directFS) Stat(name string) (isRegular bool, err error) {
+	fi, err := os.Stat(fs.prefix + name)
+	if err != nil {
+		return false, err
+	}
+	return fi.Mode().IsRegular(), nil
+}
+
+func (fs directFS) Rename(oldName, newName string) error {
+	return os.Rename(fs.prefix+oldName, fs.prefix+newName)
+}
+func (fs directFS) Symlink(oldName, newName string) error {
+	return os.Symlink(fs.prefix+oldName, fs.prefix+newName)
+}
+
+func (fs directFS) Remove(name string) error { return os.Remove(fs.prefix + name) }
+
+func (fs directFS) ReadFile(name string) ([]byte, error) {
+	return ioutil.ReadFile(fs.prefix + name)
+}
+
+func (fs directFS) WriteFile(name string, contents []byte, perm os.FileMode) error {
+	return ioutil.WriteFile(fs.prefix+name, contents, perm)
+}

net/dns/direct_test.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 dns
+
+import (
+	"io/ioutil"
+	"os"
+	"path/filepath"
+	"testing"
+
+	"inet.af/netaddr"
+	"tailscale.com/util/dnsname"
+)
+
+func TestSetDNS(t *testing.T) {
+	const orig = "nameserver 9.9.9.9 # orig"
+	tmp := t.TempDir()
+	resolvPath := filepath.Join(tmp, "etc", "resolv.conf")
+	backupPath := filepath.Join(tmp, "etc", "resolv.pre-tailscale-backup.conf")
+
+	if err := os.MkdirAll(filepath.Dir(resolvPath), 0777); err != nil {
+		t.Fatal(err)
+	}
+	if err := ioutil.WriteFile(resolvPath, []byte(orig), 0644); err != nil {
+		t.Fatal(err)
+	}
+
+	readFile := func(t *testing.T, path string) string {
+		t.Helper()
+		b, err := ioutil.ReadFile(path)
+		if err != nil {
+			t.Fatal(err)
+		}
+		return string(b)
+	}
+	assertBaseState := func(t *testing.T) {
+		if got := readFile(t, resolvPath); got != orig {
+			t.Fatalf("resolv.conf:\n%s, want:\n%s", got, orig)
+		}
+		if _, err := os.Stat(backupPath); !os.IsNotExist(err) {
+			t.Fatalf("resolv.conf backup: want it to be gone but: %v", err)
+		}
+	}
+
+	m := directManager{fs: directFS{prefix: tmp}}
+	if err := m.SetDNS(OSConfig{
+		Nameservers:   []netaddr.IP{netaddr.MustParseIP("8.8.8.8"), netaddr.MustParseIP("8.8.4.4")},
+		SearchDomains: []dnsname.FQDN{"ts.net.", "ts-dns.test."},
+		MatchDomains:  []dnsname.FQDN{"ignored."},
+	}); err != nil {
+		t.Fatal(err)
+	}
+	want := `# resolv.conf(5) file generated by tailscale
+# DO NOT EDIT THIS FILE BY HAND -- CHANGES WILL BE OVERWRITTEN
+
+nameserver 8.8.8.8
+nameserver 8.8.4.4
+search ts.net ts-dns.test
+`
+	if got := readFile(t, resolvPath); got != want {
+		t.Fatalf("resolv.conf:\n%s, want:\n%s", got, want)
+	}
+	if got := readFile(t, backupPath); got != orig {
+		t.Fatalf("resolv.conf backup:\n%s, want:\n%s", got, orig)
+	}
+
+	// Test that a nil OSConfig cleans up resolv.conf.
+	if err := m.SetDNS(OSConfig{}); err != nil {
+		t.Fatal(err)
+	}
+	assertBaseState(t)
+
+	// Test that Close cleans up resolv.conf.
+	if err := m.SetDNS(OSConfig{Nameservers: []netaddr.IP{netaddr.MustParseIP("8.8.8.8")}}); err != nil {
+		t.Fatal(err)
+	}
+	if err := m.Close(); err != nil {
+		t.Fatal(err)
+	}
+	assertBaseState(t)
+}

net/dns/manager.go

@@ -6,7 +6,6 @@ package dns
 
 import (
 	"runtime"
-	"strings"
 	"time"
 
 	"inet.af/netaddr"
@@ -21,8 +20,6 @@ import (
 // the lint exception is necessary and on others it is not,
 // and plain ignore complains if the exception is unnecessary.
 
-//lint:file-ignore U1000 reconfigTimeout is used on some platforms but not others
-
 // reconfigTimeout is the time interval within which Manager.{Up,Down} should complete.
 //
 // This is particularly useful because certain conditions can cause indefinite hangs
@@ -41,11 +38,11 @@ type Manager struct {
 }
 
 // NewManagers created a new manager from the given config.
-func NewManager(logf logger.Logf, oscfg OSConfigurator, linkMon *monitor.Mon) *Manager {
+func NewManager(logf logger.Logf, oscfg OSConfigurator, linkMon *monitor.Mon, linkSel resolver.ForwardLinkSelector) *Manager {
 	logf = logger.WithPrefix(logf, "dns: ")
 	m := &Manager{
 		logf:     logf,
-		resolver: resolver.New(logf, linkMon),
+		resolver: resolver.New(logf, linkMon, linkSel),
 		os:       oscfg,
 	}
 	m.logf("using %T", m.os)
@@ -75,40 +72,40 @@ func (m *Manager) Set(cfg Config) error {
 
 // compileConfig converts cfg into a quad-100 resolver configuration
 // and an OS-level configuration.
-func (m *Manager) compileConfig(cfg Config) (resolver.Config, OSConfig, error) {
+func (m *Manager) compileConfig(cfg Config) (rcfg resolver.Config, ocfg OSConfig, err error) {
+	// The internal resolver always gets MagicDNS hosts and
+	// authoritative suffixes, even if we don't propagate MagicDNS to
+	// the OS.
+	rcfg.Hosts = cfg.Hosts
+	routes := map[dnsname.FQDN][]netaddr.IPPort{} // assigned conditionally to rcfg.Routes below.
+	for suffix, resolvers := range cfg.Routes {
+		if len(resolvers) == 0 {
+			rcfg.LocalDomains = append(rcfg.LocalDomains, suffix)
+		} else {
+			routes[suffix] = resolvers
+		}
+	}
+	// Similarly, the OS always gets search paths.
+	ocfg.SearchDomains = cfg.SearchDomains
+
 	// 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
+		return rcfg, ocfg, nil
 	case cfg.hasDefaultResolversOnly():
 		// Trivial CorpDNS configuration, just override the OS
 		// resolver.
-		return resolver.Config{}, OSConfig{
-			Nameservers:   toIPsOnly(cfg.DefaultResolvers),
-			SearchDomains: cfg.SearchDomains,
-		}, nil
+		ocfg.Nameservers = toIPsOnly(cfg.DefaultResolvers)
+		return rcfg, ocfg, 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,
-		}
+		rcfg.Routes = routes
+		rcfg.Routes["."] = cfg.DefaultResolvers
+		ocfg.Nameservers = []netaddr.IP{tsaddr.TailscaleServiceIP()}
 		return rcfg, ocfg, nil
 	}
 
@@ -116,8 +113,6 @@ func (m *Manager) compileConfig(cfg Config) (resolver.Config, OSConfig, error) {
 	// 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
 
 	// Workaround for
 	// https://github.com/tailscale/corp/issues/1662. Even though
@@ -135,35 +130,19 @@ func (m *Manager) compileConfig(cfg Config) (resolver.Config, OSConfig, error) {
 	// This bool is used in a couple of places below to implement this
 	// workaround.
 	isWindows := runtime.GOOS == "windows"
-
-	// The windows check is for
-	// . See also below
-	// for further routing workarounds there.
-	if !cfg.hasHosts() && cfg.singleResolverSet() != nil && m.os.SupportsSplitDNS() && !isWindows {
+	if cfg.singleResolverSet() != nil && m.os.SupportsSplitDNS() && !isWindows {
 		// 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
+		ocfg.Nameservers = toIPsOnly(cfg.singleResolverSet())
+		ocfg.MatchDomains = cfg.matchDomains()
+		return rcfg, ocfg, 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,
-	}
+	rcfg.Routes = routes
+	ocfg.Nameservers = []netaddr.IP{tsaddr.TailscaleServiceIP()}
 
 	// If the OS can't do native split-dns, read out the underlying
 	// resolver config and blend it into our config.
@@ -173,28 +152,7 @@ func (m *Manager) compileConfig(cfg Config) (resolver.Config, OSConfig, error) {
 	if !m.os.SupportsSplitDNS() || isWindows {
 		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)},
-			}
+			return resolver.Config{}, OSConfig{}, err
 		}
 		rcfg.Routes["."] = toIPPorts(bcfg.Nameservers)
 		ocfg.SearchDomains = append(ocfg.SearchDomains, bcfg.SearchDomains...)
@@ -249,7 +207,7 @@ func Cleanup(logf logger.Logf, interfaceName string) {
 		logf("creating dns cleanup: %v", err)
 		return
 	}
-	dns := NewManager(logf, oscfg, nil)
+	dns := NewManager(logf, oscfg, nil, nil)
 	if err := dns.Down(); err != nil {
 		logf("dns down: %v", err)
 	}

net/dns/manager_freebsd.go

@@ -15,7 +15,7 @@ import (
 func NewOSConfigurator(logf logger.Logf, _ string) (OSConfigurator, error) {
 	bs, err := ioutil.ReadFile("/etc/resolv.conf")
 	if os.IsNotExist(err) {
-		return newDirectManager()
+		return newDirectManager(), nil
 	}
 	if err != nil {
 		return nil, fmt.Errorf("reading /etc/resolv.conf: %w", err)
@@ -25,6 +25,6 @@ func NewOSConfigurator(logf logger.Logf, _ string) (OSConfigurator, error) {
 	case "resolvconf":
 		return newResolvconfManager(logf)
 	default:
-		return newDirectManager()
+		return newDirectManager(), nil
 	}
 }

net/dns/manager_linux.go

@@ -5,7 +5,6 @@
 package dns
 
 import (
-	"bytes"
 	"context"
 	"errors"
 	"fmt"
@@ -15,6 +14,7 @@ import (
 	"time"
 
 	"github.com/godbus/dbus/v5"
+	"inet.af/netaddr"
 	"tailscale.com/types/logger"
 	"tailscale.com/util/cmpver"
 )
@@ -42,7 +42,7 @@ func NewOSConfigurator(logf logger.Logf, interfaceName string) (ret OSConfigurat
 	bs, err := ioutil.ReadFile("/etc/resolv.conf")
 	if os.IsNotExist(err) {
 		dbg("rc", "missing")
-		return newDirectManager()
+		return newDirectManager(), nil
 	}
 	if err != nil {
 		return nil, fmt.Errorf("reading /etc/resolv.conf: %w", err)
@@ -51,9 +51,18 @@ func NewOSConfigurator(logf logger.Logf, interfaceName string) (ret OSConfigurat
 	switch resolvOwner(bs) {
 	case "systemd-resolved":
 		dbg("rc", "resolved")
+		// Some systems, for reasons known only to them, have a
+		// resolv.conf that has the word "systemd-resolved" in its
+		// header, but doesn't actually point to resolved. We mustn't
+		// try to program resolved in that case.
+		// https://github.com/tailscale/tailscale/issues/2136
+		if err := resolvedIsActuallyResolver(); err != nil {
+			dbg("resolved", "not-in-use")
+			return newDirectManager(), nil
+		}
 		if err := dbusPing("org.freedesktop.resolve1", "/org/freedesktop/resolve1"); err != nil {
 			dbg("resolved", "no")
-			return newDirectManager()
+			return newDirectManager(), nil
 		}
 		if err := dbusPing("org.freedesktop.NetworkManager", "/org/freedesktop/NetworkManager/DnsManager"); err != nil {
 			dbg("nm", "no")
@@ -79,109 +88,69 @@ func NewOSConfigurator(logf logger.Logf, interfaceName string) (ret OSConfigurat
 		// "unmanaged" interfaces - meaning NM 1.26.6 and later
 		// actively ignore DNS configuration we give it. So, for those
 		// NM versions, we can and must use resolved directly.
-		old, err := nmVersionOlderThan("1.26.6")
+		//
+		// Even more fun, even-older versions of NM won't let us set
+		// DNS settings if the interface isn't managed by NM, with a
+		// hard failure on DBus requests. Empirically, NM 1.22 does
+		// this. Based on the versions popular distros shipped, we
+		// conservatively decree that only 1.26.0 through 1.26.5 are
+		// "safe" to use for our purposes. This roughly matches
+		// distros released in the latter half of 2020.
+		//
+		// In a perfect world, we'd avoid this by replacing
+		// configuration out from under NM entirely (e.g. using
+		// directManager to overwrite resolv.conf), but in a world
+		// where resolved runs, we need to get correct configuration
+		// into resolved regardless of what's in resolv.conf (because
+		// resolved can also be queried over dbus, or via an NSS
+		// module that bypasses /etc/resolv.conf). Given that we must
+		// get correct configuration into resolved, we have no choice
+		// but to use NM, and accept the loss of IPv6 configuration
+		// that comes with it (see
+		// https://github.com/tailscale/tailscale/issues/1699,
+		// https://github.com/tailscale/tailscale/pull/1945)
+		safe, err := nmVersionBetween("1.26.0", "1.26.5")
 		if err != nil {
 			// Failed to figure out NM's version, can't make a correct
 			// decision.
 			return nil, fmt.Errorf("checking NetworkManager version: %v", err)
 		}
-		if old {
-			dbg("nm-old", "yes")
+		if safe {
+			dbg("nm-safe", "yes")
 			return newNMManager(interfaceName)
 		}
-		dbg("nm-old", "no")
+		dbg("nm-safe", "no")
 		return newResolvedManager(logf, 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")
-				old, err := nmVersionOlderThan("1.26.6")
-				if err != nil {
-					return nil, fmt.Errorf("checking NetworkManager version: %v", err)
-				}
-				if old {
-					dbg("nm-old", "yes")
-					return newNMManager(interfaceName)
-				} else {
-					dbg("nm-old", "no")
-				}
-			} else {
-				dbg("nm", "no")
-			}
-		} else {
-			dbg("src-is-nm", "no")
-		}
 		if _, err := exec.LookPath("resolvconf"); err != nil {
 			dbg("resolvconf", "no")
-			return newDirectManager()
+			return newDirectManager(), nil
 		}
 		dbg("resolvconf", "yes")
 		return newResolvconfManager(logf)
 	case "NetworkManager":
+		// You'd think we would use newNMManager somewhere in
+		// here. However, as explained in
+		// https://github.com/tailscale/tailscale/issues/1699 , using
+		// NetworkManager for DNS configuration carries with it the
+		// cost of losing IPv6 configuration on the Tailscale network
+		// interface. So, when we can avoid it, we bypass
+		// NetworkManager by replacing resolv.conf directly.
+		//
+		// If you ever try to put NMManager back here, keep in mind
+		// that versions >=1.26.6 will ignore DNS configuration
+		// anyway, so you still need a fallback path that uses
+		// directManager.
 		dbg("rc", "nm")
-		if err := dbusPing("org.freedesktop.NetworkManager", "/org/freedesktop/NetworkManager/DnsManager"); err != nil {
-			dbg("nm", "no")
-			return newDirectManager()
-		}
-		dbg("nm", "yes")
-		old, err := nmVersionOlderThan("1.26.6")
-		if err != nil {
-			return nil, fmt.Errorf("checking NetworkManager version: %v", err)
-		}
-		if old {
-			dbg("nm-old", "yes")
-			return newNMManager(interfaceName)
-		}
-		dbg("nm-old", "no")
-		return newDirectManager()
+		return newDirectManager(), nil
 	default:
 		dbg("rc", "unknown")
-		return newDirectManager()
+		return newDirectManager(), nil
 	}
 }
 
