NAPI broken on old kernels and also not helpful

wait for goroutines to finish and for tun to actually be closed

remove yellow squiggles

nbio for tun

bugz

Remove more os.Exit calls and give a more reliable wait for stop function

checkpt

tun_linux.go: stdlib too slow, but can't use blocking IO and clean shutdown

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)
+		}
+	}
+}

cmd/nebula-service/main.go

@@ -50,16 +50,10 @@ func main() {
 		os.Exit(0)
 	}
 
-	l := logrus.New()
-	l.Out = os.Stdout
-
 	if *serviceFlag != "" {
-		if err := doService(configPath, configTest, Build, serviceFlag); err != nil {
+		doService(configPath, configTest, Build, serviceFlag)
-			l.WithError(err).Error("Service command failed")
 		os.Exit(1)
 	}
-		return
-	}
 
 	if *configPath == "" {
 		fmt.Println("-config flag must be set")
@@ -67,6 +61,9 @@ func main() {
 		os.Exit(1)
 	}
 
+	l := logrus.New()
+	l.Out = os.Stdout
+
 	c := config.NewC(l)
 	err := c.Load(*configPath)
 	if err != nil {

cmd/nebula-service/service.go

@@ -57,11 +57,11 @@ func fileExists(filename string) bool {
 	return true
 }
 
-func doService(configPath *string, configTest *bool, build string, serviceFlag *string) error {
+func doService(configPath *string, configTest *bool, build string, serviceFlag *string) {
 	if *configPath == "" {
 		ex, err := os.Executable()
 		if err != nil {
-			return err
+			panic(err)
 		}
 		*configPath = filepath.Dir(ex) + "/config.yaml"
 		if !fileExists(*configPath) {
@@ -88,13 +88,13 @@ func doService(configPath *string, configTest *bool, build string, serviceFlag *
 	// - above, in `Run` we create a `logrus.Logger` which is what nebula expects to use
 	s, err := service.New(prg, svcConfig)
 	if err != nil {
-		return err
+		log.Fatal(err)
 	}
 
 	errs := make(chan error, 5)
 	logger, err = s.Logger(errs)
 	if err != nil {
-		return err
+		log.Fatal(err)
 	}
 
 	go func() {
@@ -109,16 +109,18 @@ func doService(configPath *string, configTest *bool, build string, serviceFlag *
 
 	switch *serviceFlag {
 	case "run":
-		if err := s.Run(); err != nil {
+		err = s.Run()
+		if err != nil {
 			// Route any errors to the system logger
 			logger.Error(err)
 		}
 	default:
-		if err := service.Control(s, *serviceFlag); err != nil {
+		err := service.Control(s, *serviceFlag)
+		if err != nil {
 			log.Printf("Valid actions: %q\n", service.ControlAction)
-			return err
+			log.Fatal(err)
 		}
+		return
 	}
 
-	return nil
 }

connection_manager_test.go

@@ -10,6 +10,7 @@ import (
 	"github.com/flynn/noise"
 	"github.com/slackhq/nebula/cert"
 	"github.com/slackhq/nebula/config"
+	"github.com/slackhq/nebula/overlay"
 	"github.com/slackhq/nebula/test"
 	"github.com/slackhq/nebula/udp"
 	"github.com/stretchr/testify/assert"
@@ -52,7 +53,7 @@ func Test_NewConnectionManagerTest(t *testing.T) {
 	lh := newTestLighthouse()
 	ifce := &Interface{
 		hostMap:          hostMap,
-		inside:           &test.NoopTun{},
+		inside:           &overlay.NoopTun{},
 		outside:          &udp.NoopConn{},
 		firewall:         &Firewall{},
 		lightHouse:       lh,
@@ -135,7 +136,7 @@ func Test_NewConnectionManagerTest2(t *testing.T) {
 	lh := newTestLighthouse()
 	ifce := &Interface{
 		hostMap:          hostMap,
-		inside:           &test.NoopTun{},
+		inside:           &overlay.NoopTun{},
 		outside:          &udp.NoopConn{},
 		firewall:         &Firewall{},
 		lightHouse:       lh,
@@ -220,7 +221,7 @@ func Test_NewConnectionManager_DisconnectInactive(t *testing.T) {
 	lh := newTestLighthouse()
 	ifce := &Interface{
 		hostMap:          hostMap,
-		inside:           &test.NoopTun{},
+		inside:           &overlay.NoopTun{},
 		outside:          &udp.NoopConn{},
 		firewall:         &Firewall{},
 		lightHouse:       lh,
@@ -347,7 +348,7 @@ func Test_NewConnectionManagerTest_DisconnectInvalid(t *testing.T) {
 	lh := newTestLighthouse()
 	ifce := &Interface{
 		hostMap:          hostMap,
-		inside:           &test.NoopTun{},
+		inside:           &overlay.NoopTun{},
 		outside:          &udp.NoopConn{},
 		firewall:         &Firewall{},
 		lightHouse:       lh,

control.go

@@ -11,7 +11,6 @@ import (
 
 	"github.com/sirupsen/logrus"
 	"github.com/slackhq/nebula/cert"
-	"github.com/slackhq/nebula/firewall/events"
 	"github.com/slackhq/nebula/header"
 	"github.com/slackhq/nebula/overlay"
 )
@@ -341,36 +340,6 @@ func (c *Control) Device() overlay.Device {
 	return c.f.inside
 }
 
-// SetFirewallEventReporter installs an event reporter on the current firewall.
-// Passing nil clears any installed reporter. The reporter is carried across
-// firewall rule reloads. Report* methods are invoked while nebula holds
-// internal locks and must be non-blocking; in particular they must not call
-// back into *Control methods that touch the firewall, or deadlock will
-// result.
-//
-// Installation is performed by shallow-copying the current *Firewall,
-// setting the reporter field on the copy, and swapping the pointer under
-// the conntrack lock. Every Firewall the data path sees therefore has an
-// immutable reporter slot, and emit sites can read it without any
-// synchronization of their own.
-func (c *Control) SetFirewallEventReporter(r events.Reporter) {
-	old := c.f.firewall
-	if old == nil {
-		return
-	}
-	old.Conntrack.Lock()
-	defer old.Conntrack.Unlock()
-
-	// Re-read under the lock in case a concurrent reload swapped in a new
-	// Firewall between the unlocked load above and here. Both Firewalls share
-	// the same Conntrack pointer in the normal (non-overflow) reload path,
-	// so the lock we hold is the right one for whichever we see now.
-	current := c.f.firewall
-	fw := *current
-	fw.reporter = r
-	c.f.firewall = &fw
-}
-
 func copyHostInfo(h *HostInfo, preferredRanges []netip.Prefix) ControlHostInfo {
 	chi := ControlHostInfo{
 		VpnAddrs:               make([]netip.Addr, len(h.vpnAddrs)),

dns_server.go

@@ -1,14 +1,12 @@
 package nebula
 
 import (
-	"context"
 	"fmt"
 	"net"
 	"net/netip"
 	"strconv"
 	"strings"
 	"sync"
-	"sync/atomic"
 
 	"github.com/gaissmai/bart"
 	"github.com/miekg/dns"
@@ -16,207 +14,32 @@ import (
 	"github.com/slackhq/nebula/config"
 )
 
-type dnsServer struct {
+// This whole thing should be rewritten to use context
+
+var dnsR *dnsRecords
+var dnsServer *dns.Server
+var dnsAddr string
+
+type dnsRecords struct {
 	sync.RWMutex
 	l               *logrus.Logger
-	ctx             context.Context
 	dnsMap4         map[string]netip.Addr
 	dnsMap6         map[string]netip.Addr
 	hostMap         *HostMap
 	myVpnAddrsTable *bart.Lite
-
-	mux *dns.ServeMux
-
-	// enabled mirrors `lighthouse.serve_dns && lighthouse.am_lighthouse`.
-	// Start, Add, and reload consult it so callers don't need to know the
-	// gating rules. When it toggles off via reload, accumulated records are
-	// cleared so a later re-enable starts with a fresh map populated from
-	// new handshakes.
-	enabled atomic.Bool
-
-	serverMu sync.Mutex
-	server   *dns.Server
-	// started is closed once `server` has finished binding (or after
-	// ListenAndServe returns on a bind failure). Stop waits on it before
-	// calling Shutdown to avoid the miekg/dns "server not started" race
-	// where a Shutdown that arrives before bind completes is silently
-	// ignored, leaving the listener running forever.
-	started chan struct{}
-	addr    string
 }
 
-// newDnsServerFromConfig builds a dnsServer, applies the initial config, and
+func newDnsRecords(l *logrus.Logger, cs *CertState, hostMap *HostMap) *dnsRecords {
-// registers a reload callback. The reload callback is registered before the
+	return &dnsRecords{
-// initial config is applied, so a SIGHUP can later enable, fix, or disable
-// DNS even if the initial application failed.
-//
-// The dnsServer internally gates on `lighthouse.serve_dns &&
-// lighthouse.am_lighthouse`. Start and Add are safe to call unconditionally,
-// they no-op when DNS isn't enabled. Each Start invocation owns a ctx-cancel
-// watcher that tears the listener down on nebula shutdown. The returned
-// pointer is always non-nil, even on error.
-func newDnsServerFromConfig(ctx context.Context, l *logrus.Logger, cs *CertState, hostMap *HostMap, c *config.C) (*dnsServer, error) {
-	ds := &dnsServer{
 		l:               l,
-		ctx:             ctx,
 		dnsMap4:         make(map[string]netip.Addr),
 		dnsMap6:         make(map[string]netip.Addr),
 		hostMap:         hostMap,
 		myVpnAddrsTable: cs.myVpnAddrsTable,
 	}
-	ds.mux = dns.NewServeMux()
-	ds.mux.HandleFunc(".", ds.handleDnsRequest)
-
-	c.RegisterReloadCallback(func(c *config.C) {
-		if err := ds.reload(c, false); err != nil {
-			l.WithError(err).Error("Failed to reload DNS responder from config")
-		}
-	})
-
-	if err := ds.reload(c, true); err != nil {
-		return ds, err
-	}
-	return ds, nil
 }
 
-// reload applies the latest config and reconciles the running state with it:
+func (d *dnsRecords) Query(q uint16, data string) netip.Addr {
-//   - enabled toggled on  -> spawn a runner
-//   - enabled toggled off -> stop the runner
-//   - listen address changed (while running) -> restart on the new address
-//   - everything else     -> no-op
-//
-// On the initial call it only records configuration; Control.Start is what
-// launches the first runner via dnsStart.
-func (d *dnsServer) reload(c *config.C, initial bool) error {
-	wantsDns := c.GetBool("lighthouse.serve_dns", false)
-	amLighthouse := c.GetBool("lighthouse.am_lighthouse", false)
-	enabled := wantsDns && amLighthouse
-	newAddr := getDnsServerAddr(c)
-
-	d.serverMu.Lock()
-	running := d.server
-	runningStarted := d.started
-	sameAddr := d.addr == newAddr
-	d.addr = newAddr
-	d.enabled.Store(enabled)
-	d.serverMu.Unlock()
-
-	if initial {
-		if wantsDns && !amLighthouse {
-			d.l.Warn("DNS server refusing to run because this host is not a lighthouse.")
-		}
-		return nil
-	}
-
-	if !enabled {
-		if running != nil {
-			d.Stop()
-		}
-		// Drop any records that accumulated while enabled; a later re-enable
-		// will repopulate from fresh handshakes.
-		d.clearRecords()
-		return nil
-	}
-
-	if running == nil {
-		// Was disabled (or never started); bring it up now.
-		go d.Start()
-		return nil
-	}
-
-	if sameAddr {
-		return nil
-	}
-
-	d.shutdownServer(running, runningStarted, "reload")
-	// Old Start goroutine has now exited; bring up a fresh listener on the
-	// new address.
-	go d.Start()
-	return nil
-}
-
-// shutdownServer waits for the server to finish binding (so Shutdown actually
-// stops it rather than no-oping) and then shuts it down.
-func (d *dnsServer) shutdownServer(srv *dns.Server, started chan struct{}, reason string) {
-	if srv == nil {
-		return
-	}
-	if started != nil {
-		<-started
-	}
-	if err := srv.Shutdown(); err != nil {
-		d.l.WithError(err).WithField("reason", reason).Warn("Failed to shut down the DNS responder")
-	}
-}
-
-// Start binds and serves the DNS responder. Blocks until Stop is called or
-// the listener errors. Safe to call when DNS is disabled (returns
-// immediately). This is what Control.dnsStart points at.
-//
-// Must be invoked after the tun device is active so that lighthouse.dns.host
-// may bind to a nebula IP.
-func (d *dnsServer) Start() {
-	if !d.enabled.Load() {
-		return
-	}
-
-	started := make(chan struct{})
-	d.serverMu.Lock()
-	if d.ctx.Err() != nil {
-		d.serverMu.Unlock()
-		return
-	}
-	addr := d.addr
-	server := &dns.Server{
-		Addr:              addr,
-		Net:               "udp",
-		Handler:           d.mux,
-		NotifyStartedFunc: func() { close(started) },
-	}
-	d.server = server
-	d.started = started
-	d.serverMu.Unlock()
-
-	// Per-invocation ctx watcher. Exits when Start does, so we don't leak a
-	// watcher per reload-driven restart.
-	done := make(chan struct{})
-	go func() {
-		select {
-		case <-d.ctx.Done():
-			d.shutdownServer(server, started, "shutdown")
-		case <-done:
-		}
-	}()
-
-	d.l.WithField("dnsListener", addr).Info("Starting DNS responder")
-	err := server.ListenAndServe()
-	close(done)
-
-	// If the listener never bound (bind error) NotifyStartedFunc never fires,
-	// so close started here to release any Stop caller waiting on it.
-	select {
-	case <-started:
-	default:
-		close(started)
-	}
-
-	if err != nil {
-		d.l.WithError(err).Warn("Failed to run the DNS responder")
-	}
-}
-
-// Stop shuts down the active server, if any. Idempotent.
-func (d *dnsServer) Stop() {
-	d.serverMu.Lock()
-	srv := d.server
-	started := d.started
-	d.server = nil
-	d.started = nil
-	d.serverMu.Unlock()
-	d.shutdownServer(srv, started, "stop")
-}
-
-func (d *dnsServer) Query(q uint16, data string) netip.Addr {
 	data = strings.ToLower(data)
 	d.RLock()
 	defer d.RUnlock()
@@ -234,7 +57,7 @@ func (d *dnsServer) Query(q uint16, data string) netip.Addr {
 	return netip.Addr{}
 }
 
-func (d *dnsServer) QueryCert(data string) string {
+func (d *dnsRecords) QueryCert(data string) string {
 	ip, err := netip.ParseAddr(data[:len(data)-1])
 	if err != nil {
 		return ""
@@ -257,19 +80,8 @@ func (d *dnsServer) QueryCert(data string) string {
 	return string(b)
 }
 
-// clearRecords drops all DNS records.
-func (d *dnsServer) clearRecords() {
-	d.Lock()
-	defer d.Unlock()
-	clear(d.dnsMap4)
-	clear(d.dnsMap6)
-}
-
 // Add adds the first IPv4 and IPv6 address that appears in `addresses` as the record for `host`
-func (d *dnsServer) Add(host string, addresses []netip.Addr) {
+func (d *dnsRecords) Add(host string, addresses []netip.Addr) {
-	if !d.enabled.Load() {
-		return
-	}
 	host = strings.ToLower(host)
 	d.Lock()
 	defer d.Unlock()
@@ -289,7 +101,7 @@ func (d *dnsServer) Add(host string, addresses []netip.Addr) {
 	}
 }
 
-func (d *dnsServer) isSelfNebulaOrLocalhost(addr string) bool {
+func (d *dnsRecords) isSelfNebulaOrLocalhost(addr string) bool {
 	a, _, _ := net.SplitHostPort(addr)
 	b, err := netip.ParseAddr(a)
 	if err != nil {
@@ -304,7 +116,7 @@ func (d *dnsServer) isSelfNebulaOrLocalhost(addr string) bool {
 	return d.myVpnAddrsTable.Contains(b)
 }
 
-func (d *dnsServer) parseQuery(m *dns.Msg, w dns.ResponseWriter) {
+func (d *dnsRecords) parseQuery(m *dns.Msg, w dns.ResponseWriter) {
 	for _, q := range m.Question {
 		switch q.Qtype {
 		case dns.TypeA, dns.TypeAAAA:
@@ -338,7 +150,7 @@ func (d *dnsServer) parseQuery(m *dns.Msg, w dns.ResponseWriter) {
 	}
 }
 
-func (d *dnsServer) handleDnsRequest(w dns.ResponseWriter, r *dns.Msg) {
+func (d *dnsRecords) handleDnsRequest(w dns.ResponseWriter, r *dns.Msg) {
 	m := new(dns.Msg)
 	m.SetReply(r)
 	m.Compress = false
@@ -351,6 +163,21 @@ func (d *dnsServer) handleDnsRequest(w dns.ResponseWriter, r *dns.Msg) {
 	w.WriteMsg(m)
 }
 
+func dnsMain(l *logrus.Logger, cs *CertState, hostMap *HostMap, c *config.C) func() {
+	dnsR = newDnsRecords(l, cs, hostMap)
+
+	// attach request handler func
+	dns.HandleFunc(".", dnsR.handleDnsRequest)
+
+	c.RegisterReloadCallback(func(c *config.C) {
+		reloadDns(l, c)
+	})
+
+	return func() {
+		startDns(l, c)
+	}
+}
+
 func getDnsServerAddr(c *config.C) string {
 	dnsHost := strings.TrimSpace(c.GetString("lighthouse.dns.host", ""))
 	// Old guidance was to provide the literal `[::]` in `lighthouse.dns.host` but that won't resolve.
@@ -359,3 +186,25 @@ func getDnsServerAddr(c *config.C) string {
 	}
 	return net.JoinHostPort(dnsHost, strconv.Itoa(c.GetInt("lighthouse.dns.port", 53)))
 }
+
+func startDns(l *logrus.Logger, c *config.C) {
+	dnsAddr = getDnsServerAddr(c)
+	dnsServer = &dns.Server{Addr: dnsAddr, Net: "udp"}
+	l.WithField("dnsListener", dnsAddr).Info("Starting DNS responder")
+	err := dnsServer.ListenAndServe()
+	defer dnsServer.Shutdown()
+	if err != nil {
+		l.Errorf("Failed to start server: %s\n ", err.Error())
+	}
+}
+
+func reloadDns(l *logrus.Logger, c *config.C) {
+	if dnsAddr == getDnsServerAddr(c) {
+		l.Debug("No DNS server config change detected")
+		return
+	}
+
+	l.Debug("Restarting DNS server")
+	dnsServer.Shutdown()
+	go startDns(l, c)
+}

dns_server_test.go

@@ -1,31 +1,19 @@
 package nebula
 
 import (
-	"context"
-	"io"
-	"net"
 	"net/netip"
-	"strconv"
 	"testing"
-	"time"
 
 	"github.com/miekg/dns"
 	"github.com/sirupsen/logrus"
 	"github.com/slackhq/nebula/config"
 	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
 )
 
 func TestParsequery(t *testing.T) {
 	l := logrus.New()
 	hostMap := &HostMap{}
-	ds := &dnsServer{
+	ds := newDnsRecords(l, &CertState{}, hostMap)
-		l:       l,
-		dnsMap4: make(map[string]netip.Addr),
-		dnsMap6: make(map[string]netip.Addr),
-		hostMap: hostMap,
-	}
-	ds.enabled.Store(true)
 	addrs := []netip.Addr{
 		netip.MustParseAddr("1.2.3.4"),
 		netip.MustParseAddr("1.2.3.5"),
@@ -83,209 +71,3 @@ func Test_getDnsServerAddr(t *testing.T) {
 	}
 	assert.Equal(t, "[::]:1", getDnsServerAddr(c))
 }
-
-func newTestDnsServer(t *testing.T) (*dnsServer, *config.C) {
-	t.Helper()
-	l := logrus.New()
-	l.Out = io.Discard
-	ds := &dnsServer{
-		l:       l,
-		ctx:     context.Background(),
-		dnsMap4: make(map[string]netip.Addr),
-		dnsMap6: make(map[string]netip.Addr),
-		hostMap: &HostMap{},
-	}
-	ds.mux = dns.NewServeMux()
-	ds.mux.HandleFunc(".", ds.handleDnsRequest)
-	return ds, config.NewC(l)
-}
-
-func setDnsConfig(c *config.C, host string, port string, amLighthouse, serveDns bool) {
-	c.Settings["lighthouse"] = map[string]any{
-		"am_lighthouse": amLighthouse,
-		"serve_dns":     serveDns,
-		"dns": map[string]any{
-			"host": host,
-			"port": port,
-		},
-	}
-}
-
-func TestDnsServer_reload_initial_disabled(t *testing.T) {
-	ds, c := newTestDnsServer(t)
-	setDnsConfig(c, "127.0.0.1", "0", true, false)
-
-	require.NoError(t, ds.reload(c, true))
-	assert.False(t, ds.enabled.Load())
-	assert.Equal(t, "127.0.0.1:0", ds.addr)
-	assert.Nil(t, ds.server)
-}
-
-func TestDnsServer_reload_initial_enabled(t *testing.T) {
-	ds, c := newTestDnsServer(t)
-	setDnsConfig(c, "127.0.0.1", "0", true, true)
-
-	require.NoError(t, ds.reload(c, true))
-	assert.True(t, ds.enabled.Load())
-	assert.Equal(t, "127.0.0.1:0", ds.addr)
-	// initial never starts a runner; that's Control.Start's job
-	assert.Nil(t, ds.server)
-}
-
-func TestDnsServer_reload_initial_serveDnsWithoutLighthouse(t *testing.T) {
-	ds, c := newTestDnsServer(t)
-	setDnsConfig(c, "127.0.0.1", "0", false, true)
-
-	require.NoError(t, ds.reload(c, true))
-	// Wants DNS but isn't a lighthouse: gated off, no runner.
-	assert.False(t, ds.enabled.Load())
-}
-
-func TestDnsServer_reload_sameAddr_noOp(t *testing.T) {
-	ds, c := newTestDnsServer(t)
-	setDnsConfig(c, "127.0.0.1", "0", true, true)
-
-	require.NoError(t, ds.reload(c, true))
-	// No server running yet, no addr change. Reload should not spawn anything.
-	require.NoError(t, ds.reload(c, false))
-	assert.True(t, ds.enabled.Load())
-	assert.Nil(t, ds.server)
-}
-
-func TestDnsServer_StartStop_lifecycle(t *testing.T) {
-	// Bind to a real (random) UDP port so we exercise the actual
-	// ListenAndServe + Shutdown plumbing including the started-chan race fix.
-	port := freeUDPPort(t)
-
-	ds, c := newTestDnsServer(t)
-	setDnsConfig(c, "127.0.0.1", port, true, true)
-	require.NoError(t, ds.reload(c, true))
-
-	done := make(chan struct{})
-	go func() {
-		ds.Start()
-		close(done)
-	}()
-
-	waitFor(t, func() bool {
-		ds.serverMu.Lock()
-		started := ds.started
-		ds.serverMu.Unlock()
-		if started == nil {
-			return false
-		}
-		select {
-		case <-started:
-			return true
-		default:
-			return false
-		}
-	})
-
-	ds.Stop()
-	select {
-	case <-done:
-	case <-time.After(5 * time.Second):
-		t.Fatal("Start did not return after Stop")
-	}
-}
-
-func TestDnsServer_Stop_beforeBind_doesNotHang(t *testing.T) {
-	// Stop called immediately after Start should not deadlock even if bind
-	// hasn't completed yet. This exercises the started-chan close-on-bind-fail
-	// path: by binding to an obviously bad port (privileged) we get a fast
-	// bind error before NotifyStartedFunc fires.
-	ds, c := newTestDnsServer(t)
-	// Use a port that should fail to bind (negative would be invalid, use a
-	// host that won't resolve to ensure listenUDP fails quickly).
-	setDnsConfig(c, "256.256.256.256", "53", true, true)
-	require.NoError(t, ds.reload(c, true))
-
-	done := make(chan struct{})
-	go func() {
-		ds.Start()
-		close(done)
-	}()
-
-	// Give Start a moment to attempt the bind and fail.
-	select {
-	case <-done:
-		// Bind failed and Start returned; Stop should be a no-op.
-	case <-time.After(time.Second):
-		t.Fatal("Start did not return after a bad bind")
-	}
-
-	stopped := make(chan struct{})
-	go func() {
-		ds.Stop()
-		close(stopped)
-	}()
-	select {
-	case <-stopped:
-	case <-time.After(time.Second):
-		t.Fatal("Stop hung after a failed bind")
-	}
-}
-
-func TestDnsServer_reload_disable_stopsRunningServer(t *testing.T) {
-	port := freeUDPPort(t)
-	ds, c := newTestDnsServer(t)
-	setDnsConfig(c, "127.0.0.1", port, true, true)
-	require.NoError(t, ds.reload(c, true))
-
-	startReturned := make(chan struct{})
-	go func() {
-		ds.Start()
-		close(startReturned)
-	}()
-	waitForBind(t, ds)
-
-	// Toggle serve_dns off; reload should shut the running server down.
-	setDnsConfig(c, "127.0.0.1", port, true, false)
-	require.NoError(t, ds.reload(c, false))
-	select {
-	case <-startReturned:
-	case <-time.After(5 * time.Second):
-		t.Fatal("Start did not return after reload disabled DNS")
-	}
-	assert.False(t, ds.enabled.Load())
-}
-
-func freeUDPPort(t *testing.T) string {
-	t.Helper()
-	conn, err := net.ListenPacket("udp", "127.0.0.1:0")
-	require.NoError(t, err)
-	port := conn.LocalAddr().(*net.UDPAddr).Port
-	require.NoError(t, conn.Close())
-	return strconv.Itoa(port)
-}
-
-func waitForBind(t *testing.T, ds *dnsServer) {
-	t.Helper()
-	waitFor(t, func() bool {
-		ds.serverMu.Lock()
-		started := ds.started
-		ds.serverMu.Unlock()
-		if started == nil {
-			return false
-		}
-		select {
-		case <-started:
-			return true
-		default:
-			return false
-		}
-	})
-}
-
-func waitFor(t *testing.T, cond func() bool) {
-	t.Helper()
-	deadline := time.Now().Add(5 * time.Second)
-	for time.Now().Before(deadline) {
-		if cond() {
-			return
-		}
-		time.Sleep(5 * time.Millisecond)
-	}
-	t.Fatal("timed out waiting for condition")
-}

firewall.go

@@ -20,7 +20,6 @@ import (
 	"github.com/slackhq/nebula/cert"
 	"github.com/slackhq/nebula/config"
 	"github.com/slackhq/nebula/firewall"
-	"github.com/slackhq/nebula/firewall/events"
 )
 
 type FirewallInterface interface {
@@ -68,14 +67,6 @@ type Firewall struct {
 	incomingMetrics     firewallMetrics
 	outgoingMetrics     firewallMetrics
 
