checkpt3

CHANGELOG.md

@@ -7,6 +7,13 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+### Added
+
+- Experimental Linux UDP offload support: enable `listen.enable_gso` and
+  `listen.enable_gro` to activate UDP_SEGMENT batching and GRO receive
+  splitting. Includes automatic capability probing, per-packet fallbacks, and
+  runtime metrics/logs for visibility.
+
 ### Changed
 
 - `default_local_cidr_any` now defaults to false, meaning that any firewall rule

cert_test/cert.go

@@ -114,33 +114,6 @@ func NewTestCert(v cert.Version, curve cert.Curve, ca cert.Certificate, key []by
 	return c, pub, cert.MarshalPrivateKeyToPEM(curve, priv), pem
 }
 
-func NewTestCertDifferentVersion(c cert.Certificate, v cert.Version, ca cert.Certificate, key []byte) (cert.Certificate, []byte) {
-	nc := &cert.TBSCertificate{
-		Version:        v,
-		Curve:          c.Curve(),
-		Name:           c.Name(),
-		Networks:       c.Networks(),
-		UnsafeNetworks: c.UnsafeNetworks(),
-		Groups:         c.Groups(),
-		NotBefore:      time.Unix(c.NotBefore().Unix(), 0),
-		NotAfter:       time.Unix(c.NotAfter().Unix(), 0),
-		PublicKey:      c.PublicKey(),
-		IsCA:           false,
-	}
-
-	c, err := nc.Sign(ca, ca.Curve(), key)
-	if err != nil {
-		panic(err)
-	}
-
-	pem, err := c.MarshalPEM()
-	if err != nil {
-		panic(err)
-	}
-
-	return c, pem
-}
-
 func X25519Keypair() ([]byte, []byte) {
 	privkey := make([]byte, 32)
 	if _, err := io.ReadFull(rand.Reader, privkey); err != nil {

connection_manager.go

@@ -354,6 +354,7 @@ func (cm *connectionManager) makeTrafficDecision(localIndex uint32, now time.Tim
 
 		if mainHostInfo {
 			decision = tryRehandshake
+
 		} else {
 			if cm.shouldSwapPrimary(hostinfo) {
 				decision = swapPrimary
@@ -460,10 +461,6 @@ func (cm *connectionManager) shouldSwapPrimary(current *HostInfo) bool {
 	}
 
 	crt := cm.intf.pki.getCertState().getCertificate(current.ConnectionState.myCert.Version())
-	if crt == nil {
-		//my cert was reloaded away. We should definitely swap from this tunnel
-		return true
-	}
 	// If this tunnel is using the latest certificate then we should swap it to primary for a bit and see if things
 	// settle down.
 	return bytes.Equal(current.ConnectionState.myCert.Signature(), crt.Signature())
@@ -478,34 +475,31 @@ func (cm *connectionManager) swapPrimary(current, primary *HostInfo) {
 	cm.hostMap.Unlock()
 }
 
-// isInvalidCertificate decides if we should destroy a tunnel.
-// returns true if pki.disconnect_invalid is true and the certificate is no longer valid.
-// Blocklisted certificates will skip the pki.disconnect_invalid check and return true.
+// isInvalidCertificate will check if we should destroy a tunnel if pki.disconnect_invalid is true and
+// the certificate is no longer valid. Block listed certificates will skip the pki.disconnect_invalid
+// check and return true.
 func (cm *connectionManager) isInvalidCertificate(now time.Time, hostinfo *HostInfo) bool {
 	remoteCert := hostinfo.GetCert()
 	if remoteCert == nil {
-		return false //don't tear down tunnels for handshakes in progress
+		return false
 	}
 
 	caPool := cm.intf.pki.GetCAPool()
 	err := caPool.VerifyCachedCertificate(now, remoteCert)
 	if err == nil {
-		return false //cert is still valid! yay!
-	} else if err == cert.ErrBlockListed { //avoiding errors.Is for speed
+		return false
+	}
+
+	if !cm.intf.disconnectInvalid.Load() && err != cert.ErrBlockListed {
 		// Block listed certificates should always be disconnected
-		hostinfo.logger(cm.l).WithError(err).
-			WithField("fingerprint", remoteCert.Fingerprint).
-			Info("Remote certificate is blocked, tearing down the tunnel")
-		return true
-	} else if cm.intf.disconnectInvalid.Load() {
+		return false
+	}
+
 	hostinfo.logger(cm.l).WithError(err).
 		WithField("fingerprint", remoteCert.Fingerprint).
 		Info("Remote certificate is no longer valid, tearing down the tunnel")
+
 	return true
-	} else {
-		//if we reach here, the cert is no longer valid, but we're configured to keep tunnels from now-invalid certs open
-		return false
-	}
 }
 
 func (cm *connectionManager) sendPunch(hostinfo *HostInfo) {
@@ -536,45 +530,15 @@ func (cm *connectionManager) sendPunch(hostinfo *HostInfo) {
 func (cm *connectionManager) tryRehandshake(hostinfo *HostInfo) {
 	cs := cm.intf.pki.getCertState()
 	curCrt := hostinfo.ConnectionState.myCert
-	curCrtVersion := curCrt.Version()
-	myCrt := cs.getCertificate(curCrtVersion)
-	if myCrt == nil {
-		cm.l.WithField("vpnAddrs", hostinfo.vpnAddrs).
-			WithField("version", curCrtVersion).
-			WithField("reason", "local certificate removed").
-			Info("Re-handshaking with remote")
-		cm.intf.handshakeManager.StartHandshake(hostinfo.vpnAddrs[0], nil)
+	myCrt := cs.getCertificate(curCrt.Version())
+	if curCrt.Version() >= cs.initiatingVersion && bytes.Equal(curCrt.Signature(), myCrt.Signature()) == true {
+		// The current tunnel is using the latest certificate and version, no need to rehandshake.
 		return
 	}
-	peerCrt := hostinfo.ConnectionState.peerCert
-	if peerCrt != nil && curCrtVersion < peerCrt.Certificate.Version() {
-		// if our certificate version is less than theirs, and we have a matching version available, rehandshake?
-		if cs.getCertificate(peerCrt.Certificate.Version()) != nil {
-			cm.l.WithField("vpnAddrs", hostinfo.vpnAddrs).
-				WithField("version", curCrtVersion).
-				WithField("peerVersion", peerCrt.Certificate.Version()).
-				WithField("reason", "local certificate version lower than peer, attempting to correct").
-				Info("Re-handshaking with remote")
-			cm.intf.handshakeManager.StartHandshake(hostinfo.vpnAddrs[0], func(hh *HandshakeHostInfo) {
-				hh.initiatingVersionOverride = peerCrt.Certificate.Version()
-			})
-			return
-		}
-	}
-	if !bytes.Equal(curCrt.Signature(), myCrt.Signature()) {
+
 	cm.l.WithField("vpnAddrs", hostinfo.vpnAddrs).
 		WithField("reason", "local certificate is not current").
 		Info("Re-handshaking with remote")
 
 	cm.intf.handshakeManager.StartHandshake(hostinfo.vpnAddrs[0], nil)
-		return
-	}
-	if curCrtVersion < cs.initiatingVersion {
-		cm.l.WithField("vpnAddrs", hostinfo.vpnAddrs).
-			WithField("reason", "current cert version < pki.initiatingVersion").
-			Info("Re-handshaking with remote")
-
-		cm.intf.handshakeManager.StartHandshake(hostinfo.vpnAddrs[0], nil)
-		return
-	}
 }

e2e/helpers_test.go

@@ -129,109 +129,6 @@ func newSimpleServer(v cert.Version, caCrt cert.Certificate, caKey []byte, name
 	return control, vpnNetworks, udpAddr, c
 }
 
-// newServer creates a nebula instance with fewer assumptions
-func newServer(caCrt []cert.Certificate, certs []cert.Certificate, key []byte, overrides m) (*nebula.Control, []netip.Prefix, netip.AddrPort, *config.C) {
-	l := NewTestLogger()
-
-	vpnNetworks := certs[len(certs)-1].Networks()
-
-	var udpAddr netip.AddrPort
-	if vpnNetworks[0].Addr().Is4() {
-		budpIp := vpnNetworks[0].Addr().As4()
-		budpIp[1] -= 128
-		udpAddr = netip.AddrPortFrom(netip.AddrFrom4(budpIp), 4242)
-	} else {
-		budpIp := vpnNetworks[0].Addr().As16()
-		// beef for funsies
-		budpIp[2] = 190
-		budpIp[3] = 239
-		udpAddr = netip.AddrPortFrom(netip.AddrFrom16(budpIp), 4242)
-	}
-
-	caStr := ""
-	for _, ca := range caCrt {
-		x, err := ca.MarshalPEM()
-		if err != nil {
-			panic(err)
-		}
-		caStr += string(x)
-	}
-	certStr := ""
-	for _, c := range certs {
-		x, err := c.MarshalPEM()
-		if err != nil {
-			panic(err)
-		}
-		certStr += string(x)
-	}
-
-	mc := m{
-		"pki": m{
-			"ca":   caStr,
-			"cert": certStr,
-			"key":  string(key),
-		},
-		//"tun": m{"disabled": true},
-		"firewall": m{
-			"outbound": []m{{
-				"proto": "any",
-				"port":  "any",
-				"host":  "any",
-			}},
-			"inbound": []m{{
-				"proto": "any",
-				"port":  "any",
-				"host":  "any",
-			}},
-		},
-		//"handshakes": m{
-		//	"try_interval": "1s",
-		//},
-		"listen": m{
-			"host": udpAddr.Addr().String(),
-			"port": udpAddr.Port(),
-		},
-		"logging": m{
-			"timestamp_format": fmt.Sprintf("%v 15:04:05.000000", certs[0].Name()),
-			"level":            l.Level.String(),
-		},
-		"timers": m{
-			"pending_deletion_interval": 2,
-			"connection_alive_interval": 2,
-		},
-	}
-
-	if overrides != nil {
-		final := m{}
-		err := mergo.Merge(&final, overrides, mergo.WithAppendSlice)
-		if err != nil {
-			panic(err)
-		}
-		err = mergo.Merge(&final, mc, mergo.WithAppendSlice)
-		if err != nil {
-			panic(err)
-		}
-		mc = final
-	}
-
-	cb, err := yaml.Marshal(mc)
-	if err != nil {
-		panic(err)
-	}
-
-	c := config.NewC(l)
-	cStr := string(cb)
-	c.LoadString(cStr)
-
-	control, err := nebula.Main(c, false, "e2e-test", l, nil)
-
-	if err != nil {
-		panic(err)
-	}
-
-	return control, vpnNetworks, udpAddr, c
-}
-
 type doneCb func()
 
 func deadline(t *testing.T, seconds time.Duration) doneCb {

e2e/tunnels_test.go

@@ -4,16 +4,12 @@
 package e2e
 
 import (
-	"fmt"
-	"net/netip"
 	"testing"
 	"time"
 
 	"github.com/slackhq/nebula/cert"
 	"github.com/slackhq/nebula/cert_test"
 	"github.com/slackhq/nebula/e2e/router"
-	"github.com/stretchr/testify/assert"
-	"gopkg.in/yaml.v3"
 )
 
