wgengine/magicsock: use learned DERP route as send path of last resort

If we get a packet in over some DERP and don't otherwise know how to
reply (no known DERP home or UDP endpoint), this makes us use the
DERP connection on which we received the packet to reply. This will
almost always be our own home DERP region.

This is particularly useful for large one-way nodes (such as
hello.ts.net) that don't actively reach out to other nodes, so don't
need to be told the DERP home of peers. They can instead learn the
DERP home upon getting the first connection.

This can also help nodes from a slow or misbehaving control plane.

Updates tailscale/corp#26438

Change-Id: I6241ec92828bf45982e0eb83ad5c7404df5968bc
Signed-off-by: Brad Fitzpatrick <[email protected]>

control/controlclient/direct.go

@@ -1255,6 +1255,7 @@ type devKnobs struct {
 	DumpNetMapsVerbose func() bool
 	ForceProxyDNS      func() bool
 	StripEndpoints     func() bool // strip endpoints from control (only use disco messages)
+	StripHomeDERP      func() bool // strip Home DERP from control
 	StripCaps          func() bool // strip all local node's control-provided capabilities
 }
 
@@ -1266,6 +1267,7 @@ func initDevKnob() devKnobs {
 		DumpRegister:       envknob.RegisterBool("TS_DEBUG_REGISTER"),
 		ForceProxyDNS:      envknob.RegisterBool("TS_DEBUG_PROXY_DNS"),
 		StripEndpoints:     envknob.RegisterBool("TS_DEBUG_STRIP_ENDPOINTS"),
+		StripHomeDERP:      envknob.RegisterBool("TS_DEBUG_STRIP_HOME_DERP"),
 		StripCaps:          envknob.RegisterBool("TS_DEBUG_STRIP_CAPS"),
 	}
 }

control/controlclient/map.go

@@ -240,6 +240,9 @@ func upgradeNode(n *tailcfg.Node) {
 		}
 		n.LegacyDERPString = ""
 	}
+	if DevKnob.StripHomeDERP() {
+		n.HomeDERP = 0
+	}
 
 	if n.AllowedIPs == nil {
 		n.AllowedIPs = slices.Clone(n.Addresses)

tstest/integration/nat/nat_test.go

@@ -236,6 +236,22 @@ func hard(c *vnet.Config) *vnet.Node {
 		fmt.Sprintf("10.0.%d.1/24", n), vnet.HardNAT))
 }
 
+func hardNoDERPOrEndoints(c *vnet.Config) *vnet.Node {
+	n := c.NumNodes() + 1
+	return c.AddNode(c.AddNetwork(
+		fmt.Sprintf("2.%d.%d.%d", n, n, n), // public IP
+		fmt.Sprintf("10.0.%d.1/24", n), vnet.HardNAT),
+		vnet.TailscaledEnv{
+			Key:   "TS_DEBUG_STRIP_ENDPOINTS",
+			Value: "1",
+		},
+		vnet.TailscaledEnv{
+			Key:   "TS_DEBUG_STRIP_HOME_DERP",
+			Value: "1",
+		},
+	)
+}
+
 func hardPMP(c *vnet.Config) *vnet.Node {
 	n := c.NumNodes() + 1
 	return c.AddNode(c.AddNetwork(
@@ -510,6 +526,26 @@ func TestEasyEasy(t *testing.T) {
 	nt.want(routeDirect)
 }
 
+// Issue tailscale/corp#26438: use learned DERP route as send path of last
+// resort
+//
+// See (*magicsock.Conn).fallbackDERPRegionForPeer and its comment for
+// background.
+//
+// This sets up a test with two nodes that must use DERP to communicate but the
+// target of the ping (the second node) additionally is not getting DERP or
+// Endpoint updates from the control plane. (Or rather, it's getting them but is
+// configured to scrub them right when they come off the network before being
+// processed) This then tests whether node2, upon receiving a packet, will be
+// able to reply to node1 since it knows neither node1's endpoints nor its home
+// DERP. The only reply route it can use is that fact that it just received a
+// packet over a particular DERP from that peer.
+func TestFallbackDERPRegionForPeer(t *testing.T) {
+	nt := newNatTest(t)
+	nt.runTest(hard, hardNoDERPOrEndoints)
+	nt.want(routeDERP)
+}
+
 func TestSingleJustIPv6(t *testing.T) {
 	nt := newNatTest(t)
 	nt.runTest(just6)

wgengine/magicsock/derp.go

@@ -64,10 +64,30 @@ func (c *Conn) removeDerpPeerRoute(peer key.NodePublic, regionID int, dc *derpht
 // addDerpPeerRoute adds a DERP route entry, noting that peer was seen
 // on DERP node derpID, at least on the connection identified by dc.
 // See issue 150 for details.
-func (c *Conn) addDerpPeerRoute(peer key.NodePublic, derpID int, dc *derphttp.Client) {
+func (c *Conn) addDerpPeerRoute(peer key.NodePublic, regionID int, dc *derphttp.Client) {
 	c.mu.Lock()
 	defer c.mu.Unlock()
-	mak.Set(&c.derpRoute, peer, derpRoute{derpID, dc})
+	mak.Set(&c.derpRoute, peer, derpRoute{regionID, dc})
+}
+
+// fallbackDERPRegionForPeer returns the DERP region ID we might be able to use
+// to contact peer, learned from observing recent DERP traffic from them.
+//
+// This is used as a fallback when a peer receives a packet from a peer
+// over DERP but doesn't known that peer's home DERP or any UDP endpoints.
+// This is particularly useful for large one-way nodes (such as hello.ts.net)
+// that don't actively reach out to other nodes, so don't need to be told
+// the DERP home of peers. They can instead learn the DERP home upon getting the
+// first connection.
+//
+// This can also help nodes from a slow or misbehaving control plane.
+func (c *Conn) fallbackDERPRegionForPeer(peer key.NodePublic) (regionID int) {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	if dr, ok := c.derpRoute[peer]; ok {
+		return dr.regionID
+	}
+	return 0
 }
 
 // activeDerp contains fields for an active DERP connection.

wgengine/magicsock/endpoint.go

@@ -948,7 +948,15 @@ func (de *endpoint) send(buffs [][]byte) error {
 	de.mu.Unlock()
 
 	if !udpAddr.IsValid() && !derpAddr.IsValid() {
-		return errNoUDPOrDERP
+		// Make a last ditch effort to see if we have a DERP route for them. If
+		// they contacted us over DERP and we don't know their UDP endpoints or
+		// their DERP home, we can at least assume they're reachable over the
+		// DERP they used to contact us.
+		if rid := de.c.fallbackDERPRegionForPeer(de.publicKey); rid != 0 {
+			derpAddr = netip.AddrPortFrom(tailcfg.DerpMagicIPAddr, uint16(rid))
+		} else {
+			return errNoUDPOrDERP
+		}
 	}
 	var err error
 	if udpAddr.IsValid() {