conn, device: use UDP GSO and GRO on Linux

StdNetBind probes for UDP GSO and GRO support at runtime. UDP GSO is
dependent on checksum offload support on the egress netdev. UDP GSO
will be disabled in the event sendmmsg() returns EIO, which is a strong
signal that the egress netdev does not support checksum offload.

The iperf3 results below demonstrate the effect of this commit between
two Linux computers with i5-12400 CPUs. There is roughly ~13us of round
trip latency between them.

The first result is from commit 052af4a without UDP GSO or GRO.

Starting Test: protocol: TCP, 1 streams, 131072 byte blocks
[ ID] Interval           Transfer     Bitrate         Retr  Cwnd
[  5]   0.00-10.00  sec  9.85 GBytes  8.46 Gbits/sec  1139   3.01 MBytes
- - - - - - - - - - - - - - - - - - - - - - - - -
Test Complete. Summary Results:
[ ID] Interval           Transfer     Bitrate         Retr
[  5]   0.00-10.00  sec  9.85 GBytes  8.46 Gbits/sec  1139  sender
[  5]   0.00-10.04  sec  9.85 GBytes  8.42 Gbits/sec        receiver

The second result is with UDP GSO and GRO.

Starting Test: protocol: TCP, 1 streams, 131072 byte blocks
[ ID] Interval           Transfer     Bitrate         Retr  Cwnd
[  5]   0.00-10.00  sec  12.3 GBytes  10.6 Gbits/sec  232   3.15 MBytes
- - - - - - - - - - - - - - - - - - - - - - - - -
Test Complete. Summary Results:
[ ID] Interval           Transfer     Bitrate         Retr
[  5]   0.00-10.00  sec  12.3 GBytes  10.6 Gbits/sec  232   sender
[  5]   0.00-10.04  sec  12.3 GBytes  10.6 Gbits/sec        receiver

Reviewed-by: Adrian Dewhurst <adrian@tailscale.com>
Signed-off-by: Jordan Whited <jordan@tailscale.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>

13 files changed, 673 insertions, 147 deletions

diff --git a/conn/bind_std.go b/conn/bind_std.go
index c701ef8..9886c91 100644
--- a/conn/bind_std.go
+++ b/conn/bind_std.go
@@ -8,6 +8,7 @@ package conn
 import (
 	"context"
 	"errors"
+	"fmt"
 	"net"
 	"net/netip"
 	"runtime"
@@ -29,16 +30,19 @@ var (
 // methods for sending and receiving multiple datagrams per-syscall. See the
 // proposal in https://github.com/golang/go/issues/45886#issuecomment-1218301564.
 type StdNetBind struct {
-	mu     sync.Mutex // protects all fields except as specified
-	ipv4   *net.UDPConn
-	ipv6   *net.UDPConn
-	ipv4PC *ipv4.PacketConn // will be nil on non-Linux
-	ipv6PC *ipv6.PacketConn // will be nil on non-Linux
-
-	// these three fields are not guarded by mu
-	udpAddrPool  sync.Pool
-	ipv4MsgsPool sync.Pool
-	ipv6MsgsPool sync.Pool
+	mu            sync.Mutex // protects all fields except as specified
+	ipv4          *net.UDPConn
+	ipv6          *net.UDPConn
+	ipv4PC        *ipv4.PacketConn // will be nil on non-Linux
+	ipv6PC        *ipv6.PacketConn // will be nil on non-Linux
+	ipv4TxOffload bool
+	ipv4RxOffload bool
+	ipv6TxOffload bool
+	ipv6RxOffload bool
+
+	// these two fields are not guarded by mu
+	udpAddrPool sync.Pool
+	msgsPool    sync.Pool
 
 	blackhole4 bool
 	blackhole6 bool
@@ -54,23 +58,14 @@ func NewStdNetBind() Bind {
 			},
 		},
 
-		ipv4MsgsPool: sync.Pool{
-			New: func() any {
-				msgs := make([]ipv4.Message, IdealBatchSize)
-				for i := range msgs {
-					msgs[i].Buffers = make(net.Buffers, 1)
-					msgs[i].OOB = make([]byte, srcControlSize)
-				}
-				return &msgs
-			},
-		},
-
-		ipv6MsgsPool: sync.Pool{
+		msgsPool: sync.Pool{
 			New: func() any {
+				// ipv6.Message and ipv4.Message are interchangeable as they are
+				// both aliases for x/net/internal/socket.Message.
 				msgs := make([]ipv6.Message, IdealBatchSize)
 				for i := range msgs {
 					msgs[i].Buffers = make(net.Buffers, 1)
-					msgs[i].OOB = make([]byte, srcControlSize)
+					msgs[i].OOB = make([]byte, controlSize)
 				}
 				return &msgs
 			},
@@ -113,7 +108,7 @@ func (e *StdNetEndpoint) DstIP() netip.Addr {
 	return e.AddrPort.Addr()
 }
 
-// See sticky_default,linux, etc for implementations of SrcIP and SrcIfidx.
+// See control_default,linux, etc for implementations of SrcIP and SrcIfidx.
 
