net/tsdial: move more weirdo dialing into new tsdial package, plumb

Not done yet, but this move more of the outbound dial special casing
from random packages into tsdial, which aspires to be the one unified
place for all outbound dialing shenanigans.

Then this plumbs it all around, so everybody is ultimately
holding on to the same dialer.

As of this commit, macOS/iOS using an exit node should be able to
reach to the exit node's DoH DNS proxy over peerapi, doing the sockopt
to stay within the Network Extension.

A number of steps remain, including but limited to:

* move a bunch more random dialing stuff

* make netstack-mode tailscaled be able to use exit node's DNS proxy,
  teaching tsdial's resolver to use it when an exit node is in use.

Updates #1713

Change-Id: I1e8ee378f125421c2b816f47bc2c6d913ddcd2f5
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>

cmd/tailscaled/depaware.txt

@@ -194,7 +194,7 @@ tailscale.com/cmd/tailscaled dependencies: (generated by github.com/tailscale/de
         tailscale.com/net/flowtrack                                  from tailscale.com/net/packet+
      💣 tailscale.com/net/interfaces                                 from tailscale.com/cmd/tailscaled+
         tailscale.com/net/netcheck                                   from tailscale.com/wgengine/magicsock
-        tailscale.com/net/netknob                                    from tailscale.com/ipn/localapi+
+        tailscale.com/net/netknob                                    from tailscale.com/logpolicy+
         tailscale.com/net/netns                                      from tailscale.com/cmd/tailscaled+
      💣 tailscale.com/net/netstat                                    from tailscale.com/ipn/ipnserver
         tailscale.com/net/packet                                     from tailscale.com/net/tstun+

cmd/tailscaled/tailscaled.go

@@ -43,7 +43,6 @@ import (
 	"tailscale.com/safesocket"
 	"tailscale.com/types/flagtype"
 	"tailscale.com/types/logger"
-	"tailscale.com/types/netmap"
 	"tailscale.com/util/clientmetric"
 	"tailscale.com/util/multierr"
 	"tailscale.com/util/osshare"
@@ -308,7 +307,8 @@ func run() error {
 
 	socksListener, httpProxyListener := mustStartProxyListeners(args.socksAddr, args.httpProxyAddr)
 
-	e, useNetstack, err := createEngine(logf, linkMon)
+	dialer := new(tsdial.Dialer) // mutated below (before used)
+	e, useNetstack, err := createEngine(logf, linkMon, dialer)
 	if err != nil {
 		logf("wgengine.New: %v", err)
 		return err
@@ -327,7 +327,6 @@ func run() error {
 		log.Fatalf("failed to start netstack: %v", err)
 	}
 
-	dialer := new(tsdial.Dialer)
 	if useNetstack {
 		dialer.UseNetstackForIP = func(ip netaddr.IP) bool {
 			_, ok := e.PeerForIP(ip)
@@ -337,13 +336,10 @@ func run() error {
 			return ns.DialContextTCP(ctx, dst.String())
 		}
 	}
-	e.AddNetworkMapCallback(func(nm *netmap.NetworkMap) {
-		dialer.SetDNSMap(tsdial.DNSMapFromNetworkMap(nm))
-	})
 
 	if socksListener != nil || httpProxyListener != nil {
 		if httpProxyListener != nil {
-			hs := &http.Server{Handler: httpProxyHandler(dialer.DialContext)}
+			hs := &http.Server{Handler: httpProxyHandler(dialer.UserDial)}
 			go func() {
 				log.Fatalf("HTTP proxy exited: %v", hs.Serve(httpProxyListener))
 			}()
@@ -351,7 +347,7 @@ func run() error {
 		if socksListener != nil {
 			ss := &socks5.Server{
 				Logf:   logger.WithPrefix(logf, "socks5: "),
-				Dialer: dialer.DialContext,
+				Dialer: dialer.UserDial,
 			}
 			go func() {
 				log.Fatalf("SOCKS5 server exited: %v", ss.Serve(socksListener))
@@ -387,7 +383,7 @@ func run() error {
 		logf("ipnserver.StateStore: %v", err)
 		return err
 	}
-	srv, err := ipnserver.New(logf, pol.PublicID.String(), store, e, nil, opts)
+	srv, err := ipnserver.New(logf, pol.PublicID.String(), store, e, dialer, nil, opts)
 	if err != nil {
 		logf("ipnserver.New: %v", err)
 		return err
@@ -412,14 +408,14 @@ func run() error {
 	return nil
 }
 
