GRO

cruft

fix tests

haha yep faster

checksum speed

haha

save pennies

fix

typo!

checkpt

GSO again

batch.go

@@ -0,0 +1,70 @@
+package nebula
+
+import "net/netip"
+
+// sendBatchCap is the maximum number of encrypted packets accumulated before a
+// flush is forced. TSO superpackets segment to at most ~45 packets on
+// reasonable MTUs, so 128 leaves headroom without bloating the backing
+// allocation.
+const sendBatchCap = 128
+
+// sendBatch accumulates encrypted UDP packets for a single sendmmsg flush.
+// One sendBatch is owned by each listenIn goroutine; no locking is needed.
+// The backing storage holds up to batchCap packets of slotCap bytes each;
+// bufs and dsts are parallel slices of committed slots.
+type sendBatch struct {
+	bufs     [][]byte
+	dsts     []netip.AddrPort
+	backing  []byte
+	slotCap  int
+	batchCap int
+	nextSlot int
+}
+
+func newSendBatch(batchCap, slotCap int) *sendBatch {
+	return &sendBatch{
+		bufs:     make([][]byte, 0, batchCap),
+		dsts:     make([]netip.AddrPort, 0, batchCap),
+		backing:  make([]byte, batchCap*slotCap),
+		slotCap:  slotCap,
+		batchCap: batchCap,
+	}
+}
+
+// Next returns a zero-length slice with slotCap capacity over the next unused
+// slot's backing bytes. The caller writes into the returned slice and then
+// calls Commit with the final length and destination. Next returns nil when
+// the batch is full.
+func (b *sendBatch) Next() []byte {
+	if b.nextSlot >= b.batchCap {
+		return nil
+	}
+	start := b.nextSlot * b.slotCap
+	return b.backing[start : start : start+b.slotCap]
+}
+
+// Commit records the slot just returned by Next as a packet of length n
+// destined for dst.
+func (b *sendBatch) Commit(n int, dst netip.AddrPort) {
+	start := b.nextSlot * b.slotCap
+	b.bufs = append(b.bufs, b.backing[start:start+n])
+	b.dsts = append(b.dsts, dst)
+	b.nextSlot++
+}
+
+// Reset clears committed slots; backing storage is retained for reuse.
+func (b *sendBatch) Reset() {
+	b.bufs = b.bufs[:0]
+	b.dsts = b.dsts[:0]
+	b.nextSlot = 0
+}
+
+// Len returns the number of committed packets.
+func (b *sendBatch) Len() int {
+	return len(b.bufs)
+}
+
+// Cap returns the maximum number of slots in the batch.
+func (b *sendBatch) Cap() int {
+	return b.batchCap
+}

batch_test.go

@@ -0,0 +1,69 @@
+package nebula
+
+import (
+	"net/netip"
+	"testing"
+)
+
+func TestSendBatchBookkeeping(t *testing.T) {
+	b := newSendBatch(4, 32)
+	if b.Len() != 0 || b.Cap() != 4 {
+		t.Fatalf("fresh batch: len=%d cap=%d", b.Len(), b.Cap())
+	}
+
+	ap := netip.MustParseAddrPort("10.0.0.1:4242")
+	for i := 0; i < 4; i++ {
+		slot := b.Next()
+		if slot == nil {
+			t.Fatalf("slot %d: Next returned nil before cap", i)
+		}
+		if cap(slot) != 32 || len(slot) != 0 {
+			t.Fatalf("slot %d: got len=%d cap=%d want len=0 cap=32", i, len(slot), cap(slot))
+		}
+		// Write a marker byte.
+		slot = append(slot, byte(i), byte(i+1), byte(i+2))
+		b.Commit(len(slot), ap)
+	}
+	if b.Next() != nil {
+		t.Fatalf("Next should return nil when full")
+	}
+	if b.Len() != 4 {
+		t.Fatalf("Len=%d want 4", b.Len())
+	}
+	for i, buf := range b.bufs {
+		if len(buf) != 3 || buf[0] != byte(i) {
+			t.Errorf("buf %d: %x", i, buf)
+		}
+		if b.dsts[i] != ap {
+			t.Errorf("dst %d: got %v want %v", i, b.dsts[i], ap)
+		}
+	}
+
+	// Reset returns empty and Next works again.
+	b.Reset()
+	if b.Len() != 0 {
+		t.Fatalf("after Reset Len=%d want 0", b.Len())
+	}
+	slot := b.Next()
+	if slot == nil || cap(slot) != 32 {
+		t.Fatalf("after Reset Next nil or wrong cap: %v cap=%d", slot == nil, cap(slot))
+	}
+}
+
+func TestSendBatchSlotsDoNotOverlap(t *testing.T) {
+	b := newSendBatch(3, 8)
+	ap := netip.MustParseAddrPort("10.0.0.1:80")
+
+	// Fill three slots, each with its own sentinel byte.
+	for i := 0; i < 3; i++ {
+		s := b.Next()
+		s = append(s, byte(0xA0+i), byte(0xB0+i))
+		b.Commit(len(s), ap)
+	}
+
+	for i, buf := range b.bufs {
+		if buf[0] != byte(0xA0+i) || buf[1] != byte(0xB0+i) {
+			t.Errorf("slot %d corrupted: %x", i, buf)
+		}
+	}
+}