-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 nmVersionOlderThan(want string) (bool, error) {
+func nmVersionBetween(first, last string) (bool, error) {
 	conn, err := dbus.SystemBus()
 	if err != nil {
 		// DBus probably not running.
@@ -199,7 +168,8 @@ func nmVersionOlderThan(want string) (bool, error) {
 		return false, fmt.Errorf("unexpected type %T for NM version", v.Value())
 	}
 
-	return cmpver.Compare(version, want) < 0, nil
+	outside := cmpver.Compare(version, first) < 0 || cmpver.Compare(version, last) > 0
+	return !outside, nil
 }
 
 func nmIsUsingResolved() error {
@@ -224,6 +194,17 @@ func nmIsUsingResolved() error {
 	return nil
 }
 
+func resolvedIsActuallyResolver() error {
+	cfg, err := newDirectManager().readResolvConf()
+	if err != nil {
+		return err
+	}
+	if len(cfg.Nameservers) != 1 || cfg.Nameservers[0] != netaddr.IPv4(127, 0, 0, 53) {
+		return errors.New("resolv.conf doesn't point to systemd-resolved")
+	}
+	return nil
+}
+
 func dbusPing(name, objectPath string) error {
 	ctx, cancel := context.WithTimeout(context.Background(), time.Second)
 	defer cancel()

net/dns/manager_openbsd.go

@@ -7,5 +7,5 @@ package dns
 import "tailscale.com/types/logger"
 
 func NewOSConfigurator(logger.Logf, string) (OSConfigurator, error) {
-	return newDirectManager()
+	return newDirectManager(), nil
 }

net/dns/manager_test.go

@@ -76,6 +76,20 @@ func TestManager(t *testing.T) {
 				SearchDomains: fqdns("tailscale.com", "universe.tf"),
 			},
 		},
+		{
+			// Regression test for https://github.com/tailscale/tailscale/issues/1886
+			name: "hosts-only",
+			in: Config{
+				Hosts: hosts(
+					"dave.ts.com.", "1.2.3.4",
+					"bradfitz.ts.com.", "2.3.4.5"),
+			},
+			rs: resolver.Config{
+				Hosts: hosts(
+					"dave.ts.com.", "1.2.3.4",
+					"bradfitz.ts.com.", "2.3.4.5"),
+			},
+		},
 		{
 			name: "corp",
 			in: Config{
@@ -104,10 +118,10 @@ func TestManager(t *testing.T) {
 			in: Config{
 				DefaultResolvers: mustIPPs("1.1.1.1:53", "9.9.9.9:53"),
 				SearchDomains:    fqdns("tailscale.com", "universe.tf"),
+				Routes:           upstreams("ts.com", ""),
 				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"),
@@ -126,10 +140,10 @@ func TestManager(t *testing.T) {
 			in: Config{
 				DefaultResolvers: mustIPPs("1.1.1.1:53", "9.9.9.9:53"),
 				SearchDomains:    fqdns("tailscale.com", "universe.tf"),
+				Routes:           upstreams("ts.com", ""),
 				Hosts: hosts(
 					"dave.ts.com.", "1.2.3.4",
 					"bradfitz.ts.com.", "2.3.4.5"),
-				AuthoritativeSuffixes: fqdns("ts.com"),
 			},
 			split: true,
 			os: OSConfig{
@@ -261,8 +275,8 @@ func TestManager(t *testing.T) {
 				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"),
+				Routes:        upstreams("ts.com", ""),
+				SearchDomains: fqdns("tailscale.com", "universe.tf"),
 			},
 			bs: OSConfig{
 				Nameservers:   mustIPs("8.8.8.8"),
@@ -286,8 +300,8 @@ func TestManager(t *testing.T) {
 				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"),
+				Routes:        upstreams("ts.com", ""),
+				SearchDomains: fqdns("tailscale.com", "universe.tf"),
 			},
 			split: true,
 			os: OSConfig{
@@ -305,12 +319,11 @@ func TestManager(t *testing.T) {
 		{
 			name: "routes-magic",
 			in: Config{
-				Routes: upstreams("corp.com", "2.2.2.2:53"),
+				Routes: upstreams("corp.com", "2.2.2.2:53", "ts.com", ""),
 				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"),
+				SearchDomains: fqdns("tailscale.com", "universe.tf"),
 			},
 			bs: OSConfig{
 				Nameservers:   mustIPs("8.8.8.8"),
@@ -333,12 +346,13 @@ func TestManager(t *testing.T) {
 		{
 			name: "routes-magic-split",
 			in: Config{
-				Routes: upstreams("corp.com", "2.2.2.2:53"),
+				Routes: upstreams(
+					"corp.com", "2.2.2.2:53",
+					"ts.com", ""),
 				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"),
+				SearchDomains: fqdns("tailscale.com", "universe.tf"),
 			},
 			split: true,
 			os: OSConfig{
@@ -362,7 +376,7 @@ func TestManager(t *testing.T) {
 				SplitDNS:   test.split,
 				BaseConfig: test.bs,
 			}
-			m := NewManager(t.Logf, &f, nil)
+			m := NewManager(t.Logf, &f, nil, nil)
 			m.resolver.TestOnlySetHook(f.SetResolver)
 
 			if err := m.Set(test.in); err != nil {
@@ -429,7 +443,12 @@ 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 s == "" {
+			if key == "" {
+				panic("IPPort provided before suffix")
+			}
+			ret[key] = nil
+		} else if ipp, err := netaddr.ParseIPPort(s); err == nil {
 			if key == "" {
 				panic("IPPort provided before suffix")
 			}

net/dns/manager_windows.go

@@ -35,16 +35,17 @@ const (
 )
 
 type windowsManager struct {
-	logf      logger.Logf
-	guid      string
-	nrptWorks bool
+	logf       logger.Logf
+	guid       string
+	nrptWorks  bool
+	wslManager *wslManager
 }
 
 func NewOSConfigurator(logf logger.Logf, interfaceName string) (OSConfigurator, error) {
 	ret := windowsManager{
 		logf:      logf,
 		guid:      interfaceName,
-		nrptWorks: !isWindows7(),
+		nrptWorks: isWindows10OrBetter(),
 	}
 
 	// Best-effort: if our NRPT rule exists, try to delete it. Unlike
@@ -57,6 +58,10 @@ func NewOSConfigurator(logf logger.Logf, interfaceName string) (OSConfigurator,
 		ret.delKey(nrptBase)
 	}
 
+	if wslExists() {
+		ret.wslManager = newWSLManager(logf)
+	}
+
 	return ret, nil
 }
 
@@ -294,6 +299,14 @@ func (m windowsManager) SetDNS(cfg OSConfig) error {
 		} else {
 			m.logf("ran ipconfig /flushdns in %v", d)
 		}
+
+		if m.wslManager != nil {
+			if err := m.wslManager.SetDNS(cfg); err != nil {
+				m.logf("WSL SetDNS: %v", err) // continue
+			} else {
+				m.logf("WSL SetDNS was successful")
+			}
+		}
 	}()
 
 	return nil
@@ -407,22 +420,16 @@ var siteLocalResolvers = []netaddr.IP{
 	netaddr.MustParseIP("fec0:0:0:ffff::3"),
 }
 
-func isWindows7() bool {
+func isWindows10OrBetter() bool {
 	key, err := registry.OpenKey(registry.LOCAL_MACHINE, versionKey, registry.READ)
 	if err != nil {
-		// Fail safe, assume Windows 7.
-		return true
+		// Fail safe, assume old Windows.
+		return false
 	}
-	ver, _, err := key.GetStringValue("CurrentVersion")
-	if err != nil {
-		return true
+	// This key above only exists in Windows 10 and above. Its mere
+	// presence is good enough.
+	if _, _, err := key.GetIntegerValue("CurrentMajorVersionNumber"); err != nil {
+		return false
 	}
-	// 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"
+	return true
 }

net/dns/nm.go

@@ -4,8 +4,6 @@
 
 // +build linux
 
-//lint:file-ignore U1000 refactoring, temporarily unused code.
-
 package dns
 
 import (

net/dns/resolved.go

@@ -4,8 +4,6 @@
 
 // +build linux
 
-//lint:file-ignore U1000 refactoring, temporarily unused code.
-
 package dns
 
 import (
@@ -69,7 +67,7 @@ func isResolvedActive() bool {
 		return false
 	}
 
-	config, err := readResolvConf()
+	config, err := newDirectManager().readResolvConf()
 	if err != nil {
 		return false
 	}
@@ -82,7 +80,7 @@ func isResolvedActive() bool {
 	return false
 }
 
-// resolvedManager uses the systemd-resolved DBus API.
+// resolvedManager is an OSConfigurator which uses the systemd-resolved DBus API.
 type resolvedManager struct {
 	logf     logger.Logf
 	ifidx    int
@@ -107,7 +105,6 @@ func newResolvedManager(logf logger.Logf, interfaceName string) (*resolvedManage
 	}, nil
 }
 
-// Up implements managerImpl.
 func (m *resolvedManager) SetDNS(config OSConfig) error {
 	ctx, cancel := context.WithTimeout(context.Background(), reconfigTimeout)
 	defer cancel()

net/dns/resolver/forwarder.go

@@ -9,8 +9,8 @@ import (
 	"context"
 	"encoding/binary"
 	"errors"
-	"fmt"
 	"hash/crc32"
+	"io"
 	"math/rand"
 	"net"
 	"sync"
@@ -18,31 +18,21 @@ import (
 
 	dns "golang.org/x/net/dns/dnsmessage"
 	"inet.af/netaddr"
-	"tailscale.com/logtail/backoff"
 	"tailscale.com/types/logger"
 	"tailscale.com/util/dnsname"
+	"tailscale.com/wgengine/monitor"
 )
 
 // headerBytes is the number of bytes in a DNS message header.
 const headerBytes = 12
 
-// connCount is the number of UDP connections to use for forwarding.
-const connCount = 32
-
 const (
-	// cleanupInterval is the interval between purged of timed-out entries from txMap.
-	cleanupInterval = 30 * time.Second
 	// responseTimeout is the maximal amount of time to wait for a DNS response.
 	responseTimeout = 5 * time.Second
 )
 
 var errNoUpstreams = errors.New("upstream nameservers not set")
 
-type forwardingRecord struct {
-	src       netaddr.IPPort
-	createdAt time.Time
-}
-
 // txid identifies a DNS transaction.
 //
 // As the standard DNS Request ID is only 16 bits, we extend it:
@@ -98,160 +88,218 @@ func getTxID(packet []byte) txid {
 	return (txid(hash) << 32) | txid(dnsid)
 }
 
+// clampEDNSSize attempts to limit the maximum EDNS response size. This is not
+// an exhaustive solution, instead only easy cases are currently handled in the
+// interest of speed and reduced complexity. Only OPT records at the very end of
+// the message with no option codes are addressed.
+// TODO: handle more situations if we discover that they happen often
+func clampEDNSSize(packet []byte, maxSize uint16) {
+	// optFixedBytes is the size of an OPT record with no option codes.
+	const optFixedBytes = 11
+	const edns0Version = 0
+
+	if len(packet) < headerBytes+optFixedBytes {
+		return
+	}
+
+	arCount := binary.BigEndian.Uint16(packet[10:12])
+	if arCount == 0 {
+		// OPT shows up in an AR, so there must be no OPT
+		return
+	}
+
+	opt := packet[len(packet)-optFixedBytes:]
+
+	if opt[0] != 0 {
+		// OPT NAME must be 0 (root domain)
+		return
+	}
+	if dns.Type(binary.BigEndian.Uint16(opt[1:3])) != dns.TypeOPT {
+		// Not an OPT record
+		return
+	}
+	requestedSize := binary.BigEndian.Uint16(opt[3:5])
+	// Ignore extended RCODE in opt[5]
+	if opt[6] != edns0Version {
+		// Be conservative and don't touch unknown versions.
+		return
+	}
+	// Ignore flags in opt[7:9]
+	if binary.BigEndian.Uint16(opt[10:12]) != 0 {
+		// RDLEN must be 0 (no variable length data). We're at the end of the
+		// packet so this should be 0 anyway)..
+		return
+	}
+
+	if requestedSize <= maxSize {
+		return
+	}
+
+	// Clamp the maximum size
+	binary.BigEndian.PutUint16(opt[3:5], maxSize)
+}
+
 type route struct {
-	suffix    dnsname.FQDN
-	resolvers []netaddr.IPPort
+	Suffix    dnsname.FQDN
+	Resolvers []netaddr.IPPort
 }
 