-	// reporter is the optional embedder-supplied event sink. Immutable for
-	// the lifetime of this Firewall; Control.SetFirewallEventReporter
-	// installs it by shallow-copying the Firewall under the conntrack lock
-	// and swapping the pointer, and reloadFirewall carries it forward.
-	// Read unsynchronized on the data path: the preceding Firewall-pointer
-	// read pins the field's value for the duration of that call.
-	reporter events.Reporter
-
 	l *logrus.Logger
 }
 
@@ -425,27 +416,23 @@ var ErrNoMatchingRule = errors.New("no matching rule in firewall table")
 
 // Drop returns an error if the packet should be dropped, explaining why. It
 // returns nil if the packet should not be dropped.
-func (f *Firewall) Drop(fp firewall.Packet, ctx firewall.PacketContext, incoming bool, h *HostInfo, caPool *cert.CAPool, localCache firewall.ConntrackCache) error {
+func (f *Firewall) Drop(fp firewall.Packet, incoming bool, h *HostInfo, caPool *cert.CAPool, localCache firewall.ConntrackCache) error {
 	// Check if we spoke to this tuple, if we did then allow this packet
 	if f.inConns(fp, h, caPool, localCache) {
 		return nil
 	}
 
-	peerCert := h.ConnectionState.peerCert
-
 	// Make sure remote address matches nebula certificate, and determine how to treat it
 	if h.networks == nil {
 		// Simple case: Certificate has one address and no unsafe networks
 		if h.vpnAddrs[0] != fp.RemoteAddr {
 			f.metrics(incoming).droppedRemoteAddr.Inc(1)
-			f.reportDrop(incoming, events.DropInvalidRemoteIP, fp, ctx, peerCert)
 			return ErrInvalidRemoteIP
 		}
 	} else {
 		nwType, ok := h.networks.Lookup(fp.RemoteAddr)
 		if !ok {
 			f.metrics(incoming).droppedRemoteAddr.Inc(1)
-			f.reportDrop(incoming, events.DropInvalidRemoteIP, fp, ctx, peerCert)
 			return ErrInvalidRemoteIP
 		}
 		switch nwType {
@@ -453,13 +440,11 @@ func (f *Firewall) Drop(fp firewall.Packet, ctx firewall.PacketContext, incoming
 			break // nothing special
 		case NetworkTypeVPNPeer:
 			f.metrics(incoming).droppedRemoteAddr.Inc(1)
-			f.reportDrop(incoming, events.DropPeerRejected, fp, ctx, peerCert)
 			return ErrPeerRejected // reject for now, one day this may have different FW rules
 		case NetworkTypeUnsafe:
 			break // nothing special, one day this may have different FW rules
 		default:
 			f.metrics(incoming).droppedRemoteAddr.Inc(1)
-			f.reportDrop(incoming, events.DropUnknownNetwork, fp, ctx, peerCert)
 			return ErrUnknownNetworkType //should never happen
 		}
 	}
@@ -467,7 +452,6 @@ func (f *Firewall) Drop(fp firewall.Packet, ctx firewall.PacketContext, incoming
 	// Make sure we are supposed to be handling this local ip address
 	if !f.routableNetworks.Contains(fp.LocalAddr) {
 		f.metrics(incoming).droppedLocalAddr.Inc(1)
-		f.reportDrop(incoming, events.DropInvalidLocalIP, fp, ctx, peerCert)
 		return ErrInvalidLocalIP
 	}
 
@@ -477,14 +461,13 @@ func (f *Firewall) Drop(fp firewall.Packet, ctx firewall.PacketContext, incoming
 	}
 
 	// We now know which firewall table to check against
-	if !table.match(fp, incoming, peerCert, caPool) {
+	if !table.match(fp, incoming, h.ConnectionState.peerCert, caPool) {
 		f.metrics(incoming).droppedNoRule.Inc(1)
-		f.reportDrop(incoming, events.DropNoMatchingRule, fp, ctx, peerCert)
 		return ErrNoMatchingRule
 	}
 
 	// We always want to conntrack since it is a faster operation
-	f.addConn(fp, ctx, incoming, peerCert)
+	f.addConn(fp, incoming)
 
 	return nil
 }
@@ -503,59 +486,6 @@ func (f *Firewall) Destroy() {
 	//TODO: clean references if/when needed
 }
 
-func (f *Firewall) reportDrop(incoming bool, reason events.DropReason, fp firewall.Packet, ctx firewall.PacketContext, peerCert *cert.CachedCertificate) {
-	r := f.reporter
-	if r == nil {
-		return
-	}
-	r.ReportDrop(events.DropEvent{
-		Incoming:     incoming,
-		Reason:       reason,
-		Packet:       fp,
-		Context:      ctx,
-		PeerCert:     peerCert,
-		RulesVersion: f.rulesVersion,
-	})
-}
-
-func (f *Firewall) reportFlowCreate(incoming bool, fp firewall.Packet, ctx firewall.PacketContext, peerCert *cert.CachedCertificate) {
-	r := f.reporter
-	if r == nil {
-		return
-	}
-	r.ReportFlowCreate(events.FlowCreateEvent{
-		Incoming:     incoming,
-		Packet:       fp,
-		Context:      ctx,
-		PeerCert:     peerCert,
-		RulesVersion: f.rulesVersion,
-	})
-}
-
-func (f *Firewall) reportFlowEvict(incoming bool, fp firewall.Packet, rulesVersion uint16, expired bool) {
-	r := f.reporter
-	if r == nil {
-		return
-	}
-	r.ReportFlowEvict(events.FlowEvictEvent{
-		Incoming:     incoming,
-		Packet:       fp,
-		RulesVersion: rulesVersion,
-		Expired:      expired,
-	})
-}
-
-func (f *Firewall) reportRulesReload(oldVersion, newVersion uint16) {
-	r := f.reporter
-	if r == nil {
-		return
-	}
-	r.ReportRulesReload(events.RulesReloadEvent{
-		OldVersion: oldVersion,
-		NewVersion: newVersion,
-	})
-}
-
 func (f *Firewall) EmitStats() {
 	conntrack := f.Conntrack
 	conntrack.Lock()
@@ -606,9 +536,7 @@ func (f *Firewall) inConns(fp firewall.Packet, h *HostInfo, caPool *cert.CAPool,
 					WithField("oldRulesVersion", c.rulesVersion).
 					Debugln("dropping old conntrack entry, does not match new ruleset")
 			}
-			oldRulesVersion := c.rulesVersion
 			delete(conntrack.Conns, fp)
-			f.reportFlowEvict(c.incoming, fp, oldRulesVersion, false)
 			conntrack.Unlock()
 			return false
 		}
@@ -643,7 +571,7 @@ func (f *Firewall) inConns(fp firewall.Packet, h *HostInfo, caPool *cert.CAPool,
 	return true
 }
 
-func (f *Firewall) addConn(fp firewall.Packet, ctx firewall.PacketContext, incoming bool, peerCert *cert.CachedCertificate) {
+func (f *Firewall) addConn(fp firewall.Packet, incoming bool) {
 	var timeout time.Duration
 	c := &conn{}
 
@@ -658,8 +586,7 @@ func (f *Firewall) addConn(fp firewall.Packet, ctx firewall.PacketContext, incom
 
 	conntrack := f.Conntrack
 	conntrack.Lock()
-	_, existing := conntrack.Conns[fp]
+	if _, ok := conntrack.Conns[fp]; !ok {
-	if !existing {
 		conntrack.TimerWheel.Advance(time.Now())
 		conntrack.TimerWheel.Add(fp, timeout)
 	}
@@ -670,13 +597,6 @@ func (f *Firewall) addConn(fp firewall.Packet, ctx firewall.PacketContext, incom
 	c.rulesVersion = f.rulesVersion
 	c.Expires = time.Now().Add(timeout)
 	conntrack.Conns[fp] = c
-
-	// Report only when this represents a genuinely new flow. Fires under the
-	// conntrack lock so FlowCreate/FlowEvict events stay ordered relative to
-	// RulesReloadEvent, which also fires under this lock.
-	if !existing {
-		f.reportFlowCreate(incoming, fp, ctx, peerCert)
-	}
 	conntrack.Unlock()
 }
 
@@ -700,10 +620,7 @@ func (f *Firewall) evict(p firewall.Packet) {
 	}
 
 	// This conn is done
-	rulesVersion := t.rulesVersion
-	incoming := t.incoming
 	delete(conntrack.Conns, p)
-	f.reportFlowEvict(incoming, p, rulesVersion, true)
 }
 
 func (ft *FirewallTable) match(p firewall.Packet, incoming bool, c *cert.CachedCertificate, caPool *cert.CAPool) bool {

firewall/cache.go

@@ -1,7 +1,6 @@
 package firewall
 
 import (
-	"context"
 	"sync/atomic"
 	"time"
 
@@ -19,7 +18,7 @@ type ConntrackCacheTicker struct {
 	cache ConntrackCache
 }
 
-func NewConntrackCacheTicker(ctx context.Context, d time.Duration) *ConntrackCacheTicker {
+func NewConntrackCacheTicker(d time.Duration) *ConntrackCacheTicker {
 	if d == 0 {
 		return nil
 	}
@@ -28,23 +27,17 @@ func NewConntrackCacheTicker(ctx context.Context, d time.Duration) *ConntrackCac
 		cache: ConntrackCache{},
 	}
 
-	go c.tick(ctx, d)
+	go c.tick(d)
 
 	return c
 }
 
-func (c *ConntrackCacheTicker) tick(ctx context.Context, d time.Duration) {
+func (c *ConntrackCacheTicker) tick(d time.Duration) {
-	t := time.NewTicker(d)
-	defer t.Stop()
 	for {
-		select {
+		time.Sleep(d)
-		case <-ctx.Done():
-			return
-		case <-t.C:
 		c.cacheTick.Add(1)
 	}
 }
-}
 
 // Get checks if the cache ticker has moved to the next version before returning
 // the map. If it has moved, we reset the map.

firewall/events/events.go

@@ -1,102 +0,0 @@
-// Package events defines the opt-in firewall event reporting interface.
-//
-// Nebula emits raw packet-level events (drops, flow creations, flow evictions,
-// rule reloads) and does no aggregation, counting, batching, rule-description,
-// transport, or timestamping. Embedders correlate events back to yaml rules
-// out of band and capture whatever clock they need themselves. All Report*
-// methods are invoked while nebula holds internal locks and must be
-// non-blocking.
-//
-// Events are passed to Report* methods by value. Implementations must not
-// take the address of a received event: doing so forces Go's escape
-// analysis to move the event to the heap and costs one allocation per call.
-// To forward an event, either copy its fields into the reporter's own
-// pooled record or send it through a value-typed channel (chan DropEvent,
-// not chan *DropEvent).
-package events
-
-import (
-	"github.com/slackhq/nebula/cert"
-	"github.com/slackhq/nebula/firewall"
-)
-
-type DropReason uint8
-
-const (
-	DropInvalidLocalIP DropReason = iota
-	DropInvalidRemoteIP
-	DropPeerRejected
-	DropUnknownNetwork
-	DropNoMatchingRule
-)
-
-func (r DropReason) String() string {
-	switch r {
-	case DropInvalidLocalIP:
-		return "invalid_local_ip"
-	case DropInvalidRemoteIP:
-		return "invalid_remote_ip"
-	case DropPeerRejected:
-		return "peer_rejected"
-	case DropUnknownNetwork:
-		return "unknown_network"
-	case DropNoMatchingRule:
-		return "no_matching_rule"
-	default:
-		return "unknown"
-	}
-}
-
-// DropEvent is emitted for every packet that fails the firewall check. Drops
-// are not aggregated; every drop produces one event.
-type DropEvent struct {
-	Incoming     bool
-	Reason       DropReason
-	Packet       firewall.Packet
-	Context      firewall.PacketContext
-	PeerCert     *cert.CachedCertificate
-	RulesVersion uint16
-}
-
-// FlowCreateEvent is emitted when a packet is allowed and a new conntrack
-// entry is created. Subsequent packets in the same flow do not re-emit.
-type FlowCreateEvent struct {
-	Incoming     bool
-	Packet       firewall.Packet
-	Context      firewall.PacketContext
-	PeerCert     *cert.CachedCertificate
-	RulesVersion uint16
-}
-
-// FlowEvictEvent is emitted when a conntrack entry is removed. Context is
-// not carried: timer-wheel eviction has no packet in hand, and reload
-// revalidation evicts the OLD flow rather than the triggering packet.
-// RulesVersion is the version under which the flow was originally allowed,
-// which may differ from the current firewall version.
-type FlowEvictEvent struct {
-	Incoming     bool
-	Packet       firewall.Packet
-	RulesVersion uint16
-	// Expired is true when eviction was due to conntrack timeout; false when
-	// the entry was removed because it failed re-validation after a reload.
-	Expired bool
-}
-
-// RulesReloadEvent is emitted once after each successful firewall reload.
-// Reporters that bucket state by RulesVersion should close the old bucket
-// and open a new one on receipt.
-type RulesReloadEvent struct {
-	OldVersion uint16
-	NewVersion uint16
-}
-
-// Reporter is the embedder-supplied sink for firewall events. Implementations
-// that want a timestamp should call time.Now() themselves at the top of the
-// method; nebula does not provide one. See the package doc for the
-// do-not-take-address rule.
-type Reporter interface {
-	ReportDrop(DropEvent)
-	ReportFlowCreate(FlowCreateEvent)
-	ReportFlowEvict(FlowEvictEvent)
-	ReportRulesReload(RulesReloadEvent)
-}

firewall/packet.go

@@ -31,27 +31,6 @@ type Packet struct {
 	Fragment   bool
 }
 
-// PacketContext carries additional parsed details about a packet that are
-// useful for event reporting but deliberately kept out of Packet so Packet
-// can keep being used as a conntrack map key. Populated alongside Packet by
-// newPacket.
-//
-// Fields are interpreted based on Packet.Protocol:
-//   - ProtoTCP: TCPFlags is meaningful; ICMPType / ICMPCode are zero
-//   - ProtoICMP, ProtoICMPv6: ICMPType / ICMPCode are meaningful; TCPFlags is zero
-//   - ProtoUDP and others: only Length is meaningful
-type PacketContext struct {
-	// Length is the total IP packet length in bytes, including headers.
-	Length uint16
-	// TCPFlags is the flag byte from the TCP header (bits for FIN, SYN, RST,
-	// PSH, ACK, URG, ECE, CWR).
-	TCPFlags uint8
-	// ICMPType is the type field of the ICMP / ICMPv6 header.
-	ICMPType uint8
-	// ICMPCode is the code field of the ICMP / ICMPv6 header.
-	ICMPCode uint8
-}
-
 func (fp *Packet) Copy() *Packet {
 	return &Packet{
 		LocalAddr:  fp.LocalAddr,

