claude does TUN virtio header support

* specifically report if nebula start fails because of tun device name

* close all routines when closing the tun

inside.go

@@ -9,8 +9,75 @@ import (
 	"github.com/slackhq/nebula/iputil"
 	"github.com/slackhq/nebula/noiseutil"
 	"github.com/slackhq/nebula/routing"
+	"github.com/slackhq/nebula/udp"
 )
 
+// consumeInsidePacketBatched is a variant of consumeInsidePacket that queues
+// outgoing packets into pendingPackets instead of sending them immediately.
+// The caller is responsible for flushing pendingPackets with WriteBatch.
+func (f *Interface) consumeInsidePacketBatched(packet []byte, fwPacket *firewall.Packet, nb, out []byte, q int, localCache firewall.ConntrackCache, pendingPackets *[]udp.BatchPacket) {
+	err := newPacket(packet, false, fwPacket)
+	if err != nil {
+		if f.l.Level >= logrus.DebugLevel {
+			f.l.WithField("packet", packet).Debugf("Error while validating outbound packet: %s", err)
+		}
+		return
+	}
+
+	// Ignore local broadcast packets
+	if f.dropLocalBroadcast {
+		if f.myBroadcastAddrsTable.Contains(fwPacket.RemoteAddr) {
+			return
+		}
+	}
+
+	if f.myVpnAddrsTable.Contains(fwPacket.RemoteAddr) {
+		if immediatelyForwardToSelf {
+			_, err := f.readers[q].Write(packet)
+			if err != nil {
+				f.l.WithError(err).Error("Failed to forward to tun")
+			}
+		}
+		return
+	}
+
+	// Ignore multicast packets
+	if f.dropMulticast && fwPacket.RemoteAddr.IsMulticast() {
+		return
+	}
+
+	hostinfo, ready := f.getOrHandshakeConsiderRouting(fwPacket, func(hh *HandshakeHostInfo) {
+		hh.cachePacket(f.l, header.Message, 0, packet, f.sendMessageNow, f.cachedPacketMetrics)
+	})
+
+	if hostinfo == nil {
+		f.rejectInside(packet, out, q)
+		if f.l.Level >= logrus.DebugLevel {
+			f.l.WithField("vpnAddr", fwPacket.RemoteAddr).
+				WithField("fwPacket", fwPacket).
+				Debugln("dropping outbound packet, vpnAddr not in our vpn networks or in unsafe networks")
+		}
+		return
+	}
+
+	if !ready {
+		return
+	}
+
+	dropReason := f.firewall.Drop(*fwPacket, false, hostinfo, f.pki.GetCAPool(), localCache)
+	if dropReason == nil {
+		f.sendNoMetricsBatched(header.Message, 0, hostinfo.ConnectionState, hostinfo, netip.AddrPort{}, packet, nb, out, q, pendingPackets)
+	} else {
+		f.rejectInside(packet, out, q)
+		if f.l.Level >= logrus.DebugLevel {
+			hostinfo.logger(f.l).
+				WithField("fwPacket", fwPacket).
+				WithField("reason", dropReason).
+				Debugln("dropping outbound packet")
+		}
+	}
+}
+
 func (f *Interface) consumeInsidePacket(packet []byte, fwPacket *firewall.Packet, nb, out []byte, q int, localCache firewall.ConntrackCache) {
 	err := newPacket(packet, false, fwPacket)
 	if err != nil {
@@ -69,7 +136,6 @@ func (f *Interface) consumeInsidePacket(packet []byte, fwPacket *firewall.Packet
 	dropReason := f.firewall.Drop(*fwPacket, false, hostinfo, f.pki.GetCAPool(), localCache)
 	if dropReason == nil {
 		f.sendNoMetrics(header.Message, 0, hostinfo.ConnectionState, hostinfo, netip.AddrPort{}, packet, nb, out, q)
-
 	} else {
 		f.rejectInside(packet, out, q)
 		if f.l.Level >= logrus.DebugLevel {
@@ -410,3 +476,75 @@ func (f *Interface) sendNoMetrics(t header.MessageType, st header.MessageSubType
 		}
 	}
 }
+
+// sendNoMetricsBatched is like sendNoMetrics but queues the packet for batched sending
+// instead of sending immediately. The caller must flush pendingPackets with WriteBatch.
+func (f *Interface) sendNoMetricsBatched(t header.MessageType, st header.MessageSubType, ci *ConnectionState, hostinfo *HostInfo, remote netip.AddrPort, p, nb, out []byte, q int, pendingPackets *[]udp.BatchPacket) {
+	if ci.eKey == nil {
+		return
+	}
+	useRelay := !remote.IsValid() && !hostinfo.remote.IsValid()
+	fullOut := out
+
+	if useRelay {
+		if len(out) < header.Len {
+			out = out[:header.Len]
+		}
+		out = out[header.Len:]
+	}
+
+	if noiseutil.EncryptLockNeeded {
+		ci.writeLock.Lock()
+	}
+	c := ci.messageCounter.Add(1)
+
+	out = header.Encode(out, header.Version, t, st, hostinfo.remoteIndexId, c)
+	f.connectionManager.Out(hostinfo)
+
+	if t != header.CloseTunnel && hostinfo.lastRebindCount != f.rebindCount {
+		f.lightHouse.QueryServer(hostinfo.vpnAddrs[0])
+		hostinfo.lastRebindCount = f.rebindCount
+		if f.l.Level >= logrus.DebugLevel {
+			f.l.WithField("vpnAddrs", hostinfo.vpnAddrs).Debug("Lighthouse update triggered for punch due to rebind counter")
+		}
+	}
+
+	var err error
+	out, err = ci.eKey.EncryptDanger(out, out, p, c, nb)
+	if noiseutil.EncryptLockNeeded {
+		ci.writeLock.Unlock()
+	}
+	if err != nil {
+		hostinfo.logger(f.l).WithError(err).
+			WithField("udpAddr", remote).WithField("counter", c).
+			WithField("attemptedCounter", c).
+			Error("Failed to encrypt outgoing packet")
+		return
+	}
+
+	// Queue the packet for batched sending
+	var addr netip.AddrPort
+	if remote.IsValid() {
+		addr = remote
+	} else if hostinfo.remote.IsValid() {
+		addr = hostinfo.remote
+	} else {
+		// Relay path - send immediately, not batched
+		for _, relayIP := range hostinfo.relayState.CopyRelayIps() {
+			relayHostInfo, relay, err := f.hostMap.QueryVpnAddrsRelayFor(hostinfo.vpnAddrs, relayIP)
+			if err != nil {
+				hostinfo.relayState.DeleteRelay(relayIP)
+				hostinfo.logger(f.l).WithField("relay", relayIP).WithError(err).Info("sendNoMetricsBatched failed to find HostInfo")
+				continue
+			}
+			f.SendVia(relayHostInfo, relay, out, nb, fullOut[:header.Len+len(out)], true)
+			break
+		}
+		return
+	}
+
+	// Copy the payload since the buffer will be reused
+	payload := make([]byte, len(out))
+	copy(payload, out)
+	*pendingPackets = append(*pendingPackets, udp.BatchPacket{Payload: payload, Addr: addr})
+}

interface.go

@@ -48,6 +48,8 @@ type InterfaceConfig struct {
 
 	ConntrackCacheTimeout time.Duration
 	l                     *logrus.Logger
+
+	tunBatchSize int // batch size for TUN read/write batching, 0 to disable
 }
 
 type Interface struct {
@@ -86,8 +88,9 @@ type Interface struct {
 
 	conntrackCacheTimeout time.Duration
 
-	writers []udp.Conn
+	writers      []udp.Conn
-	readers []io.ReadWriteCloser
+	readers      []io.ReadWriteCloser
+	tunBatchSize int // batch size for TUN read/write batching
 
 	metricHandshakes    metrics.Histogram
 	messageMetrics      *MessageMetrics
@@ -187,6 +190,7 @@ func NewInterface(ctx context.Context, c *InterfaceConfig) (*Interface, error) {
 		relayManager:          c.relayManager,
 		connectionManager:     c.connectionManager,
 		conntrackCacheTimeout: c.ConntrackCacheTimeout,
+		tunBatchSize:          c.tunBatchSize,
 
 		metricHandshakes: metrics.GetOrRegisterHistogram("handshakes", nil, metrics.NewExpDecaySample(1028, 0.015)),
 		messageMetrics:   c.MessageMetrics,
@@ -244,6 +248,15 @@ func (f *Interface) activate() {
 		f.readers[i] = reader
 	}
 