 // forwarder forwards DNS packets to a number of upstream nameservers.
 type forwarder struct {
-	logf logger.Logf
+	logf    logger.Logf
+	linkMon *monitor.Mon
+	linkSel ForwardLinkSelector
+
+	ctx       context.Context    // good until Close
+	ctxCancel context.CancelFunc // closes ctx
 
 	// responses is a channel by which responses are returned.
 	responses chan packet
-	// closed signals all goroutines to stop.
-	closed chan struct{}
-	// wg signals when all goroutines have stopped.
-	wg sync.WaitGroup
 
-	// conns are the UDP connections used for forwarding.
-	// A random one is selected for each request, regardless of the target upstream.
-	conns []*fwdConn
+	mu sync.Mutex // guards following
 
-	mu sync.Mutex
-	// routes are per-suffix resolvers to use.
-	routes []route                   // most specific routes first
-	txMap  map[txid]forwardingRecord // txids to in-flight requests
+	// routes are per-suffix resolvers to use, with
+	// the most specific routes first.
+	routes []route
 }
 
 func init() {
 	rand.Seed(time.Now().UnixNano())
 }
 
-func newForwarder(logf logger.Logf, responses chan packet) *forwarder {
-	ret := &forwarder{
+func newForwarder(logf logger.Logf, responses chan packet, linkMon *monitor.Mon, linkSel ForwardLinkSelector) *forwarder {
+	f := &forwarder{
 		logf:      logger.WithPrefix(logf, "forward: "),
+		linkMon:   linkMon,
+		linkSel:   linkSel,
 		responses: responses,
-		closed:    make(chan struct{}),
-		conns:     make([]*fwdConn, connCount),
-		txMap:     make(map[txid]forwardingRecord),
 	}
-
-	ret.wg.Add(connCount + 1)
-	for idx := range ret.conns {
-		ret.conns[idx] = newFwdConn(ret.logf, idx)
-		go ret.recv(ret.conns[idx])
-	}
-	go ret.cleanMap()
-
-	return ret
+	f.ctx, f.ctxCancel = context.WithCancel(context.Background())
+	return f
 }
 
-func (f *forwarder) Close() {
-	select {
-	case <-f.closed:
-		return
-	default:
-		// continue
-	}
-	close(f.closed)
-
-	for _, conn := range f.conns {
-		conn.close()
-	}
-
-	f.wg.Wait()
-}
-
-func (f *forwarder) rebindFromNetworkChange() {
-	for _, c := range f.conns {
-		c.mu.Lock()
-		c.reconnectLocked()
-		c.mu.Unlock()
-	}
+func (f *forwarder) Close() error {
+	f.ctxCancel()
+	return nil
 }
 
 func (f *forwarder) setRoutes(routes []route) {
 	f.mu.Lock()
+	defer f.mu.Unlock()
 	f.routes = routes
-	f.mu.Unlock()
+}
+
+var stdNetPacketListener packetListener = new(net.ListenConfig)
+
+type packetListener interface {
+	ListenPacket(ctx context.Context, network, address string) (net.PacketConn, error)
+}
+
+func (f *forwarder) packetListener(ip netaddr.IP) (packetListener, error) {
+	if f.linkSel == nil || initListenConfig == nil {
+		return stdNetPacketListener, nil
+	}
+	linkName := f.linkSel.PickLink(ip)
+	if linkName == "" {
+		return stdNetPacketListener, nil
+	}
+	lc := new(net.ListenConfig)
+	if err := initListenConfig(lc, f.linkMon, linkName); err != nil {
+		return nil, err
+	}
+	return lc, nil
 }
 
 // send sends packet to dst. It is best effort.
-func (f *forwarder) send(packet []byte, dst netaddr.IPPort) {
-	connIdx := rand.Intn(connCount)
-	conn := f.conns[connIdx]
-	conn.send(packet, dst)
-}
+//
+// send expects the reply to have the same txid as txidOut.
+//
+// The provided closeOnCtxDone lets send register values to Close if
+// the caller's ctx expires. This avoids send from allocating its own
+// waiting goroutine to interrupt the ReadFrom, as memory is tight on
+// iOS and we want the number of pending DNS lookups to be bursty
+// without too much associated goroutine/memory cost.
+func (f *forwarder) send(ctx context.Context, txidOut txid, closeOnCtxDone *closePool, packet []byte, dst netaddr.IPPort) ([]byte, error) {
+	// TODO(bradfitz): if dst.IP is 8.8.8.8 or 8.8.4.4 or 1.1.1.1, etc, or
+	// something dynamically probed earlier to support DoH or DoT,
+	// do that here instead.
 
-func (f *forwarder) recv(conn *fwdConn) {
-	defer f.wg.Done()
+	ln, err := f.packetListener(dst.IP())
+	if err != nil {
+		return nil, err
+	}
+	conn, err := ln.ListenPacket(ctx, "udp", ":0")
+	if err != nil {
+		f.logf("ListenPacket failed: %v", err)
+		return nil, err
+	}
+	defer conn.Close()
 
-	for {
-		select {
-		case <-f.closed:
-			return
-		default:
+	closeOnCtxDone.Add(conn)
+	defer closeOnCtxDone.Remove(conn)
+
+	if _, err := conn.WriteTo(packet, dst.UDPAddr()); err != nil {
+		if err := ctx.Err(); err != nil {
+			return nil, err
 		}
-		out := make([]byte, maxResponseBytes)
-		n := conn.read(out)
-		if n == 0 {
-			continue
+		return nil, err
+	}
+
+	// The 1 extra byte is to detect packet truncation.
+	out := make([]byte, maxResponseBytes+1)
+	n, _, err := conn.ReadFrom(out)
+	if err != nil {
+		if err := ctx.Err(); err != nil {
+			return nil, err
 		}
-		if n < headerBytes {
-			f.logf("recv: packet too small (%d bytes)", n)
-		}
-
-		out = out[:n]
-		txid := getTxID(out)
-
-		f.mu.Lock()
-
-		record, found := f.txMap[txid]
-		// At most one nameserver will return a response:
-		// the first one to do so will delete txid from the map.
-		if !found {
-			f.mu.Unlock()
-			continue
-		}
-		delete(f.txMap, txid)
-
-		f.mu.Unlock()
-
-		pkt := packet{out, record.src}
-		select {
-		case <-f.closed:
-			return
-		case f.responses <- pkt:
-			// continue
+		if packetWasTruncated(err) {
+			err = nil
+		} else {
+			return nil, err
 		}
 	}
+	truncated := n > maxResponseBytes
+	if truncated {
+		n = maxResponseBytes
+	}
+	if n < headerBytes {
+		f.logf("recv: packet too small (%d bytes)", n)
+	}
+	out = out[:n]
+	txid := getTxID(out)
+	if txid != txidOut {
+		return nil, errors.New("txid doesn't match")
+	}
+
+	if truncated {
+		const dnsFlagTruncated = 0x200
+		flags := binary.BigEndian.Uint16(out[2:4])
+		flags |= dnsFlagTruncated
+		binary.BigEndian.PutUint16(out[2:4], flags)
+
+		// TODO(#2067): Remove any incomplete records? RFC 1035 section 6.2
+		// states that truncation should head drop so that the authority
+		// section can be preserved if possible. However, the UDP read with
+		// a too-small buffer has already dropped the end, so that's the
+		// best we can do.
+	}
+
+	clampEDNSSize(out, maxResponseBytes)
+
+	return out, nil
 }
 
-// cleanMap periodically deletes timed-out forwarding records from f.txMap to bound growth.
-func (f *forwarder) cleanMap() {
-	defer f.wg.Done()
-
-	t := time.NewTicker(cleanupInterval)
-	defer t.Stop()
-
-	var now time.Time
-	for {
-		select {
-		case <-f.closed:
-			return
-		case now = <-t.C:
-			// continue
+// resolvers returns the resolvers to use for domain.
+func (f *forwarder) resolvers(domain dnsname.FQDN) []netaddr.IPPort {
+	f.mu.Lock()
+	routes := f.routes
+	f.mu.Unlock()
+	for _, route := range routes {
+		if route.Suffix == "." || route.Suffix.Contains(domain) {
+			return route.Resolvers
 		}
-
-		f.mu.Lock()
-		for k, v := range f.txMap {
-			if now.Sub(v.createdAt) > responseTimeout {
-				delete(f.txMap, k)
-			}
-		}
-		f.mu.Unlock()
 	}
+	return nil
 }
 
 // forward forwards the query to all upstream nameservers and returns the first response.
@@ -262,218 +310,62 @@ func (f *forwarder) forward(query packet) error {
 	}
 
 	txid := getTxID(query.bs)
+	clampEDNSSize(query.bs, maxResponseBytes)
 
-	f.mu.Lock()
-	routes := f.routes
-	f.mu.Unlock()
-
-	var resolvers []netaddr.IPPort
-	for _, route := range routes {
-		if route.suffix != "." && !route.suffix.Contains(domain) {
-			continue
-		}
-		resolvers = route.resolvers
-		break
-	}
+	resolvers := f.resolvers(domain)
 	if len(resolvers) == 0 {
 		return errNoUpstreams
 	}
 
-	f.mu.Lock()
-	f.txMap[txid] = forwardingRecord{
-		src:       query.addr,
-		createdAt: time.Now(),
-	}
-	f.mu.Unlock()
+	closeOnCtxDone := new(closePool)
+	defer closeOnCtxDone.Close()
 
-	for _, resolver := range resolvers {
-		f.send(query.bs, resolver)
-	}
+	ctx, cancel := context.WithTimeout(f.ctx, responseTimeout)
+	defer cancel()
 
-	return nil
-}
+	resc := make(chan []byte, 1)
+	var (
+		mu       sync.Mutex
+		firstErr error
+	)
 
-// A fwdConn manages a single connection used to forward DNS requests.
-// Net link changes can cause a *net.UDPConn to become permanently unusable, particularly on macOS.
-// fwdConn detects such situations and transparently creates new connections.
-type fwdConn struct {
-	// logf allows a fwdConn to log.
-	logf logger.Logf
-
-	// change allows calls to read to block until a the network connection has been replaced.
-	change *sync.Cond
-
-	// mu protects fields that follow it; it is also change's Locker.
-	mu sync.Mutex
-	// closed tracks whether fwdConn has been permanently closed.
-	closed bool
-	// conn is the current active connection.
-	conn net.PacketConn
-}
-
-func newFwdConn(logf logger.Logf, idx int) *fwdConn {
-	c := new(fwdConn)
-	c.logf = logger.WithPrefix(logf, fmt.Sprintf("fwdConn %d: ", idx))
-	c.change = sync.NewCond(&c.mu)
-	// c.conn is created lazily in send
-	return c
-}
-
-// 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 netaddr.IPPort) {
-	var b *backoff.Backoff // lazily initialized, since it is not needed in the common case
-	backOff := func(err error) {
-		if b == nil {
-			b = backoff.NewBackoff("dns-fwdConn-send", c.logf, 30*time.Second)
-		}
-		b.BackOff(context.Background(), err)
-	}
-
-	for {
-		// Gather the current connection.
-		// We can't hold the lock while we call WriteTo.
-		c.mu.Lock()
-		conn := c.conn
-		closed := c.closed
-		if closed {
-			c.mu.Unlock()
-			return
-		}
-		if conn == nil {
-			c.reconnectLocked()
-			c.mu.Unlock()
-			continue
-		}
-		c.mu.Unlock()
-
-		_, err := conn.WriteTo(packet, dst.UDPAddr())
-		if err == nil {
-			// Success
-			return
-		}
-		if errors.Is(err, net.ErrClosed) {
-			// We intentionally closed this connection.
-			// It has been replaced by a new connection. Try again.
-			continue
-		}
-		// Something else went wrong.
-		// We have three choices here: try again, give up, or create a new connection.
-		var opErr *net.OpError
-		if !errors.As(err, &opErr) {
-			// Weird. All errors from the net package should be *net.OpError. Bail.
-			c.logf("send: non-*net.OpErr %v (%T)", err, err)
-			return
-		}
-		if opErr.Temporary() || opErr.Timeout() {
-			// I doubt that either of these can happen (this is UDP),
-			// but go ahead and try again.
-			backOff(err)
-			continue
-		}
-		if networkIsDown(err) {
-			// Fail.
-			c.logf("send: network is down")
-			return
-		}
-		if networkIsUnreachable(err) {
-			// This can be caused by a link change.
-			// Replace the existing connection with a new one.
-			c.mu.Lock()
-			// It's possible that multiple senders discovered simultaneously
-			// that the network is unreachable. Avoid reconnecting multiple times:
-			// Only reconnect if the current connection is the one that we
-			// discovered to be problematic.
-			if c.conn == conn {
-				backOff(err)
-				c.reconnectLocked()
+	for _, ipp := range resolvers {
+		go func(ipp netaddr.IPPort) {
+			resb, err := f.send(ctx, txid, closeOnCtxDone, query.bs, ipp)
+			if err != nil {
+				mu.Lock()
+				defer mu.Unlock()
+				if firstErr == nil {
+					firstErr = err
+				}
+				return
 			}
-			c.mu.Unlock()
-			// Try again with our new network connection.
-			continue
+			select {
+			case resc <- resb:
+			default:
+			}
+		}(ipp)
+	}
+
+	select {
+	case v := <-resc:
+		select {
+		case <-ctx.Done():
+			return ctx.Err()
+		case f.responses <- packet{v, query.addr}:
+			return nil
 		}
-		// Unrecognized error. Fail.
-		c.logf("send: unrecognized error: %v", err)
-		return
+	case <-ctx.Done():
+		mu.Lock()
+		defer mu.Unlock()
+		if firstErr != nil {
+			return firstErr
+		}
+		return ctx.Err()
 	}
 }
 
