netstack/tcp: software segmentation offload

Right now, we send each tcp packet separately, we call one system
call per-packet. This patch allows to generate multiple tcp packets
and send them by sendmmsg.

The arguable part of this CL is a way how to handle multiple headers.
This CL adds the next field to the Prepandable buffer.

Nginx test results:

Server Software:        nginx/1.15.9
Server Hostname:        10.138.0.2
Server Port:            8080

Document Path:          /10m.txt
Document Length:        10485760 bytes

w/o gso:
Concurrency Level:      5
Time taken for tests:   5.491 seconds
Complete requests:      100
Failed requests:        0
Total transferred:      1048600200 bytes
HTML transferred:       1048576000 bytes
Requests per second:    18.21 [#/sec] (mean)
Time per request:       274.525 [ms] (mean)
Time per request:       54.905 [ms] (mean, across all concurrent requests)
Transfer rate:          186508.03 [Kbytes/sec] received

sw-gso:

Concurrency Level:      5
Time taken for tests:   3.852 seconds
Complete requests:      100
Failed requests:        0
Total transferred:      1048600200 bytes
HTML transferred:       1048576000 bytes
Requests per second:    25.96 [#/sec] (mean)
Time per request:       192.576 [ms] (mean)
Time per request:       38.515 [ms] (mean, across all concurrent requests)
Transfer rate:          265874.92 [Kbytes/sec] received

w/o gso:
$ ./tcp_benchmark --client --duration 15  --ideal
[SUM]  0.0-15.1 sec  2.20 GBytes  1.25 Gbits/sec

software gso:
$ tcp_benchmark --client --duration 15  --ideal --gso $((1<<16)) --swgso
[SUM]  0.0-15.1 sec  3.99 GBytes  2.26 Gbits/sec

PiperOrigin-RevId: 276112677

25 files changed, 608 insertions, 58 deletions

diff --git a/pkg/tcpip/buffer/prependable.go b/pkg/tcpip/buffer/prependable.go
index 4287464f3..48a2a2713 100644
--- a/pkg/tcpip/buffer/prependable.go
+++ b/pkg/tcpip/buffer/prependable.go
@@ -41,6 +41,11 @@ func NewPrependableFromView(v View) Prependable {
 	return Prependable{buf: v, usedIdx: 0}
 }
 
+// NewEmptyPrependableFromView creates a new prependable buffer from a View.
+func NewEmptyPrependableFromView(v View) Prependable {
+	return Prependable{buf: v, usedIdx: len(v)}
+}
+
 // View returns a View of the backing buffer that contains all prepended
 // data so far.
 func (p Prependable) View() View {
diff --git a/pkg/tcpip/header/BUILD b/pkg/tcpip/header/BUILD
index de3c7e2c3..7b3189303 100644
--- a/pkg/tcpip/header/BUILD
+++ b/pkg/tcpip/header/BUILD
@@ -36,10 +36,14 @@ go_test(
     name = "header_x_test",
     size = "small",
     srcs = [
+        "checksum_test.go",
         "ipversion_test.go",
         "tcp_test.go",
     ],
-    deps = [":header"],
+    deps = [
+        ":header",
+        "//pkg/tcpip/buffer",
+    ],
 )
 
 go_test(
diff --git a/pkg/tcpip/header/checksum.go b/pkg/tcpip/header/checksum.go
index 39a4d69be..9749c7f4d 100644
--- a/pkg/tcpip/header/checksum.go
+++ b/pkg/tcpip/header/checksum.go
@@ -23,11 +23,17 @@ import (
 	"gvisor.dev/gvisor/pkg/tcpip/buffer"
 )
 
-func calculateChecksum(buf []byte, initial uint32) uint16 {
+func calculateChecksum(buf []byte, odd bool, initial uint32) (uint16, bool) {
 	v := initial
 
+	if odd {
+		v += uint32(buf[0])
+		buf = buf[1:]
+	}
+
 	l := len(buf)
-	if l&1 != 0 {
+	odd = l&1 != 0
+	if odd {
 		l--
 		v += uint32(buf[l]) << 8
 	}
@@ -36,7 +42,7 @@ func calculateChecksum(buf []byte, initial uint32) uint16 {
 		v += (uint32(buf[i]) << 8) + uint32(buf[i+1])
 	}
 
-	return ChecksumCombine(uint16(v), uint16(v>>16))
+	return ChecksumCombine(uint16(v), uint16(v>>16)), odd
 }
 
 // Checksum calculates the checksum (as defined in RFC 1071) of the bytes in the
@@ -44,7 +50,8 @@ func calculateChecksum(buf []byte, initial uint32) uint16 {
 //
 // The initial checksum must have been computed on an even number of bytes.
 func Checksum(buf []byte, initial uint16) uint16 {
-	return calculateChecksum(buf, uint32(initial))
+	s, _ := calculateChecksum(buf, false, uint32(initial))
+	return s
 }
 
