Add support for setting TCP segment hash.

This allows the link layer endpoints to consistenly hash a TCP
segment to a single underlying queue in case a link layer endpoint
does support multiple underlying queues.

Updates #231

PiperOrigin-RevId: 302760664

pkg/tcpip/link/fdbased/endpoint.go

@@ -405,6 +405,7 @@ func (e *endpoint) WritePacket(r *stack.Route, gso *stack.GSO, protocol tcpip.Ne
 		eth.Encode(ethHdr)
 	}
 
+	fd := e.fds[pkt.Hash%uint32(len(e.fds))]
 	if e.Capabilities()&stack.CapabilityHardwareGSO != 0 {
 		vnetHdr := virtioNetHdr{}
 		if gso != nil {
@@ -428,14 +429,14 @@ func (e *endpoint) WritePacket(r *stack.Route, gso *stack.GSO, protocol tcpip.Ne
 		}
 
 		vnetHdrBuf := vnetHdrToByteSlice(&vnetHdr)
-		return rawfile.NonBlockingWrite3(e.fds[0], vnetHdrBuf, pkt.Header.View(), pkt.Data.ToView())
+		return rawfile.NonBlockingWrite3(fd, vnetHdrBuf, pkt.Header.View(), pkt.Data.ToView())
 	}
 
 	if pkt.Data.Size() == 0 {
-		return rawfile.NonBlockingWrite(e.fds[0], pkt.Header.View())
+		return rawfile.NonBlockingWrite(fd, pkt.Header.View())
 	}
 
-	return rawfile.NonBlockingWrite3(e.fds[0], pkt.Header.View(), pkt.Data.ToView(), nil)
+	return rawfile.NonBlockingWrite3(fd, pkt.Header.View(), pkt.Data.ToView(), nil)
 }
 
 // WritePackets writes outbound packets to the file descriptor. If it is not
@@ -551,7 +552,8 @@ func (e *endpoint) WritePackets(r *stack.Route, gso *stack.GSO, pkts []stack.Pac
 
 	packets := 0
 	for packets < n {
-		sent, err := rawfile.NonBlockingSendMMsg(e.fds[0], mmsgHdrs)
+		fd := e.fds[pkts[packets].Hash%uint32(len(e.fds))]
+		sent, err := rawfile.NonBlockingSendMMsg(fd, mmsgHdrs)
 		if err != nil {
 			return packets, err
 		}

pkg/tcpip/link/fdbased/endpoint_test.go

@@ -49,36 +49,42 @@ type packetInfo struct {
 }
 
 type context struct {
-	t    *testing.T
-	fds  [2]int
-	ep   stack.LinkEndpoint
-	ch   chan packetInfo
-	done chan struct{}
+	t        *testing.T
+	readFDs  []int
+	writeFDs []int
+	ep       stack.LinkEndpoint
+	ch       chan packetInfo
+	done     chan struct{}
 }
 
 func newContext(t *testing.T, opt *Options) *context {
-	fds, err := syscall.Socketpair(syscall.AF_UNIX, syscall.SOCK_SEQPACKET, 0)
+	firstFDPair, err := syscall.Socketpair(syscall.AF_UNIX, syscall.SOCK_SEQPACKET, 0)
+	if err != nil {
+		t.Fatalf("Socketpair failed: %v", err)
+	}
+	secondFDPair, err := syscall.Socketpair(syscall.AF_UNIX, syscall.SOCK_SEQPACKET, 0)
 	if err != nil {
 		t.Fatalf("Socketpair failed: %v", err)
 	}
 
-	done := make(chan struct{}, 1)
+	done := make(chan struct{}, 2)
 	opt.ClosedFunc = func(*tcpip.Error) {
 		done <- struct{}{}
 	}
 