-func createEngine(logf logger.Logf, linkMon *monitor.Mon) (e wgengine.Engine, useNetstack bool, err error) {
+func createEngine(logf logger.Logf, linkMon *monitor.Mon, dialer *tsdial.Dialer) (e wgengine.Engine, useNetstack bool, err error) {
 	if args.tunname == "" {
 		return nil, false, errors.New("no --tun value specified")
 	}
 	var errs []error
 	for _, name := range strings.Split(args.tunname, ",") {
 		logf("wgengine.NewUserspaceEngine(tun %q) ...", name)
-		e, useNetstack, err = tryEngine(logf, linkMon, name)
+		e, useNetstack, err = tryEngine(logf, linkMon, dialer, name)
 		if err == nil {
 			return e, useNetstack, nil
 		}
@@ -451,10 +447,11 @@ func shouldWrapNetstack() bool {
 	return false
 }
 
-func tryEngine(logf logger.Logf, linkMon *monitor.Mon, name string) (e wgengine.Engine, useNetstack bool, err error) {
+func tryEngine(logf logger.Logf, linkMon *monitor.Mon, dialer *tsdial.Dialer, name string) (e wgengine.Engine, useNetstack bool, err error) {
 	conf := wgengine.Config{
 		ListenPort:  args.port,
 		LinkMonitor: linkMon,
+		Dialer:      dialer,
 	}
 
 	useNetstack = name == "userspace-networking"

ipn/ipnlocal/local.go

@@ -36,6 +36,7 @@ import (
 	"tailscale.com/net/dns"
 	"tailscale.com/net/interfaces"
 	"tailscale.com/net/tsaddr"
+	"tailscale.com/net/tsdial"
 	"tailscale.com/paths"
 	"tailscale.com/portlist"
 	"tailscale.com/tailcfg"
@@ -88,6 +89,7 @@ type LocalBackend struct {
 	statsLogf             logger.Logf        // for printing peers stats on change
 	e                     wgengine.Engine
 	store                 ipn.StateStore
+	dialer                *tsdial.Dialer // non-nil
 	backendLogID          string
 	unregisterLinkMon     func()
 	unregisterHealthWatch func()
@@ -155,10 +157,18 @@ type clientGen func(controlclient.Options) (controlclient.Client, error)
 
 // NewLocalBackend returns a new LocalBackend that is ready to run,
 // but is not actually running.
-func NewLocalBackend(logf logger.Logf, logid string, store ipn.StateStore, e wgengine.Engine) (*LocalBackend, error) {
+//
+// If dialer is nil, a new one is made.
+func NewLocalBackend(logf logger.Logf, logid string, store ipn.StateStore, dialer *tsdial.Dialer, e wgengine.Engine) (*LocalBackend, error) {
 	if e == nil {
-		panic("ipn.NewLocalBackend: wgengine must not be nil")
+		panic("ipn.NewLocalBackend: engine must not be nil")
+	}
+	if dialer == nil {
+		dialer = new(tsdial.Dialer)
 	}
+	e.AddNetworkMapCallback(func(nm *netmap.NetworkMap) {
+		dialer.SetDNSMap(tsdial.DNSMapFromNetworkMap(nm))
+	})
 
 	osshare.SetFileSharingEnabled(false, logf)
 
@@ -176,11 +186,14 @@ func NewLocalBackend(logf logger.Logf, logid string, store ipn.StateStore, e wge
 		statsLogf:      logger.LogOnChange(logf, 5*time.Minute, time.Now),
 		e:              e,
 		store:          store,
+		dialer:         dialer,
 		backendLogID:   logid,
 		state:          ipn.NoState,
 		portpoll:       portpoll,
 		gotPortPollRes: make(chan struct{}),
 	}
+	dialer.SetPeerDialControlFuncGetter(b.peerDialControlFunc)
+
 	// Default filter blocks everything and logs nothing, until Start() is called.
 	b.setFilter(filter.NewAllowNone(logf, &netaddr.IPSet{}))
 
@@ -210,6 +223,11 @@ func NewLocalBackend(logf logger.Logf, logid string, store ipn.StateStore, e wge
 	return b, nil
 }
 