 func TestDropInactiveTunnels(t *testing.T) {
@@ -59,262 +55,3 @@ func TestDropInactiveTunnels(t *testing.T) {
 	myControl.Stop()
 	theirControl.Stop()
 }
-
-func TestCertUpgrade(t *testing.T) {
-	// The goal of this test is to ensure the shortest inactivity timeout will close the tunnel on both sides
-	// under ideal conditions
-	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version1, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
-	caB, err := ca.MarshalPEM()
-	if err != nil {
-		panic(err)
-	}
-	ca2, _, caKey2, _ := cert_test.NewTestCaCert(cert.Version2, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
-
-	ca2B, err := ca2.MarshalPEM()
-	if err != nil {
-		panic(err)
-	}
-	caStr := fmt.Sprintf("%s\n%s", caB, ca2B)
-
-	myCert, _, myPrivKey, _ := cert_test.NewTestCert(cert.Version1, cert.Curve_CURVE25519, ca, caKey, "me", time.Now(), time.Now().Add(5*time.Minute), []netip.Prefix{netip.MustParsePrefix("10.128.0.1/24")}, nil, []string{})
-	_, myCert2Pem := cert_test.NewTestCertDifferentVersion(myCert, cert.Version2, ca2, caKey2)
-
-	theirCert, _, theirPrivKey, _ := cert_test.NewTestCert(cert.Version1, cert.Curve_CURVE25519, ca, caKey, "them", time.Now(), time.Now().Add(5*time.Minute), []netip.Prefix{netip.MustParsePrefix("10.128.0.2/24")}, nil, []string{})
-	theirCert2, _ := cert_test.NewTestCertDifferentVersion(theirCert, cert.Version2, ca2, caKey2)
-
-	myControl, myVpnIpNet, myUdpAddr, myC := newServer([]cert.Certificate{ca, ca2}, []cert.Certificate{myCert}, myPrivKey, m{})
-	theirControl, theirVpnIpNet, theirUdpAddr, _ := newServer([]cert.Certificate{ca, ca2}, []cert.Certificate{theirCert, theirCert2}, theirPrivKey, m{})
-
-	// Share our underlay information
-	myControl.InjectLightHouseAddr(theirVpnIpNet[0].Addr(), theirUdpAddr)
-	theirControl.InjectLightHouseAddr(myVpnIpNet[0].Addr(), myUdpAddr)
-
-	// Start the servers
-	myControl.Start()
-	theirControl.Start()
-
-	r := router.NewR(t, myControl, theirControl)
-	defer r.RenderFlow()
-
-	r.Log("Assert the tunnel between me and them works")
-	assertTunnel(t, myVpnIpNet[0].Addr(), theirVpnIpNet[0].Addr(), myControl, theirControl, r)
-	r.Log("yay")
-	//todo ???
-	time.Sleep(1 * time.Second)
-	r.FlushAll()
-
-	mc := m{
-		"pki": m{
-			"ca":   caStr,
-			"cert": string(myCert2Pem),
-			"key":  string(myPrivKey),
-		},
-		//"tun": m{"disabled": true},
-		"firewall": myC.Settings["firewall"],
-		//"handshakes": m{
-		//	"try_interval": "1s",
-		//},
-		"listen":  myC.Settings["listen"],
-		"logging": myC.Settings["logging"],
-		"timers":  myC.Settings["timers"],
-	}
-
-	cb, err := yaml.Marshal(mc)
-	if err != nil {
-		panic(err)
-	}
-
-	r.Logf("reload new v2-only config")
-	err = myC.ReloadConfigString(string(cb))
-	assert.NoError(t, err)
-	r.Log("yay, spin until their sees it")
-	waitStart := time.Now()
-	for {
-		assertTunnel(t, myVpnIpNet[0].Addr(), theirVpnIpNet[0].Addr(), myControl, theirControl, r)
-		c := theirControl.GetHostInfoByVpnAddr(myVpnIpNet[0].Addr(), false)
-		if c == nil {
-			r.Log("nil")
-		} else {
-			version := c.Cert.Version()
-			r.Logf("version %d", version)
-			if version == cert.Version2 {
-				break
-			}
-		}
-		since := time.Since(waitStart)
-		if since > time.Second*10 {
-			t.Fatal("Cert should be new by now")
-		}
-		time.Sleep(time.Second)
-	}
-
-	r.RenderHostmaps("Final hostmaps", myControl, theirControl)
-
-	myControl.Stop()
-	theirControl.Stop()
-}
-
-func TestCertDowngrade(t *testing.T) {
-	// The goal of this test is to ensure the shortest inactivity timeout will close the tunnel on both sides
-	// under ideal conditions
-	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version1, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
-	caB, err := ca.MarshalPEM()
-	if err != nil {
-		panic(err)
-	}
-	ca2, _, caKey2, _ := cert_test.NewTestCaCert(cert.Version2, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
-
-	ca2B, err := ca2.MarshalPEM()
-	if err != nil {
-		panic(err)
-	}
-	caStr := fmt.Sprintf("%s\n%s", caB, ca2B)
-
-	myCert, _, myPrivKey, myCertPem := cert_test.NewTestCert(cert.Version1, cert.Curve_CURVE25519, ca, caKey, "me", time.Now(), time.Now().Add(5*time.Minute), []netip.Prefix{netip.MustParsePrefix("10.128.0.1/24")}, nil, []string{})
-	myCert2, _ := cert_test.NewTestCertDifferentVersion(myCert, cert.Version2, ca2, caKey2)
-
-	theirCert, _, theirPrivKey, _ := cert_test.NewTestCert(cert.Version1, cert.Curve_CURVE25519, ca, caKey, "them", time.Now(), time.Now().Add(5*time.Minute), []netip.Prefix{netip.MustParsePrefix("10.128.0.2/24")}, nil, []string{})
-	theirCert2, _ := cert_test.NewTestCertDifferentVersion(theirCert, cert.Version2, ca2, caKey2)
-
-	myControl, myVpnIpNet, myUdpAddr, myC := newServer([]cert.Certificate{ca, ca2}, []cert.Certificate{myCert2}, myPrivKey, m{})
-	theirControl, theirVpnIpNet, theirUdpAddr, _ := newServer([]cert.Certificate{ca, ca2}, []cert.Certificate{theirCert, theirCert2}, theirPrivKey, m{})
-
-	// Share our underlay information
-	myControl.InjectLightHouseAddr(theirVpnIpNet[0].Addr(), theirUdpAddr)
-	theirControl.InjectLightHouseAddr(myVpnIpNet[0].Addr(), myUdpAddr)
-
-	// Start the servers
-	myControl.Start()
-	theirControl.Start()
-
-	r := router.NewR(t, myControl, theirControl)
-	defer r.RenderFlow()
-
-	r.Log("Assert the tunnel between me and them works")
-	//assertTunnel(t, theirVpnIpNet[0].Addr(), myVpnIpNet[0].Addr(), theirControl, myControl, r)
-	//r.Log("yay")
-	assertTunnel(t, myVpnIpNet[0].Addr(), theirVpnIpNet[0].Addr(), myControl, theirControl, r)
-	r.Log("yay")
-	//todo ???
-	time.Sleep(1 * time.Second)
-	r.FlushAll()
-
-	mc := m{
-		"pki": m{
-			"ca":   caStr,
-			"cert": string(myCertPem),
-			"key":  string(myPrivKey),
-		},
-		"firewall": myC.Settings["firewall"],
-		"listen":   myC.Settings["listen"],
-		"logging":  myC.Settings["logging"],
-		"timers":   myC.Settings["timers"],
-	}
-
-	cb, err := yaml.Marshal(mc)
-	if err != nil {
-		panic(err)
-	}
-
-	r.Logf("reload new v1-only config")
-	err = myC.ReloadConfigString(string(cb))
-	assert.NoError(t, err)
-	r.Log("yay, spin until their sees it")
-	waitStart := time.Now()
-	for {
-		assertTunnel(t, myVpnIpNet[0].Addr(), theirVpnIpNet[0].Addr(), myControl, theirControl, r)
-		c := theirControl.GetHostInfoByVpnAddr(myVpnIpNet[0].Addr(), false)
-		c2 := myControl.GetHostInfoByVpnAddr(theirVpnIpNet[0].Addr(), false)
-		if c == nil || c2 == nil {
-			r.Log("nil")
-		} else {
-			version := c.Cert.Version()
-			theirVersion := c2.Cert.Version()
-			r.Logf("version %d,%d", version, theirVersion)
-			if version == cert.Version1 {
-				break
-			}
-		}
-		since := time.Since(waitStart)
-		if since > time.Second*5 {
-			r.Log("it is unusual that the cert is not new yet, but not a failure yet")
-		}
-		if since > time.Second*10 {
-			r.Log("wtf")
-			t.Fatal("Cert should be new by now")
-		}
-		time.Sleep(time.Second)
-	}
-
-	r.RenderHostmaps("Final hostmaps", myControl, theirControl)
-
-	myControl.Stop()
-	theirControl.Stop()
-}
-
-func TestCertMismatchCorrection(t *testing.T) {
-	// The goal of this test is to ensure the shortest inactivity timeout will close the tunnel on both sides
-	// under ideal conditions
-	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version1, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
-	ca2, _, caKey2, _ := cert_test.NewTestCaCert(cert.Version2, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
-
-	myCert, _, myPrivKey, _ := cert_test.NewTestCert(cert.Version1, cert.Curve_CURVE25519, ca, caKey, "me", time.Now(), time.Now().Add(5*time.Minute), []netip.Prefix{netip.MustParsePrefix("10.128.0.1/24")}, nil, []string{})
-	myCert2, _ := cert_test.NewTestCertDifferentVersion(myCert, cert.Version2, ca2, caKey2)
-
-	theirCert, _, theirPrivKey, _ := cert_test.NewTestCert(cert.Version1, cert.Curve_CURVE25519, ca, caKey, "them", time.Now(), time.Now().Add(5*time.Minute), []netip.Prefix{netip.MustParsePrefix("10.128.0.2/24")}, nil, []string{})
-	theirCert2, _ := cert_test.NewTestCertDifferentVersion(theirCert, cert.Version2, ca2, caKey2)
-
-	myControl, myVpnIpNet, myUdpAddr, _ := newServer([]cert.Certificate{ca, ca2}, []cert.Certificate{myCert2}, myPrivKey, m{})
-	theirControl, theirVpnIpNet, theirUdpAddr, _ := newServer([]cert.Certificate{ca, ca2}, []cert.Certificate{theirCert, theirCert2}, theirPrivKey, m{})
-
-	// Share our underlay information
-	myControl.InjectLightHouseAddr(theirVpnIpNet[0].Addr(), theirUdpAddr)
-	theirControl.InjectLightHouseAddr(myVpnIpNet[0].Addr(), myUdpAddr)
-
-	// Start the servers
-	myControl.Start()
-	theirControl.Start()
-
-	r := router.NewR(t, myControl, theirControl)
-	defer r.RenderFlow()
-
-	r.Log("Assert the tunnel between me and them works")
-	//assertTunnel(t, theirVpnIpNet[0].Addr(), myVpnIpNet[0].Addr(), theirControl, myControl, r)
-	//r.Log("yay")
-	assertTunnel(t, myVpnIpNet[0].Addr(), theirVpnIpNet[0].Addr(), myControl, theirControl, r)
-	r.Log("yay")
-	//todo ???
-	time.Sleep(1 * time.Second)
-	r.FlushAll()
-
-	waitStart := time.Now()
-	for {
-		assertTunnel(t, myVpnIpNet[0].Addr(), theirVpnIpNet[0].Addr(), myControl, theirControl, r)
-		c := theirControl.GetHostInfoByVpnAddr(myVpnIpNet[0].Addr(), false)
-		c2 := myControl.GetHostInfoByVpnAddr(theirVpnIpNet[0].Addr(), false)
-		if c == nil || c2 == nil {
-			r.Log("nil")
-		} else {
-			version := c.Cert.Version()
-			theirVersion := c2.Cert.Version()
-			r.Logf("version %d,%d", version, theirVersion)
-			if version == theirVersion {
-				break
-			}
-		}
-		since := time.Since(waitStart)
-		if since > time.Second*5 {
-			r.Log("wtf")
-		}
-		if since > time.Second*10 {
-			r.Log("wtf")
-			t.Fatal("Cert should be new by now")
-		}
-		time.Sleep(time.Second)
-	}
-
-	r.RenderHostmaps("Final hostmaps", myControl, theirControl)
-
-	myControl.Stop()
-	theirControl.Stop()
-}

handshake_ix.go

@@ -23,19 +23,15 @@ func ixHandshakeStage0(f *Interface, hh *HandshakeHostInfo) bool {
 		return false
 	}
 
+	// If we're connecting to a v6 address we must use a v2 cert
 	cs := f.pki.getCertState()
 	v := cs.initiatingVersion
-	if hh.initiatingVersionOverride != cert.VersionPre1 {
-		v = hh.initiatingVersionOverride
-	} else if v < cert.Version2 {
-		// If we're connecting to a v6 address we should encourage use of a V2 cert
 	for _, a := range hh.hostinfo.vpnAddrs {
 		if a.Is6() {
 			v = cert.Version2
 			break
 		}
 	}
-	}
 
 	crt := cs.getCertificate(v)
 	if crt == nil {
@@ -52,7 +48,6 @@ func ixHandshakeStage0(f *Interface, hh *HandshakeHostInfo) bool {
 			WithField("handshake", m{"stage": 0, "style": "ix_psk0"}).
 			WithField("certVersion", v).
 			Error("Unable to handshake with host because no certificate handshake bytes is available")
-		return false
 	}
 
 	ci, err := NewConnectionState(f.l, cs, crt, true, noise.HandshakeIX)
@@ -108,7 +103,6 @@ func ixHandshakeStage1(f *Interface, addr netip.AddrPort, via *ViaSender, packet
 			WithField("handshake", m{"stage": 0, "style": "ix_psk0"}).
 			WithField("certVersion", cs.initiatingVersion).
 			Error("Unable to handshake with host because no certificate is available")
-		return
 	}
 
 	ci, err := NewConnectionState(f.l, cs, crt, false, noise.HandshakeIX)
@@ -149,8 +143,8 @@ func ixHandshakeStage1(f *Interface, addr netip.AddrPort, via *ViaSender, packet
 
 	remoteCert, err := f.pki.GetCAPool().VerifyCertificate(time.Now(), rc)
 	if err != nil {
-		fp, fperr := rc.Fingerprint()
-		if fperr != nil {
+		fp, err := rc.Fingerprint()
+		if err != nil {
 			fp = "<error generating certificate fingerprint>"
 		}
 
@@ -169,19 +163,16 @@ func ixHandshakeStage1(f *Interface, addr netip.AddrPort, via *ViaSender, packet
 
 	if remoteCert.Certificate.Version() != ci.myCert.Version() {
 		// We started off using the wrong certificate version, lets see if we can match the version that was sent to us
-		myCertOtherVersion := cs.getCertificate(remoteCert.Certificate.Version())
-		if myCertOtherVersion == nil {
-			if f.l.Level >= logrus.DebugLevel {
-				f.l.WithError(err).WithFields(m{
-					"udpAddr":   addr,
-					"handshake": m{"stage": 1, "style": "ix_psk0"},
-					"cert":      remoteCert,
-				}).Debug("Might be unable to handshake with host due to missing certificate version")
+		rc := cs.getCertificate(remoteCert.Certificate.Version())
+		if rc == nil {
+			f.l.WithError(err).WithField("udpAddr", addr).
+				WithField("handshake", m{"stage": 1, "style": "ix_psk0"}).WithField("cert", remoteCert).
+				Info("Unable to handshake with host due to missing certificate version")
+			return
 		}
-		} else {
+
 		// Record the certificate we are actually using
-			ci.myCert = myCertOtherVersion
-		}
+		ci.myCert = rc
 	}
 
 	if len(remoteCert.Certificate.Networks()) == 0 {

handshake_manager.go

@@ -70,7 +70,6 @@ type HandshakeHostInfo struct {
 
 	startTime   time.Time        // Time that we first started trying with this handshake
 	ready       bool             // Is the handshake ready
-	initiatingVersionOverride cert.Version     // Should we use a non-default cert version for this handshake?
 	counter     int64            // How many attempts have we made so far
 	lastRemotes []netip.AddrPort // Remotes that we sent to during the previous attempt
 	packetStore []*cachedPacket  // A set of packets to be transmitted once the handshake completes

inside.go

@@ -11,149 +11,6 @@ import (
 	"github.com/slackhq/nebula/routing"
 )
 