-	opt.FDs = []int{fds[1]}
+	opt.FDs = []int{firstFDPair[1], secondFDPair[1]}
 	ep, err := New(opt)
 	if err != nil {
 		t.Fatalf("Failed to create FD endpoint: %v", err)
 	}
 
 	c := &context{
-		t:    t,
-		fds:  fds,
-		ep:   ep,
-		ch:   make(chan packetInfo, 100),
-		done: done,
+		t:        t,
+		readFDs:  []int{firstFDPair[0], secondFDPair[0]},
+		writeFDs: opt.FDs,
+		ep:       ep,
+		ch:       make(chan packetInfo, 100),
+		done:     done,
 	}
 
 	ep.Attach(c)
@@ -87,9 +93,14 @@ func newContext(t *testing.T, opt *Options) *context {
 }
 
 func (c *context) cleanup() {
-	syscall.Close(c.fds[0])
+	for _, fd := range c.readFDs {
+		syscall.Close(fd)
+	}
 	<-c.done
-	syscall.Close(c.fds[1])
+	<-c.done
+	for _, fd := range c.writeFDs {
+		syscall.Close(fd)
+	}
 }
 
 func (c *context) DeliverNetworkPacket(linkEP stack.LinkEndpoint, remote tcpip.LinkAddress, local tcpip.LinkAddress, protocol tcpip.NetworkProtocolNumber, pkt stack.PacketBuffer) {
@@ -136,7 +147,7 @@ func TestAddress(t *testing.T) {
 	}
 }
 
-func testWritePacket(t *testing.T, plen int, eth bool, gsoMaxSize uint32) {
+func testWritePacket(t *testing.T, plen int, eth bool, gsoMaxSize uint32, hash uint32) {
 	c := newContext(t, &Options{Address: laddr, MTU: mtu, EthernetHeader: eth, GSOMaxSize: gsoMaxSize})
 	defer c.cleanup()
 
@@ -171,13 +182,15 @@ func testWritePacket(t *testing.T, plen int, eth bool, gsoMaxSize uint32) {
 	if err := c.ep.WritePacket(r, gso, proto, stack.PacketBuffer{
 		Header: hdr,
 		Data:   payload.ToVectorisedView(),
+		Hash:   hash,
 	}); err != nil {
 		t.Fatalf("WritePacket failed: %v", err)
 	}
 
-	// Read from fd, then compare with what we wrote.
+	// Read from the corresponding FD, then compare with what we wrote.
 	b = make([]byte, mtu)
-	n, err := syscall.Read(c.fds[0], b)
+	fd := c.readFDs[hash%uint32(len(c.readFDs))]
+	n, err := syscall.Read(fd, b)
 	if err != nil {
 		t.Fatalf("Read failed: %v", err)
 	}
@@ -238,7 +251,7 @@ func TestWritePacket(t *testing.T) {
 				t.Run(
 					fmt.Sprintf("Eth=%v,PayloadLen=%v,GSOMaxSize=%v", eth, plen, gso),
 					func(t *testing.T) {
-						testWritePacket(t, plen, eth, gso)
+						testWritePacket(t, plen, eth, gso, 0)
 					},
 				)
 			}
@@ -246,6 +259,27 @@ func TestWritePacket(t *testing.T) {
 	}
 }
 
+func TestHashedWritePacket(t *testing.T) {
+	lengths := []int{0, 100, 1000}
+	eths := []bool{true, false}
+	gsos := []uint32{0, 32768}
+	hashes := []uint32{0, 1}
+	for _, eth := range eths {
+		for _, plen := range lengths {
+			for _, gso := range gsos {
+				for _, hash := range hashes {
+					t.Run(
+						fmt.Sprintf("Eth=%v,PayloadLen=%v,GSOMaxSize=%v,Hash=%d", eth, plen, gso, hash),
+						func(t *testing.T) {
+							testWritePacket(t, plen, eth, gso, hash)
+						},
+					)
+				}
+			}
+		}
+	}
+}
+
 func TestPreserveSrcAddress(t *testing.T) {
 	baddr := tcpip.LinkAddress("\xcc\xbb\xaa\x77\x88\x99")
 
