tell stupid ai not to meddle with the paths on other platforms by adding a copy

CHANGELOG.md

@@ -7,12 +7,30 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+### Added
+
+- Experimental Linux UDP offload support: enable `listen.enable_gso` and
+  `listen.enable_gro` to activate UDP_SEGMENT batching and GRO receive
+  splitting. Includes automatic capability probing, per-packet fallbacks, and
+  runtime metrics/logs for visibility.
+- Optional Linux TUN `virtio_net_hdr` support: set `tun.enable_vnet_hdr` to
+  have Nebula negotiate VNET headers and offload flags so future batches can
+  be delivered to the kernel with metadata instead of per-packet writes.
+- Linux UDP send sharding can now be tuned with `listen.send_shards`; defaults
+  to `GOMAXPROCS` but can be increased to stripe heavy peers across more
+  goroutines.
+
 ### Changed
 
 - `default_local_cidr_any` now defaults to false, meaning that any firewall rule
   intended to target an `unsafe_routes` entry must explicitly declare it via the
   `local_cidr` field. This is almost always the intended behavior. This flag is
   deprecated and will be removed in a future release.
+- UDP receive path now enqueues into per-worker lock-free rings, restoring the
+  `listen.decrypt_workers`/`listen.decrypt_queue_depth` tuning knobs while
+  eliminating the mutex contention from the old shared channel.
+- Increased replay protection window to 32k packets so high-throughput links
+  tolerate larger bursts of reordering without tripping the anti-replay logic.
 
 ## [1.9.4] - 2024-09-09
 

connection_state.go

@@ -13,7 +13,10 @@ import (
 	"github.com/slackhq/nebula/noiseutil"
 )
 
-const ReplayWindow = 1024
+// ReplayWindow controls the size of the sliding window used to detect replays.
+// High-bandwidth links with GRO/GSO can reorder more than a thousand packets in
+// flight, so keep this comfortably above the largest expected burst.
+const ReplayWindow = 32768
 
 type ConnectionState struct {
 	eKey           *NebulaCipherState

interface.go

@@ -5,9 +5,11 @@ import (
 	"errors"
 	"fmt"
 	"io"
+	"math/bits"
 	"net/netip"
 	"os"
 	"runtime"
+	"sync"
 	"sync/atomic"
 	"time"
 
@@ -21,7 +23,12 @@ import (
 	"github.com/slackhq/nebula/udp"
 )
 
-const mtu = 9001
+const (
+	mtu                        = 9001
+	tunReadBufferSize          = mtu * 8
+	defaultDecryptWorkerFactor = 2
+	defaultInboundQueueDepth   = 1024
+)
 
 type InterfaceConfig struct {
 	HostMap            *HostMap
@@ -48,6 +55,8 @@ type InterfaceConfig struct {
 
 	ConntrackCacheTimeout time.Duration
 	l                     *logrus.Logger
+	DecryptWorkers        int
+	DecryptQueueDepth     int
 }
 
 type Interface struct {
@@ -93,6 +102,166 @@ type Interface struct {
 	cachedPacketMetrics *cachedPacketMetrics
 