 func (e *StdNetEndpoint) DstToBytes() []byte {
 	b, _ := e.AddrPort.MarshalBinary()
@@ -179,19 +174,21 @@ again:
 	}
 	var fns []ReceiveFunc
 	if v4conn != nil {
+		s.ipv4TxOffload, s.ipv4RxOffload = supportsUDPOffload(v4conn)
 		if runtime.GOOS == "linux" {
 			v4pc = ipv4.NewPacketConn(v4conn)
 			s.ipv4PC = v4pc
 		}
-		fns = append(fns, s.makeReceiveIPv4(v4pc, v4conn))
+		fns = append(fns, s.makeReceiveIPv4(v4pc, v4conn, s.ipv4RxOffload))
 		s.ipv4 = v4conn
 	}
 	if v6conn != nil {
+		s.ipv6TxOffload, s.ipv6RxOffload = supportsUDPOffload(v6conn)
 		if runtime.GOOS == "linux" {
 			v6pc = ipv6.NewPacketConn(v6conn)
 			s.ipv6PC = v6pc
 		}
-		fns = append(fns, s.makeReceiveIPv6(v6pc, v6conn))
+		fns = append(fns, s.makeReceiveIPv6(v6pc, v6conn, s.ipv6RxOffload))
 		s.ipv6 = v6conn
 	}
 	if len(fns) == 0 {
@@ -201,69 +198,93 @@ again:
 	return fns, uint16(port), nil
 }
 
-func (s *StdNetBind) makeReceiveIPv4(pc *ipv4.PacketConn, conn *net.UDPConn) ReceiveFunc {
-	return func(bufs [][]byte, sizes []int, eps []Endpoint) (n int, err error) {
-		msgs := s.ipv4MsgsPool.Get().(*[]ipv4.Message)
-		defer s.ipv4MsgsPool.Put(msgs)
-		for i := range bufs {
-			(*msgs)[i].Buffers[0] = bufs[i]
-		}
-		var numMsgs int
-		if runtime.GOOS == "linux" {
-			numMsgs, err = pc.ReadBatch(*msgs, 0)
+func (s *StdNetBind) putMessages(msgs *[]ipv6.Message) {
+	for i := range *msgs {
+		(*msgs)[i] = ipv6.Message{Buffers: (*msgs)[i].Buffers, OOB: (*msgs)[i].OOB}
+	}
+	s.msgsPool.Put(msgs)
+}
+
+func (s *StdNetBind) getMessages() *[]ipv6.Message {
+	return s.msgsPool.Get().(*[]ipv6.Message)
+}
+
+var (
+	// If compilation fails here these are no longer the same underlying type.
+	_ ipv6.Message = ipv4.Message{}
+)
+
+type batchReader interface {
+	ReadBatch([]ipv6.Message, int) (int, error)
+}
+
+type batchWriter interface {
+	WriteBatch([]ipv6.Message, int) (int, error)
+}
+
+func (s *StdNetBind) receiveIP(
+	br batchReader,
+	conn *net.UDPConn,
+	rxOffload bool,
+	bufs [][]byte,
+	sizes []int,
+	eps []Endpoint,
+) (n int, err error) {
+	msgs := s.getMessages()
+	for i := range bufs {
+		(*msgs)[i].Buffers[0] = bufs[i]
+		(*msgs)[i].OOB = (*msgs)[i].OOB[:cap((*msgs)[i].OOB)]
+	}
+	defer s.putMessages(msgs)
+	var numMsgs int
+	if runtime.GOOS == "linux" {
+		if rxOffload {
+			readAt := len(*msgs) - (IdealBatchSize / udpSegmentMaxDatagrams)
+			numMsgs, err = br.ReadBatch((*msgs)[readAt:], 0)
+			if err != nil {
+				return 0, err
+			}
+			numMsgs, err = splitCoalescedMessages(*msgs, readAt, getGSOSize)
 			if err != nil {
 				return 0, err
 			}
 		} else {
-			msg := &(*msgs)[0]
-			msg.N, msg.NN, _, msg.Addr, err = conn.ReadMsgUDP(msg.Buffers[0], msg.OOB)
+			numMsgs, err = br.ReadBatch(*msgs, 0)
 			if err != nil {
 				return 0, err
 			}
-			numMsgs = 1
 		}
-		for i := 0; i < numMsgs; i++ {
-			msg := &(*msgs)[i]
-			sizes[i] = msg.N
-			addrPort := msg.Addr.(*net.UDPAddr).AddrPort()
-			ep := &StdNetEndpoint{AddrPort: addrPort} // TODO: remove allocation
-			getSrcFromControl(msg.OOB[:msg.NN], ep)
-			eps[i] = ep
+	} else {
+		msg := &(*msgs)[0]
+		msg.N, msg.NN, _, msg.Addr, err = conn.ReadMsgUDP(msg.Buffers[0], msg.OOB)
+		if err != nil {
+			return 0, err
+		}
+		numMsgs = 1
+	}
+	for i := 0; i < numMsgs; i++ {
+		msg := &(*msgs)[i]
+		sizes[i] = msg.N
+		if sizes[i] == 0 {
+			continue
 		}
-		return numMsgs, nil
+		addrPort := msg.Addr.(*net.UDPAddr).AddrPort()
+		ep := &StdNetEndpoint{AddrPort: addrPort} // TODO: remove allocation
+		getSrcFromControl(msg.OOB[:msg.NN], ep)
+		eps[i] = ep
 	}
+	return numMsgs, nil
 }
 