@@ -270,7 +304,7 @@ func TestPreserveSrcAddress(t *testing.T) {
 
 	// Read from the FD, then compare with what we wrote.
 	b := make([]byte, mtu)
-	n, err := syscall.Read(c.fds[0], b)
+	n, err := syscall.Read(c.readFDs[0], b)
 	if err != nil {
 		t.Fatalf("Read failed: %v", err)
 	}
@@ -314,7 +348,7 @@ func TestDeliverPacket(t *testing.T) {
 				}
 
 				// Write packet via the file descriptor.
-				if _, err := syscall.Write(c.fds[0], all); err != nil {
+				if _, err := syscall.Write(c.readFDs[0], all); err != nil {
 					t.Fatalf("Write failed: %v", err)
 				}
 

pkg/tcpip/stack/BUILD

@@ -30,6 +30,7 @@ go_library(
         "nic.go",
         "packet_buffer.go",
         "packet_buffer_state.go",
+        "rand.go",
         "registration.go",
         "route.go",
         "stack.go",

pkg/tcpip/stack/packet_buffer.go

@@ -10,6 +10,7 @@
 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 // See the License for the specific language governing permissions and
 // limitations under the License.
+
 package stack
 
 import "gvisor.dev/gvisor/pkg/tcpip/buffer"
@@ -54,6 +55,10 @@ type PacketBuffer struct {
 	LinkHeader      buffer.View
 	NetworkHeader   buffer.View
 	TransportHeader buffer.View
+
+	// Hash is the transport layer hash of this packet. A value of zero
+	// indicates no valid hash has been set.
+	Hash uint32
 }
 
 // Clone makes a copy of pk. It clones the Data field, which creates a new

pkg/tcpip/stack/packet_buffer_state.go

@@ -11,6 +11,7 @@
 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 // See the License for the specific language governing permissions and
 // limitations under the License.
+
 package stack
 
 import "gvisor.dev/gvisor/pkg/tcpip/buffer"

pkg/tcpip/stack/rand.go

@@ -0,0 +1,40 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package stack
+
+import (
+	mathrand "math/rand"
+
+	"gvisor.dev/gvisor/pkg/sync"
+)
+
+// lockedRandomSource provides a threadsafe rand.Source.
+type lockedRandomSource struct {
+	mu  sync.Mutex
+	src mathrand.Source
+}
+
+func (r *lockedRandomSource) Int63() (n int64) {
+	r.mu.Lock()
+	n = r.src.Int63()
+	r.mu.Unlock()
+	return n
+}
+
+func (r *lockedRandomSource) Seed(seed int64) {
+	r.mu.Lock()
+	r.src.Seed(seed)
+	r.mu.Unlock()
+}

pkg/tcpip/stack/stack.go

@@ -20,7 +20,9 @@
 package stack
 
 import (
+	"bytes"
 	"encoding/binary"
+	mathrand "math/rand"
 	"sync/atomic"
 	"time"
 
@@ -465,6 +467,10 @@ type Stack struct {
 	// forwarder holds the packets that wait for their link-address resolutions
 	// to complete, and forwards them when each resolution is done.
 	forwarder *forwardQueue
+
+	// randomGenerator is an injectable pseudo random generator that can be
+	// used when a random number is required.
+	randomGenerator *mathrand.Rand
 }
 
 // UniqueID is an abstract generator of unique identifiers.
@@ -525,9 +531,16 @@ type Options struct {
 	// this is non-nil.
 	RawFactory RawFactory
 