+// Dialer returns the backend's dialer.
+func (b *LocalBackend) Dialer() *tsdial.Dialer {
+	return b.dialer
+}
+
 // SetDirectFileRoot sets the directory to download files to directly,
 // without buffering them through an intermediate daemon-owned
 // tailcfg.UserID-specific directory.
@@ -3035,9 +3053,9 @@ func disabledSysctls(sysctls ...string) (disabled []string, err error) {
 // bind to dial out to other peers.
 var peerDialControlFunc func(*LocalBackend) func(network, address string, c syscall.RawConn) error
 
-// PeerDialControlFunc returns a net.Dialer.Control func (possibly nil) to use to
+// peerDialControlFunc returns a net.Dialer.Control func (possibly nil) to use to
 // dial other Tailscale peers from the current environment.
-func (b *LocalBackend) PeerDialControlFunc() func(network, address string, c syscall.RawConn) error {
+func (b *LocalBackend) peerDialControlFunc() func(network, address string, c syscall.RawConn) error {
 	if peerDialControlFunc != nil {
 		return peerDialControlFunc(b)
 	}

ipn/ipnlocal/local_test.go

@@ -445,7 +445,7 @@ func TestLazyMachineKeyGeneration(t *testing.T) {
 		t.Fatalf("NewFakeUserspaceEngine: %v", err)
 	}
 	t.Cleanup(eng.Close)
-	lb, err := NewLocalBackend(logf, "logid", store, eng)
+	lb, err := NewLocalBackend(logf, "logid", store, nil, eng)
 	if err != nil {
 		t.Fatalf("NewLocalBackend: %v", err)
 	}

ipn/ipnlocal/loglines_test.go

@@ -54,7 +54,7 @@ func TestLocalLogLines(t *testing.T) {
 	}
 	t.Cleanup(e.Close)
 
-	lb, err := NewLocalBackend(logf, idA.String(), store, e)
+	lb, err := NewLocalBackend(logf, idA.String(), store, nil, e)
 	if err != nil {
 		t.Fatal(err)
 	}

ipn/ipnlocal/state_test.go

@@ -284,7 +284,7 @@ func TestStateMachine(t *testing.T) {
 	t.Cleanup(e.Close)
 
 	cc := newMockControl(t)
-	b, err := NewLocalBackend(logf, "logid", store, e)
+	b, err := NewLocalBackend(logf, "logid", store, nil, e)
 	if err != nil {
 		t.Fatalf("NewLocalBackend: %v", err)
 	}
@@ -941,7 +941,7 @@ func TestWGEngineStatusRace(t *testing.T) {
 	eng, err := wgengine.NewFakeUserspaceEngine(logf, 0)
 	c.Assert(err, qt.IsNil)
 	t.Cleanup(eng.Close)
-	b, err := NewLocalBackend(logf, "logid", new(ipn.MemoryStore), eng)
+	b, err := NewLocalBackend(logf, "logid", new(ipn.MemoryStore), nil, eng)
 	c.Assert(err, qt.IsNil)
 
 	cc := newMockControl(t)

ipn/ipnserver/server.go

@@ -38,10 +38,12 @@ import (
 	"tailscale.com/log/filelogger"
 	"tailscale.com/logtail/backoff"
 	"tailscale.com/net/netstat"
+	"tailscale.com/net/tsdial"
 	"tailscale.com/paths"
 	"tailscale.com/safesocket"
 	"tailscale.com/smallzstd"
 	"tailscale.com/types/logger"
+	"tailscale.com/types/netmap"
 	"tailscale.com/util/groupmember"
 	"tailscale.com/util/pidowner"
 	"tailscale.com/util/systemd"
@@ -735,7 +737,12 @@ func Run(ctx context.Context, logf logger.Logf, ln net.Listener, store ipn.State
 		}
 	}
 