 	l              *logrus.Logger
+	ctx            context.Context
+	udpListenWG    sync.WaitGroup
+	inboundPool    sync.Pool
+	decryptWG      sync.WaitGroup
+	decryptQueues  []*inboundRing
+	decryptWorkers int
+	decryptStates  []decryptWorkerState
+	decryptCounter atomic.Uint32
+}
+
+type inboundPacket struct {
+	addr    netip.AddrPort
+	payload []byte
+	release func()
+	queue   int
+}
+
+type decryptWorkerState struct {
+	queue  *inboundRing
+	notify chan struct{}
+}
+
+type decryptContext struct {
+	ctTicker *firewall.ConntrackCacheTicker
+	plain    []byte
+	head     header.H
+	fwPacket firewall.Packet
+	light    *LightHouseHandler
+	nebula   []byte
+}
+
+type inboundCell struct {
+	seq atomic.Uint64
+	pkt *inboundPacket
+}
+
+type inboundRing struct {
+	mask       uint64
+	cells      []inboundCell
+	enqueuePos atomic.Uint64
+	dequeuePos atomic.Uint64
+}
+
+func newInboundRing(capacity int) *inboundRing {
+	if capacity < 2 {
+		capacity = 2
+	}
+	size := nextPowerOfTwo(uint32(capacity))
+	if size < 2 {
+		size = 2
+	}
+	ring := &inboundRing{
+		mask:  uint64(size - 1),
+		cells: make([]inboundCell, size),
+	}
+	for i := range ring.cells {
+		ring.cells[i].seq.Store(uint64(i))
+	}
+	return ring
+}
+
+func nextPowerOfTwo(v uint32) uint32 {
+	if v == 0 {
+		return 1
+	}
+	return 1 << (32 - bits.LeadingZeros32(v-1))
+}
+
+func (r *inboundRing) Enqueue(pkt *inboundPacket) bool {
+	var cell *inboundCell
+	pos := r.enqueuePos.Load()
+	for {
+		cell = &r.cells[pos&r.mask]
+		seq := cell.seq.Load()
+		diff := int64(seq) - int64(pos)
+		if diff == 0 {
+			if r.enqueuePos.CompareAndSwap(pos, pos+1) {
+				break
+			}
+		} else if diff < 0 {
+			return false
+		} else {
+			pos = r.enqueuePos.Load()
+		}
+	}
+	cell.pkt = pkt
+	cell.seq.Store(pos + 1)
+	return true
+}
+
+func (r *inboundRing) Dequeue() (*inboundPacket, bool) {
+	var cell *inboundCell
+	pos := r.dequeuePos.Load()
+	for {
+		cell = &r.cells[pos&r.mask]
+		seq := cell.seq.Load()
+		diff := int64(seq) - int64(pos+1)
+		if diff == 0 {
+			if r.dequeuePos.CompareAndSwap(pos, pos+1) {
+				break
+			}
+		} else if diff < 0 {
+			return nil, false
+		} else {
+			pos = r.dequeuePos.Load()
+		}
+	}
+	pkt := cell.pkt
+	cell.pkt = nil
+	cell.seq.Store(pos + r.mask + 1)
+	return pkt, true
+}
+
+func (f *Interface) getInboundPacket() *inboundPacket {
+	if pkt, ok := f.inboundPool.Get().(*inboundPacket); ok && pkt != nil {
+		return pkt
+	}
+	return &inboundPacket{}
+}
+
+func (f *Interface) putInboundPacket(pkt *inboundPacket) {
+	if pkt == nil {
+		return
+	}
+	pkt.addr = netip.AddrPort{}
+	pkt.payload = nil
+	pkt.release = nil
+	pkt.queue = 0
+	f.inboundPool.Put(pkt)
+}
+
+func newDecryptContext(f *Interface) *decryptContext {
+	return &decryptContext{
+		ctTicker: firewall.NewConntrackCacheTicker(f.conntrackCacheTimeout),
+		plain:    make([]byte, udp.MTU),
+		head:     header.H{},
+		fwPacket: firewall.Packet{},
+		light:    f.lightHouse.NewRequestHandler(),
+		nebula:   make([]byte, 12, 12),
+	}
+}
+
+func (f *Interface) processInboundPacket(pkt *inboundPacket, ctx *decryptContext) {
+	if pkt == nil {
+		return
+	}
+	defer func() {
+		if pkt.release != nil {
+			pkt.release()
+		}
+		f.putInboundPacket(pkt)
+	}()
+
+	ctx.head = header.H{}
+	ctx.fwPacket = firewall.Packet{}
+	var cache firewall.ConntrackCache
+	if ctx.ctTicker != nil {
+		cache = ctx.ctTicker.Get(f.l)
+	}
+	f.readOutsidePackets(pkt.addr, nil, ctx.plain[:0], pkt.payload, &ctx.head, &ctx.fwPacket, ctx.light, ctx.nebula, pkt.queue, cache)
 }
 
 type EncWriter interface {
@@ -162,6 +331,35 @@ func NewInterface(ctx context.Context, c *InterfaceConfig) (*Interface, error) {
 	}
 
 	cs := c.pki.getCertState()
+	decryptWorkers := c.DecryptWorkers
+	if decryptWorkers < 0 {
+		decryptWorkers = 0
+	}
+	if decryptWorkers == 0 {
+		decryptWorkers = c.routines * defaultDecryptWorkerFactor
+		if decryptWorkers < c.routines {
+			decryptWorkers = c.routines
+		}
+	}
+	if decryptWorkers < 0 {
+		decryptWorkers = 0
+	}
+	if runtime.GOOS != "linux" {
+		decryptWorkers = 0
+	}
+
+	queueDepth := c.DecryptQueueDepth
+	if queueDepth <= 0 {
+		queueDepth = defaultInboundQueueDepth
+	}
+	minDepth := c.routines * 64
+	if minDepth <= 0 {
+		minDepth = 64
+	}
+	if queueDepth < minDepth {
+		queueDepth = minDepth
+	}
+
 	ifce := &Interface{
 		pki:                   c.pki,
 		hostMap:               c.HostMap,
@@ -195,6 +393,9 @@ func NewInterface(ctx context.Context, c *InterfaceConfig) (*Interface, error) {
 		},
 
 		l:              c.l,
+		ctx:            ctx,
+		inboundPool:    sync.Pool{New: func() any { return &inboundPacket{} }},
+		decryptWorkers: decryptWorkers,
 	}
 
