net/udprelay: use batching.Conn (#16866)

This significantly improves throughput of a peer relay server on Linux.

Server.packetReadLoop no longer passes sockets down the stack. Instead,
packet handling methods return a netip.AddrPort and []byte, which
packetReadLoop gathers together for eventual batched writes on the
appropriate socket(s).

Updates tailscale/corp#31164

Signed-off-by: Jordan Whited <jordan@tailscale.com>

cmd/tailscaled/depaware.txt

@@ -311,7 +311,7 @@ tailscale.com/cmd/tailscaled dependencies: (generated by github.com/tailscale/de
         tailscale.com/logtail/filch                                  from tailscale.com/log/sockstatlog+
         tailscale.com/metrics                                        from tailscale.com/derp+
         tailscale.com/net/bakedroots                                 from tailscale.com/net/tlsdial+
-     💣 tailscale.com/net/batching                                   from tailscale.com/wgengine/magicsock
+     💣 tailscale.com/net/batching                                   from tailscale.com/wgengine/magicsock+
         tailscale.com/net/captivedetection                           from tailscale.com/ipn/ipnlocal+
         tailscale.com/net/connstats                                  from tailscale.com/net/tstun+
         tailscale.com/net/dns                                        from tailscale.com/cmd/tailscaled+

net/batching/conn.go

@@ -32,7 +32,6 @@ type Conn interface {
 	// message may fall on either side of a nonzero.
 	//
 	// Each [ipv6.Message.OOB] must be sized to at least MinControlMessageSize().
-	// len(msgs) must be at least MinReadBatchMsgsLen().
 	ReadBatch(msgs []ipv6.Message, flags int) (n int, err error)
 	// WriteBatchTo writes buffs to addr.
 	//

net/batching/conn_default.go

@@ -19,3 +19,5 @@ var controlMessageSize = 0
 func MinControlMessageSize() int {
 	return controlMessageSize
 }
+
+const IdealBatchSize = 1

net/batching/conn_linux.go

@@ -384,7 +384,7 @@ func setGSOSizeInControl(control *[]byte, gsoSize uint16) {
 }
 
 // TryUpgradeToConn probes the capabilities of the OS and pconn, and upgrades
-// pconn to a [Conn] if appropriate. A batch size of MinReadBatchMsgsLen() is
+// pconn to a [Conn] if appropriate. A batch size of [IdealBatchSize] is
 // suggested for the best performance.
 func TryUpgradeToConn(pconn nettype.PacketConn, network string, batchSize int) nettype.PacketConn {
 	if runtime.GOOS != "linux" {
@@ -457,6 +457,4 @@ func MinControlMessageSize() int {
 	return controlMessageSize
 }
 
-func MinReadBatchMsgsLen() int {
+const IdealBatchSize = 128
-	return 128
-}

net/batching/conn_linux_test.go

@@ -310,7 +310,7 @@ func TestMinReadBatchMsgsLen(t *testing.T) {
 	// So long as magicsock uses [Conn], and [wireguard-go/conn.Bind] API is
 	// shaped for wireguard-go to control packet memory, these values should be
 	// aligned.
-	if MinReadBatchMsgsLen() != conn.IdealBatchSize {
+	if IdealBatchSize != conn.IdealBatchSize {
-		t.Fatalf("MinReadBatchMsgsLen():%d != conn.IdealBatchSize(): %d", MinReadBatchMsgsLen(), conn.IdealBatchSize)
+		t.Fatalf("IdealBatchSize: %d != conn.IdealBatchSize(): %d", IdealBatchSize, conn.IdealBatchSize)
 	}
 }

net/udprelay/server.go

@@ -20,8 +20,11 @@ import (
 	"time"
 