-// read waits for a response from c's connection.
-// It returns the number of bytes read, which may be 0
-// in case of an error or a closed connection.
-func (c *fwdConn) read(out []byte) int {
-	for {
-		// Gather the current connection.
-		// We can't hold the lock while we call ReadFrom.
-		c.mu.Lock()
-		conn := c.conn
-		closed := c.closed
-		if closed {
-			c.mu.Unlock()
-			return 0
-		}
-		if conn == nil {
-			// There is no current connection.
-			// Wait for the connection to change, then try again.
-			c.change.Wait()
-			c.mu.Unlock()
-			continue
-		}
-		c.mu.Unlock()
-
-		n, _, err := conn.ReadFrom(out)
-		if err == nil {
-			// Success.
-			return n
-		}
-		if errors.Is(err, net.ErrClosed) {
-			// We intentionally closed this connection.
-			// It has been replaced by a new connection. Try again.
-			continue
-		}
-
-		c.logf("read: unrecognized error: %v", err)
-		return 0
-	}
-}
-
-// reconnectLocked replaces the current connection with a new one.
-// c.mu must be locked.
-func (c *fwdConn) reconnectLocked() {
-	c.closeConnLocked()
-	// Make a new connection.
-	conn, err := net.ListenPacket("udp", "")
-	if err != nil {
-		c.logf("ListenPacket failed: %v", err)
-	} else {
-		c.conn = conn
-	}
-	// Broadcast that a new connection is available.
-	c.change.Broadcast()
-}
-
-// closeCurrentConn closes the current connection.
-// c.mu must be locked.
-func (c *fwdConn) closeConnLocked() {
-	if c.conn == nil {
-		return
-	}
-	c.conn.Close() // unblocks all readers/writers, they'll pick up the next connection.
-	c.conn = nil
-}
-
-// close permanently closes c.
-func (c *fwdConn) close() {
-	c.mu.Lock()
-	defer c.mu.Unlock()
-	if c.closed {
-		return
-	}
-	c.closed = true
-	c.closeConnLocked()
-	// Unblock any remaining readers.
-	c.change.Broadcast()
-}
+var initListenConfig func(_ *net.ListenConfig, _ *monitor.Mon, tunName string) error
 
 // nameFromQuery extracts the normalized query name from bs.
 func nameFromQuery(bs []byte) (dnsname.FQDN, error) {
@@ -495,3 +387,48 @@ func nameFromQuery(bs []byte) (dnsname.FQDN, error) {
 	n := q.Name.Data[:q.Name.Length]
 	return dnsname.ToFQDN(rawNameToLower(n))
 }
+
+// closePool is a dynamic set of io.Closers to close as a group.
+// It's intended to be Closed at most once.
+//
+// The zero value is ready for use.
+type closePool struct {
+	mu     sync.Mutex
+	m      map[io.Closer]bool
+	closed bool
+}
+
+func (p *closePool) Add(c io.Closer) {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	if p.closed {
+		c.Close()
+		return
+	}
+	if p.m == nil {
+		p.m = map[io.Closer]bool{}
+	}
+	p.m[c] = true
+}
+
+func (p *closePool) Remove(c io.Closer) {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	if p.closed {
+		return
+	}
+	delete(p.m, c)
+}
+
+func (p *closePool) Close() error {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	if p.closed {
+		return nil
+	}
+	p.closed = true
+	for c := range p.m {
+		c.Close()
+	}
+	return nil
+}

net/dns/resolver/macios_ext.go

@@ -0,0 +1,27 @@
+// 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 darwin,ts_macext ios,ts_macext
+
+package resolver
+
+import (
+	"errors"
+	"net"
+
+	"tailscale.com/net/netns"
+	"tailscale.com/wgengine/monitor"
+)
+
+func init() {
+	initListenConfig = initListenConfigNetworkExtension
+}
+
+func initListenConfigNetworkExtension(nc *net.ListenConfig, mon *monitor.Mon, tunName string) error {
+	nif, ok := mon.InterfaceState().Interface[tunName]
+	if !ok {
+		return errors.New("utun not found")
+	}
+	return netns.SetListenConfigInterfaceIndex(nc, nif.Interface.Index)
+}

net/dns/resolver/neterr_darwin.go

@@ -23,3 +23,8 @@ func networkIsDown(err error) bool {
 func networkIsUnreachable(err error) bool {
 	return errors.Is(err, networkUnreachable)
 }
+
+// packetWasTruncated returns true if err indicates truncation but the RecvFrom
+// that generated err was otherwise successful. It always returns false on this
+// platform.
+func packetWasTruncated(err error) bool { return false }

net/dns/resolver/neterr_other.go

@@ -8,3 +8,8 @@ package resolver
 
 func networkIsDown(err error) bool        { return false }
 func networkIsUnreachable(err error) bool { return false }
+
+// packetWasTruncated returns true if err indicates truncation but the RecvFrom
+// that generated err was otherwise successful. It always returns false on this
+// platform.
+func packetWasTruncated(err error) bool   { return false }

net/dns/resolver/neterr_windows.go

@@ -5,6 +5,7 @@
 package resolver
 
 import (
+	"errors"
 	"net"
 	"os"
 
@@ -27,3 +28,16 @@ func networkIsUnreachable(err error) bool {
 	// difference between down and unreachable? Add comments.
 	return false
 }
+
+// packetWasTruncated returns true if err indicates truncation but the RecvFrom
+// that generated err was otherwise successful. On Windows, Go's UDP RecvFrom
+// calls WSARecvFrom which returns the WSAEMSGSIZE error code when the received
+// datagram is larger than the provided buffer. When that happens, both a valid
+// size and an error are returned (as per the partial fix for golang/go#14074).
+// If the WSAEMSGSIZE error is returned, then we ignore the error to get
+// semantics similar to the POSIX operating systems. One caveat is that it
+// appears that the source address is not returned when WSAEMSGSIZE occurs, but
+// we do not currently look at the source address.
+func packetWasTruncated(err error) bool {
+	return errors.Is(err, windows.WSAEMSGSIZE)
+}

net/dns/resolver/tsdns.go

@@ -9,26 +9,39 @@ package resolver
 import (
 	"encoding/hex"
 	"errors"
+	"runtime"
 	"sort"
 	"strings"
 	"sync"
+	"sync/atomic"
 	"time"
 
 	dns "golang.org/x/net/dns/dnsmessage"
 	"inet.af/netaddr"
-	"tailscale.com/net/interfaces"
 	"tailscale.com/types/logger"
 	"tailscale.com/util/dnsname"
 	"tailscale.com/wgengine/monitor"
 )
 
-// maxResponseBytes is the maximum size of a response from a Resolver.
-const maxResponseBytes = 512
+// maxResponseBytes is the maximum size of a response from a Resolver. The
+// actual buffer size will be one larger than this so that we can detect
+// truncation in a platform-agnostic way.
+const maxResponseBytes = 4095
 
-// queueSize is the maximal number of DNS requests that can await polling.
+// maxActiveQueries returns the maximal number of DNS requests that be
+// can running.
 // If EnqueueRequest is called when this many requests are already pending,
 // the request will be dropped to avoid blocking the caller.
-const queueSize = 64
+func maxActiveQueries() int32 {
+	if runtime.GOOS == "ios" {
+		// For memory paranoia reasons on iOS, match the
+		// historical Tailscale 1.x..1.8 behavior for now
+		// (just before the 1.10 release).
+		return 64
+	}
+	// But for other platforms, allow more burstiness:
+	return 256
+}
 
 // defaultTTL is the TTL of all responses from Resolver.
 const defaultTTL = 600 * time.Second
@@ -73,13 +86,12 @@ type Config struct {
 type Resolver struct {
 	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
+	activeQueriesAtomic int32 // number of DNS queries in flight
+
 	// responses is an unbuffered channel to which responses are returned.
 	responses chan packet
 	// errors is an unbuffered channel to which errors are returned.
@@ -96,27 +108,26 @@ type Resolver struct {
 	ipToHost     map[netaddr.IP]dnsname.FQDN
 }
 
+type ForwardLinkSelector interface {
+	// PickLink returns which network device should be used to query
+	// the DNS server at the given IP.
+	// The empty string means to use an unspecified default.
+	PickLink(netaddr.IP) (linkName string)
+}
+
 // 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 {
+func New(logf logger.Logf, linkMon *monitor.Mon, linkSel ForwardLinkSelector) *Resolver {
 	r := &Resolver{
 		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{},
 	}
-	r.forwarder = newForwarder(r.logf, r.responses)
-	if r.linkMon != nil {
-		r.unregLinkMon = r.linkMon.RegisterChangeCallback(r.onLinkMonitorChange)
-	}
-
-	r.wg.Add(1)
-	go r.poll()
-
+	r.forwarder = newForwarder(r.logf, r.responses, linkMon, linkSel)
 	return r
 }
 
@@ -138,13 +149,13 @@ func (r *Resolver) SetConfig(cfg Config) error {
 
 	for suffix, ips := range cfg.Routes {
 		routes = append(routes, route{
-			suffix:    suffix,
-			resolvers: ips,
+			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()
+		return routes[i].Suffix.NumLabels() > routes[j].Suffix.NumLabels()
 	})
 
 	r.forwarder.setRoutes(routes)
@@ -168,19 +179,7 @@ func (r *Resolver) Close() {
 	}
 	close(r.closed)
 
-	if r.unregLinkMon != nil {
-		r.unregLinkMon()
-	}
-
 	r.forwarder.Close()
-	r.wg.Wait()
-}
-
-func (r *Resolver) onLinkMonitorChange(changed bool, state *interfaces.State) {
-	if !changed {
-		return
-	}
-	r.forwarder.rebindFromNetworkChange()
 }
 
 // EnqueueRequest places the given DNS request in the resolver's queue.
@@ -190,11 +189,14 @@ func (r *Resolver) EnqueueRequest(bs []byte, from netaddr.IPPort) error {
 	select {
 	case <-r.closed:
 		return ErrClosed
-	case r.queue <- packet{bs, from}:
-		return nil
 	default:
+	}
+	if n := atomic.AddInt32(&r.activeQueriesAtomic, 1); n > maxActiveQueries() {
+		atomic.AddInt32(&r.activeQueriesAtomic, -1)
 		return errFullQueue
 	}
+	go r.handleQuery(packet{bs, from})
+	return nil
 }
 
 // NextResponse returns a DNS response to a previously enqueued request.
@@ -289,53 +291,34 @@ func (r *Resolver) resolveLocal(domain dnsname.FQDN, typ dns.Type) (netaddr.IP,
 // 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()
-	ips := r.ipToHost
-	r.mu.Unlock()
-
-	name, found := ips[ip]
-	if !found {
+	defer r.mu.Unlock()
+	name, ok := r.ipToHost[ip]
+	if !ok {
 		return "", dns.RCodeNameError
 	}
 	return name, dns.RCodeSuccess
 }
 
-func (r *Resolver) poll() {
-	defer r.wg.Done()
+func (r *Resolver) handleQuery(pkt packet) {
+	defer atomic.AddInt32(&r.activeQueriesAtomic, -1)
 
-	var pkt packet
-	for {
+	out, err := r.respond(pkt.bs)
+	if err == errNotOurName {
+		err = r.forwarder.forward(pkt)
+		if err == nil {
+			// forward will send response into r.responses, nothing to do.
+			return
+		}
+	}
+	if err != nil {
 		select {
 		case <-r.closed:
-			return
-		case pkt = <-r.queue:
-			// continue
+		case r.errors <- err:
 		}
-
-		out, err := r.respond(pkt.bs)
-
-		if err == errNotOurName {
-			err = r.forwarder.forward(pkt)
-			if err == nil {
-				// forward will send response into r.responses, nothing to do.
-				continue
-			}
-		}
-
-		if err != nil {
-			select {
-			case <-r.closed:
-				return
-			case r.errors <- err:
-				// continue
-			}
-		} else {
-			pkt.bs = out
-			select {
-			case <-r.closed:
-				return
-			case r.responses <- pkt:
-				// continue
-			}
+	} else {
+		select {
+		case <-r.closed:
+		case r.responses <- packet{out, pkt.addr}:
 		}
 	}
 }
@@ -349,28 +332,44 @@ type response struct {
 	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
+var dnsParserPool = &sync.Pool{
+	New: func() interface{} {
+		return new(dnsParser)
+	},
+}
 
-	resp.Header, err = parser.Start(query)
+// dnsParser parses DNS queries using x/net/dns/dnsmessage.
+// These structs are pooled with dnsParserPool.
+type dnsParser struct {
+	Header   dns.Header
+	Question dns.Question
+
+	parser dns.Parser
+}
+
+func (p *dnsParser) response() *response {
+	return &response{Header: p.Header, Question: p.Question}
+}
+
+// zeroParser clears parser so it doesn't retain its most recently
+// parsed DNS query's []byte while it's sitting in a sync.Pool.
+// It's not useful to keep anyway: the next Start will do the same.
+func (p *dnsParser) zeroParser() { p.parser = dns.Parser{} }
+
+// parseQuery parses the query in given packet into p.Header and
+// p.Question.
+func (p *dnsParser) parseQuery(query []byte) error {
+	defer p.zeroParser()
+	var err error
+	p.Header, err = p.parser.Start(query)
 	if err != nil {
 		return err
 	}
-
-	if resp.Header.Response {
+	if p.Header.Response {
 		return errNotQuery
 	}
-
-	resp.Question, err = parser.Question()
-	if err != nil {
-		return err
-	}
-
-	return nil
+	p.Question, err = p.parser.Question()
+	return err
 }
 