 // ChecksumVV calculates the checksum (as defined in RFC 1071) of the bytes in
@@ -52,19 +59,40 @@ func Checksum(buf []byte, initial uint16) uint16 {
 //
 // The initial checksum must have been computed on an even number of bytes.
 func ChecksumVV(vv buffer.VectorisedView, initial uint16) uint16 {
-	var odd bool
+	return ChecksumVVWithOffset(vv, initial, 0, vv.Size())
+}
+
+// ChecksumVVWithOffset calculates the checksum (as defined in RFC 1071) of the
+// bytes in the given VectorizedView.
+//
+// The initial checksum must have been computed on an even number of bytes.
+func ChecksumVVWithOffset(vv buffer.VectorisedView, initial uint16, off int, size int) uint16 {
+	odd := false
 	sum := initial
 	for _, v := range vv.Views() {
 		if len(v) == 0 {
 			continue
 		}
-		s := uint32(sum)
-		if odd {
-			s += uint32(v[0])
-			v = v[1:]
+
+		if off >= len(v) {
+			off -= len(v)
+			continue
+		}
+		v = v[off:]
+
+		l := len(v)
+		if l > size {
+			l = size
+		}
+		v = v[:l]
+
+		sum, odd = calculateChecksum(v, odd, uint32(sum))
+
+		size -= len(v)
+		if size == 0 {
+			break
 		}
-		odd = len(v)&1 != 0
-		sum = calculateChecksum(v, s)
+		off = 0
 	}
 	return sum
 }
diff --git a/pkg/tcpip/header/checksum_test.go b/pkg/tcpip/header/checksum_test.go
new file mode 100644
index 000000000..86b466c1c
--- /dev/null
+++ b/pkg/tcpip/header/checksum_test.go
@@ -0,0 +1,109 @@
+// Copyright 2019 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 header provides the implementation of the encoding and decoding of
+// network protocol headers.
+package header_test
+
+import (
+	"testing"
+
+	"gvisor.dev/gvisor/pkg/tcpip/buffer"
+	"gvisor.dev/gvisor/pkg/tcpip/header"
+)
+
+func TestChecksumVVWithOffset(t *testing.T) {
+	testCases := []struct {
+		name      string
+		vv        buffer.VectorisedView
+		off, size int
+		initial   uint16
+		want      uint16
+	}{
+		{
+			name: "empty",
+			vv: buffer.NewVectorisedView(0, []buffer.View{
+				buffer.NewViewFromBytes([]byte{1, 9, 0, 5, 4}),
+			}),
+			off:  0,
+			size: 0,
+			want: 0,
+		},
+		{
+			name: "OneView",
+			vv: buffer.NewVectorisedView(0, []buffer.View{
+				buffer.NewViewFromBytes([]byte{1, 9, 0, 5, 4}),
+			}),
+			off:  0,
+			size: 5,
+			want: 1294,
+		},
+		{
+			name: "TwoViews",
+			vv: buffer.NewVectorisedView(0, []buffer.View{
+				buffer.NewViewFromBytes([]byte{1, 9, 0, 5, 4}),
+				buffer.NewViewFromBytes([]byte{4, 3, 7, 1, 2, 123}),
+			}),
+			off:  0,
+			size: 11,
+			want: 33819,
+		},
+		{
+			name: "TwoViewsWithOffset",
+			vv: buffer.NewVectorisedView(0, []buffer.View{
+				buffer.NewViewFromBytes([]byte{98, 1, 9, 0, 5, 4}),
+				buffer.NewViewFromBytes([]byte{4, 3, 7, 1, 2, 123}),
+			}),
+			off:  1,
+			size: 11,
+			want: 33819,
+		},
+		{
+			name: "ThreeViewsWithOffset",
+			vv: buffer.NewVectorisedView(0, []buffer.View{
+				buffer.NewViewFromBytes([]byte{98, 1, 9, 0, 5, 4}),
+				buffer.NewViewFromBytes([]byte{98, 1, 9, 0, 5, 4}),
+				buffer.NewViewFromBytes([]byte{4, 3, 7, 1, 2, 123}),
+			}),
+			off:  7,
+			size: 11,
+			want: 33819,
+		},
+		{
+			name: "ThreeViewsWithInitial",
+			vv: buffer.NewVectorisedView(0, []buffer.View{
+				buffer.NewViewFromBytes([]byte{77, 11, 33, 0, 55, 44}),
+				buffer.NewViewFromBytes([]byte{98, 1, 9, 0, 5, 4}),
+				buffer.NewViewFromBytes([]byte{4, 3, 7, 1, 2, 123, 99}),
+			}),
+			initial: 77,
+			off:     7,
+			size:    11,
+			want:    33896,
+		},
+	}
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			if got, want := header.ChecksumVVWithOffset(tc.vv, tc.initial, tc.off, tc.size), tc.want; got != want {
+				t.Errorf("header.ChecksumVVWithOffset(%v) = %v, want: %v", tc, got, tc.want)
+			}
+			v := tc.vv.ToView()
+			v.TrimFront(tc.off)
+			v.CapLength(tc.size)
+			if got, want := header.Checksum(v, tc.initial), tc.want; got != want {
+				t.Errorf("header.Checksum(%v) = %v, want: %v", tc, got, tc.want)
+			}
+		})
+	}
+}
diff --git a/pkg/tcpip/link/channel/channel.go b/pkg/tcpip/link/channel/channel.go
index 95c6f0828..14f197a77 100644
--- a/pkg/tcpip/link/channel/channel.go
+++ b/pkg/tcpip/link/channel/channel.go
@@ -96,7 +96,7 @@ func (e *Endpoint) MTU() uint32 {
 func (e *Endpoint) Capabilities() stack.LinkEndpointCapabilities {
 	caps := stack.LinkEndpointCapabilities(0)
 	if e.GSO {
-		caps |= stack.CapabilityGSO
+		caps |= stack.CapabilityHardwareGSO
 	}
 	return caps
 }
@@ -134,6 +134,33 @@ func (e *Endpoint) WritePacket(_ *stack.Route, gso *stack.GSO, hdr buffer.Prepen
 	return nil
 }
 
+// WritePackets stores outbound packets into the channel.
+func (e *Endpoint) WritePackets(_ *stack.Route, gso *stack.GSO, hdrs []stack.PacketDescriptor, payload buffer.VectorisedView, protocol tcpip.NetworkProtocolNumber) (int, *tcpip.Error) {
+	payloadView := payload.ToView()
+	n := 0
+packetLoop:
+	for i := range hdrs {
+		hdr := &hdrs[i].Hdr
+		off := hdrs[i].Off
+		size := hdrs[i].Size
+		p := PacketInfo{
+			Header:  hdr.View(),
+			Proto:   protocol,
+			Payload: buffer.NewViewFromBytes(payloadView[off : off+size]),
+			GSO:     gso,
+		}
+
+		select {
+		case e.C <- p:
+			n++
+		default:
+			break packetLoop
+		}
+	}
+
+	return n, nil
+}
+
 // WriteRawPacket implements stack.LinkEndpoint.WriteRawPacket.
 func (e *Endpoint) WriteRawPacket(packet buffer.VectorisedView) *tcpip.Error {
 	p := PacketInfo{
diff --git a/pkg/tcpip/link/fdbased/endpoint.go b/pkg/tcpip/link/fdbased/endpoint.go
index 255220d1a..ae4858529 100644
--- a/pkg/tcpip/link/fdbased/endpoint.go
+++ b/pkg/tcpip/link/fdbased/endpoint.go
@@ -165,6 +165,9 @@ type Options struct {
 	// disabled.
 	GSOMaxSize uint32
 
+	// SoftwareGSOEnabled indicates whether software GSO is enabled or not.
+	SoftwareGSOEnabled bool
+
 	// PacketDispatchMode specifies the type of inbound dispatcher to be
 	// used for this endpoint.
 	PacketDispatchMode PacketDispatchMode
@@ -242,7 +245,11 @@ func New(opts *Options) (stack.LinkEndpoint, error) {
 		}
 		if isSocket {
 			if opts.GSOMaxSize != 0 {
-				e.caps |= stack.CapabilityGSO
+				if opts.SoftwareGSOEnabled {
+					e.caps |= stack.CapabilitySoftwareGSO
+				} else {
+					e.caps |= stack.CapabilityHardwareGSO
+				}
 				e.gsoMaxSize = opts.GSOMaxSize
 			}
 		}
@@ -397,7 +404,7 @@ func (e *endpoint) WritePacket(r *stack.Route, gso *stack.GSO, hdr buffer.Prepen
 		eth.Encode(ethHdr)
 	}
 
-	if e.Capabilities()&stack.CapabilityGSO != 0 {
+	if e.Capabilities()&stack.CapabilityHardwareGSO != 0 {
 		vnetHdr := virtioNetHdr{}
 		vnetHdrBuf := vnetHdrToByteSlice(&vnetHdr)
 		if gso != nil {
@@ -430,6 +437,123 @@ func (e *endpoint) WritePacket(r *stack.Route, gso *stack.GSO, hdr buffer.Prepen
 	return rawfile.NonBlockingWrite3(e.fds[0], hdr.View(), payload.ToView(), nil)
 }
 
