remove gso, leave nice pretty GRO

cmd/gso/gso.go

@@ -0,0 +1,191 @@
+package main
+
+import (
+	"encoding/binary"
+	"errors"
+	"flag"
+	"fmt"
+	"log"
+	"net"
+	"net/netip"
+	"time"
+	"unsafe"
+
+	"golang.org/x/sys/unix"
+)
+
+const (
+	// UDP_SEGMENT enables GSO segmentation
+	UDP_SEGMENT = 103
+	// Maximum GSO segment size (typical MTU - headers)
+	maxGSOSize = 1400
+)
+
+func main() {
+	destAddr := flag.String("dest", "10.4.0.16:4202", "Destination address")
+	gsoSize := flag.Int("gso", 1400, "GSO segment size")
+	totalSize := flag.Int("size", 14000, "Total payload size to send")
+	count := flag.Int("count", 1, "Number of packets to send")
+	flag.Parse()
+
+	if *gsoSize > maxGSOSize {
+		log.Fatalf("GSO size %d exceeds maximum %d", *gsoSize, maxGSOSize)
+	}
+
+	// Resolve destination address
+	_, err := net.ResolveUDPAddr("udp", *destAddr)
+	if err != nil {
+		log.Fatalf("Failed to resolve address: %v", err)
+	}
+
+	// Create a raw UDP socket with GSO support
+	fd, err := unix.Socket(unix.AF_INET, unix.SOCK_DGRAM, unix.IPPROTO_UDP)
+	if err != nil {
+		log.Fatalf("Failed to create socket: %v", err)
+	}
+	defer unix.Close(fd)
+
+	// Bind to a local address
+	localAddr := &unix.SockaddrInet4{
+		Port: 0, // Let the system choose a port
+	}
+	if err := unix.Bind(fd, localAddr); err != nil {
+		log.Fatalf("Failed to bind socket: %v", err)
+	}
+
+	fmt.Printf("Sending UDP packets with GSO enabled\n")
+	fmt.Printf("Destination: %s\n", *destAddr)
+	fmt.Printf("GSO segment size: %d bytes\n", *gsoSize)
+	fmt.Printf("Total payload size: %d bytes\n", *totalSize)
+	fmt.Printf("Number of packets: %d\n\n", *count)
+
+	// Create payload
+	payload := make([]byte, *totalSize)
+	for i := range payload {
+		payload[i] = byte(i % 256)
+	}
+
+	dest := netip.MustParseAddrPort(*destAddr)
+
+	//if err := unix.SetsockoptInt(fd, unix.SOL_UDP, unix.UDP_SEGMENT, 1400); err != nil {
+	//	panic(err)
+	//}
+
+	for i := 0; i < *count; i++ {
+		err := WriteBatch(fd, payload, dest, uint16(*gsoSize), true)
+		if err != nil {
+			log.Printf("Send error on packet %d: %v", i, err)
+			continue
+		}
+
+		if (i+1)%100 == 0 || i == *count-1 {
+			fmt.Printf("Sent %d packets\n", i+1)
+		}
+	}
+	fmt.Printf("now, let's send without the correct ctrl header\n")
+	time.Sleep(time.Second)
+	for i := 0; i < *count; i++ {
+		err := WriteBatch(fd, payload, dest, uint16(*gsoSize), false)
+		if err != nil {
+			log.Printf("Send error on packet %d: %v", i, err)
+			continue
+		}
+
+		if (i+1)%100 == 0 || i == *count-1 {
+			fmt.Printf("Sent %d packets\n", i+1)
+		}
+	}
+
+}
+
+func WriteBatch(fd int, payload []byte, addr netip.AddrPort, segSize uint16, withHeader bool) error {
+	msgs := make([]rawMessage, 0, 1)
+	iovs := make([]iovec, 0, 1)
+	names := make([][unix.SizeofSockaddrInet6]byte, 0, 1)
+
+	sent := 0
+
+	pkts := []BatchPacket{
+		{
+			Payload: payload,
+			Addr:    addr,
+		},
+	}
+
+	for _, pkt := range pkts {
+		if len(pkt.Payload) == 0 {
+			sent++
+			continue
+		}
+
+		msgs = append(msgs, rawMessage{})
+		iovs = append(iovs, iovec{})
+		names = append(names, [unix.SizeofSockaddrInet6]byte{})
+
+		idx := len(msgs) - 1
+		msg := &msgs[idx]
+		iov := &iovs[idx]
+		name := &names[idx]
+
+		setIovecSlice(iov, pkt.Payload)
+		msg.Hdr.Iov = iov
+		msg.Hdr.Iovlen = 1
+
+		if withHeader {
+			setRawMessageControl(msg, buildGSOControlMessage(segSize)) //
+		} else {
+			setRawMessageControl(msg, nil) //
+		}
+
+		msg.Hdr.Flags = 0
+
+		nameLen, err := encodeSockaddr(name[:], pkt.Addr)
+		if err != nil {
+			return err
+		}
+		msg.Hdr.Name = &name[0]
+		msg.Hdr.Namelen = nameLen
+	}
+
+	if len(msgs) == 0 {
+		return errors.New("nothing to write")
+	}
+
+	offset := 0
+	for offset < len(msgs) {
+		n, _, errno := unix.Syscall6(
+			unix.SYS_SENDMMSG,
+			uintptr(fd),
+			uintptr(unsafe.Pointer(&msgs[offset])),
+			uintptr(len(msgs)-offset),
+			0,
+			0,
+			0,
+		)
+
+		if errno != 0 {
+			if errno == unix.EINTR {
+				continue
+			}
+			return &net.OpError{Op: "sendmmsg", Err: errno}
+		}
+
+		if n == 0 {
+			break
+		}
+		offset += int(n)
+	}
+
+	return nil
+}
+
+func buildGSOControlMessage(segSize uint16) []byte {
+	control := make([]byte, unix.CmsgSpace(2))
+	hdr := (*unix.Cmsghdr)(unsafe.Pointer(&control[0]))
+	hdr.Level = unix.SOL_UDP
+	hdr.Type = unix.UDP_SEGMENT
+	setCmsgLen(hdr, unix.CmsgLen(2))
+	binary.NativeEndian.PutUint16(control[unix.CmsgLen(0):unix.CmsgLen(0)+2], uint16(segSize))
+
+	return control
+}