 // marshalARecord serializes an A record into an active builder.
@@ -622,12 +621,13 @@ func (r *Resolver) respondReverse(query []byte, name dnsname.FQDN, resp *respons
 // respond returns a DNS response to query if it can be resolved locally.
 // Otherwise, it returns errNotOurName.
 func (r *Resolver) respond(query []byte) ([]byte, error) {
-	resp := new(response)
+	parser := dnsParserPool.Get().(*dnsParser)
+	defer dnsParserPool.Put(parser)
 
 	// ParseQuery is sufficiently fast to run on every DNS packet.
 	// This is considerably simpler than extracting the name by hand
 	// to shave off microseconds in case of delegation.
-	err := parseQuery(query, resp)
+	err := parser.parseQuery(query)
 	// We will not return this error: it is the sender's fault.
 	if err != nil {
 		if errors.Is(err, dns.ErrSectionDone) {
@@ -635,13 +635,15 @@ func (r *Resolver) respond(query []byte) ([]byte, error) {
 		} else {
 			r.logf("parseQuery(%02x): %v", query, err)
 		}
+		resp := parser.response()
 		resp.Header.RCode = dns.RCodeFormatError
 		return marshalResponse(resp)
 	}
-	rawName := resp.Question.Name.Data[:resp.Question.Name.Length]
+	rawName := parser.Question.Name.Data[:parser.Question.Name.Length]
 	name, err := dnsname.ToFQDN(rawNameToLower(rawName))
 	if err != nil {
 		// DNS packet unexpectedly contains an invalid FQDN.
+		resp := parser.response()
 		resp.Header.RCode = dns.RCodeFormatError
 		return marshalResponse(resp)
 	}
@@ -649,15 +651,17 @@ func (r *Resolver) respond(query []byte) ([]byte, error) {
 	// Always try to handle reverse lookups; delegate inside when not found.
 	// This way, queries for existent nodes do not leak,
 	// but we behave gracefully if non-Tailscale nodes exist in CGNATRange.
-	if resp.Question.Type == dns.TypePTR {
-		return r.respondReverse(query, name, resp)
+	if parser.Question.Type == dns.TypePTR {
+		return r.respondReverse(query, name, parser.response())
 	}
 
-	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
+	ip, rcode := r.resolveLocal(name, parser.Question.Type)
+	if rcode == dns.RCodeRefused {
+		return nil, errNotOurName // sentinel error return value: it requests forwarding
 	}
 
+	resp := parser.response()
+	resp.Header.RCode = rcode
+	resp.IP = ip
 	return marshalResponse(resp)
 }

net/dns/resolver/tsdns_server_test.go

@@ -66,6 +66,60 @@ func resolveToIP(ipv4, ipv6 netaddr.IP, ns string) dns.HandlerFunc {
 	}
 }
 
+// resolveToTXT returns a handler function which responds to queries of type TXT
+// it receives with the strings in txts.
+func resolveToTXT(txts []string, ednsMaxSize uint16) dns.HandlerFunc {
+	return func(w dns.ResponseWriter, req *dns.Msg) {
+		m := new(dns.Msg)
+		m.SetReply(req)
+
+		if len(req.Question) != 1 {
+			panic("not a single-question request")
+		}
+		question := req.Question[0]
+
+		if question.Qtype != dns.TypeTXT {
+			w.WriteMsg(m)
+			return
+		}
+
+		ans := &dns.TXT{
+			Hdr: dns.RR_Header{
+				Name:   question.Name,
+				Rrtype: dns.TypeTXT,
+				Class:  dns.ClassINET,
+			},
+			Txt: txts,
+		}
+
+		m.Answer = append(m.Answer, ans)
+
+		queryInfo := &dns.TXT{
+			Hdr: dns.RR_Header{
+				Name:   "query-info.test.",
+				Rrtype: dns.TypeTXT,
+				Class:  dns.ClassINET,
+			},
+		}
+
+		if edns := req.IsEdns0(); edns == nil {
+			queryInfo.Txt = []string{"EDNS=false"}
+		} else {
+			queryInfo.Txt = []string{"EDNS=true", fmt.Sprintf("maxSize=%v", edns.UDPSize())}
+		}
+
+		m.Extra = append(m.Extra, queryInfo)
+
+		if ednsMaxSize > 0 {
+			m.SetEdns0(ednsMaxSize, false)
+		}
+
+		if err := w.WriteMsg(m); err != nil {
+			panic(err)
+		}
+	}
+}
+
 var resolveToNXDOMAIN = dns.HandlerFunc(func(w dns.ResponseWriter, req *dns.Msg) {
 	m := new(dns.Msg)
 	m.SetRcode(req, dns.RcodeNameError)

net/dns/resolver/tsdns_test.go

@@ -6,14 +6,21 @@ package resolver
 
 import (
 	"bytes"
+	"encoding/hex"
 	"errors"
+	"fmt"
+	"math/rand"
 	"net"
+	"runtime"
+	"strconv"
+	"strings"
 	"testing"
 
 	dns "golang.org/x/net/dns/dnsmessage"
 	"inet.af/netaddr"
 	"tailscale.com/tstest"
 	"tailscale.com/util/dnsname"
+	"tailscale.com/wgengine/monitor"
 )
 
 var testipv4 = netaddr.MustParseIP("1.2.3.4")
@@ -27,7 +34,9 @@ var dnsCfg = Config{
 	LocalDomains: []dnsname.FQDN{"ipn.dev."},
 }
 
-func dnspacket(domain dnsname.FQDN, tp dns.Type) []byte {
+const noEdns = 0
+
+func dnspacket(domain dnsname.FQDN, tp dns.Type, ednsSize uint16) []byte {
 	var dnsHeader dns.Header
 	question := dns.Question{
 		Name:  dns.MustNewName(domain.WithTrailingDot()),
@@ -36,17 +45,44 @@ func dnspacket(domain dnsname.FQDN, tp dns.Type) []byte {
 	}
 
 	builder := dns.NewBuilder(nil, dnsHeader)
-	builder.StartQuestions()
-	builder.Question(question)
+	if err := builder.StartQuestions(); err != nil {
+		panic(err)
+	}
+	if err := builder.Question(question); err != nil {
+		panic(err)
+	}
+
+	if ednsSize != noEdns {
+		if err := builder.StartAdditionals(); err != nil {
+			panic(err)
+		}
+
+		ednsHeader := dns.ResourceHeader{
+			Name:  dns.MustNewName("."),
+			Type:  dns.TypeOPT,
+			Class: dns.Class(ednsSize),
+		}
+
+		if err := builder.OPTResource(ednsHeader, dns.OPTResource{}); err != nil {
+			panic(err)
+		}
+	}
+
 	payload, _ := builder.Finish()
 
 	return payload
 }
 
 type dnsResponse struct {
-	ip    netaddr.IP
-	name  dnsname.FQDN
-	rcode dns.RCode
+	ip               netaddr.IP
+	txt              []string
+	name             dnsname.FQDN
+	rcode            dns.RCode
+	truncated        bool
+	requestEdns      bool
+	requestEdnsSize  uint16
+	responseEdns     bool
+	responseEdnsSize uint16
 }
 
 func unpackResponse(payload []byte) (dnsResponse, error) {
@@ -67,47 +103,122 @@ func unpackResponse(payload []byte) (dnsResponse, error) {
 		return response, nil
 	}
 
+	response.truncated = h.Truncated
+	if response.truncated {
+		// TODO(#2067): Ideally, answer processing should still succeed when
+		// dealing with a truncated message, but currently when we truncate
+		// a packet, it's caused by the buffer being too small and usually that
+		// means the data runs out mid-record. dns.Parser does not like it when
+		// that happens. We can improve this by trimming off incomplete records.
+		return response, nil
+	}
+
 	err = parser.SkipAllQuestions()
 	if err != nil {
 		return response, err
 	}
 
-	ah, err := parser.AnswerHeader()
+	for {
+		ah, err := parser.AnswerHeader()
+		if err == dns.ErrSectionDone {
+			break
+		}
+		if err != nil {
+			return response, err
+		}
+
+		switch ah.Type {
+		case dns.TypeA:
+			res, err := parser.AResource()
+			if err != nil {
+				return response, err
+			}
+			response.ip = netaddr.IPv4(res.A[0], res.A[1], res.A[2], res.A[3])
+		case dns.TypeAAAA:
+			res, err := parser.AAAAResource()
+			if err != nil {
+				return response, err
+			}
+			response.ip = netaddr.IPv6Raw(res.AAAA)
+		case dns.TypeTXT:
+			res, err := parser.TXTResource()
+			if err != nil {
+				return response, err
+			}
+			response.txt = res.TXT
+		case dns.TypeNS:
+			res, err := parser.NSResource()
+			if err != nil {
+				return response, err
+			}
+			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}")
+		}
+	}
+
+	err = parser.SkipAllAuthorities()
 	if err != nil {
 		return response, err
 	}
 
-	switch ah.Type {
-	case dns.TypeA:
-		res, err := parser.AResource()
+	for {
+		ah, err := parser.AdditionalHeader()
+		if err == dns.ErrSectionDone {
+			break
+		}
 		if err != nil {
 			return response, err
 		}
-		response.ip = netaddr.IPv4(res.A[0], res.A[1], res.A[2], res.A[3])
-	case dns.TypeAAAA:
-		res, err := parser.AAAAResource()
-		if err != nil {
-			return response, err
+
+		switch ah.Type {
+		case dns.TypeOPT:
+			_, err := parser.OPTResource()
+			if err != nil {
+				return response, err
+			}
+			response.responseEdns = true
+			response.responseEdnsSize = uint16(ah.Class)
+		case dns.TypeTXT:
+			res, err := parser.TXTResource()
+			if err != nil {
+				return response, err
+			}
+			switch ah.Name.String() {
+			case "query-info.test.":
+				for _, msg := range res.TXT {
+					s := strings.SplitN(msg, "=", 2)
+					if len(s) != 2 {
+						continue
+					}
+					switch s[0] {
+					case "EDNS":
+						response.requestEdns, err = strconv.ParseBool(s[1])
+						if err != nil {
+							return response, err
+						}
+					case "maxSize":
+						sz, err := strconv.ParseUint(s[1], 10, 16)
+						if err != nil {
+							return response, err
+						}
+						response.requestEdnsSize = uint16(sz)
+					}
+				}
+			}
 		}
-		response.ip = netaddr.IPv6Raw(res.AAAA)
-	case dns.TypeNS:
-		res, err := parser.NSResource()
-		if err != nil {
-			return response, err
-		}
-		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}")
 	}
 
 	return response, nil
 }
 
 func syncRespond(r *Resolver, query []byte) ([]byte, error) {
-	r.EnqueueRequest(query, netaddr.IPPort{})
+	if err := r.EnqueueRequest(query, netaddr.IPPort{}); err != nil {
+		return nil, fmt.Errorf("EnqueueRequest: %w", err)
+	}
 	payload, _, err := r.NextResponse()
 	return payload, err
 }
@@ -190,8 +301,12 @@ func TestRDNSNameToIPv6(t *testing.T) {
 	}
 }
 
+func newResolver(t testing.TB) *Resolver {
+	return New(t.Logf, nil /* no link monitor */, nil /* no link selector */)
+}
+
 func TestResolveLocal(t *testing.T) {
-	r := New(t.Logf, nil)
+	r := newResolver(t)
 	defer r.Close()
 
 	r.SetConfig(dnsCfg)
@@ -231,7 +346,7 @@ func TestResolveLocal(t *testing.T) {
 }
 
 func TestResolveLocalReverse(t *testing.T) {
-	r := New(t.Logf, nil)
+	r := newResolver(t)
 	defer r.Close()
 
 	r.SetConfig(dnsCfg)
@@ -269,6 +384,32 @@ func ipv6Works() bool {
 	return true
 }
 
+func generateTXT(size int, source rand.Source) []string {
+	const sizePerTXT = 120
+
+	if size%2 != 0 {
+		panic("even lengths only")
+	}
+
+	rng := rand.New(source)
+
+	txts := make([]string, 0, size/sizePerTXT+1)
+
+	raw := make([]byte, sizePerTXT/2)
+
+	rem := size
+	for ; rem > sizePerTXT; rem -= sizePerTXT {
+		rng.Read(raw)
+		txts = append(txts, hex.EncodeToString(raw))
+	}
+	if rem > 0 {
+		rng.Read(raw[:rem/2])
+		txts = append(txts, hex.EncodeToString(raw[:rem/2]))
+	}
+
+	return txts
+}
+
 func TestDelegate(t *testing.T) {
 	tstest.ResourceCheck(t)
 
@@ -276,16 +417,43 @@ func TestDelegate(t *testing.T) {
 		t.Skip("skipping test that requires localhost IPv6")
 	}
 