+	// Enable batch reading on all readers if batch size > 1
+	if f.tunBatchSize > 1 {
+		for i := 0; i < f.routines; i++ {
+			if err := overlay.EnableBatchReading(f.readers[i]); err != nil {
+				f.l.WithError(err).WithField("routine", i).Warn("Failed to enable batch reading, falling back to single reads")
+			}
+		}
+	}
+
 	if err := f.inside.Activate(); err != nil {
 		f.inside.Close()
 		f.l.Fatal(err)
@@ -287,13 +300,21 @@ func (f *Interface) listenOut(i int) {
 func (f *Interface) listenIn(reader io.ReadWriteCloser, i int) {
 	runtime.LockOSThread()
 
+	conntrackCache := firewall.NewConntrackCacheTicker(f.conntrackCacheTimeout)
+
+	// Check if batch reading is available and enabled
+	batchReader := overlay.AsBatchReader(reader)
+	if batchReader != nil && f.tunBatchSize > 1 {
+		f.listenInBatched(reader, batchReader, i, conntrackCache)
+		return
+	}
+
+	// Fallback to single-packet reading
 	packet := make([]byte, mtu)
 	out := make([]byte, mtu)
 	fwPacket := &firewall.Packet{}
 	nb := make([]byte, 12, 12)
 
-	conntrackCache := firewall.NewConntrackCacheTicker(f.conntrackCacheTimeout)
-
 	for {
 		n, err := reader.Read(packet)
 		if err != nil {
@@ -310,6 +331,54 @@ func (f *Interface) listenIn(reader io.ReadWriteCloser, i int) {
 	}
 }
 
+func (f *Interface) listenInBatched(reader io.ReadWriteCloser, batchReader overlay.BatchReader, i int, conntrackCache *firewall.ConntrackCacheTicker) {
+	batchSize := f.tunBatchSize
+
+	// Pre-allocate buffers for batch reading
+	packets := make([][]byte, batchSize)
+	for j := range packets {
+		packets[j] = make([]byte, mtu)
+	}
+	sizes := make([]int, batchSize)
+
+	// Pre-allocate buffers for packet processing
+	out := make([]byte, mtu)
+	fwPacket := &firewall.Packet{}
+	nb := make([]byte, 12, 12)
+
+	// Pre-allocate buffer for batched UDP writes
+	pendingPackets := make([]udp.BatchPacket, 0, batchSize)
+
+	for {
+		// Read a batch of packets from TUN
+		n, err := batchReader.ReadBatch(packets, sizes)
+		if err != nil {
+			if errors.Is(err, os.ErrClosed) && f.closed.Load() {
+				return
+			}
+
+			f.l.WithError(err).Error("Error while reading outbound packets")
+			os.Exit(2)
+		}
+
+		if n == 0 {
+			continue
+		}
+
+		// Process all packets in the batch
+		cache := conntrackCache.Get(f.l)
+		for j := 0; j < n; j++ {
+			f.consumeInsidePacketBatched(packets[j][:sizes[j]], fwPacket, nb, out, i, cache, &pendingPackets)
+		}
+
+		// Flush all pending UDP writes
+		if len(pendingPackets) > 0 {
+			f.writers[i].WriteBatch(pendingPackets)
+			pendingPackets = pendingPackets[:0]
+		}
+	}
+}
+
 func (f *Interface) RegisterConfigChangeCallbacks(c *config.C) {
 	c.RegisterReloadCallback(f.reloadFirewall)
 	c.RegisterReloadCallback(f.reloadSendRecvError)
@@ -490,6 +559,14 @@ func (f *Interface) Close() error {
 			f.l.WithError(err).Error("Error while closing udp socket")
 		}
 	}
+	for i, r := range f.readers {
+		if i == 0 {
+			continue // f.readers[0] is f.inside, which we want to save for last
+		}
+		if err := r.Close(); err != nil {
+			f.l.WithError(err).Error("Error while closing tun reader")
+		}
+	}
 
 	// Release the tun device
 	return f.inside.Close()

lighthouse_test.go

@@ -1,7 +1,6 @@
 package nebula
 
 import (
-	"context"
 	"encoding/binary"
 	"fmt"
 	"net/netip"
@@ -42,14 +41,14 @@ func Test_lhStaticMapping(t *testing.T) {
 	c := config.NewC(l)
 	c.Settings["lighthouse"] = map[string]any{"hosts": []any{lh1}}
 	c.Settings["static_host_map"] = map[string]any{lh1: []any{"1.1.1.1:4242"}}
-	_, err := NewLightHouseFromConfig(context.Background(), l, c, cs, nil, nil)
+	_, err := NewLightHouseFromConfig(t.Context(), l, c, cs, nil, nil)
 	require.NoError(t, err)
 
 	lh2 := "10.128.0.3"
 	c = config.NewC(l)
 	c.Settings["lighthouse"] = map[string]any{"hosts": []any{lh1, lh2}}
 	c.Settings["static_host_map"] = map[string]any{lh1: []any{"100.1.1.1:4242"}}
-	_, err = NewLightHouseFromConfig(context.Background(), l, c, cs, nil, nil)
+	_, err = NewLightHouseFromConfig(t.Context(), l, c, cs, nil, nil)
 	require.EqualError(t, err, "lighthouse 10.128.0.3 does not have a static_host_map entry")
 }
 
@@ -71,7 +70,7 @@ func TestReloadLighthouseInterval(t *testing.T) {
 	}
 
 	c.Settings["static_host_map"] = map[string]any{lh1: []any{"1.1.1.1:4242"}}
-	lh, err := NewLightHouseFromConfig(context.Background(), l, c, cs, nil, nil)
+	lh, err := NewLightHouseFromConfig(t.Context(), l, c, cs, nil, nil)
 	require.NoError(t, err)
 	lh.ifce = &mockEncWriter{}
 
@@ -99,7 +98,7 @@ func BenchmarkLighthouseHandleRequest(b *testing.B) {
 	}
 
 	c := config.NewC(l)
-	lh, err := NewLightHouseFromConfig(context.Background(), l, c, cs, nil, nil)
+	lh, err := NewLightHouseFromConfig(b.Context(), l, c, cs, nil, nil)
 	require.NoError(b, err)
 
 	hAddr := netip.MustParseAddrPort("4.5.6.7:12345")
@@ -202,7 +201,7 @@ func TestLighthouse_Memory(t *testing.T) {
 		myVpnNetworks:      []netip.Prefix{myVpnNet},
 		myVpnNetworksTable: nt,
 	}
-	lh, err := NewLightHouseFromConfig(context.Background(), l, c, cs, nil, nil)
+	lh, err := NewLightHouseFromConfig(t.Context(), l, c, cs, nil, nil)
 	lh.ifce = &mockEncWriter{}
 	require.NoError(t, err)
 	lhh := lh.NewRequestHandler()
@@ -288,7 +287,7 @@ func TestLighthouse_reload(t *testing.T) {
 		myVpnNetworksTable: nt,
 	}
 