overlay/coalesce/tcp_coalesce.go

@@ -0,0 +1,484 @@
+package coalesce
+
+import (
+	"bytes"
+	"encoding/binary"
+	"io"
+
+	"github.com/slackhq/nebula/overlay/tio"
+)
+
+// ipProtoTCP is the IANA protocol number for TCP. Hardcoded instead of
+// reaching for golang.org/x/sys/unix — that package doesn't define the
+// constant on Windows, which would break cross-compiles even though this
+// file runs unchanged on every platform.
+const ipProtoTCP = 6
+
+// tcpCoalesceBufSize caps total bytes per superpacket. Mirrors the kernel's
+// sk_gso_max_size of ~64KiB; anything beyond this would be rejected anyway.
+const tcpCoalesceBufSize = 65535
+
+// tcpCoalesceMaxSegs caps how many segments we'll coalesce into a single
+// superpacket. Keeping this well below the kernel's TSO ceiling bounds
+// latency.
+const tcpCoalesceMaxSegs = 64
+
+// tcpCoalesceHdrCap is the scratch space we copy a seed's IP+TCP header
+// into. IPv6 (40) + TCP with full options (60) = 100 bytes.
+const tcpCoalesceHdrCap = 100
+
+// initialSlots is the starting capacity of the slot pool. One flow per
+// packet is the worst case so this matches a typical UDP recvmmsg batch.
+const initialSlots = 64
+
+// flowKey identifies a TCP flow by {src, dst, sport, dport, family}.
+// Comparable, so linear scans over the slot list stay tight.
+type flowKey struct {
+	src, dst     [16]byte
+	sport, dport uint16
+	isV6         bool
+}
+
+// coalesceSlot is one entry in the coalescer's ordered event queue. When
+// passthrough is true the slot holds a single borrowed packet that must be
+// emitted verbatim (non-TCP, non-admissible TCP, or oversize seed). When
+// passthrough is false the slot is an in-progress coalesced superpacket:
+// hdrBuf is a mutable copy of the seed's IP+TCP header (we patch total
+// length and pseudo-header partial at flush), and payIovs are *borrowed*
+// slices from the caller's plaintext buffers — no payload is ever copied.
+// The caller (listenOut) must keep those buffers alive until Flush.
+type coalesceSlot struct {
+	passthrough bool
+	rawPkt      []byte // borrowed when passthrough
+
+	fk       flowKey
+	hdrBuf   [tcpCoalesceHdrCap]byte
+	hdrLen   int
+	ipHdrLen int
+	isV6     bool
+	gsoSize  int
+	numSeg   int
+	totalPay int
+	nextSeq  uint32
+	// psh closes the chain: set when the last-accepted segment had PSH or
+	// was sub-gsoSize. No further appends after that.
+	psh     bool
+	payIovs [][]byte
+}
+
+// TCPCoalescer accumulates adjacent in-flow TCP data segments across
+// multiple concurrent flows and emits each flow's run as a single TSO
+// superpacket via tio.GSOWriter. All output — coalesced or not — is
+// deferred until Flush so arrival order is preserved on the wire. Owns
+// no locks; one coalescer per TUN write queue.
+type TCPCoalescer struct {
+	plainW io.Writer
+	gsoW   tio.GSOWriter // nil when the queue doesn't support TSO
+
+	// slots is the ordered event queue. Flush walks it once and emits each
+	// entry as either a WriteGSO (coalesced) or a plainW.Write (passthrough).
+	slots []*coalesceSlot
+	// openSlots maps a flow key to its most recent non-sealed slot, so new
+	// segments can extend an in-progress superpacket in O(1). Slots are
+	// removed from this map when they close (PSH or short-last-segment),
+	// when a non-admissible packet for that flow arrives, or in Flush.
+	openSlots map[flowKey]*coalesceSlot
+	pool      []*coalesceSlot // free list for reuse
+}
+
+func NewTCPCoalescer(w io.Writer) *TCPCoalescer {
+	c := &TCPCoalescer{
+		plainW:    w,
+		slots:     make([]*coalesceSlot, 0, initialSlots),
+		openSlots: make(map[flowKey]*coalesceSlot, initialSlots),
+		pool:      make([]*coalesceSlot, 0, initialSlots),
+	}
+	if gw, ok := w.(tio.GSOWriter); ok && gw.GSOSupported() {
+		c.gsoW = gw
+	}
+	return c
+}
+
+// parsedTCP holds the fields extracted from a single parse so later steps
+// (admission, slot lookup, canAppend) don't re-walk the header.
+type parsedTCP struct {
+	fk        flowKey
+	ipHdrLen  int
+	tcpHdrLen int
+	hdrLen    int
+	payLen    int
+	seq       uint32
+	flags     byte
+}
+
+// parseTCPBase extracts the flow key and IP/TCP offsets for any TCP packet,
+// regardless of whether it's admissible for coalescing. Returns ok=false
+// for non-TCP or malformed input. Accepts IPv4 (no options, no fragmentation)
+// and IPv6 (no extension headers).
+func parseTCPBase(pkt []byte) (parsedTCP, bool) {
+	var p parsedTCP
+	if len(pkt) < 20 {
+		return p, false
+	}
+	v := pkt[0] >> 4
+	switch v {
+	case 4:
+		ihl := int(pkt[0]&0x0f) * 4
+		if ihl != 20 {
+			return p, false
+		}
+		if pkt[9] != ipProtoTCP {
+			return p, false
+		}
+		// Reject actual fragmentation (MF or non-zero frag offset).
+		if binary.BigEndian.Uint16(pkt[6:8])&0x3fff != 0 {
+			return p, false
+		}
+		totalLen := int(binary.BigEndian.Uint16(pkt[2:4]))
+		if totalLen > len(pkt) || totalLen < ihl {
+			return p, false
+		}
+		p.ipHdrLen = 20
+		p.fk.isV6 = false
+		copy(p.fk.src[:4], pkt[12:16])
+		copy(p.fk.dst[:4], pkt[16:20])
+		pkt = pkt[:totalLen]
+	case 6:
+		if len(pkt) < 40 {
+			return p, false
+		}
+		if pkt[6] != ipProtoTCP {
+			return p, false
+		}
+		payloadLen := int(binary.BigEndian.Uint16(pkt[4:6]))
+		if 40+payloadLen > len(pkt) {
+			return p, false
+		}
+		p.ipHdrLen = 40
+		p.fk.isV6 = true
+		copy(p.fk.src[:], pkt[8:24])
+		copy(p.fk.dst[:], pkt[24:40])
+		pkt = pkt[:40+payloadLen]
+	default:
+		return p, false
+	}
+
+	if len(pkt) < p.ipHdrLen+20 {
+		return p, false
+	}
+	tcpOff := int(pkt[p.ipHdrLen+12]>>4) * 4
+	if tcpOff < 20 || tcpOff > 60 {
+		return p, false
+	}
+	if len(pkt) < p.ipHdrLen+tcpOff {
+		return p, false
+	}
+	p.tcpHdrLen = tcpOff
+	p.hdrLen = p.ipHdrLen + tcpOff
+	p.payLen = len(pkt) - p.hdrLen
+	p.seq = binary.BigEndian.Uint32(pkt[p.ipHdrLen+4 : p.ipHdrLen+8])
+	p.flags = pkt[p.ipHdrLen+13]
+	p.fk.sport = binary.BigEndian.Uint16(pkt[p.ipHdrLen : p.ipHdrLen+2])
+	p.fk.dport = binary.BigEndian.Uint16(pkt[p.ipHdrLen+2 : p.ipHdrLen+4])
+	return p, true
+}
+
+// coalesceable reports whether a parsed TCP segment is eligible for
+// coalescing. Accepts only ACK or ACK|PSH with a non-empty payload.
+func (p parsedTCP) coalesceable() bool {
+	const ack = 0x10
+	const psh = 0x08
+	if p.flags&^(ack|psh) != 0 || p.flags&ack == 0 {
+		return false
+	}
+	return p.payLen > 0
+}
+
+// Add borrows pkt. The caller must keep pkt valid until the next Flush,
+// whether or not the packet was coalesced — passthrough (non-admissible)
+// packets are queued and written at Flush time, not synchronously.
+func (c *TCPCoalescer) Add(pkt []byte) error {
+	if c.gsoW == nil {
+		c.addPassthrough(pkt)
+		return nil
+	}
+
+	info, ok := parseTCPBase(pkt)
+	if !ok {
+		// Non-TCP or malformed — can't possibly collide with an open flow.
+		c.addPassthrough(pkt)
+		return nil
+	}
+	if !info.coalesceable() {
+		// TCP but not admissible (SYN/FIN/RST/URG/CWR/ECE or zero-payload).
+		// Seal this flow's open slot so later in-flow packets don't extend
+		// it and accidentally reorder past this passthrough.
+		delete(c.openSlots, info.fk)
+		c.addPassthrough(pkt)
+		return nil
+	}
+
+	if open := c.openSlots[info.fk]; open != nil {
+		if c.canAppend(open, pkt, info) {
+			c.appendPayload(open, pkt, info)
+			if open.psh {
+				delete(c.openSlots, info.fk)
+			}
+			return nil
+		}
+		// Can't extend — seal it and fall through to seed a fresh slot.
+		delete(c.openSlots, info.fk)
+	}
+	c.seed(pkt, info)
+	return nil
+}
+
+// Flush emits every queued event in arrival order. Coalesced slots go out
+// via WriteGSO; passthrough slots go out via plainW.Write. Returns the
+// first error observed; keeps draining so one bad packet doesn't hold up
+// the rest. After Flush returns, borrowed payload slices may be recycled.
+func (c *TCPCoalescer) Flush() error {
+	var first error
+	for _, s := range c.slots {
+		var err error
+		if s.passthrough {
+			_, err = c.plainW.Write(s.rawPkt)
+		} else {
+			err = c.flushSlot(s)
+		}
+		if err != nil && first == nil {
+			first = err
+		}
+		c.release(s)
+	}
+	for i := range c.slots {
+		c.slots[i] = nil
+	}
+	c.slots = c.slots[:0]
+	for k := range c.openSlots {
+		delete(c.openSlots, k)
+	}
+	return first
+}
+
+func (c *TCPCoalescer) addPassthrough(pkt []byte) {
+	s := c.take()
+	s.passthrough = true
+	s.rawPkt = pkt
+	c.slots = append(c.slots, s)
+}
+
+func (c *TCPCoalescer) seed(pkt []byte, info parsedTCP) {
+	if info.hdrLen > tcpCoalesceHdrCap || info.hdrLen+info.payLen > tcpCoalesceBufSize {
+		// Pathological shape — can't fit our scratch, emit as-is.
+		c.addPassthrough(pkt)
+		return
+	}
+	s := c.take()
+	s.passthrough = false
+	s.rawPkt = nil
+	copy(s.hdrBuf[:], pkt[:info.hdrLen])
+	s.hdrLen = info.hdrLen
+	s.ipHdrLen = info.ipHdrLen
+	s.isV6 = info.fk.isV6
+	s.fk = info.fk
+	s.gsoSize = info.payLen
+	s.numSeg = 1
+	s.totalPay = info.payLen
+	s.nextSeq = info.seq + uint32(info.payLen)
+	s.psh = info.flags&0x08 != 0
+	s.payIovs = append(s.payIovs[:0], pkt[info.hdrLen:info.hdrLen+info.payLen])
+	c.slots = append(c.slots, s)
+	if !s.psh {
+		c.openSlots[info.fk] = s
+	}
+}
+
+// canAppend reports whether info's packet extends the slot's seed: same
+// header shape and stable contents, adjacent seq, not oversized, chain not
+// closed.
+func (c *TCPCoalescer) canAppend(s *coalesceSlot, pkt []byte, info parsedTCP) bool {
+	if s.psh {
+		return false
+	}
+	if info.hdrLen != s.hdrLen {
+		return false
+	}
+	if info.seq != s.nextSeq {
+		return false
+	}
+	if s.numSeg >= tcpCoalesceMaxSegs {
+		return false
+	}
+	if info.payLen > s.gsoSize {
+		return false
+	}
+	if s.hdrLen+s.totalPay+info.payLen > tcpCoalesceBufSize {
+		return false
+	}
+	if !headersMatch(s.hdrBuf[:s.hdrLen], pkt[:info.hdrLen], s.isV6, s.ipHdrLen) {
+		return false
+	}
+	return true
+}
+
+func (c *TCPCoalescer) appendPayload(s *coalesceSlot, pkt []byte, info parsedTCP) {
+	s.payIovs = append(s.payIovs, pkt[info.hdrLen:info.hdrLen+info.payLen])
+	s.numSeg++
+	s.totalPay += info.payLen
+	s.nextSeq = info.seq + uint32(info.payLen)
+	if info.payLen < s.gsoSize || info.flags&0x08 != 0 {
+		s.psh = true
+	}
+}
+
+func (c *TCPCoalescer) take() *coalesceSlot {
+	if n := len(c.pool); n > 0 {
+		s := c.pool[n-1]
+		c.pool[n-1] = nil
+		c.pool = c.pool[:n-1]
+		return s
+	}
+	return &coalesceSlot{}
+}
+
+func (c *TCPCoalescer) release(s *coalesceSlot) {
+	s.passthrough = false
+	s.rawPkt = nil
+	for i := range s.payIovs {
+		s.payIovs[i] = nil
+	}
+	s.payIovs = s.payIovs[:0]
+	s.numSeg = 0
+	s.totalPay = 0
+	s.psh = false
+	c.pool = append(c.pool, s)
+}
+
+// flushSlot patches the header and calls WriteGSO. Does not remove the
+// slot from c.slots.
+func (c *TCPCoalescer) flushSlot(s *coalesceSlot) error {
+	total := s.hdrLen + s.totalPay
+	l4Len := total - s.ipHdrLen
+	hdr := s.hdrBuf[:s.hdrLen]
+
+	if s.isV6 {
+		binary.BigEndian.PutUint16(hdr[4:6], uint16(l4Len))
+	} else {
+		binary.BigEndian.PutUint16(hdr[2:4], uint16(total))
+		hdr[10] = 0
+		hdr[11] = 0
+		binary.BigEndian.PutUint16(hdr[10:12], ipv4HdrChecksum(hdr[:s.ipHdrLen]))
+	}
+
+	var psum uint32
+	if s.isV6 {
+		psum = pseudoSumIPv6(hdr[8:24], hdr[24:40], ipProtoTCP, l4Len)
+	} else {
+		psum = pseudoSumIPv4(hdr[12:16], hdr[16:20], ipProtoTCP, l4Len)
+	}
+	tcsum := s.ipHdrLen + 16
+	binary.BigEndian.PutUint16(hdr[tcsum:tcsum+2], foldOnceNoInvert(psum))
+
+	return c.gsoW.WriteGSO(hdr, s.payIovs, uint16(s.gsoSize), s.isV6, uint16(s.ipHdrLen))
+}
+
+// headersMatch compares two IP+TCP header prefixes for byte-for-byte
+// equality on every field that must be identical across coalesced
+// segments. Size/IPID/IPCsum/seq/flags/tcpCsum are masked out.
+func headersMatch(a, b []byte, isV6 bool, ipHdrLen int) bool {
+	if len(a) != len(b) {
+		return false
+	}
+	if isV6 {
+		// IPv6: bytes [0:4] = version/TC/flow-label, [6:8] = next_hdr/hop,
+		// [8:40] = src+dst. Skip [4:6] payload length.
+		if !bytes.Equal(a[0:4], b[0:4]) {
+			return false
+		}
+		if !bytes.Equal(a[6:40], b[6:40]) {
+			return false
+		}
+	} else {
+		// IPv4: [0:2] version/IHL/TOS, [6:10] flags/fragoff/TTL/proto,
+		// [12:20] src+dst. Skip [2:4] total len, [4:6] id, [10:12] csum.
+		if !bytes.Equal(a[0:2], b[0:2]) {
+			return false
+		}
+		if !bytes.Equal(a[6:10], b[6:10]) {
+			return false
+		}
+		if !bytes.Equal(a[12:20], b[12:20]) {
+			return false
+		}
+	}
+	// TCP: compare [0:4] ports, [8:13] ack+dataoff, [14:16] window,
+	// [18:tcpHdrLen] options (incl. urgent).
+	tcp := ipHdrLen
+	if !bytes.Equal(a[tcp:tcp+4], b[tcp:tcp+4]) {
+		return false
+	}
+	if !bytes.Equal(a[tcp+8:tcp+13], b[tcp+8:tcp+13]) {
+		return false
+	}
+	if !bytes.Equal(a[tcp+14:tcp+16], b[tcp+14:tcp+16]) {
+		return false
+	}
+	if !bytes.Equal(a[tcp+18:], b[tcp+18:]) {
+		return false
+	}
+	return true
+}
+
+// ipv4HdrChecksum computes the IPv4 header checksum over hdr (which must
+// already have its checksum field zeroed) and returns the folded/inverted
+// 16-bit value to store.
+func ipv4HdrChecksum(hdr []byte) uint16 {
+	var sum uint32
+	for i := 0; i+1 < len(hdr); i += 2 {
+		sum += uint32(binary.BigEndian.Uint16(hdr[i : i+2]))
+	}
+	if len(hdr)%2 == 1 {
+		sum += uint32(hdr[len(hdr)-1]) << 8
+	}
+	for sum>>16 != 0 {
+		sum = (sum & 0xffff) + (sum >> 16)
+	}
+	return ^uint16(sum)
+}
+
+// pseudoSumIPv4 / pseudoSumIPv6 build the TCP pseudo-header partial sum
+// expected by the virtio NEEDS_CSUM kernel path: the 32-bit accumulator
+// before folding.
+func pseudoSumIPv4(src, dst []byte, proto byte, l4Len int) uint32 {
+	var sum uint32
+	sum += uint32(binary.BigEndian.Uint16(src[0:2]))
+	sum += uint32(binary.BigEndian.Uint16(src[2:4]))
+	sum += uint32(binary.BigEndian.Uint16(dst[0:2]))
+	sum += uint32(binary.BigEndian.Uint16(dst[2:4]))
+	sum += uint32(proto)
+	sum += uint32(l4Len)
+	return sum
+}
+
+func pseudoSumIPv6(src, dst []byte, proto byte, l4Len int) uint32 {
+	var sum uint32
+	for i := 0; i < 16; i += 2 {
+		sum += uint32(binary.BigEndian.Uint16(src[i : i+2]))
+		sum += uint32(binary.BigEndian.Uint16(dst[i : i+2]))
+	}
+	sum += uint32(l4Len >> 16)
+	sum += uint32(l4Len & 0xffff)
+	sum += uint32(proto)
+	return sum
+}
+
+// foldOnceNoInvert folds the 32-bit accumulator to 16 bits and returns it
+// unchanged (no one's complement). This is what virtio NEEDS_CSUM wants in
+// the L4 checksum field — the kernel will add the payload sum and invert.
+func foldOnceNoInvert(sum uint32) uint16 {
+	for sum>>16 != 0 {
+		sum = (sum & 0xffff) + (sum >> 16)
+	}
+	return uint16(sum)
+}