-// consumeInsidePackets processes multiple packets in a batch for improved performance
-// packets: slice of packet buffers to process
-// sizes: slice of packet sizes
-// count: number of packets to process
-// outs: slice of output buffers (one per packet) with virtio headroom
-// q: queue index
-// localCache: firewall conntrack cache
-// batchPackets: pre-allocated slice for accumulating encrypted packets
-// batchAddrs: pre-allocated slice for accumulating destination addresses
-func (f *Interface) consumeInsidePackets(packets [][]byte, sizes []int, count int, outs [][]byte, nb []byte, q int, localCache firewall.ConntrackCache, batchPackets *[][]byte, batchAddrs *[]netip.AddrPort) {
-	// Reusable per-packet state
-	fwPacket := &firewall.Packet{}
-
-	// Reset batch accumulation slices (reuse capacity)
-	*batchPackets = (*batchPackets)[:0]
-	*batchAddrs = (*batchAddrs)[:0]
-
-	// Process each packet in the batch
-	for i := 0; i < count; i++ {
-		packet := packets[i][:sizes[i]]
-		out := outs[i]
-
-		// Inline the consumeInsidePacket logic for better performance
-		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)
-			}
-			continue
-		}
-
-		// Ignore local broadcast packets
-		if f.dropLocalBroadcast {
-			if f.myBroadcastAddrsTable.Contains(fwPacket.RemoteAddr) {
-				continue
-			}
-		}
-
-		if f.myVpnAddrsTable.Contains(fwPacket.RemoteAddr) {
-			// Immediately forward packets from self to self.
-			if immediatelyForwardToSelf {
-				_, err := f.readers[q].Write(packet)
-				if err != nil {
-					f.l.WithError(err).Error("Failed to forward to tun")
-				}
-			}
-			continue
-		}
-
-		// Ignore multicast packets
-		if f.dropMulticast && fwPacket.RemoteAddr.IsMulticast() {
-			continue
-		}
-
-		hostinfo, ready := f.getOrHandshakeConsiderRouting(fwPacket, func(hh *HandshakeHostInfo) {
-			hh.cachePacket(f.l, header.Message, 0, packet, f.sendMessageNow, f.cachedPacketMetrics)
-		})
-
-		if hostinfo == nil {
-			f.rejectInside(packet, out, q)
-			if f.l.Level >= logrus.DebugLevel {
-				f.l.WithField("vpnAddr", fwPacket.RemoteAddr).
-					WithField("fwPacket", fwPacket).
-					Debugln("dropping outbound packet, vpnAddr not in our vpn networks or in unsafe networks")
-			}
-			continue
-		}
-
-		if !ready {
-			continue
-		}
-
-		dropReason := f.firewall.Drop(*fwPacket, false, hostinfo, f.pki.GetCAPool(), localCache)
-		if dropReason != nil {
-			f.rejectInside(packet, out, q)
-			if f.l.Level >= logrus.DebugLevel {
-				hostinfo.logger(f.l).
-					WithField("fwPacket", fwPacket).
-					WithField("reason", dropReason).
-					Debugln("dropping outbound packet")
-			}
-			continue
-		}
-
-		// Encrypt and prepare packet for batch sending
-		ci := hostinfo.ConnectionState
-		if ci.eKey == nil {
-			continue
-		}
-
-		// Check if this needs relay - if so, send immediately and skip batching
-		useRelay := !hostinfo.remote.IsValid()
-		if useRelay {
-			// Handle relay sends individually (less common path)
-			f.sendNoMetrics(header.Message, 0, ci, hostinfo, netip.AddrPort{}, packet, nb, out, q)
-			continue
-		}
-
-		// Encrypt the packet for batch sending
-		if noiseutil.EncryptLockNeeded {
-			ci.writeLock.Lock()
-		}
-		c := ci.messageCounter.Add(1)
-		out = header.Encode(out, header.Version, header.Message, 0, hostinfo.remoteIndexId, c)
-		f.connectionManager.Out(hostinfo)
-
-		// Query lighthouse if needed
-		if hostinfo.lastRebindCount != f.rebindCount {
-			f.lightHouse.QueryServer(hostinfo.vpnAddrs[0])
-			hostinfo.lastRebindCount = f.rebindCount
-			if f.l.Level >= logrus.DebugLevel {
-				f.l.WithField("vpnAddrs", hostinfo.vpnAddrs).Debug("Lighthouse update triggered for punch due to rebind counter")
-			}
-		}
-
-		out, err = ci.eKey.EncryptDanger(out, out, packet, c, nb)
-		if noiseutil.EncryptLockNeeded {
-			ci.writeLock.Unlock()
-		}
-		if err != nil {
-			hostinfo.logger(f.l).WithError(err).
-				WithField("counter", c).
-				Error("Failed to encrypt outgoing packet")
-			continue
-		}
-
-		// Add to batch
-		*batchPackets = append(*batchPackets, out)
-		*batchAddrs = append(*batchAddrs, hostinfo.remote)
-	}
-
-	// Send all accumulated packets in one batch
-	if len(*batchPackets) > 0 {
-		batchSize := len(*batchPackets)
-		f.batchMetrics.udpWriteSize.Update(int64(batchSize))
-
-		n, err := f.writers[q].WriteMulti(*batchPackets, *batchAddrs)
-		if err != nil {
-			f.l.WithError(err).WithField("sent", n).WithField("total", batchSize).Error("Failed to send batch")
-		}
-	}
-}
-
 func (f *Interface) consumeInsidePacket(packet []byte, fwPacket *firewall.Packet, nb, out []byte, q int, localCache firewall.ConntrackCache) {
 	err := newPacket(packet, false, fwPacket)
 	if err != nil {

interface.go

@@ -4,6 +4,7 @@ import (
 	"context"
 	"errors"
 	"fmt"
+	"io"
 	"net/netip"
 	"os"
 	"runtime"
@@ -21,7 +22,6 @@ import (
 )
 
 const mtu = 9001
-const virtioNetHdrLen = overlay.VirtioNetHdrLen
 
 type InterfaceConfig struct {
 	HostMap            *HostMap
@@ -50,13 +50,6 @@ type InterfaceConfig struct {
 	l                     *logrus.Logger
 }
 
-type batchMetrics struct {
-	udpReadSize   metrics.Histogram
-	tunReadSize   metrics.Histogram
-	udpWriteSize  metrics.Histogram
-	tunWriteSize  metrics.Histogram
-}
-
 type Interface struct {
 	hostMap               *HostMap
 	outside               udp.Conn
@@ -93,12 +86,11 @@ type Interface struct {
 	conntrackCacheTimeout time.Duration
 
 	writers []udp.Conn
-	readers []overlay.BatchReadWriter
+	readers []io.ReadWriteCloser
 
 	metricHandshakes    metrics.Histogram
 	messageMetrics      *MessageMetrics
 	cachedPacketMetrics *cachedPacketMetrics
-	batchMetrics        *batchMetrics
 
 	l *logrus.Logger
 }
@@ -185,7 +177,7 @@ func NewInterface(ctx context.Context, c *InterfaceConfig) (*Interface, error) {
 		routines:              c.routines,
 		version:               c.version,
 		writers:               make([]udp.Conn, c.routines),
-		readers:               make([]overlay.BatchReadWriter, c.routines),
+		readers:               make([]io.ReadWriteCloser, c.routines),
 		myVpnNetworks:         cs.myVpnNetworks,
 		myVpnNetworksTable:    cs.myVpnNetworksTable,
 		myVpnAddrs:            cs.myVpnAddrs,
@@ -201,12 +193,6 @@ func NewInterface(ctx context.Context, c *InterfaceConfig) (*Interface, error) {
 			sent:    metrics.GetOrRegisterCounter("hostinfo.cached_packets.sent", nil),
 			dropped: metrics.GetOrRegisterCounter("hostinfo.cached_packets.dropped", nil),
 		},
-		batchMetrics: &batchMetrics{
-			udpReadSize:  metrics.GetOrRegisterHistogram("batch.udp_read_size", nil, metrics.NewUniformSample(1024)),
-			tunReadSize:  metrics.GetOrRegisterHistogram("batch.tun_read_size", nil, metrics.NewUniformSample(1024)),
-			udpWriteSize: metrics.GetOrRegisterHistogram("batch.udp_write_size", nil, metrics.NewUniformSample(1024)),
-			tunWriteSize: metrics.GetOrRegisterHistogram("batch.tun_write_size", nil, metrics.NewUniformSample(1024)),
-		},
 
 		l: c.l,
 	}
@@ -239,7 +225,7 @@ func (f *Interface) activate() {
 	metrics.GetOrRegisterGauge("routines", nil).Update(int64(f.routines))
 
 	// Prepare n tun queues
-	var reader overlay.BatchReadWriter = f.inside
+	var reader io.ReadWriteCloser = f.inside
 	for i := 0; i < f.routines; i++ {
 		if i > 0 {
 			reader, err = f.inside.NewMultiQueueReader()
@@ -280,69 +266,39 @@ func (f *Interface) listenOut(i int) {
 
 	ctCache := firewall.NewConntrackCacheTicker(f.conntrackCacheTimeout)
 	lhh := f.lightHouse.NewRequestHandler()
-
-	// Pre-allocate output buffers for batch processing
-	batchSize := li.BatchSize()
-	outs := make([][]byte, batchSize)
-	for idx := range outs {
-		// Allocate full buffer with virtio header space
-		outs[idx] = make([]byte, virtioNetHdrLen, virtioNetHdrLen+udp.MTU)
-	}
-
+	plaintext := make([]byte, udp.MTU)
 	h := &header.H{}
 	fwPacket := &firewall.Packet{}
-	nb := make([]byte, 12)
+	nb := make([]byte, 12, 12)
 
-	li.ListenOutBatch(func(addrs []netip.AddrPort, payloads [][]byte, count int) {
-		f.readOutsidePacketsBatch(addrs, payloads, count, outs[:count], nb, i, h, fwPacket, lhh, ctCache.Get(f.l))
+	li.ListenOut(func(fromUdpAddr netip.AddrPort, payload []byte) {
+		f.readOutsidePackets(fromUdpAddr, nil, plaintext[:0], payload, h, fwPacket, lhh, nb, i, ctCache.Get(f.l))
 	})
 }
 
-func (f *Interface) listenIn(reader overlay.BatchReadWriter, i int) {
+func (f *Interface) listenIn(reader io.ReadWriteCloser, i int) {
 	runtime.LockOSThread()
 
-	batchSize := reader.BatchSize()
-
-	// Allocate buffers for batch reading
-	bufs := make([][]byte, batchSize)
-	for idx := range bufs {
-		bufs[idx] = make([]byte, mtu)
-	}
-	sizes := make([]int, batchSize)
-
-	// Allocate output buffers for batch processing (one per packet)
-	// Each has virtio header headroom to avoid copies on write
-	outs := make([][]byte, batchSize)
-	for idx := range outs {
-		outBuf := make([]byte, virtioNetHdrLen+mtu)
-		outs[idx] = outBuf[virtioNetHdrLen:] // Slice starting after headroom
-	}
-
-	// Pre-allocate batch accumulation buffers for sending
-	batchPackets := make([][]byte, 0, batchSize)
-	batchAddrs := make([]netip.AddrPort, 0, batchSize)
-
-	// Pre-allocate nonce buffer (reused for all encryptions)
-	nb := make([]byte, 12)
+	packet := make([]byte, mtu)
+	out := make([]byte, mtu)
+	fwPacket := &firewall.Packet{}
+	nb := make([]byte, 12, 12)
 
 	conntrackCache := firewall.NewConntrackCacheTicker(f.conntrackCacheTimeout)
 
 	for {
-		n, err := reader.BatchRead(bufs, sizes)
+		n, err := reader.Read(packet)
 		if err != nil {
 			if errors.Is(err, os.ErrClosed) && f.closed.Load() {
 				return
 			}
 
-			f.l.WithError(err).Error("Error while batch reading outbound packets")
+			f.l.WithError(err).Error("Error while reading outbound packet")
 			// This only seems to happen when something fatal happens to the fd, so exit.
 			os.Exit(2)
 		}
 
-		f.batchMetrics.tunReadSize.Update(int64(n))
-
-		// Process all packets in the batch at once
-		f.consumeInsidePackets(bufs, sizes, n, outs, nb, i, conntrackCache.Get(f.l), &batchPackets, &batchAddrs)
+		f.consumeInsidePacket(packet[:n], fwPacket, nb, out, i, conntrackCache.Get(f.l))
 	}
 }
 

lighthouse.go

@@ -1337,19 +1337,12 @@ func (lhh *LightHouseHandler) handleHostPunchNotification(n *NebulaMeta, fromVpn
 		}
 	}
 
-	remoteAllowList := lhh.lh.GetRemoteAllowList()
 	for _, a := range n.Details.V4AddrPorts {
-		b := protoV4AddrPortToNetAddrPort(a)
-		if remoteAllowList.Allow(detailsVpnAddr, b.Addr()) {
-			punch(b, detailsVpnAddr)
-		}
+		punch(protoV4AddrPortToNetAddrPort(a), detailsVpnAddr)
 	}
 
 	for _, a := range n.Details.V6AddrPorts {
-		b := protoV6AddrPortToNetAddrPort(a)
-		if remoteAllowList.Allow(detailsVpnAddr, b.Addr()) {
-			punch(b, detailsVpnAddr)
-		}
+		punch(protoV6AddrPortToNetAddrPort(a), detailsVpnAddr)
 	}
 
 	// This sends a nebula test packet to the host trying to contact us. In the case

main.go

@@ -75,8 +75,7 @@ func Main(c *config.C, configTest bool, buildVersion string, logger *logrus.Logg
 	if c.GetBool("sshd.enabled", false) {
 		sshStart, err = configSSH(l, ssh, c)
 		if err != nil {
-			l.WithError(err).Warn("Failed to configure sshd, ssh debugging will not be available")
-			sshStart = nil
+			return nil, util.ContextualizeIfNeeded("Error while configuring the sshd", err)
 		}
 	}
 
@@ -165,7 +164,7 @@ func Main(c *config.C, configTest bool, buildVersion string, logger *logrus.Logg
 
 		for i := 0; i < routines; i++ {
 			l.Infof("listening on %v", netip.AddrPortFrom(listenHost, uint16(port)))
-			udpServer, err := udp.NewListener(l, listenHost, port, routines > 1, c.GetInt("listen.batch", 128))
+			udpServer, err := udp.NewListener(l, listenHost, port, routines > 1, c.GetInt("listen.batch", 64))
 			if err != nil {
 				return nil, util.NewContextualError("Failed to open udp listener", m{"queue": i}, err)
 			}

outside.go

@@ -95,7 +95,8 @@ func (f *Interface) readOutsidePackets(ip netip.AddrPort, via *ViaSender, out []
 			switch relay.Type {
 			case TerminalType:
 				// If I am the target of this relay, process the unwrapped packet
-				f.readOutsidePackets(netip.AddrPort{}, &ViaSender{relayHI: hostinfo, remoteIdx: relay.RemoteIndex, relay: relay}, out[:virtioNetHdrLen], signedPayload, h, fwPacket, lhf, nb, q, localCache)
+				// From this recursive point, all these variables are 'burned'. We shouldn't rely on them again.
+				f.readOutsidePackets(netip.AddrPort{}, &ViaSender{relayHI: hostinfo, remoteIdx: relay.RemoteIndex, relay: relay}, out[:0], signedPayload, h, fwPacket, lhf, nb, q, localCache)
 				return
 			case ForwardingType:
 				// Find the target HostInfo relay object
@@ -137,7 +138,7 @@ func (f *Interface) readOutsidePackets(ip netip.AddrPort, via *ViaSender, out []
 			return
 		}
 
-		lhf.HandleRequest(ip, hostinfo.vpnAddrs, d[virtioNetHdrLen:], f)
+		lhf.HandleRequest(ip, hostinfo.vpnAddrs, d, f)
 
 		// Fallthrough to the bottom to record incoming traffic
 
@@ -159,7 +160,7 @@ func (f *Interface) readOutsidePackets(ip netip.AddrPort, via *ViaSender, out []
 			// This testRequest might be from TryPromoteBest, so we should roam
 			// to the new IP address before responding
 			f.handleHostRoaming(hostinfo, ip)
-			f.send(header.Test, header.TestReply, ci, hostinfo, d[virtioNetHdrLen:], nb, out)
+			f.send(header.Test, header.TestReply, ci, hostinfo, d, nb, out)
 		}
 
 		// Fallthrough to the bottom to record incoming traffic
@@ -202,7 +203,7 @@ func (f *Interface) readOutsidePackets(ip netip.AddrPort, via *ViaSender, out []
 			return
 		}
 
-		f.relayManager.HandleControlMsg(hostinfo, d[virtioNetHdrLen:], f)
+		f.relayManager.HandleControlMsg(hostinfo, d, f)
 
 	default:
 		f.messageMetrics.Rx(h.Type, h.Subtype, 1)
@@ -473,11 +474,9 @@ func (f *Interface) decryptToTun(hostinfo *HostInfo, messageCounter uint64, out
 		return false
 	}
 