+// WritePackets writes outbound packets to the file descriptor. If it is not
+// currently writable, the packet is dropped.
+func (e *endpoint) WritePackets(r *stack.Route, gso *stack.GSO, hdrs []stack.PacketDescriptor, payload buffer.VectorisedView, protocol tcpip.NetworkProtocolNumber) (int, *tcpip.Error) {
+	var ethHdrBuf []byte
+	// hdr + data
+	iovLen := 2
+	if e.hdrSize > 0 {
+		// Add ethernet header if needed.
+		ethHdrBuf = make([]byte, header.EthernetMinimumSize)
+		eth := header.Ethernet(ethHdrBuf)
+		ethHdr := &header.EthernetFields{
+			DstAddr: r.RemoteLinkAddress,
+			Type:    protocol,
+		}
+
+		// Preserve the src address if it's set in the route.
+		if r.LocalLinkAddress != "" {
+			ethHdr.SrcAddr = r.LocalLinkAddress
+		} else {
+			ethHdr.SrcAddr = e.addr
+		}
+		eth.Encode(ethHdr)
+		iovLen++
+	}
+
+	n := len(hdrs)
+
+	views := payload.Views()
+	/*
+	 * Each bondary in views can add one more iovec.
+	 *
+	 * payload |      |          |         |
+	 *         -----------------------------
+	 * packets |    |    |    |    |    |  |
+	 *         -----------------------------
+	 * iovecs  |    | |  |    |  | |    |  |
+	 */
+	iovec := make([]syscall.Iovec, n*iovLen+len(views)-1)
+	mmsgHdrs := make([]rawfile.MMsgHdr, n)
+
+	iovecIdx := 0
+	viewIdx := 0
+	viewOff := 0
+	off := 0
+	nextOff := 0
+	for i := range hdrs {
+		prevIovecIdx := iovecIdx
+		mmsgHdr := &mmsgHdrs[i]
+		mmsgHdr.Msg.Iov = &iovec[iovecIdx]
+		packetSize := hdrs[i].Size
+		hdr := &hdrs[i].Hdr
+
+		off = hdrs[i].Off
+		if off != nextOff {
+			// We stop in a different point last time.
+			size := packetSize
+			viewIdx = 0
+			viewOff = 0
+			for size > 0 {
+				if size >= len(views[viewIdx]) {
+					viewIdx++
+					viewOff = 0
+					size -= len(views[viewIdx])
+				} else {
+					viewOff = size
+					size = 0
+				}
+			}
+		}
+		nextOff = off + packetSize
+
+		if ethHdrBuf != nil {
+			v := &iovec[iovecIdx]
+			v.Base = &ethHdrBuf[0]
+			v.Len = uint64(len(ethHdrBuf))
+			iovecIdx++
+		}
+
+		v := &iovec[iovecIdx]
+		hdrView := hdr.View()
+		v.Base = &hdrView[0]
+		v.Len = uint64(len(hdrView))
+		iovecIdx++
+
+		for packetSize > 0 {
+			vec := &iovec[iovecIdx]
+			iovecIdx++
+
+			v := views[viewIdx]
+			vec.Base = &v[viewOff]
+			s := len(v) - viewOff
+			if s <= packetSize {
+				viewIdx++
+				viewOff = 0
+			} else {
+				s = packetSize
+				viewOff += s
+			}
+			vec.Len = uint64(s)
+			packetSize -= s
+		}
+
+		mmsgHdr.Msg.Iovlen = uint64(iovecIdx - prevIovecIdx)
+	}
+
+	packets := 0
+	for packets < n {
+		sent, err := rawfile.NonBlockingSendMMsg(e.fds[0], mmsgHdrs)
+		if err != nil {
+			return packets, err
+		}
+		packets += sent
+		mmsgHdrs = mmsgHdrs[sent:]
+	}
+	return packets, nil
+}
+
 // WriteRawPacket implements stack.LinkEndpoint.WriteRawPacket.
 func (e *endpoint) WriteRawPacket(packet buffer.VectorisedView) *tcpip.Error {
 	return rawfile.NonBlockingWrite(e.fds[0], packet.ToView())
diff --git a/pkg/tcpip/link/fdbased/packet_dispatchers.go b/pkg/tcpip/link/fdbased/packet_dispatchers.go
index 9fb760891..12168a1dc 100644
--- a/pkg/tcpip/link/fdbased/packet_dispatchers.go
+++ b/pkg/tcpip/link/fdbased/packet_dispatchers.go
@@ -53,7 +53,7 @@ func newReadVDispatcher(fd int, e *endpoint) (linkDispatcher, error) {
 	d := &readVDispatcher{fd: fd, e: e}
 	d.views = make([]buffer.View, len(BufConfig))
 	iovLen := len(BufConfig)
-	if d.e.Capabilities()&stack.CapabilityGSO != 0 {
+	if d.e.Capabilities()&stack.CapabilityHardwareGSO != 0 {
 		iovLen++
 	}
 	d.iovecs = make([]syscall.Iovec, iovLen)
@@ -63,7 +63,7 @@ func newReadVDispatcher(fd int, e *endpoint) (linkDispatcher, error) {
 func (d *readVDispatcher) allocateViews(bufConfig []int) {
 	var vnetHdr [virtioNetHdrSize]byte
 	vnetHdrOff := 0
-	if d.e.Capabilities()&stack.CapabilityGSO != 0 {
+	if d.e.Capabilities()&stack.CapabilityHardwareGSO != 0 {
 		// The kernel adds virtioNetHdr before each packet, but
 		// we don't use it, so so we allocate a buffer for it,
 		// add it in iovecs but don't add it in a view.
@@ -106,7 +106,7 @@ func (d *readVDispatcher) dispatch() (bool, *tcpip.Error) {
 	if err != nil {
 		return false, err
 	}
-	if d.e.Capabilities()&stack.CapabilityGSO != 0 {
+	if d.e.Capabilities()&stack.CapabilityHardwareGSO != 0 {
 		// Skip virtioNetHdr which is added before each packet, it
 		// isn't used and it isn't in a view.
 		n -= virtioNetHdrSize
@@ -195,7 +195,7 @@ func newRecvMMsgDispatcher(fd int, e *endpoint) (linkDispatcher, error) {
 	}
 	d.iovecs = make([][]syscall.Iovec, MaxMsgsPerRecv)
 	iovLen := len(BufConfig)
-	if d.e.Capabilities()&stack.CapabilityGSO != 0 {
+	if d.e.Capabilities()&stack.CapabilityHardwareGSO != 0 {
 		// virtioNetHdr is prepended before each packet.
 		iovLen++
 	}
@@ -226,7 +226,7 @@ func (d *recvMMsgDispatcher) allocateViews(bufConfig []int) {
 	for k := 0; k < len(d.views); k++ {
 		var vnetHdr [virtioNetHdrSize]byte
 		vnetHdrOff := 0
-		if d.e.Capabilities()&stack.CapabilityGSO != 0 {
+		if d.e.Capabilities()&stack.CapabilityHardwareGSO != 0 {
 			// The kernel adds virtioNetHdr before each packet, but
 			// we don't use it, so so we allocate a buffer for it,
 			// add it in iovecs but don't add it in a view.
@@ -262,7 +262,7 @@ func (d *recvMMsgDispatcher) dispatch() (bool, *tcpip.Error) {
 	// Process each of received packets.
 	for k := 0; k < nMsgs; k++ {
 		n := int(d.msgHdrs[k].Len)
-		if d.e.Capabilities()&stack.CapabilityGSO != 0 {
+		if d.e.Capabilities()&stack.CapabilityHardwareGSO != 0 {
 			n -= virtioNetHdrSize
 		}
 		if n <= d.e.hdrSize {
diff --git a/pkg/tcpip/link/loopback/loopback.go b/pkg/tcpip/link/loopback/loopback.go
index 5f5bacf27..a3b48fa73 100644
--- a/pkg/tcpip/link/loopback/loopback.go
+++ b/pkg/tcpip/link/loopback/loopback.go
@@ -90,6 +90,11 @@ func (e *endpoint) WritePacket(_ *stack.Route, _ *stack.GSO, hdr buffer.Prependa
 	return nil
 }
 