-func (s *StdNetBind) makeReceiveIPv6(pc *ipv6.PacketConn, conn *net.UDPConn) ReceiveFunc {
+func (s *StdNetBind) makeReceiveIPv4(pc *ipv4.PacketConn, conn *net.UDPConn, rxOffload bool) ReceiveFunc {
 	return func(bufs [][]byte, sizes []int, eps []Endpoint) (n int, err error) {
-		msgs := s.ipv6MsgsPool.Get().(*[]ipv6.Message)
-		defer s.ipv6MsgsPool.Put(msgs)
-		for i := range bufs {
-			(*msgs)[i].Buffers[0] = bufs[i]
-		}
-		var numMsgs int
-		if runtime.GOOS == "linux" {
-			numMsgs, err = pc.ReadBatch(*msgs, 0)
-			if err != nil {
-				return 0, err
-			}
-		} else {
-			msg := &(*msgs)[0]
-			msg.N, msg.NN, _, msg.Addr, err = conn.ReadMsgUDP(msg.Buffers[0], msg.OOB)
-			if err != nil {
-				return 0, err
-			}
-			numMsgs = 1
-		}
-		for i := 0; i < numMsgs; i++ {
-			msg := &(*msgs)[i]
-			sizes[i] = msg.N
-			addrPort := msg.Addr.(*net.UDPAddr).AddrPort()
-			ep := &StdNetEndpoint{AddrPort: addrPort} // TODO: remove allocation
-			getSrcFromControl(msg.OOB[:msg.NN], ep)
-			eps[i] = ep
-		}
-		return numMsgs, nil
+		return s.receiveIP(pc, conn, rxOffload, bufs, sizes, eps)
+	}
+}
+
+func (s *StdNetBind) makeReceiveIPv6(pc *ipv6.PacketConn, conn *net.UDPConn, rxOffload bool) ReceiveFunc {
+	return func(bufs [][]byte, sizes []int, eps []Endpoint) (n int, err error) {
+		return s.receiveIP(pc, conn, rxOffload, bufs, sizes, eps)
 	}
 }
 
@@ -293,28 +314,42 @@ func (s *StdNetBind) Close() error {
 	}
 	s.blackhole4 = false
 	s.blackhole6 = false
+	s.ipv4TxOffload = false
+	s.ipv4RxOffload = false
+	s.ipv6TxOffload = false
+	s.ipv6RxOffload = false
 	if err1 != nil {
 		return err1
 	}
 	return err2
 }
 
+type ErrUDPGSODisabled struct {
+	onLaddr  string
+	RetryErr error
+}
+
+func (e ErrUDPGSODisabled) Error() string {
+	return fmt.Sprintf("disabled UDP GSO on %s, NIC(s) may not support checksum offload", e.onLaddr)
+}
+
+func (e ErrUDPGSODisabled) Unwrap() error {
+	return e.RetryErr
+}
+
 func (s *StdNetBind) Send(bufs [][]byte, endpoint Endpoint) error {
 	s.mu.Lock()
 	blackhole := s.blackhole4
 	conn := s.ipv4
-	var (
-		pc4 *ipv4.PacketConn
-		pc6 *ipv6.PacketConn
-	)
+	offload := s.ipv4TxOffload
+	br := batchWriter(s.ipv4PC)
 	is6 := false
 	if endpoint.DstIP().Is6() {
 		blackhole = s.blackhole6
 		conn = s.ipv6
-		pc6 = s.ipv6PC
+		br = s.ipv6PC
 		is6 = true
-	} else {
-		pc4 = s.ipv4PC
+		offload = s.ipv6TxOffload
 	}
 	s.mu.Unlock()
 
@@ -324,25 +359,56 @@ func (s *StdNetBind) Send(bufs [][]byte, endpoint Endpoint) error {
 	if conn == nil {
 		return syscall.EAFNOSUPPORT
 	}
+
+	msgs := s.getMessages()
+	defer s.putMessages(msgs)
+	ua := s.udpAddrPool.Get().(*net.UDPAddr)
+	defer s.udpAddrPool.Put(ua)
 	if is6 {
-		return s.send6(conn, pc6, endpoint, bufs)
+		as16 := endpoint.DstIP().As16()
+		copy(ua.IP, as16[:])
+		ua.IP = ua.IP[:16]
 	} else {
-		return s.send4(conn, pc4, endpoint, bufs)
+		as4 := endpoint.DstIP().As4()
+		copy(ua.IP, as4[:])
+		ua.IP = ua.IP[:4]
 	}
+	ua.Port = int(endpoint.(*StdNetEndpoint).Port())
+	var (
+		retried bool
+		err     error
+	)
+retry:
+	if offload {
+		n := coalesceMessages(ua, endpoint.(*StdNetEndpoint), bufs, *msgs, setGSOSize)
+		err = s.send(conn, br, (*msgs)[:n])
+		if err != nil && offload && errShouldDisableUDPGSO(err) {
+			offload = false
+			s.mu.Lock()
+			if is6 {
+				s.ipv6TxOffload = false
+			} else {
+				s.ipv4TxOffload = false
+			}
+			s.mu.Unlock()
+			retried = true
+			goto retry
+		}
+	} else {
+		for i := range bufs {
+			(*msgs)[i].Addr = ua
+			(*msgs)[i].Buffers[0] = bufs[i]
+			setSrcControl(&(*msgs)[i].OOB, endpoint.(*StdNetEndpoint))
+		}
+		err = s.send(conn, br, (*msgs)[:len(bufs)])
+	}
+	if retried {
+		return ErrUDPGSODisabled{onLaddr: conn.LocalAddr().String(), RetryErr: err}
+	}
+	return err
 }
 
-func (s *StdNetBind) send4(conn *net.UDPConn, pc *ipv4.PacketConn, ep Endpoint, bufs [][]byte) error {
-	ua := s.udpAddrPool.Get().(*net.UDPAddr)
-	as4 := ep.DstIP().As4()
-	copy(ua.IP, as4[:])
-	ua.IP = ua.IP[:4]
-	ua.Port = int(ep.(*StdNetEndpoint).Port())
-	msgs := s.ipv4MsgsPool.Get().(*[]ipv4.Message)
-	for i, buf := range bufs {
-		(*msgs)[i].Buffers[0] = buf
-		(*msgs)[i].Addr = ua
-		setSrcControl(&(*msgs)[i].OOB, ep.(*StdNetEndpoint))
-	}
+func (s *StdNetBind) send(conn *net.UDPConn, pc batchWriter, msgs []ipv6.Message) error {
 	var (
 		n     int
 		err   error
@@ -350,59 +416,128 @@ func (s *StdNetBind) send4(conn *net.UDPConn, pc *ipv4.PacketConn, ep Endpoint,
 	)
 	if runtime.GOOS == "linux" {
 		for {
-			n, err = pc.WriteBatch((*msgs)[start:len(bufs)], 0)
-			if err != nil || n == len((*msgs)[start:len(bufs)]) {
+			n, err = pc.WriteBatch(msgs[start:], 0)
+			if err != nil || n == len(msgs[start:]) {
 				break
 			}
 			start += n
 		}
 	} else {
-		for i, buf := range bufs {
-			_, _, err = conn.WriteMsgUDP(buf, (*msgs)[i].OOB, ua)
+		for _, msg := range msgs {
+			_, _, err = conn.WriteMsgUDP(msg.Buffers[0], msg.OOB, msg.Addr.(*net.UDPAddr))
 			if err != nil {
 				break
 			}
 		}
 	}
-	s.udpAddrPool.Put(ua)
-	s.ipv4MsgsPool.Put(msgs)
 	return err
 }
 