-	packetData := out[virtioNetHdrLen:]
-
-	err = newPacket(packetData, true, fwPacket)
+	err = newPacket(out, true, fwPacket)
 	if err != nil {
-		hostinfo.logger(f.l).WithError(err).WithField("packet", packetData).
+		hostinfo.logger(f.l).WithError(err).WithField("packet", out).
 			Warnf("Error while validating inbound packet")
 		return false
 	}
@@ -492,7 +491,7 @@ func (f *Interface) decryptToTun(hostinfo *HostInfo, messageCounter uint64, out
 	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
-		f.rejectOutside(packetData, hostinfo.ConnectionState, hostinfo, nb, packet, q)
+		f.rejectOutside(out, hostinfo.ConnectionState, hostinfo, nb, packet, q)
 		if f.l.Level >= logrus.DebugLevel {
 			hostinfo.logger(f.l).WithField("fwPacket", fwPacket).
 				WithField("reason", dropReason).
@@ -549,108 +548,3 @@ func (f *Interface) handleRecvError(addr netip.AddrPort, h *header.H) {
 	// We also delete it from pending hostmap to allow for fast reconnect.
 	f.handshakeManager.DeleteHostInfo(hostinfo)
 }
-
-// readOutsidePacketsBatch processes multiple packets received from UDP in a batch
-// and writes all successfully decrypted packets to TUN in a single operation
-func (f *Interface) readOutsidePacketsBatch(addrs []netip.AddrPort, payloads [][]byte, count int, outs [][]byte, nb []byte, q int, h *header.H, fwPacket *firewall.Packet, lhf *LightHouseHandler, localCache firewall.ConntrackCache) {
-	// Pre-allocate slice for accumulating successful decryptions
-	tunPackets := make([][]byte, 0, count)
-
-	for i := 0; i < count; i++ {
-		payload := payloads[i]
-		addr := addrs[i]
-		out := outs[i]
-
-		// Parse header
-		err := h.Parse(payload)
-		if err != nil {
-			if len(payload) > 1 {
-				f.l.WithField("packet", payload).Infof("Error while parsing inbound packet from %s: %s", addr, err)
-			}
-			continue
-		}
-
-		if addr.IsValid() {
-			if f.myVpnNetworksTable.Contains(addr.Addr()) {
-				if f.l.Level >= logrus.DebugLevel {
-					f.l.WithField("udpAddr", addr).Debug("Refusing to process double encrypted packet")
-				}
-				continue
-			}
-		}
-
-		var hostinfo *HostInfo
-		if h.Type == header.Message && h.Subtype == header.MessageRelay {
-			hostinfo = f.hostMap.QueryRelayIndex(h.RemoteIndex)
-		} else {
-			hostinfo = f.hostMap.QueryIndex(h.RemoteIndex)
-		}
-
-		var ci *ConnectionState
-		if hostinfo != nil {
-			ci = hostinfo.ConnectionState
-		}
-
-		switch h.Type {
-		case header.Message:
-			if !f.handleEncrypted(ci, addr, h) {
-				continue
-			}
-
-			switch h.Subtype {
-			case header.MessageNone:
-				// Decrypt packet
-				out, err = hostinfo.ConnectionState.dKey.DecryptDanger(out, payload[:header.Len], payload[header.Len:], h.MessageCounter, nb)
-				if err != nil {
-					hostinfo.logger(f.l).WithError(err).Error("Failed to decrypt packet")
-					continue
-				}
-
-				packetData := out[virtioNetHdrLen:]
-
-				err = newPacket(packetData, true, fwPacket)
-				if err != nil {
-					hostinfo.logger(f.l).WithError(err).WithField("packet", packetData).Warnf("Error while validating inbound packet")
-					continue
-				}
-
-				if !hostinfo.ConnectionState.window.Update(f.l, h.MessageCounter) {
-					hostinfo.logger(f.l).WithField("fwPacket", fwPacket).Debugln("dropping out of window packet")
-					continue
-				}
-
-				dropReason := f.firewall.Drop(*fwPacket, true, hostinfo, f.pki.GetCAPool(), localCache)
-				if dropReason != nil {
-					f.rejectOutside(packetData, hostinfo.ConnectionState, hostinfo, nb, payload, q)
-					if f.l.Level >= logrus.DebugLevel {
-						hostinfo.logger(f.l).WithField("fwPacket", fwPacket).WithField("reason", dropReason).Debugln("dropping inbound packet")
-					}
-					continue
-				}
-
-				f.connectionManager.In(hostinfo)
-				// Add to batch for TUN write
-				tunPackets = append(tunPackets, out)
-
-			case header.MessageRelay:
-				// Skip relay packets in batch mode for now (less common path)
-				f.readOutsidePackets(addr, nil, out, payload, h, fwPacket, lhf, nb, q, localCache)
-
-			default:
-				hostinfo.logger(f.l).Debugf("unexpected message subtype %d", h.Subtype)
-			}
-
-		default:
-			// Handle non-Message types using single-packet path
-			f.readOutsidePackets(addr, nil, out, payload, h, fwPacket, lhf, nb, q, localCache)
-		}
-	}
-
-	if len(tunPackets) > 0 {
-		n, err := f.readers[q].WriteBatch(tunPackets, virtioNetHdrLen)
-		if err != nil {
-			f.l.WithError(err).WithField("sent", n).WithField("total", len(tunPackets)).Error("Failed to batch write to tun")
-		}
-		f.batchMetrics.tunWriteSize.Update(int64(len(tunPackets)))
-	}
-}

overlay/device.go

@@ -7,25 +7,11 @@ import (
 	"github.com/slackhq/nebula/routing"
 )
 
-// BatchReadWriter extends io.ReadWriteCloser with batch I/O operations
-type BatchReadWriter interface {
-	io.ReadWriteCloser
-
-	// BatchRead reads multiple packets at once
-	BatchRead(bufs [][]byte, sizes []int) (int, error)
-
-	// WriteBatch writes multiple packets at once
-	WriteBatch(bufs [][]byte, offset int) (int, error)
-
-	// BatchSize returns the optimal batch size for this device
-	BatchSize() int
-}
-
 type Device interface {
-	BatchReadWriter
+	io.ReadWriteCloser
 	Activate() error
 	Networks() []netip.Prefix
 	Name() string
 	RoutesFor(netip.Addr) routing.Gateways
-	NewMultiQueueReader() (BatchReadWriter, error)
+	NewMultiQueueReader() (io.ReadWriteCloser, error)
 }

overlay/tun.go

@@ -11,7 +11,6 @@ import (
 )
 
 const DefaultMTU = 1300
-const VirtioNetHdrLen = 10 // Size of virtio_net_hdr structure
 
 // TODO: We may be able to remove routines
 type DeviceFactory func(c *config.C, l *logrus.Logger, vpnNetworks []netip.Prefix, routines int) (Device, error)

overlay/tun_android.go

@@ -95,29 +95,6 @@ func (t *tun) Name() string {
 	return "android"
 }
 
-func (t *tun) NewMultiQueueReader() (BatchReadWriter, error) {
+func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
 	return nil, fmt.Errorf("TODO: multiqueue not implemented for android")
 }
-
-func (t *tun) BatchRead(bufs [][]byte, sizes []int) (int, error) {
-	n, err := t.Read(bufs[0])
-	if err != nil {
-		return 0, err
-	}
-	sizes[0] = n
-	return 1, nil
-}
-
-func (t *tun) WriteBatch(bufs [][]byte, offset int) (int, error) {
-	for i, buf := range bufs {
-		_, err := t.Write(buf[offset:])
-		if err != nil {
-			return i, err
-		}
-	}
-	return len(bufs), nil
-}
-
-func (t *tun) BatchSize() int {
-	return 1
-}

overlay/tun_darwin.go

@@ -549,32 +549,6 @@ func (t *tun) Name() string {
 	return t.Device
 }
 
-func (t *tun) NewMultiQueueReader() (BatchReadWriter, error) {
+func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
 	return nil, fmt.Errorf("TODO: multiqueue not implemented for darwin")
 }
-
-// BatchRead reads a single packet (batch size 1 for non-Linux platforms)
-func (t *tun) BatchRead(bufs [][]byte, sizes []int) (int, error) {
-	n, err := t.Read(bufs[0])
-	if err != nil {
-		return 0, err
-	}
-	sizes[0] = n
-	return 1, nil
-}
-
-// WriteBatch writes packets individually (no batching for non-Linux platforms)
-func (t *tun) WriteBatch(bufs [][]byte, offset int) (int, error) {
-	for i, buf := range bufs {
-		_, err := t.Write(buf[offset:])
-		if err != nil {
-			return i, err
-		}
-	}
-	return len(bufs), nil
-}
-
-// BatchSize returns 1 for non-Linux platforms (no batching)
-func (t *tun) BatchSize() int {
-	return 1
-}

overlay/tun_disabled.go

@@ -105,36 +105,10 @@ func (t *disabledTun) Write(b []byte) (int, error) {
 	return len(b), nil
 }
 
-func (t *disabledTun) NewMultiQueueReader() (BatchReadWriter, error) {
+func (t *disabledTun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
 	return t, nil
 }
 
-// BatchRead reads a single packet (batch size 1 for disabled tun)
-func (t *disabledTun) BatchRead(bufs [][]byte, sizes []int) (int, error) {
-	n, err := t.Read(bufs[0])
-	if err != nil {
-		return 0, err
-	}
-	sizes[0] = n
-	return 1, nil
-}
-
-// WriteBatch writes packets individually (no batching for disabled tun)
-func (t *disabledTun) WriteBatch(bufs [][]byte, offset int) (int, error) {
-	for i, buf := range bufs {
-		_, err := t.Write(buf[offset:])
-		if err != nil {
-			return i, err
-		}
-	}
-	return len(bufs), nil
-}
-
-// BatchSize returns 1 for disabled tun (no batching)
-func (t *disabledTun) BatchSize() int {
-	return 1
-}
-
 func (t *disabledTun) Close() error {
 	if t.read != nil {
 		close(t.read)

overlay/tun_freebsd.go

@@ -450,36 +450,10 @@ func (t *tun) Name() string {
 	return t.Device
 }
 
-func (t *tun) NewMultiQueueReader() (BatchReadWriter, error) {
+func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
 	return nil, fmt.Errorf("TODO: multiqueue not implemented for freebsd")
 }
 
-// BatchRead reads a single packet (batch size 1 for FreeBSD)
-func (t *tun) BatchRead(bufs [][]byte, sizes []int) (int, error) {
-	n, err := t.Read(bufs[0])
-	if err != nil {
-		return 0, err
-	}
-	sizes[0] = n
-	return 1, nil
-}
-
-// WriteBatch writes packets individually (no batching for FreeBSD)
-func (t *tun) WriteBatch(bufs [][]byte, offset int) (int, error) {
-	for i, buf := range bufs {
-		_, err := t.Write(buf[offset:])
-		if err != nil {
-			return i, err
-		}
-	}
-	return len(bufs), nil
-}
-
-// BatchSize returns 1 for FreeBSD (no batching)
-func (t *tun) BatchSize() int {
-	return 1
-}
-
 func (t *tun) addRoutes(logErrors bool) error {
 	routes := *t.Routes.Load()
 	for _, r := range routes {

overlay/tun_ios.go

@@ -151,29 +151,6 @@ func (t *tun) Name() string {
 	return "iOS"
 }
 
-func (t *tun) NewMultiQueueReader() (BatchReadWriter, error) {
+func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
 	return nil, fmt.Errorf("TODO: multiqueue not implemented for ios")
 }
-
-func (t *tun) BatchRead(bufs [][]byte, sizes []int) (int, error) {
-	n, err := t.Read(bufs[0])
-	if err != nil {
-		return 0, err
-	}
-	sizes[0] = n
-	return 1, nil
-}
-
-func (t *tun) WriteBatch(bufs [][]byte, offset int) (int, error) {
-	for i, buf := range bufs {
-		_, err := t.Write(buf[offset:])
-		if err != nil {
-			return i, err
-		}
-	}
-	return len(bufs), nil
-}
-
-func (t *tun) BatchSize() int {
-	return 1
-}

overlay/tun_linux.go

@@ -9,6 +9,7 @@ import (
 	"net"
 	"net/netip"
 	"os"
+	"strings"
 	"sync/atomic"
 	"time"
 	"unsafe"
@@ -20,12 +21,10 @@ import (
 	"github.com/slackhq/nebula/util"
 	"github.com/vishvananda/netlink"
 	"golang.org/x/sys/unix"
-	wgtun "golang.zx2c4.com/wireguard/tun"
 )
 
 type tun struct {
 	io.ReadWriteCloser
-	wgDevice    wgtun.Device
 	fd          int
 	Device      string
 	vpnNetworks []netip.Prefix
@@ -66,154 +65,59 @@ type ifreqQLEN struct {
 	pad   [8]byte
 }
 
-// wgDeviceWrapper wraps a wireguard Device to implement io.ReadWriteCloser
-// This allows multiqueue readers to use the same wireguard Device batching as the main device
-type wgDeviceWrapper struct {
-	dev wgtun.Device
-	buf []byte // Reusable buffer for single packet reads
-}
-
-func (w *wgDeviceWrapper) Read(b []byte) (int, error) {
-	// Use wireguard Device's batch API for single packet
-	bufs := [][]byte{b}
-	sizes := make([]int, 1)
-	n, err := w.dev.Read(bufs, sizes, 0)
-	if err != nil {
-		return 0, err
-	}
-	if n == 0 {
-		return 0, io.EOF
-	}
-	return sizes[0], nil
-}
-
-func (w *wgDeviceWrapper) Write(b []byte) (int, error) {
-	// Buffer b should have virtio header space (10 bytes) at the beginning
-	// The decrypted packet data starts at offset 10
-	// Pass the full buffer to WireGuard with offset=virtioNetHdrLen
-	bufs := [][]byte{b}
-	n, err := w.dev.Write(bufs, VirtioNetHdrLen)
-	if err != nil {
-		return 0, err
-	}
-	if n == 0 {
-		return 0, io.ErrShortWrite
-	}
-	return len(b), nil
-}
-
-func (w *wgDeviceWrapper) WriteBatch(bufs [][]byte, offset int) (int, error) {
-	// Pass all buffers to WireGuard's batch write
-	return w.dev.Write(bufs, offset)
-}
-
-func (w *wgDeviceWrapper) Close() error {
-	return w.dev.Close()
-}
-
-// BatchRead implements batching for multiqueue readers
-func (w *wgDeviceWrapper) BatchRead(bufs [][]byte, sizes []int) (int, error) {
-	// The zero here is offset.
-	return w.dev.Read(bufs, sizes, 0)
-}
-
-// BatchSize returns the optimal batch size
-func (w *wgDeviceWrapper) BatchSize() int {
-	return w.dev.BatchSize()
-}
-
 func newTunFromFd(c *config.C, l *logrus.Logger, deviceFd int, vpnNetworks []netip.Prefix) (*tun, error) {
-	wgDev, name, err := wgtun.CreateUnmonitoredTUNFromFD(deviceFd)
-	if err != nil {
-		return nil, fmt.Errorf("failed to create TUN from FD: %w", err)
-	}
+	file := os.NewFile(uintptr(deviceFd), "/dev/net/tun")
 
-	file := wgDev.File()
 	t, err := newTunGeneric(c, l, file, vpnNetworks)
 	if err != nil {
-		_ = wgDev.Close()
 		return nil, err
 	}
 