cmd/gso/helper.go

@@ -0,0 +1,85 @@
+package main
+
+import (
+	"encoding/binary"
+	"fmt"
+	"net/netip"
+	"unsafe"
+
+	"golang.org/x/sys/unix"
+)
+
+type iovec struct {
+	Base *byte
+	Len  uint64
+}
+
+type msghdr struct {
+	Name       *byte
+	Namelen    uint32
+	Pad0       [4]byte
+	Iov        *iovec
+	Iovlen     uint64
+	Control    *byte
+	Controllen uint64
+	Flags      int32
+	Pad1       [4]byte
+}
+
+type rawMessage struct {
+	Hdr  msghdr
+	Len  uint32
+	Pad0 [4]byte
+}
+
+type BatchPacket struct {
+	Payload []byte
+	Addr    netip.AddrPort
+}
+
+func encodeSockaddr(dst []byte, addr netip.AddrPort) (uint32, error) {
+	if addr.Addr().Is4() {
+		if !addr.Addr().Is4() {
+			return 0, fmt.Errorf("Listener is IPv4, but writing to IPv6 remote")
+		}
+		var sa unix.RawSockaddrInet4
+		sa.Family = unix.AF_INET
+		sa.Addr = addr.Addr().As4()
+		binary.BigEndian.PutUint16((*[2]byte)(unsafe.Pointer(&sa.Port))[:], addr.Port())
+		size := unix.SizeofSockaddrInet4
+		copy(dst[:size], (*(*[unix.SizeofSockaddrInet4]byte)(unsafe.Pointer(&sa)))[:])
+		return uint32(size), nil
+	}
+
+	var sa unix.RawSockaddrInet6
+	sa.Family = unix.AF_INET6
+	sa.Addr = addr.Addr().As16()
+	binary.BigEndian.PutUint16((*[2]byte)(unsafe.Pointer(&sa.Port))[:], addr.Port())
+	size := unix.SizeofSockaddrInet6
+	copy(dst[:size], (*(*[unix.SizeofSockaddrInet6]byte)(unsafe.Pointer(&sa)))[:])
+	return uint32(size), nil
+}
+
+func setRawMessageControl(msg *rawMessage, buf []byte) {
+	if len(buf) == 0 {
+		msg.Hdr.Control = nil
+		msg.Hdr.Controllen = 0
+		return
+	}
+	msg.Hdr.Control = &buf[0]
+	msg.Hdr.Controllen = uint64(len(buf))
+}
+
+func setCmsgLen(h *unix.Cmsghdr, l int) {
+	h.Len = uint64(l)
+}
+
+func setIovecSlice(iov *iovec, b []byte) {
+	if len(b) == 0 {
+		iov.Base = nil
+		iov.Len = 0
+		return
+	}
+	iov.Base = &b[0]
+	iov.Len = uint64(len(b))
+}