-func (s *StdNetBind) send6(conn *net.UDPConn, pc *ipv6.PacketConn, ep Endpoint, bufs [][]byte) error {
-	ua := s.udpAddrPool.Get().(*net.UDPAddr)
-	as16 := ep.DstIP().As16()
-	copy(ua.IP, as16[:])
-	ua.IP = ua.IP[:16]
-	ua.Port = int(ep.(*StdNetEndpoint).Port())
-	msgs := s.ipv6MsgsPool.Get().(*[]ipv6.Message)
-	for i, buf := range bufs {
-		(*msgs)[i].Buffers[0] = buf
-		(*msgs)[i].Addr = ua
-		setSrcControl(&(*msgs)[i].OOB, ep.(*StdNetEndpoint))
-	}
+const (
+	// Exceeding these values results in EMSGSIZE. They account for layer3 and
+	// layer4 headers. IPv6 does not need to account for itself as the payload
+	// length field is self excluding.
+	maxIPv4PayloadLen = 1<<16 - 1 - 20 - 8
+	maxIPv6PayloadLen = 1<<16 - 1 - 8
+
+	// This is a hard limit imposed by the kernel.
+	udpSegmentMaxDatagrams = 64
+)
+
+type setGSOFunc func(control *[]byte, gsoSize uint16)
+
+func coalesceMessages(addr *net.UDPAddr, ep *StdNetEndpoint, bufs [][]byte, msgs []ipv6.Message, setGSO setGSOFunc) int {
 	var (
-		n     int
-		err   error
-		start int
+		base     = -1 // index of msg we are currently coalescing into
+		gsoSize  int  // segmentation size of msgs[base]
+		dgramCnt int  // number of dgrams coalesced into msgs[base]
+		endBatch bool // tracking flag to start a new batch on next iteration of bufs
 	)
-	if runtime.GOOS == "linux" {
-		for {
-			n, err = pc.WriteBatch((*msgs)[start:len(bufs)], 0)
-			if err != nil || n == len((*msgs)[start:len(bufs)]) {
-				break
+	maxPayloadLen := maxIPv4PayloadLen
+	if ep.DstIP().Is6() {
+		maxPayloadLen = maxIPv6PayloadLen
+	}
+	for i, buf := range bufs {
+		if i > 0 {
+			msgLen := len(buf)
+			baseLenBefore := len(msgs[base].Buffers[0])
+			freeBaseCap := cap(msgs[base].Buffers[0]) - baseLenBefore
+			if msgLen+baseLenBefore <= maxPayloadLen &&
+				msgLen <= gsoSize &&
+				msgLen <= freeBaseCap &&
+				dgramCnt < udpSegmentMaxDatagrams &&
+				!endBatch {
+				msgs[base].Buffers[0] = append(msgs[base].Buffers[0], buf...)
+				if i == len(bufs)-1 {
+					setGSO(&msgs[base].OOB, uint16(gsoSize))
+				}
+				dgramCnt++
+				if msgLen < gsoSize {
+					// A smaller than gsoSize packet on the tail is legal, but
+					// it must end the batch.
+					endBatch = true
+				}
+				continue
 			}
-			start += n
 		}
-	} else {
-		for i, buf := range bufs {
-			_, _, err = conn.WriteMsgUDP(buf, (*msgs)[i].OOB, ua)
-			if err != nil {
-				break
+		if dgramCnt > 1 {
+			setGSO(&msgs[base].OOB, uint16(gsoSize))
+		}
+		// Reset prior to incrementing base since we are preparing to start a
+		// new potential batch.
+		endBatch = false
+		base++
+		gsoSize = len(buf)
+		setSrcControl(&msgs[base].OOB, ep)
+		msgs[base].Buffers[0] = buf
+		msgs[base].Addr = addr
+		dgramCnt = 1
+	}
+	return base + 1
+}
+
+type getGSOFunc func(control []byte) (int, error)
+
+func splitCoalescedMessages(msgs []ipv6.Message, firstMsgAt int, getGSO getGSOFunc) (n int, err error) {
+	for i := firstMsgAt; i < len(msgs); i++ {
+		msg := &msgs[i]
+		if msg.N == 0 {
+			return n, err
+		}
+		var (
+			gsoSize    int
+			start      int
+			end        = msg.N
+			numToSplit = 1
+		)
+		gsoSize, err = getGSO(msg.OOB[:msg.NN])
+		if err != nil {
+			return n, err
+		}
+		if gsoSize > 0 {
+			numToSplit = (msg.N + gsoSize - 1) / gsoSize
+			end = gsoSize
+		}
+		for j := 0; j < numToSplit; j++ {
+			if n > i {
+				return n, errors.New("splitting coalesced packet resulted in overflow")
 			}
+			copied := copy(msgs[n].Buffers[0], msg.Buffers[0][start:end])
+			msgs[n].N = copied
+			msgs[n].Addr = msg.Addr
+			start = end
+			end += gsoSize
+			if end > msg.N {
+				end = msg.N
+			}
+			n++
+		}
+		if i != n-1 {
+			// It is legal for bytes to move within msg.Buffers[0] as a result
+			// of splitting, so we only zero the source msg len when it is not
+			// the destination of the last split operation above.
+			msg.N = 0
 		}
 	}
-	s.udpAddrPool.Put(ua)
-	s.ipv6MsgsPool.Put(msgs)
-	return err
+	return n, nil
 }
diff --git a/conn/bind_std_test.go b/conn/bind_std_test.go
index 1e46776..34a3c9a 100644
--- a/conn/bind_std_test.go
+++ b/conn/bind_std_test.go
@@ -1,6 +1,12 @@
 package conn
 
