hmmmmmm it works i guess maybe

overlay/device.go

@@ -3,6 +3,7 @@ package overlay
 import (
 	"io"
 	"net/netip"
+	"sync"
 
 	"github.com/slackhq/nebula/routing"
 )
@@ -15,3 +16,84 @@ type Device interface {
 	RoutesFor(netip.Addr) routing.Gateways
 	NewMultiQueueReader() (io.ReadWriteCloser, error)
 }
+
+// Packet represents a single packet buffer with optional headroom to carry
+// metadata (for example virtio-net headers).
+type Packet struct {
+	Buf     []byte
+	Offset  int
+	Len     int
+	release func()
+}
+
+func (p *Packet) Payload() []byte {
+	return p.Buf[p.Offset : p.Offset+p.Len]
+}
+
+func (p *Packet) Reset() {
+	p.Len = 0
+	p.Offset = 0
+	p.release = nil
+}
+
+func (p *Packet) Release() {
+	if p.release != nil {
+		p.release()
+		p.release = nil
+	}
+}
+
+func (p *Packet) Capacity() int {
+	return len(p.Buf) - p.Offset
+}
+
+// PacketPool manages reusable buffers with headroom.
+type PacketPool struct {
+	headroom int
+	blksz    int
+	pool     sync.Pool
+}
+
+func NewPacketPool(headroom, payload int) *PacketPool {
+	p := &PacketPool{headroom: headroom, blksz: headroom + payload}
+	p.pool.New = func() any {
+		buf := make([]byte, p.blksz)
+		return &Packet{Buf: buf, Offset: headroom}
+	}
+	return p
+}
+
+func (p *PacketPool) Get() *Packet {
+	pkt := p.pool.Get().(*Packet)
+	pkt.Offset = p.headroom
+	pkt.Len = 0
+	pkt.release = func() { p.put(pkt) }
+	return pkt
+}
+
+func (p *PacketPool) put(pkt *Packet) {
+	pkt.Reset()
+	p.pool.Put(pkt)
+}
+
+// BatchReader allows reading multiple packets into a shared pool with
+// preallocated headroom (e.g. virtio-net headers).
+type BatchReader interface {
+	ReadIntoBatch(pool *PacketPool) ([]*Packet, error)
+}
+
+// BatchWriter writes a slice of packets that carry their own metadata.
+type BatchWriter interface {
+	WriteBatch(packets []*Packet) (int, error)
+}
+
+// BatchCapableDevice describes a device that can efficiently read and write
+// batches of packets with virtio headroom.
+type BatchCapableDevice interface {
+	Device
+	BatchReader
+	BatchWriter
+	BatchHeadroom() int
+	BatchPayloadCap() int
+	BatchSize() int
+}

overlay/tun_linux_batch.go

@@ -0,0 +1,56 @@
+//go:build linux && !android && !e2e_testing
+
+package overlay
+
+import "fmt"
+
+func (t *tun) batchIO() (*wireguardTunIO, bool) {
+	io, ok := t.ReadWriteCloser.(*wireguardTunIO)
+	return io, ok
+}
+
+func (t *tun) ReadIntoBatch(pool *PacketPool) ([]*Packet, error) {
+	io, ok := t.batchIO()
+	if !ok {
+		return nil, fmt.Errorf("wireguard batch I/O not enabled")
+	}
+	return io.ReadIntoBatch(pool)
+}
+
+func (t *tun) WriteBatch(packets []*Packet) (int, error) {
+	io, ok := t.batchIO()
+	if ok {
+		return io.WriteBatch(packets)
+	}
+	for _, pkt := range packets {
+		if pkt == nil {
+			continue
+		}
+		if _, err := t.Write(pkt.Payload()[:pkt.Len]); err != nil {
+			return 0, err
+		}
+		pkt.Release()
+	}
+	return len(packets), nil
+}
+
+func (t *tun) BatchHeadroom() int {
+	if io, ok := t.batchIO(); ok {
+		return io.BatchHeadroom()
+	}
+	return 0
+}
+
+func (t *tun) BatchPayloadCap() int {
+	if io, ok := t.batchIO(); ok {
+		return io.BatchPayloadCap()
+	}
+	return 0
+}
+
+func (t *tun) BatchSize() int {
+	if io, ok := t.batchIO(); ok {
+		return io.BatchSize()
+	}
+	return 1
+}