-	lh, err := NewLightHouseFromConfig(context.Background(), l, c, cs, nil, nil)
+	lh, err := NewLightHouseFromConfig(t.Context(), l, c, cs, nil, nil)
 	require.NoError(t, err)
 
 	nc := map[string]any{
@@ -523,7 +522,7 @@ func TestLighthouse_Dont_Delete_Static_Hosts(t *testing.T) {
 		myVpnNetworks:      []netip.Prefix{myVpnNet},
 		myVpnNetworksTable: nt,
 	}
-	lh, err := NewLightHouseFromConfig(context.Background(), l, c, cs, nil, nil)
+	lh, err := NewLightHouseFromConfig(t.Context(), l, c, cs, nil, nil)
 	require.NoError(t, err)
 	lh.ifce = &mockEncWriter{}
 
@@ -589,7 +588,7 @@ func TestLighthouse_DeletesWork(t *testing.T) {
 		myVpnNetworks:      []netip.Prefix{myVpnNet},
 		myVpnNetworksTable: nt,
 	}
-	lh, err := NewLightHouseFromConfig(context.Background(), l, c, cs, nil, nil)
+	lh, err := NewLightHouseFromConfig(t.Context(), l, c, cs, nil, nil)
 	require.NoError(t, err)
 	lh.ifce = &mockEncWriter{}
 

main.go

@@ -250,6 +250,7 @@ func Main(c *config.C, configTest bool, buildVersion string, logger *logrus.Logg
 		punchy:                punchy,
 		ConntrackCacheTimeout: conntrackCacheTimeout,
 		l:                     l,
+		tunBatchSize:          c.GetInt("listen.batch", 64),
 	}
 
 	var ifce *Interface

overlay/device.go

@@ -16,3 +16,38 @@ type Device interface {
 	SupportsMultiqueue() bool
 	NewMultiQueueReader() (io.ReadWriteCloser, error)
 }
+
+// BatchReader is an optional interface that devices can implement
+// to support reading multiple packets in a single batch operation.
+// This can significantly reduce syscall overhead under high load.
+type BatchReader interface {
+	// ReadBatch reads up to len(packets) packets into the provided buffers.
+	// Each packet is read into packets[i] and its length is stored in sizes[i].
+	// Returns the number of packets read, or an error.
+	// A return of (0, nil) indicates no packets were available (non-blocking).
+	ReadBatch(packets [][]byte, sizes []int) (int, error)
+}
+
+// AsBatchReader returns a BatchReader if the reader supports batch operations,
+// otherwise returns nil.
+func AsBatchReader(r io.ReadWriteCloser) BatchReader {
+	if br, ok := r.(BatchReader); ok {
+		return br
+	}
+	return nil
+}
+
+// BatchEnabler is an optional interface for devices that need explicit
+// enabling of batch read support (e.g., setting non-blocking mode).
+type BatchEnabler interface {
+	EnableBatchReading() error
+}
+
+// EnableBatchReading enables batch reading on the device if supported.
+// Returns nil if the device doesn't support or need explicit enabling.
+func EnableBatchReading(d interface{}) error {
+	if be, ok := d.(BatchEnabler); ok {
+		return be.EnableBatchReading()
+	}
+	return nil
+}

overlay/tun.go

@@ -12,6 +12,15 @@ import (
 
 const DefaultMTU = 1300
 
+type NameError struct {
+	Name       string
+	Underlying error
+}
+
+func (e *NameError) Error() string {
+	return fmt.Sprintf("could not set tun device name: %s because %s", e.Name, e.Underlying)
+}
+
 // TODO: We may be able to remove routines
 type DeviceFactory func(c *config.C, l *logrus.Logger, vpnNetworks []netip.Prefix, routines int) (Device, error)
 

overlay/tun_freebsd.go

@@ -266,7 +266,7 @@ func newTun(c *config.C, l *logrus.Logger, vpnNetworks []netip.Prefix, _ bool) (
 		}
 
 		// Set the device name
-		ioctl(fd, syscall.SIOCSIFNAME, uintptr(unsafe.Pointer(&ifrr)))
+		_ = ioctl(fd, syscall.SIOCSIFNAME, uintptr(unsafe.Pointer(&ifrr)))
 	}
 
 	t := &tun{

overlay/tun_linux.go

@@ -24,6 +24,11 @@ import (
 	"golang.org/x/sys/unix"
 )
 
+const (
+	// virtioNetHdrLen is the length of virtio_net_hdr (without mergeable buffers)
+	virtioNetHdrLen = 10
+)
+
 type tun struct {
 	io.ReadWriteCloser
 	fd          int
@@ -34,6 +39,13 @@ type tun struct {
 	TXQueueLen  int
 	deviceIndex int
 	ioctlFd     uintptr
+	nonBlocking bool // true if fd is in non-blocking mode
+	vnetHdr     bool // true if IFF_VNET_HDR is enabled on the TUN device
+
+	// readBuf is used when vnetHdr is enabled to read the full packet+header
+	// before stripping the header. This is needed because caller-provided
+	// buffers are sized for MTU but kernel writes MTU+10 with virtio header.
+	readBuf []byte
 
 	Routes                    atomic.Pointer[[]Route]
 	routeTree                 atomic.Pointer[bart.Table[routing.Gateways]]
@@ -53,6 +65,23 @@ func (t *tun) Networks() []netip.Prefix {
 	return t.vpnNetworks
 }
 
+// tunVnetHdrSupported checks if the kernel supports IFF_VNET_HDR on TUN devices
+func tunVnetHdrSupported() bool {
+	fd, err := unix.Open("/dev/net/tun", unix.O_RDONLY, 0)
+	if err != nil {
+		return false
+	}
+	defer unix.Close(fd)
+
+	var features uint32
+	err = ioctl(uintptr(fd), uintptr(unix.TUNGETFEATURES), uintptr(unsafe.Pointer(&features)))
+	if err != nil {
+		return false
+	}
+
+	return features&unix.IFF_VNET_HDR != 0
+}
+
 type ifReq struct {
 	Name  [16]byte
 	Flags uint16
@@ -107,17 +136,35 @@ func newTun(c *config.C, l *logrus.Logger, vpnNetworks []netip.Prefix, multiqueu
 		}
 	}
 
+	// Check if VNET_HDR is supported before trying to use it
+	useVnetHdr := tunVnetHdrSupported()
+
 	var req ifReq
 	req.Flags = uint16(unix.IFF_TUN | unix.IFF_NO_PI)
 	if multiqueue {
 		req.Flags |= unix.IFF_MULTI_QUEUE
 	}
-	copy(req.Name[:], c.GetString("tun.dev", ""))
+	if useVnetHdr {
+		req.Flags |= unix.IFF_VNET_HDR
+	}
+
+	nameStr := c.GetString("tun.dev", "")
+	copy(req.Name[:], nameStr)
 	if err = ioctl(uintptr(fd), uintptr(unix.TUNSETIFF), uintptr(unsafe.Pointer(&req))); err != nil {
-		return nil, err
+		return nil, &NameError{
+			Name:       nameStr,
+			Underlying: err,
+		}
 	}
 	name := strings.Trim(string(req.Name[:]), "\x00")
 