 	"go4.org/mem"
+	"golang.org/x/net/ipv6"
 	"tailscale.com/client/local"
 	"tailscale.com/disco"
+	"tailscale.com/net/batching"
+	"tailscale.com/net/netaddr"
 	"tailscale.com/net/netcheck"
 	"tailscale.com/net/netmon"
 	"tailscale.com/net/packet"
@@ -57,9 +60,9 @@ type Server struct {
 	bindLifetime        time.Duration
 	steadyStateLifetime time.Duration
 	bus                 *eventbus.Bus
-	uc4                 *net.UDPConn // always non-nil
+	uc4                 batching.Conn // always non-nil
 	uc4Port             uint16        // always nonzero
-	uc6                 *net.UDPConn // may be nil if IPv6 bind fails during initialization
+	uc6                 batching.Conn // may be nil if IPv6 bind fails during initialization
 	uc6Port             uint16        // may be zero if IPv6 bind fails during initialization
 	closeOnce           sync.Once
 	wg                  sync.WaitGroup
@@ -96,9 +99,9 @@ type serverEndpoint struct {
 	allocatedAt time.Time
 }
 
-func (e *serverEndpoint) handleDiscoControlMsg(from netip.AddrPort, senderIndex int, discoMsg disco.Message, conn *net.UDPConn, serverDisco key.DiscoPublic) {
+func (e *serverEndpoint) handleDiscoControlMsg(from netip.AddrPort, senderIndex int, discoMsg disco.Message, serverDisco key.DiscoPublic) (write []byte, to netip.AddrPort) {
 	if senderIndex != 0 && senderIndex != 1 {
-		return
+		return nil, netip.AddrPort{}
 	}
 
 	otherSender := 0
@@ -121,15 +124,15 @@ func (e *serverEndpoint) handleDiscoControlMsg(from netip.AddrPort, senderIndex
 		err := validateVNIAndRemoteKey(discoMsg.BindUDPRelayEndpointCommon)
 		if err != nil {
 			// silently drop
-			return
+			return nil, netip.AddrPort{}
 		}
 		if discoMsg.Generation == 0 {
 			// Generation must be nonzero, silently drop
-			return
+			return nil, netip.AddrPort{}
 		}
 		if e.handshakeGeneration[senderIndex] == discoMsg.Generation {
 			// we've seen this generation before, silently drop
-			return
+			return nil, netip.AddrPort{}
 		}
 		e.handshakeGeneration[senderIndex] = discoMsg.Generation
 		e.handshakeAddrPorts[senderIndex] = from
@@ -144,19 +147,18 @@ func (e *serverEndpoint) handleDiscoControlMsg(from netip.AddrPort, senderIndex
 		gh.VNI.Set(e.vni)
 		err = gh.Encode(reply)
 		if err != nil {
-			return
+			return nil, netip.AddrPort{}
 		}
 		reply = append(reply, disco.Magic...)
 		reply = serverDisco.AppendTo(reply)
 		box := e.discoSharedSecrets[senderIndex].Seal(m.AppendMarshal(nil))
 		reply = append(reply, box...)
-		conn.WriteMsgUDPAddrPort(reply, nil, from)
+		return reply, from
-		return
 	case *disco.BindUDPRelayEndpointAnswer:
 		err := validateVNIAndRemoteKey(discoMsg.BindUDPRelayEndpointCommon)
 		if err != nil {
 			// silently drop
-			return
+			return nil, netip.AddrPort{}
 		}
 		generation := e.handshakeGeneration[senderIndex]
 		if generation == 0 || // we have no active handshake
@@ -164,23 +166,23 @@ func (e *serverEndpoint) handleDiscoControlMsg(from netip.AddrPort, senderIndex
 			e.handshakeAddrPorts[senderIndex] != from || // mismatching source for the active handshake
 			!bytes.Equal(e.challenge[senderIndex][:], discoMsg.Challenge[:]) { // mismatching answer for the active handshake
 			// silently drop
-			return
+			return nil, netip.AddrPort{}
 		}
 		// Handshake complete. Update the binding for this sender.
 		e.boundAddrPorts[senderIndex] = from
 		e.lastSeen[senderIndex] = time.Now() // record last seen as bound time
-		return
+		return nil, netip.AddrPort{}
 	default:
 		// unexpected message types, silently drop
-		return
+		return nil, netip.AddrPort{}
 	}
 }
 