overlay/coalesce/tcp_coalesce_test.go

@@ -0,0 +1,576 @@
+package coalesce
+
+import (
+	"encoding/binary"
+	"testing"
+)
+
+// fakeTunWriter records plain Writes and WriteGSO calls without touching a
+// real TUN fd. WriteGSO preserves the split between hdr and borrowed pays
+// so tests can inspect each independently.
+type fakeTunWriter struct {
+	gsoEnabled bool
+	writes     [][]byte
+	gsoWrites  []fakeGSOWrite
+}
+
+type fakeGSOWrite struct {
+	hdr       []byte
+	pays      [][]byte
+	gsoSize   uint16
+	isV6      bool
+	csumStart uint16
+}
+
+// total returns hdrLen + sum of pay lens.
+func (g fakeGSOWrite) total() int {
+	n := len(g.hdr)
+	for _, p := range g.pays {
+		n += len(p)
+	}
+	return n
+}
+
+// payLen sums the pays.
+func (g fakeGSOWrite) payLen() int {
+	var n int
+	for _, p := range g.pays {
+		n += len(p)
+	}
+	return n
+}
+
+func (w *fakeTunWriter) Write(p []byte) (int, error) {
+	buf := make([]byte, len(p))
+	copy(buf, p)
+	w.writes = append(w.writes, buf)
+	return len(p), nil
+}
+
+func (w *fakeTunWriter) WriteGSO(hdr []byte, pays [][]byte, gsoSize uint16, isV6 bool, csumStart uint16) error {
+	hcopy := make([]byte, len(hdr))
+	copy(hcopy, hdr)
+	paysCopy := make([][]byte, len(pays))
+	for i, p := range pays {
+		pc := make([]byte, len(p))
+		copy(pc, p)
+		paysCopy[i] = pc
+	}
+	w.gsoWrites = append(w.gsoWrites, fakeGSOWrite{
+		hdr:       hcopy,
+		pays:      paysCopy,
+		gsoSize:   gsoSize,
+		isV6:      isV6,
+		csumStart: csumStart,
+	})
+	return nil
+}
+
+func (w *fakeTunWriter) GSOSupported() bool { return w.gsoEnabled }
+
+// buildTCPv4 constructs a minimal IPv4+TCP packet with the given payload,
+// seq, and flags. Assumes no IP options and a 20-byte TCP header.
+func buildTCPv4(seq uint32, flags byte, payload []byte) []byte {
+	return buildTCPv4Ports(1000, 2000, seq, flags, payload)
+}
+
+// buildTCPv4Ports is buildTCPv4 with caller-specified ports so tests can
+// build distinct flows.
+func buildTCPv4Ports(sport, dport uint16, seq uint32, flags byte, payload []byte) []byte {
+	const ipHdrLen = 20
+	const tcpHdrLen = 20
+	total := ipHdrLen + tcpHdrLen + len(payload)
+	pkt := make([]byte, total)
+
+	pkt[0] = 0x45
+	pkt[1] = 0x00
+	binary.BigEndian.PutUint16(pkt[2:4], uint16(total))
+	binary.BigEndian.PutUint16(pkt[4:6], 0)
+	binary.BigEndian.PutUint16(pkt[6:8], 0x4000)
+	pkt[8] = 64
+	pkt[9] = ipProtoTCP
+	copy(pkt[12:16], []byte{10, 0, 0, 1})
+	copy(pkt[16:20], []byte{10, 0, 0, 2})
+
+	binary.BigEndian.PutUint16(pkt[20:22], sport)
+	binary.BigEndian.PutUint16(pkt[22:24], dport)
+	binary.BigEndian.PutUint32(pkt[24:28], seq)
+	binary.BigEndian.PutUint32(pkt[28:32], 12345)
+	pkt[32] = 0x50
+	pkt[33] = flags
+	binary.BigEndian.PutUint16(pkt[34:36], 0xffff)
+
+	copy(pkt[40:], payload)
+	return pkt
+}
+
+const (
+	tcpAck    = 0x10
+	tcpPsh    = 0x08
+	tcpSyn    = 0x02
+	tcpFin    = 0x01
+	tcpAckPsh = tcpAck | tcpPsh
+)
+
+func TestCoalescerPassthroughWhenGSOUnavailable(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: false}
+	c := NewTCPCoalescer(w)
+	pkt := buildTCPv4(1000, tcpAck, []byte("hello"))
+	if err := c.Add(pkt); err != nil {
+		t.Fatal(err)
+	}
+	// No sync write — passthrough is deferred to Flush.
+	if len(w.writes) != 0 || len(w.gsoWrites) != 0 {
+		t.Fatalf("no Add-time writes: got writes=%d gso=%d", len(w.writes), len(w.gsoWrites))
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	if len(w.writes) != 1 || len(w.gsoWrites) != 0 {
+		t.Fatalf("want single plain write, got writes=%d gso=%d", len(w.writes), len(w.gsoWrites))
+	}
+}
+
+func TestCoalescerNonTCPPassthrough(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := NewTCPCoalescer(w)
+	pkt := make([]byte, 28)
+	pkt[0] = 0x45
+	binary.BigEndian.PutUint16(pkt[2:4], 28)
+	pkt[9] = 1
+	copy(pkt[12:16], []byte{10, 0, 0, 1})
+	copy(pkt[16:20], []byte{10, 0, 0, 2})
+	if err := c.Add(pkt); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	if len(w.writes) != 1 || len(w.gsoWrites) != 0 {
+		t.Fatalf("ICMP should pass through unchanged")
+	}
+}
+
+func TestCoalescerSeedThenFlushAlone(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := NewTCPCoalescer(w)
+	pkt := buildTCPv4(1000, tcpAck, make([]byte, 1000))
+	if err := c.Add(pkt); err != nil {
+		t.Fatal(err)
+	}
+	if len(w.writes) != 0 || len(w.gsoWrites) != 0 {
+		t.Fatalf("unexpected output before flush")
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	// Single-segment flush now goes through WriteGSO with GSO_NONE
+	// (virtio NEEDS_CSUM lets the kernel fill in the L4 csum).
+	if len(w.gsoWrites) != 1 || len(w.writes) != 0 {
+		t.Fatalf("single-seg flush: writes=%d gso=%d", len(w.writes), len(w.gsoWrites))
+	}
+	g := w.gsoWrites[0]
+	if g.total() != 40+1000 {
+		t.Errorf("super total=%d want %d", g.total(), 40+1000)
+	}
+	if g.payLen() != 1000 {
+		t.Errorf("payLen=%d want 1000", g.payLen())
+	}
+}
+
+func TestCoalescerCoalescesAdjacentACKs(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := NewTCPCoalescer(w)
+	pay := make([]byte, 1200)
+	if err := c.Add(buildTCPv4(1000, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Add(buildTCPv4(2200, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Add(buildTCPv4(3400, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	if len(w.gsoWrites) != 1 {
+		t.Fatalf("want 1 gso write, got %d (plain=%d)", len(w.gsoWrites), len(w.writes))
+	}
+	g := w.gsoWrites[0]
+	if g.gsoSize != 1200 {
+		t.Errorf("gsoSize=%d want 1200", g.gsoSize)
+	}
+	if len(g.hdr) != 40 {
+		t.Errorf("hdrLen=%d want 40", len(g.hdr))
+	}
+	if g.csumStart != 20 {
+		t.Errorf("csumStart=%d want 20", g.csumStart)
+	}
+	if len(g.pays) != 3 {
+		t.Errorf("pay count=%d want 3", len(g.pays))
+	}
+	if g.total() != 40+3*1200 {
+		t.Errorf("superpacket len=%d want %d", g.total(), 40+3*1200)
+	}
+	if tot := binary.BigEndian.Uint16(g.hdr[2:4]); int(tot) != g.total() {
+		t.Errorf("ip total_length=%d want %d", tot, g.total())
+	}
+}
+
+func TestCoalescerRejectsSeqGap(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := NewTCPCoalescer(w)
+	pay := make([]byte, 1200)
+	if err := c.Add(buildTCPv4(1000, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Add(buildTCPv4(3000, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	// Each packet flushes as its own single-segment WriteGSO now.
+	if len(w.gsoWrites) != 2 || len(w.writes) != 0 {
+		t.Fatalf("seq gap: want 2 gso writes got writes=%d gso=%d", len(w.writes), len(w.gsoWrites))
+	}
+}
+
+func TestCoalescerRejectsFlagMismatch(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := NewTCPCoalescer(w)
+	pay := make([]byte, 1200)
+	if err := c.Add(buildTCPv4(1000, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	// SYN|ACK is non-admissible. Must flush matching flow's slot (gso)
+	// and then plain-write the SYN packet itself.
+	syn := buildTCPv4(2200, tcpSyn|tcpAck, pay)
+	if err := c.Add(syn); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	if len(w.writes) != 1 || len(w.gsoWrites) != 1 {
+		t.Fatalf("flag mismatch: want 1 plain + 1 gso, got writes=%d gso=%d", len(w.writes), len(w.gsoWrites))
+	}
+}
+
+func TestCoalescerRejectsFIN(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := NewTCPCoalescer(w)
+	fin := buildTCPv4(1000, tcpAck|tcpFin, []byte("x"))
+	if err := c.Add(fin); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	// FIN isn't admissible — passthrough as plain, no slot, no gso.
+	if len(w.writes) != 1 || len(w.gsoWrites) != 0 {
+		t.Fatalf("FIN should be passthrough, got writes=%d gso=%d", len(w.writes), len(w.gsoWrites))
+	}
+}
+
+func TestCoalescerShortLastSegmentClosesChain(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := NewTCPCoalescer(w)
+	full := make([]byte, 1200)
+	half := make([]byte, 500)
+	if err := c.Add(buildTCPv4(1000, tcpAck, full)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Add(buildTCPv4(2200, tcpAck, half)); err != nil {
+		t.Fatal(err)
+	}
+	// Chain now closed; next packet seeds a new slot on the same flow
+	// after flushing the old one.
+	if err := c.Add(buildTCPv4(2700, tcpAck, full)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	// Expect two gso writes: the first two packets coalesced, then the
+	// third flushed alone (single-seg via GSO_NONE).
+	if len(w.gsoWrites) != 2 {
+		t.Fatalf("want 2 gso writes got %d", len(w.gsoWrites))
+	}
+	if len(w.writes) != 0 {
+		t.Fatalf("want 0 plain writes got %d", len(w.writes))
+	}
+	if w.gsoWrites[0].gsoSize != 1200 {
+		t.Errorf("gsoSize=%d want 1200", w.gsoWrites[0].gsoSize)
+	}
+	if got, want := w.gsoWrites[0].total(), 40+1200+500; got != want {
+		t.Errorf("super len=%d want %d", got, want)
+	}
+}
+
+func TestCoalescerPSHFinalizesChain(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := NewTCPCoalescer(w)
+	pay := make([]byte, 1200)
+	if err := c.Add(buildTCPv4(1000, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Add(buildTCPv4(2200, tcpAckPsh, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Add(buildTCPv4(3400, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	// First two coalesce; the third seeds a fresh slot that flushes alone.
+	if len(w.gsoWrites) != 2 {
+		t.Fatalf("want 2 gso writes got %d", len(w.gsoWrites))
+	}
+	if len(w.writes) != 0 {
+		t.Fatalf("want 0 plain writes got %d", len(w.writes))
+	}
+}
+
+func TestCoalescerRejectsDifferentFlow(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := NewTCPCoalescer(w)
+	pay := make([]byte, 1200)
+	p1 := buildTCPv4(1000, tcpAck, pay)
+	p2 := buildTCPv4(2200, tcpAck, pay)
+	binary.BigEndian.PutUint16(p2[20:22], 9999)
+	if err := c.Add(p1); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Add(p2); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	// Two independent flows, each flushes its own single-segment WriteGSO.
+	if len(w.gsoWrites) != 2 || len(w.writes) != 0 {
+		t.Fatalf("diff flow: want 2 gso writes got writes=%d gso=%d", len(w.writes), len(w.gsoWrites))
+	}
+}
+
+func TestCoalescerRejectsIPOptions(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := NewTCPCoalescer(w)
+	pay := make([]byte, 500)
+	pkt := buildTCPv4(1000, tcpAck, pay)
+	// Bump IHL to 6 to simulate 4 bytes of IP options. Don't actually add
+	// bytes — parser should bail before it matters.
+	pkt[0] = 0x46
+	if err := c.Add(pkt); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	// Non-admissible parse → passthrough as plain.
+	if len(w.writes) != 1 || len(w.gsoWrites) != 0 {
+		t.Fatalf("IP options should passthrough, got writes=%d gso=%d", len(w.writes), len(w.gsoWrites))
+	}
+}
+
+func TestCoalescerCapBySegments(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := NewTCPCoalescer(w)
+	pay := make([]byte, 512)
+	seq := uint32(1000)
+	for i := 0; i < tcpCoalesceMaxSegs+5; i++ {
+		if err := c.Add(buildTCPv4(seq, tcpAck, pay)); err != nil {
+			t.Fatal(err)
+		}
+		seq += uint32(len(pay))
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	for _, g := range w.gsoWrites {
+		segs := len(g.pays)
+		if segs > tcpCoalesceMaxSegs {
+			t.Fatalf("super exceeded seg cap: %d > %d", segs, tcpCoalesceMaxSegs)
+		}
+	}
+}
+
+// TestCoalescerMultipleFlowsInSameBatch proves two interleaved bulk TCP
+// flows coalesce independently in a single Flush.
+func TestCoalescerMultipleFlowsInSameBatch(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := NewTCPCoalescer(w)
+	pay := make([]byte, 1200)
+
+	// Flow A: sport 1000. Flow B: sport 3000.
+	if err := c.Add(buildTCPv4Ports(1000, 2000, 100, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Add(buildTCPv4Ports(3000, 2000, 500, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Add(buildTCPv4Ports(1000, 2000, 1300, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Add(buildTCPv4Ports(3000, 2000, 1700, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Add(buildTCPv4Ports(1000, 2000, 2500, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Add(buildTCPv4Ports(3000, 2000, 2900, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+
+	if len(w.gsoWrites) != 2 {
+		t.Fatalf("want 2 gso writes (one per flow), got %d", len(w.gsoWrites))
+	}
+	if len(w.writes) != 0 {
+		t.Fatalf("want no plain writes, got %d", len(w.writes))
+	}
+	// Each superpacket should carry 3 segments.
+	for i, g := range w.gsoWrites {
+		if len(g.pays) != 3 {
+			t.Errorf("gso[%d]: segs=%d want 3", i, len(g.pays))
+		}
+		if g.gsoSize != 1200 {
+			t.Errorf("gso[%d]: gsoSize=%d want 1200", i, g.gsoSize)
+		}
+	}
+	// Verify each superpacket carries the source port it was seeded with.
+	seenSports := map[uint16]bool{}
+	for _, g := range w.gsoWrites {
+		sp := binary.BigEndian.Uint16(g.hdr[20:22])
+		seenSports[sp] = true
+	}
+	if !seenSports[1000] || !seenSports[3000] {
+		t.Errorf("expected superpackets for sports 1000 and 3000, got %v", seenSports)
+	}
+}
+
+// TestCoalescerPreservesArrivalOrder confirms that with passthrough and
+// coalesced events both queued, Flush emits them in Add order rather than
+// writing passthrough packets synchronously.
+func TestCoalescerPreservesArrivalOrder(t *testing.T) {
+	w := &orderedFakeWriter{gsoEnabled: true}
+	c := NewTCPCoalescer(w)
+	// Sequence: coalesceable TCP, ICMP (passthrough), coalesceable TCP on
+	// a different flow. Expected emit order: gso(X), plain(ICMP), gso(Y).
+	pay := make([]byte, 1200)
+	if err := c.Add(buildTCPv4Ports(1000, 2000, 100, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	icmp := make([]byte, 28)
+	icmp[0] = 0x45
+	binary.BigEndian.PutUint16(icmp[2:4], 28)
+	icmp[9] = 1
+	copy(icmp[12:16], []byte{10, 0, 0, 1})
+	copy(icmp[16:20], []byte{10, 0, 0, 3})
+	if err := c.Add(icmp); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Add(buildTCPv4Ports(3000, 2000, 500, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	// Nothing should have hit the writer synchronously.
+	if len(w.events) != 0 {
+		t.Fatalf("Add emitted events synchronously: %v", w.events)
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	if got, want := w.events, []string{"gso", "plain", "gso"}; !stringSliceEq(got, want) {
+		t.Fatalf("flush order=%v want %v", got, want)
+	}
+}
+
+// orderedFakeWriter records only the sequence of call types so tests can
+// assert arrival order without inspecting bytes.
+type orderedFakeWriter struct {
+	gsoEnabled bool
+	events     []string
+}
+
+func (w *orderedFakeWriter) Write(p []byte) (int, error) {
+	w.events = append(w.events, "plain")
+	return len(p), nil
+}
+
+func (w *orderedFakeWriter) WriteGSO(hdr []byte, pays [][]byte, gsoSize uint16, isV6 bool, csumStart uint16) error {
+	w.events = append(w.events, "gso")
+	return nil
+}
+
+func (w *orderedFakeWriter) GSOSupported() bool { return w.gsoEnabled }
+
+func stringSliceEq(a, b []string) bool {
+	if len(a) != len(b) {
+		return false
+	}
+	for i := range a {
+		if a[i] != b[i] {
+			return false
+		}
+	}
+	return true
+}
+
+// TestCoalescerInterleavedFlowsPreserveOrdering checks that a non-admissible
+// packet (SYN) mid-flow only flushes its own flow, not others.
+func TestCoalescerInterleavedFlowsPreserveOrdering(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := NewTCPCoalescer(w)
+	pay := make([]byte, 1200)
+
+	// Flow A two segments.
+	if err := c.Add(buildTCPv4Ports(1000, 2000, 100, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Add(buildTCPv4Ports(1000, 2000, 1300, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	// Flow B two segments.
+	if err := c.Add(buildTCPv4Ports(3000, 2000, 500, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Add(buildTCPv4Ports(3000, 2000, 1700, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	// Flow A SYN (non-admissible) — must flush only flow A's slot.
+	syn := buildTCPv4Ports(1000, 2000, 9999, tcpSyn|tcpAck, pay)
+	if err := c.Add(syn); err != nil {
+		t.Fatal(err)
+	}
+	// Flow B continues — should still be coalesced with its seed.
+	if err := c.Add(buildTCPv4Ports(3000, 2000, 2900, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+
+	// Expected:
+	//   - 1 gso for flow A (first 2 segments)
+	//   - 1 plain for flow A SYN
+	//   - 1 gso for flow B (3 segments)
+	if len(w.gsoWrites) != 2 {
+		t.Fatalf("want 2 gso writes, got %d", len(w.gsoWrites))
+	}
+	if len(w.writes) != 1 {
+		t.Fatalf("want 1 plain write (SYN), got %d", len(w.writes))
+	}
+	// Find the 3-segment gso (flow B) and the 2-segment gso (flow A).
+	var segCounts []int
+	for _, g := range w.gsoWrites {
+		segCounts = append(segCounts, len(g.pays))
+	}
+	if !(segCounts[0] == 2 && segCounts[1] == 3) && !(segCounts[0] == 3 && segCounts[1] == 2) {
+		t.Errorf("unexpected segment counts: %v (want 2 and 3)", segCounts)
+	}
+}

overlay/tio/container_gso_linux.go

@@ -0,0 +1,70 @@
+package tio
+
+import (
+	"encoding/binary"
+	"errors"
+	"fmt"
+
+	"golang.org/x/sys/unix"
+)
+
+type offloadContainer struct {
+	pq []*Offload
+	// pqi is exactly the same as pq, but stored as the interface type
+	pqi        []Queue
+	shutdownFd int
+}
+
+func NewOffloadContainer() (Container, error) {
+	shutdownFd, err := unix.Eventfd(0, unix.EFD_NONBLOCK|unix.EFD_CLOEXEC)
+	if err != nil {
+		return nil, fmt.Errorf("failed to create eventfd: %w", err)
+	}
+
+	out := &offloadContainer{
+		pq:         []*Offload{},
+		pqi:        []Queue{},
+		shutdownFd: shutdownFd,
+	}
+
+	return out, nil
+}
+
+func (c *offloadContainer) Queues() []Queue {
+	return c.pqi
+}
+
+func (c *offloadContainer) Add(fd int) error {
+	x, err := newOffload(fd, c.shutdownFd)
+	if err != nil {
+		return err
+	}
+	c.pq = append(c.pq, x)
+	c.pqi = append(c.pqi, x)
+
+	return nil
+}
+
+func (c *offloadContainer) wakeForShutdown() error {
+	var buf [8]byte
+	binary.NativeEndian.PutUint64(buf[:], 1)
+	_, err := unix.Write(c.shutdownFd, buf[:])
+	return err
+}
+
+func (c *offloadContainer) Close() error {
+	errs := []error{}
+
+	// Signal all readers blocked in poll to wake up and exit
+	if err := c.wakeForShutdown(); err != nil {
+		errs = append(errs, err)
+	}
+
+	for _, x := range c.pq {
+		if err := x.Close(); err != nil {
+			errs = append(errs, err)
+		}
+	}
+
+	return errors.Join(errs...)
+}

overlay/tio/tio.go

@@ -13,6 +13,8 @@ type Container interface {
 
 	// Add takes a tun fd, adds it to the container, and prepares it for use as a Queue
 	Add(fd int) error
+
+	io.Closer
 }
 