 	ifce.tryPromoteEvery.Store(c.tryPromoteEvery)
@@ -203,6 +404,19 @@ func NewInterface(ctx context.Context, c *InterfaceConfig) (*Interface, error) {
 
 	ifce.connectionManager.intf = ifce
 
+	if decryptWorkers > 0 {
+		ifce.decryptQueues = make([]*inboundRing, decryptWorkers)
+		ifce.decryptStates = make([]decryptWorkerState, decryptWorkers)
+		for i := 0; i < decryptWorkers; i++ {
+			queue := newInboundRing(queueDepth)
+			ifce.decryptQueues[i] = queue
+			ifce.decryptStates[i] = decryptWorkerState{
+				queue:  queue,
+				notify: make(chan struct{}, 1),
+			}
+		}
+	}
+
 	return ifce, nil
 }
 
@@ -242,8 +456,68 @@ func (f *Interface) activate() {
 	}
 }
 
+func (f *Interface) startDecryptWorkers() {
+	if f.decryptWorkers <= 0 || len(f.decryptQueues) == 0 {
+		return
+	}
+	f.decryptWG.Add(f.decryptWorkers)
+	for i := 0; i < f.decryptWorkers; i++ {
+		go f.decryptWorker(i)
+	}
+}
+
+func (f *Interface) decryptWorker(id int) {
+	defer f.decryptWG.Done()
+	if id < 0 || id >= len(f.decryptStates) {
+		return
+	}
+	state := f.decryptStates[id]
+	if state.queue == nil {
+		return
+	}
+	ctx := newDecryptContext(f)
+	for {
+		for {
+			pkt, ok := state.queue.Dequeue()
+			if !ok {
+				break
+			}
+			f.processInboundPacket(pkt, ctx)
+		}
+		if f.closed.Load() || f.ctx.Err() != nil {
+			for {
+				pkt, ok := state.queue.Dequeue()
+				if !ok {
+					return
+				}
+				f.processInboundPacket(pkt, ctx)
+			}
+		}
+		select {
+		case <-f.ctx.Done():
+		case <-state.notify:
+		}
+	}
+}
+
+func (f *Interface) notifyDecryptWorker(idx int) {
+	if idx < 0 || idx >= len(f.decryptStates) {
+		return
+	}
+	state := f.decryptStates[idx]
+	if state.notify == nil {
+		return
+	}
+	select {
+	case state.notify <- struct{}{}:
+	default:
+	}
+}
+
 func (f *Interface) run() {
+	f.startDecryptWorkers()
 	// Launch n queues to read packets from udp
+	f.udpListenWG.Add(f.routines)
 	for i := 0; i < f.routines; i++ {
 		go f.listenOut(i)
 	}
@@ -256,6 +530,7 @@ func (f *Interface) run() {
 
 func (f *Interface) listenOut(i int) {
 	runtime.LockOSThread()
+	defer f.udpListenWG.Done()
 
 	var li udp.Conn
 	if i > 0 {
@@ -264,23 +539,78 @@ func (f *Interface) listenOut(i int) {
 		li = f.outside
 	}
 
-	ctCache := firewall.NewConntrackCacheTicker(f.conntrackCacheTimeout)
+	useWorkers := f.decryptWorkers > 0 && len(f.decryptQueues) > 0
-	lhh := f.lightHouse.NewRequestHandler()
+	var (
-	plaintext := make([]byte, udp.MTU)
+		inlineTicker  *firewall.ConntrackCacheTicker
-	h := &header.H{}
+		inlineHandler *LightHouseHandler
-	fwPacket := &firewall.Packet{}
+		inlinePlain   []byte
-	nb := make([]byte, 12, 12)
+		inlineHeader  header.H
+		inlinePacket  firewall.Packet
+		inlineNB      []byte
+		inlineCtx     *decryptContext
+	)
 