-	// OpaqueIIDOpts hold the options for generating opaque interface identifiers
-	// (IIDs) as outlined by RFC 7217.
+	// OpaqueIIDOpts hold the options for generating opaque interface
+	// identifiers (IIDs) as outlined by RFC 7217.
 	OpaqueIIDOpts OpaqueInterfaceIdentifierOptions
+
+	// RandSource is an optional source to use to generate random
+	// numbers. If omitted it defaults to a Source seeded by the data
+	// returned by rand.Read().
+	//
+	// RandSource must be thread-safe.
+	RandSource mathrand.Source
 }
 
 // TransportEndpointInfo holds useful information about a transport endpoint
@@ -623,6 +636,13 @@ func New(opts Options) *Stack {
 		opts.UniqueID = new(uniqueIDGenerator)
 	}
 
+	randSrc := opts.RandSource
+	if randSrc == nil {
+		// Source provided by mathrand.NewSource is not thread-safe so
+		// we wrap it in a simple thread-safe version.
+		randSrc = &lockedRandomSource{src: mathrand.NewSource(generateRandInt64())}
+	}
+
 	// Make sure opts.NDPConfigs contains valid values only.
 	opts.NDPConfigs.validate()
 
@@ -645,6 +665,7 @@ func New(opts Options) *Stack {
 		ndpDisp:              opts.NDPDisp,
 		opaqueIIDOpts:        opts.OpaqueIIDOpts,
 		forwarder:            newForwardQueue(),
+		randomGenerator:      mathrand.New(randSrc),
 	}
 
 	// Add specified network protocols.
@@ -1818,6 +1839,12 @@ func (s *Stack) Seed() uint32 {
 	return s.seed
 }
 
+// Rand returns a reference to a pseudo random generator that can be used
+// to generate random numbers as required.
+func (s *Stack) Rand() *mathrand.Rand {
+	return s.randomGenerator
+}
+
 func generateRandUint32() uint32 {
 	b := make([]byte, 4)
 	if _, err := rand.Read(b); err != nil {
@@ -1825,3 +1852,16 @@ func generateRandUint32() uint32 {
 	}
 	return binary.LittleEndian.Uint32(b)
 }
+
+func generateRandInt64() int64 {
+	b := make([]byte, 8)
+	if _, err := rand.Read(b); err != nil {
+		panic(err)
+	}
+	buf := bytes.NewReader(b)
+	var v int64
+	if err := binary.Read(buf, binary.LittleEndian, &v); err != nil {
+		panic(err)
+	}
+	return v
+}

pkg/tcpip/transport/tcp/accept.go

@@ -445,7 +445,15 @@ func (e *endpoint) handleListenSegment(ctx *listenContext, s *segment) {
 		// RFC 793 section 3.4 page 35 (figure 12) outlines that a RST
 		// must be sent in response to a SYN-ACK while in the listen
 		// state to prevent completing a handshake from an old SYN.
-		e.sendTCP(&s.route, s.id, buffer.VectorisedView{}, e.ttl, e.sendTOS, header.TCPFlagRst, s.ackNumber, 0, 0, nil, nil)
+		e.sendTCP(&s.route, tcpFields{
+			id:     s.id,
+			ttl:    e.ttl,
+			tos:    e.sendTOS,
+			flags:  header.TCPFlagRst,
+			seq:    s.ackNumber,
+			ack:    0,
+			rcvWnd: 0,
+		}, buffer.VectorisedView{}, nil)
 		return
 	}
 
@@ -493,7 +501,15 @@ func (e *endpoint) handleListenSegment(ctx *listenContext, s *segment) {
 				TSEcr: opts.TSVal,
 				MSS:   mssForRoute(&s.route),
 			}
-			e.sendSynTCP(&s.route, s.id, e.ttl, e.sendTOS, header.TCPFlagSyn|header.TCPFlagAck, cookie, s.sequenceNumber+1, ctx.rcvWnd, synOpts)
+			e.sendSynTCP(&s.route, tcpFields{
+				id:     s.id,
+				ttl:    e.ttl,
+				tos:    e.sendTOS,
+				flags:  header.TCPFlagSyn | header.TCPFlagAck,
+				seq:    cookie,
+				ack:    s.sequenceNumber + 1,
+				rcvWnd: ctx.rcvWnd,
+			}, synOpts)
 			e.stack.Stats().TCP.ListenOverflowSynCookieSent.Increment()
 		}
 