 // Queue is a readable/writable Poll queue. One Queue is driven by a single

overlay/tio/tio_gso_linux.go

@@ -0,0 +1,374 @@
+package tio
+
+import (
+	"fmt"
+	"io"
+	"os"
+	"sync/atomic"
+	"syscall"
+	"unsafe"
+
+	"golang.org/x/sys/unix"
+)
+
+// Space for segmented output. Worst case is many small segments, each paying
+// an IP+TCP header. Should be a multiple of 64KiB.
+// const tunSegBufSize = 0xffff * 8 TODO larger? config?
+const tunSegBufSize = 131072
+
+// tunSegBufCap is the total size we allocate for the per-reader segment
+// buffer. It is sized as one worst-case TSO superpacket (tunSegBufSize) plus
+// the same again as drain headroom so a Read wake can accumulate
+// additional packets after an initial big read without overflowing.
+const tunSegBufCap = tunSegBufSize * 2
+
+// tunDrainCap caps how many packets a single Read will accumulate via
+// the post-wake drain loop. Sized to soak up a burst of small ACKs while
+// bounding how much work a single caller holds before handing off.
+const tunDrainCap = 64 //256
+
+// gsoInitialPayIovs is the starting capacity (in payload fragments) of
+// Offload.gsoIovs. Sized to cover the default coalesce segment cap without
+// any reallocations.
+const gsoInitialPayIovs = 66
+
+// validVnetHdr is the 10-byte virtio_net_hdr we prepend to every non-GSO TUN
+// write. Only flag set is VIRTIO_NET_HDR_F_DATA_VALID, which marks the skb
+// CHECKSUM_UNNECESSARY so the receiving network stack skips L4 checksum
+// verification. All packets that reach the plain Write / WriteReject paths
+// already carry a valid L4 checksum (either supplied by a remote peer whose
+// ciphertext we AEAD-authenticated, or produced by finishChecksum during TSO
+// segmentation, or built locally by CreateRejectPacket), so trusting them is
+// safe.
+var validVnetHdr = [virtioNetHdrLen]byte{unix.VIRTIO_NET_HDR_F_DATA_VALID}
+
+// Offload wraps a TUN file descriptor with poll-based reads. The FD provided will be changed to non-blocking.
+// A shared eventfd allows Close to wake all readers blocked in poll.
+type Offload struct {
+	fd         int
+	shutdownFd int
+	readPoll   [2]unix.PollFd
+	writePoll  [2]unix.PollFd
+	closed     atomic.Bool
+	readBuf    []byte        // scratch for a single raw read (virtio hdr + superpacket)
+	segBuf     []byte        // backing store for segmented output
+	segOff     int           // cursor into segBuf for the current Read drain
+	pending    [][]byte      // segments returned from the most recent Read
+	writeIovs  [2]unix.Iovec // preallocated iovecs for Write (coalescer passthrough); iovs[0] is fixed to validVnetHdr
+	// rejectIovs is a second preallocated iovec scratch used exclusively by
+	// WriteReject (reject + self-forward from the inside path). It mirrors
+	// writeIovs but lets listenIn goroutines emit reject packets without
+	// racing with the listenOut coalescer that owns writeIovs.
+	rejectIovs [2]unix.Iovec
+
+	// gsoHdrBuf is a per-queue 10-byte scratch for the virtio_net_hdr emitted
+	// by WriteGSO. Separate from validVnetHdr so a concurrent non-GSO Write on
+	// another queue never observes a half-written header.
+	gsoHdrBuf [virtioNetHdrLen]byte
+	// gsoIovs is the writev iovec scratch for WriteGSO. Sized to hold the
+	// virtio header + IP/TCP header + up to gsoInitialPayIovs payload
+	// fragments; grown on demand if a coalescer pushes more.
+	gsoIovs []unix.Iovec
+}
+
+func newOffload(fd int, shutdownFd int) (*Offload, error) {
+	if err := unix.SetNonblock(fd, true); err != nil {
+		return nil, fmt.Errorf("failed to set tun fd non-blocking: %w", err)
+	}
+
+	out := &Offload{
+		fd:         fd,
+		shutdownFd: shutdownFd,
+		closed:     atomic.Bool{},
+		readBuf:    make([]byte, tunReadBufSize),
+		readPoll: [2]unix.PollFd{
+			{Fd: int32(fd), Events: unix.POLLIN},
+			{Fd: int32(shutdownFd), Events: unix.POLLIN},
+		},
+		writePoll: [2]unix.PollFd{
+			{Fd: int32(fd), Events: unix.POLLOUT},
+			{Fd: int32(shutdownFd), Events: unix.POLLIN},
+		},
+
+		segBuf:  make([]byte, tunSegBufCap),
+		gsoIovs: make([]unix.Iovec, 2, 2+gsoInitialPayIovs),
+	}
+
+	out.writeIovs[0].Base = &validVnetHdr[0]
+	out.writeIovs[0].SetLen(virtioNetHdrLen)
+	out.rejectIovs[0].Base = &validVnetHdr[0]
+	out.rejectIovs[0].SetLen(virtioNetHdrLen)
+	out.gsoIovs[0].Base = &out.gsoHdrBuf[0]
+	out.gsoIovs[0].SetLen(virtioNetHdrLen)
+
+	return out, nil
+}
+
+func (r *Offload) blockOnRead() error {
+	const problemFlags = unix.POLLHUP | unix.POLLNVAL | unix.POLLERR
+	var err error
+	for {
+		_, err = unix.Poll(r.readPoll[:], -1)
+		if err != unix.EINTR {
+			break
+		}
+	}
+	//always reset these!
+	tunEvents := r.readPoll[0].Revents
+	shutdownEvents := r.readPoll[1].Revents
+	r.readPoll[0].Revents = 0
+	r.readPoll[1].Revents = 0
+	//do the err check before trusting the potentially bogus bits we just got
+	if err != nil {
+		return err
+	}
+	if shutdownEvents&(unix.POLLIN|problemFlags) != 0 {
+		return os.ErrClosed
+	} else if tunEvents&problemFlags != 0 {
+		return os.ErrClosed
+	}
+	return nil
+}
+
+func (r *Offload) blockOnWrite() error {
+	const problemFlags = unix.POLLHUP | unix.POLLNVAL | unix.POLLERR
+	var err error
+	for {
+		_, err = unix.Poll(r.writePoll[:], -1)
+		if err != unix.EINTR {
+			break
+		}
+	}
+	//always reset these!
+	tunEvents := r.writePoll[0].Revents
+	shutdownEvents := r.writePoll[1].Revents
+	r.writePoll[0].Revents = 0
+	r.writePoll[1].Revents = 0
+	//do the err check before trusting the potentially bogus bits we just got
+	if err != nil {
+		return err
+	}
+	if shutdownEvents&(unix.POLLIN|problemFlags) != 0 {
+		return os.ErrClosed
+	} else if tunEvents&problemFlags != 0 {
+		return os.ErrClosed
+	}
+	return nil
+}
+
+func (r *Offload) readRaw(buf []byte) (int, error) {
+	for {
+		if n, err := unix.Read(r.fd, buf); err == nil {
+			return n, nil
+		} else if err == unix.EAGAIN {
+			if err = r.blockOnRead(); err != nil {
+				return 0, err
+			}
+			continue
+		} else if err == unix.EINTR {
+			continue
+		} else if err == unix.EBADF {
+			return 0, os.ErrClosed
+		} else {
+			return 0, err
+		}
+	}
+}
+
+// Read reads one or more superpackets from the tun and returns the
+// resulting packets. The first read blocks via poll; once the fd is known
+// readable we drain additional packets non-blocking until the kernel queue
+// is empty (EAGAIN), we've collected tunDrainCap packets, or we're out of
+// segBuf headroom. This amortizes the poll wake over bursts of small
+// packets (e.g. TCP ACKs). Slices point into the Offload's internal buffers
+// and are only valid until the next Read or Close on this Queue.
+func (r *Offload) Read() ([][]byte, error) {
+	r.pending = r.pending[:0]
+	r.segOff = 0
+
+	// Initial (blocking) read. Retry on decode errors so a single bad
+	// packet does not stall the reader.
+	for {
+		n, err := r.readRaw(r.readBuf)
+		if err != nil {
+			return nil, err
+		}
+		if err := r.decodeRead(n); err != nil {
+			// Drop and read again — a bad packet should not kill the reader.
+			continue
+		}
+		break
+	}
+
+	// Drain: non-blocking reads until the kernel queue is empty, the drain
+	// cap is reached, or segBuf no longer has room for another worst-case
+	// superpacket.
+	for len(r.pending) < tunDrainCap && tunSegBufCap-r.segOff >= tunSegBufSize {
+		n, err := unix.Read(r.fd, r.readBuf)
+		if err != nil {
+			// EAGAIN / EINTR / anything else: stop draining. We already
+			// have a valid batch from the first read.
+			break
+		}
+		if n <= 0 {
+			break
+		}
+		if err := r.decodeRead(n); err != nil {
+			// Drop this packet and stop the drain; we'd rather hand off
+			// what we have than keep spinning here.
+			break
+		}
+	}
+
+	return r.pending, nil
+}
+
+// decodeRead decodes the virtio header plus payload in r.readBuf[:n], appends
+// the segments to r.pending, and advances r.segOff by the total scratch used.
+// Caller must have already ensured r.vnetHdr is true.
+func (r *Offload) decodeRead(n int) error {
+	if n < virtioNetHdrLen {
+		return fmt.Errorf("short tun read: %d < %d", n, virtioNetHdrLen)
+	}
+	var hdr VirtioNetHdr
+	hdr.decode(r.readBuf[:virtioNetHdrLen])
+	before := len(r.pending)
+	if err := segmentInto(r.readBuf[virtioNetHdrLen:n], hdr, &r.pending, r.segBuf[r.segOff:]); err != nil {
+		return err
+	}
+	for k := before; k < len(r.pending); k++ {
+		r.segOff += len(r.pending[k])
+	}
+	return nil
+}
+
+func (r *Offload) Write(buf []byte) (int, error) {
+	return r.writeWithScratch(buf, &r.writeIovs)
+}
+
+// WriteReject emits a packet using a dedicated iovec scratch (rejectIovs)
+// distinct from the one used by the coalescer's Write path. This avoids a
+// data race between the inside (listenIn) goroutine emitting reject or
+// self-forward packets and the outside (listenOut) goroutine flushing TCP
+// coalescer passthroughs on the same Offload.
+func (r *Offload) WriteReject(buf []byte) (int, error) {
+	return r.writeWithScratch(buf, &r.rejectIovs)
+}
+
+func (r *Offload) writeWithScratch(buf []byte, iovs *[2]unix.Iovec) (int, error) {
+	if len(buf) == 0 {
+		return 0, nil
+	}
+	// Point the payload iovec at the caller's buffer. iovs[0] is pre-wired
+	// to validVnetHdr during Offload construction so we don't rebuild it here.
+	iovs[1].Base = &buf[0]
+	iovs[1].SetLen(len(buf))
+	iovPtr := unsafe.Pointer(&iovs[0])
+	return r.rawWrite(iovPtr, 2)
+}
+
+func (r *Offload) rawWrite(iovs unsafe.Pointer, iovcnt int) (int, error) {
+	for {
+		n, _, errno := syscall.Syscall(unix.SYS_WRITEV, uintptr(r.fd), uintptr(iovs), uintptr(iovcnt))
+		if errno == 0 {
+			if int(n) < virtioNetHdrLen {
+				return 0, io.ErrShortWrite
+			}
+			return int(n) - virtioNetHdrLen, nil
+		}
+		if errno == unix.EAGAIN {
+			if err := r.blockOnWrite(); err != nil {
+				return 0, err
+			}
+			continue
+		}
+		if errno == unix.EINTR {
+			continue
+		}
+		if errno == unix.EBADF {
+			return 0, os.ErrClosed
+		}
+		return 0, errno
+	}
+}
+
+// GSOSupported reports whether this queue was opened with IFF_VNET_HDR and
+// can accept WriteGSO. When false, callers should fall back to per-segment
+// Write calls.
+func (r *Offload) GSOSupported() bool { return true }
+
+// WriteGSO emits a TCP TSO superpacket in a single writev. hdr is the
+// IPv4/IPv6 + TCP header prefix (already finalized — total length, IP csum,
+// and TCP pseudo-header partial set by the caller). pays are payload
+// fragments whose concatenation forms the full coalesced payload; each
+// slice is read-only and must stay valid until return. gsoSize is the MSS;
+// every segment except possibly the last is exactly gsoSize bytes.
+// csumStart is the byte offset where the TCP header begins within hdr.
+func (r *Offload) WriteGSO(hdr []byte, pays [][]byte, gsoSize uint16, isV6 bool, csumStart uint16) error {
+	if len(hdr) == 0 || len(pays) == 0 {
+		return nil
+	}
+
+	// Build the virtio_net_hdr. When pays total to <= gsoSize the kernel
+	// would produce a single segment; keep NEEDS_CSUM semantics but skip
+	// the GSO type so the kernel doesn't spuriously mark this as TSO.
+	vhdr := VirtioNetHdr{
+		Flags:      unix.VIRTIO_NET_HDR_F_NEEDS_CSUM,
+		HdrLen:     uint16(len(hdr)),
+		GSOSize:    gsoSize,
+		CsumStart:  csumStart,
+		CsumOffset: 16, // TCP checksum field lives 16 bytes into the TCP header
+	}
+	var totalPay int
+	for _, p := range pays {
+		totalPay += len(p)
+	}
+	if totalPay > int(gsoSize) {
+		if isV6 {
+			vhdr.GSOType = unix.VIRTIO_NET_HDR_GSO_TCPV6
+		} else {
+			vhdr.GSOType = unix.VIRTIO_NET_HDR_GSO_TCPV4
+		}
+	} else {
+		vhdr.GSOType = unix.VIRTIO_NET_HDR_GSO_NONE
+		vhdr.GSOSize = 0
+	}
+	vhdr.encode(r.gsoHdrBuf[:])
+
+	// Build the iovec array: [virtio_hdr, hdr, pays...]. r.gsoIovs[0] is
+	// wired to gsoHdrBuf at construction and never changes.
+	need := 2 + len(pays)
+	if cap(r.gsoIovs) < need {
+		grown := make([]unix.Iovec, need)
+		grown[0] = r.gsoIovs[0]
+		r.gsoIovs = grown
+	} else {
+		r.gsoIovs = r.gsoIovs[:need]
+	}
+	r.gsoIovs[1].Base = &hdr[0]
+	r.gsoIovs[1].SetLen(len(hdr))
+	for i, p := range pays {
+		r.gsoIovs[2+i].Base = &p[0]
+		r.gsoIovs[2+i].SetLen(len(p))
+	}
+
+	iovPtr := unsafe.Pointer(&r.gsoIovs[0])
+	iovCnt := len(r.gsoIovs)
+	_, err := r.rawWrite(iovPtr, iovCnt)
+	return err
+}
+
+func (r *Offload) Close() error {
+	if r.closed.Swap(true) {
+		return nil
+	}
+
+	//shutdownFd is owned by the container, so we should not close it
+	var err error
+	if r.fd >= 0 {
+		err = unix.Close(r.fd)
+		r.fd = -1
+	}
+
+	return err
+}