-func (e *serverEndpoint) handleSealedDiscoControlMsg(from netip.AddrPort, b []byte, conn *net.UDPConn, serverDisco key.DiscoPublic) {
+func (e *serverEndpoint) handleSealedDiscoControlMsg(from netip.AddrPort, b []byte, serverDisco key.DiscoPublic) (write []byte, to netip.AddrPort) {
 	senderRaw, isDiscoMsg := disco.Source(b)
 	if !isDiscoMsg {
 		// Not a Disco message
-		return
+		return nil, netip.AddrPort{}
 	}
 	sender := key.DiscoPublicFromRaw32(mem.B(senderRaw))
 	senderIndex := -1
@@ -191,63 +193,51 @@ func (e *serverEndpoint) handleSealedDiscoControlMsg(from netip.AddrPort, b []by
 		senderIndex = 1
 	default:
 		// unknown Disco public key
-		return
+		return nil, netip.AddrPort{}
 	}
 
 	const headerLen = len(disco.Magic) + key.DiscoPublicRawLen
 	discoPayload, ok := e.discoSharedSecrets[senderIndex].Open(b[headerLen:])
 	if !ok {
 		// unable to decrypt the Disco payload
-		return
+		return nil, netip.AddrPort{}
 	}
 
 	discoMsg, err := disco.Parse(discoPayload)
 	if err != nil {
 		// unable to parse the Disco payload
-		return
+		return nil, netip.AddrPort{}
 	}
 
-	e.handleDiscoControlMsg(from, senderIndex, discoMsg, conn, serverDisco)
+	return e.handleDiscoControlMsg(from, senderIndex, discoMsg, serverDisco)
 }
 
-func (e *serverEndpoint) handlePacket(from netip.AddrPort, gh packet.GeneveHeader, b []byte, rxSocket, otherAFSocket *net.UDPConn, serverDisco key.DiscoPublic) {
+func (e *serverEndpoint) handlePacket(from netip.AddrPort, gh packet.GeneveHeader, b []byte, serverDisco key.DiscoPublic) (write []byte, to netip.AddrPort) {
 	if !gh.Control {
 		if !e.isBound() {
 			// not a control packet, but serverEndpoint isn't bound
-			return
+			return nil, netip.AddrPort{}
 		}
-		var to netip.AddrPort
 		switch {
 		case from == e.boundAddrPorts[0]:
 			e.lastSeen[0] = time.Now()
-			to = e.boundAddrPorts[1]
+			return b, e.boundAddrPorts[1]
 		case from == e.boundAddrPorts[1]:
 			e.lastSeen[1] = time.Now()
-			to = e.boundAddrPorts[0]
+			return b, e.boundAddrPorts[0]
 		default:
 			// unrecognized source
-			return
+			return nil, netip.AddrPort{}
 		}
-		// Relay the packet towards the other party via the socket associated
-		// with the destination's address family. If source and destination
-		// address families are matching we tx on the same socket the packet
-		// was received (rxSocket), otherwise we use the "other" socket
-		// (otherAFSocket). [Server] makes no use of dual-stack sockets.
-		if from.Addr().Is4() == to.Addr().Is4() {
-			rxSocket.WriteMsgUDPAddrPort(b, nil, to)
-		} else if otherAFSocket != nil {
-			otherAFSocket.WriteMsgUDPAddrPort(b, nil, to)
-		}
-		return
 	}
 
 	if gh.Protocol != packet.GeneveProtocolDisco {
 		// control packet, but not Disco
-		return
+		return nil, netip.AddrPort{}
 	}
 
 	msg := b[packet.GeneveFixedHeaderLength:]
-	e.handleSealedDiscoControlMsg(from, msg, rxSocket, serverDisco)
+	return e.handleSealedDiscoControlMsg(from, msg, serverDisco)
 }
 
 func (e *serverEndpoint) isExpired(now time.Time, bindLifetime, steadyStateLifetime time.Duration) bool {
@@ -338,10 +328,10 @@ func NewServer(logf logger.Logf, port int, overrideAddrs []netip.Addr) (s *Serve
 	}
 
 	s.wg.Add(1)
-	go s.packetReadLoop(s.uc4, s.uc6)
+	go s.packetReadLoop(s.uc4, s.uc6, true)
 	if s.uc6 != nil {
 		s.wg.Add(1)
-		go s.packetReadLoop(s.uc6, s.uc4)
+		go s.packetReadLoop(s.uc6, s.uc4, false)
 	}
 	s.wg.Add(1)
 	go s.endpointGCLoop()
@@ -425,6 +415,41 @@ func (s *Server) addrDiscoveryLoop() {
 	}
 }
 