interface.go

@@ -207,7 +207,7 @@ func NewInterface(ctx context.Context, c *InterfaceConfig) (*Interface, error) {
 		l: c.l,
 	}
 
-	ifce.pktPool = packet.NewPool()
+	ifce.pktPool = packet.GetPool()
 
 	ifce.tryPromoteEvery.Store(c.tryPromoteEvery)
 	ifce.reQueryEvery.Store(c.reQueryEvery)
@@ -338,7 +338,18 @@ func (f *Interface) workerIn(i int, ctx context.Context) {
 	for {
 		select {
 		case p := <-f.inbound:
-			f.readOutsidePackets(p.Addr, nil, result2[:0], p.Payload, h, fwPacket2, lhh, nb2, i, conntrackCache.Get(f.l))
+			if p.SegSize > 0 && p.SegSize < len(p.Payload) {
+				for offset := 0; offset < len(p.Payload); offset += p.SegSize {
+					end := offset + p.SegSize
+					if end > len(p.Payload) {
+						end = len(p.Payload)
+					}
+					f.readOutsidePackets(p.Addr, nil, result2[:0], p.Payload[offset:end], h, fwPacket2, lhh, nb2, i, conntrackCache.Get(f.l))
+				}
+			} else {
+				f.readOutsidePackets(p.Addr, nil, result2[:0], p.Payload, h, fwPacket2, lhh, nb2, i, conntrackCache.Get(f.l))
+			}
+
 			f.pktPool.Put(p)
 		case <-ctx.Done():
 			f.wg.Done()

packet/packet.go

@@ -3,27 +3,36 @@ package packet
 import (
 	"net/netip"
 	"sync"
+
+	"golang.org/x/sys/unix"
 )
 
-const Size = 9001
+const Size = 0xffff
 
 type Packet struct {
 	Payload []byte
+	Control []byte
+	SegSize int
 	Addr    netip.AddrPort
 }
 
 func New() *Packet {
-	return &Packet{Payload: make([]byte, Size)}
+	return &Packet{
+		Payload: make([]byte, Size),
+		Control: make([]byte, unix.CmsgSpace(2)),
+	}
 }
 
 type Pool struct {
 	pool sync.Pool
 }
 
-func NewPool() *Pool {
-	return &Pool{
-		pool: sync.Pool{New: func() any { return New() }},
-	}
+var bigPool = &Pool{
+	pool: sync.Pool{New: func() any { return New() }},
+}
+
+func GetPool() *Pool {
+	return bigPool
 }
 
 func (p *Pool) Get() *Packet {

udp/udp_linux.go

@@ -22,10 +22,13 @@ import (
 var readTimeout = unix.NsecToTimeval(int64(time.Millisecond * 500))
 
 type StdConn struct {
-	sysFd int
-	isV4  bool
-	l     *logrus.Logger
-	batch int
+	sysFd     int
+	isV4      bool
+	l         *logrus.Logger
+	batch     int
+	enableGRO bool
+	enableGSO bool
+	//gso       gsoState
 }
 
 func NewListener(l *logrus.Logger, ip netip.Addr, port int, multi bool, batch int) (Conn, error) {
@@ -145,15 +148,47 @@ func (u *StdConn) ListenOut(pg PacketBufferGetter, pc chan *packet.Packet) error
 				ip, _ = netip.AddrFromSlice(names[i][8:24])
 			}
 			out.Addr = netip.AddrPortFrom(ip.Unmap(), binary.BigEndian.Uint16(names[i][2:4]))
+			ctrlLen := getRawMessageControlLen(&msgs[i])
+			if ctrlLen > 0 {
+				packets[i].SegSize = parseGROControl(packets[i].Control[:ctrlLen])
+			} else {
+				packets[i].SegSize = 0
+			}
+
 			pc <- out
 
