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))
+}

connection_state.go

@@ -15,7 +15,7 @@ import (
 
 // TODO: In a 5Gbps test, 1024 is not sufficient. With a 1400 MTU this is about 1.4Gbps of window, assuming full packets.
 // 4092 should be sufficient for 5Gbps
-const ReplayWindow = 1024
+const ReplayWindow = 8192
 
 type ConnectionState struct {
 	eKey           *NebulaCipherState

interface.go

@@ -96,11 +96,9 @@ type Interface struct {
 
 	l *logrus.Logger
 
-	inPool  sync.Pool
+	pktPool  *packet.Pool
 	inbound  chan *packet.Packet
-
-	outPool  sync.Pool
-	outbound chan *[]byte
+	outbound chan *packet.Packet
 }
 
 type EncWriter interface {
@@ -203,20 +201,13 @@ func NewInterface(ctx context.Context, c *InterfaceConfig) (*Interface, error) {
 		},
 
 		//TODO: configurable size
-		inbound:  make(chan *packet.Packet, 1028),
-		outbound: make(chan *[]byte, 1028),
+		inbound:  make(chan *packet.Packet, 2048),
+		outbound: make(chan *packet.Packet, 2048),
 
 		l: c.l,
 	}
 
-	ifce.inPool = sync.Pool{New: func() any {
-		return packet.New()
-	}}
-
-	ifce.outPool = sync.Pool{New: func() any {
-		t := make([]byte, mtu)
-		return &t
-	}}
+	ifce.pktPool = packet.GetPool()
 
 	ifce.tryPromoteEvery.Store(c.tryPromoteEvery)
 	ifce.reQueryEvery.Store(c.reQueryEvery)
@@ -267,19 +258,21 @@ func (f *Interface) activate() error {
 
 func (f *Interface) run(c context.Context) (func(), error) {
 	for i := 0; i < f.routines; i++ {
-		// Launch n queues to read packets from udp
+		// read packets from udp and queue to f.inbound
 		f.wg.Add(1)
 		go f.listenOut(i)
 
-		// Launch n queues to read packets from tun dev
-		f.wg.Add(1)
+		// Launch n queues to read packets from inside tun dev and queue to f.outbound
+		//todo this never stops f.wg.Add(1)
 		go f.listenIn(f.readers[i], i)
 
-		// Launch n queues to read packets from tun dev
+		// Launch n workers to process traffic from f.inbound and smash it onto the inside of the tun
+		f.wg.Add(1)
+		go f.workerIn(i, c)
 		f.wg.Add(1)
 		go f.workerIn(i, c)
 
-		// Launch n queues to read packets from tun dev
+		// read from f.outbound and write to UDP (outside the tun)
 		f.wg.Add(1)
 		go f.workerOut(i, c)
 	}
@@ -296,22 +289,7 @@ func (f *Interface) listenOut(i int) {
 		li = f.outside
 	}
 
-	err := li.ListenOut(func(fromUdpAddr netip.AddrPort, payload []byte) {
-		p := f.inPool.Get().(*packet.Packet)
-		//TODO: have the listener store this in the msgs array after a read instead of doing a copy
-
-		p.Payload = p.Payload[:mtu]
-		copy(p.Payload, payload)
-		p.Payload = p.Payload[:len(payload)]
-		p.Addr = fromUdpAddr
-		f.inbound <- p
-		//select {
-		//case f.inbound <- p:
-		//default:
-		//	f.l.Error("Dropped packet from inbound channel")
-		//}
-	})
-
+	err := li.ListenOut(f.pktPool.Get, f.inbound)
 	if err != nil && !f.closed.Load() {
 		f.l.WithError(err).Error("Error while reading packet inbound packet, closing")
 		//TODO: Trigger Control to close
@@ -325,9 +303,8 @@ func (f *Interface) listenIn(reader io.ReadWriteCloser, i int) {
 	runtime.LockOSThread()
 
 	for {
-		p := f.outPool.Get().(*[]byte)
-		*p = (*p)[:mtu]
-		n, err := reader.Read(*p)
+		p := f.pktPool.Get()
+		n, err := reader.Read(p.Payload)
 		if err != nil {
 			if !f.closed.Load() {
 				f.l.WithError(err).Error("Error while reading outbound packet, closing")
@@ -336,7 +313,7 @@ func (f *Interface) listenIn(reader io.ReadWriteCloser, i int) {
 			break
 		}
 
-		*p = (*p)[:n]
+		p.Payload = (p.Payload)[:n]
 		//TODO: nonblocking channel write
 		f.outbound <- p
 		//select {
@@ -361,9 +338,19 @@ func (f *Interface) workerIn(i int, ctx context.Context) {
 	for {
 		select {
 		case p := <-f.inbound:
+			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))
-			p.Payload = p.Payload[:mtu]
-			f.inPool.Put(p)
+			}
+
+			f.pktPool.Put(p)
 		case <-ctx.Done():
 			f.wg.Done()
 			return
@@ -380,9 +367,8 @@ func (f *Interface) workerOut(i int, ctx context.Context) {
 	for {
 		select {
 		case data := <-f.outbound:
-			f.consumeInsidePacket(*data, fwPacket1, nb1, result1, i, conntrackCache.Get(f.l))
-			*data = (*data)[:mtu]
-			f.outPool.Put(data)
+			f.consumeInsidePacket(data.Payload, fwPacket1, nb1, result1, i, conntrackCache.Get(f.l))
+			f.pktPool.Put(data)
 		case <-ctx.Done():
 			f.wg.Done()
 			return