-import "testing"
+import (
+	"encoding/binary"
+	"net"
+	"testing"
+
+	"golang.org/x/net/ipv6"
+)
 
 func TestStdNetBindReceiveFuncAfterClose(t *testing.T) {
 	bind := NewStdNetBind().(*StdNetBind)
@@ -20,3 +26,225 @@ func TestStdNetBindReceiveFuncAfterClose(t *testing.T) {
 		fn(bufs, sizes, eps)
 	}
 }
+
+func mockSetGSOSize(control *[]byte, gsoSize uint16) {
+	*control = (*control)[:cap(*control)]
+	binary.LittleEndian.PutUint16(*control, gsoSize)
+}
+
+func Test_coalesceMessages(t *testing.T) {
+	cases := []struct {
+		name     string
+		buffs    [][]byte
+		wantLens []int
+		wantGSO  []int
+	}{
+		{
+			name: "one message no coalesce",
+			buffs: [][]byte{
+				make([]byte, 1, 1),
+			},
+			wantLens: []int{1},
+			wantGSO:  []int{0},
+		},
+		{
+			name: "two messages equal len coalesce",
+			buffs: [][]byte{
+				make([]byte, 1, 2),
+				make([]byte, 1, 1),
+			},
+			wantLens: []int{2},
+			wantGSO:  []int{1},
+		},
+		{
+			name: "two messages unequal len coalesce",
+			buffs: [][]byte{
+				make([]byte, 2, 3),
+				make([]byte, 1, 1),
+			},
+			wantLens: []int{3},
+			wantGSO:  []int{2},
+		},
+		{
+			name: "three messages second unequal len coalesce",
+			buffs: [][]byte{
+				make([]byte, 2, 3),
+				make([]byte, 1, 1),
+				make([]byte, 2, 2),
+			},
+			wantLens: []int{3, 2},
+			wantGSO:  []int{2, 0},
+		},
+		{
+			name: "three messages limited cap coalesce",
+			buffs: [][]byte{
+				make([]byte, 2, 4),
+				make([]byte, 2, 2),
+				make([]byte, 2, 2),
+			},
+			wantLens: []int{4, 2},
+			wantGSO:  []int{2, 0},
+		},
+	}
+
+	for _, tt := range cases {
+		t.Run(tt.name, func(t *testing.T) {
+			addr := &net.UDPAddr{
+				IP:   net.ParseIP("127.0.0.1").To4(),
+				Port: 1,
+			}
+			msgs := make([]ipv6.Message, len(tt.buffs))
+			for i := range msgs {
+				msgs[i].Buffers = make([][]byte, 1)
+				msgs[i].OOB = make([]byte, 0, 2)
+			}
+			got := coalesceMessages(addr, &StdNetEndpoint{AddrPort: addr.AddrPort()}, tt.buffs, msgs, mockSetGSOSize)
+			if got != len(tt.wantLens) {
+				t.Fatalf("got len %d want: %d", got, len(tt.wantLens))
+			}
+			for i := 0; i < got; i++ {
+				if msgs[i].Addr != addr {
+					t.Errorf("msgs[%d].Addr != passed addr", i)
+				}
+				gotLen := len(msgs[i].Buffers[0])
+				if gotLen != tt.wantLens[i] {
+					t.Errorf("len(msgs[%d].Buffers[0]) %d != %d", i, gotLen, tt.wantLens[i])
+				}
+				gotGSO, err := mockGetGSOSize(msgs[i].OOB)
+				if err != nil {
+					t.Fatalf("msgs[%d] getGSOSize err: %v", i, err)
+				}
+				if gotGSO != tt.wantGSO[i] {
+					t.Errorf("msgs[%d] gsoSize %d != %d", i, gotGSO, tt.wantGSO[i])
+				}
+			}
+		})
+	}
+}
+
+func mockGetGSOSize(control []byte) (int, error) {
+	if len(control) < 2 {
+		return 0, nil
+	}
+	return int(binary.LittleEndian.Uint16(control)), nil
+}
+
+func Test_splitCoalescedMessages(t *testing.T) {
+	newMsg := func(n, gso int) ipv6.Message {
+		msg := ipv6.Message{
+			Buffers: [][]byte{make([]byte, 1<<16-1)},
+			N:       n,
+			OOB:     make([]byte, 2),
+		}
+		binary.LittleEndian.PutUint16(msg.OOB, uint16(gso))
+		if gso > 0 {
+			msg.NN = 2
+		}
+		return msg
+	}
+
+	cases := []struct {
+		name        string
+		msgs        []ipv6.Message
+		firstMsgAt  int
+		wantNumEval int
+		wantMsgLens []int
+		wantErr     bool
+	}{
+		{
+			name: "second last split last empty",
+			msgs: []ipv6.Message{
+				newMsg(0, 0),
+				newMsg(0, 0),
+				newMsg(3, 1),
+				newMsg(0, 0),
+			},
+			firstMsgAt:  2,
+			wantNumEval: 3,
+			wantMsgLens: []int{1, 1, 1, 0},
+			wantErr:     false,
+		},
+		{
+			name: "second last no split last empty",
+			msgs: []ipv6.Message{
+				newMsg(0, 0),
+				newMsg(0, 0),
+				newMsg(1, 0),
+				newMsg(0, 0),
+			},
+			firstMsgAt:  2,
+			wantNumEval: 1,
+			wantMsgLens: []int{1, 0, 0, 0},
+			wantErr:     false,
+		},
+		{
+			name: "second last no split last no split",
+			msgs: []ipv6.Message{
+				newMsg(0, 0),
+				newMsg(0, 0),
+				newMsg(1, 0),
+				newMsg(1, 0),
+			},
+			firstMsgAt:  2,
+			wantNumEval: 2,
+			wantMsgLens: []int{1, 1, 0, 0},
+			wantErr:     false,
+		},
+		{
+			name: "second last no split last split",
+			msgs: []ipv6.Message{
+				newMsg(0, 0),
+				newMsg(0, 0),
+				newMsg(1, 0),
+				newMsg(3, 1),
+			},
+			firstMsgAt:  2,
+			wantNumEval: 4,
+			wantMsgLens: []int{1, 1, 1, 1},
+			wantErr:     false,
+		},
+		{
+			name: "second last split last split",
+			msgs: []ipv6.Message{
+				newMsg(0, 0),
+				newMsg(0, 0),
+				newMsg(2, 1),
+				newMsg(2, 1),
+			},
+			firstMsgAt:  2,
+			wantNumEval: 4,
+			wantMsgLens: []int{1, 1, 1, 1},
+			wantErr:     false,
+		},
+		{
+			name: "second last no split last split overflow",
+			msgs: []ipv6.Message{
+				newMsg(0, 0),
+				newMsg(0, 0),
+				newMsg(1, 0),
+				newMsg(4, 1),
+			},
+			firstMsgAt:  2,
+			wantNumEval: 4,
+			wantMsgLens: []int{1, 1, 1, 1},
+			wantErr:     true,
+		},
+	}
+
+	for _, tt := range cases {
+		t.Run(tt.name, func(t *testing.T) {
+			got, err := splitCoalescedMessages(tt.msgs, 2, mockGetGSOSize)
+			if err != nil && !tt.wantErr {
+				t.Fatalf("err: %v", err)
+			}
+			if got != tt.wantNumEval {
+				t.Fatalf("got to eval: %d want: %d", got, tt.wantNumEval)
+			}
+			for i, msg := range tt.msgs {
+				if msg.N != tt.wantMsgLens[i] {
+					t.Fatalf("msg[%d].N: %d want: %d", i, msg.N, tt.wantMsgLens[i])
+				}
+			}
+		})
+	}
+}
diff --git a/conn/sticky_default.go b/conn/control_default.go
index 1fa8a0c..9459da5 100644
--- a/conn/sticky_default.go
+++ b/conn/control_default.go
@@ -21,8 +21,9 @@ func (e *StdNetEndpoint) SrcToString() string {
 	return ""
 }
 