+	// Track if VNET_HDR is in use
+	// Note: We don't call TUNSETOFFLOAD - just handle the headers manually
+	vnetHdrEnabled := useVnetHdr
+	if vnetHdrEnabled {
+		l.Info("TUN VNET_HDR enabled")
+	}
+
 	file := os.NewFile(uintptr(fd), "/dev/net/tun")
 	t, err := newTunGeneric(c, l, file, vpnNetworks)
 	if err != nil {
@@ -125,6 +172,13 @@ func newTun(c *config.C, l *logrus.Logger, vpnNetworks []netip.Prefix, multiqueu
 	}
 
 	t.Device = name
+	t.vnetHdr = vnetHdrEnabled
+
+	// Allocate read buffer for virtio header handling
+	// Buffer needs to be large enough for virtio header + max packet
+	if t.vnetHdr {
+		t.readBuf = make([]byte, t.MaxMTU+virtioNetHdrLen)
+	}
 
 	return t, nil
 }
@@ -235,21 +289,172 @@ func (t *tun) SupportsMultiqueue() bool {
 }
 
 func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
-	fd, err := unix.Open("/dev/net/tun", os.O_RDWR, 0)
+	fd, err := unix.Open("/dev/net/tun", os.O_RDWR|unix.O_NONBLOCK, 0)
 	if err != nil {
 		return nil, err
 	}
 
 	var req ifReq
 	req.Flags = uint16(unix.IFF_TUN | unix.IFF_NO_PI | unix.IFF_MULTI_QUEUE)
+	if t.vnetHdr {
+		req.Flags |= unix.IFF_VNET_HDR
+	}
 	copy(req.Name[:], t.Device)
 	if err = ioctl(uintptr(fd), uintptr(unix.TUNSETIFF), uintptr(unsafe.Pointer(&req))); err != nil {
 		return nil, err
 	}
 
-	file := os.NewFile(uintptr(fd), "/dev/net/tun")
+	reader := &tunBatchReader{fd: fd, device: t.Device, vnetHdr: t.vnetHdr}
+	if t.vnetHdr {
+		reader.readBuf = make([]byte, t.MaxMTU+virtioNetHdrLen)
+	}
+	return reader, nil
+}
 
-	return file, nil
+// tunBatchReader implements BatchReader for efficient batch packet reading
+type tunBatchReader struct {
+	fd      int
+	device  string
+	vnetHdr bool
+	readBuf []byte // internal buffer for virtio header handling
+}
+
+func (r *tunBatchReader) Read(b []byte) (int, error) {
+	// Choose buffer: use internal buffer for vnetHdr, caller's buffer otherwise
+	readBuf := b
+	if r.vnetHdr {
+		readBuf = r.readBuf
+	}
+
+	// Use poll to wait for data, then read
+	for {
+		n, err := unix.Read(r.fd, readBuf)
+		if err == nil {
+			if r.vnetHdr && n > virtioNetHdrLen {
+				packetLen := n - virtioNetHdrLen
+				copy(b, readBuf[virtioNetHdrLen:n])
+				return packetLen, nil
+			}
+			if r.vnetHdr {
+				return 0, nil // No packet data
+			}
+			return n, nil
+		}
+		if err == unix.EAGAIN || err == unix.EWOULDBLOCK {
+			// Wait for data
+			pfds := []unix.PollFd{{Fd: int32(r.fd), Events: unix.POLLIN}}
+			_, err = unix.Poll(pfds, -1)
+			if err != nil {
+				if err == unix.EINTR {
+					continue
+				}
+				return 0, err
+			}
+			continue
+		}
+		return n, err
+	}
+}
+
+func (r *tunBatchReader) Write(b []byte) (int, error) {
+	if !r.vnetHdr {
+		return unix.Write(r.fd, b)
+	}
+
+	// Use writev to prepend virtio header without copying the packet data
+	// Header is all zeros = no GSO, no checksum offload
+	var hdr [virtioNetHdrLen]byte
+	bufs := [][]byte{hdr[:], b}
+
+	n, err := unix.Writev(r.fd, bufs)
+	if err != nil {
+		return 0, err
+	}
+	// Return only the packet bytes written (exclude header)
+	if n > virtioNetHdrLen {
+		return n - virtioNetHdrLen, nil
+	}
+	return 0, nil
+}
+
+func (r *tunBatchReader) Close() error {
+	return unix.Close(r.fd)
+}
+
+// ReadBatch reads up to len(packets) packets from the TUN device.
+// It drains all available packets without blocking, using poll() only
+// when no packets have been read yet.
+func (r *tunBatchReader) ReadBatch(packets [][]byte, sizes []int) (int, error) {
+	count := 0
+	maxPackets := len(packets)
+	if len(sizes) < maxPackets {
+		maxPackets = len(sizes)
+	}
+
+	// Choose read buffer based on vnetHdr
+	readBuf := packets[0] // Will be updated in loop for non-vnetHdr
+	if r.vnetHdr {
+		readBuf = r.readBuf
+	}
+
+	for count < maxPackets {
+		if !r.vnetHdr {
+			readBuf = packets[count]
+		}
+
+		n, err := unix.Read(r.fd, readBuf)
+		if err == nil && n > 0 {
+			if r.vnetHdr {
+				if n > virtioNetHdrLen {
+					packetLen := n - virtioNetHdrLen
+					copy(packets[count], readBuf[virtioNetHdrLen:n])
+					sizes[count] = packetLen
+				} else {
+					// Malformed packet (no data after header), skip
+					continue
+				}
+			} else {
+				sizes[count] = n
+			}
+			count++
+			continue
+		}
+
+		if err == unix.EAGAIN || err == unix.EWOULDBLOCK {
+			// No more packets available
+			if count > 0 {
+				// We have some packets, return them
+				return count, nil
+			}
+			// No packets yet, wait for at least one
+			pfds := []unix.PollFd{{Fd: int32(r.fd), Events: unix.POLLIN}}
+			_, err = unix.Poll(pfds, -1)
+			if err != nil {
+				if err == unix.EINTR {
+					continue
+				}
+				return 0, err
+			}
+			continue
+		}
+
+		if err != nil {
+			if count > 0 {
+				// Return what we have
+				return count, nil
+			}
+			return 0, err
+		}
+
+		if n == 0 {
+			if count > 0 {
+				return count, nil
+			}
+			return 0, io.EOF
+		}
+	}
+
+	return count, nil
 }
 
 func (t *tun) RoutesFor(ip netip.Addr) routing.Gateways {
@@ -258,6 +463,27 @@ func (t *tun) RoutesFor(ip netip.Addr) routing.Gateways {
 }
 
 func (t *tun) Write(b []byte) (int, error) {
+	if !t.vnetHdr {
+		return t.writeSimple(b)
+	}
+
+	// Use writev to prepend virtio header without copying the packet data
+	// Header is all zeros = no GSO, no checksum offload
+	var hdr [virtioNetHdrLen]byte
+	bufs := [][]byte{hdr[:], b}
+
+	n, err := unix.Writev(t.fd, bufs)
+	if err != nil {
+		return 0, err
+	}
+	// Return only the packet bytes written (exclude header)
+	if n > virtioNetHdrLen {
+		return n - virtioNetHdrLen, nil
+	}
+	return 0, nil
+}
+
+func (t *tun) writeSimple(b []byte) (int, error) {
 	var nn int
 	maximum := len(b)
 