pkg/tcpip/transport/tcp/connect.go

@@ -308,7 +308,15 @@ func (h *handshake) synSentState(s *segment) *tcpip.Error {
 	if ttl == 0 {
 		ttl = s.route.DefaultTTL()
 	}
-	h.ep.sendSynTCP(&s.route, h.ep.ID, ttl, h.ep.sendTOS, h.flags, h.iss, h.ackNum, h.rcvWnd, synOpts)
+	h.ep.sendSynTCP(&s.route, tcpFields{
+		id:     h.ep.ID,
+		ttl:    ttl,
+		tos:    h.ep.sendTOS,
+		flags:  h.flags,
+		seq:    h.iss,
+		ack:    h.ackNum,
+		rcvWnd: h.rcvWnd,
+	}, synOpts)
 	return nil
 }
 
@@ -361,7 +369,15 @@ func (h *handshake) synRcvdState(s *segment) *tcpip.Error {
 			SACKPermitted: h.ep.sackPermitted,
 			MSS:           h.ep.amss,
 		}
-		h.ep.sendSynTCP(&s.route, h.ep.ID, h.ep.ttl, h.ep.sendTOS, h.flags, h.iss, h.ackNum, h.rcvWnd, synOpts)
+		h.ep.sendSynTCP(&s.route, tcpFields{
+			id:     h.ep.ID,
+			ttl:    h.ep.ttl,
+			tos:    h.ep.sendTOS,
+			flags:  h.flags,
+			seq:    h.iss,
+			ack:    h.ackNum,
+			rcvWnd: h.rcvWnd,
+		}, synOpts)
 		return nil
 	}
 
@@ -550,7 +566,16 @@ func (h *handshake) execute() *tcpip.Error {
 			synOpts.WS = -1
 		}
 	}
-	h.ep.sendSynTCP(&h.ep.route, h.ep.ID, h.ep.ttl, h.ep.sendTOS, h.flags, h.iss, h.ackNum, h.rcvWnd, synOpts)
+
+	h.ep.sendSynTCP(&h.ep.route, tcpFields{
+		id:     h.ep.ID,
+		ttl:    h.ep.ttl,
+		tos:    h.ep.sendTOS,
+		flags:  h.flags,
+		seq:    h.iss,
+		ack:    h.ackNum,
+		rcvWnd: h.rcvWnd,
+	}, synOpts)
 
 	for h.state != handshakeCompleted {
 		h.ep.mu.Unlock()
@@ -573,7 +598,15 @@ func (h *handshake) execute() *tcpip.Error {
 			// the connection with another ACK or data (as ACKs are never
 			// retransmitted on their own).
 			if h.active || !h.acked || h.deferAccept != 0 && time.Since(h.startTime) > h.deferAccept {
-				h.ep.sendSynTCP(&h.ep.route, h.ep.ID, h.ep.ttl, h.ep.sendTOS, h.flags, h.iss, h.ackNum, h.rcvWnd, synOpts)
+				h.ep.sendSynTCP(&h.ep.route, tcpFields{
+					id:     h.ep.ID,
+					ttl:    h.ep.ttl,
+					tos:    h.ep.sendTOS,
+					flags:  h.flags,
+					seq:    h.iss,
+					ack:    h.ackNum,
+					rcvWnd: h.rcvWnd,
+				}, synOpts)
 			}
 
 		case wakerForNotification:
@@ -686,18 +719,33 @@ func makeSynOptions(opts header.TCPSynOptions) []byte {
 	return options[:offset]
 }
 