-// TODO: macOS, FreeBSD and other BSDs likely do support this feature set, but
-// use alternatively named flags and need ports and require testing.
+// TODO: macOS, FreeBSD and other BSDs likely do support the sticky sockets
+// {get,set}srcControl feature set, but use alternatively named flags and need
+// ports and require testing.
 
 // getSrcFromControl parses the control for PKTINFO and if found updates ep with
 // the source information found.
@@ -34,8 +35,17 @@ func getSrcFromControl(control []byte, ep *StdNetEndpoint) {
 func setSrcControl(control *[]byte, ep *StdNetEndpoint) {
 }
 
-// srcControlSize returns the recommended buffer size for pooling sticky control
-// data.
-const srcControlSize = 0
+// getGSOSize parses control for UDP_GRO and if found returns its GSO size data.
+func getGSOSize(control []byte) (int, error) {
+	return 0, nil
+}
+
+// setGSOSize sets a UDP_SEGMENT in control based on gsoSize.
+func setGSOSize(control *[]byte, gsoSize uint16) {
+}
+
+// controlSize returns the recommended buffer size for pooling sticky and UDP
+// offloading control data.
+const controlSize = 0
 
 const StdNetSupportsStickySockets = false
diff --git a/conn/sticky_linux.go b/conn/control_linux.go
index a30ccc7..44a94e6 100644
--- a/conn/sticky_linux.go
+++ b/conn/control_linux.go
@@ -8,6 +8,7 @@
 package conn
 
 import (
+	"fmt"
 	"net/netip"
 	"unsafe"
 
@@ -105,6 +106,54 @@ func setSrcControl(control *[]byte, ep *StdNetEndpoint) {
 	*control = append(*control, ep.src...)
 }
 