+// This is a compile-time assertion that [singlePacketConn] implements the
+// [batching.Conn] interface.
+var _ batching.Conn = (*singlePacketConn)(nil)
+
+// singlePacketConn implements [batching.Conn] with single packet syscall
+// operations.
+type singlePacketConn struct {
+	*net.UDPConn
+}
+
+func (c *singlePacketConn) ReadBatch(msgs []ipv6.Message, _ int) (int, error) {
+	n, ap, err := c.UDPConn.ReadFromUDPAddrPort(msgs[0].Buffers[0])
+	if err != nil {
+		return 0, err
+	}
+	msgs[0].N = n
+	msgs[0].Addr = net.UDPAddrFromAddrPort(netaddr.Unmap(ap))
+	return 1, nil
+}
+
+func (c *singlePacketConn) WriteBatchTo(buffs [][]byte, addr netip.AddrPort, geneve packet.GeneveHeader, offset int) error {
+	for _, buff := range buffs {
+		if geneve.VNI.IsSet() {
+			geneve.Encode(buff)
+		} else {
+			buff = buff[offset:]
+		}
+		_, err := c.UDPConn.WriteToUDPAddrPort(buff, addr)
+		if err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
 // listenOn binds an IPv4 and IPv6 socket to port. We consider it successful if
 // we manage to bind the IPv4 socket.
 //
@@ -433,7 +458,10 @@ func (s *Server) addrDiscoveryLoop() {
 // across IPv4 and IPv6 if the requested port is zero.
 //
 // TODO: make these "re-bindable" in similar fashion to magicsock as a means to
-// deal with EDR software closing them. http://go/corp/30118
+// deal with EDR software closing them. http://go/corp/30118. We could re-use
+// [magicsock.RebindingConn], which would also remove the need for
+// [singlePacketConn], as [magicsock.RebindingConn] also handles fallback to
+// single packet syscall operations.
 func (s *Server) listenOn(port int) error {
 	for _, network := range []string{"udp4", "udp6"} {
 		uc, err := net.ListenUDP(network, &net.UDPAddr{Port: port})
@@ -462,11 +490,16 @@ func (s *Server) listenOn(port int) error {
 			}
 			return err
 		}
+		pc := batching.TryUpgradeToConn(uc, network, batching.IdealBatchSize)
+		bc, ok := pc.(batching.Conn)
+		if !ok {
+			bc = &singlePacketConn{uc}
+		}
 		if network == "udp4" {
-			s.uc4 = uc
+			s.uc4 = bc
 			s.uc4Port = uint16(portUint)
 		} else {
-			s.uc6 = uc
+			s.uc6 = bc
 			s.uc6Port = uint16(portUint)
 		}
 	}
@@ -526,18 +559,18 @@ func (s *Server) endpointGCLoop() {
 	}
 }
 
-func (s *Server) handlePacket(from netip.AddrPort, b []byte, rxSocket, otherAFSocket *net.UDPConn) {
+func (s *Server) handlePacket(from netip.AddrPort, b []byte) (write []byte, to netip.AddrPort) {
 	if stun.Is(b) && b[1] == 0x01 {
 		// A b[1] value of 0x01 (STUN method binding) is sufficiently
 		// non-overlapping with the Geneve header where the LSB is always 0
 		// (part of 6 "reserved" bits).
 		s.netChecker.ReceiveSTUNPacket(b, from)
-		return
+		return nil, netip.AddrPort{}
 	}
 	gh := packet.GeneveHeader{}
 	err := gh.Decode(b)
 	if err != nil {
-		return
+		return nil, netip.AddrPort{}
 	}
 	// TODO: consider performance implications of holding s.mu for the remainder
 	// of this method, which does a bunch of disco/crypto work depending. Keep
@@ -547,13 +580,13 @@ func (s *Server) handlePacket(from netip.AddrPort, b []byte, rxSocket, otherAFSo
 	e, ok := s.byVNI[gh.VNI.Get()]
 	if !ok {
 		// unknown VNI
-		return
+		return nil, netip.AddrPort{}
 	}
 