-func (e *endpoint) sendSynTCP(r *stack.Route, id stack.TransportEndpointID, ttl, tos uint8, flags byte, seq, ack seqnum.Value, rcvWnd seqnum.Size, opts header.TCPSynOptions) *tcpip.Error {
-	options := makeSynOptions(opts)
+// tcpFields is a struct to carry different parameters required by the
+// send*TCP variant functions below.
+type tcpFields struct {
+	id     stack.TransportEndpointID
+	ttl    uint8
+	tos    uint8
+	flags  byte
+	seq    seqnum.Value
+	ack    seqnum.Value
+	rcvWnd seqnum.Size
+	opts   []byte
+	txHash uint32
+}
+
+func (e *endpoint) sendSynTCP(r *stack.Route, tf tcpFields, opts header.TCPSynOptions) *tcpip.Error {
+	tf.opts = makeSynOptions(opts)
 	// We ignore SYN send errors and let the callers re-attempt send.
-	if err := e.sendTCP(r, id, buffer.VectorisedView{}, ttl, tos, flags, seq, ack, rcvWnd, options, nil); err != nil {
+	if err := e.sendTCP(r, tf, buffer.VectorisedView{}, nil); err != nil {
 		e.stats.SendErrors.SynSendToNetworkFailed.Increment()
 	}
-	putOptions(options)
+	putOptions(tf.opts)
 	return nil
 }
 
-func (e *endpoint) sendTCP(r *stack.Route, id stack.TransportEndpointID, data buffer.VectorisedView, ttl, tos uint8, flags byte, seq, ack seqnum.Value, rcvWnd seqnum.Size, opts []byte, gso *stack.GSO) *tcpip.Error {
-	if err := sendTCP(r, id, data, ttl, tos, flags, seq, ack, rcvWnd, opts, gso); err != nil {
+func (e *endpoint) sendTCP(r *stack.Route, tf tcpFields, data buffer.VectorisedView, gso *stack.GSO) *tcpip.Error {
+	tf.txHash = e.txHash
+	if err := sendTCP(r, tf, data, gso); err != nil {
 		e.stats.SendErrors.SegmentSendToNetworkFailed.Increment()
 		return err
 	}
@@ -705,8 +753,8 @@ func (e *endpoint) sendTCP(r *stack.Route, id stack.TransportEndpointID, data bu
 	return nil
 }
 
-func buildTCPHdr(r *stack.Route, id stack.TransportEndpointID, pkt *stack.PacketBuffer, flags byte, seq, ack seqnum.Value, rcvWnd seqnum.Size, opts []byte, gso *stack.GSO) {
-	optLen := len(opts)
+func buildTCPHdr(r *stack.Route, tf tcpFields, pkt *stack.PacketBuffer, gso *stack.GSO) {
+	optLen := len(tf.opts)
 	hdr := &pkt.Header
 	packetSize := pkt.DataSize
 	off := pkt.DataOffset
@@ -714,15 +762,15 @@ func buildTCPHdr(r *stack.Route, id stack.TransportEndpointID, pkt *stack.Packet
 	tcp := header.TCP(hdr.Prepend(header.TCPMinimumSize + optLen))
 	pkt.TransportHeader = buffer.View(tcp)
 	tcp.Encode(&header.TCPFields{
-		SrcPort:    id.LocalPort,
-		DstPort:    id.RemotePort,
-		SeqNum:     uint32(seq),
-		AckNum:     uint32(ack),
+		SrcPort:    tf.id.LocalPort,
+		DstPort:    tf.id.RemotePort,
+		SeqNum:     uint32(tf.seq),
+		AckNum:     uint32(tf.ack),
 		DataOffset: uint8(header.TCPMinimumSize + optLen),
-		Flags:      flags,
-		WindowSize: uint16(rcvWnd),
+		Flags:      tf.flags,
+		WindowSize: uint16(tf.rcvWnd),
 	})
-	copy(tcp[header.TCPMinimumSize:], opts)
+	copy(tcp[header.TCPMinimumSize:], tf.opts)
 
 	length := uint16(hdr.UsedLength() + packetSize)
 	xsum := r.PseudoHeaderChecksum(ProtocolNumber, length)
@@ -737,13 +785,12 @@ func buildTCPHdr(r *stack.Route, id stack.TransportEndpointID, pkt *stack.Packet
 		xsum = header.ChecksumVVWithOffset(pkt.Data, xsum, off, packetSize)
 		tcp.SetChecksum(^tcp.CalculateChecksum(xsum))
 	}
-
 }
 