-var srcControlSize = unix.CmsgSpace(unix.SizeofInet6Pktinfo)
+const (
+	sizeOfGSOData = 2
+)
+
+// getGSOSize parses control for UDP_GRO and if found returns its GSO size data.
+func getGSOSize(control []byte) (int, error) {
+	var (
+		hdr  unix.Cmsghdr
+		data []byte
+		rem  = control
+		err  error
+	)
+
+	for len(rem) > unix.SizeofCmsghdr {
+		hdr, data, rem, err = unix.ParseOneSocketControlMessage(rem)
+		if err != nil {
+			return 0, fmt.Errorf("error parsing socket control message: %w", err)
+		}
+		if hdr.Level == unix.SOL_UDP && hdr.Type == unix.UDP_GRO && len(data) >= sizeOfGSOData {
+			var gso uint16
+			copy(unsafe.Slice((*byte)(unsafe.Pointer(&gso)), sizeOfGSOData), data[:sizeOfGSOData])
+			return int(gso), nil
+		}
+	}
+	return 0, nil
+}
+
+// setGSOSize sets a UDP_SEGMENT in control based on gsoSize. It leaves existing
+// data in control untouched.
+func setGSOSize(control *[]byte, gsoSize uint16) {
+	existingLen := len(*control)
+	avail := cap(*control) - existingLen
+	space := unix.CmsgSpace(sizeOfGSOData)
+	if avail < space {
+		return
+	}
+	*control = (*control)[:cap(*control)]
+	gsoControl := (*control)[existingLen:]
+	hdr := (*unix.Cmsghdr)(unsafe.Pointer(&(gsoControl)[0]))
+	hdr.Level = unix.SOL_UDP
+	hdr.Type = unix.UDP_SEGMENT
+	hdr.SetLen(unix.CmsgLen(sizeOfGSOData))
+	copy((gsoControl)[unix.SizeofCmsghdr:], unsafe.Slice((*byte)(unsafe.Pointer(&gsoSize)), sizeOfGSOData))
+	*control = (*control)[:existingLen+space]
+}
+
+// controlSize returns the recommended buffer size for pooling sticky and UDP
+// offloading control data.
+var controlSize = unix.CmsgSpace(unix.SizeofInet6Pktinfo) + unix.CmsgSpace(sizeOfGSOData)
 
 const StdNetSupportsStickySockets = true
diff --git a/conn/sticky_linux_test.go b/conn/control_linux_test.go
index 679213a..96f9da2 100644
--- a/conn/sticky_linux_test.go
+++ b/conn/control_linux_test.go
@@ -60,7 +60,7 @@ func Test_setSrcControl(t *testing.T) {
 		}
 		setSrc(ep, netip.MustParseAddr("127.0.0.1"), 5)
 
-		control := make([]byte, srcControlSize)
+		control := make([]byte, controlSize)
 
 		setSrcControl(&control, ep)
 
@@ -89,7 +89,7 @@ func Test_setSrcControl(t *testing.T) {
 		}
 		setSrc(ep, netip.MustParseAddr("::1"), 5)
 
-		control := make([]byte, srcControlSize)
+		control := make([]byte, controlSize)
 
 		setSrcControl(&control, ep)
 
@@ -113,7 +113,7 @@ func Test_setSrcControl(t *testing.T) {
 	})
 
 	t.Run("ClearOnNoSrc", func(t *testing.T) {
-		control := make([]byte, unix.CmsgLen(0))
+		control := make([]byte, controlSize)
 		hdr := (*unix.Cmsghdr)(unsafe.Pointer(&control[0]))
 		hdr.Level = 1
 		hdr.Type = 2
@@ -129,7 +129,7 @@ func Test_setSrcControl(t *testing.T) {
 
 func Test_getSrcFromControl(t *testing.T) {
 	t.Run("IPv4", func(t *testing.T) {
-		control := make([]byte, unix.CmsgSpace(unix.SizeofInet4Pktinfo))
+		control := make([]byte, controlSize)
 		hdr := (*unix.Cmsghdr)(unsafe.Pointer(&control[0]))
 		hdr.Level = unix.IPPROTO_IP
 		hdr.Type = unix.IP_PKTINFO
@@ -149,7 +149,7 @@ func Test_getSrcFromControl(t *testing.T) {
 		}
 	})
 	t.Run("IPv6", func(t *testing.T) {
-		control := make([]byte, unix.CmsgSpace(unix.SizeofInet6Pktinfo))
+		control := make([]byte, controlSize)
 		hdr := (*unix.Cmsghdr)(unsafe.Pointer(&control[0]))
 		hdr.Level = unix.IPPROTO_IPV6
 		hdr.Type = unix.IPV6_PKTINFO
diff --git a/conn/controlfns_linux.go b/conn/controlfns_linux.go
index a2396fe..f6ab1d2 100644
--- a/conn/controlfns_linux.go
+++ b/conn/controlfns_linux.go
@@ -57,5 +57,13 @@ func init() {
 			}
 			return err
 		},
+
+		// Attempt to enable UDP_GRO
+		func(network, address string, c syscall.RawConn) error {
+			c.Control(func(fd uintptr) {
+				_ = unix.SetsockoptInt(int(fd), unix.IPPROTO_UDP, unix.UDP_GRO, 1)
+			})
+			return nil
+		},
 	)
 }
diff --git a/conn/errors_default.go b/conn/errors_default.go
new file mode 100644
index 0000000..f1e5b90
--- /dev/null
+++ b/conn/errors_default.go
@@ -0,0 +1,12 @@
+//go:build !linux
+
+/* SPDX-License-Identifier: MIT
+ *
+ * Copyright (C) 2017-2023 WireGuard LLC. All Rights Reserved.
+ */
+
+package conn
+
+func errShouldDisableUDPGSO(err error) bool {
+	return false
+}
diff --git a/conn/errors_linux.go b/conn/errors_linux.go
new file mode 100644
index 0000000..8e61000
--- /dev/null
+++ b/conn/errors_linux.go
@@ -0,0 +1,26 @@
+/* SPDX-License-Identifier: MIT
+ *
+ * Copyright (C) 2017-2023 WireGuard LLC. All Rights Reserved.
+ */
+
+package conn
+
+import (
+	"errors"
+	"os"
+
+	"golang.org/x/sys/unix"
+)
+
+func errShouldDisableUDPGSO(err error) bool {
+	var serr *os.SyscallError
+	if errors.As(err, &serr) {
+		// EIO is returned by udp_send_skb() if the device driver does not have
+		// tx checksumming enabled, which is a hard requirement of UDP_SEGMENT.
+		// See:
+		// https://git.kernel.org/pub/scm/docs/man-pages/man-pages.git/tree/man7/udp.7?id=806eabd74910447f21005160e90957bde4db0183#n228
+		// https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/net/ipv4/udp.c?h=v6.2&id=c9c3395d5e3dcc6daee66c6908354d47bf98cb0c#n942
+		return serr.Err == unix.EIO
+	}
+	return false
+}
diff --git a/conn/features_default.go b/conn/features_default.go
new file mode 100644
index 0000000..d53ff5f
--- /dev/null
+++ b/conn/features_default.go
@@ -0,0 +1,15 @@
+//go:build !linux
+// +build !linux
+
+/* SPDX-License-Identifier: MIT
+ *
+ * Copyright (C) 2017-2023 WireGuard LLC. All Rights Reserved.
+ */
+
+package conn
+
+import "net"
+
+func supportsUDPOffload(conn *net.UDPConn) (txOffload, rxOffload bool) {
+	return
+}
diff --git a/conn/features_linux.go b/conn/features_linux.go
new file mode 100644
index 0000000..e1fb57f
--- /dev/null
+++ b/conn/features_linux.go
@@ -0,0 +1,35 @@
+/* SPDX-License-Identifier: MIT
+ *
+ * Copyright (C) 2017-2023 WireGuard LLC. All Rights Reserved.
+ */
+
+package conn
+
+import (
+	"net"
+
+	"golang.org/x/sys/unix"
+)
+
+func supportsUDPOffload(conn *net.UDPConn) (txOffload, rxOffload bool) {
+	rc, err := conn.SyscallConn()
+	if err != nil {
+		return
+	}
+	err = rc.Control(func(fd uintptr) {
+		_, errSyscall := unix.GetsockoptInt(int(fd), unix.IPPROTO_UDP, unix.UDP_SEGMENT)
+		if errSyscall != nil {
+			return
+		}
+		txOffload = true
+		opt, errSyscall := unix.GetsockoptInt(int(fd), unix.IPPROTO_UDP, unix.UDP_GRO)
+		if errSyscall != nil {
+			return
+		}
+		rxOffload = opt == 1
+	})
+	if err != nil {
+		return false, false
+	}
+	return txOffload, rxOffload
+}
diff --git a/device/send.go b/device/send.go
index d22bf26..cd8a2a0 100644
--- a/device/send.go
+++ b/device/send.go
@@ -17,6 +17,7 @@ import (
 	"golang.org/x/crypto/chacha20poly1305"
 	"golang.org/x/net/ipv4"
 	"golang.org/x/net/ipv6"
+	"golang.zx2c4.com/wireguard/conn"
 	"golang.zx2c4.com/wireguard/tun"
 )
 