@@ -280,6 +506,177 @@ func (t *tun) Write(b []byte) (int, error) {
 	}
 }
 
+// EnableBatchReading sets the TUN fd to non-blocking mode to enable batch reading.
+// This should be called before using ReadBatch.
+func (t *tun) EnableBatchReading() error {
+	if t.nonBlocking {
+		return nil
+	}
+	err := unix.SetNonblock(t.fd, true)
+	if err != nil {
+		return err
+	}
+	t.nonBlocking = true
+	return nil
+}
+
+// Read overrides the default Read to handle non-blocking mode and virtio headers
+func (t *tun) Read(b []byte) (int, error) {
+	if !t.vnetHdr {
+		return t.readSimple(b)
+	}
+
+	// With VNET_HDR, read into internal buffer (which has space for header)
+	// then copy packet data to caller's buffer
+	if !t.nonBlocking {
+		n, err := t.ReadWriteCloser.Read(t.readBuf)
+		if err != nil {
+			return 0, err
+		}
+		if n <= virtioNetHdrLen {
+			return 0, nil // No packet data
+		}
+		packetLen := n - virtioNetHdrLen
+		copy(b, t.readBuf[virtioNetHdrLen:n])
+		return packetLen, nil
+	}
+
+	// Non-blocking read with poll
+	for {
+		n, err := unix.Read(t.fd, t.readBuf)
+		if err == nil {
+			if n <= virtioNetHdrLen {
+				return 0, nil // No packet data
+			}
+			packetLen := n - virtioNetHdrLen
+			copy(b, t.readBuf[virtioNetHdrLen:n])
+			return packetLen, nil
+		}
+		if err == unix.EAGAIN || err == unix.EWOULDBLOCK {
+			pfds := []unix.PollFd{{Fd: int32(t.fd), Events: unix.POLLIN}}
+			_, err = unix.Poll(pfds, -1)
+			if err != nil {
+				if err == unix.EINTR {
+					continue
+				}
+				return 0, err
+			}
+			continue
+		}
+		return n, err
+	}
+}
+
+func (t *tun) readSimple(b []byte) (int, error) {
+	if !t.nonBlocking {
+		return t.ReadWriteCloser.Read(b)
+	}
+
+	for {
+		n, err := unix.Read(t.fd, b)
+		if err == nil {
+			return n, nil
+		}
+		if err == unix.EAGAIN || err == unix.EWOULDBLOCK {
+			pfds := []unix.PollFd{{Fd: int32(t.fd), Events: unix.POLLIN}}
+			_, err = unix.Poll(pfds, -1)
+			if err != nil {
+				if err == unix.EINTR {
+					continue
+				}
+				return 0, err
+			}
+			continue
+		}
+		return n, err
+	}
+}
+
+// ReadBatch reads up to len(packets) packets from the TUN device.
+// EnableBatchReading must be called first.
+func (t *tun) ReadBatch(packets [][]byte, sizes []int) (int, error) {
+	if !t.nonBlocking {
+		// Fallback to single read if non-blocking not enabled
+		n, err := t.Read(packets[0])
+		if err != nil {
+			return 0, err
+		}
+		sizes[0] = n
+		return 1, nil
+	}
+
+	count := 0
+	maxPackets := len(packets)
+	if len(sizes) < maxPackets {
+		maxPackets = len(sizes)
+	}
+
+	// Choose read buffer based on vnetHdr
+	// With vnetHdr, we need to read into internal buffer (has space for header)
+	// then copy packet data to caller's buffer
+	readBuf := packets[0] // Will be updated in the loop
+	if t.vnetHdr {
+		readBuf = t.readBuf
+	}
+
+	for count < maxPackets {
+		if !t.vnetHdr {
+			readBuf = packets[count]
+		}
+
+		n, err := unix.Read(t.fd, readBuf)
+		if err == nil && n > 0 {
+			if t.vnetHdr {
+				if n > virtioNetHdrLen {
+					packetLen := n - virtioNetHdrLen
+					copy(packets[count], readBuf[virtioNetHdrLen:n])
+					sizes[count] = packetLen
+				} else {
+					// Malformed packet (no data after header), skip
+					continue
+				}
+			} else {
+				sizes[count] = n
+			}
+			count++
+			continue
+		}
+
+		if err == unix.EAGAIN || err == unix.EWOULDBLOCK {
+			// No more packets available
+			if count > 0 {
+				return count, nil
+			}
+			// No packets yet, wait for at least one
+			pfds := []unix.PollFd{{Fd: int32(t.fd), Events: unix.POLLIN}}
+			_, err = unix.Poll(pfds, -1)
+			if err != nil {
+				if err == unix.EINTR {
+					continue
+				}
+				return 0, err
+			}
+			continue
+		}
+
+		if err != nil {
+			if count > 0 {
+				return count, nil
+			}
+			return 0, err
+		}
+
+		if n == 0 {
+			if count > 0 {
+				return count, nil
+			}
+			return 0, io.EOF
+		}
+	}
+
+	return count, nil
+}
+
 func (t *tun) deviceBytes() (o [16]byte) {
 	for i, c := range t.Device {
 		o[i] = byte(c)
@@ -713,6 +1110,7 @@ func (t *tun) Close() error {
 
 	if t.ioctlFd > 0 {
 		_ = os.NewFile(t.ioctlFd, "ioctlFd").Close()
+		t.ioctlFd = 0
 	}
 
 	return nil

overlay/tun_windows.go

@@ -74,7 +74,10 @@ func newTun(c *config.C, l *logrus.Logger, vpnNetworks []netip.Prefix, _ bool) (
 		l.WithError(err).Debug("Failed to create wintun device, retrying")
 		tunDevice, err = wintun.CreateTUNWithRequestedGUID(deviceName, guid, t.MTU)
 		if err != nil {
-			return nil, fmt.Errorf("create TUN device failed: %w", err)
+			return nil, &NameError{
+				Name:       deviceName,
+				Underlying: fmt.Errorf("create TUN device failed: %w", err),
+			}
 		}
 	}
 	t.tun = tunDevice.(*wintun.NativeTun)

udp/conn.go

@@ -13,13 +13,22 @@ type EncReader func(
 	payload []byte,
 )
 
+// BatchPacket represents a single packet in a batch write operation
+type BatchPacket struct {
+	Payload []byte
+	Addr    netip.AddrPort
+}
+
 type Conn interface {
 	Rebind() error
 	LocalAddr() (netip.AddrPort, error)
 	ListenOut(r EncReader)
 	WriteTo(b []byte, addr netip.AddrPort) error
+	WriteBatch(pkts []BatchPacket) (int, error)
 	ReloadConfig(c *config.C)
 	SupportsMultipleReaders() bool
+	SupportsGSO() bool
+	SupportsGRO() bool
 	Close() error
 }
 
@@ -37,9 +46,18 @@ func (NoopConn) ListenOut(_ EncReader) {
 func (NoopConn) SupportsMultipleReaders() bool {
 	return false
 }
+func (NoopConn) SupportsGSO() bool {
+	return false
+}
+func (NoopConn) SupportsGRO() bool {
+	return false
+}
 func (NoopConn) WriteTo(_ []byte, _ netip.AddrPort) error {
 	return nil
 }
+func (NoopConn) WriteBatch(pkts []BatchPacket) (int, error) {
+	return len(pkts), nil
+}
 func (NoopConn) ReloadConfig(_ *config.C) {
 	return
 }

udp/udp_darwin.go

@@ -188,6 +188,14 @@ func (u *StdConn) SupportsMultipleReaders() bool {
 	return false
 }
 