-	v4server := serveDNS(t, "127.0.0.1:0",
-		"test.site.", resolveToIP(testipv4, testipv6, "dns.test.site."),
-		"nxdomain.site.", resolveToNXDOMAIN)
+	randSource := rand.NewSource(4)
+
+	// smallTXT does not require EDNS
+	smallTXT := generateTXT(300, randSource)
+
+	// medTXT and largeTXT are responses that require EDNS but we would like to
+	// support these sizes of response without truncation because they are
+	// moderately common.
+	medTXT := generateTXT(1200, randSource)
+	largeTXT := generateTXT(3900, randSource)
+
+	// xlargeTXT is slightly above the maximum response size that we support,
+	// so there should be truncation.
+	xlargeTXT := generateTXT(5000, randSource)
+
+	// hugeTXT is significantly larger than any typical MTU and will require
+	// significant fragmentation. For buffer management reasons, we do not
+	// intend to handle responses this large, so there should be truncation.
+	hugeTXT := generateTXT(64000, randSource)
+
+	records := []interface{}{
+		"test.site.",
+		resolveToIP(testipv4, testipv6, "dns.test.site."),
+		"nxdomain.site.", resolveToNXDOMAIN,
+		"small.txt.", resolveToTXT(smallTXT, noEdns),
+		"smalledns.txt.", resolveToTXT(smallTXT, 512),
+		"med.txt.", resolveToTXT(medTXT, 1500),
+		"large.txt.", resolveToTXT(largeTXT, maxResponseBytes),
+		"xlarge.txt.", resolveToTXT(xlargeTXT, 8000),
+		"huge.txt.", resolveToTXT(hugeTXT, 65527),
+	}
+	v4server := serveDNS(t, "127.0.0.1:0", records...)
 	defer v4server.Shutdown()
-	v6server := serveDNS(t, "[::1]:0",
-		"test.site.", resolveToIP(testipv4, testipv6, "dns.test.site."),
-		"nxdomain.site.", resolveToNXDOMAIN)
+	v6server := serveDNS(t, "[::1]:0", records...)
 	defer v6server.Shutdown()
 
-	r := New(t.Logf, nil)
+	r := newResolver(t)
 	defer r.Close()
 
 	cfg := dnsCfg
@@ -304,28 +472,92 @@ func TestDelegate(t *testing.T) {
 	}{
 		{
 			"ipv4",
-			dnspacket("test.site.", dns.TypeA),
+			dnspacket("test.site.", dns.TypeA, noEdns),
 			dnsResponse{ip: testipv4, rcode: dns.RCodeSuccess},
 		},
 		{
 			"ipv6",
-			dnspacket("test.site.", dns.TypeAAAA),
+			dnspacket("test.site.", dns.TypeAAAA, noEdns),
 			dnsResponse{ip: testipv6, rcode: dns.RCodeSuccess},
 		},
 		{
 			"ns",
-			dnspacket("test.site.", dns.TypeNS),
+			dnspacket("test.site.", dns.TypeNS, noEdns),
 			dnsResponse{name: "dns.test.site.", rcode: dns.RCodeSuccess},
 		},
 		{
 			"nxdomain",
-			dnspacket("nxdomain.site.", dns.TypeA),
+			dnspacket("nxdomain.site.", dns.TypeA, noEdns),
 			dnsResponse{rcode: dns.RCodeNameError},
 		},
+		{
+			"smalltxt",
+			dnspacket("small.txt.", dns.TypeTXT, 8000),
+			dnsResponse{txt: smallTXT, rcode: dns.RCodeSuccess, requestEdns: true, requestEdnsSize: maxResponseBytes},
+		},
+		{
+			"smalltxtedns",
+			dnspacket("smalledns.txt.", dns.TypeTXT, 512),
+			dnsResponse{
+				txt:              smallTXT,
+				rcode:            dns.RCodeSuccess,
+				requestEdns:      true,
+				requestEdnsSize:  512,
+				responseEdns:     true,
+				responseEdnsSize: 512,
+			},
+		},
+		{
+			"medtxt",
+			dnspacket("med.txt.", dns.TypeTXT, 2000),
+			dnsResponse{
+				txt:              medTXT,
+				rcode:            dns.RCodeSuccess,
+				requestEdns:      true,
+				requestEdnsSize:  2000,
+				responseEdns:     true,
+				responseEdnsSize: 1500,
+			},
+		},
+		{
+			"largetxt",
+			dnspacket("large.txt.", dns.TypeTXT, maxResponseBytes),
+			dnsResponse{
+				txt:              largeTXT,
+				rcode:            dns.RCodeSuccess,
+				requestEdns:      true,
+				requestEdnsSize:  maxResponseBytes,
+				responseEdns:     true,
+				responseEdnsSize: maxResponseBytes,
+			},
+		},
+		{
+			"xlargetxt",
+			dnspacket("xlarge.txt.", dns.TypeTXT, 8000),
+			dnsResponse{
+				rcode:     dns.RCodeSuccess,
+				truncated: true,
+				// request/response EDNS fields will be unset because of
+				// they were truncated away
+			},
+		},
+		{
+			"hugetxt",
+			dnspacket("huge.txt.", dns.TypeTXT, 8000),
+			dnsResponse{
+				rcode:     dns.RCodeSuccess,
+				truncated: true,
+				// request/response EDNS fields will be unset because of
+				// they were truncated away
+			},
+		},
 	}
 
 	for _, tt := range tests {
 		t.Run(tt.title, func(t *testing.T) {
+			if tt.title == "hugetxt" && runtime.GOOS == "darwin" {
+				t.Skip("known to not work on macOS: https://github.com/tailscale/tailscale/issues/2229")
+			}
 			payload, err := syncRespond(r, tt.query)
 			if err != nil {
 				t.Errorf("err = %v; want nil", err)
@@ -345,6 +577,27 @@ func TestDelegate(t *testing.T) {
 			if response.name != tt.response.name {
 				t.Errorf("name = %v; want %v", response.name, tt.response.name)
 			}
+			if len(response.txt) != len(tt.response.txt) {
+				t.Errorf("%v txt records, want %v txt records", len(response.txt), len(tt.response.txt))
+			} else {
+				for i := range response.txt {
+					if response.txt[i] != tt.response.txt[i] {
+						t.Errorf("txt record %v is %s, want %s", i, response.txt[i], tt.response.txt[i])
+					}
+				}
+			}
+			if response.requestEdns != tt.response.requestEdns {
+				t.Errorf("requestEdns = %v; want %v", response.requestEdns, tt.response.requestEdns)
+			}
+			if response.requestEdnsSize != tt.response.requestEdnsSize {
+				t.Errorf("requestEdnsSize = %v; want %v", response.requestEdnsSize, tt.response.requestEdnsSize)
+			}
+			if response.responseEdns != tt.response.responseEdns {
+				t.Errorf("responseEdns = %v; want %v", response.requestEdns, tt.response.requestEdns)
+			}
+			if response.responseEdnsSize != tt.response.responseEdnsSize {
+				t.Errorf("responseEdnsSize = %v; want %v", response.responseEdnsSize, tt.response.responseEdnsSize)
+			}
 		})
 	}
 }
@@ -360,7 +613,7 @@ func TestDelegateSplitRoute(t *testing.T) {
 		"test.other.", resolveToIP(test4, test6, "dns.other."))
 	defer server2.Shutdown()
 
-	r := New(t.Logf, nil)
+	r := newResolver(t)
 	defer r.Close()
 
 	cfg := dnsCfg
@@ -377,12 +630,12 @@ func TestDelegateSplitRoute(t *testing.T) {
 	}{
 		{
 			"general",
-			dnspacket("test.site.", dns.TypeA),
+			dnspacket("test.site.", dns.TypeA, noEdns),
 			dnsResponse{ip: testipv4, rcode: dns.RCodeSuccess},
 		},
 		{
 			"override",
-			dnspacket("test.other.", dns.TypeA),
+			dnspacket("test.other.", dns.TypeA, noEdns),
 			dnsResponse{ip: test4, rcode: dns.RCodeSuccess},
 		},
 	}
@@ -417,7 +670,7 @@ func TestDelegateCollision(t *testing.T) {
 		"test.site.", resolveToIP(testipv4, testipv6, "dns.test.site."))
 	defer server.Shutdown()
 
-	r := New(t.Logf, nil)
+	r := newResolver(t)
 	defer r.Close()
 
 	cfg := dnsCfg
@@ -439,7 +692,7 @@ func TestDelegateCollision(t *testing.T) {
 
 	// packets will have the same dns txid.
 	for _, p := range packets {
-		payload := dnspacket(p.qname, p.qtype)
+		payload := dnspacket(p.qname, p.qtype, noEdns)
 		err := r.EnqueueRequest(payload, p.addr)
 		if err != nil {
 			t.Error(err)
@@ -631,7 +884,7 @@ var emptyResponse = []byte{
 }
 
 func TestFull(t *testing.T) {
-	r := New(t.Logf, nil)
+	r := newResolver(t)
 	defer r.Close()
 
 	r.SetConfig(dnsCfg)
@@ -642,15 +895,15 @@ func TestFull(t *testing.T) {
 		request  []byte
 		response []byte
 	}{
-		{"all", dnspacket("test1.ipn.dev.", dns.TypeALL), allResponse},
-		{"ipv4", dnspacket("test1.ipn.dev.", dns.TypeA), ipv4Response},
-		{"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},
-		{"ptr4", dnspacket("4.3.2.1.in-addr.arpa.", dns.TypePTR), ptrResponse},
+		{"all", dnspacket("test1.ipn.dev.", dns.TypeALL, noEdns), allResponse},
+		{"ipv4", dnspacket("test1.ipn.dev.", dns.TypeA, noEdns), ipv4Response},
+		{"ipv6", dnspacket("test2.ipn.dev.", dns.TypeAAAA, noEdns), ipv6Response},
+		{"no-ipv6", dnspacket("test1.ipn.dev.", dns.TypeAAAA, noEdns), emptyResponse},
+		{"upper", dnspacket("TEST1.IPN.DEV.", dns.TypeA, noEdns), ipv4UppercaseResponse},
+		{"ptr4", dnspacket("4.3.2.1.in-addr.arpa.", dns.TypePTR, noEdns), 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},
+			dns.TypePTR, noEdns), ptrResponse6},
+		{"nxdomain", dnspacket("test3.ipn.dev.", dns.TypeA, noEdns), nxdomainResponse},
 	}
 
 	for _, tt := range tests {
@@ -667,7 +920,7 @@ func TestFull(t *testing.T) {
 }
 
 func TestAllocs(t *testing.T) {
-	r := New(t.Logf, nil)
+	r := newResolver(t)
 	defer r.Close()
 	r.SetConfig(dnsCfg)
 
@@ -679,9 +932,9 @@ func TestAllocs(t *testing.T) {
 		want  int
 	}{
 		// Name lowercasing and response slice created by dns.NewBuilder.
-		{"forward", dnspacket("test1.ipn.dev.", dns.TypeA), 2},
+		{"forward", dnspacket("test1.ipn.dev.", dns.TypeA, noEdns), 2},
 		// 3 extra allocs in rdnsNameToIPv4 and one in marshalPTRRecord (dns.NewName).
-		{"reverse", dnspacket("4.3.2.1.in-addr.arpa.", dns.TypePTR), 5},
+		{"reverse", dnspacket("4.3.2.1.in-addr.arpa.", dns.TypePTR, noEdns), 5},
 	}
 
 	for _, tt := range tests {
@@ -721,7 +974,7 @@ func BenchmarkFull(b *testing.B) {
 		"test.site.", resolveToIP(testipv4, testipv6, "dns.test.site."))
 	defer server.Shutdown()
 
-	r := New(b.Logf, nil)
+	r := newResolver(b)
 	defer r.Close()
 
 	cfg := dnsCfg
@@ -735,9 +988,9 @@ func BenchmarkFull(b *testing.B) {
 		name    string
 		request []byte
 	}{
-		{"forward", dnspacket("test1.ipn.dev.", dns.TypeA)},
-		{"reverse", dnspacket("4.3.2.1.in-addr.arpa.", dns.TypePTR)},
-		{"delegated", dnspacket("test.site.", dns.TypeA)},
+		{"forward", dnspacket("test1.ipn.dev.", dns.TypeA, noEdns)},
+		{"reverse", dnspacket("4.3.2.1.in-addr.arpa.", dns.TypePTR, noEdns)},
+		{"delegated", dnspacket("test.site.", dns.TypeA, noEdns)},
 	}
 
 	for _, tt := range tests {
@@ -758,3 +1011,58 @@ func TestMarshalResponseFormatError(t *testing.T) {
 	}
 	t.Logf("response: %q", v)
 }
+
+func TestForwardLinkSelection(t *testing.T) {
+	old := initListenConfig
+	defer func() { initListenConfig = old }()
+
+	configCall := make(chan string, 1)
+	initListenConfig = func(nc *net.ListenConfig, mon *monitor.Mon, tunName string) error {
+		select {
+		case configCall <- tunName:
+			return nil
+		default:
+			t.Error("buffer full")
+			return errors.New("buffer full")
+		}
+	}
+
+	// specialIP is some IP we pretend that our link selector
+	// routes differently.
+	specialIP := netaddr.IPv4(1, 2, 3, 4)
+
+	fwd := newForwarder(t.Logf, nil, nil, linkSelFunc(func(ip netaddr.IP) string {
+		if ip == netaddr.IPv4(1, 2, 3, 4) {
+			return "special"
+		}
+		return ""
+	}))
+
+	// Test non-special IP.
+	if got, err := fwd.packetListener(netaddr.IP{}); err != nil {
+		t.Fatal(err)
+	} else if got != stdNetPacketListener {
+		t.Errorf("for IP zero value, didn't get expected packet listener")
+	}
+	select {
+	case v := <-configCall:
+		t.Errorf("unexpected ListenConfig call, with tunName %q", v)
+	default:
+	}
+
+	// Test that our special IP generates a call to initListenConfig.
+	if got, err := fwd.packetListener(specialIP); err != nil {
+		t.Fatal(err)
+	} else if got == stdNetPacketListener {
+		t.Errorf("special IP returned std packet listener; expected unique one")
+	}
+	if v, ok := <-configCall; !ok {
+		t.Errorf("didn't get ListenConfig call")
+	} else if v != "special" {
+		t.Errorf("got tunName %q; want 'special'", v)
+	}
+}
+
+type linkSelFunc func(ip netaddr.IP) string
+
+func (f linkSelFunc) PickLink(ip netaddr.IP) string { return f(ip) }