-	t.wgDevice = wgDev
-	t.Device = name
+	t.Device = "tun0"
 
 	return t, nil
 }
 
 func newTun(c *config.C, l *logrus.Logger, vpnNetworks []netip.Prefix, multiqueue bool) (*tun, error) {
-	// Check if /dev/net/tun exists, create if needed (for docker containers)
-	if _, err := os.Stat("/dev/net/tun"); os.IsNotExist(err) {
-		if err := os.MkdirAll("/dev/net", 0755); err != nil {
-			return nil, 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 nil, fmt.Errorf("failed to create /dev/net/tun: %w", err)
-		}
-	}
-
-	devName := c.GetString("tun.dev", "")
-	mtu := c.GetInt("tun.mtu", DefaultMTU)
-
-	// Create TUN device manually to support multiqueue
 	fd, err := unix.Open("/dev/net/tun", os.O_RDWR, 0)
 	if err != nil {
+		// If /dev/net/tun doesn't exist, try to create it (will happen in docker)
+		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 err != nil {
+				return nil, 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)
+			}
+		} else {
 			return nil, err
 		}
+	}
 
 	var req ifReq
-	req.Flags = uint16(unix.IFF_TUN | unix.IFF_NO_PI | unix.IFF_VNET_HDR)
+	req.Flags = uint16(unix.IFF_TUN | unix.IFF_NO_PI)
 	if multiqueue {
 		req.Flags |= unix.IFF_MULTI_QUEUE
 	}
-	copy(req.Name[:], devName)
+	copy(req.Name[:], c.GetString("tun.dev", ""))
 	if err = ioctl(uintptr(fd), uintptr(unix.TUNSETIFF), uintptr(unsafe.Pointer(&req))); err != nil {
-		unix.Close(fd)
 		return nil, err
 	}
-
-	// Set nonblocking
-	if err = unix.SetNonblock(fd, true); err != nil {
-		unix.Close(fd)
-		return nil, err
-	}
-
-	// Enable TCP and UDP offload (TSO/GRO) for performance
-	// This allows the kernel to handle segmentation/coalescing
-	const (
-		tunTCPOffloads = unix.TUN_F_CSUM | unix.TUN_F_TSO4 | unix.TUN_F_TSO6
-		tunUDPOffloads = unix.TUN_F_USO4 | unix.TUN_F_USO6
-	)
-	offloads := tunTCPOffloads | tunUDPOffloads
-	if err = unix.IoctlSetInt(fd, unix.TUNSETOFFLOAD, offloads); err != nil {
-		// Log warning but don't fail - offload is optional
-		l.WithError(err).Warn("Failed to enable TUN offload (TSO/GRO), performance may be reduced")
-	}
+	name := strings.Trim(string(req.Name[:]), "\x00")
 
 	file := os.NewFile(uintptr(fd), "/dev/net/tun")
-
-	// Create wireguard device from file descriptor
-	wgDev, err := wgtun.CreateTUNFromFile(file, mtu)
-	if err != nil {
-		file.Close()
-		return nil, fmt.Errorf("failed to create TUN from file: %w", err)
-	}
-
-	name, err := wgDev.Name()
-	if err != nil {
-		_ = wgDev.Close()
-		return nil, fmt.Errorf("failed to get TUN device name: %w", err)
-	}
-
-	// file is now owned by wgDev, get a new reference
-	file = wgDev.File()
 	t, err := newTunGeneric(c, l, file, vpnNetworks)
 	if err != nil {
-		_ = wgDev.Close()
 		return nil, err
 	}
 
-	t.wgDevice = wgDev
 	t.Device = name
 
 	return t, nil
@@ -312,44 +216,22 @@ func (t *tun) reload(c *config.C, initial bool) error {
 	return nil
 }
 
-func (t *tun) NewMultiQueueReader() (BatchReadWriter, error) {
+func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
 	fd, err := unix.Open("/dev/net/tun", os.O_RDWR, 0)
 	if err != nil {
 		return nil, err
 	}
 
 	var req ifReq
-	// MUST match the flags used in newTun - includes IFF_VNET_HDR
-	req.Flags = uint16(unix.IFF_TUN | unix.IFF_NO_PI | unix.IFF_VNET_HDR | unix.IFF_MULTI_QUEUE)
+	req.Flags = uint16(unix.IFF_TUN | unix.IFF_NO_PI | unix.IFF_MULTI_QUEUE)
 	copy(req.Name[:], t.Device)
 	if err = ioctl(uintptr(fd), uintptr(unix.TUNSETIFF), uintptr(unsafe.Pointer(&req))); err != nil {
-		unix.Close(fd)
 		return nil, err
 	}
 
-	// Set nonblocking mode - CRITICAL for proper netpoller integration
-	if err = unix.SetNonblock(fd, true); err != nil {
-		unix.Close(fd)
-		return nil, err
-	}
-
-	// Get MTU from main device
-	mtu := t.MaxMTU
-	if mtu == 0 {
-		mtu = DefaultMTU
-	}
-
 	file := os.NewFile(uintptr(fd), "/dev/net/tun")
 
-	// Create wireguard Device from the file descriptor (just like the main device)
-	wgDev, err := wgtun.CreateTUNFromFile(file, mtu)
-	if err != nil {
-		file.Close()
-		return nil, fmt.Errorf("failed to create multiqueue TUN device: %w", err)
-	}
-
-	// Return a wrapper that uses the wireguard Device for all I/O
-	return &wgDeviceWrapper{dev: wgDev}, nil
+	return file, nil
 }
 
 func (t *tun) RoutesFor(ip netip.Addr) routing.Gateways {
@@ -357,68 +239,7 @@ func (t *tun) RoutesFor(ip netip.Addr) routing.Gateways {
 	return r
 }
 
-func (t *tun) Read(b []byte) (int, error) {
-	if t.wgDevice != nil {
-		// Use wireguard device which handles virtio headers internally
-		bufs := [][]byte{b}
-		sizes := make([]int, 1)
-		n, err := t.wgDevice.Read(bufs, sizes, 0)
-		if err != nil {
-			return 0, err
-		}
-		if n == 0 {
-			return 0, io.EOF
-		}
-		return sizes[0], nil
-	}
-
-	// Fallback: direct read from file (shouldn't happen in normal operation)
-	return t.ReadWriteCloser.Read(b)
-}
-
-// BatchRead reads multiple packets at once for improved performance
-// bufs: slice of buffers to read into
-// sizes: slice that will be filled with packet sizes
-// Returns number of packets read
-func (t *tun) BatchRead(bufs [][]byte, sizes []int) (int, error) {
-	if t.wgDevice != nil {
-		return t.wgDevice.Read(bufs, sizes, 0)
-	}
-
-	// Fallback: single packet read
-	n, err := t.ReadWriteCloser.Read(bufs[0])
-	if err != nil {
-		return 0, err
-	}
-	sizes[0] = n
-	return 1, nil
-}
-
-// BatchSize returns the optimal number of packets to read/write in a batch
-func (t *tun) BatchSize() int {
-	if t.wgDevice != nil {
-		return t.wgDevice.BatchSize()
-	}
-	return 1
-}
-
 func (t *tun) Write(b []byte) (int, error) {
-	if t.wgDevice != nil {
-		// Buffer b should have virtio header space (10 bytes) at the beginning
-		// The decrypted packet data starts at offset 10
-		// Pass the full buffer to WireGuard with offset=virtioNetHdrLen
-		bufs := [][]byte{b}
-		n, err := t.wgDevice.Write(bufs, VirtioNetHdrLen)
-		if err != nil {
-			return 0, err
-		}
-		if n == 0 {
-			return 0, io.ErrShortWrite
-		}
-		return len(b), nil
-	}
-
-	// Fallback: direct write (shouldn't happen in normal operation)
 	var nn int
 	maximum := len(b)
 
@@ -441,22 +262,6 @@ func (t *tun) Write(b []byte) (int, error) {
 	}
 }
 
-// WriteBatch writes multiple packets to the TUN device in a single syscall
-func (t *tun) WriteBatch(bufs [][]byte, offset int) (int, error) {
-	if t.wgDevice != nil {
-		return t.wgDevice.Write(bufs, offset)
-	}
-
-	// Fallback: write individually (shouldn't happen in normal operation)
-	for i, buf := range bufs {
-		_, err := t.Write(buf)
-		if err != nil {
-			return i, err
-		}
-	}
-	return len(bufs), nil
-}
-
 func (t *tun) deviceBytes() (o [16]byte) {
 	for i, c := range t.Device {
 		o[i] = byte(c)
@@ -869,10 +674,6 @@ func (t *tun) Close() error {
 		close(t.routeChan)
 	}
 
-	if t.wgDevice != nil {
-		_ = t.wgDevice.Close()
-	}
-
 	if t.ReadWriteCloser != nil {
 		_ = t.ReadWriteCloser.Close()
 	}

overlay/tun_netbsd.go

@@ -390,33 +390,10 @@ func (t *tun) Name() string {
 	return t.Device
 }
 
-func (t *tun) NewMultiQueueReader() (BatchReadWriter, error) {
+func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
 	return nil, fmt.Errorf("TODO: multiqueue not implemented for netbsd")
 }
 
-func (t *tun) BatchRead(bufs [][]byte, sizes []int) (int, error) {
-	n, err := t.Read(bufs[0])
-	if err != nil {
-		return 0, err
-	}
-	sizes[0] = n
-	return 1, nil
-}
-
-func (t *tun) WriteBatch(bufs [][]byte, offset int) (int, error) {
-	for i, buf := range bufs {
-		_, err := t.Write(buf[offset:])
-		if err != nil {
-			return i, err
-		}
-	}
-	return len(bufs), nil
-}
-
-func (t *tun) BatchSize() int {
-	return 1
-}
-
 func (t *tun) addRoutes(logErrors bool) error {
 	routes := *t.Routes.Load()
 

overlay/tun_openbsd.go

@@ -310,33 +310,10 @@ func (t *tun) Name() string {
 	return t.Device
 }
 
-func (t *tun) NewMultiQueueReader() (BatchReadWriter, error) {
+func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
 	return nil, fmt.Errorf("TODO: multiqueue not implemented for openbsd")
 }
 
-func (t *tun) BatchRead(bufs [][]byte, sizes []int) (int, error) {
-	n, err := t.Read(bufs[0])
-	if err != nil {
-		return 0, err
-	}
-	sizes[0] = n
-	return 1, nil
-}
-
-func (t *tun) WriteBatch(bufs [][]byte, offset int) (int, error) {
-	for i, buf := range bufs {
-		_, err := t.Write(buf[offset:])
-		if err != nil {
-			return i, err
-		}
-	}
-	return len(bufs), nil
-}
-
-func (t *tun) BatchSize() int {
-	return 1
-}
-
 func (t *tun) addRoutes(logErrors bool) error {
 	routes := *t.Routes.Load()
 

overlay/tun_tester.go

@@ -132,29 +132,6 @@ func (t *TestTun) Read(b []byte) (int, error) {
 	return len(p), nil
 }
 
-func (t *TestTun) NewMultiQueueReader() (BatchReadWriter, error) {
+func (t *TestTun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
 	return nil, fmt.Errorf("TODO: multiqueue not implemented")
 }
-
-func (t *TestTun) BatchRead(bufs [][]byte, sizes []int) (int, error) {
-	n, err := t.Read(bufs[0])
-	if err != nil {
-		return 0, err
-	}
-	sizes[0] = n
-	return 1, nil
-}
-
-func (t *TestTun) WriteBatch(bufs [][]byte, offset int) (int, error) {
-	for i, buf := range bufs {
-		_, err := t.Write(buf[offset:])
-		if err != nil {
-			return i, err
-		}
-	}
-	return len(bufs), nil
-}
-
-func (t *TestTun) BatchSize() int {
-	return 1
-}

overlay/tun_windows.go

@@ -6,6 +6,7 @@ package overlay
 import (
 	"crypto"
 	"fmt"
+	"io"
 	"net/netip"
 	"os"
 	"path/filepath"
@@ -233,36 +234,10 @@ func (t *winTun) Write(b []byte) (int, error) {
 	return t.tun.Write(b, 0)
 }
 
-func (t *winTun) NewMultiQueueReader() (BatchReadWriter, error) {
+func (t *winTun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
 	return nil, fmt.Errorf("TODO: multiqueue not implemented for windows")
 }
 
-// BatchRead reads a single packet (batch size 1 for Windows)
-func (t *winTun) BatchRead(bufs [][]byte, sizes []int) (int, error) {
-	n, err := t.Read(bufs[0])
-	if err != nil {
-		return 0, err
-	}
-	sizes[0] = n
-	return 1, nil
-}
-
-// WriteBatch writes packets individually (no batching for Windows)
-func (t *winTun) WriteBatch(bufs [][]byte, offset int) (int, error) {
-	for i, buf := range bufs {
-		_, err := t.Write(buf[offset:])
-		if err != nil {
-			return i, err
-		}
-	}
-	return len(bufs), nil
-}
-
-// BatchSize returns 1 for Windows (no batching)
-func (t *winTun) BatchSize() int {
-	return 1
-}
-
 func (t *winTun) Close() error {
 	// It seems that the Windows networking stack doesn't like it when we destroy interfaces that have active routes,
 	// so to be certain, just remove everything before destroying.

overlay/user.go

@@ -46,36 +46,10 @@ func (d *UserDevice) RoutesFor(ip netip.Addr) routing.Gateways {
 	return routing.Gateways{routing.NewGateway(ip, 1)}
 }
 
-func (d *UserDevice) NewMultiQueueReader() (BatchReadWriter, error) {
+func (d *UserDevice) NewMultiQueueReader() (io.ReadWriteCloser, error) {
 	return d, nil
 }
 