 			//rotate this packet out so we don't overwrite it
 			packets[i] = pg()
 			msgs[i].Hdr.Iov.Base = &packets[i].Payload[0]
+			if u.enableGRO {
+				msgs[i].Hdr.Control = &packets[i].Control[0]
+				msgs[i].Hdr.Controllen = uint64(cap(packets[i].Control))
+			}
+
 		}
 	}
 }
 
+func parseGROControl(control []byte) int {
+	if len(control) == 0 {
+		return 0
+	}
+
+	cmsgs, err := unix.ParseSocketControlMessage(control)
+	if err != nil {
+		return 0
+	}
+
+	for _, c := range cmsgs {
+		if c.Header.Level == unix.SOL_UDP && c.Header.Type == unix.UDP_GRO && len(c.Data) >= 2 {
+			segSize := int(binary.LittleEndian.Uint16(c.Data[:2]))
+			return segSize
+		}
+	}
+
+	return 0
+}
+
 func (u *StdConn) ReadSingle(msgs []rawMessage) (int, error) {
 	for {
 		n, _, err := unix.Syscall6(
@@ -308,6 +343,28 @@ func (u *StdConn) ReloadConfig(c *config.C) {
 			u.l.WithError(err).Error("Failed to set listen.so_mark")
 		}
 	}
+	u.configureGRO(true)
+}
+
+func (u *StdConn) configureGRO(enable bool) {
+	if enable == u.enableGRO {
+		return
+	}
+
+	if enable {
+		if err := unix.SetsockoptInt(u.sysFd, unix.SOL_UDP, unix.UDP_GRO, 1); err != nil {
+			u.l.WithError(err).Warn("Failed to enable UDP GRO")
+			return
+		}
+		u.enableGRO = true
+		u.l.Info("UDP GRO enabled")
+		return
+	}
+
+	if err := unix.SetsockoptInt(u.sysFd, unix.SOL_UDP, unix.UDP_GRO, 0); err != nil && err != unix.ENOPROTOOPT {
+		u.l.WithError(err).Warn("Failed to disable UDP GRO")
+	}
+	u.enableGRO = false
 }
 
 func (u *StdConn) getMemInfo(meminfo *[unix.SK_MEMINFO_VARS]uint32) error {

udp/udp_linux_64.go

@@ -34,6 +34,24 @@ type rawMessage struct {
 	Pad0 [4]byte
 }
 
+func setRawMessageControl(msg *rawMessage, buf []byte) {
+	if len(buf) == 0 {
+		msg.Hdr.Control = nil
+		msg.Hdr.Controllen = 0
+		return
+	}
+	msg.Hdr.Control = &buf[0]
+	msg.Hdr.Controllen = uint64(len(buf))
+}
+
+func getRawMessageControlLen(msg *rawMessage) int {
+	return int(msg.Hdr.Controllen)
+}
+
+func setCmsgLen(h *unix.Cmsghdr, l int) {
+	h.Len = uint64(l)
+}
+
 func (u *StdConn) PrepareRawMessages(n int, pg PacketBufferGetter) ([]rawMessage, []*packet.Packet, [][]byte) {
 	msgs := make([]rawMessage, n)
 	names := make([][]byte, n)
@@ -42,6 +60,7 @@ func (u *StdConn) PrepareRawMessages(n int, pg PacketBufferGetter) ([]rawMessage
 	for i := range packets {
 		packets[i] = pg()
 	}
+	//todo?
 
 	for i := range msgs {
 		names[i] = make([]byte, unix.SizeofSockaddrInet6)
@@ -55,6 +74,13 @@ func (u *StdConn) PrepareRawMessages(n int, pg PacketBufferGetter) ([]rawMessage
 
 		msgs[i].Hdr.Name = &names[i][0]
 		msgs[i].Hdr.Namelen = uint32(len(names[i]))
+		if u.enableGRO {
+			msgs[i].Hdr.Control = &packets[i].Control[0]
+			msgs[i].Hdr.Controllen = uint64(len(packets[i].Control))
+		} else {
+			msgs[i].Hdr.Control = nil
+			msgs[i].Hdr.Controllen = 0
+		}
 	}
 
 	return msgs, packets, names