net/dns/wsl_windows.go

@@ -0,0 +1,169 @@
+// 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"
+	"os/exec"
+	"regexp"
+	"strings"
+	"syscall"
+
+	"tailscale.com/types/logger"
+)
+
+func wslExists() bool {
+	cmd := exec.Command("wsl.exe", "-l", "-q")
+	cmd.SysProcAttr = &syscall.SysProcAttr{HideWindow: true}
+	b, err := cmd.CombinedOutput()
+	if err != nil {
+		return false
+	}
+	return strings.TrimSpace(string(b)) != "" // linux is installed
+}
+
+type wslManager struct {
+	logf logger.Logf
+	m    directManager
+}
+
+func newWSLManager(logf logger.Logf) *wslManager {
+	return &wslManager{
+		logf: logf,
+		m:    newDirectManagerOnFS(wslFS{user: "root"}),
+	}
+}
+
+func (m wslManager) SetDNS(cfg OSConfig) error {
+	if !cfg.IsZero() {
+		changed, err := m.setWSLConf()
+		if err != nil {
+			m.logf("WSL SetDNS wsl.conf: %v", err)
+		} else if changed {
+			// What's this? So glad you asked.
+			//
+			// WSL2 writes the /etc/resolv.conf.
+			// It is aggressive about it. Every time you execute wsl.exe,
+			// it writes it. (Opening a terminal is done by running wsl.exe.)
+			// You can turn this off using /etc/wsl.conf! But: this wsl.conf
+			// file is only parsed when the VM boots up. To do that, we
+			// have to shut down WSL2.
+			//
+			// So we do it here, before we call wsl.exe to write resolv.conf.
+			if b, err := wslCommand("--shutdown").CombinedOutput(); err != nil {
+				m.logf("WSL SetDNS shutdown: %v: %s", err, b)
+			}
+		}
+
+		// TODO(crawshaw): use "ip r" to get the default gateway and
+		// add it to the nameserver list. This is the default resolv.conf
+		// entry for WSL2, and that way it gets used as a backup when
+		// tailscale is off.
+	}
+	if err := m.m.SetDNS(cfg); err != nil {
+		m.logf("WSL SetDNS: %v", err)
+	}
+	return nil
+}
+
+const wslConf = "/etc/wsl.conf"
+const wslConfSection = `[network]
+generateResolvConf = false  # added by tailscale
+`
+
+func (m wslManager) setWSLConf() (changed bool, err error) {
+	b, err := m.m.fs.ReadFile(wslConf)
+	m.logf("setWSLConf XXX b=%s, err=%v", b, err)
+	if err != nil && !os.IsNotExist(err) {
+		return false, err
+	}
+	ini := parseIni(string(b))
+	m.logf("wsl.conf ini: %v", ini)
+	if v := ini["network"]["generateResolvConf"]; v == "" {
+		b = append(b, wslConfSection...)
+		if err := m.m.fs.WriteFile(wslConf, b, 0644); err != nil {
+			return false, err
+		}
+		return true, nil
+	}
+	return false, nil
+}
+
+func parseIni(data string) map[string]map[string]string {
+	sectionRE := regexp.MustCompile(`^\[([^]]+)\]\s*`)
+	kvRE := regexp.MustCompile(`^\s*(\w*)\s*=\s*([^#]*)\s*`)
+
+	ini := map[string]map[string]string{}
+	var section string
+	for _, line := range strings.Split(data, "\n") {
+		if res := sectionRE.FindStringSubmatch(line); len(res) > 1 {
+			section = res[1]
+			ini[section] = map[string]string{}
+		} else if res := kvRE.FindStringSubmatch(line); len(res) > 2 {
+			k, v := res[1], res[2]
+			ini[section][k] = v
+		}
+	}
+	return ini
+}
+
+func (m wslManager) SupportsSplitDNS() bool { return false }
+func (m wslManager) Close() error           { return m.SetDNS(OSConfig{}) }
+
+// wslFS is a pinholeFS implemented on top of wsl.exe.
+//
+// We access WSL2 file systems via wsl.exe instead of \\wsl$\ because
+// the netpath appears to operate as the standard user, not root.
+type wslFS struct {
+	user string
+}
+
+func (fs wslFS) Stat(name string) (isRegular bool, err error) {
+	err = wslCommand("-u", fs.user, "-e", "test", "-f", name).Run()
+	if ee, _ := err.(*exec.ExitError); ee != nil {
+		if ee.ExitCode() == 1 {
+			return false, os.ErrNotExist
+		}
+		return false, err
+	}
+	return true, nil
+}
+
+func (fs wslFS) Rename(oldName, newName string) error {
+	return wslCommand("-u", fs.user, "-e", "mv", oldName, newName).Run()
+}
+
+func (fs wslFS) Symlink(oldName, newName string) error {
+	return wslCommand("-u", fs.user, "-e", "ln", "-s", "-f", oldName, newName).Run()
+}
+
+func (fs wslFS) Remove(name string) error { return wslCommand("-u", fs.user, "-e", "rm", name).Run() }
+
+func (fs wslFS) ReadFile(name string) ([]byte, error) {
+	b, err := wslCommand("-u", fs.user, "-e", "cat", name).CombinedOutput()
+	if ee, _ := err.(*exec.ExitError); ee != nil && ee.ExitCode() == 1 {
+		return nil, os.ErrNotExist
+	}
+	return b, err
+}
+
+func (fs wslFS) WriteFile(name string, contents []byte, perm os.FileMode) error {
+	cmd := wslCommand("-u", fs.user, "-e", "tee", name)
+	cmd.Stdin = bytes.NewReader(contents)
+	cmd.Stdout = nil
+	if err := cmd.Run(); err != nil {
+		return err
+	}
+	return wslCommand("-u", fs.user, "-e", "chmod", "0644", name).Run() // TODO perm
+}
+
+func wslCommand(args ...string) *exec.Cmd {
+	cmd := exec.Command("wsl.exe", args...)
+	fmt.Printf("wslCommand: %v\n", cmd.Args)
+	cmd.SysProcAttr = &syscall.SysProcAttr{HideWindow: true}
+	return cmd
+}

net/interfaces/interfaces.go

@@ -15,6 +15,7 @@ import (
 	"strings"
 
 	"inet.af/netaddr"
+	"tailscale.com/hostinfo"
 	"tailscale.com/net/tsaddr"
 	"tailscale.com/net/tshttpproxy"
 )
@@ -81,13 +82,16 @@ func isProblematicInterface(nif *net.Interface) bool {
 }
 
 // LocalAddresses returns the machine's IP addresses, separated by
-// whether they're loopback addresses.
+// whether they're loopback addresses. If there are no regular addresses
+// it will return any IPv4 linklocal or IPv6 unique local addresses because we
+// know of environments where these are used with NAT to provide connectivity.
 func LocalAddresses() (regular, loopback []netaddr.IP, err error) {
 	// TODO(crawshaw): don't serve interface addresses that we are routing
 	ifaces, err := net.Interfaces()
 	if err != nil {
 		return nil, nil, err
 	}
+	var regular4, regular6, linklocal4, ula6 []netaddr.IP
 	for i := range ifaces {
 		iface := &ifaces[i]
 		if !isUp(iface) || isProblematicInterface(iface) {
@@ -117,17 +121,44 @@ func LocalAddresses() (regular, loopback []netaddr.IP, err error) {
 				if tsaddr.IsTailscaleIP(ip) {
 					continue
 				}
-				if ip.IsLinkLocalUnicast() {
-					continue
-				}
 				if ip.IsLoopback() || ifcIsLoopback {
 					loopback = append(loopback, ip)
+				} else if ip.IsLinkLocalUnicast() {
+					if ip.Is4() {
+						linklocal4 = append(linklocal4, ip)
+					}
+
+					// We know of no cases where the IPv6 fe80:: addresses
+					// are used to provide WAN connectivity. It is also very
+					// common for users to have no IPv6 WAN connectivity,
+					// but their OS supports IPv6 so they have an fe80::
+					// address. We don't want to report all of those
+					// IPv6 LL to Control.
+				} else if ip.Is6() && tsaddr.IsULA(ip) {
+					// Google Cloud Run uses NAT with IPv6 Unique
+					// Local Addresses to provide IPv6 connectivity.
+					ula6 = append(ula6, ip)
 				} else {
-					regular = append(regular, ip)
+					if ip.Is4() {
+						regular4 = append(regular4, ip)
+					} else {
+						regular6 = append(regular6, ip)
+					}
 				}
 			}
 		}
 	}
+	if len(regular4) == 0 && len(regular6) == 0 {
+		// if we have no usable IP addresses then be willing to accept
+		// addresses we otherwise wouldn't, like:
+		//   + 169.254.x.x (AWS Lambda uses NAT with these)
+		//   + IPv6 ULA (Google Cloud Run uses these with address translation)
+		if hostinfo.GetEnvType() == hostinfo.AWSLambda {
+			regular4 = linklocal4
+		}
+		regular6 = ula6
+	}
+	regular = append(regular4, regular6...)
 	sortIPs(regular)
 	sortIPs(loopback)
 	return regular, loopback, nil
@@ -213,9 +244,9 @@ type State struct {
 	InterfaceIPs map[string][]netaddr.IPPrefix
 	Interface    map[string]Interface
 
-	// HaveV6Global is whether this machine has an IPv6 global address
-	// on some non-Tailscale interface that's up.
-	HaveV6Global bool
+	// HaveV6 is whether this machine has an IPv6 Global or Unique Local Address
+	// which might provide connectivity on a non-Tailscale interface that's up.
+	HaveV6 bool
 
 	// HaveV4 is whether the machine has some non-localhost,
 	// non-link-local IPv4 address on a non-Tailscale interface that's up.
@@ -289,7 +320,7 @@ func (s *State) String() string {
 	if s.PAC != "" {
 		fmt.Fprintf(&sb, " pac=%s", s.PAC)
 	}
-	fmt.Fprintf(&sb, " v4=%v v6global=%v}", s.HaveV4, s.HaveV6Global)
+	fmt.Fprintf(&sb, " v4=%v v6=%v}", s.HaveV4, s.HaveV6)
 	return sb.String()
 }
 
@@ -302,7 +333,7 @@ func (s *State) EqualFiltered(s2 *State, filter func(i Interface, ips []netaddr.
 	if s == nil || s2 == nil {
 		return false
 	}
-	if s.HaveV6Global != s2.HaveV6Global ||
+	if s.HaveV6 != s2.HaveV6 ||
 		s.HaveV4 != s2.HaveV4 ||
 		s.IsExpensive != s2.IsExpensive ||
 		s.DefaultRouteInterface != s2.DefaultRouteInterface ||
@@ -362,7 +393,7 @@ func (s *State) HasPAC() bool { return s != nil && s.PAC != "" }
 
 // AnyInterfaceUp reports whether any interface seems like it has Internet access.
 func (s *State) AnyInterfaceUp() bool {
-	return s != nil && (s.HaveV4 || s.HaveV6Global)
+	return s != nil && (s.HaveV4 || s.HaveV6)
 }
 
 func hasTailscaleIP(pfxs []netaddr.IPPrefix) bool {
@@ -407,11 +438,11 @@ func GetState() (*State, error) {
 			return
 		}
 		for _, pfx := range pfxs {
-			if pfx.IP().IsLoopback() || pfx.IP().IsLinkLocalUnicast() {
+			if pfx.IP().IsLoopback() {
 				continue
 			}
-			s.HaveV6Global = s.HaveV6Global || isGlobalV6(pfx.IP())
-			s.HaveV4 = s.HaveV4 || pfx.IP().Is4()
+			s.HaveV6 = s.HaveV6 || isUsableV6(pfx.IP())
+			s.HaveV4 = s.HaveV4 || isUsableV4(pfx.IP())
 		}
 	}); err != nil {
 		return nil, err
@@ -503,7 +534,25 @@ func isPrivateIP(ip netaddr.IP) bool {
 	return private1.Contains(ip) || private2.Contains(ip) || private3.Contains(ip)
 }
 
-func isGlobalV6(ip netaddr.IP) bool {
+// isUsableV4 reports whether ip is a usable IPv4 address which could
+// conceivably be used to get Internet connectivity. Globally routable and
+// private IPv4 addresses are always Usable, and link local 169.254.x.x
+// addresses are in some environments.
+func isUsableV4(ip netaddr.IP) bool {
+	if !ip.Is4() || ip.IsLoopback() {
+		return false
+	}
+	if ip.IsLinkLocalUnicast() {
+		return hostinfo.GetEnvType() == hostinfo.AWSLambda
+	}
+	return true
+}
+
+// isUsableV6 reports whether ip is a usable IPv6 address which could
+// conceivably be used to get Internet connectivity. Globally routable
+// IPv6 addresses are always Usable, and Unique Local Addresses
+// (fc00::/7) are in some environments used with address translation.
+func isUsableV6(ip netaddr.IP) bool {
 	return v6Global1.Contains(ip) ||
 		(tsaddr.IsULA(ip) && !tsaddr.TailscaleULARange().Contains(ip))
 }

net/interfaces/interfaces_default_route_test.go

@@ -2,15 +2,11 @@
 // Use of this source code is governed by a BSD-style
 // license that can be found in the LICENSE file.
 