-// BatchRead reads a single packet (batch size 1 for UserDevice)
-func (d *UserDevice) BatchRead(bufs [][]byte, sizes []int) (int, error) {
-	n, err := d.Read(bufs[0])
-	if err != nil {
-		return 0, err
-	}
-	sizes[0] = n
-	return 1, nil
-}
-
-// WriteBatch writes packets individually (no batching for UserDevice)
-func (d *UserDevice) WriteBatch(bufs [][]byte, offset int) (int, error) {
-	for i, buf := range bufs {
-		_, err := d.Write(buf[offset:])
-		if err != nil {
-			return i, err
-		}
-	}
-	return len(bufs), nil
-}
-
-// BatchSize returns 1 for UserDevice (no batching)
-func (d *UserDevice) BatchSize() int {
-	return 1
-}
-
 func (d *UserDevice) Pipe() (*io.PipeReader, *io.PipeWriter) {
 	return d.inboundReader, d.outboundWriter
 }

pki.go

@@ -100,36 +100,41 @@ func (p *PKI) reloadCerts(c *config.C, initial bool) *util.ContextualError {
 		currentState := p.cs.Load()
 		if newState.v1Cert != nil {
 			if currentState.v1Cert == nil {
-				//adding certs is fine, actually. Networks-in-common confirmed in newCertState().
-			} else {
+				return util.NewContextualError("v1 certificate was added, restart required", nil, err)
+			}
+
 			// did IP in cert change? if so, don't set
 			if !slices.Equal(currentState.v1Cert.Networks(), newState.v1Cert.Networks()) {
 				return util.NewContextualError(
 					"Networks in new cert was different from old",
-						m{"new_networks": newState.v1Cert.Networks(), "old_networks": currentState.v1Cert.Networks(), "cert_version": cert.Version1},
+					m{"new_networks": newState.v1Cert.Networks(), "old_networks": currentState.v1Cert.Networks()},
 					nil,
 				)
 			}
 
 			if currentState.v1Cert.Curve() != newState.v1Cert.Curve() {
 				return util.NewContextualError(
-						"Curve in new v1 cert was different from old",
-						m{"new_curve": newState.v1Cert.Curve(), "old_curve": currentState.v1Cert.Curve(), "cert_version": cert.Version1},
+					"Curve in new cert was different from old",
+					m{"new_curve": newState.v1Cert.Curve(), "old_curve": currentState.v1Cert.Curve()},
 					nil,
 				)
 			}
-			}
+
+		} else if currentState.v1Cert != nil {
+			//TODO: CERT-V2 we should be able to tear this down
+			return util.NewContextualError("v1 certificate was removed, restart required", nil, err)
 		}
 
 		if newState.v2Cert != nil {
 			if currentState.v2Cert == nil {
-				//adding certs is fine, actually
-			} else {
+				return util.NewContextualError("v2 certificate was added, restart required", nil, err)
+			}
+
 			// did IP in cert change? if so, don't set
 			if !slices.Equal(currentState.v2Cert.Networks(), newState.v2Cert.Networks()) {
 				return util.NewContextualError(
 					"Networks in new cert was different from old",
-						m{"new_networks": newState.v2Cert.Networks(), "old_networks": currentState.v2Cert.Networks(), "cert_version": cert.Version2},
+					m{"new_networks": newState.v2Cert.Networks(), "old_networks": currentState.v2Cert.Networks()},
 					nil,
 				)
 			}
@@ -137,25 +142,13 @@ func (p *PKI) reloadCerts(c *config.C, initial bool) *util.ContextualError {
 			if currentState.v2Cert.Curve() != newState.v2Cert.Curve() {
 				return util.NewContextualError(
 					"Curve in new cert was different from old",
-						m{"new_curve": newState.v2Cert.Curve(), "old_curve": currentState.v2Cert.Curve(), "cert_version": cert.Version2},
+					m{"new_curve": newState.v2Cert.Curve(), "old_curve": currentState.v2Cert.Curve()},
 					nil,
 				)
 			}
-			}
 
 		} else if currentState.v2Cert != nil {
-			//newState.v1Cert is non-nil bc empty certstates aren't permitted
-			if newState.v1Cert == nil {
-				return util.NewContextualError("v1 and v2 certs are nil, this should be impossible", nil, err)
-			}
-			//if we're going to v1-only, we need to make sure we didn't orphan any v2-cert vpnaddrs
-			if !slices.Equal(currentState.v2Cert.Networks(), newState.v1Cert.Networks()) {
-				return util.NewContextualError(
-					"Removing a V2 cert is not permitted unless it has identical networks to the new V1 cert",
-					m{"new_v1_networks": newState.v1Cert.Networks(), "old_v2_networks": currentState.v2Cert.Networks()},
-					nil,
-				)
-			}
+			return util.NewContextualError("v2 certificate was removed, restart required", nil, err)
 		}
 
 		// Cipher cant be hot swapped so just leave it at what it was before

stats.go

@@ -6,7 +6,6 @@ import (
 	"log"
 	"net"
 	"net/http"
-	_ "net/http/pprof"
 	"runtime"
 	"strconv"
 	"time"

udp/conn.go

@@ -13,21 +13,12 @@ type EncReader func(
 	payload []byte,
 )
 
-type EncBatchReader func(
-	addrs []netip.AddrPort,
-	payloads [][]byte,
-	count int,
-)
-
 type Conn interface {
 	Rebind() error
 	LocalAddr() (netip.AddrPort, error)
 	ListenOut(r EncReader)
-	ListenOutBatch(r EncBatchReader)
 	WriteTo(b []byte, addr netip.AddrPort) error
-	WriteMulti(packets [][]byte, addrs []netip.AddrPort) (int, error)
 	ReloadConfig(c *config.C)
-	BatchSize() int
 	Close() error
 }
 
@@ -42,21 +33,12 @@ func (NoopConn) LocalAddr() (netip.AddrPort, error) {
 func (NoopConn) ListenOut(_ EncReader) {
 	return
 }
-func (NoopConn) ListenOutBatch(_ EncBatchReader) {
-	return
-}
 func (NoopConn) WriteTo(_ []byte, _ netip.AddrPort) error {
 	return nil
 }
-func (NoopConn) WriteMulti(_ [][]byte, _ []netip.AddrPort) (int, error) {
-	return 0, nil
-}
 func (NoopConn) ReloadConfig(_ *config.C) {
 	return
 }
-func (NoopConn) BatchSize() int {
-	return 1
-}
 func (NoopConn) Close() error {
 	return nil
 }

udp/msghdr_helper_linux_32.go

@@ -0,0 +1,17 @@
+//go:build linux && (386 || amd64p32 || arm || mips || mipsle) && !android && !e2e_testing
+// +build linux
+// +build 386 amd64p32 arm mips mipsle
+// +build !android
+// +build !e2e_testing
+
+package udp
+
+import "golang.org/x/sys/unix"
+
+func controllen(n int) uint32 {
+	return uint32(n)
+}
+
+func setCmsgLen(h *unix.Cmsghdr, n int) {
+	h.Len = uint32(unix.CmsgLen(n))
+}

udp/msghdr_helper_linux_64.go

@@ -0,0 +1,17 @@
+//go:build linux && (amd64 || arm64 || ppc64 || ppc64le || mips64 || mips64le || s390x || riscv64 || loong64) && !android && !e2e_testing
+// +build linux
+// +build amd64 arm64 ppc64 ppc64le mips64 mips64le s390x riscv64 loong64
+// +build !android
+// +build !e2e_testing
+
+package udp
+
+import "golang.org/x/sys/unix"
+
+func controllen(n int) uint64 {
+	return uint64(n)
+}
+
+func setCmsgLen(h *unix.Cmsghdr, n int) {
+	h.Len = uint64(unix.CmsgLen(n))
+}

udp/udp_darwin.go

@@ -140,17 +140,6 @@ func (u *StdConn) WriteTo(b []byte, ap netip.AddrPort) error {
 	}
 }
 
-// WriteMulti sends multiple packets - fallback implementation without sendmmsg
-func (u *StdConn) WriteMulti(packets [][]byte, addrs []netip.AddrPort) (int, error) {
-	for i := range packets {
-		err := u.WriteTo(packets[i], addrs[i])
-		if err != nil {
-			return i, err
-		}
-	}
-	return len(packets), nil
-}
-
 func (u *StdConn) LocalAddr() (netip.AddrPort, error) {
 	a := u.UDPConn.LocalAddr()
 
@@ -195,34 +184,6 @@ func (u *StdConn) ListenOut(r EncReader) {
 	}
 }
 
-// ListenOutBatch - fallback to single-packet reads for Darwin
-func (u *StdConn) ListenOutBatch(r EncBatchReader) {
-	buffer := make([]byte, MTU)
-	addrs := make([]netip.AddrPort, 1)
-	payloads := make([][]byte, 1)
-
-	for {
-		// Just read one packet at a time and call batch callback with count=1
-		n, rua, err := u.ReadFromUDPAddrPort(buffer)
-		if err != nil {
-			if errors.Is(err, net.ErrClosed) {
-				u.l.WithError(err).Debug("udp socket is closed, exiting read loop")
-				return
-			}
-
-			u.l.WithError(err).Error("unexpected udp socket receive error")
-		}
-
-		addrs[0] = netip.AddrPortFrom(rua.Addr().Unmap(), rua.Port())
-		payloads[0] = buffer[:n]
-		r(addrs, payloads, 1)
-	}
-}
-
-func (u *StdConn) BatchSize() int {
-	return 1
-}
-
 func (u *StdConn) Rebind() error {
 	var err error
 	if u.isV4 {

udp/udp_generic.go

@@ -85,42 +85,3 @@ func (u *GenericConn) ListenOut(r EncReader) {
 		r(netip.AddrPortFrom(rua.Addr().Unmap(), rua.Port()), buffer[:n])
 	}
 }
-
-// ListenOutBatch - fallback to single-packet reads for generic platforms
-func (u *GenericConn) ListenOutBatch(r EncBatchReader) {
-	buffer := make([]byte, MTU)
-	addrs := make([]netip.AddrPort, 1)
-	payloads := make([][]byte, 1)
-
-	for {
-		// Just read one packet at a time and call batch callback with count=1
-		n, rua, err := u.ReadFromUDPAddrPort(buffer)
-		if err != nil {
-			u.l.WithError(err).Debug("udp socket is closed, exiting read loop")
-			return
-		}
-
-		addrs[0] = netip.AddrPortFrom(rua.Addr().Unmap(), rua.Port())
-		payloads[0] = buffer[:n]
-		r(addrs, payloads, 1)
-	}
-}
-
-// WriteMulti sends multiple packets - fallback implementation
-func (u *GenericConn) WriteMulti(packets [][]byte, addrs []netip.AddrPort) (int, error) {
-	for i := range packets {
-		err := u.WriteTo(packets[i], addrs[i])
-		if err != nil {
-			return i, err
-		}
-	}
-	return len(packets), nil
-}
-
-func (u *GenericConn) BatchSize() int {
-	return 1
-}
-
-func (u *GenericConn) Rebind() error {
-	return nil
-}

udp/udp_linux.go

@@ -5,10 +5,14 @@ package udp
 
 import (
 	"encoding/binary"
+	"errors"
 	"fmt"
 	"net"
 	"net/netip"
+	"sync"
+	"sync/atomic"
 	"syscall"
+	"time"
 	"unsafe"
 
 	"github.com/rcrowley/go-metrics"
@@ -17,16 +21,38 @@ import (
 	"golang.org/x/sys/unix"
 )
 
+const (
+	defaultGSOMaxSegments    = 64
+	defaultGSOMaxBytes       = 64000
+	defaultGROReadBufferSize = 2 * defaultGSOMaxBytes
+	defaultGSOFlushTimeout   = 100 * time.Microsecond
+)
+
 type StdConn struct {
 	sysFd int
 	isV4  bool
 	l     *logrus.Logger
 	batch int
 
-	// Pre-allocated buffers for batch writes (sized for IPv6, works for both)
-	writeMsgs   []rawMessage
-	writeIovecs []iovec
-	writeNames  [][]byte
+	enableGRO bool
+	enableGSO bool
+
+	controlLen atomic.Int32
+
+	gsoMu              sync.Mutex
+	gsoPendingBuf      []byte
+	gsoPendingSegments int
+	gsoPendingAddr     netip.AddrPort
+	gsoPendingSegSize  int
+	gsoMaxSegments     int
+	gsoMaxBytes        int
+	gsoFlushTimeout    time.Duration
+	gsoFlushTimer      *time.Timer
+	gsoControlBuf      []byte
+
+	gsoBatches  metrics.Counter
+	gsoSegments metrics.Counter
+	groSegments metrics.Counter
 }
 
 func maybeIPV4(ip net.IP) (net.IP, bool) {
@@ -74,26 +100,18 @@ func NewListener(l *logrus.Logger, ip netip.Addr, port int, multi bool, batch in
 		return nil, fmt.Errorf("unable to bind to socket: %s", err)
 	}
 
-	c := &StdConn{sysFd: fd, isV4: ip.Is4(), l: l, batch: batch}
-
-	// Pre-allocate write message structures for batching (sized for IPv6, works for both)
-	c.writeMsgs = make([]rawMessage, batch)
-	c.writeIovecs = make([]iovec, batch)
-	c.writeNames = make([][]byte, batch)
-
-	for i := range c.writeMsgs {
-		// Allocate for IPv6 size (larger than IPv4, works for both)
-		c.writeNames[i] = make([]byte, unix.SizeofSockaddrInet6)
-
-		// Point to the iovec in the slice
-		c.writeMsgs[i].Hdr.Iov = &c.writeIovecs[i]
-		c.writeMsgs[i].Hdr.Iovlen = 1
-
-		c.writeMsgs[i].Hdr.Name = &c.writeNames[i][0]
-		// Namelen will be set appropriately in writeMulti4/writeMulti6
-	}
-
-	return c, err
+	return &StdConn{
+		sysFd:           fd,
+		isV4:            ip.Is4(),
+		l:               l,
+		batch:           batch,
+		gsoMaxSegments:  defaultGSOMaxSegments,
+		gsoMaxBytes:     defaultGSOMaxBytes,
+		gsoFlushTimeout: defaultGSOFlushTimeout,
+		gsoBatches:      metrics.GetOrRegisterCounter("udp.gso.batches", nil),
+		gsoSegments:     metrics.GetOrRegisterCounter("udp.gso.segments", nil),
+		groSegments:     metrics.GetOrRegisterCounter("udp.gro.segments", nil),
+	}, err
 }
 
 func (u *StdConn) Rebind() error {
@@ -145,72 +163,67 @@ func (u *StdConn) LocalAddr() (netip.AddrPort, error) {
 func (u *StdConn) ListenOut(r EncReader) {
 	var ip netip.Addr
 
-	msgs, buffers, names := u.PrepareRawMessages(u.batch)
+	msgs, buffers, names, controls := u.PrepareRawMessages(u.batch)
 	read := u.ReadMulti
 	if u.batch == 1 {
 		read = u.ReadSingle
 	}
 
-	udpBatchHist := metrics.GetOrRegisterHistogram("batch.udp_read_size", nil, metrics.NewUniformSample(1024))
-
 	for {
+		//desiredControl := int(u.controlLen.Load())
+		//hasControl := len(controls) > 0
+		//if (desiredControl > 0) != hasControl || (desiredControl > 0 && hasControl && len(controls[0]) != desiredControl) {
+		//	msgs, buffers, names, controls = u.PrepareRawMessages(u.batch)
+		//	hasControl = len(controls) > 0
+		//}
+		//
+		for i := range msgs {
+			if len(controls) <= i || len(controls[i]) == 0 {
+				continue
+			}
+			msgs[i].Hdr.Controllen = controllen(len(controls[i]))
+		}
+
 		n, err := read(msgs)
 		if err != nil {
 			u.l.WithError(err).Debug("udp socket is closed, exiting read loop")
 			return
 		}
 
-		udpBatchHist.Update(int64(n))
-
 		for i := 0; i < n; i++ {
+			payloadLen := int(msgs[i].Len)
+			if payloadLen == 0 {
+				continue
+			}
+
 			// Its ok to skip the ok check here, the slicing is the only error that can occur and it will panic
 			if u.isV4 {
 				ip, _ = netip.AddrFromSlice(names[i][4:8])
 			} else {
 				ip, _ = netip.AddrFromSlice(names[i][8:24])
 			}
-			r(netip.AddrPortFrom(ip.Unmap(), binary.BigEndian.Uint16(names[i][2:4])), buffers[i][:msgs[i].Len])
+			addr := netip.AddrPortFrom(ip.Unmap(), binary.BigEndian.Uint16(names[i][2:4]))
+
+			if len(controls) > i && len(controls[i]) > 0 {
+				if segSize, segCount := u.parseGROSegment(&msgs[i], controls[i]); segSize > 0 && segSize < payloadLen {
+					if u.emitSegments(r, addr, buffers[i][:payloadLen], segSize, segCount) {
+						continue
+					}
+					if segCount > 1 {
+						u.l.WithFields(logrus.Fields{
+							"tag":         "gro-debug",
+							"stage":       "listen_out",
+							"reason":      "emit_failed",
+							"payload_len": payloadLen,
+							"seg_size":    segSize,
+							"seg_count":   segCount,
+						}).Debug("gro-debug fallback to single packet")
 					}
 				}
-}
-
-func (u *StdConn) ListenOutBatch(r EncBatchReader) {
-	var ip netip.Addr
-
-	msgs, buffers, names := u.PrepareRawMessages(u.batch)
-	read := u.ReadMulti
-	if u.batch == 1 {
-		read = u.ReadSingle
 			}
 