+// WritePackets implements stack.LinkEndpoint.WritePackets.
+func (e *endpoint) WritePackets(_ *stack.Route, _ *stack.GSO, hdrs []stack.PacketDescriptor, payload buffer.VectorisedView, protocol tcpip.NetworkProtocolNumber) (int, *tcpip.Error) {
+	panic("not implemented")
+}
+
 // WriteRawPacket implements stack.LinkEndpoint.WriteRawPacket.
 func (e *endpoint) WriteRawPacket(packet buffer.VectorisedView) *tcpip.Error {
 	// Reject the packet if it's shorter than an ethernet header.
diff --git a/pkg/tcpip/link/muxed/injectable.go b/pkg/tcpip/link/muxed/injectable.go
index 41524e271..682b60291 100644
--- a/pkg/tcpip/link/muxed/injectable.go
+++ b/pkg/tcpip/link/muxed/injectable.go
@@ -84,12 +84,23 @@ func (m *InjectableEndpoint) InjectInbound(protocol tcpip.NetworkProtocolNumber,
 	m.dispatcher.DeliverNetworkPacket(m, "" /* remote */, "" /* local */, protocol, vv, nil /* linkHeader */)
 }
 
+// WritePackets writes outbound packets to the appropriate
+// LinkInjectableEndpoint based on the RemoteAddress. HandleLocal only works if
+// r.RemoteAddress has a route registered in this endpoint.
+func (m *InjectableEndpoint) WritePackets(r *stack.Route, gso *stack.GSO, hdrs []stack.PacketDescriptor, payload buffer.VectorisedView, protocol tcpip.NetworkProtocolNumber) (int, *tcpip.Error) {
+	endpoint, ok := m.routes[r.RemoteAddress]
+	if !ok {
+		return 0, tcpip.ErrNoRoute
+	}
+	return endpoint.WritePackets(r, gso, hdrs, payload, protocol)
+}
+
 // WritePacket writes outbound packets to the appropriate LinkInjectableEndpoint
 // based on the RemoteAddress. HandleLocal only works if r.RemoteAddress has a
 // route registered in this endpoint.
-func (m *InjectableEndpoint) WritePacket(r *stack.Route, _ *stack.GSO, hdr buffer.Prependable, payload buffer.VectorisedView, protocol tcpip.NetworkProtocolNumber) *tcpip.Error {
+func (m *InjectableEndpoint) WritePacket(r *stack.Route, gso *stack.GSO, hdr buffer.Prependable, payload buffer.VectorisedView, protocol tcpip.NetworkProtocolNumber) *tcpip.Error {
 	if endpoint, ok := m.routes[r.RemoteAddress]; ok {
-		return endpoint.WritePacket(r, nil /* gso */, hdr, payload, protocol)
+		return endpoint.WritePacket(r, gso, hdr, payload, protocol)
 	}
 	return tcpip.ErrNoRoute
 }
diff --git a/pkg/tcpip/link/rawfile/BUILD b/pkg/tcpip/link/rawfile/BUILD
index 2e8bc772a..05c7b8024 100644
--- a/pkg/tcpip/link/rawfile/BUILD
+++ b/pkg/tcpip/link/rawfile/BUILD
@@ -16,5 +16,8 @@ go_library(
     visibility = [
         "//visibility:public",
     ],
-    deps = ["//pkg/tcpip"],
+    deps = [
+        "//pkg/tcpip",
+        "@org_golang_x_sys//unix:go_default_library",
+    ],
 )
diff --git a/pkg/tcpip/link/rawfile/rawfile_unsafe.go b/pkg/tcpip/link/rawfile/rawfile_unsafe.go
index 7e286a3a6..44e25d475 100644
--- a/pkg/tcpip/link/rawfile/rawfile_unsafe.go
+++ b/pkg/tcpip/link/rawfile/rawfile_unsafe.go
@@ -22,6 +22,7 @@ import (
 	"syscall"
 	"unsafe"
 
+	"golang.org/x/sys/unix"
 	"gvisor.dev/gvisor/pkg/tcpip"
 )
 
@@ -101,6 +102,16 @@ func NonBlockingWrite3(fd int, b1, b2, b3 []byte) *tcpip.Error {
 	return nil
 }
 
+// NonBlockingSendMMsg sends multiple messages on a socket.
+func NonBlockingSendMMsg(fd int, msgHdrs []MMsgHdr) (int, *tcpip.Error) {
+	n, _, e := syscall.RawSyscall6(unix.SYS_SENDMMSG, uintptr(fd), uintptr(unsafe.Pointer(&msgHdrs[0])), uintptr(len(msgHdrs)), syscall.MSG_DONTWAIT, 0, 0)
+	if e != 0 {
+		return 0, TranslateErrno(e)
+	}
+
+	return int(n), nil
+}
+
 // PollEvent represents the pollfd structure passed to a poll() system call.
 type PollEvent struct {
 	FD      int32
diff --git a/pkg/tcpip/link/sharedmem/sharedmem.go b/pkg/tcpip/link/sharedmem/sharedmem.go
index 4393403c5..279e2b457 100644
--- a/pkg/tcpip/link/sharedmem/sharedmem.go
+++ b/pkg/tcpip/link/sharedmem/sharedmem.go
@@ -212,6 +212,11 @@ func (e *endpoint) WritePacket(r *stack.Route, _ *stack.GSO, hdr buffer.Prependa
 	return nil
 }
 
+// WritePackets implements stack.LinkEndpoint.WritePackets.
+func (e *endpoint) WritePackets(r *stack.Route, _ *stack.GSO, hdrs []stack.PacketDescriptor, payload buffer.VectorisedView, protocol tcpip.NetworkProtocolNumber) (int, *tcpip.Error) {
+	panic("not implemented")
+}
+
 // WriteRawPacket implements stack.LinkEndpoint.WriteRawPacket.
 func (e *endpoint) WriteRawPacket(packet buffer.VectorisedView) *tcpip.Error {
 	v := packet.ToView()
diff --git a/pkg/tcpip/link/sniffer/sniffer.go b/pkg/tcpip/link/sniffer/sniffer.go
index de9ff98cb..39757ea2a 100644
--- a/pkg/tcpip/link/sniffer/sniffer.go
+++ b/pkg/tcpip/link/sniffer/sniffer.go
@@ -193,10 +193,7 @@ func (e *endpoint) GSOMaxSize() uint32 {
 	return 0
 }
 