-	li.ListenOut(func(fromUdpAddr netip.AddrPort, payload []byte) {
+	if useWorkers {
-		f.readOutsidePackets(fromUdpAddr, nil, plaintext[:0], payload, h, fwPacket, lhh, nb, i, ctCache.Get(f.l))
+		inlineCtx = newDecryptContext(f)
+	} else {
+		inlineTicker = firewall.NewConntrackCacheTicker(f.conntrackCacheTimeout)
+		inlineHandler = f.lightHouse.NewRequestHandler()
+		inlinePlain = make([]byte, udp.MTU)
+		inlineNB = make([]byte, 12, 12)
+	}
+
+	li.ListenOut(func(fromUdpAddr netip.AddrPort, payload []byte, release func()) {
+		if !useWorkers {
+			if release != nil {
+				defer release()
+			}
+			select {
+			case <-f.ctx.Done():
+				return
+			default:
+			}
+			inlineHeader = header.H{}
+			inlinePacket = firewall.Packet{}
+			var cache firewall.ConntrackCache
+			if inlineTicker != nil {
+				cache = inlineTicker.Get(f.l)
+			}
+			f.readOutsidePackets(fromUdpAddr, nil, inlinePlain[:0], payload, &inlineHeader, &inlinePacket, inlineHandler, inlineNB, i, cache)
+			return
+		}
+
+		if f.ctx.Err() != nil {
+			if release != nil {
+				release()
+			}
+			return
+		}
+
+		pkt := f.getInboundPacket()
+		pkt.addr = fromUdpAddr
+		pkt.payload = payload
+		pkt.release = release
+		pkt.queue = i
+
+		queueCount := len(f.decryptQueues)
+		if queueCount == 0 {
+			f.processInboundPacket(pkt, inlineCtx)
+			return
+		}
+		w := int(f.decryptCounter.Add(1)-1) % queueCount
+		if w < 0 || w >= queueCount || !f.decryptQueues[w].Enqueue(pkt) {
+			f.processInboundPacket(pkt, inlineCtx)
+			return
+		}
+		f.notifyDecryptWorker(w)
 	})
 }
 
 func (f *Interface) listenIn(reader io.ReadWriteCloser, i int) {
 	runtime.LockOSThread()
 
-	packet := make([]byte, mtu)
+	packet := make([]byte, tunReadBufferSize)
-	out := make([]byte, mtu)
+	out := make([]byte, tunReadBufferSize)
 	fwPacket := &firewall.Packet{}
 	nb := make([]byte, 12, 12)
 
@@ -458,6 +788,19 @@ func (f *Interface) Close() error {
 		}
 	}
 
+	f.udpListenWG.Wait()
+	if f.decryptWorkers > 0 {
+		for _, state := range f.decryptStates {
+			if state.notify != nil {
+				select {
+				case state.notify <- struct{}{}:
+				default:
+				}
+			}
+		}
+		f.decryptWG.Wait()
+	}
+
 	// Release the tun device
 	return f.inside.Close()
 }

main.go

@@ -120,6 +120,8 @@ func Main(c *config.C, configTest bool, buildVersion string, logger *logrus.Logg
 		l.WithField("duration", conntrackCacheTimeout).Info("Using routine-local conntrack cache")
 	}
 
+	udp.SetDisableUDPCsum(c.GetBool("listen.disable_udp_checksum", false))
+
 	var tun overlay.Device
 	if !configTest {
 		c.CatchHUP(ctx)
@@ -221,6 +223,9 @@ func Main(c *config.C, configTest bool, buildVersion string, logger *logrus.Logg
 		}
 	}
 
+	decryptWorkers := c.GetInt("listen.decrypt_workers", 0)
+	decryptQueueDepth := c.GetInt("listen.decrypt_queue_depth", 0)
+
 	ifConfig := &InterfaceConfig{
 		HostMap:               hostMap,
 		Inside:                tun,
@@ -243,6 +248,8 @@ func Main(c *config.C, configTest bool, buildVersion string, logger *logrus.Logg
 		punchy:                punchy,
 		ConntrackCacheTimeout: conntrackCacheTimeout,
 		l:                     l,
+		DecryptWorkers:        decryptWorkers,
+		DecryptQueueDepth:     decryptQueueDepth,
 	}
 
 	var ifce *Interface

outside.go

@@ -470,7 +470,13 @@ func (f *Interface) decryptToTun(hostinfo *HostInfo, messageCounter uint64, out
 
 	out, err = hostinfo.ConnectionState.dKey.DecryptDanger(out, packet[:header.Len], packet[header.Len:], messageCounter, nb)
 	if err != nil {
-		hostinfo.logger(f.l).WithError(err).Error("Failed to decrypt packet")
+		hostinfo.logger(f.l).
+			WithError(err).
+			WithField("tag", "decrypt-debug").
+			WithField("remoteIndexLocal", hostinfo.localIndexId).
+			WithField("messageCounter", messageCounter).
+			WithField("packet_len", len(packet)).
+			Error("Failed to decrypt packet")
 		return false
 	}
 

overlay/tun_linux.go

@@ -33,6 +33,9 @@ type tun struct {
 	TXQueueLen    int
 	deviceIndex   int
 	ioctlFd       uintptr
+	enableVnetHdr bool
+	vnetHdrLen    int
+	queues        []*tunQueue
 
 	Routes                    atomic.Pointer[[]Route]
 	routeTree                 atomic.Pointer[bart.Table[routing.Gateways]]
@@ -65,10 +68,90 @@ type ifreqQLEN struct {
 	pad   [8]byte
 }
 