overlay/wireguard_tun_linux.go

@@ -14,15 +14,15 @@ type wireguardTunIO struct {
 	mtu       int
 	batchSize int
 
-	readMu   sync.Mutex
-	readBufs [][]byte
-	readLens []int
-	pending  [][]byte
-	pendIdx  int
+	readMu      sync.Mutex
+	readBuffers [][]byte
+	readLens    []int
+	legacyBuf   []byte
 
-	writeMu   sync.Mutex
-	writeBuf  []byte
-	writeWrap [][]byte
+	writeMu      sync.Mutex
+	writeBuf     []byte
+	writeWrap    [][]byte
+	writeBuffers [][]byte
 }
 
 func newWireguardTunIO(dev wgtun.Device, mtu int) *wireguardTunIO {
@@ -33,17 +33,12 @@ func newWireguardTunIO(dev wgtun.Device, mtu int) *wireguardTunIO {
 	if mtu <= 0 {
 		mtu = DefaultMTU
 	}
-	bufs := make([][]byte, batch)
-	for i := range bufs {
-		bufs[i] = make([]byte, wgtun.VirtioNetHdrLen+mtu)
-	}
 	return &wireguardTunIO{
 		dev:       dev,
 		mtu:       mtu,
 		batchSize: batch,
-		readBufs:  bufs,
 		readLens:  make([]int, batch),
-		pending:   make([][]byte, 0, batch),
+		legacyBuf: make([]byte, wgtun.VirtioNetHdrLen+mtu),
 		writeBuf:  make([]byte, wgtun.VirtioNetHdrLen+mtu),
 		writeWrap: make([][]byte, 1),
 	}
@@ -53,29 +48,21 @@ func (w *wireguardTunIO) Read(p []byte) (int, error) {
 	w.readMu.Lock()
 	defer w.readMu.Unlock()
 
-	for {
-		if w.pendIdx < len(w.pending) {
-			segment := w.pending[w.pendIdx]
-			w.pendIdx++
-			n := copy(p, segment)
-			return n, nil
-		}
-
-		n, err := w.dev.Read(w.readBufs, w.readLens, wgtun.VirtioNetHdrLen)
-		if err != nil {
-			return 0, err
-		}
-		w.pending = w.pending[:0]
-		w.pendIdx = 0
-		for i := 0; i < n; i++ {
-			length := w.readLens[i]
-			if length == 0 {
-				continue
-			}
-			segment := w.readBufs[i][wgtun.VirtioNetHdrLen : wgtun.VirtioNetHdrLen+length]
-			w.pending = append(w.pending, segment)
-		}
+	bufs := w.readBuffers
+	if len(bufs) == 0 {
+		bufs = [][]byte{w.legacyBuf}
+		w.readBuffers = bufs
 	}
+	n, err := w.dev.Read(bufs[:1], w.readLens[:1], wgtun.VirtioNetHdrLen)
+	if err != nil {
+		return 0, err
+	}
+	if n == 0 {
+		return 0, nil
+	}
+	length := w.readLens[0]
+	copy(p, w.legacyBuf[wgtun.VirtioNetHdrLen:wgtun.VirtioNetHdrLen+length])
+	return length, nil
 }
 
 func (w *wireguardTunIO) Write(p []byte) (int, error) {
@@ -97,6 +84,134 @@ func (w *wireguardTunIO) Write(p []byte) (int, error) {
 	return len(p), nil
 }
 