-// WritePacket implements the stack.LinkEndpoint interface. It is called by
-// higher-level protocols to write packets; it just logs the packet and forwards
-// the request to the lower endpoint.
-func (e *endpoint) WritePacket(r *stack.Route, gso *stack.GSO, hdr buffer.Prependable, payload buffer.VectorisedView, protocol tcpip.NetworkProtocolNumber) *tcpip.Error {
+func (e *endpoint) dumpPacket(gso *stack.GSO, hdr buffer.Prependable, payload buffer.VectorisedView, protocol tcpip.NetworkProtocolNumber) {
 	if atomic.LoadUint32(&LogPackets) == 1 && e.file == nil {
 		logPacket("send", protocol, hdr.View(), gso)
 	}
@@ -223,9 +220,27 @@ func (e *endpoint) WritePacket(r *stack.Route, gso *stack.GSO, hdr buffer.Prepen
 			panic(err)
 		}
 	}
+}
+
+// WritePacket implements the stack.LinkEndpoint interface. It is called by
+// higher-level protocols to write packets; it just logs the packet and
+// forwards the request to the lower endpoint.
+func (e *endpoint) WritePacket(r *stack.Route, gso *stack.GSO, hdr buffer.Prependable, payload buffer.VectorisedView, protocol tcpip.NetworkProtocolNumber) *tcpip.Error {
+	e.dumpPacket(gso, hdr, payload, protocol)
 	return e.lower.WritePacket(r, gso, hdr, payload, protocol)
 }
 
+// WritePackets implements the stack.LinkEndpoint interface. It is called by
+// higher-level protocols to write packets; it just logs the packet and
+// forwards the request to the lower endpoint.
+func (e *endpoint) WritePackets(r *stack.Route, gso *stack.GSO, hdrs []stack.PacketDescriptor, payload buffer.VectorisedView, protocol tcpip.NetworkProtocolNumber) (int, *tcpip.Error) {
+	view := payload.ToView()
+	for _, d := range hdrs {
+		e.dumpPacket(gso, d.Hdr, buffer.NewVectorisedView(d.Size, []buffer.View{view[d.Off:][:d.Size]}), protocol)
+	}
+	return e.lower.WritePackets(r, gso, hdrs, payload, protocol)
+}
+
 // WriteRawPacket implements stack.LinkEndpoint.WriteRawPacket.
 func (e *endpoint) WriteRawPacket(packet buffer.VectorisedView) *tcpip.Error {
 	if atomic.LoadUint32(&LogPackets) == 1 && e.file == nil {
diff --git a/pkg/tcpip/link/waitable/waitable.go b/pkg/tcpip/link/waitable/waitable.go
index e9b50b802..a04fc1062 100644
--- a/pkg/tcpip/link/waitable/waitable.go
+++ b/pkg/tcpip/link/waitable/waitable.go
@@ -109,6 +109,19 @@ func (e *Endpoint) WritePacket(r *stack.Route, gso *stack.GSO, hdr buffer.Prepen
 	return err
 }
 
+// WritePackets implements stack.LinkEndpoint.WritePackets. It is called by
+// higher-level protocols to write packets. It only forwards packets to the
+// lower endpoint if Wait or WaitWrite haven't been called.
+func (e *Endpoint) WritePackets(r *stack.Route, gso *stack.GSO, hdrs []stack.PacketDescriptor, payload buffer.VectorisedView, protocol tcpip.NetworkProtocolNumber) (int, *tcpip.Error) {
+	if !e.writeGate.Enter() {
+		return len(hdrs), nil
+	}
+
+	n, err := e.lower.WritePackets(r, gso, hdrs, payload, protocol)
+	e.writeGate.Leave()
+	return n, err
+}
+
 // WriteRawPacket implements stack.LinkEndpoint.WriteRawPacket.
 func (e *Endpoint) WriteRawPacket(packet buffer.VectorisedView) *tcpip.Error {
 	if !e.writeGate.Enter() {
diff --git a/pkg/tcpip/link/waitable/waitable_test.go b/pkg/tcpip/link/waitable/waitable_test.go
index f03bbceec..5f0f8fa2d 100644
--- a/pkg/tcpip/link/waitable/waitable_test.go
+++ b/pkg/tcpip/link/waitable/waitable_test.go
@@ -70,6 +70,12 @@ func (e *countedEndpoint) WritePacket(r *stack.Route, _ *stack.GSO, hdr buffer.P
 	return nil
 }
 
+// WritePackets implements stack.LinkEndpoint.WritePackets.
+func (e *countedEndpoint) WritePackets(r *stack.Route, _ *stack.GSO, hdrs []stack.PacketDescriptor, payload buffer.VectorisedView, protocol tcpip.NetworkProtocolNumber) (int, *tcpip.Error) {
+	e.writeCount += len(hdrs)
+	return len(hdrs), nil
+}
+
 func (e *countedEndpoint) WriteRawPacket(packet buffer.VectorisedView) *tcpip.Error {
 	e.writeCount++
 	return nil
diff --git a/pkg/tcpip/network/arp/arp.go b/pkg/tcpip/network/arp/arp.go
index 6b1e854dc..46178459e 100644
--- a/pkg/tcpip/network/arp/arp.go
+++ b/pkg/tcpip/network/arp/arp.go
@@ -83,6 +83,11 @@ func (e *endpoint) WritePacket(*stack.Route, *stack.GSO, buffer.Prependable, buf
 	return tcpip.ErrNotSupported
 }
 
+// WritePackets implements stack.NetworkEndpoint.WritePackets.
+func (e *endpoint) WritePackets(*stack.Route, *stack.GSO, []stack.PacketDescriptor, buffer.VectorisedView, stack.NetworkHeaderParams, stack.PacketLooping) (int, *tcpip.Error) {
+	return 0, tcpip.ErrNotSupported
+}
+
 func (e *endpoint) WriteHeaderIncludedPacket(r *stack.Route, payload buffer.VectorisedView, loop stack.PacketLooping) *tcpip.Error {
 	return tcpip.ErrNotSupported
 }
diff --git a/pkg/tcpip/network/ip_test.go b/pkg/tcpip/network/ip_test.go
index bd09b9a1c..8d74497ba 100644
--- a/pkg/tcpip/network/ip_test.go
+++ b/pkg/tcpip/network/ip_test.go
@@ -171,6 +171,11 @@ func (t *testObject) WritePacket(_ *stack.Route, _ *stack.GSO, hdr buffer.Prepen
 	return nil
 }
 
+// WritePackets implements stack.LinkEndpoint.WritePackets.
+func (t *testObject) WritePackets(_ *stack.Route, _ *stack.GSO, hdr []stack.PacketDescriptor, payload buffer.VectorisedView, protocol tcpip.NetworkProtocolNumber) (int, *tcpip.Error) {
+	panic("not implemented")
+}
+
 func (t *testObject) WriteRawPacket(_ buffer.VectorisedView) *tcpip.Error {
 	return tcpip.ErrNotSupported
 }
diff --git a/pkg/tcpip/network/ipv4/ipv4.go b/pkg/tcpip/network/ipv4/ipv4.go
index 5cd895ff0..1339f8474 100644
--- a/pkg/tcpip/network/ipv4/ipv4.go
+++ b/pkg/tcpip/network/ipv4/ipv4.go
@@ -198,10 +198,9 @@ func (e *endpoint) writePacketFragments(r *stack.Route, gso *stack.GSO, hdr buff
 	return nil
 }
 