+const (
+	virtioNetHdrLen     = 12
+	tunDefaultMaxPacket = 65536
+)
+
+type tunQueue struct {
+	file          *os.File
+	fd            int
+	enableVnetHdr bool
+	vnetHdrLen    int
+	maxPacket     int
+	writeScratch  []byte
+	readScratch   []byte
+	l             *logrus.Logger
+}
+
+func newTunQueue(file *os.File, enableVnetHdr bool, vnetHdrLen, maxPacket int, l *logrus.Logger) *tunQueue {
+	if maxPacket <= 0 {
+		maxPacket = tunDefaultMaxPacket
+	}
+	q := &tunQueue{
+		file:          file,
+		fd:            int(file.Fd()),
+		enableVnetHdr: enableVnetHdr,
+		vnetHdrLen:    vnetHdrLen,
+		maxPacket:     maxPacket,
+		l:             l,
+	}
+	if enableVnetHdr {
+		q.growReadScratch(maxPacket)
+	}
+	return q
+}
+
+func (q *tunQueue) growReadScratch(packetSize int) {
+	needed := q.vnetHdrLen + packetSize
+	if needed < q.vnetHdrLen+DefaultMTU {
+		needed = q.vnetHdrLen + DefaultMTU
+	}
+	if q.readScratch == nil || cap(q.readScratch) < needed {
+		q.readScratch = make([]byte, needed)
+	} else {
+		q.readScratch = q.readScratch[:needed]
+	}
+}
+
+func (q *tunQueue) setMaxPacket(packet int) {
+	if packet <= 0 {
+		packet = DefaultMTU
+	}
+	q.maxPacket = packet
+	if q.enableVnetHdr {
+		q.growReadScratch(packet)
+	}
+}
+
+func configureVnetHdr(fd int, hdrLen int, l *logrus.Logger) error {
+	features, err := unix.IoctlGetInt(fd, unix.TUNGETFEATURES)
+	if err == nil && features&unix.IFF_VNET_HDR == 0 {
+		return fmt.Errorf("kernel does not support IFF_VNET_HDR")
+	}
+	if err := unix.IoctlSetInt(fd, unix.TUNSETVNETHDRSZ, hdrLen); err != nil {
+		return err
+	}
+	offload := unix.TUN_F_CSUM | unix.TUN_F_UFO
+	if err := unix.IoctlSetInt(fd, unix.TUNSETOFFLOAD, offload); err != nil {
+		if l != nil {
+			l.WithError(err).Warn("Failed to enable TUN offload features")
+		}
+	}
+	return nil
+}
+
 func newTunFromFd(c *config.C, l *logrus.Logger, deviceFd int, vpnNetworks []netip.Prefix) (*tun, error) {
 	file := os.NewFile(uintptr(deviceFd), "/dev/net/tun")
+	enableVnetHdr := c.GetBool("tun.enable_vnet_hdr", false)
+	if enableVnetHdr {
+		if err := configureVnetHdr(deviceFd, virtioNetHdrLen, l); err != nil {
+			l.WithError(err).Warn("Failed to configure VNET header support on provided tun fd; disabling")
+			enableVnetHdr = false
+		}
+	}
 
-	t, err := newTunGeneric(c, l, file, vpnNetworks)
+	t, err := newTunGeneric(c, l, file, vpnNetworks, enableVnetHdr)
 	if err != nil {
 		return nil, err
 	}
@@ -106,14 +189,25 @@ func newTun(c *config.C, l *logrus.Logger, vpnNetworks []netip.Prefix, multiqueu
 	if multiqueue {
 		req.Flags |= unix.IFF_MULTI_QUEUE
 	}
+	enableVnetHdr := c.GetBool("tun.enable_vnet_hdr", false)
+	if enableVnetHdr {
+		req.Flags |= unix.IFF_VNET_HDR
+	}
 	copy(req.Name[:], c.GetString("tun.dev", ""))
 	if err = ioctl(uintptr(fd), uintptr(unix.TUNSETIFF), uintptr(unsafe.Pointer(&req))); err != nil {
 		return nil, err
 	}
 	name := strings.Trim(string(req.Name[:]), "\x00")
 
+	if enableVnetHdr {
+		if err := configureVnetHdr(fd, virtioNetHdrLen, l); err != nil {
+			l.WithError(err).Warn("Failed to configure VNET header support on tun device; disabling")
+			enableVnetHdr = false
+		}
+	}
+
 	file := os.NewFile(uintptr(fd), "/dev/net/tun")
-	t, err := newTunGeneric(c, l, file, vpnNetworks)
+	t, err := newTunGeneric(c, l, file, vpnNetworks, enableVnetHdr)
 	if err != nil {
 		return nil, err
 	}
@@ -123,21 +217,30 @@ func newTun(c *config.C, l *logrus.Logger, vpnNetworks []netip.Prefix, multiqueu
 	return t, nil
 }
 