+func (u *StdConn) SupportsGSO() bool {
+	return false
+}
+
+func (u *StdConn) SupportsGRO() bool {
+	return false
+}
+
 func (u *StdConn) Rebind() error {
 	var err error
 	if u.isV4 {
@@ -202,3 +210,12 @@ func (u *StdConn) Rebind() error {
 
 	return nil
 }
+
+func (u *StdConn) WriteBatch(pkts []BatchPacket) (int, error) {
+	for i := range pkts {
+		if err := u.WriteTo(pkts[i].Payload, pkts[i].Addr); err != nil {
+			return i, err
+		}
+	}
+	return len(pkts), nil
+}

udp/udp_generic.go

@@ -101,3 +101,20 @@ func (u *GenericConn) ListenOut(r EncReader) {
 func (u *GenericConn) SupportsMultipleReaders() bool {
 	return false
 }
+
+func (u *GenericConn) SupportsGSO() bool {
+	return false
+}
+
+func (u *GenericConn) SupportsGRO() bool {
+	return false
+}
+
+func (u *GenericConn) WriteBatch(pkts []BatchPacket) (int, error) {
+	for i := range pkts {
+		if err := u.WriteTo(pkts[i].Payload, pkts[i].Addr); err != nil {
+			return i, err
+		}
+	}
+	return len(pkts), nil
+}

udp/udp_linux.go

@@ -5,6 +5,7 @@ package udp
 
 import (
 	"encoding/binary"
+	"errors"
 	"fmt"
 	"net"
 	"net/netip"
@@ -18,10 +19,12 @@ import (
 )
 
 type StdConn struct {
-	sysFd int
+	sysFd        int
-	isV4  bool
+	isV4         bool
-	l     *logrus.Logger
+	l            *logrus.Logger
-	batch int
+	batch        int
+	gsoSupported bool
+	groSupported bool
 }
 
 func maybeIPV4(ip net.IP) (net.IP, bool) {
@@ -32,6 +35,78 @@ func maybeIPV4(ip net.IP) (net.IP, bool) {
 	return ip, false
 }
 
+// supportsUDPOffload checks if the kernel supports UDP GSO (Generic Segmentation Offload)
+// by attempting to get the UDP_SEGMENT socket option.
+func supportsUDPOffload(fd int) bool {
+	_, err := unix.GetsockoptInt(fd, unix.IPPROTO_UDP, unix.UDP_SEGMENT)
+	return err == nil
+}
+
+// supportsUDPGRO checks if the kernel supports UDP GRO (Generic Receive Offload)
+// and attempts to enable it on the socket.
+func supportsUDPGRO(fd int) bool {
+	// Try to enable UDP_GRO
+	err := unix.SetsockoptInt(fd, unix.IPPROTO_UDP, unix.UDP_GRO, 1)
+	return err == nil
+}
+
+const (
+	// Maximum number of datagrams that can be coalesced with GSO/GRO
+	udpSegmentMaxDatagrams = 64
+
+	// Maximum size of a GRO coalesced packet (64KB is the practical limit)
+	// This is udpSegmentMaxDatagrams * MTU but capped at 65535
+	groMaxPacketSize = 65535
+)
+
+// setGSOSize writes a UDP_SEGMENT control message to the provided buffer
+// with the given segment size. Returns the actual control message length.
+func setGSOSize(control []byte, gsoSize uint16) int {
+	// Build the cmsghdr structure
+	cmsgLen := unix.CmsgLen(2) // 2 bytes for uint16 segment size
+	cmsg := (*unix.Cmsghdr)(unsafe.Pointer(&control[0]))
+	cmsg.Level = unix.IPPROTO_UDP
+	cmsg.Type = unix.UDP_SEGMENT
+	cmsg.SetLen(cmsgLen)
+
+	// Write the segment size after the header (after cmsghdr)
+	binary.NativeEndian.PutUint16(control[unix.SizeofCmsghdr:], gsoSize)
+
+	return unix.CmsgSpace(2) // aligned size
+}
+
+// getGROSize parses a control message buffer to extract the UDP_GRO segment size.
+// Returns 0 if no GRO control message is present (meaning the packet is not coalesced).
+func getGROSize(control []byte, controlLen int) uint16 {
+	if controlLen < unix.SizeofCmsghdr {
+		return 0
+	}
+
+	// Parse control messages
+	for offset := 0; offset < controlLen; {
+		if offset+unix.SizeofCmsghdr > controlLen {
+			break
+		}
+
+		cmsg := (*unix.Cmsghdr)(unsafe.Pointer(&control[offset]))
+		cmsgDataLen := int(cmsg.Len) - unix.SizeofCmsghdr
+		if cmsgDataLen < 0 {
+			break
+		}
+
+		if cmsg.Level == unix.IPPROTO_UDP && cmsg.Type == unix.UDP_GRO {
+			if cmsgDataLen >= 2 {
+				return binary.NativeEndian.Uint16(control[offset+unix.SizeofCmsghdr:])
+			}
+		}
+
+		// Move to next control message (aligned)
+		offset += unix.CmsgSpace(cmsgDataLen)
+	}
+
+	return 0
+}
+
 func NewListener(l *logrus.Logger, ip netip.Addr, port int, multi bool, batch int) (Conn, error) {
 	af := unix.AF_INET6
 	if ip.Is4() {
@@ -69,13 +144,31 @@ func NewListener(l *logrus.Logger, ip netip.Addr, port int, multi bool, batch in
 		return nil, fmt.Errorf("unable to bind to socket: %s", err)
 	}
 
-	return &StdConn{sysFd: fd, isV4: ip.Is4(), l: l, batch: batch}, err
+	gsoSupported := supportsUDPOffload(fd)
+	if gsoSupported {
+		l.Info("UDP GSO offload is supported")
+	}
+
+	groSupported := supportsUDPGRO(fd)
+	if groSupported {
+		l.Info("UDP GRO offload is supported and enabled")
+	}
+
+	return &StdConn{sysFd: fd, isV4: ip.Is4(), l: l, batch: batch, gsoSupported: gsoSupported, groSupported: groSupported}, err
 }
 
 func (u *StdConn) SupportsMultipleReaders() bool {
 	return true
 }
 
+func (u *StdConn) SupportsGSO() bool {
+	return u.gsoSupported
+}
+
+func (u *StdConn) SupportsGRO() bool {
+	return u.groSupported
+}
+
 func (u *StdConn) Rebind() error {
 	return nil
 }
@@ -125,13 +218,27 @@ func (u *StdConn) LocalAddr() (netip.AddrPort, error) {
 func (u *StdConn) ListenOut(r EncReader) {
 	var ip netip.Addr
 
-	msgs, buffers, names := u.PrepareRawMessages(u.batch)
+	msgs, buffers, names, controls := u.PrepareRawMessages(u.batch)
 	read := u.ReadMulti
 	if u.batch == 1 {
 		read = u.ReadSingle
 	}
 
+	// Store the original control buffer size for resetting after each read
+	controlLen := 0
+	if u.groSupported && len(controls) > 0 && len(controls[0]) > 0 {
+		controlLen = len(controls[0])
+	}
+
 	for {
+		// Reset Controllen before each read - the kernel updates this field
+		// after recvmsg to indicate actual received control data length
+		if controlLen > 0 {
+			for i := range msgs {
+				setMsghdrControllen(&msgs[i].Hdr, controlLen)
+			}
+		}
+
 		n, err := read(msgs)
 		if err != nil {
 			u.l.WithError(err).Debug("udp socket is closed, exiting read loop")
@@ -139,13 +246,36 @@ func (u *StdConn) ListenOut(r EncReader) {
 		}
 
 		for i := 0; i < n; i++ {
-			// Its ok to skip the ok check here, the slicing is the only error that can occur and it will panic
+			// Extract source address
 			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])
+			srcAddr := netip.AddrPortFrom(ip.Unmap(), binary.BigEndian.Uint16(names[i][2:4]))
+
+			// Check for GRO coalesced packet
+			totalLen := int(msgs[i].Len)
+			segmentSize := uint16(0)
+			if controlLen > 0 {
+				segmentSize = getGROSize(controls[i], getMsghdrControllen(&msgs[i].Hdr))
+			}
+
+			if segmentSize > 0 && totalLen > int(segmentSize) {
+				// This is a GRO coalesced packet - split it into individual datagrams
+				for offset := 0; offset < totalLen; {
+					packetLen := int(segmentSize)
+					if offset+packetLen > totalLen {
+						// Last packet may be smaller
+						packetLen = totalLen - offset
+					}
+					r(srcAddr, buffers[i][offset:offset+packetLen])
+					offset += packetLen
+				}
+			} else {
+				// Single packet, no coalescing
+				r(srcAddr, buffers[i][:totalLen])
+			}
 		}
 	}
 }
@@ -313,6 +443,226 @@ func (u *StdConn) Close() error {
 	return syscall.Close(u.sysFd)
 }
 