firewall_events_test.go

@@ -1,731 +0,0 @@
-package nebula
-
-import (
-	"net"
-	"net/netip"
-	"sync"
-	"testing"
-	"time"
-
-	"github.com/gaissmai/bart"
-	"github.com/google/gopacket"
-	"github.com/google/gopacket/layers"
-	"github.com/slackhq/nebula/cert"
-	"github.com/slackhq/nebula/firewall"
-	"github.com/slackhq/nebula/firewall/events"
-	"github.com/slackhq/nebula/test"
-	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
-)
-
-// recordingReporter captures every event fired against it. Its methods take
-// the conntrack lock implicitly (via the firewall code path that invokes
-// them), so we synchronize accumulator mutations with a small mutex to keep
-// the race detector happy across goroutines in case a test introduces any.
-type recordingReporter struct {
-	mu      sync.Mutex
-	drops   []recordedDrop
-	creates []recordedCreate
-	evicts  []recordedEvict
-	reloads []recordedReload
-}
-
-type recordedDrop struct {
-	incoming     bool
-	reason       events.DropReason
-	remote       netip.Addr
-	local        netip.Addr
-	peerName     string
-	rulesVersion uint16
-	ctx          firewall.PacketContext
-}
-
-type recordedCreate struct {
-	incoming     bool
-	remote       netip.Addr
-	local        netip.Addr
-	peerName     string
-	rulesVersion uint16
-	ctx          firewall.PacketContext
-}
-
-type recordedEvict struct {
-	incoming     bool
-	remote       netip.Addr
-	local        netip.Addr
-	rulesVersion uint16
-	expired      bool
-}
-
-type recordedReload struct {
-	oldVersion uint16
-	newVersion uint16
-}
-
-func (r *recordingReporter) ReportDrop(e events.DropEvent) {
-	r.mu.Lock()
-	defer r.mu.Unlock()
-	name := ""
-	if e.PeerCert != nil && e.PeerCert.Certificate != nil {
-		name = e.PeerCert.Certificate.Name()
-	}
-	r.drops = append(r.drops, recordedDrop{
-		incoming:     e.Incoming,
-		reason:       e.Reason,
-		remote:       e.Packet.RemoteAddr,
-		local:        e.Packet.LocalAddr,
-		peerName:     name,
-		rulesVersion: e.RulesVersion,
-		ctx:          e.Context,
-	})
-}
-
-func (r *recordingReporter) ReportFlowCreate(e events.FlowCreateEvent) {
-	r.mu.Lock()
-	defer r.mu.Unlock()
-	name := ""
-	if e.PeerCert != nil && e.PeerCert.Certificate != nil {
-		name = e.PeerCert.Certificate.Name()
-	}
-	r.creates = append(r.creates, recordedCreate{
-		incoming:     e.Incoming,
-		remote:       e.Packet.RemoteAddr,
-		local:        e.Packet.LocalAddr,
-		peerName:     name,
-		rulesVersion: e.RulesVersion,
-		ctx:          e.Context,
-	})
-}
-
-func (r *recordingReporter) ReportFlowEvict(e events.FlowEvictEvent) {
-	r.mu.Lock()
-	defer r.mu.Unlock()
-	r.evicts = append(r.evicts, recordedEvict{
-		incoming:     e.Incoming,
-		remote:       e.Packet.RemoteAddr,
-		local:        e.Packet.LocalAddr,
-		rulesVersion: e.RulesVersion,
-		expired:      e.Expired,
-	})
-}
-
-func (r *recordingReporter) ReportRulesReload(e events.RulesReloadEvent) {
-	r.mu.Lock()
-	defer r.mu.Unlock()
-	r.reloads = append(r.reloads, recordedReload{
-		oldVersion: e.OldVersion,
-		newVersion: e.NewVersion,
-	})
-}
-
-// eventFixture builds a Firewall wired to a Control plus a packet/hostinfo
-// pair that a test can reuse. By default the ruleset allows the packet;
-// callers mutate fw / p / h as needed before invoking Drop.
-type eventFixture struct {
-	ctl *Control
-	fw  *Firewall
-	p   firewall.Packet
-	h   *HostInfo
-	cp  *cert.CAPool
-}
-
-func newEventFixture(t *testing.T) *eventFixture {
-	t.Helper()
-	l := test.NewLogger()
-
-	// myVpnNetworksTable covers our single peer address so buildNetworks takes
-	// the "simple case" path (h.networks stays nil); tests that want a populated
-	// BART table overwrite h.networks directly.
-	vpnNetworks := new(bart.Lite)
-	vpnNetworks.Insert(netip.MustParsePrefix("1.2.3.0/24"))
-
-	// Use the same cert for "peer" and "local" endpoints, matching the
-	// TestFirewall_Drop fixture style: LocalAddr == RemoteAddr == peer vpn addr.
-	c := &dummyCert{
-		name:     "host1",
-		networks: []netip.Prefix{netip.MustParsePrefix("1.2.3.4/24")},
-		groups:   []string{"default-group"},
-		issuer:   "signer-shasum",
-	}
-	h := &HostInfo{
-		ConnectionState: &ConnectionState{
-			peerCert: &cert.CachedCertificate{
-				Certificate:    c,
-				InvertedGroups: map[string]struct{}{"default-group": {}},
-			},
-		},
-		vpnAddrs: []netip.Addr{netip.MustParseAddr("1.2.3.4")},
-	}
-	h.buildNetworks(vpnNetworks, c)
-
-	fw := NewFirewall(l, time.Minute, time.Minute, time.Minute, c)
-	require.NoError(t, fw.AddRule(true, firewall.ProtoAny, 0, 0, []string{"any"}, "", "", "", "", ""))
-	require.NoError(t, fw.AddRule(false, firewall.ProtoAny, 0, 0, []string{"any"}, "", "", "", "", ""))
-
-	ctl := &Control{
-		f: &Interface{firewall: fw},
-		l: l,
-	}
-
-	return &eventFixture{
-		ctl: ctl,
-		fw:  fw,
-		p: firewall.Packet{
-			LocalAddr:  netip.MustParseAddr("1.2.3.4"),
-			RemoteAddr: netip.MustParseAddr("1.2.3.4"),
-			LocalPort:  10,
-			RemotePort: 90,
-			Protocol:   firewall.ProtoUDP,
-		},
-		h:  h,
-		cp: cert.NewCAPool(),
-	}
-}
-
-// firewall() returns the currently-installed firewall. Needed because
-// SetFirewallEventReporter replaces it via shallow-copy swap.
-func (f *eventFixture) firewall() *Firewall {
-	return f.ctl.f.firewall
-}
-
-func TestEvents_ReportDrop_InvalidRemoteIP(t *testing.T) {
-	f := newEventFixture(t)
-	r := &recordingReporter{}
-	f.ctl.SetFirewallEventReporter(r)
-
-	// Packet to an address not in the cert's networks.
-	f.p.RemoteAddr = netip.MustParseAddr("9.9.9.9")
-	assert.Equal(t, ErrInvalidRemoteIP, f.firewall().Drop(f.p, firewall.PacketContext{}, false, f.h, f.cp, nil))
-
-	require.Len(t, r.drops, 1)
-	assert.Equal(t, events.DropInvalidRemoteIP, r.drops[0].reason)
-	assert.False(t, r.drops[0].incoming)
-	assert.Equal(t, "host1", r.drops[0].peerName)
-	assert.Empty(t, r.creates)
-	assert.Empty(t, r.evicts)
-}
-
-func TestEvents_ReportDrop_InvalidLocalIP(t *testing.T) {
-	f := newEventFixture(t)
-	r := &recordingReporter{}
-	f.ctl.SetFirewallEventReporter(r)
-
-	// LocalAddr outside our routable networks.
-	f.p.LocalAddr = netip.MustParseAddr("9.9.9.9")
-	assert.Equal(t, ErrInvalidLocalIP, f.firewall().Drop(f.p, firewall.PacketContext{}, true, f.h, f.cp, nil))
-
-	require.Len(t, r.drops, 1)
-	assert.Equal(t, events.DropInvalidLocalIP, r.drops[0].reason)
-	assert.True(t, r.drops[0].incoming)
-}
-
-func TestEvents_ReportDrop_NoMatchingRule(t *testing.T) {
-	f := newEventFixture(t)
-	// Reset to a firewall with no matching rule.
-	l := test.NewLogger()
-	fw := NewFirewall(l, time.Minute, time.Minute, time.Minute, f.h.ConnectionState.peerCert.Certificate)
-	// Rule that won't match (group not in peer's groups).
-	require.NoError(t, fw.AddRule(true, firewall.ProtoAny, 0, 0, []string{"nope"}, "", "", "", "", ""))
-	require.NoError(t, fw.AddRule(false, firewall.ProtoAny, 0, 0, []string{"nope"}, "", "", "", "", ""))
-	f.ctl.f.firewall = fw
-
-	r := &recordingReporter{}
-	f.ctl.SetFirewallEventReporter(r)
-
-	assert.Equal(t, ErrNoMatchingRule, f.firewall().Drop(f.p, firewall.PacketContext{}, true, f.h, f.cp, nil))
-	require.Len(t, r.drops, 1)
-	assert.Equal(t, events.DropNoMatchingRule, r.drops[0].reason)
-}
-
-func TestEvents_ReportDrop_PeerRejected(t *testing.T) {
-	f := newEventFixture(t)
-	// Re-classify the remote as VPNPeer so it triggers DropPeerRejected.
-	f.h.networks = new(bart.Table[NetworkType])
-	f.h.networks.Insert(netip.MustParsePrefix("1.2.3.0/24"), NetworkTypeVPNPeer)
-
-	r := &recordingReporter{}
-	f.ctl.SetFirewallEventReporter(r)
-
-	assert.Equal(t, ErrPeerRejected, f.firewall().Drop(f.p, firewall.PacketContext{}, true, f.h, f.cp, nil))
-	require.Len(t, r.drops, 1)
-	assert.Equal(t, events.DropPeerRejected, r.drops[0].reason)
-}
-
-func TestEvents_ReportDrop_UnknownNetwork(t *testing.T) {
-	f := newEventFixture(t)
-	// Insert an unrecognized NetworkType value to hit the default branch.
-	f.h.networks = new(bart.Table[NetworkType])
-	f.h.networks.Insert(netip.MustParsePrefix("1.2.3.0/24"), NetworkTypeUnknown)
-
-	r := &recordingReporter{}
-	f.ctl.SetFirewallEventReporter(r)
-
-	assert.Equal(t, ErrUnknownNetworkType, f.firewall().Drop(f.p, firewall.PacketContext{}, true, f.h, f.cp, nil))
-	require.Len(t, r.drops, 1)
-	assert.Equal(t, events.DropUnknownNetwork, r.drops[0].reason)
-}
-
-func TestEvents_ReportFlowCreate_OnceOnly(t *testing.T) {
-	f := newEventFixture(t)
-	r := &recordingReporter{}
-	f.ctl.SetFirewallEventReporter(r)
-
-	// First allowed packet creates the conntrack entry.
-	require.NoError(t, f.firewall().Drop(f.p, firewall.PacketContext{}, true, f.h, f.cp, nil))
-	// Second matching packet on the same tuple is short-circuited by conntrack
-	// and must not fire another FlowCreate.
-	require.NoError(t, f.firewall().Drop(f.p, firewall.PacketContext{}, true, f.h, f.cp, nil))
-
-	require.Len(t, r.creates, 1)
-	assert.True(t, r.creates[0].incoming)
-	assert.Equal(t, f.p.RemoteAddr, r.creates[0].remote)
-	assert.Empty(t, r.drops)
-}
-
-func TestEvents_ReportFlowEvict_OnReloadPurge(t *testing.T) {
-	f := newEventFixture(t)
-	r := &recordingReporter{}
-	f.ctl.SetFirewallEventReporter(r)
-
-	// Create a flow under the current rules.
-	require.NoError(t, f.firewall().Drop(f.p, firewall.PacketContext{}, true, f.h, f.cp, nil))
-	require.Len(t, r.creates, 1)
-
-	// Simulate a reload that produces rules the existing flow no longer
-	// matches. Bump rulesVersion and replace InRules with an empty table so
-	// revalidation fails.
-	fw := f.firewall()
-	fw.Conntrack.Lock()
-	fw.rulesVersion++
-	fw.InRules = newFirewallTable()
-	fw.Conntrack.Unlock()
-
-	// Next packet triggers re-validation, which fails and evicts the entry.
-	err := fw.Drop(f.p, firewall.PacketContext{}, true, f.h, f.cp, nil)
-	assert.Equal(t, ErrNoMatchingRule, err)
-
-	require.Len(t, r.evicts, 1)
-	assert.False(t, r.evicts[0].expired, "evict from reload purge is not expiration")
-	assert.True(t, r.evicts[0].incoming)
-}
-
-func TestEvents_ReportFlowEvict_OnTimeout(t *testing.T) {
-	f := newEventFixture(t)
-	r := &recordingReporter{}
-	f.ctl.SetFirewallEventReporter(r)
-
-	require.NoError(t, f.firewall().Drop(f.p, firewall.PacketContext{}, true, f.h, f.cp, nil))
-	require.Len(t, r.creates, 1)
-
-	// Force expiration by rewinding the entry's deadline.
-	fw := f.firewall()
-	fw.Conntrack.Lock()
-	c := fw.Conntrack.Conns[f.p]
-	require.NotNil(t, c)
-	c.Expires = time.Now().Add(-time.Hour)
-	fw.evict(f.p)
-	fw.Conntrack.Unlock()
-
-	require.Len(t, r.evicts, 1)
-	assert.True(t, r.evicts[0].expired)
-}
-
-func TestEvents_SetNil_Clears(t *testing.T) {
-	f := newEventFixture(t)
-	r := &recordingReporter{}
-	f.ctl.SetFirewallEventReporter(r)
-	require.NoError(t, f.firewall().Drop(f.p, firewall.PacketContext{}, true, f.h, f.cp, nil))
-	require.Len(t, r.creates, 1)
-
-	f.ctl.SetFirewallEventReporter(nil)
-	resetConntrack(f.firewall())
-	require.NoError(t, f.firewall().Drop(f.p, firewall.PacketContext{}, true, f.h, f.cp, nil))
-	// No second create should be recorded.
-	assert.Len(t, r.creates, 1)
-}
-
-func TestEvents_ReporterSurvivesSwap(t *testing.T) {
-	f := newEventFixture(t)
-	r := &recordingReporter{}
-	f.ctl.SetFirewallEventReporter(r)
-
-	// Simulate a reload by swapping in a fresh Firewall that carries the
-	// reporter forward. Mirrors what reloadFirewall does with the shared
-	// conntrack pointer.
-	l := test.NewLogger()
-	oldFw := f.firewall()
-	newFw := NewFirewall(l, time.Minute, time.Minute, time.Minute, f.h.ConnectionState.peerCert.Certificate)
-	require.NoError(t, newFw.AddRule(true, firewall.ProtoAny, 0, 0, []string{"any"}, "", "", "", "", ""))
-	require.NoError(t, newFw.AddRule(false, firewall.ProtoAny, 0, 0, []string{"any"}, "", "", "", "", ""))
-	newFw.Conntrack = oldFw.Conntrack
-	newFw.rulesVersion = oldFw.rulesVersion + 1
-	newFw.reporter = oldFw.reporter
-	f.ctl.f.firewall = newFw
-	newFw.reportRulesReload(oldFw.rulesVersion, newFw.rulesVersion)
-
-	require.Len(t, r.reloads, 1)
-	assert.Equal(t, oldFw.rulesVersion, r.reloads[0].oldVersion)
-	assert.Equal(t, newFw.rulesVersion, r.reloads[0].newVersion)
-
-	// Events on the new firewall should still reach the same reporter.
-	require.NoError(t, newFw.Drop(f.p, firewall.PacketContext{}, true, f.h, f.cp, nil))
-	require.Len(t, r.creates, 1)
-	assert.Equal(t, newFw.rulesVersion, r.creates[0].rulesVersion)
-}
-
-func TestEvents_InstallDoesNotMutateOldFirewall(t *testing.T) {
-	f := newEventFixture(t)
-	before := f.firewall()
-
-	r := &recordingReporter{}
-	f.ctl.SetFirewallEventReporter(r)
-
-	after := f.firewall()
-	assert.NotSame(t, before, after, "SetFirewallEventReporter must replace the Firewall pointer")
-	assert.Nil(t, before.reporter, "the pre-install Firewall must remain untouched")
-	assert.NotNil(t, after.reporter)
-}
-
-// --- PacketContext parse tests --------------------------------------------
-
-func mustSerialize(t *testing.T, lrs ...gopacket.SerializableLayer) []byte {
-	t.Helper()
-	buf := gopacket.NewSerializeBuffer()
-	opt := gopacket.SerializeOptions{ComputeChecksums: false, FixLengths: true}
-	require.NoError(t, gopacket.SerializeLayers(buf, opt, lrs...))
-	return buf.Bytes()
-}
-
-func TestPacketContext_IPv4_TCPFlags(t *testing.T) {
-	ip := &layers.IPv4{
-		Version: 4, TTL: 64, Protocol: layers.IPProtocolTCP,
-		SrcIP: net.IPv4(10, 0, 0, 1), DstIP: net.IPv4(10, 0, 0, 2),
-	}
-	tcp := &layers.TCP{SrcPort: 1234, DstPort: 80, SYN: true, ACK: true}
-	require.NoError(t, tcp.SetNetworkLayerForChecksum(ip))
-	data := mustSerialize(t, ip, tcp, gopacket.Payload([]byte("hello")))
-
-	var fp firewall.Packet
-	var ctx firewall.PacketContext
-	require.NoError(t, newPacket(data, true, &fp, &ctx))
-
-	assert.Equal(t, uint8(firewall.ProtoTCP), fp.Protocol)
-	// SYN (0x02) + ACK (0x10) = 0x12
-	assert.Equal(t, uint8(0x12), ctx.TCPFlags)
-	assert.Equal(t, uint16(len(data)), ctx.Length)
-	assert.Equal(t, uint8(0), ctx.ICMPType)
-	assert.Equal(t, uint8(0), ctx.ICMPCode)
-}
-
-func TestPacketContext_IPv4_ICMPTypeCode(t *testing.T) {
-	ip := &layers.IPv4{
-		Version: 4, TTL: 64, Protocol: layers.IPProtocolICMPv4,
-		SrcIP: net.IPv4(10, 0, 0, 1), DstIP: net.IPv4(10, 0, 0, 2),
-	}
-	// Destination Unreachable, code 3 (port unreachable)
-	icmp := &layers.ICMPv4{
-		TypeCode: layers.CreateICMPv4TypeCode(layers.ICMPv4TypeDestinationUnreachable, layers.ICMPv4CodePort),
-	}
-	data := mustSerialize(t, ip, icmp, gopacket.Payload([]byte{0, 0, 0, 0}))
-
-	var fp firewall.Packet
-	var ctx firewall.PacketContext
-	require.NoError(t, newPacket(data, true, &fp, &ctx))
-
-	assert.Equal(t, uint8(firewall.ProtoICMP), fp.Protocol)
-	assert.Equal(t, uint8(layers.ICMPv4TypeDestinationUnreachable), ctx.ICMPType)
-	assert.Equal(t, uint8(layers.ICMPv4CodePort), ctx.ICMPCode)
-	assert.Equal(t, uint16(len(data)), ctx.Length)
-	assert.Equal(t, uint8(0), ctx.TCPFlags)
-}
-
-func TestPacketContext_IPv4_UDPLengthOnly(t *testing.T) {
-	ip := &layers.IPv4{
-		Version: 4, TTL: 64, Protocol: layers.IPProtocolUDP,
-		SrcIP: net.IPv4(10, 0, 0, 1), DstIP: net.IPv4(10, 0, 0, 2),
-	}
-	udp := &layers.UDP{SrcPort: 1234, DstPort: 53}
-	require.NoError(t, udp.SetNetworkLayerForChecksum(ip))
-	data := mustSerialize(t, ip, udp, gopacket.Payload([]byte("query")))
-
-	var fp firewall.Packet
-	var ctx firewall.PacketContext
-	require.NoError(t, newPacket(data, true, &fp, &ctx))
-
-	assert.Equal(t, uint8(firewall.ProtoUDP), fp.Protocol)
-	assert.Equal(t, uint16(len(data)), ctx.Length)
-	assert.Zero(t, ctx.TCPFlags)
-	assert.Zero(t, ctx.ICMPType)
-	assert.Zero(t, ctx.ICMPCode)
-}
-
-func TestPacketContext_IPv6_TCPFlags(t *testing.T) {
-	ip := &layers.IPv6{
-		Version: 6, HopLimit: 64, NextHeader: layers.IPProtocolTCP,
-		SrcIP: net.ParseIP("fd00::1"), DstIP: net.ParseIP("fd00::2"),
-	}
-	tcp := &layers.TCP{SrcPort: 1234, DstPort: 443, FIN: true, ACK: true}
-	require.NoError(t, tcp.SetNetworkLayerForChecksum(ip))
-	data := mustSerialize(t, ip, tcp, gopacket.Payload([]byte("bye")))
-
-	var fp firewall.Packet
-	var ctx firewall.PacketContext
-	require.NoError(t, newPacket(data, true, &fp, &ctx))
-
-	assert.Equal(t, uint8(firewall.ProtoTCP), fp.Protocol)
-	// FIN (0x01) + ACK (0x10) = 0x11
-	assert.Equal(t, uint8(0x11), ctx.TCPFlags)
-	assert.Equal(t, uint16(len(data)), ctx.Length)
-}
-
-func TestPacketContext_IPv6_ICMPv6TypeCode(t *testing.T) {
-	ip := &layers.IPv6{
-		Version: 6, HopLimit: 64, NextHeader: layers.IPProtocolICMPv6,
-		SrcIP: net.ParseIP("fd00::1"), DstIP: net.ParseIP("fd00::2"),
-	}
-	icmp := &layers.ICMPv6{
-		TypeCode: layers.CreateICMPv6TypeCode(layers.ICMPv6TypeDestinationUnreachable, layers.ICMPv6CodePortUnreachable),
-	}
-	require.NoError(t, icmp.SetNetworkLayerForChecksum(ip))
-	data := mustSerialize(t, ip, icmp, gopacket.Payload([]byte{0, 0, 0, 0, 0, 0, 0, 0}))
-
-	var fp firewall.Packet
-	var ctx firewall.PacketContext
-	require.NoError(t, newPacket(data, true, &fp, &ctx))
-
-	assert.Equal(t, uint8(firewall.ProtoICMPv6), fp.Protocol)
-	assert.Equal(t, uint8(layers.ICMPv6TypeDestinationUnreachable), ctx.ICMPType)
-	assert.Equal(t, uint8(layers.ICMPv6CodePortUnreachable), ctx.ICMPCode)
-	assert.Equal(t, uint16(len(data)), ctx.Length)
-}
-
-// TestPacketContext_NilOK confirms a nil context pointer is accepted by
-// newPacket (the hot path may elect not to pass one).
-func TestPacketContext_NilOK(t *testing.T) {
-	ip := &layers.IPv4{
-		Version: 4, TTL: 64, Protocol: layers.IPProtocolUDP,
-		SrcIP: net.IPv4(10, 0, 0, 1), DstIP: net.IPv4(10, 0, 0, 2),
-	}
-	udp := &layers.UDP{SrcPort: 1, DstPort: 2}
-	require.NoError(t, udp.SetNetworkLayerForChecksum(ip))
-	data := mustSerialize(t, ip, udp)
-
-	var fp firewall.Packet
-	require.NoError(t, newPacket(data, true, &fp, nil))
-}
-
-// TestPacketContext_FlowCreateCarriesContext exercises the full Drop -> addConn
-// -> ReportFlowCreate path with a realistic TCP packet and confirms the
-// context makes it into the reporter.
-func TestPacketContext_FlowCreateCarriesContext(t *testing.T) {
-	f := newEventFixture(t)
-	r := &recordingReporter{}
-	f.ctl.SetFirewallEventReporter(r)
-
-	// Hand-construct a matching TCP packet.
-	ctx := firewall.PacketContext{Length: 1500, TCPFlags: 0x12}
-	p := f.p
-	p.Protocol = firewall.ProtoTCP
-	require.NoError(t, f.firewall().Drop(p, ctx, true, f.h, f.cp, nil))
-
-	require.Len(t, r.creates, 1)
-	assert.Equal(t, uint16(1500), r.creates[0].ctx.Length)
-	assert.Equal(t, uint8(0x12), r.creates[0].ctx.TCPFlags)
-}
-
-// --- benchmarks ------------------------------------------------------------
-
-// noopReporter is the cheapest possible reporter. Methods discard the event.
-type noopReporter struct{}
-
-func (noopReporter) ReportDrop(events.DropEvent)             {}
-func (noopReporter) ReportFlowCreate(events.FlowCreateEvent) {}
-func (noopReporter) ReportFlowEvict(events.FlowEvictEvent)   {}
-func (noopReporter) ReportRulesReload(events.RulesReloadEvent) {
-}
-
-// bufferedReporter demonstrates a realistic zero-alloc reporter: each event
-// is forwarded to a value-typed channel. The channel send is a memcpy into
-// the channel's pre-allocated ring buffer -- no heap traffic. A background
-// goroutine would drain these; the bench skips draining to keep the report
-// path pure.
-type bufferedReporter struct {
-	drops  chan events.DropEvent
-	flows  chan events.FlowCreateEvent
-	evicts chan events.FlowEvictEvent
-}
-
-func newBufferedReporter(cap int) *bufferedReporter {
-	return &bufferedReporter{
-		drops:  make(chan events.DropEvent, cap),
-		flows:  make(chan events.FlowCreateEvent, cap),
-		evicts: make(chan events.FlowEvictEvent, cap),
-	}
-}
-
-func (r *bufferedReporter) ReportDrop(e events.DropEvent) {
-	select {
-	case r.drops <- e:
-	default:
-	}
-}
-
-func (r *bufferedReporter) ReportFlowCreate(e events.FlowCreateEvent) {
-	select {
-	case r.flows <- e:
-	default:
-	}
-}
-
-func (r *bufferedReporter) ReportFlowEvict(e events.FlowEvictEvent) {
-	select {
-	case r.evicts <- e:
-	default:
-	}
-}
-
-func (r *bufferedReporter) ReportRulesReload(events.RulesReloadEvent) {}
-
-// pointerReporter is the anti-pattern: it takes the address of the incoming
-// event struct, which forces the callee-side copy onto the heap. Kept for
-// comparison so we can see the alloc cost an unwary reporter would incur.
-type pointerReporter struct {
-	last *events.DropEvent
-}
-
-func (r *pointerReporter) ReportDrop(e events.DropEvent) {
-	r.last = &e
-}
-
-func (r *pointerReporter) ReportFlowCreate(events.FlowCreateEvent) {}
-func (r *pointerReporter) ReportFlowEvict(events.FlowEvictEvent)   {}
-func (r *pointerReporter) ReportRulesReload(events.RulesReloadEvent) {
-}
-
-func newBenchFixture(b *testing.B) *eventFixture {
-	b.Helper()
-	l := test.NewLogger()
-
-	vpnNetworks := new(bart.Lite)
-	vpnNetworks.Insert(netip.MustParsePrefix("1.2.3.0/24"))
-
-	c := &dummyCert{
-		name:     "host1",
-		networks: []netip.Prefix{netip.MustParsePrefix("1.2.3.4/24")},
-		groups:   []string{"default-group"},
-		issuer:   "signer-shasum",
-	}
-	h := &HostInfo{
-		ConnectionState: &ConnectionState{
-			peerCert: &cert.CachedCertificate{
-				Certificate:    c,
-				InvertedGroups: map[string]struct{}{"default-group": {}},
-			},
-		},
-		vpnAddrs: []netip.Addr{netip.MustParseAddr("1.2.3.4")},
-	}
-	h.buildNetworks(vpnNetworks, c)
-
-	fw := NewFirewall(l, time.Minute, time.Minute, time.Minute, c)
-	// Inbound rule that matches our packet; outbound has no match so we can
-	// also benchmark the no-rule drop path.
-	if err := fw.AddRule(true, firewall.ProtoAny, 0, 0, []string{"any"}, "", "", "", "", ""); err != nil {
-		b.Fatal(err)
-	}
-
-	ctl := &Control{f: &Interface{firewall: fw}, l: l}
-	return &eventFixture{
-		ctl: ctl,
-		fw:  fw,
-		p: firewall.Packet{
-			LocalAddr:  netip.MustParseAddr("1.2.3.4"),
-			RemoteAddr: netip.MustParseAddr("1.2.3.4"),
-			LocalPort:  10,
-			RemotePort: 90,
-			Protocol:   firewall.ProtoUDP,
-		},
-		h:  h,
-		cp: cert.NewCAPool(),
-	}
-}
-
-// BenchmarkFirewallDropPath measures the cost of Firewall.Drop on a packet
-// that reaches the no-matching-rule branch (the longest drop path). Compare
-// reporter shapes:
-//
-//	nilReporter      -- no reporter installed (feature cost when off)
-//	noopReporter     -- reporter installed, methods discard args (minimum on-cost)
-//	bufferedReporter -- realistic zero-alloc reporter: value-typed channels
-//	pointerReporter  -- anti-pattern that takes &composite-literal (allocates)
-func BenchmarkFirewallDropPath(b *testing.B) {
-	run := func(b *testing.B, install func(*Control)) {
-		f := newBenchFixture(b)
-		install(f.ctl)
-		b.ReportAllocs()
-		b.ResetTimer()
-		for i := 0; i < b.N; i++ {
-			_ = f.firewall().Drop(f.p, firewall.PacketContext{}, false, f.h, f.cp, nil)
-		}
-	}
-
-	b.Run("nilReporter", func(b *testing.B) { run(b, func(*Control) {}) })
-	b.Run("noopReporter", func(b *testing.B) {
-		run(b, func(c *Control) { c.SetFirewallEventReporter(noopReporter{}) })
-	})
-	b.Run("bufferedReporter", func(b *testing.B) {
-		run(b, func(c *Control) { c.SetFirewallEventReporter(newBufferedReporter(1024)) })
-	})
-	b.Run("pointerReporter", func(b *testing.B) {
-		run(b, func(c *Control) { c.SetFirewallEventReporter(&pointerReporter{}) })
-	})
-}
-
-// BenchmarkConntrackCreate measures Firewall.Drop for an allowed inbound
-// packet on a fresh conntrack (so addConn fires each iteration).
-func BenchmarkConntrackCreate(b *testing.B) {
-	run := func(b *testing.B, install func(*Control)) {
-		f := newBenchFixture(b)
-		install(f.ctl)
-		b.ReportAllocs()
-		b.ResetTimer()
-		for i := 0; i < b.N; i++ {
-			resetConntrack(f.firewall())
-			_ = f.firewall().Drop(f.p, firewall.PacketContext{}, true, f.h, f.cp, nil)
-		}
-	}
-	b.Run("nilReporter", func(b *testing.B) { run(b, func(*Control) {}) })
-	b.Run("noopReporter", func(b *testing.B) {
-		run(b, func(c *Control) { c.SetFirewallEventReporter(noopReporter{}) })
-	})
-	b.Run("bufferedReporter", func(b *testing.B) {
-		run(b, func(c *Control) { c.SetFirewallEventReporter(newBufferedReporter(1024)) })
-	})
-}
-
-// BenchmarkConntrackHit measures the hot path where a flow is already in
-// conntrack and short-circuits rule evaluation. The reporter slot is checked
-// only on create/evict, so this bench should show the reporter having zero
-// impact regardless of install state.
-func BenchmarkConntrackHit(b *testing.B) {
-	b.Run("nilReporter", func(b *testing.B) {
-		f := newBenchFixture(b)
-		// Prime conntrack.
-		require.NoError(b, f.firewall().Drop(f.p, firewall.PacketContext{}, true, f.h, f.cp, nil))
-		b.ReportAllocs()
-		b.ResetTimer()
-		for i := 0; i < b.N; i++ {
-			_ = f.firewall().Drop(f.p, firewall.PacketContext{}, true, f.h, f.cp, nil)
-		}
-	})
-
-	b.Run("noopReporter", func(b *testing.B) {
-		f := newBenchFixture(b)
-		f.ctl.SetFirewallEventReporter(noopReporter{})
-		require.NoError(b, f.firewall().Drop(f.p, firewall.PacketContext{}, true, f.h, f.cp, nil))
-		b.ReportAllocs()
-		b.ResetTimer()
-		for i := 0; i < b.N; i++ {
-			_ = f.firewall().Drop(f.p, firewall.PacketContext{}, true, f.h, f.cp, nil)
-		}
-	})
-}

firewall_test.go

@@ -213,44 +213,44 @@ func TestFirewall_Drop(t *testing.T) {
 	cp := cert.NewCAPool()
 
 	// Drop outbound
-	assert.Equal(t, ErrNoMatchingRule, fw.Drop(p, firewall.PacketContext{}, false, &h, cp, nil))
+	assert.Equal(t, ErrNoMatchingRule, fw.Drop(p, false, &h, cp, nil))
 	// Allow inbound
 	resetConntrack(fw)
-	require.NoError(t, fw.Drop(p, firewall.PacketContext{}, true, &h, cp, nil))
+	require.NoError(t, fw.Drop(p, true, &h, cp, nil))
 	// Allow outbound because conntrack
-	require.NoError(t, fw.Drop(p, firewall.PacketContext{}, false, &h, cp, nil))
+	require.NoError(t, fw.Drop(p, false, &h, cp, nil))
 
 	// test remote mismatch
 	oldRemote := p.RemoteAddr
 	p.RemoteAddr = netip.MustParseAddr("1.2.3.10")
-	assert.Equal(t, fw.Drop(p, firewall.PacketContext{}, false, &h, cp, nil), ErrInvalidRemoteIP)
+	assert.Equal(t, fw.Drop(p, false, &h, cp, nil), ErrInvalidRemoteIP)
 	p.RemoteAddr = oldRemote
 
 	// ensure signer doesn't get in the way of group checks
 	fw = NewFirewall(l, time.Second, time.Minute, time.Hour, &c)
 	require.NoError(t, fw.AddRule(true, firewall.ProtoAny, 0, 0, []string{"nope"}, "", "", "", "", "signer-shasum"))
 	require.NoError(t, fw.AddRule(true, firewall.ProtoAny, 0, 0, []string{"default-group"}, "", "", "", "", "signer-shasum-bad"))
-	assert.Equal(t, fw.Drop(p, firewall.PacketContext{}, true, &h, cp, nil), ErrNoMatchingRule)
+	assert.Equal(t, fw.Drop(p, true, &h, cp, nil), ErrNoMatchingRule)
 
 	// test caSha doesn't drop on match
 	fw = NewFirewall(l, time.Second, time.Minute, time.Hour, &c)
 	require.NoError(t, fw.AddRule(true, firewall.ProtoAny, 0, 0, []string{"nope"}, "", "", "", "", "signer-shasum-bad"))
 	require.NoError(t, fw.AddRule(true, firewall.ProtoAny, 0, 0, []string{"default-group"}, "", "", "", "", "signer-shasum"))
-	require.NoError(t, fw.Drop(p, firewall.PacketContext{}, true, &h, cp, nil))
+	require.NoError(t, fw.Drop(p, true, &h, cp, nil))
 
 	// ensure ca name doesn't get in the way of group checks
 	cp.CAs["signer-shasum"] = &cert.CachedCertificate{Certificate: &dummyCert{name: "ca-good"}}
 	fw = NewFirewall(l, time.Second, time.Minute, time.Hour, &c)
 	require.NoError(t, fw.AddRule(true, firewall.ProtoAny, 0, 0, []string{"nope"}, "", "", "", "ca-good", ""))
 	require.NoError(t, fw.AddRule(true, firewall.ProtoAny, 0, 0, []string{"default-group"}, "", "", "", "ca-good-bad", ""))
-	assert.Equal(t, fw.Drop(p, firewall.PacketContext{}, true, &h, cp, nil), ErrNoMatchingRule)
+	assert.Equal(t, fw.Drop(p, true, &h, cp, nil), ErrNoMatchingRule)
 
 	// test caName doesn't drop on match
 	cp.CAs["signer-shasum"] = &cert.CachedCertificate{Certificate: &dummyCert{name: "ca-good"}}
 	fw = NewFirewall(l, time.Second, time.Minute, time.Hour, &c)
 	require.NoError(t, fw.AddRule(true, firewall.ProtoAny, 0, 0, []string{"nope"}, "", "", "", "ca-good-bad", ""))
 	require.NoError(t, fw.AddRule(true, firewall.ProtoAny, 0, 0, []string{"default-group"}, "", "", "", "ca-good", ""))
-	require.NoError(t, fw.Drop(p, firewall.PacketContext{}, true, &h, cp, nil))
+	require.NoError(t, fw.Drop(p, true, &h, cp, nil))
 }
 
 func TestFirewall_DropV6(t *testing.T) {
@@ -292,44 +292,44 @@ func TestFirewall_DropV6(t *testing.T) {
 	cp := cert.NewCAPool()
 
