wgengine/magicsock: use wireguard-go/conn.PeerAwareEndpoint

If we get an non-disco presumably-wireguard-encrypted UDP packet from
an IP:port we don't recognize, rather than drop the packet, give it to
WireGuard anyway and let WireGuard try to figure out who it's from and
tell us.

This uses the new hook added in https://github.com/tailscale/wireguard-go/pull/27

Updates tailscale/corp#20732

Change-Id: I5c61a40143810592f9efac6c12808a87f924ecf2
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>

control/controlknobs/controlknobs.go

@@ -95,6 +95,10 @@ type Knobs struct {
 	// We began creating this rule on 2024-06-14, and this knob
 	// allows us to disable the new behavior remotely if needed.
 	DisableLocalDNSOverrideViaNRPT atomic.Bool
+
+	// DisableCryptorouting indicates that the node should not use the
+	// magicsock crypto routing feature.
+	DisableCryptorouting atomic.Bool
 }
 
 // UpdateFromNodeAttributes updates k (if non-nil) based on the provided self
@@ -122,6 +126,7 @@ func (k *Knobs) UpdateFromNodeAttributes(capMap tailcfg.NodeCapMap) {
 		userDialUseRoutes                    = has(tailcfg.NodeAttrUserDialUseRoutes)
 		disableSplitDNSWhenNoCustomResolvers = has(tailcfg.NodeAttrDisableSplitDNSWhenNoCustomResolvers)
 		disableLocalDNSOverrideViaNRPT       = has(tailcfg.NodeAttrDisableLocalDNSOverrideViaNRPT)
+		disableCryptorouting                 = has(tailcfg.NodeAttrDisableMagicSockCryptoRouting)
 	)
 
 	if has(tailcfg.NodeAttrOneCGNATEnable) {
@@ -147,6 +152,7 @@ func (k *Knobs) UpdateFromNodeAttributes(capMap tailcfg.NodeCapMap) {
 	k.UserDialUseRoutes.Store(userDialUseRoutes)
 	k.DisableSplitDNSWhenNoCustomResolvers.Store(disableSplitDNSWhenNoCustomResolvers)
 	k.DisableLocalDNSOverrideViaNRPT.Store(disableLocalDNSOverrideViaNRPT)
+	k.DisableCryptorouting.Store(disableCryptorouting)
 }
 
 // AsDebugJSON returns k as something that can be marshalled with json.Marshal
@@ -173,5 +179,6 @@ func (k *Knobs) AsDebugJSON() map[string]any {
 		"UserDialUseRoutes":                    k.UserDialUseRoutes.Load(),
 		"DisableSplitDNSWhenNoCustomResolvers": k.DisableSplitDNSWhenNoCustomResolvers.Load(),
 		"DisableLocalDNSOverrideViaNRPT":       k.DisableLocalDNSOverrideViaNRPT.Load(),
+		"DisableCryptorouting":                 k.DisableCryptorouting.Load(),
 	}
 }

tailcfg/tailcfg.go

@@ -144,7 +144,8 @@ type CapabilityVersion int
 //   - 99: 2024-06-14: Client understands NodeAttrDisableLocalDNSOverrideViaNRPT
 //   - 100: 2024-06-18: Client supports filtertype.Match.SrcCaps (issue #12542)
 //   - 101: 2024-07-01: Client supports SSH agent forwarding when handling connections with /bin/su
-const CurrentCapabilityVersion CapabilityVersion = 101
+//   - 102: 2024-07-12: NodeAttrDisableMagicSockCryptoRouting support
+const CurrentCapabilityVersion CapabilityVersion = 102
 
 type StableID string
 
@@ -2322,6 +2323,10 @@ const (
 	// We began creating this rule on 2024-06-14, and this node attribute
 	// allows us to disable the new behavior remotely if needed.
 	NodeAttrDisableLocalDNSOverrideViaNRPT NodeCapability = "disable-local-dns-override-via-nrpt"
+
+	// NodeAttrDisableMagicSockCryptoRouting disables the use of the
+	// magicsock cryptorouting hook. See tailscale/corp#20732.
+	NodeAttrDisableMagicSockCryptoRouting NodeCapability = "disable-magicsock-crypto-routing"
 )
 