overlay/tio/tun_linux_offload.go

@@ -0,0 +1,239 @@
+//go:build linux && !android && !e2e_testing
+// +build linux,!android,!e2e_testing
+
+package tio
+
+import (
+	"encoding/binary"
+	"fmt"
+
+	"golang.org/x/sys/unix"
+	"gvisor.dev/gvisor/pkg/tcpip/checksum"
+)
+
+// segmentInto splits a TUN-side packet described by hdr into one or more
+// IP packets, each appended to *out as a slice of scratch. scratch must be
+// sized to hold every segment (including replicated headers).
+func segmentInto(pkt []byte, hdr VirtioNetHdr, out *[][]byte, scratch []byte) error {
+	// When RSC_INFO is set the csum_start/csum_offset fields are repurposed to
+	// carry coalescing info rather than checksum offsets. A TUN writing via
+	// IFF_VNET_HDR should never emit this, but if it did we would silently
+	// miscompute the segment checksums — refuse the packet instead.
+	if hdr.Flags&unix.VIRTIO_NET_HDR_F_RSC_INFO != 0 {
+		return fmt.Errorf("virtio RSC_INFO flag not supported on TUN reads")
+	}
+
+	switch hdr.GSOType {
+	case unix.VIRTIO_NET_HDR_GSO_NONE:
+		if len(pkt) > len(scratch) {
+			return fmt.Errorf("packet larger than segment buffer: %d > %d", len(pkt), len(scratch))
+		}
+		copy(scratch, pkt)
+		seg := scratch[:len(pkt)]
+		if hdr.Flags&unix.VIRTIO_NET_HDR_F_NEEDS_CSUM != 0 {
+			if err := finishChecksum(seg, hdr); err != nil {
+				return err
+			}
+		}
+		*out = append(*out, seg)
+		return nil
+
+	case unix.VIRTIO_NET_HDR_GSO_TCPV4, unix.VIRTIO_NET_HDR_GSO_TCPV6:
+		return segmentTCP(pkt, hdr, out, scratch)
+
+	default:
+		return fmt.Errorf("unsupported virtio gso type: %d", hdr.GSOType)
+	}
+}
+
+// finishChecksum computes the L4 checksum for a non-GSO packet that the kernel
+// handed us with NEEDS_CSUM set. csum_start / csum_offset point at the 16-bit
+// checksum field; we zero it, fold a full sum (the field was pre-loaded with
+// the pseudo-header partial sum by the kernel), and store the result.
+func finishChecksum(seg []byte, hdr VirtioNetHdr) error {
+	cs := int(hdr.CsumStart)
+	co := int(hdr.CsumOffset)
+	if cs+co+2 > len(seg) {
+		return fmt.Errorf("csum offsets out of range: start=%d offset=%d len=%d", cs, co, len(seg))
+	}
+	// The kernel stores a partial pseudo-header sum at [cs+co:]; sum over the
+	// L4 region starting at cs, folding the prior partial in as the seed.
+	partial := binary.BigEndian.Uint16(seg[cs+co : cs+co+2])
+	seg[cs+co] = 0
+	seg[cs+co+1] = 0
+	binary.BigEndian.PutUint16(seg[cs+co:cs+co+2], ^checksum.Checksum(seg[cs:], partial))
+	return nil
+}
+
+// segmentTCP software-segments a TSO superpacket into one IP packet per MSS
+// chunk. The caller guarantees hdr.GSOType is TCPV4 or TCPV6.
+//
+// Hot-path shape: the per-segment loop only sums the payload chunk. The TCP
+// header, the IPv4 header, and the pseudo-header src/dst/proto contributions
+// are each summed once up front — every segment reuses those three pre-folded
+// uint32 values and combines them with small per-segment deltas (seq, flags,
+// tcpLen, ip_id, total_len) that are cheap to fold in.
+func segmentTCP(pkt []byte, hdr VirtioNetHdr, out *[][]byte, scratch []byte) error {
+	if hdr.GSOSize == 0 {
+		return fmt.Errorf("gso_size is zero")
+	}
+	if int(hdr.HdrLen) > len(pkt) || hdr.HdrLen == 0 {
+		return fmt.Errorf("hdr_len %d out of range (pkt %d)", hdr.HdrLen, len(pkt))
+	}
+	if hdr.CsumStart == 0 || hdr.CsumStart >= hdr.HdrLen {
+		return fmt.Errorf("csum_start %d out of range (hdr_len %d)", hdr.CsumStart, hdr.HdrLen)
+	}
+
+	isV4 := hdr.GSOType == unix.VIRTIO_NET_HDR_GSO_TCPV4
+	headerLen := int(hdr.HdrLen)
+	csumStart := int(hdr.CsumStart)
+
+	if isV4 && csumStart < 20 {
+		return fmt.Errorf("csum_start %d too small for IPv4", csumStart)
+	}
+	if !isV4 && csumStart < 40 {
+		return fmt.Errorf("csum_start %d too small for IPv6", csumStart)
+	}
+	tcpHdrLen := headerLen - csumStart
+	if tcpHdrLen < 20 {
+		return fmt.Errorf("tcp header region too small: %d", tcpHdrLen)
+	}
+
+	payload := pkt[headerLen:]
+	payLen := len(payload)
+	gso := int(hdr.GSOSize)
+	numSeg := (payLen + gso - 1) / gso
+	if numSeg == 0 {
+		numSeg = 1
+	}
+
+	need := numSeg*headerLen + payLen
+	if need > len(scratch) {
+		return fmt.Errorf("scratch too small for %d segments: need %d have %d", numSeg, need, len(scratch))
+	}
+
+	origSeq := binary.BigEndian.Uint32(pkt[csumStart+4 : csumStart+8])
+	origFlags := pkt[csumStart+13]
+	const tcpFinPsh = 0x09 // FIN(0x01) | PSH(0x08)
+
+	// Precompute the TCP header sum with seq/flags/csum zeroed. The max TCP
+	// header is 60 bytes; copy onto the stack, zero the per-segment-varying
+	// fields, sum once.
+	var tmp [60]byte
+	copy(tmp[:tcpHdrLen], pkt[csumStart:headerLen])
+	tmp[4], tmp[5], tmp[6], tmp[7] = 0, 0, 0, 0 // seq
+	tmp[13] = 0                                 // flags
+	tmp[16], tmp[17] = 0, 0                     // csum
+	baseTcpHdrSum := uint32(checksum.Checksum(tmp[:tcpHdrLen], 0))
+
+	// Pseudo-header src+dst+proto contribution (tcpLen varies per segment).
+	var baseProtoSum uint32
+	if isV4 {
+		baseProtoSum = uint32(checksum.Checksum(pkt[12:20], 0))
+	} else {
+		baseProtoSum = uint32(checksum.Checksum(pkt[8:40], 0))
+	}
+	baseProtoSum += uint32(unix.IPPROTO_TCP)
+
+	// Precompute IPv4 header sum with total_len/id/csum zeroed.
+	var origIPID uint16
+	var ihl int
+	var baseIPHdrSum uint32
+	if isV4 {
+		origIPID = binary.BigEndian.Uint16(pkt[4:6])
+		ihl = int(pkt[0]&0x0f) * 4
+		if ihl < 20 || ihl > csumStart {
+			return fmt.Errorf("bad IPv4 IHL: %d", ihl)
+		}
+		var ipTmp [60]byte
+		copy(ipTmp[:ihl], pkt[:ihl])
+		ipTmp[2], ipTmp[3] = 0, 0   // total_len
+		ipTmp[4], ipTmp[5] = 0, 0   // id
+		ipTmp[10], ipTmp[11] = 0, 0 // checksum
+		baseIPHdrSum = uint32(checksum.Checksum(ipTmp[:ihl], 0))
+	}
+
+	off := 0
+	for i := 0; i < numSeg; i++ {
+		segStart := i * gso
+		segEnd := segStart + gso
+		if segEnd > payLen {
+			segEnd = payLen
+		}
+		segPayLen := segEnd - segStart
+
+		copy(scratch[off:], pkt[:headerLen])
+		copy(scratch[off+headerLen:], payload[segStart:segEnd])
+		seg := scratch[off : off+headerLen+segPayLen]
+		off += headerLen + segPayLen
+
+		segSeq := origSeq + uint32(segStart)
+		segFlags := origFlags
+		if i != numSeg-1 {
+			segFlags = origFlags &^ tcpFinPsh
+		}
+		totalLen := headerLen + segPayLen
+
+		// Patch IP header and write the v4 header checksum from the precomputed base.
+		if isV4 {
+			segID := origIPID + uint16(i)
+			binary.BigEndian.PutUint16(seg[2:4], uint16(totalLen))
+			binary.BigEndian.PutUint16(seg[4:6], segID)
+			ipSum := baseIPHdrSum + uint32(totalLen) + uint32(segID)
+			binary.BigEndian.PutUint16(seg[10:12], foldComplement(ipSum))
+		} else {
+			// IPv6 payload length excludes the 40-byte fixed header but
+			// includes any extension headers between [40:csumStart].
+			binary.BigEndian.PutUint16(seg[4:6], uint16(headerLen-40+segPayLen))
+		}
+
+		// Patch TCP header.
+		binary.BigEndian.PutUint32(seg[csumStart+4:csumStart+8], segSeq)
+		seg[csumStart+13] = segFlags
+		// (csum is written below; its prior contents in `seg` don't affect the
+		// computation since we never sum over the segment's own header.)
+
+		tcpLen := tcpHdrLen + segPayLen
+		paySum := uint32(checksum.Checksum(payload[segStart:segEnd], 0))
+
+		// Combine pre-folded uint32s into a wider accumulator, then fold. Using
+		// uint64 guards against overflow when segSeq's high bits set.
+		wide := uint64(baseTcpHdrSum) + uint64(paySum) + uint64(baseProtoSum)
+		wide += uint64(segSeq) + uint64(segFlags) + uint64(tcpLen)
+		wide = (wide & 0xffffffff) + (wide >> 32)
+		wide = (wide & 0xffffffff) + (wide >> 32)
+		binary.BigEndian.PutUint16(seg[csumStart+16:csumStart+18], foldComplement(uint32(wide)))
+
+		*out = append(*out, seg)
+	}
+
+	return nil
+}
+
+// foldComplement folds a 32-bit one's-complement partial sum to 16 bits and
+// complements it, yielding the on-wire Internet checksum value.
+func foldComplement(sum uint32) uint16 {
+	sum = (sum & 0xffff) + (sum >> 16)
+	sum = (sum & 0xffff) + (sum >> 16)
+	return ^uint16(sum)
+}
+
+// pseudoHeaderIPv4 returns the folded pseudo-header sum used to verify a TCP
+// segment's checksum in tests. src/dst are 4 bytes each.
+func pseudoHeaderIPv4(src, dst []byte, proto byte, tcpLen int) uint16 {
+	s := uint32(checksum.Checksum(src, 0)) + uint32(checksum.Checksum(dst, 0))
+	s += uint32(proto) + uint32(tcpLen)
+	s = (s & 0xffff) + (s >> 16)
+	s = (s & 0xffff) + (s >> 16)
+	return uint16(s)
+}
+
+// pseudoHeaderIPv6 returns the folded pseudo-header sum used to verify a TCP
+// segment's checksum in tests. src/dst are 16 bytes each.
+func pseudoHeaderIPv6(src, dst []byte, proto byte, tcpLen int) uint16 {
+	s := uint32(checksum.Checksum(src, 0)) + uint32(checksum.Checksum(dst, 0))
+	s += uint32(tcpLen>>16) + uint32(tcpLen&0xffff) + uint32(proto)
+	s = (s & 0xffff) + (s >> 16)
+	s = (s & 0xffff) + (s >> 16)
+	return uint16(s)
+}