-	e.handlePacket(from, gh, b, rxSocket, otherAFSocket, s.discoPublic)
+	return e.handlePacket(from, gh, b, s.discoPublic)
 }
 
-func (s *Server) packetReadLoop(readFromSocket, otherSocket *net.UDPConn) {
+func (s *Server) packetReadLoop(readFromSocket, otherSocket batching.Conn, readFromSocketIsIPv4 bool) {
 	defer func() {
 		// We intentionally close the [Server] if we encounter a socket read
 		// error below, at least until socket "re-binding" is implemented as
@@ -564,15 +597,73 @@ func (s *Server) packetReadLoop(readFromSocket, otherSocket *net.UDPConn) {
 		s.wg.Done()
 		s.Close()
 	}()
-	b := make([]byte, 1<<16-1)
+
+	msgs := make([]ipv6.Message, batching.IdealBatchSize)
+	for i := range msgs {
+		msgs[i].OOB = make([]byte, batching.MinControlMessageSize())
+		msgs[i].Buffers = make([][]byte, 1)
+		msgs[i].Buffers[0] = make([]byte, 1<<16-1)
+	}
+	writeBuffsByDest := make(map[netip.AddrPort][][]byte, batching.IdealBatchSize)
+
 	for {
+		for i := range msgs {
+			msgs[i] = ipv6.Message{Buffers: msgs[i].Buffers, OOB: msgs[i].OOB[:cap(msgs[i].OOB)]}
+		}
+
 		// TODO: extract laddr from IP_PKTINFO for use in reply
-		n, from, err := readFromSocket.ReadFromUDPAddrPort(b)
+		// ReadBatch will split coalesced datagrams before returning, which
+		// WriteBatchTo will re-coalesce further down. We _could_ be more
+		// efficient and not split datagrams that belong to the same VNI if they
+		// are non-control/handshake packets. We pay the memmove/memcopy
+		// performance penalty for now in the interest of simple single packet
+		// handlers.
+		n, err := readFromSocket.ReadBatch(msgs, 0)
 		if err != nil {
 			s.logf("error reading from socket(%v): %v", readFromSocket.LocalAddr(), err)
 			return
 		}
-		s.handlePacket(from, b[:n], readFromSocket, otherSocket)
+
+		for _, msg := range msgs[:n] {
+			if msg.N == 0 {
+				continue
+			}
+			buf := msg.Buffers[0][:msg.N]
+			from := msg.Addr.(*net.UDPAddr).AddrPort()
+			write, to := s.handlePacket(from, buf)
+			if !to.IsValid() {
+				continue
+			}
+			if from.Addr().Is4() == to.Addr().Is4() || otherSocket != nil {
+				buffs, ok := writeBuffsByDest[to]
+				if !ok {
+					buffs = make([][]byte, 0, batching.IdealBatchSize)
+				}
+				buffs = append(buffs, write)
+				writeBuffsByDest[to] = buffs
+			} else {
+				// This is unexpected. We should never produce a packet to write
+				// to the "other" socket if the other socket is nil/unbound.
+				// [server.handlePacket] has to see a packet from a particular
+				// address family at least once in order for it to return a
+				// packet to write towards a dest for the same address family.
+				s.logf("[unexpected] packet from: %v produced packet to: %v while otherSocket is nil", from, to)
+			}
+		}
+
+		for dest, buffs := range writeBuffsByDest {
+			// Write the packet batches via the socket associated with the
+			// destination's address family. If source and destination address
+			// families are matching we tx on the same socket the packet was
+			// received, otherwise we use the "other" socket. [Server] makes no
+			// use of dual-stack sockets.
+			if dest.Addr().Is4() == readFromSocketIsIPv4 {
+				readFromSocket.WriteBatchTo(buffs, dest, packet.GeneveHeader{}, 0)
+			} else {
+				otherSocket.WriteBatchTo(buffs, dest, packet.GeneveHeader{}, 0)
+			}
+			delete(writeBuffsByDest, dest)
+		}
 	}
 }