@@ -526,6 +527,13 @@ func (peer *Peer) RoutineSequentialSender(maxBatchSize int) {
 		}
 		device.PutOutboundElementsSlice(elems)
 		if err != nil {
+			var errGSO conn.ErrUDPGSODisabled
+			if errors.As(err, &errGSO) {
+				device.log.Verbosef(err.Error())
+				err = errGSO.RetryErr
+			}
+		}
+		if err != nil {
 			device.log.Errorf("%v - Failed to send data packets: %v", peer, err)
 			continue
 		}
diff --git a/go.mod b/go.mod
index c04e1bb..758dcde 100644
--- a/go.mod
+++ b/go.mod
@@ -5,7 +5,7 @@ go 1.20
 require (
 	golang.org/x/crypto v0.6.0
 	golang.org/x/net v0.7.0
-	golang.org/x/sys v0.5.1-0.20230222185716-a3b23cc77e89
+	golang.org/x/sys v0.12.0
 	golang.zx2c4.com/wintun v0.0.0-20230126152724-0fa3db229ce2
 	gvisor.dev/gvisor v0.0.0-20221203005347-703fd9b7fbc0
 )
diff --git a/go.sum b/go.sum
index cfeaee6..fe4ca7e 100644
--- a/go.sum
+++ b/go.sum
@@ -4,8 +4,8 @@ golang.org/x/crypto v0.6.0 h1:qfktjS5LUO+fFKeJXZ+ikTRijMmljikvG68fpMMruSc=
 golang.org/x/crypto v0.6.0/go.mod h1:OFC/31mSvZgRz0V1QTNCzfAI1aIRzbiufJtkMIlEp58=
 golang.org/x/net v0.7.0 h1:rJrUqqhjsgNp7KqAIc25s9pZnjU7TUcSY7HcVZjdn1g=
 golang.org/x/net v0.7.0/go.mod h1:2Tu9+aMcznHK/AK1HMvgo6xiTLG5rD5rZLDS+rp2Bjs=
-golang.org/x/sys v0.5.1-0.20230222185716-a3b23cc77e89 h1:260HNjMTPDya+jq5AM1zZLgG9pv9GASPAGiEEJUbRg4=
-golang.org/x/sys v0.5.1-0.20230222185716-a3b23cc77e89/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
+golang.org/x/sys v0.12.0 h1:CM0HF96J0hcLAwsHPJZjfdNzs0gftsLfgKt57wWHJ0o=
+golang.org/x/sys v0.12.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/time v0.0.0-20191024005414-555d28b269f0 h1:/5xXl8Y5W96D+TtHSlonuFqGHIWVuyCkGJLwGh9JJFs=
 golang.org/x/time v0.0.0-20191024005414-555d28b269f0/go.mod h1:tRJNPiyCQ0inRvYxbN9jk5I+vvW/OXSQhTDSoE431IQ=
 golang.zx2c4.com/wintun v0.0.0-20230126152724-0fa3db229ce2 h1:B82qJJgjvYKsXS9jeunTOisW56dUokqW/FOteYJJ/yg=