 	// Drop outbound
-	assert.Equal(t, ErrNoMatchingRule, fw.Drop(p, firewall.PacketContext{}, false, &h, cp, nil))
+	assert.Equal(t, ErrNoMatchingRule, fw.Drop(p, false, &h, cp, nil))
 	// Allow inbound
 	resetConntrack(fw)
-	require.NoError(t, fw.Drop(p, firewall.PacketContext{}, true, &h, cp, nil))
+	require.NoError(t, fw.Drop(p, true, &h, cp, nil))
 	// Allow outbound because conntrack
-	require.NoError(t, fw.Drop(p, firewall.PacketContext{}, false, &h, cp, nil))
+	require.NoError(t, fw.Drop(p, false, &h, cp, nil))
 
 	// test remote mismatch
 	oldRemote := p.RemoteAddr
 	p.RemoteAddr = netip.MustParseAddr("fd12::56")
-	assert.Equal(t, fw.Drop(p, firewall.PacketContext{}, false, &h, cp, nil), ErrInvalidRemoteIP)
+	assert.Equal(t, fw.Drop(p, false, &h, cp, nil), ErrInvalidRemoteIP)
 	p.RemoteAddr = oldRemote
 
 	// ensure signer doesn't get in the way of group checks
 	fw = NewFirewall(l, time.Second, time.Minute, time.Hour, &c)
 	require.NoError(t, fw.AddRule(true, firewall.ProtoAny, 0, 0, []string{"nope"}, "", "", "", "", "signer-shasum"))
 	require.NoError(t, fw.AddRule(true, firewall.ProtoAny, 0, 0, []string{"default-group"}, "", "", "", "", "signer-shasum-bad"))
-	assert.Equal(t, fw.Drop(p, firewall.PacketContext{}, true, &h, cp, nil), ErrNoMatchingRule)
+	assert.Equal(t, fw.Drop(p, true, &h, cp, nil), ErrNoMatchingRule)
 
 	// test caSha doesn't drop on match
 	fw = NewFirewall(l, time.Second, time.Minute, time.Hour, &c)
 	require.NoError(t, fw.AddRule(true, firewall.ProtoAny, 0, 0, []string{"nope"}, "", "", "", "", "signer-shasum-bad"))
 	require.NoError(t, fw.AddRule(true, firewall.ProtoAny, 0, 0, []string{"default-group"}, "", "", "", "", "signer-shasum"))
-	require.NoError(t, fw.Drop(p, firewall.PacketContext{}, true, &h, cp, nil))
+	require.NoError(t, fw.Drop(p, true, &h, cp, nil))
 
 	// ensure ca name doesn't get in the way of group checks
 	cp.CAs["signer-shasum"] = &cert.CachedCertificate{Certificate: &dummyCert{name: "ca-good"}}
 	fw = NewFirewall(l, time.Second, time.Minute, time.Hour, &c)
 	require.NoError(t, fw.AddRule(true, firewall.ProtoAny, 0, 0, []string{"nope"}, "", "", "", "ca-good", ""))
 	require.NoError(t, fw.AddRule(true, firewall.ProtoAny, 0, 0, []string{"default-group"}, "", "", "", "ca-good-bad", ""))
-	assert.Equal(t, fw.Drop(p, firewall.PacketContext{}, true, &h, cp, nil), ErrNoMatchingRule)
+	assert.Equal(t, fw.Drop(p, true, &h, cp, nil), ErrNoMatchingRule)
 
 	// test caName doesn't drop on match
 	cp.CAs["signer-shasum"] = &cert.CachedCertificate{Certificate: &dummyCert{name: "ca-good"}}
 	fw = NewFirewall(l, time.Second, time.Minute, time.Hour, &c)
 	require.NoError(t, fw.AddRule(true, firewall.ProtoAny, 0, 0, []string{"nope"}, "", "", "", "ca-good-bad", ""))
 	require.NoError(t, fw.AddRule(true, firewall.ProtoAny, 0, 0, []string{"default-group"}, "", "", "", "ca-good", ""))
-	require.NoError(t, fw.Drop(p, firewall.PacketContext{}, true, &h, cp, nil))
+	require.NoError(t, fw.Drop(p, true, &h, cp, nil))
 }
 
 func BenchmarkFirewallTable_match(b *testing.B) {
@@ -537,10 +537,10 @@ func TestFirewall_Drop2(t *testing.T) {
 	cp := cert.NewCAPool()
 
 	// h1/c1 lacks the proper groups
-	require.ErrorIs(t, fw.Drop(p, firewall.PacketContext{}, true, &h1, cp, nil), ErrNoMatchingRule)
+	require.ErrorIs(t, fw.Drop(p, true, &h1, cp, nil), ErrNoMatchingRule)
 	// c has the proper groups
 	resetConntrack(fw)
-	require.NoError(t, fw.Drop(p, firewall.PacketContext{}, true, &h, cp, nil))
+	require.NoError(t, fw.Drop(p, true, &h, cp, nil))
 }
 
 func TestFirewall_Drop3(t *testing.T) {
@@ -618,18 +618,18 @@ func TestFirewall_Drop3(t *testing.T) {
 	cp := cert.NewCAPool()
 
 	// c1 should pass because host match
-	require.NoError(t, fw.Drop(p, firewall.PacketContext{}, true, &h1, cp, nil))
+	require.NoError(t, fw.Drop(p, true, &h1, cp, nil))
 	// c2 should pass because ca sha match
 	resetConntrack(fw)
-	require.NoError(t, fw.Drop(p, firewall.PacketContext{}, true, &h2, cp, nil))
+	require.NoError(t, fw.Drop(p, true, &h2, cp, nil))
 	// c3 should fail because no match
 	resetConntrack(fw)
-	assert.Equal(t, fw.Drop(p, firewall.PacketContext{}, true, &h3, cp, nil), ErrNoMatchingRule)
+	assert.Equal(t, fw.Drop(p, true, &h3, cp, nil), ErrNoMatchingRule)
 
 	// Test a remote address match
 	fw = NewFirewall(l, time.Second, time.Minute, time.Hour, c.Certificate)
 	require.NoError(t, fw.AddRule(true, firewall.ProtoAny, 1, 1, []string{}, "", "1.2.3.4/24", "", "", ""))
-	require.NoError(t, fw.Drop(p, firewall.PacketContext{}, true, &h1, cp, nil))
+	require.NoError(t, fw.Drop(p, true, &h1, cp, nil))
 }
 
 func TestFirewall_Drop3V6(t *testing.T) {
@@ -667,7 +667,7 @@ func TestFirewall_Drop3V6(t *testing.T) {
 	fw := NewFirewall(l, time.Second, time.Minute, time.Hour, c.Certificate)
 	cp := cert.NewCAPool()
 	require.NoError(t, fw.AddRule(true, firewall.ProtoAny, 1, 1, []string{}, "", "fd12::34/120", "", "", ""))
-	require.NoError(t, fw.Drop(p, firewall.PacketContext{}, true, &h, cp, nil))
+	require.NoError(t, fw.Drop(p, true, &h, cp, nil))
 }
 
 func TestFirewall_DropConntrackReload(t *testing.T) {
@@ -709,12 +709,12 @@ func TestFirewall_DropConntrackReload(t *testing.T) {
 	cp := cert.NewCAPool()
 
 	// Drop outbound
-	assert.Equal(t, fw.Drop(p, firewall.PacketContext{}, false, &h, cp, nil), ErrNoMatchingRule)
+	assert.Equal(t, fw.Drop(p, false, &h, cp, nil), ErrNoMatchingRule)
 	// Allow inbound
 	resetConntrack(fw)
-	require.NoError(t, fw.Drop(p, firewall.PacketContext{}, true, &h, cp, nil))
+	require.NoError(t, fw.Drop(p, true, &h, cp, nil))
 	// Allow outbound because conntrack
-	require.NoError(t, fw.Drop(p, firewall.PacketContext{}, false, &h, cp, nil))
+	require.NoError(t, fw.Drop(p, false, &h, cp, nil))
 
 	oldFw := fw
 	fw = NewFirewall(l, time.Second, time.Minute, time.Hour, c.Certificate)
@@ -723,7 +723,7 @@ func TestFirewall_DropConntrackReload(t *testing.T) {
 	fw.rulesVersion = oldFw.rulesVersion + 1
 
 	// Allow outbound because conntrack and new rules allow port 10
-	require.NoError(t, fw.Drop(p, firewall.PacketContext{}, false, &h, cp, nil))
+	require.NoError(t, fw.Drop(p, false, &h, cp, nil))
 
 	oldFw = fw
 	fw = NewFirewall(l, time.Second, time.Minute, time.Hour, c.Certificate)
@@ -732,7 +732,7 @@ func TestFirewall_DropConntrackReload(t *testing.T) {
 	fw.rulesVersion = oldFw.rulesVersion + 1
 
 	// Drop outbound because conntrack doesn't match new ruleset
-	assert.Equal(t, fw.Drop(p, firewall.PacketContext{}, false, &h, cp, nil), ErrNoMatchingRule)
+	assert.Equal(t, fw.Drop(p, false, &h, cp, nil), ErrNoMatchingRule)
 }
 
 func TestFirewall_ICMPPortBehavior(t *testing.T) {
@@ -778,12 +778,12 @@ func TestFirewall_ICMPPortBehavior(t *testing.T) {
 			p.LocalPort = 0
 			p.RemotePort = 0
 			// Drop outbound
-			assert.Equal(t, fw.Drop(*p, firewall.PacketContext{}, false, &h, cp, nil), ErrNoMatchingRule)
+			assert.Equal(t, fw.Drop(*p, false, &h, cp, nil), ErrNoMatchingRule)
 			// Allow inbound
 			resetConntrack(fw)
-			require.NoError(t, fw.Drop(*p, firewall.PacketContext{}, true, &h, cp, nil))
+			require.NoError(t, fw.Drop(*p, true, &h, cp, nil))
 			//now also allow outbound
-			require.NoError(t, fw.Drop(*p, firewall.PacketContext{}, false, &h, cp, nil))
+			require.NoError(t, fw.Drop(*p, false, &h, cp, nil))
 		})
 
 		t.Run("nonzero ports", func(t *testing.T) {
@@ -791,12 +791,12 @@ func TestFirewall_ICMPPortBehavior(t *testing.T) {
 			p.LocalPort = 0xabcd
 			p.RemotePort = 0x1234
 			// Drop outbound
-			assert.Equal(t, fw.Drop(*p, firewall.PacketContext{}, false, &h, cp, nil), ErrNoMatchingRule)
+			assert.Equal(t, fw.Drop(*p, false, &h, cp, nil), ErrNoMatchingRule)
 			// Allow inbound
 			resetConntrack(fw)
-			require.NoError(t, fw.Drop(*p, firewall.PacketContext{}, true, &h, cp, nil))
+			require.NoError(t, fw.Drop(*p, true, &h, cp, nil))
 			//now also allow outbound
-			require.NoError(t, fw.Drop(*p, firewall.PacketContext{}, false, &h, cp, nil))
+			require.NoError(t, fw.Drop(*p, false, &h, cp, nil))
 		})
 	})
 
@@ -808,12 +808,12 @@ func TestFirewall_ICMPPortBehavior(t *testing.T) {
 			p.LocalPort = 0
 			p.RemotePort = 0
 			// Drop outbound
-			assert.Equal(t, fw.Drop(*p, firewall.PacketContext{}, false, &h, cp, nil), ErrNoMatchingRule)
+			assert.Equal(t, fw.Drop(*p, false, &h, cp, nil), ErrNoMatchingRule)
 			// Allow inbound
 			resetConntrack(fw)
-			assert.Equal(t, fw.Drop(*p, firewall.PacketContext{}, true, &h, cp, nil), ErrNoMatchingRule)
+			assert.Equal(t, fw.Drop(*p, true, &h, cp, nil), ErrNoMatchingRule)
 			//now also allow outbound
-			assert.Equal(t, fw.Drop(*p, firewall.PacketContext{}, false, &h, cp, nil), ErrNoMatchingRule)
+			assert.Equal(t, fw.Drop(*p, false, &h, cp, nil), ErrNoMatchingRule)
 		})
 
 		t.Run("nonzero ports, still blocked", func(t *testing.T) {
@@ -821,12 +821,12 @@ func TestFirewall_ICMPPortBehavior(t *testing.T) {
 			p.LocalPort = 0xabcd
 			p.RemotePort = 0x1234
 			// Drop outbound
-			assert.Equal(t, fw.Drop(*p, firewall.PacketContext{}, false, &h, cp, nil), ErrNoMatchingRule)
+			assert.Equal(t, fw.Drop(*p, false, &h, cp, nil), ErrNoMatchingRule)
 			// Allow inbound
 			resetConntrack(fw)
-			assert.Equal(t, fw.Drop(*p, firewall.PacketContext{}, true, &h, cp, nil), ErrNoMatchingRule)
+			assert.Equal(t, fw.Drop(*p, true, &h, cp, nil), ErrNoMatchingRule)
 			//now also allow outbound
-			assert.Equal(t, fw.Drop(*p, firewall.PacketContext{}, false, &h, cp, nil), ErrNoMatchingRule)
+			assert.Equal(t, fw.Drop(*p, false, &h, cp, nil), ErrNoMatchingRule)
 		})
 
 		t.Run("nonzero, matching ports, still blocked", func(t *testing.T) {
@@ -834,12 +834,12 @@ func TestFirewall_ICMPPortBehavior(t *testing.T) {
 			p.LocalPort = 80
 			p.RemotePort = 80
 			// Drop outbound
-			assert.Equal(t, fw.Drop(*p, firewall.PacketContext{}, false, &h, cp, nil), ErrNoMatchingRule)
+			assert.Equal(t, fw.Drop(*p, false, &h, cp, nil), ErrNoMatchingRule)
 			// Allow inbound
 			resetConntrack(fw)
-			assert.Equal(t, fw.Drop(*p, firewall.PacketContext{}, true, &h, cp, nil), ErrNoMatchingRule)
+			assert.Equal(t, fw.Drop(*p, true, &h, cp, nil), ErrNoMatchingRule)
 			//now also allow outbound
-			assert.Equal(t, fw.Drop(*p, firewall.PacketContext{}, false, &h, cp, nil), ErrNoMatchingRule)
+			assert.Equal(t, fw.Drop(*p, false, &h, cp, nil), ErrNoMatchingRule)
 		})
 	})
 	t.Run("Any proto, any port", func(t *testing.T) {
@@ -851,12 +851,12 @@ func TestFirewall_ICMPPortBehavior(t *testing.T) {
 			p.LocalPort = 0
 			p.RemotePort = 0
 			// Drop outbound
-			assert.Equal(t, fw.Drop(*p, firewall.PacketContext{}, false, &h, cp, nil), ErrNoMatchingRule)
+			assert.Equal(t, fw.Drop(*p, false, &h, cp, nil), ErrNoMatchingRule)
 			// Allow inbound
 			resetConntrack(fw)
-			require.NoError(t, fw.Drop(*p, firewall.PacketContext{}, true, &h, cp, nil))
+			require.NoError(t, fw.Drop(*p, true, &h, cp, nil))
 			//now also allow outbound
-			require.NoError(t, fw.Drop(*p, firewall.PacketContext{}, false, &h, cp, nil))
+			require.NoError(t, fw.Drop(*p, false, &h, cp, nil))
 		})
 
 		t.Run("nonzero ports, allowed", func(t *testing.T) {
@@ -865,15 +865,15 @@ func TestFirewall_ICMPPortBehavior(t *testing.T) {
 			p.LocalPort = 0xabcd
 			p.RemotePort = 0x1234
 			// Drop outbound
-			assert.Equal(t, fw.Drop(*p, firewall.PacketContext{}, false, &h, cp, nil), ErrNoMatchingRule)
+			assert.Equal(t, fw.Drop(*p, false, &h, cp, nil), ErrNoMatchingRule)
 			// Allow inbound
 			resetConntrack(fw)
-			require.NoError(t, fw.Drop(*p, firewall.PacketContext{}, true, &h, cp, nil))
+			require.NoError(t, fw.Drop(*p, true, &h, cp, nil))
 			//now also allow outbound
-			require.NoError(t, fw.Drop(*p, firewall.PacketContext{}, false, &h, cp, nil))
+			require.NoError(t, fw.Drop(*p, false, &h, cp, nil))
 			//different ID is blocked
 			p.RemotePort++
-			require.Equal(t, fw.Drop(*p, firewall.PacketContext{}, false, &h, cp, nil), ErrNoMatchingRule)
+			require.Equal(t, fw.Drop(*p, false, &h, cp, nil), ErrNoMatchingRule)
 		})
 	})
 
@@ -922,7 +922,7 @@ func TestFirewall_DropIPSpoofing(t *testing.T) {
 		Protocol:   firewall.ProtoUDP,
 		Fragment:   false,
 	}
-	assert.Equal(t, fw.Drop(p, firewall.PacketContext{}, true, &h1, cp, nil), ErrInvalidRemoteIP)
+	assert.Equal(t, fw.Drop(p, true, &h1, cp, nil), ErrInvalidRemoteIP)
 }
 
 func BenchmarkLookup(b *testing.B) {
@@ -1336,7 +1336,7 @@ func (c *testcase) Test(t *testing.T, fw *Firewall) {
 	t.Helper()
 	cp := cert.NewCAPool()
 	resetConntrack(fw)
-	err := fw.Drop(c.p, firewall.PacketContext{}, true, c.h, cp, nil)
+	err := fw.Drop(c.p, true, c.h, cp, nil)
 	if c.err == nil {
 		require.NoError(t, err, "failed to not drop remote address %s", c.p.RemoteAddr)
 	} else {

hostmap.go

@@ -604,9 +604,9 @@ func (hm *HostMap) queryVpnAddr(vpnIp netip.Addr, promoteIfce *Interface) *HostI
 // unlockedAddHostInfo assumes you have a write-lock and will add a hostinfo object to the hostmap Indexes and RemoteIndexes maps.
 // If an entry exists for the Hosts table (vpnIp -> hostinfo) then the provided hostinfo will be made primary
 func (hm *HostMap) unlockedAddHostInfo(hostinfo *HostInfo, f *Interface) {
-	if f.dnsServer != nil {
+	if f.serveDns {
 		remoteCert := hostinfo.ConnectionState.peerCert
-		f.dnsServer.Add(remoteCert.Certificate.Name()+".", hostinfo.vpnAddrs)
+		dnsR.Add(remoteCert.Certificate.Name()+".", hostinfo.vpnAddrs)
 	}
 	for _, addr := range hostinfo.vpnAddrs {
 		hm.unlockedInnerAddHostInfo(addr, hostinfo, f)

inside.go

@@ -11,8 +11,8 @@ import (
 	"github.com/slackhq/nebula/routing"
 )
 
-func (f *Interface) consumeInsidePacket(packet []byte, fwPacket *firewall.Packet, fwCtx *firewall.PacketContext, nb, out []byte, q int, localCache firewall.ConntrackCache) {
+func (f *Interface) consumeInsidePacket(packet []byte, fwPacket *firewall.Packet, nb []byte, batch *sendBatch, rejectBuf []byte, q int, localCache firewall.ConntrackCache) {
-	err := newPacket(packet, false, fwPacket, fwCtx)
+	err := newPacket(packet, false, fwPacket)
 	if err != nil {
 		if f.l.Level >= logrus.DebugLevel {
 			f.l.WithField("packet", packet).Debugf("Error while validating outbound packet: %s", err)
@@ -33,7 +33,7 @@ func (f *Interface) consumeInsidePacket(packet []byte, fwPacket *firewall.Packet
 		// routes packets from the Nebula addr to the Nebula addr through the Nebula
 		// TUN device.
 		if immediatelyForwardToSelf {
-			_, err := f.readers[q].Write(packet)
+			_, err := f.readers[q].WriteReject(packet)
 			if err != nil {
 				f.l.WithError(err).Error("Failed to forward to tun")
 			}
@@ -53,7 +53,7 @@ func (f *Interface) consumeInsidePacket(packet []byte, fwPacket *firewall.Packet
 	})
 
 	if hostinfo == nil {
-		f.rejectInside(packet, out, q)
+		f.rejectInside(packet, rejectBuf, q)
 		if f.l.Level >= logrus.DebugLevel {
 			f.l.WithField("vpnAddr", fwPacket.RemoteAddr).
 				WithField("fwPacket", fwPacket).
@@ -66,12 +66,12 @@ func (f *Interface) consumeInsidePacket(packet []byte, fwPacket *firewall.Packet
 		return
 	}
 
-	dropReason := f.firewall.Drop(*fwPacket, *fwCtx, false, hostinfo, f.pki.GetCAPool(), localCache)
+	dropReason := f.firewall.Drop(*fwPacket, false, hostinfo, f.pki.GetCAPool(), localCache)
 	if dropReason == nil {
-		f.sendNoMetrics(header.Message, 0, hostinfo.ConnectionState, hostinfo, netip.AddrPort{}, packet, nb, out, q)
+		f.sendInsideMessage(hostinfo, packet, nb, batch, rejectBuf, q)
 
 	} else {
-		f.rejectInside(packet, out, q)
+		f.rejectInside(packet, rejectBuf, q)
 		if f.l.Level >= logrus.DebugLevel {
 			hostinfo.logger(f.l).
 				WithField("fwPacket", fwPacket).
@@ -81,6 +81,63 @@ func (f *Interface) consumeInsidePacket(packet []byte, fwPacket *firewall.Packet
 	}
 }
 
+// sendInsideMessage encrypts a firewall-approved inside packet into the
+// caller's batch slot for later sendmmsg flush. When hostinfo.remote is not
+// valid we fall through to the relay slow path via the unbatched sendNoMetrics
+// so relay behavior is unchanged.
+func (f *Interface) sendInsideMessage(hostinfo *HostInfo, p, nb []byte, batch *sendBatch, rejectBuf []byte, q int) {
+	ci := hostinfo.ConnectionState
+	if ci.eKey == nil {
+		return
+	}
+
+	if !hostinfo.remote.IsValid() {
+		// Slow path: relay fallback. Reuse rejectBuf as the ciphertext
+		// scratch; sendNoMetrics arranges header space for SendVia.
+		f.sendNoMetrics(header.Message, 0, ci, hostinfo, netip.AddrPort{}, p, nb, rejectBuf, q)
+		return
+	}
+
+	scratch := batch.Next()
+	if scratch == nil {
+		// Batch full: bypass batching and send this packet directly so we
+		// never drop traffic on over-subscribed iterations.
+		f.sendNoMetrics(header.Message, 0, ci, hostinfo, netip.AddrPort{}, p, nb, rejectBuf, q)
+		return
+	}
+
+	if noiseutil.EncryptLockNeeded {
+		ci.writeLock.Lock()
+	}
+	c := ci.messageCounter.Add(1)
+
+	out := header.Encode(scratch, header.Version, header.Message, 0, hostinfo.remoteIndexId, c)
+	f.connectionManager.Out(hostinfo)
+
+	if hostinfo.lastRebindCount != f.rebindCount {
+		//NOTE: there is an update hole if a tunnel isn't used and exactly 256 rebinds occur before the tunnel is
+		// finally used again. This tunnel would eventually be torn down and recreated if this action didn't help.
+		f.lightHouse.QueryServer(hostinfo.vpnAddrs[0])
+		hostinfo.lastRebindCount = f.rebindCount
+		if f.l.Level >= logrus.DebugLevel {
+			f.l.WithField("vpnAddrs", hostinfo.vpnAddrs).Debug("Lighthouse update triggered for punch due to rebind counter")
+		}
+	}
+
+	out, err := ci.eKey.EncryptDanger(out, out, p, c, nb)
+	if noiseutil.EncryptLockNeeded {
+		ci.writeLock.Unlock()
+	}
+	if err != nil {
+		hostinfo.logger(f.l).WithError(err).
+			WithField("udpAddr", hostinfo.remote).WithField("counter", c).
+			Error("Failed to encrypt outgoing packet")
+		return
+	}
+
+	batch.Commit(len(out), hostinfo.remote)
+}
+
 func (f *Interface) rejectInside(packet []byte, out []byte, q int) {
 	if !f.firewall.InSendReject {
 		return
@@ -91,7 +148,7 @@ func (f *Interface) rejectInside(packet []byte, out []byte, q int) {
 		return
 	}
 
-	_, err := f.readers[q].Write(out)
+	_, err := f.readers[q].WriteReject(out)
 	if err != nil {
 		f.l.WithError(err).Error("Failed to write to tun")
 	}
@@ -211,15 +268,14 @@ func (f *Interface) getOrHandshakeConsiderRouting(fwPacket *firewall.Packet, cac
 
 func (f *Interface) sendMessageNow(t header.MessageType, st header.MessageSubType, hostinfo *HostInfo, p, nb, out []byte) {
 	fp := &firewall.Packet{}
-	ctx := &firewall.PacketContext{}
+	err := newPacket(p, false, fp)
-	err := newPacket(p, false, fp, ctx)
 	if err != nil {
 		f.l.Warnf("error while parsing outgoing packet for firewall check; %v", err)
 		return
 	}
 
 	// check if packet is in outbound fw rules
-	dropReason := f.firewall.Drop(*fp, *ctx, false, hostinfo, f.pki.GetCAPool(), nil)
+	dropReason := f.firewall.Drop(*fp, false, hostinfo, f.pki.GetCAPool(), nil)
 	if dropReason != nil {
 		if f.l.Level >= logrus.DebugLevel {
 			f.l.WithField("fwPacket", fp).

interface.go

@@ -4,7 +4,6 @@ import (
 	"context"
 	"errors"
 	"fmt"
-	"io"
 	"net/netip"
 	"sync"
 	"sync/atomic"
@@ -17,6 +16,8 @@ import (
 	"github.com/slackhq/nebula/firewall"
 	"github.com/slackhq/nebula/header"
 	"github.com/slackhq/nebula/overlay"
+	"github.com/slackhq/nebula/overlay/coalesce"
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/udp"
 )
 