+func (u *StdConn) WriteBatch(pkts []BatchPacket) (int, error) {
+	if len(pkts) == 0 {
+		return 0, nil
+	}
+
+	// If GSO is supported, try to coalesce packets to the same destination
+	if u.gsoSupported {
+		return u.writeBatchGSO(pkts)
+	}
+
+	return u.writeBatchSendmmsg(pkts)
+}
+
+// writeBatchSendmmsg sends packets using sendmmsg without GSO coalescing
+func (u *StdConn) writeBatchSendmmsg(pkts []BatchPacket) (int, error) {
+	msgs := make([]rawMessage, len(pkts))
+	iovecs := make([]iovec, len(pkts))
+	var names4 []unix.RawSockaddrInet4
+	var names6 []unix.RawSockaddrInet6
+
+	if u.isV4 {
+		names4 = make([]unix.RawSockaddrInet4, len(pkts))
+	} else {
+		names6 = make([]unix.RawSockaddrInet6, len(pkts))
+	}
+
+	for i := range pkts {
+		setIovecBase(&iovecs[i], &pkts[i].Payload[0])
+		setIovecLen(&iovecs[i], len(pkts[i].Payload))
+		msgs[i].Hdr.Iov = &iovecs[i]
+		setMsghdrIovlen(&msgs[i].Hdr, 1)
+
+		if u.isV4 {
+			names4[i].Family = unix.AF_INET
+			names4[i].Addr = pkts[i].Addr.Addr().As4()
+			binary.BigEndian.PutUint16((*[2]byte)(unsafe.Pointer(&names4[i].Port))[:], pkts[i].Addr.Port())
+			msgs[i].Hdr.Name = (*byte)(unsafe.Pointer(&names4[i]))
+			msgs[i].Hdr.Namelen = unix.SizeofSockaddrInet4
+		} else {
+			names6[i].Family = unix.AF_INET6
+			names6[i].Addr = pkts[i].Addr.Addr().As16()
+			binary.BigEndian.PutUint16((*[2]byte)(unsafe.Pointer(&names6[i].Port))[:], pkts[i].Addr.Port())
+			msgs[i].Hdr.Name = (*byte)(unsafe.Pointer(&names6[i]))
+			msgs[i].Hdr.Namelen = unix.SizeofSockaddrInet6
+		}
+	}
+
+	var sent int
+	for sent < len(msgs) {
+		n, _, errno := unix.Syscall6(
+			unix.SYS_SENDMMSG,
+			uintptr(u.sysFd),
+			uintptr(unsafe.Pointer(&msgs[sent])),
+			uintptr(len(msgs)-sent),
+			0,
+			0,
+			0,
+		)
+
+		if errno == unix.EINTR {
+			continue
+		}
+
+		if errno != 0 {
+			return sent, &net.OpError{Op: "sendmmsg", Err: errno}
+		}
+
+		sent += int(n)
+	}
+
+	return sent, nil
+}
+
+// writeBatchGSO sends packets using GSO coalescing when possible.
+// Packets to the same destination with the same size are coalesced into a single
+// GSO message. Mixed destinations or sizes fall back to individual sendmmsg calls.
+func (u *StdConn) writeBatchGSO(pkts []BatchPacket) (int, error) {
+	// Group packets by destination and try to coalesce
+	totalSent := 0
+	i := 0
+
+	for i < len(pkts) {
+		// Find a run of packets to the same destination with compatible sizes
+		startIdx := i
+		dst := pkts[i].Addr
+		segmentSize := len(pkts[i].Payload)
+
+		// Count how many packets we can coalesce (same destination, same size except possibly last)
+		coalescedCount := 1
+		totalSize := segmentSize
+		for i+coalescedCount < len(pkts) && coalescedCount < udpSegmentMaxDatagrams {
+			next := pkts[i+coalescedCount]
+			if next.Addr != dst {
+				break
+			}
+			nextSize := len(next.Payload)
+			// For GSO, all packets except the last must have the same size
+			// The last packet can be smaller (but not larger)
+			if nextSize != segmentSize {
+				// Check if this could be the last packet (smaller is ok)
+				if nextSize < segmentSize && i+coalescedCount == len(pkts)-1 {
+					coalescedCount++
+					totalSize += nextSize
+				}
+				break
+			}
+			coalescedCount++
+			totalSize += nextSize
+		}
+
+		// If we have multiple packets to coalesce, use GSO
+		if coalescedCount > 1 {
+			err := u.sendGSO(pkts[startIdx:startIdx+coalescedCount], dst, segmentSize, totalSize)
+			if err != nil {
+				// If GSO fails (e.g., EIO due to NIC not supporting checksum offload),
+				// disable GSO and fall back to sendmmsg for the rest
+				if isGSOError(err) {
+					u.l.WithError(err).Warn("GSO send failed, disabling GSO for this connection")
+					u.gsoSupported = false
+					// Send remaining packets with sendmmsg
+					remaining, rerr := u.writeBatchSendmmsg(pkts[startIdx:])
+					return totalSent + remaining, rerr
+				}
+				return totalSent, err
+			}
+			totalSent += coalescedCount
+			i += coalescedCount
+		} else {
+			// Single packet, send without GSO overhead
+			err := u.WriteTo(pkts[i].Payload, pkts[i].Addr)
+			if err != nil {
+				return totalSent, err
+			}
+			totalSent++
+			i++
+		}
+	}
+
+	return totalSent, nil
+}
+
+// sendGSO sends coalesced packets using UDP GSO
+func (u *StdConn) sendGSO(pkts []BatchPacket, dst netip.AddrPort, segmentSize, totalSize int) error {
+	// Allocate a buffer large enough for all packet payloads
+	coalescedBuf := make([]byte, totalSize)
+	offset := 0
+	for _, pkt := range pkts {
+		copy(coalescedBuf[offset:], pkt.Payload)
+		offset += len(pkt.Payload)
+	}
+
+	// Prepare control message with GSO segment size
+	control := make([]byte, unix.CmsgSpace(2))
+	controlLen := setGSOSize(control, uint16(segmentSize))
+
+	// Prepare the iovec
+	iov := iovec{}
+	setIovecBase(&iov, &coalescedBuf[0])
+	setIovecLen(&iov, totalSize)
+
+	// Prepare the msghdr
+	var hdr msghdr
+	hdr.Iov = &iov
+	setMsghdrIovlen(&hdr, 1)
+	hdr.Control = &control[0]
+	setMsghdrControllen(&hdr, controlLen)
+
+	// Declare sockaddr at function scope so it remains valid for the syscall
+	// (must not go out of scope before the syscall is made)
+	var rsa4 unix.RawSockaddrInet4
+	var rsa6 unix.RawSockaddrInet6
+
+	// Set destination address
+	if u.isV4 {
+		rsa4.Family = unix.AF_INET
+		rsa4.Addr = dst.Addr().As4()
+		binary.BigEndian.PutUint16((*[2]byte)(unsafe.Pointer(&rsa4.Port))[:], dst.Port())
+		hdr.Name = (*byte)(unsafe.Pointer(&rsa4))
+		hdr.Namelen = unix.SizeofSockaddrInet4
+	} else {
+		rsa6.Family = unix.AF_INET6
+		rsa6.Addr = dst.Addr().As16()
+		binary.BigEndian.PutUint16((*[2]byte)(unsafe.Pointer(&rsa6.Port))[:], dst.Port())
+		hdr.Name = (*byte)(unsafe.Pointer(&rsa6))
+		hdr.Namelen = unix.SizeofSockaddrInet6
+	}
+
+	for {
+		_, _, errno := unix.Syscall6(
+			unix.SYS_SENDMSG,
+			uintptr(u.sysFd),
+			uintptr(unsafe.Pointer(&hdr)),
+			0,
+			0,
+			0,
+			0,
+		)
+
+		if errno == unix.EINTR {
+			continue
+		}
+
+		if errno != 0 {
+			return &net.OpError{Op: "sendmsg", Err: errno}
+		}
+
+		return nil
+	}
+}
+
+// isGSOError returns true if the error indicates GSO is not supported by the NIC
+func isGSOError(err error) bool {
+	var opErr *net.OpError
+	if !errors.As(err, &opErr) {
+		return false
+	}
+	// EIO typically means the NIC doesn't support checksum offload required for GSO
+	return errors.Is(opErr.Err, unix.EIO)
+}
+
 func NewUDPStatsEmitter(udpConns []Conn) func() {
 	// Check if our kernel supports SO_MEMINFO before registering the gauges
 	var udpGauges [][unix.SK_MEMINFO_VARS]metrics.Gauge