+func (w *wireguardTunIO) ReadIntoBatch(pool *PacketPool) ([]*Packet, error) {
+	if pool == nil {
+		return nil, fmt.Errorf("wireguard tun: packet pool is nil")
+	}
+
+	w.readMu.Lock()
+	defer w.readMu.Unlock()
+
+	if len(w.readBuffers) < w.batchSize {
+		w.readBuffers = make([][]byte, w.batchSize)
+	}
+	if len(w.readLens) < w.batchSize {
+		w.readLens = make([]int, w.batchSize)
+	}
+
+	packets := make([]*Packet, w.batchSize)
+	requiredHeadroom := w.BatchHeadroom()
+	requiredPayload := w.BatchPayloadCap()
+	headroom := 0
+	for i := 0; i < w.batchSize; i++ {
+		pkt := pool.Get()
+		if pkt == nil {
+			releasePackets(packets[:i])
+			return nil, fmt.Errorf("wireguard tun: packet pool returned nil packet")
+		}
+		if pkt.Capacity() < requiredPayload {
+			pkt.Release()
+			releasePackets(packets[:i])
+			return nil, fmt.Errorf("wireguard tun: packet capacity %d below required %d", pkt.Capacity(), requiredPayload)
+		}
+		if i == 0 {
+			headroom = pkt.Offset
+			if headroom < requiredHeadroom {
+				pkt.Release()
+				releasePackets(packets[:i])
+				return nil, fmt.Errorf("wireguard tun: packet headroom %d below virtio requirement %d", headroom, requiredHeadroom)
+			}
+		} else if pkt.Offset != headroom {
+			pkt.Release()
+			releasePackets(packets[:i])
+			return nil, fmt.Errorf("wireguard tun: inconsistent packet headroom (%d != %d)", pkt.Offset, headroom)
+		}
+		packets[i] = pkt
+		w.readBuffers[i] = pkt.Buf
+	}
+
+	n, err := w.dev.Read(w.readBuffers[:w.batchSize], w.readLens[:w.batchSize], headroom)
+	if err != nil {
+		releasePackets(packets)
+		return nil, err
+	}
+	if n == 0 {
+		releasePackets(packets)
+		return nil, nil
+	}
+	for i := 0; i < n; i++ {
+		packets[i].Len = w.readLens[i]
+	}
+	for i := n; i < w.batchSize; i++ {
+		packets[i].Release()
+		packets[i] = nil
+	}
+	return packets[:n], nil
+}
+
+func (w *wireguardTunIO) WriteBatch(packets []*Packet) (int, error) {
+	if len(packets) == 0 {
+		return 0, nil
+	}
+	requiredHeadroom := w.BatchHeadroom()
+	offset := packets[0].Offset
+	if offset < requiredHeadroom {
+		releasePackets(packets)
+		return 0, fmt.Errorf("wireguard tun: packet offset %d smaller than required headroom %d", offset, requiredHeadroom)
+	}
+	for _, pkt := range packets {
+		if pkt == nil {
+			continue
+		}
+		if pkt.Offset != offset {
+			releasePackets(packets)
+			return 0, fmt.Errorf("wireguard tun: mixed packet offsets not supported")
+		}
+		limit := pkt.Offset + pkt.Len
+		if limit > len(pkt.Buf) {
+			releasePackets(packets)
+			return 0, fmt.Errorf("wireguard tun: packet length %d exceeds buffer capacity %d", pkt.Len, len(pkt.Buf)-pkt.Offset)
+		}
+	}
+	w.writeMu.Lock()
+	defer w.writeMu.Unlock()
+
+	if len(w.writeBuffers) < len(packets) {
+		w.writeBuffers = make([][]byte, len(packets))
+	}
+	for i, pkt := range packets {
+		if pkt == nil {
+			w.writeBuffers[i] = nil
+			continue
+		}
+		limit := pkt.Offset + pkt.Len
+		w.writeBuffers[i] = pkt.Buf[:limit]
+	}
+	n, err := w.dev.Write(w.writeBuffers[:len(packets)], offset)
+	releasePackets(packets)
+	return n, err
+}
+
+func (w *wireguardTunIO) BatchHeadroom() int {
+	return wgtun.VirtioNetHdrLen
+}
+
+func (w *wireguardTunIO) BatchPayloadCap() int {
+	return w.mtu
+}
+
+func (w *wireguardTunIO) BatchSize() int {
+	return w.batchSize
+}
+
 func (w *wireguardTunIO) Close() error {
 	return nil
 }
+
+func releasePackets(pkts []*Packet) {
+	for _, pkt := range pkts {
+		if pkt != nil {
+			pkt.Release()
+		}
+	}
+}