@@ -29,7 +30,7 @@ type InterfaceConfig struct {
 	pki                *PKI
 	Cipher             string
 	Firewall           *Firewall
-	DnsServer          *dnsServer
+	ServeDns           bool
 	HandshakeManager   *HandshakeManager
 	lightHouse         *LightHouse
 	connectionManager  *connectionManager
@@ -57,7 +58,7 @@ type Interface struct {
 	firewall              *Firewall
 	connectionManager     *connectionManager
 	handshakeManager      *HandshakeManager
-	dnsServer             *dnsServer
+	serveDns              bool
 	createTime            time.Time
 	lightHouse            *LightHouse
 	myBroadcastAddrsTable *bart.Lite
@@ -85,9 +86,12 @@ type Interface struct {
 
 	conntrackCacheTimeout time.Duration
 
-	ctx     context.Context
 	writers []udp.Conn
-	readers []io.ReadWriteCloser
+	readers []tio.Queue
+	// tunCoalescers is one tcpCoalescer per tun queue, wrapping readers[i].
+	// decryptToTun sends plaintext into the coalescer; listenOut calls its
+	// Flush at the end of each UDP recvmmsg batch.
+	tunCoalescers []*coalesce.TCPCoalescer
 	wg            sync.WaitGroup
 
 	// fatalErr holds the first unexpected reader error that caused shutdown.
@@ -171,13 +175,12 @@ func NewInterface(ctx context.Context, c *InterfaceConfig) (*Interface, error) {
 
 	cs := c.pki.getCertState()
 	ifce := &Interface{
-		ctx:                   ctx,
 		pki:                   c.pki,
 		hostMap:               c.HostMap,
 		outside:               c.Outside,
 		inside:                c.Inside,
 		firewall:              c.Firewall,
-		dnsServer:             c.DnsServer,
+		serveDns:              c.ServeDns,
 		handshakeManager:      c.HandshakeManager,
 		createTime:            time.Now(),
 		lightHouse:            c.lightHouse,
@@ -186,7 +189,8 @@ func NewInterface(ctx context.Context, c *InterfaceConfig) (*Interface, error) {
 		routines:              c.routines,
 		version:               c.version,
 		writers:               make([]udp.Conn, c.routines),
-		readers:               make([]io.ReadWriteCloser, c.routines),
+		readers:               make([]tio.Queue, c.routines),
+		tunCoalescers:         make([]*coalesce.TCPCoalescer, c.routines),
 		myVpnNetworks:         cs.myVpnNetworks,
 		myVpnNetworksTable:    cs.myVpnNetworksTable,
 		myVpnAddrs:            cs.myVpnAddrs,
@@ -241,15 +245,17 @@ func (f *Interface) activate() error {
 	metrics.GetOrRegisterGauge("routines", nil).Update(int64(f.routines))
 
 	// Prepare n tun queues
-	var reader io.ReadWriteCloser = f.inside
 	for i := 0; i < f.routines; i++ {
 		if i > 0 {
-			reader, err = f.inside.NewMultiQueueReader()
+			err = f.inside.NewMultiQueueReader()
 			if err != nil {
 				return err
 			}
 		}
-		f.readers[i] = reader
+	}
+	f.readers = f.inside.Readers()
+	for i := range f.readers {
+		f.tunCoalescers[i] = coalesce.NewTCPCoalescer(f.readers[i]) //todo don't always do this
 	}
 
 	f.wg.Add(1) // for us to wait on Close() to return
@@ -305,16 +311,30 @@ func (f *Interface) listenOut(i int) {
 		li = f.outside
 	}
 
-	ctCache := firewall.NewConntrackCacheTicker(f.ctx, f.conntrackCacheTimeout)
+	ctCache := firewall.NewConntrackCacheTicker(f.conntrackCacheTimeout)
 	lhh := f.lightHouse.NewRequestHandler()
-	plaintext := make([]byte, udp.MTU)
 	h := &header.H{}
 	fwPacket := &firewall.Packet{}
-	fwCtx := &firewall.PacketContext{}
 	nb := make([]byte, 12, 12)
 
+	// plaintexts is a ring of decrypt scratches, one per packet in a UDP
+	// recvmmsg batch. The coalescer borrows payload slices from here and
+	// requires they stay valid until Flush — so we rotate each packet and
+	// reset only in the batch-end flush callback.
+	var plaintexts [][]byte
+	idx := 0
+	coalescer := f.tunCoalescers[i]
 	err := li.ListenOut(func(fromUdpAddr netip.AddrPort, payload []byte) {
-		f.readOutsidePackets(ViaSender{UdpAddr: fromUdpAddr}, plaintext[:0], payload, h, fwPacket, fwCtx, lhh, nb, i, ctCache.Get(f.l))
+		if idx >= len(plaintexts) {
+			plaintexts = append(plaintexts, make([]byte, udp.MTU))
+		}
+		f.readOutsidePackets(ViaSender{UdpAddr: fromUdpAddr}, plaintexts[idx][:0], payload, h, fwPacket, lhh, nb, i, ctCache.Get(f.l))
+		idx++
+	}, func() {
+		if err := coalescer.Flush(); err != nil {
+			f.l.WithError(err).Error("Failed to flush tun coalescer")
+		}
+		idx = 0
 	})
 
 	if err != nil && !f.closed.Load() {
@@ -325,17 +345,16 @@ func (f *Interface) listenOut(i int) {
 	f.l.Debugf("underlay reader %v is done", i)
 }
 
-func (f *Interface) listenIn(reader io.ReadWriteCloser, i int) {
+func (f *Interface) listenIn(reader tio.Queue, i int) {
-	packet := make([]byte, mtu)
+	rejectBuf := make([]byte, mtu)
-	out := make([]byte, mtu)
+	batch := newSendBatch(sendBatchCap, udp.MTU+32)
 	fwPacket := &firewall.Packet{}
-	fwCtx := &firewall.PacketContext{}
 	nb := make([]byte, 12, 12)
 
-	conntrackCache := firewall.NewConntrackCacheTicker(f.ctx, f.conntrackCacheTimeout)
+	conntrackCache := firewall.NewConntrackCacheTicker(f.conntrackCacheTimeout)
 
 	for {
-		n, err := reader.Read(packet)
+		pkts, err := reader.Read()
 		if err != nil {
 			if !f.closed.Load() {
 				f.l.WithError(err).WithField("reader", i).Error("Error while reading outbound packet, closing")
@@ -344,12 +363,28 @@ func (f *Interface) listenIn(reader io.ReadWriteCloser, i int) {
 			break
 		}
 
-		f.consumeInsidePacket(packet[:n], fwPacket, fwCtx, nb, out, i, conntrackCache.Get(f.l))
+		batch.Reset()
+		for _, pkt := range pkts {
+			if batch.Len() >= batch.Cap() {
+				f.flushBatch(batch, i)
+				batch.Reset()
+			}
+			f.consumeInsidePacket(pkt, fwPacket, nb, batch, rejectBuf, i, conntrackCache.Get(f.l))
+		}
+		if batch.Len() > 0 {
+			f.flushBatch(batch, i)
+		}
 	}
 
 	f.l.Debugf("overlay reader %v is done", i)
 }
 
+func (f *Interface) flushBatch(batch *sendBatch, q int) {
+	if err := f.writers[q].WriteBatch(batch.bufs, batch.dsts); err != nil {
+		f.l.WithError(err).WithField("writer", q).Error("Failed to write outgoing batch")
+	}
+}
+
 func (f *Interface) RegisterConfigChangeCallbacks(c *config.C) {
 	c.RegisterReloadCallback(f.reloadFirewall)
 	c.RegisterReloadCallback(f.reloadSendRecvError)
@@ -402,15 +437,8 @@ func (f *Interface) reloadFirewall(c *config.C) {
 		fw.Conntrack = conntrack
 	}
 
-	fw.reporter = oldFw.reporter
-
 	f.firewall = fw
 
-	// Fire ReportRulesReload under the conntrack lock so the reporter cannot
-	// observe a FlowCreate/FlowEvict for the new rulesVersion before it
-	// observes the reload marker. Report* must be non-blocking.
-	fw.reportRulesReload(oldFw.rulesVersion, fw.rulesVersion)
-
 	oldFw.Destroy()
 	f.l.WithField("firewallHashes", fw.GetRuleHashes()).
 		WithField("oldFirewallHashes", oldFw.GetRuleHashes()).

main.go

@@ -3,7 +3,10 @@ package nebula
 import (
 	"context"
 	"fmt"
+	"log"
 	"net"
+	"net/http"
+	_ "net/http/pprof"
 	"net/netip"
 	"runtime/debug"
 	"strings"
@@ -49,6 +52,11 @@ func Main(c *config.C, configTest bool, buildVersion string, logger *logrus.Logg
 		l.Println(string(b))
 	}
 
+	//todo!!!
+	go func() {
+		log.Println(http.ListenAndServe("0.0.0.0:6060", nil))
+	}()
+
 	err := configLogger(l, c)
 	if err != nil {
 		return nil, util.ContextualizeIfNeeded("Failed to configure the logger", err)
@@ -215,9 +223,13 @@ func Main(c *config.C, configTest bool, buildVersion string, logger *logrus.Logg
 	handshakeManager := NewHandshakeManager(l, hostMap, lightHouse, udpConns[0], handshakeConfig)
 	lightHouse.handshakeTrigger = handshakeManager.trigger
 
-	ds, err := newDnsServerFromConfig(ctx, l, pki.getCertState(), hostMap, c)
+	serveDns := false
-	if err != nil {
+	if c.GetBool("lighthouse.serve_dns", false) {
-		l.WithError(err).Warn("Failed to start DNS responder")
+		if c.GetBool("lighthouse.am_lighthouse", false) {
+			serveDns = true
+		} else {
+			l.Warn("DNS server refusing to run because this host is not a lighthouse.")
+		}
 	}
 
 	ifConfig := &InterfaceConfig{
@@ -226,7 +238,7 @@ func Main(c *config.C, configTest bool, buildVersion string, logger *logrus.Logg
 		Outside:               udpConns[0],
 		pki:                   pki,
 		Firewall:              fw,
-		DnsServer:             ds,
+		ServeDns:              serveDns,
 		HandshakeManager:      handshakeManager,
 		connectionManager:     connManager,
 		lightHouse:            lightHouse,
@@ -276,6 +288,13 @@ func Main(c *config.C, configTest bool, buildVersion string, logger *logrus.Logg
 
 	attachCommands(l, c, ssh, ifce)
 
+	// Start DNS server last to allow using the nebula IP as lighthouse.dns.host
+	var dnsStart func()
+	if lightHouse.amLighthouse && serveDns {
+		l.Debugln("Starting dns server")
+		dnsStart = dnsMain(l, pki.getCertState(), hostMap, c)
+	}
+
 	return &Control{
 		state:                  StateReady,
 		f:                      ifce,
@@ -284,7 +303,7 @@ func Main(c *config.C, configTest bool, buildVersion string, logger *logrus.Logg
 		cancel:                 cancel,
 		sshStart:               sshStart,
 		statsStart:             statsStart,
-		dnsStart:               ds.Start,
+		dnsStart:               dnsStart,
 		lighthouseStart:        lightHouse.StartUpdateWorker,
 		connectionManagerStart: connManager.Start,
 	}, nil

noise.go

@@ -15,14 +15,14 @@ type endianness interface {
 var noiseEndianness endianness = binary.BigEndian
 
 type NebulaCipherState struct {
-	c noise.Cipher
+	c cipher.AEAD
 	//k [32]byte
 	//n uint64
 }
 
 func NewNebulaCipherState(s *noise.CipherState) *NebulaCipherState {
-	return &NebulaCipherState{c: s.Cipher()}
+	x := s.Cipher()
-
+	return &NebulaCipherState{c: x.(cipher.AEAD)}
 }
 
 // EncryptDanger encrypts and authenticates a given payload.
@@ -46,7 +46,7 @@ func (s *NebulaCipherState) EncryptDanger(out, ad, plaintext []byte, n uint64, n
 		nb[2] = 0
 		nb[3] = 0
 		noiseEndianness.PutUint64(nb[4:], n)
-		out = s.c.(cipher.AEAD).Seal(out, nb, plaintext, ad)
+		out = s.c.Seal(out, nb, plaintext, ad)
 		//l.Debugf("Encryption: outlen: %d, nonce: %d, ad: %s, plainlen %d", len(out), n, ad, len(plaintext))
 		return out, nil
 	} else {
@@ -61,7 +61,7 @@ func (s *NebulaCipherState) DecryptDanger(out, ad, ciphertext []byte, n uint64,
 		nb[2] = 0
 		nb[3] = 0
 		noiseEndianness.PutUint64(nb[4:], n)
-		return s.c.(cipher.AEAD).Open(out, nb, ciphertext, ad)
+		return s.c.Open(out, nb, ciphertext, ad)
 	} else {
 		return []byte{}, nil
 	}
@@ -69,7 +69,7 @@ func (s *NebulaCipherState) DecryptDanger(out, ad, ciphertext []byte, n uint64,
 
 func (s *NebulaCipherState) Overhead() int {
 	if s != nil {
-		return s.c.(cipher.AEAD).Overhead()
+		return s.c.Overhead()
 	}
 	return 0
 }

outside.go

@@ -19,7 +19,7 @@ const (
 	minFwPacketLen = 4
 )
 
-func (f *Interface) readOutsidePackets(via ViaSender, out []byte, packet []byte, h *header.H, fwPacket *firewall.Packet, fwCtx *firewall.PacketContext, lhf *LightHouseHandler, nb []byte, q int, localCache firewall.ConntrackCache) {
+func (f *Interface) readOutsidePackets(via ViaSender, out []byte, packet []byte, h *header.H, fwPacket *firewall.Packet, lhf *LightHouseHandler, nb []byte, q int, localCache firewall.ConntrackCache) {
 	err := h.Parse(packet)
 	if err != nil {
 		// Hole punch packets are 0 or 1 byte big, so lets ignore printing those errors
@@ -60,7 +60,7 @@ func (f *Interface) readOutsidePackets(via ViaSender, out []byte, packet []byte,
 
 		switch h.Subtype {
 		case header.MessageNone:
-			if !f.decryptToTun(hostinfo, h.MessageCounter, out, packet, fwPacket, fwCtx, nb, q, localCache) {
+			if !f.decryptToTun(hostinfo, h.MessageCounter, out, packet, fwPacket, nb, q, localCache) {
 				return
 			}
 		case header.MessageRelay:
@@ -102,7 +102,7 @@ func (f *Interface) readOutsidePackets(via ViaSender, out []byte, packet []byte,
 					relay:     relay,
 					IsRelayed: true,
 				}
-				f.readOutsidePackets(via, out[:0], signedPayload, h, fwPacket, fwCtx, lhf, nb, q, localCache)
+				f.readOutsidePackets(via, out[:0], signedPayload, h, fwPacket, lhf, nb, q, localCache)
 				return
 			case ForwardingType:
 				// Find the target HostInfo relay object
@@ -295,10 +295,7 @@ var (
 )
 
 // newPacket validates and parses the interesting bits for the firewall out of the ip and sub protocol headers
-func newPacket(data []byte, incoming bool, fp *firewall.Packet, ctx *firewall.PacketContext) error {
+func newPacket(data []byte, incoming bool, fp *firewall.Packet) error {
-	if ctx != nil {
-		*ctx = firewall.PacketContext{}
-	}
 	if len(data) < 1 {
 		return ErrPacketTooShort
 	}
@@ -306,14 +303,14 @@ func newPacket(data []byte, incoming bool, fp *firewall.Packet, ctx *firewall.Pa
 	version := int((data[0] >> 4) & 0x0f)
 	switch version {
 	case ipv4.Version:
-		return parseV4(data, incoming, fp, ctx)
+		return parseV4(data, incoming, fp)
 	case ipv6.Version:
-		return parseV6(data, incoming, fp, ctx)
+		return parseV6(data, incoming, fp)
 	}
 	return ErrUnknownIPVersion
 }
 
-func parseV6(data []byte, incoming bool, fp *firewall.Packet, ctx *firewall.PacketContext) error {
+func parseV6(data []byte, incoming bool, fp *firewall.Packet) error {
 	dataLen := len(data)
 	if dataLen < ipv6.HeaderLen {
 		return ErrIPv6PacketTooShort
@@ -358,11 +355,6 @@ func parseV6(data []byte, incoming bool, fp *firewall.Packet, ctx *firewall.Pack
 				fp.RemotePort = 0
 			}
 			fp.Fragment = false
-			if ctx != nil {
-				ctx.Length = binary.BigEndian.Uint16(data[4:6]) + uint16(ipv6.HeaderLen)
-				ctx.ICMPType = data[offset]
-				ctx.ICMPCode = data[offset+1]
-			}
 			return nil
 
 		case layers.IPProtocolTCP, layers.IPProtocolUDP:
@@ -380,12 +372,6 @@ func parseV6(data []byte, incoming bool, fp *firewall.Packet, ctx *firewall.Pack
 			}
 
 			fp.Fragment = false
-			if ctx != nil {
-				ctx.Length = binary.BigEndian.Uint16(data[4:6]) + uint16(ipv6.HeaderLen)
-				if proto == layers.IPProtocolTCP && dataLen >= offset+14 {
-					ctx.TCPFlags = data[offset+13]
-				}
-			}
 			return nil
 
 		case layers.IPProtocolIPv6Fragment:
@@ -437,7 +423,7 @@ func parseV6(data []byte, incoming bool, fp *firewall.Packet, ctx *firewall.Pack
 	return ErrIPv6CouldNotFindPayload
 }
 
-func parseV4(data []byte, incoming bool, fp *firewall.Packet, ctx *firewall.PacketContext) error {
+func parseV4(data []byte, incoming bool, fp *firewall.Packet) error {
 	// Do we at least have an ipv4 header worth of data?
 	if len(data) < ipv4.HeaderLen {
 		return ErrIPv4PacketTooShort
@@ -494,21 +480,6 @@ func parseV4(data []byte, incoming bool, fp *firewall.Packet, ctx *firewall.Pack
 		fp.RemotePort = binary.BigEndian.Uint16(data[ihl+2 : ihl+4]) //dst port
 	}
 
-	if ctx != nil {
-		ctx.Length = binary.BigEndian.Uint16(data[2:4])
-		if !fp.Fragment {
-			switch fp.Protocol {
-			case firewall.ProtoICMP:
-				ctx.ICMPType = data[ihl]
-				ctx.ICMPCode = data[ihl+1]
-			case firewall.ProtoTCP:
-				if len(data) >= ihl+14 {
-					ctx.TCPFlags = data[ihl+13]
-				}
-			}
-		}
-	}
-
 	return nil
 }
 
@@ -528,7 +499,7 @@ func (f *Interface) decrypt(hostinfo *HostInfo, mc uint64, out []byte, packet []
 	return out, nil
 }
 
-func (f *Interface) decryptToTun(hostinfo *HostInfo, messageCounter uint64, out []byte, packet []byte, fwPacket *firewall.Packet, fwCtx *firewall.PacketContext, nb []byte, q int, localCache firewall.ConntrackCache) bool {
+func (f *Interface) decryptToTun(hostinfo *HostInfo, messageCounter uint64, out []byte, packet []byte, fwPacket *firewall.Packet, nb []byte, q int, localCache firewall.ConntrackCache) bool {
 	var err error
 
 	out, err = hostinfo.ConnectionState.dKey.DecryptDanger(out, packet[:header.Len], packet[header.Len:], messageCounter, nb)
@@ -537,7 +508,7 @@ func (f *Interface) decryptToTun(hostinfo *HostInfo, messageCounter uint64, out
 		return false
 	}
 
-	err = newPacket(out, true, fwPacket, fwCtx)
+	err = newPacket(out, true, fwPacket)
 	if err != nil {
 		hostinfo.logger(f.l).WithError(err).WithField("packet", out).
 			Warnf("Error while validating inbound packet")
@@ -550,7 +521,7 @@ func (f *Interface) decryptToTun(hostinfo *HostInfo, messageCounter uint64, out
 		return false
 	}
 
-	dropReason := f.firewall.Drop(*fwPacket, *fwCtx, true, hostinfo, f.pki.GetCAPool(), localCache)
+	dropReason := f.firewall.Drop(*fwPacket, true, hostinfo, f.pki.GetCAPool(), localCache)
 	if dropReason != nil {
 		// NOTE: We give `packet` as the `out` here since we already decrypted from it and we don't need it anymore
 		// This gives us a buffer to build the reject packet in
@@ -564,7 +535,7 @@ func (f *Interface) decryptToTun(hostinfo *HostInfo, messageCounter uint64, out
 	}
 
 	f.connectionManager.In(hostinfo)
-	_, err = f.readers[q].Write(out)
+	err = f.tunCoalescers[q].Add(out)
 	if err != nil {
 		f.l.WithError(err).Error("Failed to write to tun")
 	}

outside_test.go

@@ -20,13 +20,13 @@ func Test_newPacket(t *testing.T) {
 	p := &firewall.Packet{}
 
 	// length fails
-	err := newPacket([]byte{}, true, p, nil)
+	err := newPacket([]byte{}, true, p)
 	require.ErrorIs(t, err, ErrPacketTooShort)
 
-	err = newPacket([]byte{0x40}, true, p, nil)
+	err = newPacket([]byte{0x40}, true, p)
 	require.ErrorIs(t, err, ErrIPv4PacketTooShort)
 
-	err = newPacket([]byte{0x60}, true, p, nil)
+	err = newPacket([]byte{0x60}, true, p)
 	require.ErrorIs(t, err, ErrIPv6PacketTooShort)
 
 	// length fail with ip options
@@ -39,15 +39,15 @@ func Test_newPacket(t *testing.T) {
 	}
 
 	b, _ := h.Marshal()
-	err = newPacket(b, true, p, nil)
+	err = newPacket(b, true, p)
 	require.ErrorIs(t, err, ErrIPv4InvalidHeaderLength)
 
 	// not an ipv4 packet
-	err = newPacket([]byte{0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, true, p, nil)
+	err = newPacket([]byte{0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, true, p)
 	require.ErrorIs(t, err, ErrUnknownIPVersion)
 
 	// invalid ihl
-	err = newPacket([]byte{4<<4 | (8 >> 2 & 0x0f), 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, true, p, nil)
+	err = newPacket([]byte{4<<4 | (8 >> 2 & 0x0f), 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, true, p)
 	require.ErrorIs(t, err, ErrIPv4InvalidHeaderLength)
 
 	// account for variable ip header length - incoming
@@ -62,7 +62,7 @@ func Test_newPacket(t *testing.T) {
 
 	b, _ = h.Marshal()
 	b = append(b, []byte{0, 3, 0, 4}...)
-	err = newPacket(b, true, p, nil)
+	err = newPacket(b, true, p)
 
 	require.NoError(t, err)
 	assert.Equal(t, uint8(firewall.ProtoTCP), p.Protocol)
@@ -84,7 +84,7 @@ func Test_newPacket(t *testing.T) {
 
 	b, _ = h.Marshal()
 	b = append(b, []byte{0, 5, 0, 6}...)
-	err = newPacket(b, false, p, nil)
+	err = newPacket(b, false, p)
 
 	require.NoError(t, err)
 	assert.Equal(t, uint8(2), p.Protocol)
@@ -114,7 +114,7 @@ func Test_newPacket_v6(t *testing.T) {
 	err := gopacket.SerializeLayers(buffer, opt, &ip)
 	require.NoError(t, err)
 
-	err = newPacket(buffer.Bytes(), true, p, nil)
+	err = newPacket(buffer.Bytes(), true, p)
 	require.ErrorIs(t, err, ErrIPv6CouldNotFindPayload)
 
 	// A v6 packet with a hop-by-hop extension
@@ -148,12 +148,12 @@ func Test_newPacket_v6(t *testing.T) {
 
 	// A full IPv6 header and 1 byte in the first extension, but missing
 	// the length byte.
-	err = newPacket(buffer.Bytes()[:41], true, p, nil)
+	err = newPacket(buffer.Bytes()[:41], true, p)
 	require.ErrorIs(t, err, ErrIPv6CouldNotFindPayload)
 
 	// A full IPv6 header plus 1 full extension, but only 1 byte of the
 	// next layer, missing length byte
-	err = newPacket(buffer.Bytes()[:49], true, p, nil)
+	err = newPacket(buffer.Bytes()[:49], true, p)
 	require.ErrorIs(t, err, ErrIPv6CouldNotFindPayload)
 	err = nil
 
@@ -173,7 +173,7 @@ func Test_newPacket_v6(t *testing.T) {
 
 	buffer.Clear()
 	require.NoError(t, gopacket.SerializeLayers(buffer, opt, &ip, &icmp))
-	require.Error(t, newPacket(buffer.Bytes(), true, p, nil))
+	require.Error(t, newPacket(buffer.Bytes(), true, p))
 
 	buffer.Clear()
 	echo := layers.ICMPv6Echo{
@@ -181,7 +181,7 @@ func Test_newPacket_v6(t *testing.T) {
 		SeqNumber:  1234,
 	}
 	require.NoError(t, gopacket.SerializeLayers(buffer, opt, &ip, &icmp, &echo))
-	require.NoError(t, newPacket(buffer.Bytes(), true, p, nil))
+	require.NoError(t, newPacket(buffer.Bytes(), true, p))
 	assert.Equal(t, uint8(layers.IPProtocolICMPv6), p.Protocol)
 	assert.Equal(t, netip.MustParseAddr("ff02::2"), p.RemoteAddr)
 	assert.Equal(t, netip.MustParseAddr("ff02::1"), p.LocalAddr)
@@ -192,7 +192,7 @@ func Test_newPacket_v6(t *testing.T) {
 	// A good ESP packet
 	b := buffer.Bytes()
 	b[6] = byte(layers.IPProtocolESP)
-	err = newPacket(b, true, p, nil)
+	err = newPacket(b, true, p)
 	require.NoError(t, err)
 	assert.Equal(t, uint8(layers.IPProtocolESP), p.Protocol)
 	assert.Equal(t, netip.MustParseAddr("ff02::2"), p.RemoteAddr)
@@ -204,7 +204,7 @@ func Test_newPacket_v6(t *testing.T) {
 	// A good None packet
 	b = buffer.Bytes()
 	b[6] = byte(layers.IPProtocolNoNextHeader)
-	err = newPacket(b, true, p, nil)
+	err = newPacket(b, true, p)
 	require.NoError(t, err)
 	assert.Equal(t, uint8(layers.IPProtocolNoNextHeader), p.Protocol)
 	assert.Equal(t, netip.MustParseAddr("ff02::2"), p.RemoteAddr)
@@ -216,7 +216,7 @@ func Test_newPacket_v6(t *testing.T) {
 	// An unknown protocol packet
 	b = buffer.Bytes()
 	b[6] = 255 // 255 is a reserved protocol number
-	err = newPacket(b, true, p, nil)
+	err = newPacket(b, true, p)
 	require.ErrorIs(t, err, ErrIPv6CouldNotFindPayload)
 
 	// A good UDP packet
@@ -243,7 +243,7 @@ func Test_newPacket_v6(t *testing.T) {
 	b = buffer.Bytes()
 
 	// incoming
-	err = newPacket(b, true, p, nil)
+	err = newPacket(b, true, p)
 	require.NoError(t, err)
 	assert.Equal(t, uint8(firewall.ProtoUDP), p.Protocol)
 	assert.Equal(t, netip.MustParseAddr("ff02::2"), p.RemoteAddr)
@@ -253,7 +253,7 @@ func Test_newPacket_v6(t *testing.T) {
 	assert.False(t, p.Fragment)
 
 	// outgoing
-	err = newPacket(b, false, p, nil)
+	err = newPacket(b, false, p)
 	require.NoError(t, err)
 	assert.Equal(t, uint8(firewall.ProtoUDP), p.Protocol)
 	assert.Equal(t, netip.MustParseAddr("ff02::2"), p.LocalAddr)
@@ -263,14 +263,14 @@ func Test_newPacket_v6(t *testing.T) {
 	assert.False(t, p.Fragment)
 
 	// Too short UDP packet
-	err = newPacket(b[:len(b)-10], false, p, nil) // pull off the last 10 bytes
+	err = newPacket(b[:len(b)-10], false, p) // pull off the last 10 bytes
 	require.ErrorIs(t, err, ErrIPv6PacketTooShort)
 
 	// A good TCP packet
 	b[6] = byte(layers.IPProtocolTCP)
 
 	// incoming
-	err = newPacket(b, true, p, nil)
+	err = newPacket(b, true, p)
 	require.NoError(t, err)
 	assert.Equal(t, uint8(firewall.ProtoTCP), p.Protocol)
 	assert.Equal(t, netip.MustParseAddr("ff02::2"), p.RemoteAddr)
@@ -280,7 +280,7 @@ func Test_newPacket_v6(t *testing.T) {
 	assert.False(t, p.Fragment)
 
 	// outgoing
-	err = newPacket(b, false, p, nil)
+	err = newPacket(b, false, p)
 	require.NoError(t, err)
 	assert.Equal(t, uint8(firewall.ProtoTCP), p.Protocol)
 	assert.Equal(t, netip.MustParseAddr("ff02::2"), p.LocalAddr)
@@ -290,7 +290,7 @@ func Test_newPacket_v6(t *testing.T) {
 	assert.False(t, p.Fragment)
 
 	// Too short TCP packet
-	err = newPacket(b[:len(b)-10], false, p, nil) // pull off the last 10 bytes
+	err = newPacket(b[:len(b)-10], false, p) // pull off the last 10 bytes
 	require.ErrorIs(t, err, ErrIPv6PacketTooShort)
 
 	// A good UDP packet with an AH header
@@ -325,7 +325,7 @@ func Test_newPacket_v6(t *testing.T) {
 	b = append(b, ahb...)
 	b = append(b, udpHeader...)
 