-func newTunGeneric(c *config.C, l *logrus.Logger, file *os.File, vpnNetworks []netip.Prefix) (*tun, error) {
+func newTunGeneric(c *config.C, l *logrus.Logger, file *os.File, vpnNetworks []netip.Prefix, enableVnetHdr bool) (*tun, error) {
+	queue := newTunQueue(file, enableVnetHdr, virtioNetHdrLen, tunDefaultMaxPacket, l)
 	t := &tun{
-		ReadWriteCloser:           file,
+		ReadWriteCloser:           queue,
 		fd:                        int(file.Fd()),
 		vpnNetworks:               vpnNetworks,
 		TXQueueLen:                c.GetInt("tun.tx_queue", 500),
 		useSystemRoutes:           c.GetBool("tun.use_system_route_table", false),
 		useSystemRoutesBufferSize: c.GetInt("tun.use_system_route_table_buffer_size", 0),
 		l:                         l,
+		enableVnetHdr:             enableVnetHdr,
+		vnetHdrLen:                virtioNetHdrLen,
+		queues:                    []*tunQueue{queue},
 	}
 
 	err := t.reload(c, true)
 	if err != nil {
 		return nil, err
 	}
+	if enableVnetHdr {
+		for _, q := range t.queues {
+			q.setMaxPacket(t.MaxMTU)
+		}
+	}
 
 	c.RegisterReloadCallback(func(c *config.C) {
 		err := t.reload(c, false)
@@ -180,6 +283,11 @@ func (t *tun) reload(c *config.C, initial bool) error {
 
 	t.MaxMTU = newMaxMTU
 	t.DefaultMTU = newDefaultMTU
+	if t.enableVnetHdr {
+		for _, q := range t.queues {
+			q.setMaxPacket(t.MaxMTU)
+		}
+	}
 
 	// Teach nebula how to handle the routes before establishing them in the system table
 	oldRoutes := t.Routes.Swap(&routes)
@@ -224,14 +332,87 @@ func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
 
 	var req ifReq
 	req.Flags = uint16(unix.IFF_TUN | unix.IFF_NO_PI | unix.IFF_MULTI_QUEUE)
+	if t.enableVnetHdr {
+		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")
+	queue := newTunQueue(file, t.enableVnetHdr, t.vnetHdrLen, t.MaxMTU, t.l)
+	if t.enableVnetHdr {
+		if err := configureVnetHdr(fd, t.vnetHdrLen, t.l); err != nil {
+			queue.enableVnetHdr = false
+		}
+	}
+	t.queues = append(t.queues, queue)
 
-	return file, nil
+	return queue, nil
+}
+
+func (q *tunQueue) Read(p []byte) (int, error) {
+	if !q.enableVnetHdr {
+		return q.file.Read(p)
+	}
+
+	if len(p)+q.vnetHdrLen > cap(q.readScratch) {
+		q.growReadScratch(len(p))
+	}
+
+	buf := q.readScratch[:cap(q.readScratch)]
+	n, err := q.file.Read(buf)
+	if n <= 0 {
+		return n, err
+	}
+	if n < q.vnetHdrLen {
+		if err == nil {
+			err = io.ErrUnexpectedEOF
+		}
+		return 0, err
+	}
+
+	payload := buf[q.vnetHdrLen:n]
+	if len(payload) > len(p) {
+		copy(p, payload[:len(p)])
+		if err == nil {
+			err = io.ErrShortBuffer
+		}
+		return len(p), err
+	}
+	copy(p, payload)
+	return len(payload), err
+}
+
+func (q *tunQueue) Write(b []byte) (int, error) {
+	if !q.enableVnetHdr {
+		return unix.Write(q.fd, b)
+	}
+
+	total := q.vnetHdrLen + len(b)
+	if cap(q.writeScratch) < total {
+		q.writeScratch = make([]byte, total)
+	} else {
+		q.writeScratch = q.writeScratch[:total]
+	}
+
+	for i := 0; i < q.vnetHdrLen; i++ {
+		q.writeScratch[i] = 0
+	}
+	copy(q.writeScratch[q.vnetHdrLen:], b)
+
+	n, err := unix.Write(q.fd, q.writeScratch)
+	if n >= q.vnetHdrLen {
+		n -= q.vnetHdrLen
+	} else {
+		n = 0
+	}
+	return n, err
+}
+
+func (q *tunQueue) Close() error {
+	return q.file.Close()
 }
 
 func (t *tun) RoutesFor(ip netip.Addr) routing.Gateways {

udp/config.go

@@ -0,0 +1,16 @@
+package udp
+
+import "sync/atomic"
+
+var disableUDPCsum atomic.Bool
+
+// SetDisableUDPCsum controls whether IPv4 UDP sockets opt out of kernel
+// checksum calculation via SO_NO_CHECK. Only applicable on platforms that
+// support the option (Linux). IPv6 always keeps the checksum enabled.
+func SetDisableUDPCsum(disable bool) {
+	disableUDPCsum.Store(disable)
+}
+
+func udpChecksumDisabled() bool {
+	return disableUDPCsum.Load()
+}