-	server, err := New(logf, logid, store, eng, serverModeUser, opts)
+	dialer := new(tsdial.Dialer)
+	eng.AddNetworkMapCallback(func(nm *netmap.NetworkMap) {
+		dialer.SetDNSMap(tsdial.DNSMapFromNetworkMap(nm))
+	})
+
+	server, err := New(logf, logid, store, eng, nil, serverModeUser, opts)
 	if err != nil {
 		return err
 	}
@@ -748,8 +755,8 @@ func Run(ctx context.Context, logf logger.Logf, ln net.Listener, store ipn.State
 // New returns a new Server.
 //
 // To start it, use the Server.Run method.
-func New(logf logger.Logf, logid string, store ipn.StateStore, eng wgengine.Engine, serverModeUser *user.User, opts Options) (*Server, error) {
+func New(logf logger.Logf, logid string, store ipn.StateStore, eng wgengine.Engine, dialer *tsdial.Dialer, serverModeUser *user.User, opts Options) (*Server, error) {
-	b, err := ipnlocal.NewLocalBackend(logf, logid, store, eng)
+	b, err := ipnlocal.NewLocalBackend(logf, logid, store, dialer, eng)
 	if err != nil {
 		return nil, fmt.Errorf("NewLocalBackend: %v", err)
 	}

ipn/localapi/localapi.go

@@ -12,7 +12,6 @@ import (
 	"errors"
 	"fmt"
 	"io"
-	"net"
 	"net/http"
 	"net/http/httputil"
 	"net/url"
@@ -20,7 +19,6 @@ import (
 	"runtime"
 	"strconv"
 	"strings"
-	"sync"
 	"time"
 
 	"inet.af/netaddr"
@@ -28,7 +26,6 @@ import (
 	"tailscale.com/ipn"
 	"tailscale.com/ipn/ipnlocal"
 	"tailscale.com/ipn/ipnstate"
-	"tailscale.com/net/netknob"
 	"tailscale.com/tailcfg"
 	"tailscale.com/types/logger"
 	"tailscale.com/util/clientmetric"
@@ -442,7 +439,7 @@ func (h *Handler) serveFilePut(w http.ResponseWriter, r *http.Request) {
 	outReq.ContentLength = r.ContentLength
 
 	rp := httputil.NewSingleHostReverseProxy(dstURL)
-	rp.Transport = getDialPeerTransport(h.b)
+	rp.Transport = h.b.Dialer().PeerAPITransport()
 	rp.ServeHTTP(w, outReq)
 }
 
@@ -476,26 +473,6 @@ func (h *Handler) serveDERPMap(w http.ResponseWriter, r *http.Request) {
 	e.Encode(h.b.DERPMap())
 }
 
-var dialPeerTransportOnce struct {
-	sync.Once
-	v *http.Transport
-}
-
-func getDialPeerTransport(b *ipnlocal.LocalBackend) *http.Transport {
-	dialPeerTransportOnce.Do(func() {
-		t := http.DefaultTransport.(*http.Transport).Clone()
-		t.Dial = nil
-		dialer := net.Dialer{
-			Timeout:   30 * time.Second,
-			KeepAlive: netknob.PlatformTCPKeepAlive(),
-			Control:   b.PeerDialControlFunc(),
-		}
-		t.DialContext = dialer.DialContext
-		dialPeerTransportOnce.v = t
-	})
-	return dialPeerTransportOnce.v
-}
-
 func defBool(a string, def bool) bool {
 	if a == "" {
 		return def

net/dns/manager.go

@@ -13,6 +13,7 @@ import (
 	"tailscale.com/health"
 	"tailscale.com/net/dns/resolver"
 	"tailscale.com/net/tsaddr"
+	"tailscale.com/net/tsdial"
 	"tailscale.com/types/dnstype"
 	"tailscale.com/types/logger"
 	"tailscale.com/util/dnsname"
@@ -39,11 +40,14 @@ type Manager struct {
 }
 
 // NewManagers created a new manager from the given config.
-func NewManager(logf logger.Logf, oscfg OSConfigurator, linkMon *monitor.Mon, linkSel resolver.ForwardLinkSelector) *Manager {
+func NewManager(logf logger.Logf, oscfg OSConfigurator, linkMon *monitor.Mon, dialer *tsdial.Dialer, linkSel resolver.ForwardLinkSelector) *Manager {
+	if dialer == nil {
+		panic("nil Dialer")
+	}
 	logf = logger.WithPrefix(logf, "dns: ")
 	m := &Manager{
 		logf:     logf,
-		resolver: resolver.New(logf, linkMon, linkSel),
+		resolver: resolver.New(logf, linkMon, linkSel, dialer),
 		os:       oscfg,
 	}
 	m.logf("using %T", m.os)
@@ -230,7 +234,7 @@ func Cleanup(logf logger.Logf, interfaceName string) {
 		logf("creating dns cleanup: %v", err)
 		return
 	}
-	dns := NewManager(logf, oscfg, nil, nil)
+	dns := NewManager(logf, oscfg, nil, new(tsdial.Dialer), nil)
 	if err := dns.Down(); err != nil {
 		logf("dns down: %v", err)
 	}