-	err = newPacket(b, true, p, nil)
+	err = newPacket(b, true, p)
 	require.NoError(t, err)
 	assert.Equal(t, uint8(firewall.ProtoUDP), p.Protocol)
 	assert.Equal(t, netip.MustParseAddr("ff02::2"), p.RemoteAddr)
@@ -335,12 +335,12 @@ func Test_newPacket_v6(t *testing.T) {
 	assert.False(t, p.Fragment)
 
 	// Ensure buffer bounds checking during processing
-	err = newPacket(b[:41], true, p, nil)
+	err = newPacket(b[:41], true, p)
 	require.ErrorIs(t, err, ErrIPv6PacketTooShort)
 
 	// Invalid AH header
 	b = buffer.Bytes()
-	err = newPacket(b, true, p, nil)
+	err = newPacket(b, true, p)
 	require.ErrorIs(t, err, ErrIPv6CouldNotFindPayload)
 }
 
@@ -388,7 +388,7 @@ func Test_newPacket_ipv6Fragment(t *testing.T) {
 	firstFrag = append(firstFrag, []byte{0xde, 0xad, 0xbe, 0xef}...)
 
 	// Test first fragment incoming
-	err = newPacket(firstFrag, true, p, nil)
+	err = newPacket(firstFrag, true, p)
 	require.NoError(t, err)
 	assert.Equal(t, netip.MustParseAddr("ff02::2"), p.RemoteAddr)
 	assert.Equal(t, netip.MustParseAddr("ff02::1"), p.LocalAddr)
@@ -398,7 +398,7 @@ func Test_newPacket_ipv6Fragment(t *testing.T) {
 	assert.False(t, p.Fragment)
 
 	// Test first fragment outgoing
-	err = newPacket(firstFrag, false, p, nil)
+	err = newPacket(firstFrag, false, p)
 	require.NoError(t, err)
 	assert.Equal(t, netip.MustParseAddr("ff02::2"), p.LocalAddr)
 	assert.Equal(t, netip.MustParseAddr("ff02::1"), p.RemoteAddr)
@@ -427,7 +427,7 @@ func Test_newPacket_ipv6Fragment(t *testing.T) {
 	secondFrag = append(secondFrag, []byte{0xde, 0xad, 0xbe, 0xef}...)
 
 	// Test second fragment incoming
-	err = newPacket(secondFrag, true, p, nil)
+	err = newPacket(secondFrag, true, p)
 	require.NoError(t, err)
 	assert.Equal(t, netip.MustParseAddr("ff02::2"), p.RemoteAddr)
 	assert.Equal(t, netip.MustParseAddr("ff02::1"), p.LocalAddr)
@@ -437,7 +437,7 @@ func Test_newPacket_ipv6Fragment(t *testing.T) {
 	assert.True(t, p.Fragment)
 
 	// Test second fragment outgoing
-	err = newPacket(secondFrag, false, p, nil)
+	err = newPacket(secondFrag, false, p)
 	require.NoError(t, err)
 	assert.Equal(t, netip.MustParseAddr("ff02::2"), p.LocalAddr)
 	assert.Equal(t, netip.MustParseAddr("ff02::1"), p.RemoteAddr)
@@ -447,7 +447,7 @@ func Test_newPacket_ipv6Fragment(t *testing.T) {
 	assert.True(t, p.Fragment)
 
 	// Too short of a fragment packet
-	err = newPacket(secondFrag[:len(secondFrag)-10], false, p, nil)
+	err = newPacket(secondFrag[:len(secondFrag)-10], false, p)
 	require.ErrorIs(t, err, ErrIPv6PacketTooShort)
 }
 
@@ -529,7 +529,7 @@ func BenchmarkParseV6(b *testing.B) {
 
 	b.Run("Normal", func(b *testing.B) {
 		for i := 0; i < b.N; i++ {
-			if err = parseV6(normalPacket, true, fp, nil); err != nil {
+			if err = parseV6(normalPacket, true, fp); err != nil {
 				b.Fatal(err)
 			}
 		}
@@ -537,7 +537,7 @@ func BenchmarkParseV6(b *testing.B) {
 
 	b.Run("FirstFragment", func(b *testing.B) {
 		for i := 0; i < b.N; i++ {
-			if err = parseV6(firstFrag, true, fp, nil); err != nil {
+			if err = parseV6(firstFrag, true, fp); err != nil {
 				b.Fatal(err)
 			}
 		}
@@ -545,7 +545,7 @@ func BenchmarkParseV6(b *testing.B) {
 
 	b.Run("SecondFragment", func(b *testing.B) {
 		for i := 0; i < b.N; i++ {
-			if err = parseV6(secondFrag, true, fp, nil); err != nil {
+			if err = parseV6(secondFrag, true, fp); err != nil {
 				b.Fatal(err)
 			}
 		}
@@ -590,7 +590,7 @@ func BenchmarkParseV6(b *testing.B) {
 
 	b.Run("200 HopByHop headers", func(b *testing.B) {
 		for i := 0; i < b.N; i++ {
-			if err = parseV6(evilBytes, false, fp, nil); err != nil {
+			if err = parseV6(evilBytes, false, fp); err != nil {
 				b.Fatal(err)
 			}
 		}

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/device.go

@@ -4,15 +4,21 @@ import (
 	"io"
 	"net/netip"
 
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 )
 
+// defaultBatchBufSize is the per-Queue scratch size for Read on backends
+// that don't do TSO segmentation. 65535 covers any single IP packet.
+const defaultBatchBufSize = 65535
+
 type Device interface {
-	io.ReadWriteCloser
+	io.Closer
 	Activate() error
 	Networks() []netip.Prefix
 	Name() string
 	RoutesFor(netip.Addr) routing.Gateways
-	SupportsMultiqueue() bool
+	SupportsMultiqueue() bool //todo remove?
-	NewMultiQueueReader() (io.ReadWriteCloser, error)
+	NewMultiQueueReader() error
+	Readers() []tio.Queue
 }

overlay/noop.go

@@ -1,10 +1,10 @@
-package test
+package overlay
 
 import (
 	"errors"
-	"io"
 	"net/netip"
 
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 )
 
@@ -26,20 +26,28 @@ func (NoopTun) Name() string {
 	return "noop"
 }
 
-func (NoopTun) Read([]byte) (int, error) {
+func (NoopTun) Read() ([][]byte, error) {
-	return 0, nil
+	return nil, nil
 }
 
 func (NoopTun) Write([]byte) (int, error) {
 	return 0, nil
 }
 
+func (NoopTun) WriteReject(p []byte) (int, error) {
+	return 0, nil
+}
+
 func (NoopTun) SupportsMultiqueue() bool {
 	return false
 }
 
-func (NoopTun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+func (NoopTun) NewMultiQueueReader() error {
-	return nil, errors.New("unsupported")
+	return errors.New("unsupported")
+}
+
+func (NoopTun) Readers() []tio.Queue {
+	return []tio.Queue{NoopTun{}}
 }
 
 func (NoopTun) Close() error {

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/container_poll_linux.go

@@ -0,0 +1,69 @@
+package tio
+
+import (
+	"encoding/binary"
+	"errors"
+	"fmt"
+
+	"golang.org/x/sys/unix"
+)
+
+type pollContainer struct {
+	pq []*Poll
+	// pqi is exactly the same as pq, but stored as the interface type
+	pqi        []Queue
+	shutdownFd int
+}
+
+func NewPollContainer() (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 := &pollContainer{
+		pq:         []*Poll{},
+		pqi:        []Queue{},
+		shutdownFd: shutdownFd,
+	}
+
+	return out, nil
+}
+
+func (c *pollContainer) Queues() []Queue {
+	return c.pqi
+}
+
+func (c *pollContainer) Add(fd int) error {
+	x, err := newPoll(fd, c.shutdownFd)
+	if err != nil {
+		return err
+	}
+	c.pq = append(c.pq, x)
+	c.pqi = append(c.pqi, x)
+
+	return nil
+}
+
+func (c *pollContainer) wakeForShutdown() error {
+	var buf [8]byte
+	binary.NativeEndian.PutUint64(buf[:], 1)
+	_, err := unix.Write(int(c.shutdownFd), buf[:])
+	return err
+}
+
+func (c *pollContainer) Close() error {
+	errs := []error{}
+
+	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

@@ -0,0 +1,63 @@
+package tio
+
+import "io"
+
+// defaultBatchBufSize is the per-Queue scratch size for Read on backends
+// that don't do TSO segmentation. 65535 covers any single IP packet.
+const defaultBatchBufSize = 65535
+
+type Container interface {
+	Queues() []Queue
+	Add(fd int) error
+
+	io.Closer
+}
+
+// Queue is a readable/writable Poll queue. One Queue is driven by a single
+// read goroutine plus concurrent writers (see Write / WriteReject below).
+type Queue interface {
+	io.Closer
+
+	// Read returns one or more packets. The returned slices are borrowed
+	// from the Queue's internal buffer and are only valid until the next
+	// Read or Close on this Queue — callers must encrypt or copy each
+	// slice before the next call. Not safe for concurrent Reads; exactly
+	// one goroutine per Queue reads.
+	Read() ([][]byte, error)
+
+	// Write emits a single packet on the plaintext (outside→inside)
+	// delivery path. May run concurrently with WriteReject on the same
+	// Queue, but not with itself.
+	Write(p []byte) (int, error)
+
+	// WriteReject writes a single packet that originated from the inside
+	// path (reject replies or self-forward) using scratch state distinct
+	// from Write, so it can run concurrently with Write on the same Queue
+	// without a data race. On backends without a shared-scratch Write, a
+	// trivial delegation to Write is acceptable.
+	WriteReject(p []byte) (int, error)
+}
+
+// GSOWriter is implemented by Queues that can emit a TCP TSO superpacket
+// assembled from a header prefix plus one or more borrowed payload
+// fragments, in a single vectored write (writev with a leading
+// virtio_net_hdr). This lets the coalescer avoid copying payload bytes
+// between the caller's decrypt buffer and the TUN. Backends without GSO
+// support return false from GSOSupported and coalescing is skipped.
+//
+// hdr contains the IPv4/IPv6 + TCP header prefix (mutable — callers will
+// have filled in total length and pseudo-header partial). pays are
+// non-overlapping payload fragments whose concatenation is the full
+// superpacket payload; they are read-only from the writer's perspective
+// and must remain valid until the call returns. gsoSize is the MSS:
+// every segment except possibly the last is exactly that many bytes.
+// csumStart is the byte offset where the TCP header begins within hdr.
+//
+// # TODO fold into Queue
+//
+// hdr's TCP checksum field must already hold the pseudo-header partial
+// sum (single-fold, not inverted), per virtio NEEDS_CSUM semantics.
+type GSOWriter interface {
+	WriteGSO(hdr []byte, pays [][]byte, gsoSize uint16, isV6 bool, csumStart uint16) error
+	GSOSupported() bool
+}

overlay/tio/tio_gso_linux.go

@@ -0,0 +1,371 @@
+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))
+	return r.rawWrite(unsafe.Slice(&iovs[0], len(iovs)))
+}
+
+func (r *Offload) rawWrite(iovs []unix.Iovec) (int, error) {
+	for {
+		n, _, errno := syscall.Syscall(unix.SYS_WRITEV, uintptr(r.fd), uintptr(unsafe.Pointer(&iovs[0])), uintptr(len(iovs)))
+		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))
+	}
+
+	_, err := r.rawWrite(r.gsoIovs)
+	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/tio_poll_linux.go

@@ -0,0 +1,205 @@
+package tio
+
+import (
+	"fmt"
+	"os"
+	"sync/atomic"
+	"syscall"
+	"unsafe"
+
+	"golang.org/x/sys/unix"
+)
+
+// Maximum size we accept for a single read from a TUN with IFF_VNET_HDR. A
+// TSO superpacket can be up to 64KiB of payload plus a single L2/L3/L4 header
+// prefix plus the virtio header.
+const tunReadBufSize = 65535
+
+type Poll struct {
+	fd int
+
+	readPoll  [2]unix.PollFd
+	writePoll [2]unix.PollFd
+	closed    atomic.Bool
+
+	readBuf  []byte
+	batchRet [1][]byte
+}
+
+func newPoll(fd int, shutdownFd int) (*Poll, error) {
+	if err := unix.SetNonblock(fd, true); err != nil {
+		_ = unix.Close(fd)
+		return nil, fmt.Errorf("failed to set Poll device as nonblocking: %w", err)
+	}
+
+	out := &Poll{
+		fd:      fd,
+		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},
+		},
+	}
+	return out, nil
+}
+
+// blockOnRead waits until the Poll fd is readable or shutdown has been signaled.
+// Returns os.ErrClosed if Close was called.
+func (t *Poll) blockOnRead() error {
+	const problemFlags = unix.POLLHUP | unix.POLLNVAL | unix.POLLERR
+	var err error
+	for {
+		_, err = unix.Poll(t.readPoll[:], -1)
+		if err != unix.EINTR {
+			break
+		}
+	}
+	tunEvents := t.readPoll[0].Revents
+	shutdownEvents := t.readPoll[1].Revents
+	t.readPoll[0].Revents = 0
+	t.readPoll[1].Revents = 0
+	if err != nil {
+		return err
+	}
+	if shutdownEvents&(unix.POLLIN|problemFlags) != 0 {
+		return os.ErrClosed
+	}
+	if tunEvents&problemFlags != 0 {
+		return os.ErrClosed
+	}
+	return nil
+}
+
+func (t *Poll) blockOnWrite() error {
+	const problemFlags = unix.POLLHUP | unix.POLLNVAL | unix.POLLERR
+	var err error
+	for {
+		_, err = unix.Poll(t.writePoll[:], -1)
+		if err != unix.EINTR {
+			break
+		}
+	}
+	tunEvents := t.writePoll[0].Revents
+	shutdownEvents := t.writePoll[1].Revents
+	t.writePoll[0].Revents = 0
+	t.writePoll[1].Revents = 0
+	if err != nil {
+		return err
+	}
+	if shutdownEvents&(unix.POLLIN|problemFlags) != 0 {
+		return os.ErrClosed
+	}
+	if tunEvents&problemFlags != 0 {
+		return os.ErrClosed
+	}
+	return nil
+}
+
+func (t *Poll) Read() ([][]byte, error) {
+	if t.readBuf == nil {
+		t.readBuf = make([]byte, defaultBatchBufSize)
+	}
+	n, err := t.readOne(t.readBuf)
+	if err != nil {
+		return nil, err
+	}
+	t.batchRet[0] = t.readBuf[:n]
+	return t.batchRet[:], nil
+}
+
+func (t *Poll) readOne(to []byte) (int, error) {
+	// first 4 bytes is protocol family, in network byte order
+	var head [4]byte
+	iovecs := [2]syscall.Iovec{ //todo plat-specific
+		{&head[0], 4},
+		{&to[0], uint64(len(to))},
+	}
+	for {
+		n, _, errno := syscall.Syscall(syscall.SYS_READV, uintptr(t.fd), uintptr(unsafe.Pointer(&iovecs[0])), 2)
+		if errno == 0 {
+			bytesRead := int(n)
+			if bytesRead < 4 {
+				return 0, nil
+			}
+			return bytesRead - 4, nil
+		}
+		switch errno {
+		case unix.EAGAIN:
+			if err := t.blockOnRead(); err != nil {
+				return 0, err
+			}
+		case unix.EINTR:
+			// retry
+		case unix.EBADF:
+			return 0, os.ErrClosed
+		default:
+			return 0, errno
+		}
+	}
+}
+
+// Write is only valid for single threaded use
+func (t *Poll) Write(from []byte) (int, error) {
+	if len(from) <= 1 {
+		return 0, syscall.EIO
+	}
+
+	ipVer := from[0] >> 4
+	var head [4]byte
+	// first 4 bytes is protocol family, in network byte order
+	switch ipVer {
+	case 4:
+		head[3] = syscall.AF_INET
+	case 6:
+		head[3] = syscall.AF_INET6
+	default:
+		return 0, fmt.Errorf("unable to determine IP version from packet")
+	}
+
+	iovecs := [2]syscall.Iovec{ //todo plat specific
+		{&head[0], 4},
+		{&from[0], uint64(len(from))},
+	}
+	for {
+		n, _, errno := syscall.Syscall(syscall.SYS_WRITEV, uintptr(t.fd), uintptr(unsafe.Pointer(&iovecs[0])), 2)
+		if errno == 0 {
+			return int(n) - 4, nil
+		}
+		switch errno {
+		case unix.EAGAIN:
+			if err := t.blockOnWrite(); err != nil {
+				return 0, err
+			}
+		case unix.EINTR:
+			// retry
+		case unix.EBADF:
+			return 0, os.ErrClosed
+		default:
+			return 0, errno
+		}
+	}
+}
+
+func (t *Poll) Close() error {
+	if t.closed.Swap(true) {
+		return nil
+	}
+
+	//shutdownFd is owned by the container, so we should not close it
+
+	var err error
+	if t.fd >= 0 {
+		err = unix.Close(t.fd)
+		t.fd = -1
+	}
+
+	return err
+}
+
+func (t *Poll) WriteReject(p []byte) (int, error) {
+	return t.Write(p)
+}

overlay/tio/tun_file_linux_test.go

@@ -0,0 +1,86 @@
+//go:build linux && !android && !e2e_testing
+// +build linux,!android,!e2e_testing
+
+package tio
+
+import (
+	"errors"
+	"os"
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/require"
+	"golang.org/x/sys/unix"
+)
+
+// newReadPipe returns a read fd. The matching write fd is registered for cleanup.
+// The caller takes ownership of the read fd (pass it to newOffload / newFriend).
+func newReadPipe(t *testing.T) int {
+	t.Helper()
+	var fds [2]int
+	if err := unix.Pipe2(fds[:], unix.O_CLOEXEC); err != nil {
+		t.Fatalf("pipe2: %v", err)
+	}
+	t.Cleanup(func() { _ = unix.Close(fds[1]) })
+	return fds[0]
+}
+
+func TestOffload_WakeForShutdown_WakesFriends(t *testing.T) {
+	pipe1 := newReadPipe(t)
+	pipe2 := newReadPipe(t)
+	parent, err := NewOffloadContainer()
+	if err != nil {
+		t.Fatalf("newOffload: %v", err)
+	}
+	require.NoError(t, parent.Add(pipe1))
+	require.NoError(t, parent.Add(pipe2))
+	t.Cleanup(func() {
+		_ = unix.Close(pipe1)
+		_ = unix.Close(pipe2)
+	})
+
+	readers := parent.Queues()
+	errs := make([]error, len(readers))
+	var wg sync.WaitGroup
+	for i, r := range readers {
+		wg.Add(1)
+		go func(i int, r Queue) {
+			defer wg.Done()
+			_, errs[i] = r.Read()
+		}(i, r)
+	}
+
+	time.Sleep(50 * time.Millisecond)
+
+	if err := parent.Close(); err != nil {
+		t.Fatalf("Close: %v", err)
+	}
+
+	done := make(chan struct{})
+	go func() { wg.Wait(); close(done) }()
+	select {
+	case <-done:
+	case <-time.After(2 * time.Second):
+		t.Fatal("readers did not wake")
+	}
+
+	for i, err := range errs {
+		if !errors.Is(err, os.ErrClosed) {
+			t.Errorf("reader %d: expected os.ErrClosed, got %v", i, err)
+		}
+	}
+}
+
+func TestTunFile_Close_Idempotent(t *testing.T) {
+	tf, err := newOffload(newReadPipe(t), 1)
+	if err != nil {
+		t.Fatalf("newOffload: %v", err)
+	}
+	if err := tf.Close(); err != nil {
+		t.Fatalf("first Close: %v", err)
+	}
+	if err := tf.Close(); err != nil {
+		t.Fatalf("second Close should be a no-op, got %v", err)
+	}
+}

overlay/tio/tun_linux_offload.go

@@ -0,0 +1,281 @@
+//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"
+)
+
+// Protocol header size bounds used to validate / cap kernel-supplied offsets.
+const (
+	ipv4HeaderMinLen = 20 // IHL=5, no options
+	ipv4HeaderMaxLen = 60 // IHL=15, max options
+	ipv6FixedLen     = 40 // IPv6 base header; extensions would extend this
+	tcpHeaderMinLen  = 20 // data-offset=5, no options
+	tcpHeaderMaxLen  = 60 // data-offset=15, max options
+)
+
+// Byte offsets inside an IPv4 header.
+const (
+	ipv4TotalLenOff = 2
+	ipv4IDOff       = 4
+	ipv4ChecksumOff = 10
+	ipv4SrcOff      = 12
+	ipv4AddrsEnd    = 20 // end of dst address (ipv4SrcOff + 2*4)
+)
+
+// Byte offsets inside an IPv6 header.
+const (
+	ipv6PayloadLenOff = 4
+	ipv6SrcOff        = 8
+	ipv6AddrsEnd      = 40 // end of dst address (ipv6SrcOff + 2*16)
+)
+
+// Byte offsets inside a TCP header (relative to its start, i.e. csumStart).
+const (
+	tcpSeqOff      = 4
+	tcpDataOffOff  = 12 // upper nibble is header len in 32-bit words
+	tcpFlagsOff    = 13
+	tcpChecksumOff = 16
+)
+
+// tcpFinPshMask is cleared on every segment except the last of a TSO burst.
+const tcpFinPshMask = 0x09 // FIN(0x01) | PSH(0x08)
+
+// 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 hdr.CsumStart == 0 {
+		return fmt.Errorf("csum_start is zero")
+	}
+
+	isV4 := hdr.GSOType == unix.VIRTIO_NET_HDR_GSO_TCPV4
+	csumStart := int(hdr.CsumStart)
+
+	if isV4 && csumStart < ipv4HeaderMinLen {
+		return fmt.Errorf("csum_start %d too small for IPv4", csumStart)
+	}
+	if !isV4 && csumStart < ipv6FixedLen {
+		return fmt.Errorf("csum_start %d too small for IPv6", csumStart)
+	}
+
+	// Don't trust hdr.HdrLen from the kernel: on some paths it can be set
+	// to the full length of the first packet rather than the true L3+L4 header length.
+	// Instead, read the TCP data-offset field from the packet itself and derive
+	// headerLen = csum_start + tcpHdrLen. Matches wireguard-go's approach.
+	if csumStart+tcpFlagsOff+1 > len(pkt) {
+		return fmt.Errorf("packet too short for tcp header at csum_start=%d (pkt %d)", csumStart, len(pkt))
+	}
+	tcpHdrLen := int(pkt[csumStart+tcpDataOffOff]>>4) * 4
+	if tcpHdrLen < tcpHeaderMinLen || tcpHdrLen > tcpHeaderMaxLen {
+		return fmt.Errorf("tcp data-offset out of range: %d", tcpHdrLen)
+	}
+	headerLen := csumStart + tcpHdrLen
+	if headerLen > len(pkt) {
+		return fmt.Errorf("derived hdr_len %d > pkt %d", headerLen, len(pkt))
+	}
+
+	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+tcpSeqOff : csumStart+tcpSeqOff+4])
+	origFlags := pkt[csumStart+tcpFlagsOff]
+
+	// Precompute the TCP header sum with seq/flags/csum zeroed. Copy onto
+	// the stack, zero the per-segment-varying fields, sum once.
+	var tmp [tcpHeaderMaxLen]byte
+	copy(tmp[:tcpHdrLen], pkt[csumStart:headerLen])
+	tmp[tcpSeqOff], tmp[tcpSeqOff+1], tmp[tcpSeqOff+2], tmp[tcpSeqOff+3] = 0, 0, 0, 0
+	tmp[tcpFlagsOff] = 0
+	tmp[tcpChecksumOff], tmp[tcpChecksumOff+1] = 0, 0
+	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[ipv4SrcOff:ipv4AddrsEnd], 0))
+	} else {
+		baseProtoSum = uint32(checksum.Checksum(pkt[ipv6SrcOff:ipv6AddrsEnd], 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[ipv4IDOff : ipv4IDOff+2])
+		ihl = int(pkt[0]&0x0f) * 4
+		if ihl < ipv4HeaderMinLen || ihl > csumStart {
+			return fmt.Errorf("bad IPv4 IHL: %d", ihl)
+		}
+		var ipTmp [ipv4HeaderMaxLen]byte
+		copy(ipTmp[:ihl], pkt[:ihl])
+		ipTmp[ipv4TotalLenOff], ipTmp[ipv4TotalLenOff+1] = 0, 0
+		ipTmp[ipv4IDOff], ipTmp[ipv4IDOff+1] = 0, 0
+		ipTmp[ipv4ChecksumOff], ipTmp[ipv4ChecksumOff+1] = 0, 0
+		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 &^ tcpFinPshMask
+		}
+		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[ipv4TotalLenOff:ipv4TotalLenOff+2], uint16(totalLen))
+			binary.BigEndian.PutUint16(seg[ipv4IDOff:ipv4IDOff+2], segID)
+			ipSum := baseIPHdrSum + uint32(totalLen) + uint32(segID)
+			binary.BigEndian.PutUint16(seg[ipv4ChecksumOff:ipv4ChecksumOff+2], foldComplement(ipSum))
+		} else {
+			// IPv6 payload length excludes the fixed header but includes any
+			// extension headers between [ipv6FixedLen:csumStart].
+			binary.BigEndian.PutUint16(seg[ipv6PayloadLenOff:ipv6PayloadLenOff+2], uint16(headerLen-ipv6FixedLen+segPayLen))
+		}
+
+		// Patch TCP header.
+		binary.BigEndian.PutUint32(seg[csumStart+tcpSeqOff:csumStart+tcpSeqOff+4], segSeq)
+		seg[csumStart+tcpFlagsOff] = 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+tcpChecksumOff:csumStart+tcpChecksumOff+2], 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_android.go