-// WritePacket writes a packet to the given destination address and protocol.
-func (e *endpoint) WritePacket(r *stack.Route, gso *stack.GSO, hdr buffer.Prependable, payload buffer.VectorisedView, params stack.NetworkHeaderParams, loop stack.PacketLooping) *tcpip.Error {
+func (e *endpoint) addIPHeader(r *stack.Route, hdr *buffer.Prependable, payloadSize int, params stack.NetworkHeaderParams) {
 	ip := header.IPv4(hdr.Prepend(header.IPv4MinimumSize))
-	length := uint16(hdr.UsedLength() + payload.Size())
+	length := uint16(hdr.UsedLength() + payloadSize)
 	id := uint32(0)
 	if length > header.IPv4MaximumHeaderSize+8 {
 		// Packets of 68 bytes or less are required by RFC 791 to not be
@@ -219,6 +218,11 @@ func (e *endpoint) WritePacket(r *stack.Route, gso *stack.GSO, hdr buffer.Prepen
 		DstAddr:     r.RemoteAddress,
 	})
 	ip.SetChecksum(^ip.CalculateChecksum())
+}
+
+// WritePacket writes a packet to the given destination address and protocol.
+func (e *endpoint) WritePacket(r *stack.Route, gso *stack.GSO, hdr buffer.Prependable, payload buffer.VectorisedView, params stack.NetworkHeaderParams, loop stack.PacketLooping) *tcpip.Error {
+	e.addIPHeader(r, &hdr, payload.Size(), params)
 
 	if loop&stack.PacketLoop != 0 {
 		views := make([]buffer.View, 1, 1+len(payload.Views()))
@@ -242,6 +246,23 @@ func (e *endpoint) WritePacket(r *stack.Route, gso *stack.GSO, hdr buffer.Prepen
 	return nil
 }
 
+// WritePackets implements stack.NetworkEndpoint.WritePackets.
+func (e *endpoint) WritePackets(r *stack.Route, gso *stack.GSO, hdrs []stack.PacketDescriptor, payload buffer.VectorisedView, params stack.NetworkHeaderParams, loop stack.PacketLooping) (int, *tcpip.Error) {
+	if loop&stack.PacketLoop != 0 {
+		panic("multiple packets in local loop")
+	}
+	if loop&stack.PacketOut == 0 {
+		return len(hdrs), nil
+	}
+
+	for i := range hdrs {
+		e.addIPHeader(r, &hdrs[i].Hdr, hdrs[i].Size, params)
+	}
+	n, err := e.linkEP.WritePackets(r, gso, hdrs, payload, ProtocolNumber)
+	r.Stats().IP.PacketsSent.IncrementBy(uint64(n))
+	return n, err
+}
+
 // WriteHeaderIncludedPacket writes a packet already containing a network
 // header through the given route.
 func (e *endpoint) WriteHeaderIncludedPacket(r *stack.Route, payload buffer.VectorisedView, loop stack.PacketLooping) *tcpip.Error {
diff --git a/pkg/tcpip/network/ipv6/ipv6.go b/pkg/tcpip/network/ipv6/ipv6.go
index cd1e34085..5898f8f9e 100644
--- a/pkg/tcpip/network/ipv6/ipv6.go
+++ b/pkg/tcpip/network/ipv6/ipv6.go
@@ -97,9 +97,8 @@ func (e *endpoint) GSOMaxSize() uint32 {
 	return 0
 }
 
-// WritePacket writes a packet to the given destination address and protocol.
-func (e *endpoint) WritePacket(r *stack.Route, gso *stack.GSO, hdr buffer.Prependable, payload buffer.VectorisedView, params stack.NetworkHeaderParams, loop stack.PacketLooping) *tcpip.Error {
-	length := uint16(hdr.UsedLength() + payload.Size())
+func (e *endpoint) addIPHeader(r *stack.Route, hdr *buffer.Prependable, payloadSize int, params stack.NetworkHeaderParams) {
+	length := uint16(hdr.UsedLength() + payloadSize)
 	ip := header.IPv6(hdr.Prepend(header.IPv6MinimumSize))
 	ip.Encode(&header.IPv6Fields{
 		PayloadLength: length,
@@ -109,6 +108,11 @@ func (e *endpoint) WritePacket(r *stack.Route, gso *stack.GSO, hdr buffer.Prepen
 		SrcAddr:       r.LocalAddress,
 		DstAddr:       r.RemoteAddress,
 	})
+}
+
+// WritePacket writes a packet to the given destination address and protocol.
+func (e *endpoint) WritePacket(r *stack.Route, gso *stack.GSO, hdr buffer.Prependable, payload buffer.VectorisedView, params stack.NetworkHeaderParams, loop stack.PacketLooping) *tcpip.Error {
+	e.addIPHeader(r, &hdr, payload.Size(), params)
 
 	if loop&stack.PacketLoop != 0 {
 		views := make([]buffer.View, 1, 1+len(payload.Views()))
@@ -127,6 +131,26 @@ func (e *endpoint) WritePacket(r *stack.Route, gso *stack.GSO, hdr buffer.Prepen
 	return e.linkEP.WritePacket(r, gso, hdr, payload, ProtocolNumber)
 }
 
+// WritePackets implements stack.LinkEndpoint.WritePackets.
+func (e *endpoint) WritePackets(r *stack.Route, gso *stack.GSO, hdrs []stack.PacketDescriptor, payload buffer.VectorisedView, params stack.NetworkHeaderParams, loop stack.PacketLooping) (int, *tcpip.Error) {
+	if loop&stack.PacketLoop != 0 {
+		panic("not implemented")
+	}
+	if loop&stack.PacketOut == 0 {
+		return len(hdrs), nil
+	}
+
+	for i := range hdrs {
+		hdr := &hdrs[i].Hdr
+		size := hdrs[i].Size
+		e.addIPHeader(r, hdr, size, params)
+	}
+
+	n, err := e.linkEP.WritePackets(r, gso, hdrs, payload, ProtocolNumber)
+	r.Stats().IP.PacketsSent.IncrementBy(uint64(n))
+	return n, err
+}
+
 // WriteHeaderIncludedPacker implements stack.NetworkEndpoint. It is not yet
 // supported by IPv6.
 func (*endpoint) WriteHeaderIncludedPacket(r *stack.Route, payload buffer.VectorisedView, loop stack.PacketLooping) *tcpip.Error {
diff --git a/pkg/tcpip/stack/registration.go b/pkg/tcpip/stack/registration.go
index f655d011f..0869fb084 100644
--- a/pkg/tcpip/stack/registration.go
+++ b/pkg/tcpip/stack/registration.go
@@ -201,6 +201,10 @@ type NetworkEndpoint interface {
 	// protocol.
 	WritePacket(r *Route, gso *GSO, hdr buffer.Prependable, payload buffer.VectorisedView, params NetworkHeaderParams, loop PacketLooping) *tcpip.Error
 