net/dns/manager_test.go

@@ -13,6 +13,7 @@ import (
 	"github.com/google/go-cmp/cmp/cmpopts"
 	"inet.af/netaddr"
 	"tailscale.com/net/dns/resolver"
+	"tailscale.com/net/tsdial"
 	"tailscale.com/types/dnstype"
 	"tailscale.com/util/dnsname"
 )
@@ -398,7 +399,7 @@ func TestManager(t *testing.T) {
 				SplitDNS:   test.split,
 				BaseConfig: test.bs,
 			}
-			m := NewManager(t.Logf, &f, nil, nil)
+			m := NewManager(t.Logf, &f, nil, new(tsdial.Dialer), nil)
 			m.resolver.TestOnlySetHook(f.SetResolver)
 
 			if err := m.Set(test.in); err != nil {

net/dns/resolver/forwarder.go

@@ -26,6 +26,7 @@ import (
 	"inet.af/netaddr"
 	"tailscale.com/hostinfo"
 	"tailscale.com/net/netns"
+	"tailscale.com/net/tsdial"
 	"tailscale.com/types/dnstype"
 	"tailscale.com/types/logger"
 	"tailscale.com/util/dnsname"
@@ -159,7 +160,8 @@ type resolverAndDelay struct {
 type forwarder struct {
 	logf    logger.Logf
 	linkMon *monitor.Mon
-	linkSel ForwardLinkSelector
+	linkSel ForwardLinkSelector // TODO(bradfitz): remove this when tsdial.Dialer absords it
+	dialer  *tsdial.Dialer
 	dohSem  chan struct{}
 
 	ctx       context.Context    // good until Close
@@ -205,11 +207,12 @@ func maxDoHInFlight(goos string) int {
 	return 1000
 }
 
-func newForwarder(logf logger.Logf, responses chan packet, linkMon *monitor.Mon, linkSel ForwardLinkSelector) *forwarder {
+func newForwarder(logf logger.Logf, responses chan packet, linkMon *monitor.Mon, linkSel ForwardLinkSelector, dialer *tsdial.Dialer) *forwarder {
 	f := &forwarder{
 		logf:      logger.WithPrefix(logf, "forward: "),
 		linkMon:   linkMon,
 		linkSel:   linkSel,
+		dialer:    dialer,
 		responses: responses,
 		dohSem:    make(chan struct{}, maxDoHInFlight(runtime.GOOS)),
 	}
@@ -423,10 +426,7 @@ func (f *forwarder) sendDoH(ctx context.Context, urlBase string, c *http.Client,
 // send expects the reply to have the same txid as txidOut.
 func (f *forwarder) send(ctx context.Context, fq *forwardQuery, rr resolverAndDelay) ([]byte, error) {
 	if strings.HasPrefix(rr.name.Addr, "http://") {
-		// TODO(bradfitz): this only work for TUN mode right now; plumb a universal dialer
+		return f.sendDoH(ctx, rr.name.Addr, f.dialer.PeerAPIHTTPClient(), fq.packet)
-		// that can handle the dozen special cases for modes/platforms/routes.
-		TODOHTTPClient := http.DefaultClient
-		return f.sendDoH(ctx, rr.name.Addr, TODOHTTPClient, fq.packet)
 	}
 	if strings.HasPrefix(rr.name.Addr, "https://") {
 		metricDNSFwdErrorType.Add(1)

net/dns/resolver/tsdns.go

@@ -27,6 +27,7 @@ import (
 	dns "golang.org/x/net/dns/dnsmessage"
 	"inet.af/netaddr"
 	"tailscale.com/net/tsaddr"
+	"tailscale.com/net/tsdial"
 	"tailscale.com/types/dnstype"
 	"tailscale.com/types/logger"
 	"tailscale.com/util/clientmetric"
@@ -192,6 +193,7 @@ func WriteRoutes(w *bufio.Writer, routes map[dnsname.FQDN][]dnstype.Resolver) {
 type Resolver struct {
 	logf               logger.Logf
 	linkMon            *monitor.Mon     // or nil
+	dialer             *tsdial.Dialer   // non-nil
 	saveConfigForTests func(cfg Config) // used in tests to capture resolver config
 	// forwarder forwards requests to upstream nameservers.
 	forwarder *forwarder
@@ -223,7 +225,10 @@ type ForwardLinkSelector interface {
 