-// +build linux,!redo
+// +build linux darwin,!ts_macext
 
 package interfaces
 
 import (
-	"fmt"
-	"io/ioutil"
-	"os"
-	"path/filepath"
 	"testing"
 )
 
@@ -23,64 +19,3 @@ func TestDefaultRouteInterface(t *testing.T) {
 	}
 	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_test.go

@@ -4,7 +4,74 @@
 
 package interfaces
 
-import "testing"
+import (
+	"fmt"
+	"io/ioutil"
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+// 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)
+	}
+}
 
 func BenchmarkDefaultRouteInterface(b *testing.B) {
 	b.ReportAllocs()

net/interfaces/interfaces_test.go

@@ -46,7 +46,7 @@ func TestLikelyHomeRouterIP(t *testing.T) {
 	t.Logf("myIP = %v; gw = %v", my, gw)
 }
 
-func TestIsGlobalV6(t *testing.T) {
+func TestIsUsableV6(t *testing.T) {
 	tests := []struct {
 		name string
 		ip   string
@@ -61,8 +61,8 @@ func TestIsGlobalV6(t *testing.T) {
 	}
 
 	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)
+		if got := isUsableV6(netaddr.MustParseIP(test.ip)); got != test.want {
+			t.Errorf("isUsableV6(%s) = %v, want %v", test.name, got, test.want)
 		}
 	}
 }

net/netcheck/netcheck.go

@@ -336,7 +336,7 @@ func makeProbePlan(dm *tailcfg.DERPMap, ifState *interfaces.State, last *Report)
 	if last == nil || len(last.RegionLatency) == 0 {
 		return makeProbePlanInitial(dm, ifState)
 	}
-	have6if := ifState.HaveV6Global
+	have6if := ifState.HaveV6
 	have4if := ifState.HaveV4
 	plan = make(probePlan)
 	if !have4if && !have6if {
@@ -425,7 +425,7 @@ func makeProbePlanInitial(dm *tailcfg.DERPMap, ifState *interfaces.State) (plan
 			if ifState.HaveV4 && nodeMight4(n) {
 				p4 = append(p4, probe{delay: delay, node: n.Name, proto: probeIPv4})
 			}
-			if ifState.HaveV6Global && nodeMight6(n) {
+			if ifState.HaveV6 && nodeMight6(n) {
 				p6 = append(p6, probe{delay: delay, node: n.Name, proto: probeIPv6})
 			}
 		}
@@ -808,7 +808,7 @@ func (c *Client) GetReport(ctx context.Context, dm *tailcfg.DERPMap) (*Report, e
 		go c.readPackets(ctx, u4)
 	}
 
-	if ifState.HaveV6Global {
+	if ifState.HaveV6 {
 		if f := c.GetSTUNConn6; f != nil {
 			rs.pc6 = f()
 		} else {

net/netcheck/netcheck_test.go

@@ -443,8 +443,8 @@ func TestMakeProbePlan(t *testing.T) {
 	for _, tt := range tests {
 		t.Run(tt.name, func(t *testing.T) {
 			ifState := &interfaces.State{
-				HaveV6Global: tt.have6if,
-				HaveV4:       !tt.no4,
+				HaveV6: tt.have6if,
+				HaveV4: !tt.no4,
 			}
 			got := makeProbePlan(tt.dm, ifState, tt.last)
 			if !reflect.DeepEqual(got, tt.want) {

net/netns/netns_android.go

@@ -0,0 +1,64 @@
+// 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 android
+
+package netns
+
+import (
+	"fmt"
+	"sync"
+	"syscall"
+)
+
+var (
+	androidProtectFuncMu sync.Mutex
+	androidProtectFunc   func(fd int) error
+)
+
+// SetAndroidProtectFunc register a func that Android provides that JNI calls into
+// https://developer.android.com/reference/android/net/VpnService#protect(int)
+// which is documented as:
+//
+// "Protect a socket from VPN connections. After protecting, data sent
+// through this socket will go directly to the underlying network, so
+// its traffic will not be forwarded through the VPN. This method is
+// useful if some connections need to be kept outside of VPN. For
+// example, a VPN tunnel should protect itself if its destination is
+// covered by VPN routes. Otherwise its outgoing packets will be sent
+// back to the VPN interface and cause an infinite loop. This method
+// will fail if the application is not prepared or is revoked."
+//
+// A nil func disables the use the hook.
+//
+// This indirection is necessary because this is the supported, stable
+// interface to use on Android, and doing the sockopts to set the
+// fwmark return errors on Android. The actual implementation of
+// VpnService.protect ends up doing an IPC to another process on
+// Android, asking for the fwmark to be set.
+func SetAndroidProtectFunc(f func(fd int) error) {
+	androidProtectFuncMu.Lock()
+	defer androidProtectFuncMu.Unlock()
+	androidProtectFunc = f
+}
+
+// control marks c as necessary to dial in a separate network namespace.
+//
+// It's intentionally the same signature as net.Dialer.Control
+// and net.ListenConfig.Control.
+func control(network, address string, c syscall.RawConn) error {
+	var sockErr error
+	err := c.Control(func(fd uintptr) {
+		androidProtectFuncMu.Lock()
+		f := androidProtectFunc
+		androidProtectFuncMu.Unlock()
+		if f != nil {
+			sockErr = f(int(fd))
+		}
+	})
+	if err != nil {
+		return fmt.Errorf("RawConn.Control on %T: %w", c, err)
+	}
+	return sockErr
+}

net/netns/netns_darwin_tailscaled.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.
 
-// +build darwin,!redo
+// +build darwin,!ts_macext
 
 package netns
 

net/netns/netns_default.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.
 
-// +build !linux,!windows,!darwin darwin,redo
+// +build !linux,!windows,!darwin darwin,ts_macext
 
 package netns
 

net/netns/netns_linux.go

@@ -2,6 +2,8 @@
 // Use of this source code is governed by a BSD-style
 // license that can be found in the LICENSE file.
 
+// +build linux,!android
+
 package netns
 
 import (

net/netns/netns_macios.go

@@ -0,0 +1,53 @@
+// 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 darwin ios
+
+package netns
+
+import (
+	"errors"
+	"log"
+	"net"
+	"strings"
+	"syscall"
+
+	"golang.org/x/sys/unix"
+)
+
+// SetListenConfigInterfaceIndex sets lc.Control such that sockets are bound
+// to the provided interface index.
+func SetListenConfigInterfaceIndex(lc *net.ListenConfig, ifIndex int) error {
+	if lc == nil {
+		return errors.New("nil ListenConfig")
+	}
+	if lc.Control != nil {
+		return errors.New("ListenConfig.Control already set")
+	}
+	lc.Control = func(network, address string, c syscall.RawConn) error {
+		var sockErr error
+		err := c.Control(func(fd uintptr) {
+			sockErr = bindInterface(fd, network, address, ifIndex)
+			if sockErr != nil {
+				log.Printf("netns: bind(%q, %q) on index %v: %v", network, address, ifIndex, sockErr)
+			}
+		})
+		if err != nil {
+			return err
+		}
+		return sockErr
+	}
+	return nil
+}
+
+func bindInterface(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)
+}

net/packet/packet.go

@@ -12,7 +12,6 @@ import (
 
 	"inet.af/netaddr"
 	"tailscale.com/types/ipproto"
-	"tailscale.com/types/strbuilder"
 )
 
 const unknown = ipproto.Unknown
@@ -62,36 +61,17 @@ func (p *Parsed) String() string {
 		return "Unknown{???}"
 	}
 
-	sb := strbuilder.Get()
-	sb.WriteString(p.IPProto.String())
-	sb.WriteByte('{')
-	writeIPPort(sb, p.Src)
-	sb.WriteString(" > ")
-	writeIPPort(sb, p.Dst)
-	sb.WriteByte('}')
-	return sb.String()
-}
-
-// writeIPPort writes ipp.String() into sb, with fewer allocations.
-//
-// TODO: make netaddr more efficient in this area, and retire this func.
-func writeIPPort(sb *strbuilder.Builder, ipp netaddr.IPPort) {
-	if ipp.IP().Is4() {
-		raw := ipp.IP().As4()
-		sb.WriteUint(uint64(raw[0]))
-		sb.WriteByte('.')
-		sb.WriteUint(uint64(raw[1]))
-		sb.WriteByte('.')
-		sb.WriteUint(uint64(raw[2]))
-		sb.WriteByte('.')
-		sb.WriteUint(uint64(raw[3]))
-		sb.WriteByte(':')
-	} else {
-		sb.WriteByte('[')
-		sb.WriteString(ipp.IP().String()) // TODO: faster?
-		sb.WriteString("]:")
-	}
-	sb.WriteUint(uint64(ipp.Port()))
+	// max is the maximum reasonable length of the string we are constructing.
+	// It's OK to overshoot, as the temp buffer is allocated on the stack.
+	const max = len("ICMPv6{[ffff:ffff:ffff:ffff:ffff:ffff:ffff:ffff%enp5s0]:65535 > [ffff:ffff:ffff:ffff:ffff:ffff:ffff:ffff%enp5s0]:65535}")
+	b := make([]byte, 0, max)
+	b = append(b, p.IPProto.String()...)
+	b = append(b, '{')
+	b = p.Src.AppendTo(b)
+	b = append(b, ' ', '>', ' ')
+	b = p.Dst.AppendTo(b)
+	b = append(b, '}')
+	return string(b)
 }
 
 // Decode extracts data from the packet in b into q.

net/packet/packet_test.go

@@ -378,11 +378,9 @@ func TestParsedString(t *testing.T) {
 		})
 	}
 
-	var sink string
 	allocs := testing.AllocsPerRun(1000, func() {
-		sink = tests[0].qdecode.String()
+		sinkString = tests[0].qdecode.String()
 	})
-	_ = sink
 	if allocs != 1 {
 		t.Errorf("allocs = %v; want 1", allocs)
 	}
@@ -532,3 +530,33 @@ func TestMarshalResponse(t *testing.T) {
 		})
 	}
 }
+
+var sinkString string
+
+func BenchmarkString(b *testing.B) {
+	benches := []struct {
+		name string
+		buf  []byte
+	}{
+		{"tcp4", tcp4PacketBuffer},
+		{"tcp6", tcp6RequestBuffer},
+		{"udp4", udp4RequestBuffer},
+		{"udp6", udp6RequestBuffer},
+		{"icmp4", icmp4RequestBuffer},
+		{"icmp6", icmp6PacketBuffer},
+		{"igmp", igmpPacketBuffer},
+		{"unknown", unknownPacketBuffer},
+	}
+
+	for _, bench := range benches {
+		b.Run(bench.name, func(b *testing.B) {
+			b.ReportAllocs()
+			var p Parsed
+			p.Decode(bench.buf)
+			b.ResetTimer()
+			for i := 0; i < b.N; i++ {
+				sinkString = p.String()
+			}
+		})
+	}
+}

net/portmapper/portmapper.go

@@ -577,8 +577,6 @@ func pcpAnnounceRequest(myIP netaddr.IP) []byte {
 	return pkt
 }
 
-//lint:ignore U1000 moved this code from netcheck's old PCP probing; will be needed when we add PCP mapping
-
 // pcpMapRequest generates a PCP packet with a MAP opcode.
 func pcpMapRequest(myIP netaddr.IP, mapToLocalPort int, delete bool) []byte {
 	const udpProtoNumber = 17

net/socks5/socks5.go

@@ -32,7 +32,7 @@ const (
 // that represent the kind of connection the client needs.
 type commandType byte
 
-// The set of valid SOCKS5 commans as described in RFC 1928.
+// The set of valid SOCKS5 commands as described in RFC 1928.
 const (
 	connect      commandType = 1
 	bind         commandType = 2

net/tsaddr/tsaddr.go

@@ -41,12 +41,9 @@ var (
 // TailscaleServiceIP returns the listen address of services
 // provided by Tailscale itself such as the MagicDNS proxy.
 func TailscaleServiceIP() netaddr.IP {
-	serviceIP.Do(func() { mustIP(&serviceIP.v, "100.100.100.100") })
-	return serviceIP.v
+	return netaddr.IPv4(100, 100, 100, 100) // "100.100.100.100" for those grepping
 }
 
-var serviceIP onceIP
-
 // IsTailscaleIP reports whether ip is an IP address in a range that
 // Tailscale assigns from.
 func IsTailscaleIP(ip netaddr.IP) bool {
@@ -126,19 +123,6 @@ type oncePrefix struct {
 	v netaddr.IPPrefix
 }
 
-func mustIP(v *netaddr.IP, ip string) {
-	var err error
-	*v, err = netaddr.ParseIP(ip)
-	if err != nil {
-		panic(err)
-	}
-}
-
-type onceIP struct {
-	sync.Once
-	v netaddr.IP
-}
-
 // NewContainsIPFunc returns a func that reports whether ip is in addrs.
 //
 // It's optimized for the cases of addrs being empty and addrs

net/tsaddr/tsaddr_test.go

@@ -93,3 +93,11 @@ func TestNewContainsIPFunc(t *testing.T) {
 		t.Fatal("bad")
 	}
 }
+
+var sinkIP netaddr.IP
+
+func BenchmarkTailscaleServiceAddr(b *testing.B) {
+	for i := 0; i < b.N; i++ {
+		sinkIP = TailscaleServiceIP()
+	}
+}

net/tstun/fake.go

@@ -8,7 +8,7 @@ import (
 	"io"
 	"os"
 
-	"github.com/tailscale/wireguard-go/tun"
+	"golang.zx2c4.com/wireguard/tun"
 )
 
 type fakeTUN struct {

net/tstun/ifstatus_noop.go

@@ -9,7 +9,7 @@ package tstun
 import (
 	"time"
 
-	"github.com/tailscale/wireguard-go/tun"
+	"golang.zx2c4.com/wireguard/tun"
 	"tailscale.com/types/logger"
 )
 

net/tstun/ifstatus_windows.go

@@ -9,7 +9,7 @@ import (
 	"sync"
 	"time"
 
-	"github.com/tailscale/wireguard-go/tun"
+	"golang.zx2c4.com/wireguard/tun"
 	"golang.zx2c4.com/wireguard/windows/tunnel/winipcfg"
 	"tailscale.com/types/logger"
 )