udp/conn.go

@@ -11,6 +11,7 @@ const MTU = 9001
 type EncReader func(
 	addr netip.AddrPort,
 	payload []byte,
+	release func(),
 )
 
 type Conn interface {

udp/msghdr_helper_linux_32.go

@@ -0,0 +1,25 @@
+//go:build linux && (386 || amd64p32 || arm || mips || mipsle) && !android && !e2e_testing
+// +build linux
+// +build 386 amd64p32 arm mips mipsle
+// +build !android
+// +build !e2e_testing
+
+package udp
+
+import "golang.org/x/sys/unix"
+
+func controllen(n int) uint32 {
+	return uint32(n)
+}
+
+func setCmsgLen(h *unix.Cmsghdr, n int) {
+	h.Len = uint32(unix.CmsgLen(n))
+}
+
+func setIovecLen(v *unix.Iovec, n int) {
+	v.Len = uint32(n)
+}
+
+func setMsghdrIovlen(m *unix.Msghdr, n int) {
+	m.Iovlen = uint32(n)
+}

udp/msghdr_helper_linux_64.go

@@ -0,0 +1,25 @@
+//go:build linux && (amd64 || arm64 || ppc64 || ppc64le || mips64 || mips64le || s390x || riscv64 || loong64) && !android && !e2e_testing
+// +build linux
+// +build amd64 arm64 ppc64 ppc64le mips64 mips64le s390x riscv64 loong64
+// +build !android
+// +build !e2e_testing
+
+package udp
+
+import "golang.org/x/sys/unix"
+
+func controllen(n int) uint64 {
+	return uint64(n)
+}
+
+func setCmsgLen(h *unix.Cmsghdr, n int) {
+	h.Len = uint64(unix.CmsgLen(n))
+}
+
+func setIovecLen(v *unix.Iovec, n int) {
+	v.Len = uint64(n)
+}
+
+func setMsghdrIovlen(m *unix.Msghdr, n int) {
+	m.Iovlen = uint64(n)
+}

udp/sendmmsg_linux_32.go

@@ -0,0 +1,25 @@
+//go:build linux && (386 || amd64p32 || arm || mips || mipsle) && !android && !e2e_testing
+
+package udp
+
+import (
+	"unsafe"
+
+	"golang.org/x/sys/unix"
+)
+
+type linuxMmsgHdr struct {
+	Hdr unix.Msghdr
+	Len uint32
+}
+
+func sendmmsg(fd int, hdrs []linuxMmsgHdr, flags int) (int, error) {
+	if len(hdrs) == 0 {
+		return 0, nil
+	}
+	n, _, errno := unix.Syscall6(unix.SYS_SENDMMSG, uintptr(fd), uintptr(unsafe.Pointer(&hdrs[0])), uintptr(len(hdrs)), uintptr(flags), 0, 0)
+	if errno != 0 {
+		return int(n), errno
+	}
+	return int(n), nil
+}

udp/sendmmsg_linux_64.go

@@ -0,0 +1,26 @@
+//go:build linux && (amd64 || arm64 || ppc64 || ppc64le || mips64 || mips64le || s390x || riscv64 || loong64) && !android && !e2e_testing
+
+package udp
+
+import (
+	"unsafe"
+
+	"golang.org/x/sys/unix"
+)
+
+type linuxMmsgHdr struct {
+	Hdr unix.Msghdr
+	Len uint32
+	_   uint32
+}
+
+func sendmmsg(fd int, hdrs []linuxMmsgHdr, flags int) (int, error) {
+	if len(hdrs) == 0 {
+		return 0, nil
+	}
+	n, _, errno := unix.Syscall6(unix.SYS_SENDMMSG, uintptr(fd), uintptr(unsafe.Pointer(&hdrs[0])), uintptr(len(hdrs)), uintptr(flags), 0, 0)
+	if errno != 0 {
+		return int(n), errno
+	}
+	return int(n), nil
+}

udp/udp_darwin.go

@@ -180,7 +180,7 @@ func (u *StdConn) ListenOut(r EncReader) {
 			u.l.WithError(err).Error("unexpected udp socket receive error")
 		}
 