 // New returns a new resolver.
 // linkMon optionally specifies a link monitor to use for socket rebinding.
-func New(logf logger.Logf, linkMon *monitor.Mon, linkSel ForwardLinkSelector) *Resolver {
+func New(logf logger.Logf, linkMon *monitor.Mon, linkSel ForwardLinkSelector, dialer *tsdial.Dialer) *Resolver {
+	if dialer == nil {
+		panic("nil Dialer")
+	}
 	r := &Resolver{
 		logf:      logger.WithPrefix(logf, "resolver: "),
 		linkMon:   linkMon,
@@ -232,8 +237,9 @@ func New(logf logger.Logf, linkMon *monitor.Mon, linkSel ForwardLinkSelector) *R
 		closed:    make(chan struct{}),
 		hostToIP:  map[dnsname.FQDN][]netaddr.IP{},
 		ipToHost:  map[netaddr.IP]dnsname.FQDN{},
+		dialer:    dialer,
 	}
-	r.forwarder = newForwarder(r.logf, r.responses, linkMon, linkSel)
+	r.forwarder = newForwarder(r.logf, r.responses, linkMon, linkSel, dialer)
 	return r
 }
 

net/dns/resolver/tsdns_test.go

@@ -18,6 +18,7 @@ import (
 
 	dns "golang.org/x/net/dns/dnsmessage"
 	"inet.af/netaddr"
+	"tailscale.com/net/tsdial"
 	"tailscale.com/tstest"
 	"tailscale.com/types/dnstype"
 	"tailscale.com/util/dnsname"
@@ -308,7 +309,7 @@ func TestRDNSNameToIPv6(t *testing.T) {
 }
 
 func newResolver(t testing.TB) *Resolver {
-	return New(t.Logf, nil /* no link monitor */, nil /* no link selector */)
+	return New(t.Logf, nil /* no link monitor */, nil /* no link selector */, new(tsdial.Dialer))
 }
 
 func TestResolveLocal(t *testing.T) {
@@ -1062,7 +1063,7 @@ func TestForwardLinkSelection(t *testing.T) {
 			return "special"
 		}
 		return ""
-	}))
+	}), new(tsdial.Dialer))
 
 	// Test non-special IP.
 	if got, err := fwd.packetListener(netaddr.IP{}); err != nil {

net/tsdial/tsdial.go

@@ -9,9 +9,14 @@ import (
 	"context"
 	"errors"
 	"net"
+	"net/http"
 	"sync"
+	"sync/atomic"
+	"syscall"
+	"time"
 
 	"inet.af/netaddr"
+	"tailscale.com/net/netknob"
 )
 
 // Dialer dials out of tailscaled, while taking care of details while
@@ -27,11 +32,43 @@ type Dialer struct {
 	// If nil, it's not used.
 	NetstackDialTCP func(context.Context, netaddr.IPPort) (net.Conn, error)
 
+	peerDialControlFuncAtomic atomic.Value // of func() func(network, address string, c syscall.RawConn) error
+
+	peerClientOnce sync.Once
+	peerClient     *http.Client
+
 	mu  sync.Mutex
 	dns DNSMap
 }
 
+// PeerDialControlFunc returns a function
+// that can assigned to net.Dialer.Control to set sockopts or whatnot
+// to make a dial escape the current platform's network sandbox.
+//
+// On many platforms the returned func will be nil.
+//
+// Notably, this is non-nil on iOS and macOS when run as a Network or
+// System Extension (the GUI variants).
+func (d *Dialer) PeerDialControlFunc() func(network, address string, c syscall.RawConn) error {
+	gf, _ := d.peerDialControlFuncAtomic.Load().(func() func(network, address string, c syscall.RawConn) error)
+	if gf == nil {
+		return nil
+	}
+	return gf()
+}
+
+// SetPeerDialControlFuncGetter sets a function that returns, for the
+// current network configuration at the time it's called, a function
+// that can assigned to net.Dialer.Control to set sockopts or whatnot
+// to make a dial escape the current platform's network sandbox.
+func (d *Dialer) SetPeerDialControlFuncGetter(getFunc func() func(network, address string, c syscall.RawConn) error) {
+	d.peerDialControlFuncAtomic.Store(getFunc)
+}
+
+// SetDNSMap sets the current map of DNS names learned from the netmap.
 func (d *Dialer) SetDNSMap(m DNSMap) {
+	// TODO(bradfitz): update this to be aware of DNSConfig
+	// ExtraNames and CertDomains.
 	d.mu.Lock()
 	defer d.mu.Unlock()
 	d.dns = m
@@ -44,7 +81,9 @@ func (d *Dialer) resolve(ctx context.Context, addr string) (netaddr.IPPort, erro
 	return dns.Resolve(ctx, addr)
 }
 