@@ -13,17 +13,45 @@ import (
 	"github.com/gaissmai/bart"
 	"github.com/sirupsen/logrus"
 	"github.com/slackhq/nebula/config"
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 	"github.com/slackhq/nebula/util"
 )
 
 type tun struct {
-	io.ReadWriteCloser
+	rwc         io.ReadWriteCloser
 	fd          int
 	vpnNetworks []netip.Prefix
 	Routes      atomic.Pointer[[]Route]
 	routeTree   atomic.Pointer[bart.Table[routing.Gateways]]
 	l           *logrus.Logger
+
+	readBuf  []byte
+	batchRet [1][]byte
+}
+
+func (t *tun) Read() ([][]byte, error) {
+	if t.readBuf == nil {
+		t.readBuf = make([]byte, defaultBatchBufSize)
+	}
+	n, err := t.rwc.Read(t.readBuf)
+	if err != nil {
+		return nil, err
+	}
+	t.batchRet[0] = t.readBuf[:n]
+	return t.batchRet[:], nil
+}
+
+func (t *tun) Write(p []byte) (int, error) {
+	return t.rwc.Write(p)
+}
+
+func (t *tun) WriteReject(p []byte) (int, error) {
+	return t.rwc.Write(p)
+}
+
+func (t *tun) Close() error {
+	return t.rwc.Close()
 }
 
 func newTunFromFd(c *config.C, l *logrus.Logger, deviceFd int, vpnNetworks []netip.Prefix) (*tun, error) {
@@ -32,7 +60,7 @@ func newTunFromFd(c *config.C, l *logrus.Logger, deviceFd int, vpnNetworks []net
 	file := os.NewFile(uintptr(deviceFd), "/dev/net/tun")
 
 	t := &tun{
-		ReadWriteCloser: file,
+		rwc:         file,
 		fd:          deviceFd,
 		vpnNetworks: vpnNetworks,
 		l:           l,
@@ -99,6 +127,6 @@ func (t *tun) SupportsMultiqueue() bool {
 	return false
 }
 
-func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+func (t *tun) NewMultiQueueReader() (tio.Queue, error) {
 	return nil, fmt.Errorf("TODO: multiqueue not implemented for android")
 }

overlay/tun_darwin.go

@@ -16,6 +16,7 @@ import (
 	"github.com/gaissmai/bart"
 	"github.com/sirupsen/logrus"
 	"github.com/slackhq/nebula/config"
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 	"github.com/slackhq/nebula/util"
 	netroute "golang.org/x/net/route"
@@ -23,7 +24,7 @@ import (
 )
 
 type tun struct {
-	io.ReadWriteCloser
+	rwc         io.ReadWriteCloser
 	Device      string
 	vpnNetworks []netip.Prefix
 	DefaultMTU  int
@@ -34,6 +35,9 @@ type tun struct {
 
 	// cache out buffer since we need to prepend 4 bytes for tun metadata
 	out []byte
+
+	readBuf  []byte
+	batchRet [1][]byte
 }
 
 type ifReq struct {
@@ -124,7 +128,7 @@ func newTun(c *config.C, l *logrus.Logger, vpnNetworks []netip.Prefix, _ bool) (
 	}
 
 	t := &tun{
-		ReadWriteCloser: os.NewFile(uintptr(fd), ""),
+		rwc:         os.NewFile(uintptr(fd), ""),
 		Device:      name,
 		vpnNetworks: vpnNetworks,
 		DefaultMTU:  c.GetInt("tun.mtu", DefaultMTU),
@@ -158,8 +162,8 @@ func newTunFromFd(_ *config.C, _ *logrus.Logger, _ int, _ []netip.Prefix) (*tun,
 }
 
 func (t *tun) Close() error {
-	if t.ReadWriteCloser != nil {
+	if t.rwc != nil {
-		return t.ReadWriteCloser.Close()
+		return t.rwc.Close()
 	}
 	return nil
 }
@@ -503,15 +507,31 @@ func delRoute(prefix netip.Prefix, gateway netroute.Addr) error {
 	return nil
 }
 
-func (t *tun) Read(to []byte) (int, error) {
+func (t *tun) readOne(to []byte) (int, error) {
 	buf := make([]byte, len(to)+4)
 
-	n, err := t.ReadWriteCloser.Read(buf)
+	n, err := t.rwc.Read(buf)
 
 	copy(to, buf[4:])
 	return n - 4, err
 }
 
+func (t *tun) Read() ([][]byte, error) {
+	if t.readBuf == nil {
+		t.readBuf = make([]byte, defaultBatchBufSize)
+	}
+	n, err := t.readOne(t.readBuf)
+	if err != nil {
+		return nil, err
+	}
+	t.batchRet[0] = t.readBuf[:n]
+	return t.batchRet[:], nil
+}
+
+func (t *tun) WriteReject(p []byte) (int, error) {
+	return t.Write(p)
+}
+
 // Write is only valid for single threaded use
 func (t *tun) Write(from []byte) (int, error) {
 	buf := t.out
@@ -537,7 +557,7 @@ func (t *tun) Write(from []byte) (int, error) {
 
 	copy(buf[4:], from)
 
-	n, err := t.ReadWriteCloser.Write(buf)
+	n, err := t.rwc.Write(buf)
 	return n - 4, err
 }
 
@@ -553,6 +573,6 @@ func (t *tun) SupportsMultiqueue() bool {
 	return false
 }
 
-func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+func (t *tun) NewMultiQueueReader() (tio.Queue, error) {
 	return nil, fmt.Errorf("TODO: multiqueue not implemented for darwin")
 }

overlay/tun_disabled.go

@@ -9,6 +9,7 @@ import (
 	"github.com/rcrowley/go-metrics"
 	"github.com/sirupsen/logrus"
 	"github.com/slackhq/nebula/iputil"
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 )
 
@@ -20,6 +21,24 @@ type disabledTun struct {
 	tx         metrics.Counter
 	rx         metrics.Counter
 	l          *logrus.Logger
+	numReaders int
+
+	batchRet [1][]byte
+}
+
+func (t *disabledTun) Read() ([][]byte, error) {
+	r, ok := <-t.read
+	if !ok {
+		return nil, io.EOF
+	}
+
+	t.tx.Inc(1)
+	if t.l.Level >= logrus.DebugLevel {
+		t.l.WithField("raw", prettyPacket(r)).Debugf("Write payload")
+	}
+
+	t.batchRet[0] = r
+	return t.batchRet[:], nil
 }
 
 func newDisabledTun(vpnNetworks []netip.Prefix, queueLen int, metricsEnabled bool, l *logrus.Logger) *disabledTun {
@@ -27,6 +46,7 @@ func newDisabledTun(vpnNetworks []netip.Prefix, queueLen int, metricsEnabled boo
 		vpnNetworks: vpnNetworks,
 		read:        make(chan []byte, queueLen),
 		l:           l,
+		numReaders:  1,
 	}
 
 	if metricsEnabled {
@@ -56,24 +76,6 @@ func (*disabledTun) Name() string {
 	return "disabled"
 }
 
-func (t *disabledTun) Read(b []byte) (int, error) {
-	r, ok := <-t.read
-	if !ok {
-		return 0, io.EOF
-	}
-
-	if len(r) > len(b) {
-		return 0, fmt.Errorf("packet larger than mtu: %d > %d bytes", len(r), len(b))
-	}
-
-	t.tx.Inc(1)
-	if t.l.Level >= logrus.DebugLevel {
-		t.l.WithField("raw", prettyPacket(r)).Debugf("Write payload")
-	}
-
-	return copy(b, r), nil
-}
-
 func (t *disabledTun) handleICMPEchoRequest(b []byte) bool {
 	out := make([]byte, len(b))
 	out = iputil.CreateICMPEchoResponse(b, out)
@@ -105,12 +107,25 @@ func (t *disabledTun) Write(b []byte) (int, error) {
 	return len(b), nil
 }
 
+func (t *disabledTun) WriteReject(b []byte) (int, error) {
+	return t.Write(b)
+}
+
 func (t *disabledTun) SupportsMultiqueue() bool {
 	return true
 }
 
-func (t *disabledTun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+func (t *disabledTun) NewMultiQueueReader() error {
-	return t, nil
+	t.numReaders++
+	return nil
+}
+
+func (t *disabledTun) Readers() []tio.Queue {
+	out := make([]tio.Queue, t.numReaders)
+	for i := range t.numReaders {
+		out[i] = t
+	}
+	return out
 }
 
 func (t *disabledTun) Close() error {

overlay/tun_file_linux_test.go

@@ -1,120 +0,0 @@
-//go:build linux && !android && !e2e_testing
-// +build linux,!android,!e2e_testing
-
-package overlay
-
-import (
-	"errors"
-	"os"
-	"sync"
-	"testing"
-	"time"
-
-	"golang.org/x/sys/unix"
-)
-
-// newReadPipe returns a read fd. The matching write fd is registered for cleanup.
-// The caller takes ownership of the read fd (pass it to newTunFd / newFriend).
-func newReadPipe(t *testing.T) int {
-	t.Helper()
-	var fds [2]int
-	if err := unix.Pipe2(fds[:], unix.O_CLOEXEC); err != nil {
-		t.Fatalf("pipe2: %v", err)
-	}
-	t.Cleanup(func() { _ = unix.Close(fds[1]) })
-	return fds[0]
-}
-
-func TestTunFile_WakeForShutdown_UnblocksRead(t *testing.T) {
-	tf, err := newTunFd(newReadPipe(t))
-	if err != nil {
-		t.Fatalf("newTunFd: %v", err)
-	}
-	t.Cleanup(func() { _ = tf.Close() })
-
-	done := make(chan error, 1)
-	go func() {
-		_, err := tf.Read(make([]byte, 64))
-		done <- err
-	}()
-
-	// Verify Read is actually blocked in poll.
-	select {
-	case err := <-done:
-		t.Fatalf("Read returned before shutdown signal: %v", err)
-	case <-time.After(50 * time.Millisecond):
-	}
-
-	if err := tf.wakeForShutdown(); err != nil {
-		t.Fatalf("wakeForShutdown: %v", err)
-	}
-
-	select {
-	case err := <-done:
-		if !errors.Is(err, os.ErrClosed) {
-			t.Fatalf("expected os.ErrClosed, got %v", err)
-		}
-	case <-time.After(2 * time.Second):
-		t.Fatal("Read did not wake on shutdown")
-	}
-}
-
-func TestTunFile_WakeForShutdown_WakesFriends(t *testing.T) {
-	parent, err := newTunFd(newReadPipe(t))
-	if err != nil {
-		t.Fatalf("newTunFd: %v", err)
-	}
-	friend, err := parent.newFriend(newReadPipe(t))
-	if err != nil {
-		_ = parent.Close()
-		t.Fatalf("newFriend: %v", err)
-	}
-	t.Cleanup(func() {
-		_ = friend.Close()
-		_ = parent.Close()
-	})
-
-	readers := []*tunFile{parent, friend}
-	errs := make([]error, len(readers))
-	var wg sync.WaitGroup
-	for i, r := range readers {
-		wg.Add(1)
-		go func(i int, r *tunFile) {
-			defer wg.Done()
-			_, errs[i] = r.Read(make([]byte, 64))
-		}(i, r)
-	}
-
-	time.Sleep(50 * time.Millisecond)
-
-	if err := parent.wakeForShutdown(); err != nil {
-		t.Fatalf("wakeForShutdown: %v", err)
-	}
-
-	done := make(chan struct{})
-	go func() { wg.Wait(); close(done) }()
-	select {
-	case <-done:
-	case <-time.After(2 * time.Second):
-		t.Fatal("readers did not wake")
-	}
-
-	for i, err := range errs {
-		if !errors.Is(err, os.ErrClosed) {
-			t.Errorf("reader %d: expected os.ErrClosed, got %v", i, err)
-		}
-	}
-}
-
-func TestTunFile_Close_Idempotent(t *testing.T) {
-	tf, err := newTunFd(newReadPipe(t))
-	if err != nil {
-		t.Fatalf("newTunFd: %v", err)
-	}
-	if err := tf.Close(); err != nil {
-		t.Fatalf("first Close: %v", err)
-	}
-	if err := tf.Close(); err != nil {
-		t.Fatalf("second Close should be a no-op, got %v", err)
-	}
-}

overlay/tun_freebsd.go

@@ -7,7 +7,6 @@ import (
 	"bytes"
 	"errors"
 	"fmt"
-	"io"
 	"io/fs"
 	"net/netip"
 	"os"
@@ -19,6 +18,7 @@ import (
 	"github.com/gaissmai/bart"
 	"github.com/sirupsen/logrus"
 	"github.com/slackhq/nebula/config"
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 	"github.com/slackhq/nebula/util"
 	netroute "golang.org/x/net/route"
@@ -101,6 +101,9 @@ type tun struct {
 	readPoll  [2]unix.PollFd
 	writePoll [2]unix.PollFd
 	closed    atomic.Bool
+
+	readBuf  []byte
+	batchRet [1][]byte
 }
 
 // blockOnRead waits until the tun fd is readable or shutdown has been signaled.
@@ -155,7 +158,23 @@ func (t *tun) blockOnWrite() error {
 	return nil
 }
 
-func (t *tun) Read(to []byte) (int, error) {
+func (t *tun) Read() ([][]byte, error) {
+	if t.readBuf == nil {
+		t.readBuf = make([]byte, defaultBatchBufSize)
+	}
+	n, err := t.readOne(t.readBuf)
+	if err != nil {
+		return nil, err
+	}
+	t.batchRet[0] = t.readBuf[:n]
+	return t.batchRet[:], nil
+}
+
+func (t *tun) WriteReject(p []byte) (int, error) {
+	return t.Write(p)
+}
+
+func (t *tun) readOne(to []byte) (int, error) {
 	// first 4 bytes is protocol family, in network byte order
 	var head [4]byte
 	iovecs := [2]syscall.Iovec{
@@ -563,7 +582,7 @@ func (t *tun) SupportsMultiqueue() bool {
 	return false
 }
 
-func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+func (t *tun) NewMultiQueueReader() (tio.Queue, error) {
 	return nil, fmt.Errorf("TODO: multiqueue not implemented for freebsd")
 }
 

overlay/tun_ios.go

@@ -16,16 +16,44 @@ import (
 	"github.com/gaissmai/bart"
 	"github.com/sirupsen/logrus"
 	"github.com/slackhq/nebula/config"
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 	"github.com/slackhq/nebula/util"
 )
 
 type tun struct {
-	io.ReadWriteCloser
+	rwc         io.ReadWriteCloser
 	vpnNetworks []netip.Prefix
 	Routes      atomic.Pointer[[]Route]
 	routeTree   atomic.Pointer[bart.Table[routing.Gateways]]
 	l           *logrus.Logger
+
+	readBuf  []byte
+	batchRet [1][]byte
+}
+
+func (t *tun) Read() ([][]byte, error) {
+	if t.readBuf == nil {
+		t.readBuf = make([]byte, defaultBatchBufSize)
+	}
+	n, err := t.rwc.Read(t.readBuf)
+	if err != nil {
+		return nil, err
+	}
+	t.batchRet[0] = t.readBuf[:n]
+	return t.batchRet[:], nil
+}
+
+func (t *tun) Write(p []byte) (int, error) {
+	return t.rwc.Write(p)
+}
+
+func (t *tun) WriteReject(p []byte) (int, error) {
+	return t.rwc.Write(p)
+}
+
+func (t *tun) Close() error {
+	return t.rwc.Close()
 }
 
 func newTun(_ *config.C, _ *logrus.Logger, _ []netip.Prefix, _ bool) (*tun, error) {
@@ -36,7 +64,7 @@ func newTunFromFd(c *config.C, l *logrus.Logger, deviceFd int, vpnNetworks []net
 	file := os.NewFile(uintptr(deviceFd), "/dev/tun")
 	t := &tun{
 		vpnNetworks: vpnNetworks,
-		ReadWriteCloser: &tunReadCloser{f: file},
+		rwc:         &tunReadCloser{f: file},
 		l:           l,
 	}
 
@@ -155,6 +183,6 @@ func (t *tun) SupportsMultiqueue() bool {
 	return false
 }
 
-func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+func (t *tun) NewMultiQueueReader() (tio.Queue, error) {
 	return nil, fmt.Errorf("TODO: multiqueue not implemented for ios")
 }

overlay/tun_linux.go

@@ -4,9 +4,7 @@
 package overlay
 
 import (
-	"encoding/binary"
 	"fmt"
-	"io"
 	"net"
 	"net/netip"
 	"os"
@@ -19,180 +17,15 @@ import (
 	"github.com/gaissmai/bart"
 	"github.com/sirupsen/logrus"
 	"github.com/slackhq/nebula/config"
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 	"github.com/slackhq/nebula/util"
 	"github.com/vishvananda/netlink"
 	"golang.org/x/sys/unix"
 )
 
-// tunFile 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 tunFile struct {
-	fd         int
-	shutdownFd int
-	lastOne    bool
-	readPoll   [2]unix.PollFd
-	writePoll  [2]unix.PollFd
-	closed     bool
-}
-
-// newFriend makes a tunFile for a MultiQueueReader that copies the shutdown eventfd from the parent tun
-func (r *tunFile) newFriend(fd int) (*tunFile, error) {
-	if err := unix.SetNonblock(fd, true); err != nil {
-		return nil, fmt.Errorf("failed to set tun fd non-blocking: %w", err)
-	}
-	return &tunFile{
-		fd:         fd,
-		shutdownFd: r.shutdownFd,
-		readPoll: [2]unix.PollFd{
-			{Fd: int32(fd), Events: unix.POLLIN},
-			{Fd: int32(r.shutdownFd), Events: unix.POLLIN},
-		},
-		writePoll: [2]unix.PollFd{
-			{Fd: int32(fd), Events: unix.POLLOUT},
-			{Fd: int32(r.shutdownFd), Events: unix.POLLIN},
-		},
-	}, nil
-}
-
-func newTunFd(fd int) (*tunFile, error) {
-	if err := unix.SetNonblock(fd, true); err != nil {
-		return nil, fmt.Errorf("failed to set tun fd non-blocking: %w", err)
-	}
-
-	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 := &tunFile{
-		fd:         fd,
-		shutdownFd: shutdownFd,
-		lastOne:    true,
-		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},
-		},
-	}
-
-	return out, nil
-}
-
-func (r *tunFile) 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 *tunFile) 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 *tunFile) Read(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
-		}
-	}
-}
-
-func (r *tunFile) Write(buf []byte) (int, error) {
-	for {
-		if n, err := unix.Write(r.fd, buf); err == nil {
-			return n, nil
-		} else if err == unix.EAGAIN {
-			if err = r.blockOnWrite(); 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
-		}
-	}
-}
-
-func (r *tunFile) wakeForShutdown() error {
-	var buf [8]byte
-	binary.NativeEndian.PutUint64(buf[:], 1)
-	_, err := unix.Write(int(r.readPoll[1].Fd), buf[:])
-	return err
-}
-
-func (r *tunFile) Close() error {
-	if r.closed { // avoid closing more than once. Technically a fd could get re-used, which would be a problem
-		return nil
-	}
-	r.closed = true
-	if r.lastOne {
-		_ = unix.Close(r.shutdownFd)
-	}
-	return unix.Close(r.fd)
-}
-
 type tun struct {
-	*tunFile
+	readers     tio.Container
-	readers     []*tunFile
 	closeLock   sync.Mutex
 	Device      string
 	vpnNetworks []netip.Prefix
@@ -201,6 +34,7 @@ type tun struct {
 	TXQueueLen  int
 	deviceIndex int
 	ioctlFd     uintptr
+	vnetHdr     bool
 
 	Routes                    atomic.Pointer[[]Route]
 	routeTree                 atomic.Pointer[bart.Table[routing.Gateways]]
@@ -239,7 +73,9 @@ type ifreqQLEN struct {
 }
 
 func newTunFromFd(c *config.C, l *logrus.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
 	}
@@ -249,46 +85,83 @@ func newTunFromFd(c *config.C, l *logrus.Logger, deviceFd int, vpnNetworks []net
 	return t, nil
 }
 
-func newTun(c *config.C, l *logrus.Logger, vpnNetworks []netip.Prefix, multiqueue bool) (*tun, error) {
+// openTunDev opens /dev/net/tun, creating the device node first if it's
+// missing (docker containers occasionally omit it).
+func openTunDev() (int, error) {
 	fd, err := unix.Open("/dev/net/tun", os.O_RDWR, 0)
-	if err != nil {
+	if err == nil {
-		// If /dev/net/tun doesn't exist, try to create it (will happen in docker)
+		return fd, nil
-		if os.IsNotExist(err) {
-			err = os.MkdirAll("/dev/net", 0755)
-			if err != nil {
-				return nil, fmt.Errorf("/dev/net/tun doesn't exist, failed to mkdir -p /dev/net: %w", err)
 	}
-			err = unix.Mknod("/dev/net/tun", unix.S_IFCHR|0600, int(unix.Mkdev(10, 200)))
+	if !os.IsNotExist(err) {
-			if err != nil {
+		return -1, err
-				return nil, fmt.Errorf("failed to create /dev/net/tun: %w", err)
+	}
+	if err = os.MkdirAll("/dev/net", 0755); err != nil {
+		return -1, fmt.Errorf("/dev/net/tun doesn't exist, failed to mkdir -p /dev/net: %w", err)
+	}
+	if err = unix.Mknod("/dev/net/tun", unix.S_IFCHR|0600, int(unix.Mkdev(10, 200))); err != nil {
+		return -1, fmt.Errorf("failed to create /dev/net/tun: %w", err)
 	}
-
 	fd, err = unix.Open("/dev/net/tun", os.O_RDWR, 0)
 	if err != nil {
-				return nil, fmt.Errorf("created /dev/net/tun, but still failed: %w", err)
+		return -1, fmt.Errorf("created /dev/net/tun, but still failed: %w", err)
-			}
-		} else {
-			return nil, err
 	}
+	return fd, nil
 }
 
+// tunSetIff runs TUNSETIFF with the given flags and returns the kernel-chosen
+// device name on success.
+func tunSetIff(fd int, name string, flags uint16) (string, error) {
 	var req ifReq
-	req.Flags = uint16(unix.IFF_TUN | unix.IFF_NO_PI)
+	req.Flags = flags
+	copy(req.Name[:], name)
+	if err := ioctl(uintptr(fd), uintptr(unix.TUNSETIFF), uintptr(unsafe.Pointer(&req))); err != nil {
+		return "", err
+	}
+	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 *logrus.Logger, vpnNetworks []netip.Prefix, multiqueue bool) (*tun, error) {
+	baseFlags := uint16(unix.IFF_TUN | unix.IFF_NO_PI)
 	if multiqueue {
-		req.Flags |= unix.IFF_MULTI_QUEUE
+		baseFlags |= unix.IFF_MULTI_QUEUE
 	}
 	nameStr := c.GetString("tun.dev", "")
-	copy(req.Name[:], nameStr)
-	if err = ioctl(uintptr(fd), uintptr(unix.TUNSETIFF), uintptr(unsafe.Pointer(&req))); err != nil {
-		_ = unix.Close(fd)
-		return nil, &NameError{
-			Name:       nameStr,
-			Underlying: err,
-		}
-	}
-	name := strings.Trim(string(req.Name[:]), "\x00")
 