 // SetDNSRequest is a request to add a DNS record.

wgengine/magicsock/magicsock.go

@@ -1276,7 +1276,8 @@ func (c *Conn) mkReceiveFunc(ruc *RebindingUDPConn, healthItem *health.ReceiveFu
 //
 // ok is whether this read should be reported up to wireguard-go (our
 // caller).
-func (c *Conn) receiveIP(b []byte, ipp netip.AddrPort, cache *ippEndpointCache) (ep *endpoint, ok bool) {
+func (c *Conn) receiveIP(b []byte, ipp netip.AddrPort, cache *ippEndpointCache) (_ conn.Endpoint, ok bool) {
+	var ep *endpoint
 	if stun.Is(b) {
 		c.netChecker.ReceiveSTUNPacket(b, ipp)
 		return nil, false
@@ -1297,8 +1298,11 @@ func (c *Conn) receiveIP(b []byte, ipp netip.AddrPort, cache *ippEndpointCache)
 		de, ok := c.peerMap.endpointForIPPort(ipp)
 		c.mu.Unlock()
 		if !ok {
+			if c.controlKnobs != nil && c.controlKnobs.DisableCryptorouting.Load() {
 				return nil, false
 			}
+			return &lazyEndpoint{c: c, src: ipp}, true
+		}
 		cache.ipp = ipp
 		cache.de = de
 		cache.gen = de.numStopAndReset()
@@ -3114,3 +3118,35 @@ func (c *Conn) GetLastNetcheckReport(ctx context.Context) *netcheck.Report {
 func (c *Conn) SetLastNetcheckReportForTest(ctx context.Context, report *netcheck.Report) {
 	c.lastNetCheckReport.Store(report)
 }
+
+// lazyEndpoint is a wireguard conn.Endpoint for when magicsock received a
+// non-disco (presumably WireGuard) packet from a UDP address from which we
+// can't map to a Tailscale peer. But Wireguard most likely can, once it
+// decrypts it. So we implement the conn.PeerAwareEndpoint interface
+// from https://github.com/tailscale/wireguard-go/pull/27 to allow WireGuard
+// to tell us who it is later and get the correct conn.Endpoint.
+type lazyEndpoint struct {
+	c   *Conn
+	src netip.AddrPort
+}
+
+var _ conn.PeerAwareEndpoint = (*lazyEndpoint)(nil)
+var _ conn.Endpoint = (*lazyEndpoint)(nil)
+
+func (le *lazyEndpoint) ClearSrc()           {}
+func (le *lazyEndpoint) SrcIP() netip.Addr   { return le.src.Addr() }
+func (le *lazyEndpoint) DstIP() netip.Addr   { return netip.Addr{} }
+func (le *lazyEndpoint) SrcToString() string { return le.src.String() }
+func (le *lazyEndpoint) DstToString() string { return "dst" }
+func (le *lazyEndpoint) DstToBytes() []byte  { return nil }
+func (le *lazyEndpoint) GetPeerEndpoint(peerPublicKey [32]byte) conn.Endpoint {
+	pubKey := key.NodePublicFromRaw32(mem.B(peerPublicKey[:]))
+	le.c.mu.Lock()
+	defer le.c.mu.Unlock()
+	ep, ok := le.c.peerMap.endpointForNodeKey(pubKey)
+	if !ok {
+		return nil
+	}
+	le.c.logf("magicsock: lazyEndpoint.GetPeerEndpoint(%v) found: %v", pubKey.ShortString(), ep.nodeAddr)
+	return ep
+}