-func (d *Dialer) DialContext(ctx context.Context, network, addr string) (net.Conn, error) {
+// UserDial connects to the provided network address as if a user were initiating the dial.
+// (e.g. from a SOCKS or HTTP outbound proxy)
+func (d *Dialer) UserDial(ctx context.Context, network, addr string) (net.Conn, error) {
 	ipp, err := d.resolve(ctx, addr)
 	if err != nil {
 		return nil, err
@@ -59,3 +98,31 @@ func (d *Dialer) DialContext(ctx context.Context, network, addr string) (net.Con
 	var stdDialer net.Dialer
 	return stdDialer.DialContext(ctx, network, ipp.String())
 }
+
+// PeerAPIHTTPClient returns an HTTP Client to call peers' peerapi
+// endpoints.                                                                                                                                                                                                                      //
+// The returned Client must not be mutated; it's owned by the Dialer
+// and shared by callers.
+func (d *Dialer) PeerAPIHTTPClient() *http.Client {
+	d.peerClientOnce.Do(func() {
+		t := http.DefaultTransport.(*http.Transport).Clone()
+		t.Dial = nil
+		dialer := net.Dialer{
+			Timeout:   30 * time.Second,
+			KeepAlive: netknob.PlatformTCPKeepAlive(),
+			Control:   d.PeerDialControlFunc(),
+		}
+		t.DialContext = dialer.DialContext
+		d.peerClient = &http.Client{Transport: t}
+	})
+	return d.peerClient
+}
+
+// PeerAPITransport returns a Transport to call peers' peerapi
+// endpoints.
+//
+// The returned value must not be mutated; it's owned by the Dialer
+// and shared by callers.
+func (d *Dialer) PeerAPITransport() *http.Transport {
+	return d.PeerAPIHTTPClient().Transport.(*http.Transport)
+}

tsnet/tsnet.go

@@ -8,6 +8,7 @@
 package tsnet
 
 import (
+	"context"
 	"errors"
 	"fmt"
 	"log"
@@ -20,12 +21,14 @@ import (
 	"sync"
 	"time"
 
+	"inet.af/netaddr"
 	"tailscale.com/client/tailscale"
 	"tailscale.com/control/controlclient"
 	"tailscale.com/ipn"
 	"tailscale.com/ipn/ipnlocal"
 	"tailscale.com/ipn/localapi"
 	"tailscale.com/net/nettest"
+	"tailscale.com/net/tsdial"
 	"tailscale.com/smallzstd"
 	"tailscale.com/types/logger"
 	"tailscale.com/wgengine"
@@ -114,9 +117,11 @@ func (s *Server) start() error {
 		return err
 	}
 
+	dialer := new(tsdial.Dialer) // mutated below (before used)
 	eng, err := wgengine.NewUserspaceEngine(logf, wgengine.Config{
 		ListenPort:  0,
 		LinkMonitor: linkMon,
+		Dialer:      dialer,
 	})
 	if err != nil {
 		return err
@@ -136,6 +141,13 @@ func (s *Server) start() error {
 	if err := ns.Start(); err != nil {
 		return fmt.Errorf("failed to start netstack: %w", err)
 	}
+	dialer.UseNetstackForIP = func(ip netaddr.IP) bool {
+		_, ok := eng.PeerForIP(ip)
+		return ok
+	}
+	dialer.NetstackDialTCP = func(ctx context.Context, dst netaddr.IPPort) (net.Conn, error) {
+		return ns.DialContextTCP(ctx, dst.String())
+	}
 
 	statePath := filepath.Join(s.dir, "tailscaled.state")
 	store, err := ipn.NewFileStore(statePath)
@@ -144,7 +156,7 @@ func (s *Server) start() error {
 	}
 	logid := "tslib-TODO"
 