-	t, err := newTunGeneric(c, l, fd, vpnNetworks)
+	// 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
+	}
+	vnetHdr := true
+	name, err := tunSetIff(fd, nameStr, baseFlags|unix.IFF_VNET_HDR)
+	if err != nil {
+		_ = unix.Close(fd)
+		vnetHdr = false
+	} else if err = ioctl(uintptr(fd), unix.TUNSETOFFLOAD, uintptr(tsoOffloadFlags)); err != nil {
+		l.WithError(err).Warn("Failed to enable TUN offload (TSO); proceeding without virtio headers")
+		_ = unix.Close(fd)
+		vnetHdr = false
+	}
+
+	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
 	}
@@ -299,16 +172,29 @@ func newTun(c *config.C, l *logrus.Logger, vpnNetworks []netip.Prefix, multiqueu
 }
 
 // newTunGeneric does all the stuff common to different tun initialization paths. It will close your files on error.
-func newTunGeneric(c *config.C, l *logrus.Logger, fd int, vpnNetworks []netip.Prefix) (*tun, error) {
+func newTunGeneric(c *config.C, l *logrus.Logger, fd int, vnetHdr bool, vpnNetworks []netip.Prefix) (*tun, error) {
-	tfd, err := newTunFd(fd)
+	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
 	}
+	err = container.Add(fd)
+	if err != nil {
+		_ = unix.Close(fd)
+		return nil, err
+	}
+
 	t := &tun{
-		tunFile:                   tfd,
+		readers:                   container,
-		readers:                   []*tunFile{tfd},
 		closeLock:                 sync.Mutex{},
+		vnetHdr:                   vnetHdr,
 		vpnNetworks:               vpnNetworks,
 		TXQueueLen:                c.GetInt("tun.tx_queue", 500),
 		useSystemRoutes:           c.GetBool("tun.use_system_route_table", false),
@@ -410,32 +296,38 @@ func (t *tun) SupportsMultiqueue() bool {
 	return true
 }
 
-func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+func (t *tun) NewMultiQueueReader() error {
 	t.closeLock.Lock()
 	defer t.closeLock.Unlock()
 
 	fd, err := unix.Open("/dev/net/tun", os.O_RDWR, 0)
 	if err != nil {
-		return nil, err
+		return err
 	}
 
-	var req ifReq
+	flags := uint16(unix.IFF_TUN | unix.IFF_NO_PI | unix.IFF_MULTI_QUEUE)
-	req.Flags = uint16(unix.IFF_TUN | unix.IFF_NO_PI | unix.IFF_MULTI_QUEUE)
+	if t.vnetHdr {
-	copy(req.Name[:], t.Device)
+		flags |= unix.IFF_VNET_HDR
-	if err = ioctl(uintptr(fd), uintptr(unix.TUNSETIFF), uintptr(unsafe.Pointer(&req))); err != nil {
+	}
+	if _, err = tunSetIff(fd, t.Device, flags); err != nil {
 		_ = unix.Close(fd)
-		return nil, err
+		return err
 	}
 
-	out, err := t.tunFile.newFriend(fd)
+	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)
-		return nil, err
+		return err
 	}
 
-	t.readers = append(t.readers, out)
+	return nil
-
-	return out, nil
 }
 
 func (t *tun) RoutesFor(ip netip.Addr) routing.Gateways {
@@ -865,6 +757,10 @@ func (t *tun) updateRoutes(r netlink.RouteUpdate) {
 	t.routeTree.Store(newTree)
 }
 
+func (t *tun) Readers() []tio.Queue {
+	return t.readers.Queues()
+}
+
 func (t *tun) Close() error {
 	t.closeLock.Lock()
 	defer t.closeLock.Unlock()
@@ -874,32 +770,10 @@ func (t *tun) Close() error {
 		t.routeChan = nil
 	}
 
-	// Signal all readers blocked in poll to wake up and exit
-	_ = t.tunFile.wakeForShutdown()
-
 	if t.ioctlFd > 0 {
 		_ = unix.Close(int(t.ioctlFd))
 		t.ioctlFd = 0
 	}
 
-	for i := range t.readers {
+	return t.readers.Close()
-		if i == 0 {
-			continue //we want to close the zeroth reader last
-		}
-		err := t.readers[i].Close()
-		if err != nil {
-			t.l.WithField("reader", i).WithError(err).Error("error closing tun reader")
-		} else {
-			t.l.WithField("reader", i).Info("closed tun reader")
-		}
-	}
-
-	//this is t.readers[0] too
-	err := t.tunFile.Close()
-	if err != nil {
-		t.l.WithField("reader", 0).WithError(err).Error("error closing tun reader")
-	} else {
-		t.l.WithField("reader", 0).Info("closed tun reader")
-	}
-	return err
 }

overlay/tun_linux_test.go

@@ -3,7 +3,9 @@
 
 package overlay
 
-import "testing"
+import (
+	"testing"
+)
 
 var runAdvMSSTests = []struct {
 	name     string

overlay/tun_netbsd.go

@@ -6,7 +6,6 @@ package overlay
 import (
 	"errors"
 	"fmt"
-	"io"
 	"net/netip"
 	"os"
 	"regexp"
@@ -17,6 +16,7 @@ import (
 	"github.com/gaissmai/bart"
 	"github.com/sirupsen/logrus"
 	"github.com/slackhq/nebula/config"
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 	"github.com/slackhq/nebula/util"
 	netroute "golang.org/x/net/route"
@@ -66,6 +66,25 @@ type tun struct {
 	l           *logrus.Logger
 	f           *os.File
 	fd          int
+
+	readBuf  []byte
+	batchRet [1][]byte
+}
+
+func (t *tun) Read() ([][]byte, error) {
+	if t.readBuf == nil {
+		t.readBuf = make([]byte, defaultBatchBufSize)
+	}
+	n, err := t.readOne(t.readBuf)
+	if err != nil {
+		return nil, err
+	}
+	t.batchRet[0] = t.readBuf[:n]
+	return t.batchRet[:], nil
+}
+
+func (t *tun) WriteReject(p []byte) (int, error) {
+	return t.Write(p)
 }
 
 var deviceNameRE = regexp.MustCompile(`^tun[0-9]+$`)
@@ -141,7 +160,7 @@ func (t *tun) Close() error {
 	return nil
 }
 
-func (t *tun) Read(to []byte) (int, error) {
+func (t *tun) readOne(to []byte) (int, error) {
 	rc, err := t.f.SyscallConn()
 	if err != nil {
 		return 0, fmt.Errorf("failed to get syscall conn for tun: %w", err)
@@ -394,7 +413,7 @@ func (t *tun) SupportsMultiqueue() bool {
 	return false
 }
 
-func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+func (t *tun) NewMultiQueueReader() (tio.Queue, error) {
 	return nil, fmt.Errorf("TODO: multiqueue not implemented for netbsd")
 }
 

overlay/tun_openbsd.go

@@ -6,7 +6,6 @@ package overlay
 import (
 	"errors"
 	"fmt"
-	"io"
 	"net/netip"
 	"os"
 	"regexp"
@@ -17,6 +16,7 @@ import (
 	"github.com/gaissmai/bart"
 	"github.com/sirupsen/logrus"
 	"github.com/slackhq/nebula/config"
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 	"github.com/slackhq/nebula/util"
 	netroute "golang.org/x/net/route"
@@ -59,6 +59,25 @@ type tun struct {
 	fd          int
 	// cache out buffer since we need to prepend 4 bytes for tun metadata
 	out []byte
+
+	readBuf  []byte
+	batchRet [1][]byte
+}
+
+func (t *tun) Read() ([][]byte, error) {
+	if t.readBuf == nil {
+		t.readBuf = make([]byte, defaultBatchBufSize)
+	}
+	n, err := t.readOne(t.readBuf)
+	if err != nil {
+		return nil, err
+	}
+	t.batchRet[0] = t.readBuf[:n]
+	return t.batchRet[:], nil
+}
+
+func (t *tun) WriteReject(p []byte) (int, error) {
+	return t.Write(p)
 }
 
 var deviceNameRE = regexp.MustCompile(`^tun[0-9]+$`)
@@ -124,7 +143,7 @@ func (t *tun) Close() error {
 	return nil
 }
 
-func (t *tun) Read(to []byte) (int, error) {
+func (t *tun) readOne(to []byte) (int, error) {
 	buf := make([]byte, len(to)+4)
 
 	n, err := t.f.Read(buf)
@@ -314,7 +333,7 @@ func (t *tun) SupportsMultiqueue() bool {
 	return false
 }
 
-func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+func (t *tun) NewMultiQueueReader() (tio.Queue, error) {
 	return nil, fmt.Errorf("TODO: multiqueue not implemented for openbsd")
 }
 

overlay/tun_tester.go

@@ -13,6 +13,7 @@ import (
 	"github.com/gaissmai/bart"
 	"github.com/sirupsen/logrus"
 	"github.com/slackhq/nebula/config"
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 )
 
@@ -26,6 +27,17 @@ type TestTun struct {
 	closed    atomic.Bool
 	rxPackets chan []byte // Packets to receive into nebula
 	TxPackets chan []byte // Packets transmitted outside by nebula
+
+	batchRet [1][]byte
+}
+
+func (t *TestTun) Read() ([][]byte, error) {
+	p, ok := <-t.rxPackets
+	if !ok {
+		return nil, os.ErrClosed
+	}
+	t.batchRet[0] = p
+	return t.batchRet[:], nil
 }
 
 func newTun(c *config.C, l *logrus.Logger, vpnNetworks []netip.Prefix, _ bool) (*TestTun, error) {
@@ -115,6 +127,10 @@ func (t *TestTun) Write(b []byte) (n int, err error) {
 	return len(b), nil
 }
 
+func (t *TestTun) WriteReject(b []byte) (int, error) {
+	return t.Write(b)
+}
+
 func (t *TestTun) Close() error {
 	if t.closed.CompareAndSwap(false, true) {
 		close(t.rxPackets)
@@ -123,19 +139,10 @@ func (t *TestTun) Close() error {
 	return nil
 }
 
-func (t *TestTun) Read(b []byte) (int, error) {
-	p, ok := <-t.rxPackets
-	if !ok {
-		return 0, os.ErrClosed
-	}
-	copy(b, p)
-	return len(p), nil
-}
-
 func (t *TestTun) SupportsMultiqueue() bool {
 	return false
 }
 
-func (t *TestTun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+func (t *TestTun) NewMultiQueueReader() (tio.Queue, error) {
 	return nil, fmt.Errorf("TODO: multiqueue not implemented")
 }

overlay/tun_windows.go

@@ -6,7 +6,6 @@ package overlay
 import (
 	"crypto"
 	"fmt"
-	"io"
 	"net/netip"
 	"os"
 	"path/filepath"
@@ -18,6 +17,7 @@ import (
 	"github.com/gaissmai/bart"
 	"github.com/sirupsen/logrus"
 	"github.com/slackhq/nebula/config"
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 	"github.com/slackhq/nebula/util"
 	"github.com/slackhq/nebula/wintun"
@@ -36,6 +36,25 @@ type winTun struct {
 	l           *logrus.Logger
 
 	tun *wintun.NativeTun
+
+	readBuf  []byte
+	batchRet [1][]byte
+}
+
+func (t *winTun) Read() ([][]byte, error) {
+	if t.readBuf == nil {
+		t.readBuf = make([]byte, defaultBatchBufSize)
+	}
+	n, err := t.tun.Read(t.readBuf, 0)
+	if err != nil {
+		return nil, err
+	}
+	t.batchRet[0] = t.readBuf[:n]
+	return t.batchRet[:], nil
+}
+
+func (t *winTun) WriteReject(p []byte) (int, error) {
+	return t.Write(p)
 }
 
 func newTunFromFd(_ *config.C, _ *logrus.Logger, _ int, _ []netip.Prefix) (Device, error) {
@@ -229,10 +248,6 @@ func (t *winTun) Name() string {
 	return t.Device
 }
 
-func (t *winTun) Read(b []byte) (int, error) {
-	return t.tun.Read(b, 0)
-}
-
 func (t *winTun) Write(b []byte) (int, error) {
 	return t.tun.Write(b, 0)
 }
@@ -241,7 +256,7 @@ func (t *winTun) SupportsMultiqueue() bool {
 	return false
 }
 
-func (t *winTun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+func (t *winTun) NewMultiQueueReader() (tio.Queue, error) {
 	return nil, fmt.Errorf("TODO: multiqueue not implemented for windows")
 }
 

overlay/user.go

@@ -6,6 +6,7 @@ import (
 
 	"github.com/sirupsen/logrus"
 	"github.com/slackhq/nebula/config"
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 )
 
@@ -23,17 +24,34 @@ func NewUserDevice(vpnNetworks []netip.Prefix) (Device, error) {
 		outboundWriter: ow,
 		inboundReader:  ir,
 		inboundWriter:  iw,
+		numReaders:     1,
 	}, nil
 }
 
 type UserDevice struct {
 	vpnNetworks []netip.Prefix
+	numReaders  int
 
 	outboundReader *io.PipeReader
 	outboundWriter *io.PipeWriter
 
 	inboundReader *io.PipeReader
 	inboundWriter *io.PipeWriter
+
+	readBuf  []byte
+	batchRet [1][]byte
+}
+
+func (d *UserDevice) Read() ([][]byte, error) {
+	if d.readBuf == nil {
+		d.readBuf = make([]byte, defaultBatchBufSize)
+	}
+	n, err := d.outboundReader.Read(d.readBuf)
+	if err != nil {
+		return nil, err
+	}
+	d.batchRet[0] = d.readBuf[:n]
+	return d.batchRet[:], nil
 }
 
 func (d *UserDevice) Activate() error {
@@ -50,20 +68,29 @@ func (d *UserDevice) SupportsMultiqueue() bool {
 	return true
 }
 
-func (d *UserDevice) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+func (d *UserDevice) NewMultiQueueReader() error {
-	return d, nil
+	d.numReaders++
+	return nil
+}
+
+func (d *UserDevice) Readers() []tio.Queue {
+	out := make([]tio.Queue, d.numReaders)
+	for i := range d.numReaders {
+		out[i] = d
+	}
+	return out
 }
 
 func (d *UserDevice) Pipe() (*io.PipeReader, *io.PipeWriter) {
 	return d.inboundReader, d.outboundWriter
 }
 
-func (d *UserDevice) Read(p []byte) (n int, err error) {
-	return d.outboundReader.Read(p)
-}
 func (d *UserDevice) Write(p []byte) (n int, err error) {
 	return d.inboundWriter.Write(p)
 }
+func (d *UserDevice) WriteReject(p []byte) (n int, err error) {
+	return d.Write(p)
+}
 func (d *UserDevice) Close() error {
 	d.inboundWriter.Close()
 	d.outboundWriter.Close()

udp/conn.go

@@ -8,6 +8,12 @@ import (
 
 const MTU = 9001
 
+// MaxWriteBatch is the largest batch any Conn.WriteBatch implementation is
+// required to accept. Callers SHOULD NOT pass more than this per call; Linux
+// backends preallocate sendmmsg scratch sized to this value, so exceeding it
+// only costs a chunked retry.
+const MaxWriteBatch = 128
+
 type EncReader func(
 	addr netip.AddrPort,
 	payload []byte,
@@ -16,8 +22,19 @@ type EncReader func(
 type Conn interface {
 	Rebind() error
 	LocalAddr() (netip.AddrPort, error)
-	ListenOut(r EncReader) error
+	// ListenOut invokes r for each received packet. On batch-capable
+	// backends (recvmmsg), flush is called after each batch is fully
+	// delivered — callers use it to flush per-batch accumulators such as
+	// TUN write coalescers. Single-packet backends call flush after each
+	// packet. flush must not be nil.
+	ListenOut(r EncReader, flush func()) error
 	WriteTo(b []byte, addr netip.AddrPort) error
+	// WriteBatch sends a contiguous batch of packets, each with its own
+	// destination. bufs and addrs must have the same length. Linux uses
+	// sendmmsg(2) for a single syscall; other backends fall back to a
+	// WriteTo loop. Returns on the first error; callers may observe a
+	// partial send if some packets went out before the error.
+	WriteBatch(bufs [][]byte, addrs []netip.AddrPort) error
 	ReloadConfig(c *config.C)
 	SupportsMultipleReaders() bool
 	Close() error
@@ -31,7 +48,7 @@ func (NoopConn) Rebind() error {
 func (NoopConn) LocalAddr() (netip.AddrPort, error) {
 	return netip.AddrPort{}, nil
 }
-func (NoopConn) ListenOut(_ EncReader) error {
+func (NoopConn) ListenOut(_ EncReader, _ func()) error {
 	return nil
 }
 func (NoopConn) SupportsMultipleReaders() bool {
@@ -40,6 +57,9 @@ func (NoopConn) SupportsMultipleReaders() bool {
 func (NoopConn) WriteTo(_ []byte, _ netip.AddrPort) error {
 	return nil
 }
+func (NoopConn) WriteBatch(_ [][]byte, _ []netip.AddrPort) error {
+	return nil
+}
 func (NoopConn) ReloadConfig(_ *config.C) {
 	return
 }

udp/udp_darwin.go

@@ -140,6 +140,26 @@ func (u *StdConn) WriteTo(b []byte, ap netip.AddrPort) error {
 	}
 }
 
+func (u *StdConn) WriteBatch(bufs [][]byte, addrs []netip.AddrPort) error {
+	for i, b := range bufs {
+		if err := u.WriteTo(b, addrs[i]); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func (u *StdConn) WriteSegmented(bufs [][]byte, addr netip.AddrPort, _ int) error {
+	for _, b := range bufs {
+		if err := u.WriteTo(b, addr); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func (u *StdConn) SupportsGSO() bool { return false }
+
 func (u *StdConn) LocalAddr() (netip.AddrPort, error) {
 	a := u.UDPConn.LocalAddr()
 
@@ -165,7 +185,7 @@ func NewUDPStatsEmitter(udpConns []Conn) func() {
 	return func() {}
 }
 
-func (u *StdConn) ListenOut(r EncReader) error {
+func (u *StdConn) ListenOut(r EncReader, flush func()) error {
 	buffer := make([]byte, MTU)
 
 	for {
@@ -180,6 +200,7 @@ func (u *StdConn) ListenOut(r EncReader) error {
 		}
 
 		r(netip.AddrPortFrom(rua.Addr().Unmap(), rua.Port()), buffer[:n])
+		flush()
 	}
 }
 

udp/udp_generic.go

@@ -44,6 +44,26 @@ func (u *GenericConn) WriteTo(b []byte, addr netip.AddrPort) error {
 	return err
 }
 
+func (u *GenericConn) WriteBatch(bufs [][]byte, addrs []netip.AddrPort) error {
+	for i, b := range bufs {
+		if _, err := u.UDPConn.WriteToUDPAddrPort(b, addrs[i]); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func (u *GenericConn) WriteSegmented(bufs [][]byte, addr netip.AddrPort, _ int) error {
+	for _, b := range bufs {
+		if _, err := u.UDPConn.WriteToUDPAddrPort(b, addr); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func (u *GenericConn) SupportsGSO() bool { return false }
+
 func (u *GenericConn) LocalAddr() (netip.AddrPort, error) {
 	a := u.UDPConn.LocalAddr()
 
@@ -73,7 +93,7 @@ type rawMessage struct {
 	Len uint32
 }
 
-func (u *GenericConn) ListenOut(r EncReader) error {
+func (u *GenericConn) ListenOut(r EncReader, flush func()) error {
 	buffer := make([]byte, MTU)
 
 	var lastRecvErr time.Time
@@ -94,6 +114,7 @@ func (u *GenericConn) ListenOut(r EncReader) error {
 		}
 
 		r(netip.AddrPortFrom(rua.Addr().Unmap(), rua.Port()), buffer[:n])
+		flush()
 	}
 }
 

udp/udp_linux.go

@@ -24,6 +24,43 @@ type StdConn struct {
 	isV4    bool
 	l       *logrus.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 *logrus.Logger, ip netip.Addr, port int, multi bool, batch in
 	}
 	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
 }
@@ -171,7 +301,7 @@ func recvmmsg(fd uintptr, msgs []rawMessage) (int, bool, error) {
 	return int(n), true, nil
 }
 
-func (u *StdConn) listenOutSingle(r EncReader) error {
+func (u *StdConn) listenOutSingle(r EncReader, flush func()) error {
 	var err error
 	var n int
 	var from netip.AddrPort
@@ -184,15 +314,22 @@ func (u *StdConn) listenOutSingle(r EncReader) error {
 		}
 		from = netip.AddrPortFrom(from.Addr().Unmap(), from.Port())
 		r(from, buffer[:n])
+		flush()
 	}
 }
 
-func (u *StdConn) listenOutBatch(r EncReader) error {
+func (u *StdConn) listenOutBatch(r EncReader, flush func()) error {
 	var ip netip.Addr
 	var n int
 	var operr error
 
-	msgs, buffers, names := u.PrepareRawMessages(u.batch)
+	bufSize := MTU
+	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
@@ -202,6 +339,11 @@ func (u *StdConn) listenOutBatch(r EncReader) error {
 	}
 
 	for {
+		if cmsgSpace > 0 {
+			for i := range msgs {
+				setMsgControllen(&msgs[i].Hdr, cmsgSpace)
+			}
+		}
 		err := u.rawConn.Read(reader)
 		if err != nil {
 			return err
@@ -217,16 +359,72 @@ func (u *StdConn) listenOutBatch(r EncReader) error {
 			} else {
 				ip, _ = netip.AddrFromSlice(names[i][8:24])
 			}
-			r(netip.AddrPortFrom(ip.Unmap(), binary.BigEndian.Uint16(names[i][2:4])), buffers[i][:msgs[i].Len])
+			from := netip.AddrPortFrom(ip.Unmap(), binary.BigEndian.Uint16(names[i][2:4]))
+			payload := buffers[i][:msgs[i].Len]
+
+			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.
+		flush()
 	}
 }
 
-func (u *StdConn) ListenOut(r EncReader) 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)
+		return u.listenOutSingle(r, flush)
 	} else {
-		return u.listenOutBatch(r)
+		return u.listenOutBatch(r, flush)
 	}
 }
 
@@ -235,6 +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))
+	}
+
+	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

@@ -30,13 +30,18 @@ type rawMessage struct {
 	Len uint32
 }
 
-func (u *StdConn) PrepareRawMessages(n 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, MTU)
+		buffers[i] = make([]byte, bufSize)
 		names[i] = make([]byte, unix.SizeofSockaddrInet6)
 
 		vs := []iovec{
@@ -48,7 +53,28 @@ func (u *StdConn) PrepareRawMessages(n int) ([]rawMessage, [][]byte, [][]byte) {
 
 		msgs[i].Hdr.Name = &names[i][0]
 		msgs[i].Hdr.Namelen = uint32(len(names[i]))
+
+		if 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)
 }

udp/udp_linux_64.go

@@ -33,13 +33,18 @@ type rawMessage struct {
 	Pad0 [4]byte
 }
 
-func (u *StdConn) PrepareRawMessages(n 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, MTU)
+		buffers[i] = make([]byte, bufSize)
 		names[i] = make([]byte, unix.SizeofSockaddrInet6)
 
 		vs := []iovec{
@@ -51,7 +56,28 @@ func (u *StdConn) PrepareRawMessages(n int) ([]rawMessage, [][]byte, [][]byte) {
 
 		msgs[i].Hdr.Name = &names[i][0]
 		msgs[i].Hdr.Namelen = uint32(len(names[i]))
+
+		if 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)
 }

udp/udp_rio_windows.go

@@ -140,7 +140,7 @@ func (u *RIOConn) bind(l *logrus.Logger, sa windows.Sockaddr) error {
 	return nil
 }
 
-func (u *RIOConn) ListenOut(r EncReader) error {
+func (u *RIOConn) ListenOut(r EncReader, flush func()) error {
 	buffer := make([]byte, MTU)
 
 	var lastRecvErr time.Time
@@ -162,6 +162,7 @@ func (u *RIOConn) ListenOut(r EncReader) error {
 		}
 
 		r(netip.AddrPortFrom(netip.AddrFrom16(rua.Addr).Unmap(), (rua.Port>>8)|((rua.Port&0xff)<<8)), buffer[:n])
+		flush()
 	}
 }
 
@@ -316,6 +317,26 @@ func (u *RIOConn) WriteTo(buf []byte, ip netip.AddrPort) error {
 	return winrio.SendEx(u.rq, dataBuffer, 1, nil, addressBuffer, nil, nil, 0, 0)
 }
 
+func (u *RIOConn) WriteBatch(bufs [][]byte, addrs []netip.AddrPort) error {
+	for i, b := range bufs {
+		if err := u.WriteTo(b, addrs[i]); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func (u *RIOConn) WriteSegmented(bufs [][]byte, addr netip.AddrPort, _ int) error {
+	for _, b := range bufs {
+		if err := u.WriteTo(b, addr); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func (u *RIOConn) SupportsGSO() bool { return false }
+
 func (u *RIOConn) LocalAddr() (netip.AddrPort, error) {
 	sa, err := windows.Getsockname(u.sock)
 	if err != nil {

udp/udp_tester.go

@@ -107,13 +107,34 @@ func (u *TesterConn) WriteTo(b []byte, addr netip.AddrPort) error {
 	return nil
 }
 
-func (u *TesterConn) ListenOut(r EncReader) error {
+func (u *TesterConn) WriteBatch(bufs [][]byte, addrs []netip.AddrPort) error {
+	for i, b := range bufs {
+		if err := u.WriteTo(b, addrs[i]); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func (u *TesterConn) WriteSegmented(bufs [][]byte, addr netip.AddrPort, _ int) error {
+	for _, b := range bufs {
+		if err := u.WriteTo(b, addr); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func (u *TesterConn) SupportsGSO() bool { return false }
+
+func (u *TesterConn) ListenOut(r EncReader, flush func()) error {
 	for {
 		p, ok := <-u.RxPackets
 		if !ok {
 			return os.ErrClosed
 		}
 		r(p.From, p.Data)
+		flush()
 	}
 }