udp/udp_linux_32.go

@@ -30,13 +30,26 @@ type rawMessage struct {
 	Len uint32
 }
 
-func (u *StdConn) PrepareRawMessages(n int) ([]rawMessage, [][]byte, [][]byte) {
+func (u *StdConn) PrepareRawMessages(n int) ([]rawMessage, [][]byte, [][]byte, [][]byte) {
 	msgs := make([]rawMessage, n)
 	buffers := make([][]byte, n)
 	names := make([][]byte, n)
+	controls := make([][]byte, n)
+
+	// Use larger buffers if GRO is enabled to hold coalesced packets
+	bufSize := MTU
+	if u.groSupported {
+		bufSize = groMaxPacketSize
+	}
+
+	// Control buffer size for receiving UDP_GRO segment size
+	controlSize := 0
+	if u.groSupported {
+		controlSize = unix.CmsgSpace(2) // space for uint16 segment size
+	}
 
 	for i := range msgs {
-		buffers[i] = make([]byte, MTU)
+		buffers[i] = make([]byte, bufSize)
 		names[i] = make([]byte, unix.SizeofSockaddrInet6)
 
 		vs := []iovec{
@@ -48,7 +61,34 @@ func (u *StdConn) PrepareRawMessages(n int) ([]rawMessage, [][]byte, [][]byte) {
 
 		msgs[i].Hdr.Name = &names[i][0]
 		msgs[i].Hdr.Namelen = uint32(len(names[i]))
+
+		// Set up control message buffer for GRO
+		if controlSize > 0 {
+			controls[i] = make([]byte, controlSize)
+			msgs[i].Hdr.Control = &controls[i][0]
+			msgs[i].Hdr.Controllen = uint32(controlSize)
+		}
 	}
 
-	return msgs, buffers, names
+	return msgs, buffers, names, controls
+}
+
+func setIovecBase(iov *iovec, base *byte) {
+	iov.Base = base
+}
+
+func setIovecLen(iov *iovec, l int) {
+	iov.Len = uint32(l)
+}
+
+func setMsghdrIovlen(hdr *msghdr, l int) {
+	hdr.Iovlen = uint32(l)
+}
+
+func setMsghdrControllen(hdr *msghdr, l int) {
+	hdr.Controllen = uint32(l)
+}
+
+func getMsghdrControllen(hdr *msghdr) int {
+	return int(hdr.Controllen)
 }

udp/udp_linux_64.go

@@ -33,13 +33,26 @@ type rawMessage struct {
 	Pad0 [4]byte
 }
 
-func (u *StdConn) PrepareRawMessages(n int) ([]rawMessage, [][]byte, [][]byte) {
+func (u *StdConn) PrepareRawMessages(n int) ([]rawMessage, [][]byte, [][]byte, [][]byte) {
 	msgs := make([]rawMessage, n)
 	buffers := make([][]byte, n)
 	names := make([][]byte, n)
+	controls := make([][]byte, n)
+
+	// Use larger buffers if GRO is enabled to hold coalesced packets
+	bufSize := MTU
+	if u.groSupported {
+		bufSize = groMaxPacketSize
+	}
+
+	// Control buffer size for receiving UDP_GRO segment size
+	controlSize := 0
+	if u.groSupported {
+		controlSize = unix.CmsgSpace(2) // space for uint16 segment size
+	}
 
 	for i := range msgs {
-		buffers[i] = make([]byte, MTU)
+		buffers[i] = make([]byte, bufSize)
 		names[i] = make([]byte, unix.SizeofSockaddrInet6)
 
 		vs := []iovec{
@@ -51,7 +64,34 @@ func (u *StdConn) PrepareRawMessages(n int) ([]rawMessage, [][]byte, [][]byte) {
 
 		msgs[i].Hdr.Name = &names[i][0]
 		msgs[i].Hdr.Namelen = uint32(len(names[i]))
+
+		// Set up control message buffer for GRO
+		if controlSize > 0 {
+			controls[i] = make([]byte, controlSize)
+			msgs[i].Hdr.Control = &controls[i][0]
+			msgs[i].Hdr.Controllen = uint64(controlSize)
+		}
 	}
 
-	return msgs, buffers, names
+	return msgs, buffers, names, controls
+}
+
+func setIovecBase(iov *iovec, base *byte) {
+	iov.Base = base
+}
+
+func setIovecLen(iov *iovec, l int) {
+	iov.Len = uint64(l)
+}
+
+func setMsghdrIovlen(hdr *msghdr, l int) {
+	hdr.Iovlen = uint64(l)
+}
+
+func setMsghdrControllen(hdr *msghdr, l int) {
+	hdr.Controllen = uint64(l)
+}
+
+func getMsghdrControllen(hdr *msghdr) int {
+	return int(hdr.Controllen)
 }

udp/udp_rio_windows.go

@@ -332,12 +332,29 @@ func (u *RIOConn) SupportsMultipleReaders() bool {
 	return false
 }
 
+func (u *RIOConn) SupportsGSO() bool {
+	return false
+}
+
+func (u *RIOConn) SupportsGRO() bool {
+	return false
+}
+
 func (u *RIOConn) Rebind() error {
 	return nil
 }
 
 func (u *RIOConn) ReloadConfig(*config.C) {}
 
+func (u *RIOConn) WriteBatch(pkts []BatchPacket) (int, error) {
+	for i := range pkts {
+		if err := u.WriteTo(pkts[i].Payload, pkts[i].Addr); err != nil {
+			return i, err
+		}
+	}
+	return len(pkts), nil
+}
+
 func (u *RIOConn) Close() error {
 	if !u.isOpen.CompareAndSwap(true, false) {
 		return nil

udp/udp_tester.go

@@ -131,6 +131,14 @@ func (u *TesterConn) SupportsMultipleReaders() bool {
 	return false
 }
 
+func (u *TesterConn) SupportsGSO() bool {
+	return false
+}
+
+func (u *TesterConn) SupportsGRO() bool {
+	return false
+}
+
 func (u *TesterConn) Rebind() error {
 	return nil
 }
@@ -142,3 +150,12 @@ func (u *TesterConn) Close() error {
 	}
 	return nil
 }
+
+func (u *TesterConn) WriteBatch(pkts []BatchPacket) (int, error) {
+	for i := range pkts {
+		if err := u.WriteTo(pkts[i].Payload, pkts[i].Addr); err != nil {
+			return i, err
+		}
+	}
+	return len(pkts), nil
+}