-	lb, err := ipnlocal.NewLocalBackend(logf, logid, store, eng)
+	lb, err := ipnlocal.NewLocalBackend(logf, logid, store, dialer, eng)
 	if err != nil {
 		return fmt.Errorf("NewLocalBackend: %v", err)
 	}

tstest/integration/tailscaled_deps_test_darwin.go

@@ -33,7 +33,6 @@ import (
 	_ "tailscale.com/types/flagtype"
 	_ "tailscale.com/types/key"
 	_ "tailscale.com/types/logger"
-	_ "tailscale.com/types/netmap"
 	_ "tailscale.com/util/clientmetric"
 	_ "tailscale.com/util/multierr"
 	_ "tailscale.com/util/osshare"

tstest/integration/tailscaled_deps_test_freebsd.go

@@ -33,7 +33,6 @@ import (
 	_ "tailscale.com/types/flagtype"
 	_ "tailscale.com/types/key"
 	_ "tailscale.com/types/logger"
-	_ "tailscale.com/types/netmap"
 	_ "tailscale.com/util/clientmetric"
 	_ "tailscale.com/util/multierr"
 	_ "tailscale.com/util/osshare"

tstest/integration/tailscaled_deps_test_linux.go

@@ -33,7 +33,6 @@ import (
 	_ "tailscale.com/types/flagtype"
 	_ "tailscale.com/types/key"
 	_ "tailscale.com/types/logger"
-	_ "tailscale.com/types/netmap"
 	_ "tailscale.com/util/clientmetric"
 	_ "tailscale.com/util/multierr"
 	_ "tailscale.com/util/osshare"

tstest/integration/tailscaled_deps_test_openbsd.go

@@ -33,7 +33,6 @@ import (
 	_ "tailscale.com/types/flagtype"
 	_ "tailscale.com/types/key"
 	_ "tailscale.com/types/logger"
-	_ "tailscale.com/types/netmap"
 	_ "tailscale.com/util/clientmetric"
 	_ "tailscale.com/util/multierr"
 	_ "tailscale.com/util/osshare"

tstest/integration/tailscaled_deps_test_windows.go

@@ -37,7 +37,6 @@ import (
 	_ "tailscale.com/types/flagtype"
 	_ "tailscale.com/types/key"
 	_ "tailscale.com/types/logger"
-	_ "tailscale.com/types/netmap"
 	_ "tailscale.com/util/clientmetric"
 	_ "tailscale.com/util/multierr"
 	_ "tailscale.com/util/osshare"

wgengine/userspace.go

@@ -33,6 +33,7 @@ import (
 	"tailscale.com/net/interfaces"
 	"tailscale.com/net/packet"
 	"tailscale.com/net/tsaddr"
+	"tailscale.com/net/tsdial"
 	"tailscale.com/net/tshttpproxy"
 	"tailscale.com/net/tstun"
 	"tailscale.com/tailcfg"
@@ -191,6 +192,10 @@ type Config struct {
 	// If nil, a new link monitor is created.
 	LinkMonitor *monitor.Mon
 
+	// Dialer is the dialer to use for outbound connections.
+	// If nil, a new Dialer is created
+	Dialer *tsdial.Dialer
+
 	// ListenPort is the port on which the engine will listen.
 	// If zero, a port is automatically selected.
 	ListenPort uint16
@@ -268,6 +273,9 @@ func NewUserspaceEngine(logf logger.Logf, conf Config) (_ Engine, reterr error)
 		}
 		conf.DNS = d
 	}
+	if conf.Dialer == nil {
+		conf.Dialer = new(tsdial.Dialer)
+	}
 
 	var tsTUNDev *tstun.Wrapper
 	if conf.IsTAP {
@@ -310,7 +318,7 @@ func NewUserspaceEngine(logf logger.Logf, conf Config) (_ Engine, reterr error)
 	}
 
 	tunName, _ := conf.Tun.Name()
-	e.dns = dns.NewManager(logf, conf.DNS, e.linkMon, fwdDNSLinkSelector{e, tunName})
+	e.dns = dns.NewManager(logf, conf.DNS, e.linkMon, conf.Dialer, fwdDNSLinkSelector{e, tunName})
 
 	logf("link state: %+v", e.linkMon.InterfaceState())