+	// WritePackets writes packets to the given destination address and
+	// protocol.
+	WritePackets(r *Route, gso *GSO, hdrs []PacketDescriptor, payload buffer.VectorisedView, params NetworkHeaderParams, loop PacketLooping) (int, *tcpip.Error)
+
 	// WriteHeaderIncludedPacket writes a packet that includes a network
 	// header to the given destination address.
 	WriteHeaderIncludedPacket(r *Route, payload buffer.VectorisedView, loop PacketLooping) *tcpip.Error
@@ -283,7 +287,11 @@ const (
 	CapabilitySaveRestore
 	CapabilityDisconnectOk
 	CapabilityLoopback
-	CapabilityGSO
+	CapabilityHardwareGSO
+
+	// CapabilitySoftwareGSO indicates the link endpoint supports of sending
+	// multiple packets using a single call (LinkEndpoint.WritePackets).
+	CapabilitySoftwareGSO
 )
 
 // LinkEndpoint is the interface implemented by data link layer protocols (e.g.,
@@ -318,6 +326,14 @@ type LinkEndpoint interface {
 	// r.LocalLinkAddress if it is provided.
 	WritePacket(r *Route, gso *GSO, hdr buffer.Prependable, payload buffer.VectorisedView, protocol tcpip.NetworkProtocolNumber) *tcpip.Error
 
+	// WritePackets writes packets with the given protocol through the
+	// given route.
+	//
+	// Right now, WritePackets is used only when the software segmentation
+	// offload is enabled. If it will be used for something else, it may
+	// require to change syscall filters.
+	WritePackets(r *Route, gso *GSO, hdrs []PacketDescriptor, payload buffer.VectorisedView, protocol tcpip.NetworkProtocolNumber) (int, *tcpip.Error)
+
 	// WriteRawPacket writes a packet directly to the link. The packet
 	// should already have an ethernet header.
 	WriteRawPacket(packet buffer.VectorisedView) *tcpip.Error
@@ -421,8 +437,14 @@ type GSOType int
 // Types of gso segments.
 const (
 	GSONone GSOType = iota
+
+	// Hardware GSO types:
 	GSOTCPv4
 	GSOTCPv6
+
+	// GSOSW is used for software GSO segments which have to be sent by
+	// endpoint.WritePackets.
+	GSOSW
 )
 
 // GSO contains generic segmentation offload properties.
@@ -450,3 +472,7 @@ type GSOEndpoint interface {
 	// GSOMaxSize returns the maximum GSO packet size.
 	GSOMaxSize() uint32
 }
+
+// SoftwareGSOMaxSize is a maximum allowed size of a software GSO segment.
+// This isn't a hard limit, because it is never set into packet headers.
+const SoftwareGSOMaxSize = (1 << 16)
diff --git a/pkg/tcpip/stack/route.go b/pkg/tcpip/stack/route.go
index e72373964..1a0a51b57 100644
--- a/pkg/tcpip/stack/route.go
+++ b/pkg/tcpip/stack/route.go
@@ -47,8 +47,8 @@ type Route struct {
 	// starts.
 	ref *referencedNetworkEndpoint
 
-	// loop controls where WritePacket should send packets.
-	loop PacketLooping
+	// Loop controls where WritePacket should send packets.
+	Loop PacketLooping
 }
 
 // makeRoute initializes a new route. It takes ownership of the provided
@@ -69,7 +69,7 @@ func makeRoute(netProto tcpip.NetworkProtocolNumber, localAddr, remoteAddr tcpip
 		LocalLinkAddress: localLinkAddr,
 		RemoteAddress:    remoteAddr,
 		ref:              ref,
-		loop:             loop,
+		Loop:             loop,
 	}
 }
 
@@ -159,7 +159,7 @@ func (r *Route) WritePacket(gso *GSO, hdr buffer.Prependable, payload buffer.Vec
 		return tcpip.ErrInvalidEndpointState
 	}
 
-	err := r.ref.ep.WritePacket(r, gso, hdr, payload, params, r.loop)
+	err := r.ref.ep.WritePacket(r, gso, hdr, payload, params, r.Loop)
 	if err != nil {
 		r.Stats().IP.OutgoingPacketErrors.Increment()
 	} else {
@@ -169,6 +169,44 @@ func (r *Route) WritePacket(gso *GSO, hdr buffer.Prependable, payload buffer.Vec
 	return err
 }
 
+// PacketDescriptor is a packet descriptor which contains a packet header and
+// offset and size of packet data in a payload view.
+type PacketDescriptor struct {
+	Hdr  buffer.Prependable
+	Off  int
+	Size int
+}
+
+// NewPacketDescriptors allocates a set of packet descriptors.
+func NewPacketDescriptors(n int, hdrSize int) []PacketDescriptor {
+	buf := make([]byte, n*hdrSize)
+	hdrs := make([]PacketDescriptor, n)
+	for i := range hdrs {
+		hdrs[i].Hdr = buffer.NewEmptyPrependableFromView(buf[i*hdrSize:][:hdrSize])
+	}
+	return hdrs
+}
+
+// WritePackets writes the set of packets through the given route.
+func (r *Route) WritePackets(gso *GSO, hdrs []PacketDescriptor, payload buffer.VectorisedView, params NetworkHeaderParams) (int, *tcpip.Error) {
+	if !r.ref.isValidForOutgoing() {
+		return 0, tcpip.ErrInvalidEndpointState
+	}
+
+	n, err := r.ref.ep.WritePackets(r, gso, hdrs, payload, params, r.Loop)
+	if err != nil {
+		r.Stats().IP.OutgoingPacketErrors.IncrementBy(uint64(len(hdrs) - n))
+	}
+	r.ref.nic.stats.Tx.Packets.IncrementBy(uint64(n))
+	payloadSize := 0
+	for i := 0; i < n; i++ {
+		r.ref.nic.stats.Tx.Bytes.IncrementBy(uint64(hdrs[i].Hdr.UsedLength()))
+		payloadSize += hdrs[i].Size
+	}
+	r.ref.nic.stats.Tx.Bytes.IncrementBy(uint64(payloadSize))
+	return n, err
+}
+
 // WriteHeaderIncludedPacket writes a packet already containing a network
 // header through the given route.
 func (r *Route) WriteHeaderIncludedPacket(payload buffer.VectorisedView) *tcpip.Error {
@@ -176,7 +214,7 @@ func (r *Route) WriteHeaderIncludedPacket(payload buffer.VectorisedView) *tcpip.
 		return tcpip.ErrInvalidEndpointState
 	}
 
-	if err := r.ref.ep.WriteHeaderIncludedPacket(r, payload, r.loop); err != nil {
+	if err := r.ref.ep.WriteHeaderIncludedPacket(r, payload, r.Loop); err != nil {
 		r.Stats().IP.OutgoingPacketErrors.Increment()
 		return err
 	}
diff --git a/pkg/tcpip/stack/stack_test.go b/pkg/tcpip/stack/stack_test.go
index 0a184fabe..72eee9334 100644
--- a/pkg/tcpip/stack/stack_test.go
+++ b/pkg/tcpip/stack/stack_test.go
@@ -146,6 +146,11 @@ func (f *fakeNetworkEndpoint) WritePacket(r *stack.Route, gso *stack.GSO, hdr bu
 	return f.ep.WritePacket(r, gso, hdr, payload, fakeNetNumber)
 }
 