overlay/tio/tun_linux_offload_test.go

@@ -0,0 +1,330 @@
+//go:build linux && !android && !e2e_testing
+// +build linux,!android,!e2e_testing
+
+package tio
+
+import (
+	"encoding/binary"
+	"os"
+	"testing"
+
+	"golang.org/x/sys/unix"
+	"gvisor.dev/gvisor/pkg/tcpip/checksum"
+)
+
+// verifyChecksum confirms that the one's-complement sum across `b`, seeded
+// with a folded pseudo-header sum, equals all-ones (valid).
+func verifyChecksum(b []byte, pseudo uint16) bool {
+	return checksum.Checksum(b, pseudo) == 0xffff
+}
+
+// buildTSOv4 builds a synthetic IPv4/TCP TSO superpacket with a payload of
+// `payLen` bytes split at `mss`.
+func buildTSOv4(t *testing.T, payLen, mss int) ([]byte, VirtioNetHdr) {
+	t.Helper()
+	const ipLen = 20
+	const tcpLen = 20
+	pkt := make([]byte, ipLen+tcpLen+payLen)
+
+	// IPv4 header
+	pkt[0] = 0x45 // version 4, IHL 5
+	// total length is meaningless for TSO but set it anyway
+	binary.BigEndian.PutUint16(pkt[2:4], uint16(ipLen+tcpLen+payLen))
+	binary.BigEndian.PutUint16(pkt[4:6], 0x4242) // original ID
+	pkt[8] = 64                                  // TTL
+	pkt[9] = unix.IPPROTO_TCP
+	copy(pkt[12:16], []byte{10, 0, 0, 1}) // src
+	copy(pkt[16:20], []byte{10, 0, 0, 2}) // dst
+
+	// TCP header
+	binary.BigEndian.PutUint16(pkt[20:22], 12345) // sport
+	binary.BigEndian.PutUint16(pkt[22:24], 80)    // dport
+	binary.BigEndian.PutUint32(pkt[24:28], 10000) // seq
+	binary.BigEndian.PutUint32(pkt[28:32], 20000) // ack
+	pkt[32] = 0x50                                // data offset 5 words
+	pkt[33] = 0x18                                // ACK | PSH
+	binary.BigEndian.PutUint16(pkt[34:36], 65535) // window
+
+	// payload
+	for i := 0; i < payLen; i++ {
+		pkt[ipLen+tcpLen+i] = byte(i & 0xff)
+	}
+
+	return pkt, VirtioNetHdr{
+		Flags:      unix.VIRTIO_NET_HDR_F_NEEDS_CSUM,
+		GSOType:    unix.VIRTIO_NET_HDR_GSO_TCPV4,
+		HdrLen:     uint16(ipLen + tcpLen),
+		GSOSize:    uint16(mss),
+		CsumStart:  uint16(ipLen),
+		CsumOffset: 16,
+	}
+}
+
+func TestSegmentTCPv4(t *testing.T) {
+	const mss = 100
+	const numSeg = 3
+	pkt, hdr := buildTSOv4(t, mss*numSeg, mss)
+
+	scratch := make([]byte, tunSegBufSize)
+	var out [][]byte
+	if err := segmentTCP(pkt, hdr, &out, scratch); err != nil {
+		t.Fatalf("segmentTCP: %v", err)
+	}
+	if len(out) != numSeg {
+		t.Fatalf("expected %d segments, got %d", numSeg, len(out))
+	}
+
+	for i, seg := range out {
+		if len(seg) != 40+mss {
+			t.Errorf("seg %d: unexpected len %d", i, len(seg))
+		}
+		totalLen := binary.BigEndian.Uint16(seg[2:4])
+		if totalLen != uint16(40+mss) {
+			t.Errorf("seg %d: total_len=%d want %d", i, totalLen, 40+mss)
+		}
+		id := binary.BigEndian.Uint16(seg[4:6])
+		if id != 0x4242+uint16(i) {
+			t.Errorf("seg %d: ip id=%#x want %#x", i, id, 0x4242+uint16(i))
+		}
+		seq := binary.BigEndian.Uint32(seg[24:28])
+		wantSeq := uint32(10000 + i*mss)
+		if seq != wantSeq {
+			t.Errorf("seg %d: seq=%d want %d", i, seq, wantSeq)
+		}
+		flags := seg[33]
+		wantFlags := byte(0x10) // ACK only, PSH cleared
+		if i == numSeg-1 {
+			wantFlags = 0x18 // ACK | PSH preserved on last
+		}
+		if flags != wantFlags {
+			t.Errorf("seg %d: flags=%#x want %#x", i, flags, wantFlags)
+		}
+		// IPv4 header checksum must verify against itself.
+		if !verifyChecksum(seg[:20], 0) {
+			t.Errorf("seg %d: bad IPv4 header checksum", i)
+		}
+		// TCP checksum must verify against the pseudo-header.
+		psum := pseudoHeaderIPv4(seg[12:16], seg[16:20], unix.IPPROTO_TCP, 20+mss)
+		if !verifyChecksum(seg[20:], psum) {
+			t.Errorf("seg %d: bad TCP checksum", i)
+		}
+	}
+}
+
+func TestSegmentTCPv4OddTail(t *testing.T) {
+	// Payload of 250 bytes with MSS 100 → segments of 100, 100, 50.
+	pkt, hdr := buildTSOv4(t, 250, 100)
+	scratch := make([]byte, tunSegBufSize)
+	var out [][]byte
+	if err := segmentTCP(pkt, hdr, &out, scratch); err != nil {
+		t.Fatalf("segmentTCP: %v", err)
+	}
+	if len(out) != 3 {
+		t.Fatalf("want 3 segments, got %d", len(out))
+	}
+	wantPayLens := []int{100, 100, 50}
+	for i, seg := range out {
+		if len(seg)-40 != wantPayLens[i] {
+			t.Errorf("seg %d: pay len %d want %d", i, len(seg)-40, wantPayLens[i])
+		}
+		if !verifyChecksum(seg[:20], 0) {
+			t.Errorf("seg %d: bad IPv4 header checksum", i)
+		}
+		psum := pseudoHeaderIPv4(seg[12:16], seg[16:20], unix.IPPROTO_TCP, 20+wantPayLens[i])
+		if !verifyChecksum(seg[20:], psum) {
+			t.Errorf("seg %d: bad TCP checksum", i)
+		}
+	}
+}
+
+func TestSegmentTCPv6(t *testing.T) {
+	const ipLen = 40
+	const tcpLen = 20
+	const mss = 120
+	const numSeg = 2
+	payLen := mss * numSeg
+	pkt := make([]byte, ipLen+tcpLen+payLen)
+
+	// IPv6 header
+	pkt[0] = 0x60 // version 6
+	binary.BigEndian.PutUint16(pkt[4:6], uint16(tcpLen+payLen))
+	pkt[6] = unix.IPPROTO_TCP
+	pkt[7] = 64
+	// src/dst fe80::1 / fe80::2
+	pkt[8] = 0xfe
+	pkt[9] = 0x80
+	pkt[23] = 1
+	pkt[24] = 0xfe
+	pkt[25] = 0x80
+	pkt[39] = 2
+
+	// TCP header
+	binary.BigEndian.PutUint16(pkt[40:42], 12345)
+	binary.BigEndian.PutUint16(pkt[42:44], 80)
+	binary.BigEndian.PutUint32(pkt[44:48], 7)
+	binary.BigEndian.PutUint32(pkt[48:52], 99)
+	pkt[52] = 0x50
+	pkt[53] = 0x19 // FIN | ACK | PSH — exercise FIN clearing too
+	binary.BigEndian.PutUint16(pkt[54:56], 65535)
+
+	for i := 0; i < payLen; i++ {
+		pkt[ipLen+tcpLen+i] = byte(i)
+	}
+
+	hdr := VirtioNetHdr{
+		Flags:      unix.VIRTIO_NET_HDR_F_NEEDS_CSUM,
+		GSOType:    unix.VIRTIO_NET_HDR_GSO_TCPV6,
+		HdrLen:     uint16(ipLen + tcpLen),
+		GSOSize:    uint16(mss),
+		CsumStart:  uint16(ipLen),
+		CsumOffset: 16,
+	}
+
+	scratch := make([]byte, tunSegBufSize)
+	var out [][]byte
+	if err := segmentTCP(pkt, hdr, &out, scratch); err != nil {
+		t.Fatalf("segmentTCP: %v", err)
+	}
+	if len(out) != numSeg {
+		t.Fatalf("want %d segments, got %d", numSeg, len(out))
+	}
+
+	for i, seg := range out {
+		if len(seg) != ipLen+tcpLen+mss {
+			t.Errorf("seg %d: len %d want %d", i, len(seg), ipLen+tcpLen+mss)
+		}
+		pl := binary.BigEndian.Uint16(seg[4:6])
+		if pl != uint16(tcpLen+mss) {
+			t.Errorf("seg %d: payload_length=%d want %d", i, pl, tcpLen+mss)
+		}
+		seq := binary.BigEndian.Uint32(seg[44:48])
+		if seq != uint32(7+i*mss) {
+			t.Errorf("seg %d: seq=%d want %d", i, seq, 7+i*mss)
+		}
+		flags := seg[53]
+		// Original flags = 0x19 (FIN|ACK|PSH). FIN(0x01)+PSH(0x08) should be
+		// cleared on all but the last; ACK(0x10) always preserved.
+		wantFlags := byte(0x10)
+		if i == numSeg-1 {
+			wantFlags = 0x19
+		}
+		if flags != wantFlags {
+			t.Errorf("seg %d: flags=%#x want %#x", i, flags, wantFlags)
+		}
+		psum := pseudoHeaderIPv6(seg[8:24], seg[24:40], unix.IPPROTO_TCP, tcpLen+mss)
+		if !verifyChecksum(seg[ipLen:], psum) {
+			t.Errorf("seg %d: bad TCP checksum", i)
+		}
+	}
+}
+
+func TestSegmentGSONonePassesThrough(t *testing.T) {
+	pkt, hdr := buildTSOv4(t, 100, 100)
+	hdr.GSOType = unix.VIRTIO_NET_HDR_GSO_NONE
+	hdr.Flags = 0 // no NEEDS_CSUM, leave packet untouched
+
+	scratch := make([]byte, tunSegBufSize)
+	var out [][]byte
+	if err := segmentInto(pkt, hdr, &out, scratch); err != nil {
+		t.Fatalf("segmentInto: %v", err)
+	}
+	if len(out) != 1 {
+		t.Fatalf("want 1 segment, got %d", len(out))
+	}
+	if len(out[0]) != len(pkt) {
+		t.Fatalf("unexpected length: %d vs %d", len(out[0]), len(pkt))
+	}
+}
+
+func TestSegmentRejectsUDP(t *testing.T) {
+	hdr := VirtioNetHdr{GSOType: unix.VIRTIO_NET_HDR_GSO_UDP}
+	var out [][]byte
+	if err := segmentInto(nil, hdr, &out, nil); err == nil {
+		t.Fatalf("expected rejection for UDP GSO")
+	}
+}
+
+func BenchmarkSegmentTCPv4(b *testing.B) {
+	sizes := []struct {
+		name   string
+		payLen int
+		mss    int
+	}{
+		{"64KiB_MSS1460", 65000, 1460},
+		{"16KiB_MSS1460", 16384, 1460},
+		{"4KiB_MSS1460", 4096, 1460},
+	}
+	for _, sz := range sizes {
+		b.Run(sz.name, func(b *testing.B) {
+			const ipLen = 20
+			const tcpLen = 20
+			pkt := make([]byte, ipLen+tcpLen+sz.payLen)
+			pkt[0] = 0x45
+			binary.BigEndian.PutUint16(pkt[2:4], uint16(ipLen+tcpLen+sz.payLen))
+			binary.BigEndian.PutUint16(pkt[4:6], 0x4242)
+			pkt[8] = 64
+			pkt[9] = unix.IPPROTO_TCP
+			copy(pkt[12:16], []byte{10, 0, 0, 1})
+			copy(pkt[16:20], []byte{10, 0, 0, 2})
+			binary.BigEndian.PutUint16(pkt[20:22], 12345)
+			binary.BigEndian.PutUint16(pkt[22:24], 80)
+			binary.BigEndian.PutUint32(pkt[24:28], 10000)
+			binary.BigEndian.PutUint32(pkt[28:32], 20000)
+			pkt[32] = 0x50
+			pkt[33] = 0x18
+			binary.BigEndian.PutUint16(pkt[34:36], 65535)
+			for i := 0; i < sz.payLen; i++ {
+				pkt[ipLen+tcpLen+i] = byte(i)
+			}
+			hdr := VirtioNetHdr{
+				Flags:      unix.VIRTIO_NET_HDR_F_NEEDS_CSUM,
+				GSOType:    unix.VIRTIO_NET_HDR_GSO_TCPV4,
+				HdrLen:     uint16(ipLen + tcpLen),
+				GSOSize:    uint16(sz.mss),
+				CsumStart:  uint16(ipLen),
+				CsumOffset: 16,
+			}
+
+			scratch := make([]byte, tunSegBufSize)
+			out := make([][]byte, 0, 64)
+
+			b.SetBytes(int64(len(pkt)))
+			b.ResetTimer()
+			for i := 0; i < b.N; i++ {
+				out = out[:0]
+				if err := segmentTCP(pkt, hdr, &out, scratch); err != nil {
+					b.Fatal(err)
+				}
+			}
+		})
+	}
+}
+
+// TestTunFileWriteVnetHdrNoAlloc verifies the IFF_VNET_HDR fast-path write is
+// allocation-free. We write to /dev/null so every call succeeds synchronously.
+func TestTunFileWriteVnetHdrNoAlloc(t *testing.T) {
+	fd, err := unix.Open("/dev/null", os.O_WRONLY, 0)
+	if err != nil {
+		t.Fatalf("open /dev/null: %v", err)
+	}
+	t.Cleanup(func() { _ = unix.Close(fd) })
+
+	tf := &Offload{fd: fd}
+	tf.writeIovs[0].Base = &validVnetHdr[0]
+	tf.writeIovs[0].SetLen(virtioNetHdrLen)
+
+	payload := make([]byte, 1400)
+	// Warm up (first call may trigger one-time internal allocations elsewhere).
+	if _, err := tf.Write(payload); err != nil {
+		t.Fatalf("Write: %v", err)
+	}
+
+	allocs := testing.AllocsPerRun(1000, func() {
+		if _, err := tf.Write(payload); err != nil {
+			t.Fatalf("Write: %v", err)
+		}
+	})
+	if allocs != 0 {
+		t.Fatalf("Write allocated %.1f times per call, want 0", allocs)
+	}
+}

overlay/tio/vnethdr_linux.go

@@ -0,0 +1,39 @@
+package tio
+
+import "encoding/binary"
+
+// Size of the legacy struct virtio_net_hdr that the kernel prepends/expects on
+// a TUN opened with IFF_VNET_HDR (TUNSETVNETHDRSZ not set).
+const virtioNetHdrLen = 10
+
+type VirtioNetHdr struct {
+	Flags      uint8
+	GSOType    uint8
+	HdrLen     uint16
+	GSOSize    uint16
+	CsumStart  uint16
+	CsumOffset uint16
+}
+
+// decode reads a virtio_net_hdr in host byte order (TUN default; we never
+// call TUNSETVNETLE so the kernel matches our endianness).
+func (h *VirtioNetHdr) decode(b []byte) {
+	h.Flags = b[0]
+	h.GSOType = b[1]
+	h.HdrLen = binary.NativeEndian.Uint16(b[2:4])
+	h.GSOSize = binary.NativeEndian.Uint16(b[4:6])
+	h.CsumStart = binary.NativeEndian.Uint16(b[6:8])
+	h.CsumOffset = binary.NativeEndian.Uint16(b[8:10])
+}
+
+// encode is the inverse of decode: writes the virtio_net_hdr fields into b
+// (must be at least virtioNetHdrLen bytes). Used to emit a TSO superpacket
+// on egress.
+func (h *VirtioNetHdr) encode(b []byte) {
+	b[0] = h.Flags
+	b[1] = h.GSOType
+	binary.NativeEndian.PutUint16(b[2:4], h.HdrLen)
+	binary.NativeEndian.PutUint16(b[4:6], h.GSOSize)
+	binary.NativeEndian.PutUint16(b[6:8], h.CsumStart)
+	binary.NativeEndian.PutUint16(b[8:10], h.CsumOffset)
+}

overlay/tun_linux.go

@@ -34,6 +34,7 @@ type tun struct {
 	TXQueueLen  int
 	deviceIndex int
 	ioctlFd     uintptr
+	vnetHdr     bool
 