-func sendTCPBatch(r *stack.Route, id stack.TransportEndpointID, data buffer.VectorisedView, ttl, tos uint8, flags byte, seq, ack seqnum.Value, rcvWnd seqnum.Size, opts []byte, gso *stack.GSO) *tcpip.Error {
-	optLen := len(opts)
-	if rcvWnd > 0xffff {
-		rcvWnd = 0xffff
+func sendTCPBatch(r *stack.Route, tf tcpFields, data buffer.VectorisedView, gso *stack.GSO) *tcpip.Error {
+	optLen := len(tf.opts)
+	if tf.rcvWnd > 0xffff {
+		tf.rcvWnd = 0xffff
 	}
 
 	mss := int(gso.MSS)
@@ -768,14 +815,15 @@ func sendTCPBatch(r *stack.Route, id stack.TransportEndpointID, data buffer.Vect
 		pkts[i].DataOffset = off
 		pkts[i].DataSize = packetSize
 		pkts[i].Data = data
-		buildTCPHdr(r, id, &pkts[i], flags, seq, ack, rcvWnd, opts, gso)
+		pkts[i].Hash = tf.txHash
+		buildTCPHdr(r, tf, &pkts[i], gso)
 		off += packetSize
-		seq = seq.Add(seqnum.Size(packetSize))
+		tf.seq = tf.seq.Add(seqnum.Size(packetSize))
 	}
-	if ttl == 0 {
-		ttl = r.DefaultTTL()
+	if tf.ttl == 0 {
+		tf.ttl = r.DefaultTTL()
 	}
-	sent, err := r.WritePackets(gso, pkts, stack.NetworkHeaderParams{Protocol: ProtocolNumber, TTL: ttl, TOS: tos})
+	sent, err := r.WritePackets(gso, pkts, stack.NetworkHeaderParams{Protocol: ProtocolNumber, TTL: tf.ttl, TOS: tf.tos})
 	if err != nil {
 		r.Stats().TCP.SegmentSendErrors.IncrementBy(uint64(n - sent))
 	}
@@ -785,14 +833,14 @@ func sendTCPBatch(r *stack.Route, id stack.TransportEndpointID, data buffer.Vect
 
 // sendTCP sends a TCP segment with the provided options via the provided
 // network endpoint and under the provided identity.
-func sendTCP(r *stack.Route, id stack.TransportEndpointID, data buffer.VectorisedView, ttl, tos uint8, flags byte, seq, ack seqnum.Value, rcvWnd seqnum.Size, opts []byte, gso *stack.GSO) *tcpip.Error {
-	optLen := len(opts)
-	if rcvWnd > 0xffff {
-		rcvWnd = 0xffff
+func sendTCP(r *stack.Route, tf tcpFields, data buffer.VectorisedView, gso *stack.GSO) *tcpip.Error {
+	optLen := len(tf.opts)
+	if tf.rcvWnd > 0xffff {
+		tf.rcvWnd = 0xffff
 	}
 
 	if r.Loop&stack.PacketLoop == 0 && gso != nil && gso.Type == stack.GSOSW && int(gso.MSS) < data.Size() {
-		return sendTCPBatch(r, id, data, ttl, tos, flags, seq, ack, rcvWnd, opts, gso)
+		return sendTCPBatch(r, tf, data, gso)
 	}
 
 	pkt := stack.PacketBuffer{
@@ -800,18 +848,19 @@ func sendTCP(r *stack.Route, id stack.TransportEndpointID, data buffer.Vectorise
 		DataOffset: 0,
 		DataSize:   data.Size(),
 		Data:       data,
+		Hash:       tf.txHash,
 	}
-	buildTCPHdr(r, id, &pkt, flags, seq, ack, rcvWnd, opts, gso)
+	buildTCPHdr(r, tf, &pkt, gso)
 