-		r(netip.AddrPortFrom(rua.Addr().Unmap(), rua.Port()), buffer[:n])
+		r(netip.AddrPortFrom(rua.Addr().Unmap(), rua.Port()), buffer[:n], nil)
 	}
 }
 

udp/udp_generic.go

@@ -82,6 +82,6 @@ func (u *GenericConn) ListenOut(r EncReader) {
 			return
 		}
 
-		r(netip.AddrPortFrom(rua.Addr().Unmap(), rua.Port()), buffer[:n])
+		r(netip.AddrPortFrom(rua.Addr().Unmap(), rua.Port()), buffer[:n], nil)
 	}
 }

udp/udp_linux_32.go

@@ -30,17 +30,29 @@ type rawMessage struct {
 	Len uint32
 }
 
-func (u *StdConn) PrepareRawMessages(n int) ([]rawMessage, [][]byte, [][]byte) {
+func (u *StdConn) PrepareRawMessages(n int) ([]rawMessage, [][]byte, [][]byte, [][]byte) {
+	controlLen := int(u.controlLen.Load())
+
 	msgs := make([]rawMessage, n)
 	buffers := make([][]byte, n)
 	names := make([][]byte, n)
 
+	var controls [][]byte
+	if controlLen > 0 {
+		controls = make([][]byte, n)
+	}
+
 	for i := range msgs {
-		buffers[i] = make([]byte, MTU)
+		size := int(u.groBufSize.Load())
+		if size < MTU {
+			size = MTU
+		}
+		buf := u.borrowRxBuffer(size)
+		buffers[i] = buf
 		names[i] = make([]byte, unix.SizeofSockaddrInet6)
 
 		vs := []iovec{
-			{Base: &buffers[i][0], Len: uint32(len(buffers[i]))},
+			{Base: &buf[0], Len: uint32(len(buf))},
 		}
 
 		msgs[i].Hdr.Iov = &vs[0]
@@ -48,7 +60,22 @@ 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 controlLen > 0 {
+			controls[i] = make([]byte, controlLen)
+			msgs[i].Hdr.Control = &controls[i][0]
+			msgs[i].Hdr.Controllen = controllen(len(controls[i]))
+		} else {
+			msgs[i].Hdr.Control = nil
+			msgs[i].Hdr.Controllen = controllen(0)
+		}
 	}
 
-	return msgs, buffers, names
+	return msgs, buffers, names, controls
+}
+
+func setIovecBase(msg *rawMessage, buf []byte) {
+	iov := (*iovec)(msg.Hdr.Iov)
+	iov.Base = &buf[0]
+	iov.Len = uint32(len(buf))
 }

udp/udp_linux_64.go

@@ -33,25 +33,50 @@ 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) {
+	controlLen := int(u.controlLen.Load())
+
 	msgs := make([]rawMessage, n)
 	buffers := make([][]byte, n)
 	names := make([][]byte, n)
 
+	var controls [][]byte
+	if controlLen > 0 {
+		controls = make([][]byte, n)
+	}
+
 	for i := range msgs {
-		buffers[i] = make([]byte, MTU)
+		size := int(u.groBufSize.Load())
+		if size < MTU {
+			size = MTU
+		}
+		buf := u.borrowRxBuffer(size)
+		buffers[i] = buf
 		names[i] = make([]byte, unix.SizeofSockaddrInet6)
 
-		vs := []iovec{
+		vs := []iovec{{Base: &buf[0], Len: uint64(len(buf))}}
-			{Base: &buffers[i][0], Len: uint64(len(buffers[i]))},
-		}
 
 		msgs[i].Hdr.Iov = &vs[0]
 		msgs[i].Hdr.Iovlen = uint64(len(vs))
 
 		msgs[i].Hdr.Name = &names[i][0]
 		msgs[i].Hdr.Namelen = uint32(len(names[i]))
+
+		if controlLen > 0 {
+			controls[i] = make([]byte, controlLen)
+			msgs[i].Hdr.Control = &controls[i][0]
+			msgs[i].Hdr.Controllen = controllen(len(controls[i]))
+		} else {
+			msgs[i].Hdr.Control = nil
+			msgs[i].Hdr.Controllen = controllen(0)
+		}
 	}
 
-	return msgs, buffers, names
+	return msgs, buffers, names, controls
+}
+
+func setIovecBase(msg *rawMessage, buf []byte) {
+	iov := (*iovec)(msg.Hdr.Iov)
+	iov.Base = &buf[0]
+	iov.Len = uint64(len(buf))
 }

udp/udp_rio_windows.go

@@ -149,7 +149,7 @@ func (u *RIOConn) ListenOut(r EncReader) {
 			continue
 		}
 
-		r(netip.AddrPortFrom(netip.AddrFrom16(rua.Addr).Unmap(), (rua.Port>>8)|((rua.Port&0xff)<<8)), buffer[:n])
+		r(netip.AddrPortFrom(netip.AddrFrom16(rua.Addr).Unmap(), (rua.Port>>8)|((rua.Port&0xff)<<8)), buffer[:n], nil)
 	}
 }
 

udp/udp_tester.go

@@ -112,7 +112,7 @@ func (u *TesterConn) ListenOut(r EncReader) {
 		if !ok {
 			return
 		}
-		r(p.From, p.Data)
+		r(p.From, p.Data, func() {})
 	}
 }