-	udpBatchHist := metrics.GetOrRegisterHistogram("batch.udp_read_size", nil, metrics.NewUniformSample(1024))
-
-	// Pre-allocate slices for batch callback
-	addrs := make([]netip.AddrPort, u.batch)
-	payloads := make([][]byte, u.batch)
-
-	for {
-		n, err := read(msgs)
-		if err != nil {
-			u.l.WithError(err).Debug("udp socket is closed, exiting read loop")
-			return
+			r(addr, buffers[i][:payloadLen])
 		}
-
-		udpBatchHist.Update(int64(n))
-
-		// Prepare batch data
-		for i := 0; i < n; i++ {
-			if u.isV4 {
-				ip, _ = netip.AddrFromSlice(names[i][4:8])
-			} else {
-				ip, _ = netip.AddrFromSlice(names[i][8:24])
-			}
-			addrs[i] = netip.AddrPortFrom(ip.Unmap(), binary.BigEndian.Uint16(names[i][2:4]))
-			payloads[i] = buffers[i][:msgs[i].Len]
-		}
-
-		// Call batch callback with all packets
-		r(addrs, payloads, n)
 	}
 }
 
@@ -256,25 +269,19 @@ func (u *StdConn) ReadMulti(msgs []rawMessage) (int, error) {
 }
 
 func (u *StdConn) WriteTo(b []byte, ip netip.AddrPort) error {
+	if u.enableGSO {
+		if err := u.writeToGSO(b, ip); err != nil {
+			return err
+		}
+		return nil
+	}
+
 	if u.isV4 {
 		return u.writeTo4(b, ip)
 	}
 	return u.writeTo6(b, ip)
 }
 
-func (u *StdConn) WriteMulti(packets [][]byte, addrs []netip.AddrPort) (int, error) {
-	if len(packets) != len(addrs) {
-		return 0, fmt.Errorf("packets and addrs length mismatch")
-	}
-	if len(packets) == 0 {
-		return 0, nil
-	}
-	if u.isV4 {
-		return u.writeMulti4(packets, addrs)
-	}
-	return u.writeMulti6(packets, addrs)
-}
-
 func (u *StdConn) writeTo6(b []byte, ip netip.AddrPort) error {
 	var rsa unix.RawSockaddrInet6
 	rsa.Family = unix.AF_INET6
@@ -329,121 +336,492 @@ func (u *StdConn) writeTo4(b []byte, ip netip.AddrPort) error {
 	}
 }
 
-func (u *StdConn) writeMulti4(packets [][]byte, addrs []netip.AddrPort) (int, error) {
-	sent := 0
-	for sent < len(packets) {
-		// Determine batch size based on remaining packets and buffer capacity
-		batchSize := len(packets) - sent
-		if batchSize > len(u.writeMsgs) {
-			batchSize = len(u.writeMsgs)
+func (u *StdConn) writeToGSO(b []byte, addr netip.AddrPort) error {
+	if len(b) == 0 {
+		return nil
 	}
 
-		// Use pre-allocated buffers
-		msgs := u.writeMsgs[:batchSize]
-		iovecs := u.writeIovecs[:batchSize]
-		names := u.writeNames[:batchSize]
-
-		// Setup message structures for this batch
-		for i := 0; i < batchSize; i++ {
-			pktIdx := sent + i
-			if !addrs[pktIdx].Addr().Is4() {
-				return sent + i, ErrInvalidIPv6RemoteForSocket
+	if !addr.IsValid() {
+		return u.directWrite(b, addr)
 	}
 
-			// Setup the packet buffer
-			iovecs[i].Base = &packets[pktIdx][0]
-			iovecs[i].Len = uint(len(packets[pktIdx]))
+	u.gsoMu.Lock()
+	defer u.gsoMu.Unlock()
 
-			// Setup the destination address
-			rsa := (*unix.RawSockaddrInet4)(unsafe.Pointer(&names[i][0]))
-			rsa.Family = unix.AF_INET
-			rsa.Addr = addrs[pktIdx].Addr().As4()
-			binary.BigEndian.PutUint16((*[2]byte)(unsafe.Pointer(&rsa.Port))[:], addrs[pktIdx].Port())
-
-			// Set the appropriate address length for IPv4
-			msgs[i].Hdr.Namelen = unix.SizeofSockaddrInet4
+	if cap(u.gsoPendingBuf) < u.gsoMaxBytes { //I feel like this is bad?
+		u.gsoPendingBuf = make([]byte, 0, u.gsoMaxBytes)
 	}
 
-		// Send this batch
-		nsent, _, err := unix.Syscall6(
-			unix.SYS_SENDMMSG,
-			uintptr(u.sysFd),
-			uintptr(unsafe.Pointer(&msgs[0])),
-			uintptr(batchSize),
-			0,
-			0,
-			0,
-		)
-
-		if err != 0 {
-			return sent + int(nsent), &net.OpError{Op: "sendmmsg", Err: err}
-		}
-
-		sent += int(nsent)
-		if int(nsent) < batchSize {
-			// Couldn't send all packets in batch, return what we sent
-			return sent, nil
+	if u.gsoPendingSegments > 0 && u.gsoPendingAddr != addr {
+		if err := u.flushPendingLocked(); err != nil {
+			return err
 		}
 	}
 
-	return sent, nil
+	if len(b) > u.gsoMaxBytes || u.gsoMaxSegments <= 1 {
+		if err := u.flushPendingLocked(); err != nil {
+			return err
+		}
+		return u.directWrite(b, addr)
+	}
+
+	if u.gsoPendingSegments == 0 {
+		u.gsoPendingAddr = addr
+		u.gsoPendingSegSize = len(b)
+	} else {
+		if len(b) > u.gsoPendingSegSize {
+			if err := u.flushPendingLocked(); err != nil {
+				return err
+			}
+			u.gsoPendingAddr = addr
+			u.gsoPendingSegSize = len(b)
+		} else if len(b) < u.gsoPendingSegSize {
+			if err := u.flushPendingLocked(); err != nil {
+				return err
+			}
+			u.gsoPendingAddr = addr
+			u.gsoPendingSegSize = len(b)
+		}
+	}
+
+	inBuf := len(u.gsoPendingBuf) + len(b)
+	if len(u.gsoPendingBuf)+len(b) > u.gsoMaxBytes {
+		if err := u.flushPendingLocked(); err != nil {
+			return err
+		}
+		u.gsoPendingAddr = addr
+		u.gsoPendingSegSize = len(b)
+	}
+
+	u.gsoPendingBuf = append(u.gsoPendingBuf, b...)
+	u.gsoPendingSegments++
+
+	if u.gsoPendingSegments >= u.gsoMaxSegments {
+		return u.flushPendingLocked()
+	}
+
+	if u.gsoFlushTimeout <= 0 {
+		return u.flushPendingLocked()
+	}
+
+	u.scheduleFlushLocked(inBuf)
+	return nil
 }
 
-func (u *StdConn) writeMulti6(packets [][]byte, addrs []netip.AddrPort) (int, error) {
-	sent := 0
-	for sent < len(packets) {
-		// Determine batch size based on remaining packets and buffer capacity
-		batchSize := len(packets) - sent
-		if batchSize > len(u.writeMsgs) {
-			batchSize = len(u.writeMsgs)
+func (u *StdConn) flushPendingLocked() error {
+	if u.gsoPendingSegments == 0 {
+		u.stopFlushTimerLocked()
+		return nil
 	}
 
-		// Use pre-allocated buffers
-		msgs := u.writeMsgs[:batchSize]
-		iovecs := u.writeIovecs[:batchSize]
-		names := u.writeNames[:batchSize]
+	buf := u.gsoPendingBuf[:len(u.gsoPendingBuf)]
+	addr := u.gsoPendingAddr
+	segSize := u.gsoPendingSegSize
+	segments := u.gsoPendingSegments
 
-		// Setup message structures for this batch
-		for i := 0; i < batchSize; i++ {
-			pktIdx := sent + i
+	u.stopFlushTimerLocked()
 
-			// Setup the packet buffer
-			iovecs[i].Base = &packets[pktIdx][0]
-			iovecs[i].Len = uint(len(packets[pktIdx]))
-
-			// Setup the destination address
-			rsa := (*unix.RawSockaddrInet6)(unsafe.Pointer(&names[i][0]))
-			rsa.Family = unix.AF_INET6
-			rsa.Addr = addrs[pktIdx].Addr().As16()
-			binary.BigEndian.PutUint16((*[2]byte)(unsafe.Pointer(&rsa.Port))[:], addrs[pktIdx].Port())
-
-			// Set the appropriate address length for IPv6
-			msgs[i].Hdr.Namelen = unix.SizeofSockaddrInet6
-		}
-
-		// Send this batch
-		nsent, _, err := unix.Syscall6(
-			unix.SYS_SENDMMSG,
-			uintptr(u.sysFd),
-			uintptr(unsafe.Pointer(&msgs[0])),
-			uintptr(batchSize),
-			0,
-			0,
-			0,
-		)
-
-		if err != 0 {
-			return sent + int(nsent), &net.OpError{Op: "sendmmsg", Err: err}
-		}
-
-		sent += int(nsent)
-		if int(nsent) < batchSize {
-			// Couldn't send all packets in batch, return what we sent
-			return sent, nil
+	var err error
+	if segments <= 1 || !u.enableGSO {
+		err = u.directWrite(buf, addr)
+	} else {
+		err = u.sendSegmentedLocked(buf, addr, segSize)
+		if err != nil && (errors.Is(err, unix.EOPNOTSUPP) || errors.Is(err, unix.ENOTSUP)) {
+			u.enableGSO = false
+			u.l.WithError(err).Warn("UDP GSO not supported, disabling")
+			err = u.sendSequentialLocked(buf, addr, segSize)
 		}
 	}
 