+// WritePackets implements stack.LinkEndpoint.WritePackets.
+func (f *fakeNetworkEndpoint) WritePackets(r *stack.Route, gso *stack.GSO, hdrs []stack.PacketDescriptor, payload buffer.VectorisedView, params stack.NetworkHeaderParams, loop stack.PacketLooping) (int, *tcpip.Error) {
+	panic("not implemented")
+}
+
 func (*fakeNetworkEndpoint) WriteHeaderIncludedPacket(r *stack.Route, payload buffer.VectorisedView, loop stack.PacketLooping) *tcpip.Error {
 	return tcpip.ErrNotSupported
 }
diff --git a/pkg/tcpip/transport/tcp/BUILD b/pkg/tcpip/transport/tcp/BUILD
index aed70e06f..f1dbc6f91 100644
--- a/pkg/tcpip/transport/tcp/BUILD
+++ b/pkg/tcpip/transport/tcp/BUILD
@@ -44,6 +44,7 @@ go_library(
     imports = ["gvisor.dev/gvisor/pkg/tcpip/buffer"],
     visibility = ["//visibility:public"],
     deps = [
+        "//pkg/log",
         "//pkg/rand",
         "//pkg/sleep",
         "//pkg/tcpip",
diff --git a/pkg/tcpip/transport/tcp/connect.go b/pkg/tcpip/transport/tcp/connect.go
index 5ea036bea..082135374 100644
--- a/pkg/tcpip/transport/tcp/connect.go
+++ b/pkg/tcpip/transport/tcp/connect.go
@@ -607,17 +607,11 @@ func (e *endpoint) sendTCP(r *stack.Route, id stack.TransportEndpointID, data bu
 	return nil
 }
 
-// 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 {
+func buildTCPHdr(r *stack.Route, id stack.TransportEndpointID, d *stack.PacketDescriptor, data buffer.VectorisedView, flags byte, seq, ack seqnum.Value, rcvWnd seqnum.Size, opts []byte, gso *stack.GSO) {
 	optLen := len(opts)
-	// Allocate a buffer for the TCP header.
-	hdr := buffer.NewPrependable(header.TCPMinimumSize + int(r.MaxHeaderLength()) + optLen)
-
-	if rcvWnd > 0xffff {
-		rcvWnd = 0xffff
-	}
-
+	hdr := &d.Hdr
+	packetSize := d.Size
+	off := d.Off
 	// Initialize the header.
 	tcp := header.TCP(hdr.Prepend(header.TCPMinimumSize + optLen))
 	tcp.Encode(&header.TCPFields{
@@ -631,7 +625,7 @@ func sendTCP(r *stack.Route, id stack.TransportEndpointID, data buffer.Vectorise
 	})
 	copy(tcp[header.TCPMinimumSize:], opts)
 
-	length := uint16(hdr.UsedLength() + data.Size())
+	length := uint16(hdr.UsedLength() + packetSize)
 	xsum := r.PseudoHeaderChecksum(ProtocolNumber, length)
 	// Only calculate the checksum if offloading isn't supported.
 	if gso != nil && gso.NeedsCsum {
@@ -641,14 +635,71 @@ func sendTCP(r *stack.Route, id stack.TransportEndpointID, data buffer.Vectorise
 		// header and data and get the right sum of the TCP packet.
 		tcp.SetChecksum(xsum)
 	} else if r.Capabilities()&stack.CapabilityTXChecksumOffload == 0 {
-		xsum = header.ChecksumVV(data, xsum)
+		xsum = header.ChecksumVVWithOffset(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
+	}
+
+	mss := int(gso.MSS)
+	n := (data.Size() + mss - 1) / mss
+
+	hdrs := stack.NewPacketDescriptors(n, header.TCPMinimumSize+int(r.MaxHeaderLength())+optLen)
+
+	size := data.Size()
+	off := 0
+	for i := 0; i < n; i++ {
+		packetSize := mss
+		if packetSize > size {
+			packetSize = size
+		}
+		size -= packetSize
+		hdrs[i].Off = off
+		hdrs[i].Size = packetSize
+		buildTCPHdr(r, id, &hdrs[i], data, flags, seq, ack, rcvWnd, opts, gso)
+		off += packetSize
+		seq = seq.Add(seqnum.Size(packetSize))
+	}
+	if ttl == 0 {
+		ttl = r.DefaultTTL()
+	}
+	sent, err := r.WritePackets(gso, hdrs, data, stack.NetworkHeaderParams{Protocol: ProtocolNumber, TTL: ttl, TOS: tos})
+	if err != nil {
+		r.Stats().TCP.SegmentSendErrors.IncrementBy(uint64(n - sent))
+	}
+	r.Stats().TCP.SegmentsSent.IncrementBy(uint64(sent))
+	return err
+}
+
+// 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
+	}
+
+	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)
+	}
+
+	d := &stack.PacketDescriptor{
+		Hdr:  buffer.NewPrependable(header.TCPMinimumSize + int(r.MaxHeaderLength()) + optLen),
+		Off:  0,
+		Size: data.Size(),
+	}
+	buildTCPHdr(r, id, d, data, flags, seq, ack, rcvWnd, opts, gso)
+
 	if ttl == 0 {
 		ttl = r.DefaultTTL()
 	}
-	if err := r.WritePacket(gso, hdr, data, stack.NetworkHeaderParams{Protocol: ProtocolNumber, TTL: ttl, TOS: tos}); err != nil {
+	if err := r.WritePacket(gso, d.Hdr, data, stack.NetworkHeaderParams{Protocol: ProtocolNumber, TTL: ttl, TOS: tos}); err != nil {
 		r.Stats().TCP.SegmentSendErrors.Increment()
 		return err
 	}
diff --git a/pkg/tcpip/transport/tcp/endpoint.go b/pkg/tcpip/transport/tcp/endpoint.go
index a1b784b49..8b9cb4c33 100644
--- a/pkg/tcpip/transport/tcp/endpoint.go
+++ b/pkg/tcpip/transport/tcp/endpoint.go
@@ -2327,11 +2327,7 @@ func (e *endpoint) completeState() stack.TCPEndpointState {
 	return s
 }
 
-func (e *endpoint) initGSO() {
-	if e.route.Capabilities()&stack.CapabilityGSO == 0 {
-		return
-	}
-
+func (e *endpoint) initHardwareGSO() {
 	gso := &stack.GSO{}
 	switch e.route.NetProto {
 	case header.IPv4ProtocolNumber:
@@ -2349,6 +2345,18 @@ func (e *endpoint) initGSO() {
 	e.gso = gso
 }
 
+func (e *endpoint) initGSO() {
+	if e.route.Capabilities()&stack.CapabilityHardwareGSO != 0 {
+		e.initHardwareGSO()
+	} else if e.route.Capabilities()&stack.CapabilitySoftwareGSO != 0 {
+		e.gso = &stack.GSO{
+			MaxSize:   e.route.GSOMaxSize(),
+			Type:      stack.GSOSW,
+			NeedsCsum: false,
+		}
+	}
+}
+
 // State implements tcpip.Endpoint.State. It exports the endpoint's protocol
 // state for diagnostics.
 func (e *endpoint) State() uint32 {