 	Routes                    atomic.Pointer[[]Route]
 	routeTree                 atomic.Pointer[bart.Table[routing.Gateways]]
@@ -72,7 +73,9 @@ type ifreqQLEN struct {
 }
 
 func newTunFromFd(c *config.C, l *slog.Logger, deviceFd int, vpnNetworks []netip.Prefix) (*tun, error) {
-	t, err := newTunGeneric(c, l, deviceFd, vpnNetworks)
+	// We don't know what flags the caller opened this fd with and can't turn
+	// on IFF_VNET_HDR after TUNSETIFF, so skip offload on inherited fds.
+	t, err := newTunGeneric(c, l, deviceFd, false, vpnNetworks)
 	if err != nil {
 		return nil, err
 	}
@@ -117,6 +120,10 @@ func tunSetIff(fd int, name string, flags uint16) (string, error) {
 	return strings.Trim(string(req.Name[:]), "\x00"), nil
 }
 
+// tsoOffloadFlags are the TUN_F_* bits we ask the kernel to enable when a
+// TSO-capable TUN is available. CSUM is required as a prerequisite for TSO.
+const tsoOffloadFlags = unix.TUN_F_CSUM | unix.TUN_F_TSO4 | unix.TUN_F_TSO6
+
 func newTun(c *config.C, l *slog.Logger, vpnNetworks []netip.Prefix, multiqueue bool) (*tun, error) {
 	baseFlags := uint16(unix.IFF_TUN | unix.IFF_NO_PI)
 	if multiqueue {
@@ -124,17 +131,37 @@ func newTun(c *config.C, l *slog.Logger, vpnNetworks []netip.Prefix, multiqueue
 	}
 	nameStr := c.GetString("tun.dev", "")
 
+	// First try to open with IFF_VNET_HDR + TUNSETOFFLOAD so we can receive
+	// TSO superpackets. If either step fails (older kernel, unprivileged
+	// container, etc.) we close and fall back to a plain TUN.
 	fd, err := openTunDev()
 	if err != nil {
 		return nil, err
 	}
-	name, err := tunSetIff(fd, nameStr, baseFlags)
+	vnetHdr := true
+	name, err := tunSetIff(fd, nameStr, baseFlags|unix.IFF_VNET_HDR)
 	if err != nil {
 		_ = unix.Close(fd)
-		return nil, &NameError{Name: nameStr, Underlying: err}
+		vnetHdr = false
+	} else if err = ioctl(uintptr(fd), unix.TUNSETOFFLOAD, uintptr(tsoOffloadFlags)); err != nil {
+		l.Warn("Failed to enable TUN offload (TSO); proceeding without virtio headers", "error", err)
+		_ = unix.Close(fd)
+		vnetHdr = false
 	}
 
-	t, err := newTunGeneric(c, l, fd, vpnNetworks)
+	if !vnetHdr {
+		fd, err = openTunDev()
+		if err != nil {
+			return nil, err
+		}
+		name, err = tunSetIff(fd, nameStr, baseFlags)
+		if err != nil {
+			_ = unix.Close(fd)
+			return nil, &NameError{Name: nameStr, Underlying: err}
+		}
+	}
+
+	t, err := newTunGeneric(c, l, fd, vnetHdr, vpnNetworks)
 	if err != nil {
 		return nil, err
 	}
@@ -145,8 +172,15 @@ func newTun(c *config.C, l *slog.Logger, vpnNetworks []netip.Prefix, multiqueue
 }
 
 // newTunGeneric does all the stuff common to different tun initialization paths. It will close your files on error.
-func newTunGeneric(c *config.C, l *slog.Logger, fd int, vpnNetworks []netip.Prefix) (*tun, error) {
-	container, err := tio.NewPollContainer()
+func newTunGeneric(c *config.C, l *slog.Logger, fd int, vnetHdr bool, vpnNetworks []netip.Prefix) (*tun, error) {
+	var container tio.Container
+	var err error
+	if vnetHdr {
+		container, err = tio.NewOffloadContainer()
+	} else {
+		container, err = tio.NewPollContainer()
+	}
+
 	if err != nil {
 		_ = unix.Close(fd)
 		return nil, err
@@ -160,6 +194,7 @@ func newTunGeneric(c *config.C, l *slog.Logger, fd int, vpnNetworks []netip.Pref
 	t := &tun{
 		readers:                   container,
 		closeLock:                 sync.Mutex{},
+		vnetHdr:                   vnetHdr,
 		vpnNetworks:               vpnNetworks,
 		TXQueueLen:                c.GetInt("tun.tx_queue", 500),
 		useSystemRoutes:           c.GetBool("tun.use_system_route_table", false),
@@ -271,11 +306,21 @@ func (t *tun) NewMultiQueueReader() error {
 	}
 
 	flags := uint16(unix.IFF_TUN | unix.IFF_NO_PI | unix.IFF_MULTI_QUEUE)
+	if t.vnetHdr {
+		flags |= unix.IFF_VNET_HDR
+	}
 	if _, err = tunSetIff(fd, t.Device, flags); err != nil {
 		_ = unix.Close(fd)
 		return err
 	}
 
+	if t.vnetHdr {
+		if err = ioctl(uintptr(fd), unix.TUNSETOFFLOAD, uintptr(tsoOffloadFlags)); err != nil {
+			_ = unix.Close(fd)
+			return fmt.Errorf("failed to enable offload on multiqueue tun fd: %w", err)
+		}
+	}
+
 	err = t.readers.Add(fd)
 	if err != nil {
 		_ = unix.Close(fd)

udp/udp_linux.go

@@ -24,6 +24,43 @@ type StdConn struct {
 	isV4    bool
 	l       *slog.Logger
 	batch   int
+
+	// sendmmsg scratch. Each queue has its own StdConn, so no locking is
+	// needed. Sized to MaxWriteBatch at construction; WriteBatch chunks
+	// larger inputs.
+	writeMsgs  []rawMessage
+	writeIovs  []iovec
+	writeNames [][]byte
+
+	// Per-entry UDP_SEGMENT cmsg scratch. writeCmsg is one contiguous slab
+	// of MaxWriteBatch * writeCmsgSpace bytes; each entry's cmsg header is
+	// pre-filled once in prepareWriteMessages. WriteBatch only rewrites the
+	// 2-byte gso_size payload (and toggles Hdr.Control on/off) per call.
+	writeCmsg      []byte
+	writeCmsgSpace int
+
+	// writeEntryEnd[e] is the bufs index *after* the last packet packed
+	// into mmsghdr entry e. Used to rewind `i` on partial sendmmsg success.
+	writeEntryEnd []int
+
+	// Preallocated closure + in/out slots for sendmmsg, so the hot path
+	// does not heap-allocate a fresh closure per call.
+	writeChunk int
+	writeSent  int
+	writeErrno syscall.Errno
+	writeFunc  func(fd uintptr) bool
+
+	// UDP GSO (sendmsg with UDP_SEGMENT cmsg) support. gsoSupported is
+	// probed once at socket creation. When true, WriteSegmented takes a
+	// single-syscall GSO path; otherwise it falls back to a WriteTo loop.
+	gsoSupported bool
+
+	// UDP GRO (recvmsg with UDP_GRO cmsg) support. groSupported is probed
+	// once at socket creation. When true, listenOutBatch allocates larger
+	// RX buffers and a per-entry cmsg slot so the kernel can coalesce
+	// consecutive same-flow datagrams into a single recvmmsg entry; the
+	// delivered cmsg carries the gso_size used to split them back apart.
+	groSupported bool
 }
 
 func setReusePort(network, address string, c syscall.RawConn) error {
@@ -70,9 +107,102 @@ func NewListener(l *slog.Logger, ip netip.Addr, port int, multi bool, batch int)
 	}
 	out.isV4 = af == unix.AF_INET
 
+	out.prepareWriteMessages(MaxWriteBatch)
+	out.writeFunc = out.sendmmsgRawWrite
+
+	out.prepareGSO()
+	// GRO delivers coalesced superpackets that need a cmsg to split back
+	// into segments. The single-packet RX path uses ReadFromUDPAddrPort
+	// and cannot see that cmsg, so only enable GRO for the batch path.
+	if batch > 1 {
+		out.prepareGRO()
+	}
+
 	return out, nil
 }
 
+// prepareWriteMessages allocates one mmsghdr/iovec/sockaddr/cmsg scratch
+// slot per sendmmsg entry. The iovec slab is sized to the same n so a
+// single entry can fan out to up to n iovecs (needed for UDP_SEGMENT runs
+// that coalesce consecutive bufs into one entry). Hdr.Iov / Hdr.Iovlen /
+// Hdr.Control / Hdr.Controllen are wired per call since each entry can
+// span a variable number of iovecs and may or may not carry a cmsg.
+func (u *StdConn) prepareWriteMessages(n int) {
+	u.writeMsgs = make([]rawMessage, n)
+	u.writeIovs = make([]iovec, n)
+	u.writeNames = make([][]byte, n)
+	u.writeEntryEnd = make([]int, n)
+
+	u.writeCmsgSpace = unix.CmsgSpace(2)
+	u.writeCmsg = make([]byte, n*u.writeCmsgSpace)
+	for k := 0; k < n; k++ {
+		off := k * u.writeCmsgSpace
+		h := (*unix.Cmsghdr)(unsafe.Pointer(&u.writeCmsg[off]))
+		h.Level = unix.SOL_UDP
+		h.Type = unix.UDP_SEGMENT
+		setCmsgLen(h, unix.CmsgLen(2))
+	}
+
+	for i := range u.writeMsgs {
+		u.writeNames[i] = make([]byte, unix.SizeofSockaddrInet6)
+		u.writeMsgs[i].Hdr.Name = &u.writeNames[i][0]
+	}
+}
+
+// maxGSOSegments caps the per-sendmsg GSO fan-out. Linux kernels have
+// historically capped UDP_MAX_SEGMENTS at 64; newer kernels raise it to 128
+// but we stay conservative so the same code works everywhere.
+const maxGSOSegments = 64
+
+// maxGSOBytes bounds the total payload per sendmsg() when UDP_SEGMENT is
+// set. The kernel stitches all iovecs into a single skb whose length the
+// UDP length field can represent, and also enforces sk_gso_max_size (which
+// on most devices is 65536). We use 65535 so ciphertext + headers always
+// fits, avoiding EMSGSIZE on large TSO superpackets.
+const maxGSOBytes = 65535
+
+// prepareGSO probes UDP_SEGMENT support
+func (u *StdConn) prepareGSO() {
+	var probeErr error
+	if err := u.rawConn.Control(func(fd uintptr) {
+		probeErr = unix.SetsockoptInt(int(fd), unix.IPPROTO_UDP, unix.UDP_SEGMENT, 0)
+	}); err != nil {
+		return
+	}
+	if probeErr != nil {
+		return
+	}
+	u.gsoSupported = true
+}
+
+// udpGROBufferSize sizes the per-entry recvmmsg buffer when UDP_GRO is on.
+// The kernel stitches a run of same-flow datagrams into a single skb whose
+// length is bounded by sk_gso_max_size (typically 65535); anything larger
+// would be MSG_TRUNCed. We use the maximum representable UDP length so a
+// full superpacket always lands intact.
+const udpGROBufferSize = 65535
+
+// udpGROCmsgPayload is the size of the UDP_GRO cmsg data delivered by the
+// kernel: a single int (gso_size in bytes). See udp_cmsg_recv() in
+// net/ipv4/udp.c.
+const udpGROCmsgPayload = 4
+
+// prepareGRO turns on UDP_GRO so the kernel coalesces consecutive same-flow
+// datagrams into one recvmmsg entry, with a cmsg carrying the gso_size used
+// to split them back apart on the application side.
+func (u *StdConn) prepareGRO() {
+	var probeErr error
+	if err := u.rawConn.Control(func(fd uintptr) {
+		probeErr = unix.SetsockoptInt(int(fd), unix.IPPROTO_UDP, unix.UDP_GRO, 1)
+	}); err != nil {
+		return
+	}
+	if probeErr != nil {
+		return
+	}
+	u.groSupported = true
+}
+
 func (u *StdConn) SupportsMultipleReaders() bool {
 	return true
 }
@@ -194,7 +324,12 @@ func (u *StdConn) listenOutBatch(r EncReader, flush func()) error {
 	var operr error
 
 	bufSize := MTU
-	msgs, buffers, names := u.PrepareRawMessages(u.batch, bufSize)
+	cmsgSpace := 0
+	if u.groSupported {
+		bufSize = udpGROBufferSize
+		cmsgSpace = unix.CmsgSpace(udpGROCmsgPayload)
+	}
+	msgs, buffers, names, _ := u.PrepareRawMessages(u.batch, bufSize, cmsgSpace)
 
 	//reader needs to capture variables from this function, since it's used as a lambda with rawConn.Read
 	//defining it outside the loop so it gets re-used
@@ -204,6 +339,11 @@ func (u *StdConn) listenOutBatch(r EncReader, flush func()) error {
 	}
 
 	for {
+		if cmsgSpace > 0 {
+			for i := range msgs {
+				setMsgControllen(&msgs[i].Hdr, cmsgSpace)
+			}
+		}
 		err := u.rawConn.Read(reader)
 		if err != nil {
 			return err
@@ -222,7 +362,25 @@ func (u *StdConn) listenOutBatch(r EncReader, flush func()) error {
 			from := netip.AddrPortFrom(ip.Unmap(), binary.BigEndian.Uint16(names[i][2:4]))
 			payload := buffers[i][:msgs[i].Len]
 