packet/packet.go

@@ -1,12 +1,45 @@
 package packet
 
-import "net/netip"
+import (
+	"net/netip"
+	"sync"
+
+	"golang.org/x/sys/unix"
+)
+
+const Size = 0xffff
 
 type Packet struct {
 	Payload []byte
+	Control []byte
+	SegSize int
 	Addr    netip.AddrPort
 }
 
 func New() *Packet {
-	return &Packet{Payload: make([]byte, 9001)}
+	return &Packet{
+		Payload: make([]byte, Size),
+		Control: make([]byte, unix.CmsgSpace(2)),
+	}
+}
+
+type Pool struct {
+	pool sync.Pool
+}
+
+var bigPool = &Pool{
+	pool: sync.Pool{New: func() any { return New() }},
+}
+
+func GetPool() *Pool {
+	return bigPool
+}
+
+func (p *Pool) Get() *Packet {
+	return p.pool.Get().(*Packet)
+}
+
+func (p *Pool) Put(x *Packet) {
+	x.Payload = x.Payload[:Size]
+	p.pool.Put(x)
 }

udp/conn.go

@@ -4,19 +4,19 @@ import (
 	"net/netip"
 
 	"github.com/slackhq/nebula/config"
+	"github.com/slackhq/nebula/packet"
 )
 
 const MTU = 9001
 
-type EncReader func(
-	addr netip.AddrPort,
-	payload []byte,
-)
+type EncReader func(*packet.Packet)
+
+type PacketBufferGetter func() *packet.Packet
 
 type Conn interface {
 	Rebind() error
 	LocalAddr() (netip.AddrPort, error)
-	ListenOut(r EncReader) error
+	ListenOut(pg PacketBufferGetter, pc chan *packet.Packet) error
 	WriteTo(b []byte, addr netip.AddrPort) error
 	ReloadConfig(c *config.C)
 	Close() error

udp/udp_linux.go

@@ -15,6 +15,7 @@ import (
 	"github.com/rcrowley/go-metrics"
 	"github.com/sirupsen/logrus"
 	"github.com/slackhq/nebula/config"
+	"github.com/slackhq/nebula/packet"
 	"golang.org/x/sys/unix"
 )
 
@@ -25,6 +26,9 @@ type StdConn struct {
 	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) {
@@ -118,10 +122,10 @@ func (u *StdConn) LocalAddr() (netip.AddrPort, error) {
 	}
 }
 
-func (u *StdConn) ListenOut(r EncReader) error {
+func (u *StdConn) ListenOut(pg PacketBufferGetter, pc chan *packet.Packet) error {
 	var ip netip.Addr
 
-	msgs, buffers, names := u.PrepareRawMessages(u.batch)
+	msgs, packets, names := u.PrepareRawMessages(u.batch, pg)
 	read := u.ReadMulti
 	if u.batch == 1 {
 		read = u.ReadSingle
@@ -134,17 +138,57 @@ func (u *StdConn) ListenOut(r EncReader) error {
 		}
 
 		for i := 0; i < n; i++ {
+			out := packets[i]
+			out.Payload = out.Payload[:msgs[i].Len]
+
 			// Its ok to skip the ok check here, the slicing is the only error that can occur and it will panic
 			if u.isV4 {
 				ip, _ = netip.AddrFromSlice(names[i][4:8])
 			} else {
 				ip, _ = netip.AddrFromSlice(names[i][8:24])
 			}
-			r(netip.AddrPortFrom(ip.Unmap(), binary.BigEndian.Uint16(names[i][2:4])), buffers[i][:msgs[i].Len])
+			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(
@@ -299,6 +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

@@ -7,6 +7,7 @@
 package udp
 
 import (
+	"github.com/slackhq/nebula/packet"
 	"golang.org/x/sys/unix"
 )
 
@@ -33,17 +34,39 @@ type rawMessage struct {
 	Pad0 [4]byte
 }
 
-func (u *StdConn) PrepareRawMessages(n int) ([]rawMessage, [][]byte, [][]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)
-	buffers := make([][]byte, n)
 	names := make([][]byte, n)
 
+	packets := make([]*packet.Packet, n)
+	for i := range packets {
+		packets[i] = pg()
+	}
+	//todo?
+
 	for i := range msgs {
-		buffers[i] = make([]byte, MTU)
 		names[i] = make([]byte, unix.SizeofSockaddrInet6)
 
 		vs := []iovec{
-			{Base: &buffers[i][0], Len: uint64(len(buffers[i]))},
+			{Base: &packets[i].Payload[0], Len: uint64(packet.Size)},
 		}
 
 		msgs[i].Hdr.Iov = &vs[0]
@@ -51,7 +74,14 @@ func (u *StdConn) PrepareRawMessages(n int) ([]rawMessage, [][]byte, [][]byte) {
 
 		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, buffers, names
+	return msgs, packets, names
 }