-	if ttl == 0 {
-		ttl = r.DefaultTTL()
+	if tf.ttl == 0 {
+		tf.ttl = r.DefaultTTL()
 	}
-	if err := r.WritePacket(gso, stack.NetworkHeaderParams{Protocol: ProtocolNumber, TTL: ttl, TOS: tos}, pkt); err != nil {
+	if err := r.WritePacket(gso, stack.NetworkHeaderParams{Protocol: ProtocolNumber, TTL: tf.ttl, TOS: tf.tos}, pkt); err != nil {
 		r.Stats().TCP.SegmentSendErrors.Increment()
 		return err
 	}
 	r.Stats().TCP.SegmentsSent.Increment()
-	if (flags & header.TCPFlagRst) != 0 {
+	if (tf.flags & header.TCPFlagRst) != 0 {
 		r.Stats().TCP.ResetsSent.Increment()
 	}
 	return nil
@@ -863,7 +912,16 @@ func (e *endpoint) sendRaw(data buffer.VectorisedView, flags byte, seq, ack seqn
 		sackBlocks = e.sack.Blocks[:e.sack.NumBlocks]
 	}
 	options := e.makeOptions(sackBlocks)
-	err := e.sendTCP(&e.route, e.ID, data, e.ttl, e.sendTOS, flags, seq, ack, rcvWnd, options, e.gso)
+	err := e.sendTCP(&e.route, tcpFields{
+		id:     e.ID,
+		ttl:    e.ttl,
+		tos:    e.sendTOS,
+		flags:  flags,
+		seq:    seq,
+		ack:    ack,
+		rcvWnd: rcvWnd,
+		opts:   options,
+	}, data, e.gso)
 	putOptions(options)
 	return err
 }

pkg/tcpip/transport/tcp/endpoint.go

@@ -581,6 +581,10 @@ type endpoint struct {
 	// endpoint and at this point the endpoint is only around
 	// to complete the TCP shutdown.
 	closed bool
+
+	// txHash is the transport layer hash to be set on outbound packets
+	// emitted by this endpoint.
+	txHash uint32
 }
 
 // UniqueID implements stack.TransportEndpoint.UniqueID.
@@ -771,6 +775,7 @@ func newEndpoint(s *stack.Stack, netProto tcpip.NetworkProtocolNumber, waiterQue
 			count:    9,
 		},
 		uniqueID: s.UniqueID(),
+		txHash:   s.Rand().Uint32(),
 	}
 
 	var ss SendBufferSizeOption

pkg/tcpip/transport/tcp/protocol.go

@@ -191,7 +191,15 @@ func replyWithReset(s *segment) {
 		flags |= header.TCPFlagAck
 		ack = s.sequenceNumber.Add(s.logicalLen())
 	}
-	sendTCP(&s.route, s.id, buffer.VectorisedView{}, s.route.DefaultTTL(), stack.DefaultTOS, flags, seq, ack, 0 /* rcvWnd */, nil /* options */, nil /* gso */)
+	sendTCP(&s.route, tcpFields{
+		id:     s.id,
+		ttl:    s.route.DefaultTTL(),
+		tos:    stack.DefaultTOS,
+		flags:  flags,
+		seq:    seq,
+		ack:    ack,
+		rcvWnd: 0,
+	}, buffer.VectorisedView{}, nil /* gso */)
 }
 
 // SetOption implements stack.TransportProtocol.SetOption.