-			r(from, payload)
+			segSize := 0
+			if u.groSupported {
+				segSize = parseUDPGRO(&msgs[i].Hdr)
+			}
+			if segSize <= 0 || segSize >= len(payload) {
+				// No coalescing happened (or a lone datagram).
+				r(from, payload)
+				continue
+			}
+			// GRO superpacket: the kernel guarantees every segment is
+			// exactly segSize bytes except for the final one, which may be
+			// short.
+			for off := 0; off < len(payload); off += segSize {
+				end := off + segSize
+				if end > len(payload) {
+					end = len(payload)
+				}
+				r(from, payload[off:end])
+			}
 		}
 		// End-of-batch: let callers (e.g. TUN write coalescer) flush any
 		// state they accumulated across this batch.
@@ -230,6 +388,38 @@ func (u *StdConn) listenOutBatch(r EncReader, flush func()) error {
 	}
 }
 
+// parseUDPGRO walks the control buffer on hdr looking for a SOL_UDP/UDP_GRO
+// cmsg and returns the gso_size (bytes per coalesced segment) it carries.
+// Returns 0 when no UDP_GRO cmsg is present, which is the normal case for
+// lone datagrams that the kernel did not coalesce.
+func parseUDPGRO(hdr *msghdr) int {
+	controllen := int(hdr.Controllen)
+	if controllen < unix.SizeofCmsghdr || hdr.Control == nil {
+		return 0
+	}
+	ctrl := unsafe.Slice(hdr.Control, controllen)
+	off := 0
+	for off+unix.SizeofCmsghdr <= len(ctrl) {
+		ch := (*unix.Cmsghdr)(unsafe.Pointer(&ctrl[off]))
+		clen := int(ch.Len)
+		if clen < unix.SizeofCmsghdr || off+clen > len(ctrl) {
+			return 0
+		}
+		if ch.Level == unix.SOL_UDP && ch.Type == unix.UDP_GRO {
+			dataOff := off + unix.CmsgLen(0)
+			if dataOff+udpGROCmsgPayload <= len(ctrl) {
+				return int(int32(binary.NativeEndian.Uint32(ctrl[dataOff : dataOff+udpGROCmsgPayload])))
+			}
+			return 0
+		}
+		// Advance by the aligned cmsg space. CmsgSpace(n) is the stride
+		// from one header to the next (len aligned up to the platform's
+		// cmsg alignment).
+		off += unix.CmsgSpace(clen - unix.CmsgLen(0))
+	}
+	return 0
+}
+
 func (u *StdConn) ListenOut(r EncReader, flush func()) error {
 	if u.batch == 1 {
 		return u.listenOutSingle(r, flush)
@@ -243,19 +433,209 @@ func (u *StdConn) WriteTo(b []byte, ip netip.AddrPort) error {
 	return err
 }
 
+// WriteBatch sends bufs via sendmmsg(2) using the preallocated scratch on
+// StdConn. Consecutive packets to the same destination with matching segment
+// sizes (all but possibly the last) are coalesced into a single mmsghdr entry
+// carrying a UDP_SEGMENT cmsg, so one syscall can mix runs of GSO superpackets
+// with plain one-off datagrams. Without GSO support every packet is its own
+// entry, matching the prior behaviour.
+//
+// Chunks larger than the scratch are processed across multiple syscalls. If
+// sendmmsg returns a fatal error before any entry is sent we fall back to
+// per-packet WriteTo for that chunk so the caller still gets best-effort
+// delivery.
 func (u *StdConn) WriteBatch(bufs [][]byte, addrs []netip.AddrPort) error {
 	if len(bufs) != len(addrs) {
 		return fmt.Errorf("WriteBatch: len(bufs)=%d != len(addrs)=%d", len(bufs), len(addrs))
 	}
-	//todo use sendmmsg
-	for i := 0; i < len(bufs); i++ {
-		if _, err := u.udpConn.WriteToUDPAddrPort(bufs[i], addrs[i]); err != nil {
-			return err
+
+	i := 0
+	for i < len(bufs) {
+		baseI := i
+		entry := 0
+		iovIdx := 0
+
+		for entry < len(u.writeMsgs) && i < len(bufs) {
+			iovBudget := len(u.writeIovs) - iovIdx
+			if iovBudget < 1 {
+				break
+			}
+			runLen, segSize := u.planRun(bufs, addrs, i, iovBudget)
+			if runLen == 0 {
+				break
+			}
+
+			for k := 0; k < runLen; k++ {
+				b := bufs[i+k]
+				if len(b) == 0 {
+					u.writeIovs[iovIdx+k].Base = nil
+					setIovLen(&u.writeIovs[iovIdx+k], 0)
+				} else {
+					u.writeIovs[iovIdx+k].Base = &b[0]
+					setIovLen(&u.writeIovs[iovIdx+k], len(b))
+				}
+			}
+
+			nlen, err := writeSockaddr(u.writeNames[entry], addrs[i], u.isV4)
+			if err != nil {
+				return err
+			}
+
+			hdr := &u.writeMsgs[entry].Hdr
+			hdr.Iov = &u.writeIovs[iovIdx]
+			setMsgIovlen(hdr, runLen)
+			hdr.Namelen = uint32(nlen)
+
+			if runLen >= 2 {
+				off := entry * u.writeCmsgSpace
+				dataOff := off + unix.CmsgLen(0)
+				binary.NativeEndian.PutUint16(u.writeCmsg[dataOff:dataOff+2], uint16(segSize))
+				hdr.Control = &u.writeCmsg[off]
+				setMsgControllen(hdr, u.writeCmsgSpace)
+			} else {
+				hdr.Control = nil
+				setMsgControllen(hdr, 0)
+			}
+
+			i += runLen
+			iovIdx += runLen
+			u.writeEntryEnd[entry] = i
+			entry++
 		}
+
+		if entry == 0 {
+			return fmt.Errorf("sendmmsg: no progress")
+		}
+
+		sent, serr := u.sendmmsg(entry)
+		if serr != nil && sent <= 0 {
+			// Nothing went out for this chunk; fall back to WriteTo for each
+			// packet that was queued this iteration.
+			for k := baseI; k < i; k++ {
+				if werr := u.WriteTo(bufs[k], addrs[k]); werr != nil {
+					return werr
+				}
+			}
+			continue
+		}
+		if sent == 0 {
+			return fmt.Errorf("sendmmsg made no progress")
+		}
+		// Rewind i to the end of the last successfully sent entry. For a
+		// full-success send this leaves i unchanged; for a partial send it
+		// replays the remainder on the next outer-loop iteration.
+		i = u.writeEntryEnd[sent-1]
 	}
 	return nil
 }
 
+// planRun groups consecutive packets starting at `start` that can be sent as
+// a single UDP GSO superpacket (one sendmmsg entry with UDP_SEGMENT cmsg).
+// A run of length 1 means the entry carries no cmsg and the kernel treats
+// it as a plain datagram. Returns the run length and the per-segment size
+// (which equals len(bufs[start])). Without GSO support every call returns
+// runLen=1.
+func (u *StdConn) planRun(bufs [][]byte, addrs []netip.AddrPort, start, iovBudget int) (int, int) {
+	if start >= len(bufs) || iovBudget < 1 {
+		return 0, 0
+	}
+	segSize := len(bufs[start])
+	if !u.gsoSupported || segSize == 0 || segSize > maxGSOBytes {
+		return 1, segSize
+	}
+	dst := addrs[start]
+	maxLen := maxGSOSegments
+	if iovBudget < maxLen {
+		maxLen = iovBudget
+	}
+	runLen := 1
+	total := segSize
+	for runLen < maxLen && start+runLen < len(bufs) {
+		nextLen := len(bufs[start+runLen])
+		if nextLen == 0 || nextLen > segSize {
+			break
+		}
+		if addrs[start+runLen] != dst {
+			break
+		}
+		if total+nextLen > maxGSOBytes {
+			break
+		}
+		total += nextLen
+		runLen++
+		if nextLen < segSize {
+			// A short packet must be the last in the run.
+			break
+		}
+	}
+	return runLen, segSize
+}
+
+// sendmmsgRawWrite is the preallocated callback passed to rawConn.Write. It
+// reads its input (u.writeChunk) and writes its outputs (u.writeSent,
+// u.writeErrno) through StdConn fields so the closure itself does not
+// capture per-call locals and therefore does not heap-allocate.
+func (u *StdConn) sendmmsgRawWrite(fd uintptr) bool {
+	r1, _, errno := unix.Syscall6(
+		unix.SYS_SENDMMSG,
+		fd,
+		uintptr(unsafe.Pointer(&u.writeMsgs[0])),
+		uintptr(u.writeChunk),
+		0,
+		0,
+		0,
+	)
+	if errno == syscall.EAGAIN || errno == syscall.EWOULDBLOCK {
+		return false
+	}
+	u.writeSent = int(r1)
+	u.writeErrno = errno
+	return true
+}
+
+func (u *StdConn) sendmmsg(n int) (int, error) {
+	u.writeChunk = n
+	u.writeSent = 0
+	u.writeErrno = 0
+	if err := u.rawConn.Write(u.writeFunc); err != nil {
+		return u.writeSent, err
+	}
+	if u.writeErrno != 0 {
+		return u.writeSent, &net.OpError{Op: "sendmmsg", Err: u.writeErrno}
+	}
+	return u.writeSent, nil
+}
+
+// writeSockaddr encodes addr into buf (which must be at least
+// SizeofSockaddrInet6 bytes). Returns the number of bytes used. If isV4 is
+// true and addr is not a v4 (or v4-in-v6) address, returns an error.
+func writeSockaddr(buf []byte, addr netip.AddrPort, isV4 bool) (int, error) {
+	ap := addr.Addr().Unmap()
+	if isV4 {
+		if !ap.Is4() {
+			return 0, ErrInvalidIPv6RemoteForSocket
+		}
+		// struct sockaddr_in: { sa_family_t(2), in_port_t(2, BE), in_addr(4), zero(8) }
+		// sa_family is host endian.
+		binary.NativeEndian.PutUint16(buf[0:2], unix.AF_INET)
+		binary.BigEndian.PutUint16(buf[2:4], addr.Port())
+		ip4 := ap.As4()
+		copy(buf[4:8], ip4[:])
+		for j := 8; j < 16; j++ {
+			buf[j] = 0
+		}
+		return unix.SizeofSockaddrInet4, nil
+	}
+	// struct sockaddr_in6: { sa_family_t(2), in_port_t(2, BE), flowinfo(4), in6_addr(16), scope_id(4) }
+	binary.NativeEndian.PutUint16(buf[0:2], unix.AF_INET6)
+	binary.BigEndian.PutUint16(buf[2:4], addr.Port())
+	binary.NativeEndian.PutUint32(buf[4:8], 0)
+	ip6 := addr.Addr().As16()
+	copy(buf[8:24], ip6[:])
+	binary.NativeEndian.PutUint32(buf[24:28], 0)
+	return unix.SizeofSockaddrInet6, nil
+}
+
 func (u *StdConn) ReloadConfig(c *config.C) {
 	b := c.GetInt("listen.read_buffer", 0)
 	if b > 0 {

udp/udp_linux_32.go

@@ -30,11 +30,16 @@ type rawMessage struct {
 	Len uint32
 }
 
-func (u *StdConn) PrepareRawMessages(n, bufSize int) ([]rawMessage, [][]byte, [][]byte) {
+func (u *StdConn) PrepareRawMessages(n, bufSize, cmsgSpace int) ([]rawMessage, [][]byte, [][]byte, []byte) {
 	msgs := make([]rawMessage, n)
 	buffers := make([][]byte, n)
 	names := make([][]byte, n)
 
+	var cmsgs []byte
+	if cmsgSpace > 0 {
+		cmsgs = make([]byte, n*cmsgSpace)
+	}
+
 	for i := range msgs {
 		buffers[i] = make([]byte, bufSize)
 		names[i] = make([]byte, unix.SizeofSockaddrInet6)
@@ -48,9 +53,30 @@ func (u *StdConn) PrepareRawMessages(n, bufSize int) ([]rawMessage, [][]byte, []
 
 		msgs[i].Hdr.Name = &names[i][0]
 		msgs[i].Hdr.Namelen = uint32(len(names[i]))
+
+		if cmsgSpace > 0 {
+			msgs[i].Hdr.Control = &cmsgs[i*cmsgSpace]
+			msgs[i].Hdr.Controllen = uint32(cmsgSpace)
+		}
 	}
 
-	return msgs, buffers, names
+	return msgs, buffers, names, cmsgs
+}
+
+func setIovLen(v *iovec, n int) {
+	v.Len = uint32(n)
+}
+
+func setMsgIovlen(m *msghdr, n int) {
+	m.Iovlen = uint32(n)
+}
+
+func setMsgControllen(m *msghdr, n int) {
+	m.Controllen = uint32(n)
+}
+
+func setCmsgLen(h *unix.Cmsghdr, n int) {
+	h.Len = uint32(n)
 }
 
 func setIovLen(v *iovec, n int) {

udp/udp_linux_64.go

@@ -33,11 +33,16 @@ type rawMessage struct {
 	Pad0 [4]byte
 }
 
-func (u *StdConn) PrepareRawMessages(n, bufSize int) ([]rawMessage, [][]byte, [][]byte) {
+func (u *StdConn) PrepareRawMessages(n, bufSize, cmsgSpace int) ([]rawMessage, [][]byte, [][]byte, []byte) {
 	msgs := make([]rawMessage, n)
 	buffers := make([][]byte, n)
 	names := make([][]byte, n)
 
+	var cmsgs []byte
+	if cmsgSpace > 0 {
+		cmsgs = make([]byte, n*cmsgSpace)
+	}
+
 	for i := range msgs {
 		buffers[i] = make([]byte, bufSize)
 		names[i] = make([]byte, unix.SizeofSockaddrInet6)
@@ -51,9 +56,30 @@ func (u *StdConn) PrepareRawMessages(n, bufSize int) ([]rawMessage, [][]byte, []
 
 		msgs[i].Hdr.Name = &names[i][0]
 		msgs[i].Hdr.Namelen = uint32(len(names[i]))
+
+		if cmsgSpace > 0 {
+			msgs[i].Hdr.Control = &cmsgs[i*cmsgSpace]
+			msgs[i].Hdr.Controllen = uint64(cmsgSpace)
+		}
 	}
 
-	return msgs, buffers, names
+	return msgs, buffers, names, cmsgs
+}
+
+func setIovLen(v *iovec, n int) {
+	v.Len = uint64(n)
+}
+
+func setMsgIovlen(m *msghdr, n int) {
+	m.Iovlen = uint64(n)
+}
+
+func setMsgControllen(m *msghdr, n int) {
+	m.Controllen = uint64(n)
+}
+
+func setCmsgLen(h *unix.Cmsghdr, n int) {
+	h.Len = uint64(n)
 }
 
 func setIovLen(v *iovec, n int) {