-	return sent, nil
+	if err == nil && segments > 1 && u.enableGSO {
+		if u.gsoBatches != nil {
+			u.gsoBatches.Inc(1)
+		}
+		if u.gsoSegments != nil {
+			u.gsoSegments.Inc(int64(segments))
+		}
+	}
+
+	u.gsoPendingBuf = u.gsoPendingBuf[:0]
+	u.gsoPendingSegments = 0
+	u.gsoPendingSegSize = 0
+	u.gsoPendingAddr = netip.AddrPort{}
+
+	return err
+}
+
+func (u *StdConn) sendSegmentedLocked(buf []byte, addr netip.AddrPort, segSize int) error {
+	if len(buf) == 0 {
+		return nil
+	}
+	if segSize <= 0 {
+		segSize = len(buf)
+	}
+
+	if len(u.gsoControlBuf) < unix.CmsgSpace(2) {
+		u.gsoControlBuf = make([]byte, unix.CmsgSpace(2))
+	}
+	control := u.gsoControlBuf[:unix.CmsgSpace(2)]
+	for i := range control {
+		control[i] = 0
+	}
+
+	hdr := (*unix.Cmsghdr)(unsafe.Pointer(&control[0]))
+	setCmsgLen(hdr, 2)
+	hdr.Level = unix.SOL_UDP
+	hdr.Type = unix.UDP_SEGMENT
+
+	dataOff := unix.CmsgLen(0)
+	binary.NativeEndian.PutUint16(control[dataOff:dataOff+2], uint16(segSize))
+
+	var sa unix.Sockaddr
+	if u.isV4 {
+		sa4 := &unix.SockaddrInet4{Port: int(addr.Port())}
+		sa4.Addr = addr.Addr().As4()
+		sa = sa4
+	} else {
+		sa6 := &unix.SockaddrInet6{Port: int(addr.Port())}
+		sa6.Addr = addr.Addr().As16()
+		sa = sa6
+	}
+
+	for {
+		n, err := unix.SendmsgN(u.sysFd, buf, control[:unix.CmsgSpace(2)], sa, 0)
+		if err != nil {
+			if err == unix.EINTR {
+				continue
+			}
+			return &net.OpError{Op: "sendmsg", Err: err}
+		}
+		if n != len(buf) {
+			return &net.OpError{Op: "sendmsg", Err: unix.EIO}
+		}
+		return nil
+	}
+}
+
+func (u *StdConn) sendSequentialLocked(buf []byte, addr netip.AddrPort, segSize int) error {
+	if len(buf) == 0 {
+		return nil
+	}
+	if segSize <= 0 {
+		segSize = len(buf)
+	}
+
+	for offset := 0; offset < len(buf); offset += segSize {
+		end := offset + segSize
+		if end > len(buf) {
+			end = len(buf)
+		}
+		var err error
+		if u.isV4 {
+			err = u.writeTo4(buf[offset:end], addr)
+		} else {
+			err = u.writeTo6(buf[offset:end], addr)
+		}
+		if err != nil {
+			return err
+		}
+		if end == len(buf) {
+			break
+		}
+	}
+
+	return nil
+}
+
+func (u *StdConn) scheduleFlushLocked(inBuf int) {
+	if u.gsoFlushTimeout <= 0 {
+		_ = u.flushPendingLocked()
+		return
+	}
+
+	t := u.gsoFlushTimeout
+	if inBuf > u.gsoMaxBytes/2 {
+		t = t / 2
+	}
+
+	if u.gsoFlushTimer == nil {
+		u.gsoFlushTimer = time.AfterFunc(t, u.flushTimerHandler)
+		return
+	}
+	if !u.gsoFlushTimer.Stop() {
+		// timer already fired or running; allow handler to exit if no data
+	}
+	u.gsoFlushTimer.Reset(t)
+}
+
+func (u *StdConn) stopFlushTimerLocked() {
+	if u.gsoFlushTimer != nil {
+		u.gsoFlushTimer.Stop()
+	}
+}
+
+func (u *StdConn) flushTimerHandler() {
+	//u.l.Warn("timer hit")
+	u.gsoMu.Lock()
+	defer u.gsoMu.Unlock()
+
+	if u.gsoPendingSegments == 0 {
+		return
+	}
+
+	if err := u.flushPendingLocked(); err != nil {
+		u.l.WithError(err).Warn("Failed to flush GSO batch")
+	}
+}
+
+func (u *StdConn) directWrite(b []byte, addr netip.AddrPort) error {
+	if u.isV4 {
+		return u.writeTo4(b, addr)
+	}
+	return u.writeTo6(b, addr)
+}
+
+func (u *StdConn) emitSegments(r EncReader, addr netip.AddrPort, payload []byte, segSize, segCount int) bool {
+	if segSize <= 0 || segSize >= len(payload) {
+		if u.l.Level >= logrus.DebugLevel {
+			u.l.WithFields(logrus.Fields{
+				"tag":         "gro-debug",
+				"stage":       "emit",
+				"reason":      "invalid_seg_size",
+				"payload_len": len(payload),
+				"seg_size":    segSize,
+				"seg_count":   segCount,
+			}).Debug("gro-debug skip emit")
+		}
+		return false
+	}
+
+	totalLen := len(payload)
+	if segCount <= 0 {
+		segCount = (totalLen + segSize - 1) / segSize
+	}
+	if segCount <= 1 {
+		if u.l.Level >= logrus.DebugLevel {
+			u.l.WithFields(logrus.Fields{
+				"tag":         "gro-debug",
+				"stage":       "emit",
+				"reason":      "single_segment",
+				"payload_len": totalLen,
+				"seg_size":    segSize,
+				"seg_count":   segCount,
+			}).Debug("gro-debug skip emit")
+		}
+		return false
+	}
+
+	//segments := make([][]byte, 0, segCount)
+	start := 0
+	//var firstHeader header.H
+	//firstParsed := false
+	//var firstCounter uint64
+	//var firstRemote uint32
+	numSegments := 0
+	//for start < totalLen && len(segments) < segCount {
+	for start < totalLen && numSegments < segCount {
+		end := start + segSize
+		if end > totalLen {
+			end = totalLen
+		}
+
+		//segment := append([]byte(nil), payload[start:end]...)
+		//q := numSegments % 4 //TODO
+		r(addr, payload[start:end])
+		numSegments++
+		//segments = append(segments, segment)
+		start = end
+
+		//if !firstParsed {
+		//	if err := firstHeader.Parse(segment); err == nil {
+		//		firstParsed = true
+		//		firstCounter = firstHeader.MessageCounter
+		//		firstRemote = firstHeader.RemoteIndex
+		//	} else if u.l.IsLevelEnabled(logrus.DebugLevel) {
+		//		u.l.WithFields(logrus.Fields{
+		//			"tag":         "gro-debug",
+		//			"stage":       "emit",
+		//			"event":       "parse_fail",
+		//			"seg_index":   len(segments) - 1,
+		//			"seg_size":    segSize,
+		//			"seg_count":   segCount,
+		//			"payload_len": totalLen,
+		//			"err":         err,
+		//		}).Debug("gro-debug segment parse failed")
+		//	}
+		//}
+	}
+
+	//for idx, segment := range segments {
+	//	r(addr, segment)
+	//if idx == len(segments)-1 && len(segment) < segSize && u.l.IsLevelEnabled(logrus.DebugLevel) {
+	//	var tail header.H
+	//	if err := tail.Parse(segment); err == nil {
+	//		u.l.WithFields(logrus.Fields{
+	//			"tag":             "gro-debug",
+	//			"stage":           "emit",
+	//			"event":           "tail_segment",
+	//			"segment_len":     len(segment),
+	//			"remote_index":    tail.RemoteIndex,
+	//			"message_counter": tail.MessageCounter,
+	//		}).Debug("gro-debug tail segment metadata")
+	//	} else {
+	//		u.l.WithError(err).Warn("Failed to parse tail segment")
+	//	}
+	//}
+	//}
+
+	if u.groSegments != nil {
+		//u.groSegments.Inc(int64(len(segments)))
+		u.groSegments.Inc(int64(numSegments))
+	}
+
+	//if len(segments) > 0 {
+	//	lastLen := len(segments[len(segments)-1])
+	//	if u.l.IsLevelEnabled(logrus.DebugLevel) {
+	//		u.l.WithFields(logrus.Fields{
+	//			"tag":           "gro-debug",
+	//			"stage":         "emit",
+	//			"event":         "success",
+	//			"payload_len":   totalLen,
+	//			"seg_size":      segSize,
+	//			"seg_count":     segCount,
+	//			"actual_segs":   len(segments),
+	//			"last_seg_len":  lastLen,
+	//			"addr":          addr.String(),
+	//			"first_remote":  firstRemote,
+	//			"first_counter": firstCounter,
+	//		}).Debug("gro-debug emit")
+	//	}
+	//}
+
+	return true
+}
+
+func (u *StdConn) parseGROSegment(msg *rawMessage, control []byte) (int, int) {
+	ctrlLen := int(msg.Hdr.Controllen)
+	if ctrlLen <= 0 {
+		return 0, 0
+	}
+	if ctrlLen > len(control) {
+		ctrlLen = len(control)
+	}
+
+	cmsgs, err := unix.ParseSocketControlMessage(control[:ctrlLen])
+	if err != nil {
+		u.l.WithError(err).Debug("failed to parse UDP GRO control message")
+		return 0, 0
+	}
+
+	for _, c := range cmsgs {
+		if c.Header.Level == unix.SOL_UDP && c.Header.Type == unix.UDP_GRO && len(c.Data) >= 2 {
+			segSize := int(binary.NativeEndian.Uint16(c.Data[:2]))
+			segCount := 0
+			if len(c.Data) >= 4 {
+				segCount = int(binary.NativeEndian.Uint16(c.Data[2:4]))
+			}
+			if u.l.Level >= logrus.DebugLevel {
+				u.l.WithFields(logrus.Fields{
+					"tag":       "gro-debug",
+					"stage":     "parse",
+					"seg_size":  segSize,
+					"seg_count": segCount,
+				}).Debug("gro-debug control parsed")
+			}
+			return segSize, segCount
+		}
+	}
+
+	return 0, 0
+}
+
+func (u *StdConn) configureGRO(enable bool) {
+	if enable == u.enableGRO {
+		if enable {
+			u.controlLen.Store(int32(unix.CmsgSpace(2)))
+		} else {
+			u.controlLen.Store(0)
+		}
+		return
+	}
+
+	if enable {
+		if err := unix.SetsockoptInt(u.sysFd, unix.SOL_UDP, unix.UDP_GRO, 1); err != nil {
+			u.l.WithError(err).Warn("Failed to enable UDP GRO")
+			u.enableGRO = false
+			u.controlLen.Store(0)
+			return
+		}
+		u.enableGRO = true
+		u.controlLen.Store(int32(unix.CmsgSpace(2)))
+		u.l.Info("UDP GRO enabled")
+	} else {
+		if u.enableGRO {
+			if err := unix.SetsockoptInt(u.sysFd, unix.SOL_UDP, unix.UDP_GRO, 0); err != nil {
+				u.l.WithError(err).Warn("Failed to disable UDP GRO")
+			}
+		}
+		u.enableGRO = false
+		u.controlLen.Store(0)
+	}
+}
+
+func (u *StdConn) configureGSO(enable bool, c *config.C) {
+	u.gsoMu.Lock()
+	defer u.gsoMu.Unlock()
+
+	if !enable {
+		if u.enableGSO {
+			if err := u.flushPendingLocked(); err != nil {
+				u.l.WithError(err).Warn("Failed to flush GSO buffers while disabling")
+			}
+			u.enableGSO = false
+			if u.gsoFlushTimer != nil {
+				u.gsoFlushTimer.Stop()
+			}
+			u.l.Info("UDP GSO disabled")
+		}
+		return
+	}
+
+	maxSegments := c.GetInt("listen.gso_max_segments", defaultGSOMaxSegments)
+	if maxSegments < 2 {
+		maxSegments = 2
+	}
+
+	maxBytes := c.GetInt("listen.gso_max_bytes", 0)
+	if maxBytes <= 0 {
+		maxBytes = defaultGSOMaxBytes
+	}
+	if maxBytes < MTU {
+		maxBytes = MTU
+	}
+
+	flushTimeout := c.GetDuration("listen.gso_flush_timeout", defaultGSOFlushTimeout)
+	if flushTimeout < 0 {
+		flushTimeout = 0
+	}
+
+	u.enableGSO = true
+	u.gsoMaxSegments = maxSegments
+	u.gsoMaxBytes = maxBytes
+	u.gsoFlushTimeout = flushTimeout
+
+	if cap(u.gsoPendingBuf) < u.gsoMaxBytes {
+		u.gsoPendingBuf = make([]byte, 0, u.gsoMaxBytes)
+	} else {
+		u.gsoPendingBuf = u.gsoPendingBuf[:0]
+	}
+
+	if len(u.gsoControlBuf) < unix.CmsgSpace(2) {
+		u.gsoControlBuf = make([]byte, unix.CmsgSpace(2))
+	}
+
+	u.l.WithFields(logrus.Fields{
+		"segments":      u.gsoMaxSegments,
+		"bytes":         u.gsoMaxBytes,
+		"flush_timeout": u.gsoFlushTimeout,
+	}).Info("UDP GSO configured")
 }
 
 func (u *StdConn) ReloadConfig(c *config.C) {
@@ -492,6 +870,9 @@ func (u *StdConn) ReloadConfig(c *config.C) {
 			u.l.WithError(err).Error("Failed to set listen.so_mark")
 		}
 	}
+
+	u.configureGRO(c.GetBool("listen.enable_gro", false))
+	u.configureGSO(c.GetBool("listen.enable_gso", false), c)
 }
 
 func (u *StdConn) getMemInfo(meminfo *[unix.SK_MEMINFO_VARS]uint32) error {
@@ -503,12 +884,16 @@ func (u *StdConn) getMemInfo(meminfo *[unix.SK_MEMINFO_VARS]uint32) error {
 	return nil
 }
 
-func (u *StdConn) BatchSize() int {
-	return u.batch
-}
-
 func (u *StdConn) Close() error {
-	return syscall.Close(u.sysFd)
+	u.gsoMu.Lock()
+	flushErr := u.flushPendingLocked()
+	u.gsoMu.Unlock()
+
+	closeErr := syscall.Close(u.sysFd)
+	if flushErr != nil {
+		return flushErr
+	}
+	return closeErr
 }
 
 func NewUDPStatsEmitter(udpConns []Conn) func() {

udp/udp_linux_32.go

@@ -12,7 +12,7 @@ import (
 
 type iovec struct {
 	Base *byte
-	Len  uint
+	Len  uint32
 }
 
 type msghdr struct {
@@ -30,17 +30,28 @@ type rawMessage struct {
 	Len uint32
 }
 
-func (u *StdConn) PrepareRawMessages(n int) ([]rawMessage, [][]byte, [][]byte) {
+func (u *StdConn) PrepareRawMessages(n int) ([]rawMessage, [][]byte, [][]byte, [][]byte) {
+	controlLen := int(u.controlLen.Load())
+
 	msgs := make([]rawMessage, n)
 	buffers := make([][]byte, n)
 	names := make([][]byte, n)
 
+	var controls [][]byte
+	if controlLen > 0 {
+		controls = make([][]byte, n)
+	}
+
 	for i := range msgs {
-		buffers[i] = make([]byte, MTU)
+		size := MTU
+		if defaultGROReadBufferSize > size {
+			size = defaultGROReadBufferSize
+		}
+		buffers[i] = make([]byte, size)
 		names[i] = make([]byte, unix.SizeofSockaddrInet6)
 
 		vs := []iovec{
-			{Base: &buffers[i][0], Len: uint(len(buffers[i]))},
+			{Base: &buffers[i][0], Len: uint32(len(buffers[i]))},
 		}
 
 		msgs[i].Hdr.Iov = &vs[0]
@@ -48,7 +59,16 @@ func (u *StdConn) PrepareRawMessages(n int) ([]rawMessage, [][]byte, [][]byte) {
 
 		msgs[i].Hdr.Name = &names[i][0]
 		msgs[i].Hdr.Namelen = uint32(len(names[i]))
+
+		if controlLen > 0 {
+			controls[i] = make([]byte, controlLen)
+			msgs[i].Hdr.Control = &controls[i][0]
+			msgs[i].Hdr.Controllen = controllen(len(controls[i]))
+		} else {
+			msgs[i].Hdr.Control = nil
+			msgs[i].Hdr.Controllen = controllen(0)
+		}
 	}
 
-	return msgs, buffers, names
+	return msgs, buffers, names, controls
 }

udp/udp_linux_64.go

@@ -12,7 +12,7 @@ import (
 
 type iovec struct {
 	Base *byte
-	Len  uint
+	Len  uint64
 }
 
 type msghdr struct {
@@ -33,25 +33,43 @@ type rawMessage struct {
 	Pad0 [4]byte
 }
 
-func (u *StdConn) PrepareRawMessages(n int) ([]rawMessage, [][]byte, [][]byte) {
+func (u *StdConn) PrepareRawMessages(n int) ([]rawMessage, [][]byte, [][]byte, [][]byte) {
+	controlLen := int(u.controlLen.Load())
+
 	msgs := make([]rawMessage, n)
 	buffers := make([][]byte, n)
 	names := make([][]byte, n)
 
+	var controls [][]byte
+	if controlLen > 0 {
+		controls = make([][]byte, n)
+	}
+
 	for i := range msgs {
-		buffers[i] = make([]byte, MTU)
+		size := MTU
+		if defaultGROReadBufferSize > size {
+			size = defaultGROReadBufferSize
+		}
+		buffers[i] = make([]byte, size)
 		names[i] = make([]byte, unix.SizeofSockaddrInet6)
 
-		vs := []iovec{
-			{Base: &buffers[i][0], Len: uint(len(buffers[i]))},
-		}
+		vs := []iovec{{Base: &buffers[i][0], Len: uint64(len(buffers[i]))}}
 
 		msgs[i].Hdr.Iov = &vs[0]
 		msgs[i].Hdr.Iovlen = uint64(len(vs))
 
 		msgs[i].Hdr.Name = &names[i][0]
 		msgs[i].Hdr.Namelen = uint32(len(names[i]))
+
+		if controlLen > 0 {
+			controls[i] = make([]byte, controlLen)
+			msgs[i].Hdr.Control = &controls[i][0]
+			msgs[i].Hdr.Controllen = controllen(len(controls[i]))
+		} else {
+			msgs[i].Hdr.Control = nil
+			msgs[i].Hdr.Controllen = controllen(0)
+		}
 	}
 
-	return msgs, buffers, names
+	return msgs, buffers, names, controls
 }

udp/udp_tester.go

@@ -116,31 +116,6 @@ func (u *TesterConn) ListenOut(r EncReader) {
 	}
 }
 
-func (u *TesterConn) ListenOutBatch(r EncBatchReader) {
-	addrs := make([]netip.AddrPort, 1)
-	payloads := make([][]byte, 1)
-
-	for {
-		p, ok := <-u.RxPackets
-		if !ok {
-			return
-		}
-		addrs[0] = p.From
-		payloads[0] = p.Data
-		r(addrs, payloads, 1)
-	}
-}
-
-func (u *TesterConn) WriteMulti(packets [][]byte, addrs []netip.AddrPort) (int, error) {
-	for i := range packets {
-		err := u.WriteTo(packets[i], addrs[i])
-		if err != nil {
-			return i, err
-		}
-	}
-	return len(packets), nil
-}
-
 func (u *TesterConn) ReloadConfig(*config.C) {}
 
 func NewUDPStatsEmitter(_ []Conn) func() {
@@ -152,10 +127,6 @@ func (u *TesterConn) LocalAddr() (netip.AddrPort, error) {
 	return u.Addr, nil
 }
 
-func (u *TesterConn) BatchSize() int {
-	return 1
-}
-
 func (u *TesterConn) Rebind() error {
 	return nil
 }