window

cruft

fix tests

haha yep faster

checksum speed

haha

save pennies

fix

typo!

checkpt

GSO again

cert_test/cert.go

@@ -163,55 +163,3 @@ func P256Keypair() ([]byte, []byte) {
 	pubkey := privkey.PublicKey()
 	return pubkey.Bytes(), privkey.Bytes()
 }
-
-// DummyCert is a minimal cert.Certificate implementation for testing error paths.
-type DummyCert struct {
-	Version_        cert.Version
-	Curve_          cert.Curve
-	Groups_         []string
-	IsCA_           bool
-	Issuer_         string
-	Name_           string
-	Networks_       []netip.Prefix
-	NotAfter_       time.Time
-	NotBefore_      time.Time
-	PublicKey_      []byte
-	Signature_      []byte
-	UnsafeNetworks_ []netip.Prefix
-}
-
-func (d *DummyCert) Version() cert.Version                         { return d.Version_ }
-func (d *DummyCert) Curve() cert.Curve                             { return d.Curve_ }
-func (d *DummyCert) Groups() []string                              { return d.Groups_ }
-func (d *DummyCert) IsCA() bool                                    { return d.IsCA_ }
-func (d *DummyCert) Issuer() string                                { return d.Issuer_ }
-func (d *DummyCert) Name() string                                  { return d.Name_ }
-func (d *DummyCert) Networks() []netip.Prefix                      { return d.Networks_ }
-func (d *DummyCert) NotAfter() time.Time                           { return d.NotAfter_ }
-func (d *DummyCert) NotBefore() time.Time                          { return d.NotBefore_ }
-func (d *DummyCert) PublicKey() []byte                             { return d.PublicKey_ }
-func (d *DummyCert) Signature() []byte                             { return d.Signature_ }
-func (d *DummyCert) UnsafeNetworks() []netip.Prefix                { return d.UnsafeNetworks_ }
-func (d *DummyCert) Fingerprint() (string, error)                  { return "", nil }
-func (d *DummyCert) CheckSignature(key []byte) bool                { return false }
-func (d *DummyCert) MarshalForHandshakes() ([]byte, error)         { return nil, nil }
-func (d *DummyCert) MarshalPEM() ([]byte, error)                   { return nil, nil }
-func (d *DummyCert) MarshalJSON() ([]byte, error)                  { return nil, nil }
-func (d *DummyCert) Marshal() ([]byte, error)                      { return nil, nil }
-func (d *DummyCert) String() string                                { return "dummy" }
-func (d *DummyCert) Copy() cert.Certificate                        { return d }
-func (d *DummyCert) VerifyPrivateKey(c cert.Curve, k []byte) error { return nil }
-func (d *DummyCert) Expired(time.Time) bool                        { return false }
-func (d *DummyCert) MarshalPublicKeyPEM() []byte                   { return nil }
-func (d *DummyCert) PublicKeyPEM() []byte                          { return nil }
-
-// NewTestCAPool creates a CAPool from the given CA certificates, panicking on error.
-func NewTestCAPool(cas ...cert.Certificate) *cert.CAPool {
-	pool := cert.NewCAPool()
-	for _, ca := range cas {
-		if err := pool.AddCA(ca); err != nil {
-			panic(err)
-		}
-	}
-	return pool
-}

connection_manager.go

@@ -145,7 +145,6 @@ func (cm *connectionManager) getAndResetTrafficCheck(h *HostInfo, now time.Time)
 func (cm *connectionManager) AddTrafficWatch(h *HostInfo) {
 	if h.out.Swap(true) == false {
 		cm.trafficTimer.Add(h.localIndexId, cm.checkInterval)
-		cm.intf.pmtudManager.OnTunnelUp(h)
 	}
 }
 
@@ -181,7 +180,6 @@ func (cm *connectionManager) doTrafficCheck(localIndex uint32, p, nb, out []byte
 
 	switch decision {
 	case deleteTunnel:
-		cm.intf.pmtudManager.OnTunnelDown(hostinfo)
 		if cm.hostMap.DeleteHostInfo(hostinfo) {
 			// Only clearing the lighthouse cache if this is the last hostinfo for this vpn ip in the hostmap
 			cm.intf.lightHouse.DeleteVpnAddrs(hostinfo.vpnAddrs)
@@ -201,15 +199,8 @@ func (cm *connectionManager) doTrafficCheck(localIndex uint32, p, nb, out []byte
 		cm.tryRehandshake(hostinfo)
 
 	case sendTestPacket:
-		// Defer to pmtud if it has a confirmed PMTU > floor for this peer:
-		// the probe at the confirmed size verifies both liveness AND that
-		// the discovered PMTU still fits, so we don't burn a separate test
-		// packet on top of it. If pmtud declines (disabled, peer unsupported,
-		// or no confirmed size yet) we fall back to the regular test.
-		if !cm.intf.pmtudManager.MaybeProbeAsTest(hostinfo) {
 		cm.intf.SendMessageToHostInfo(header.Test, header.TestRequest, hostinfo, p, nb, out)
 	}
-	}
 
 	cm.resetRelayTrafficCheck(hostinfo)
 }

connection_manager_test.go

@@ -7,6 +7,7 @@ import (
 	"testing"
 	"time"
 
+	"github.com/flynn/noise"
 	"github.com/slackhq/nebula/cert"
 	"github.com/slackhq/nebula/config"
 	"github.com/slackhq/nebula/overlay/overlaytest"
@@ -46,7 +47,7 @@ func Test_NewConnectionManagerTest(t *testing.T) {
 		initiatingVersion: cert.Version1,
 		privateKey:        []byte{},
 		v1Cert:            &dummyCert{version: cert.Version1},
-		v1Credential:      nil,
+		v1HandshakeBytes:  []byte{},
 	}
 
 	lh := newTestLighthouse()
@@ -67,8 +68,6 @@ func Test_NewConnectionManagerTest(t *testing.T) {
 	punchy := NewPunchyFromConfig(test.NewLogger(), conf)
 	nc := newConnectionManagerFromConfig(test.NewLogger(), conf, hostMap, punchy)
 	nc.intf = ifce
-	ifce.pmtudManager = newPMTUDManagerFromConfig(test.NewLogger(), conf, ifce.inside)
-	ifce.pmtudManager.intf = ifce
 	p := []byte("")
 	nb := make([]byte, 12, 12)
 	out := make([]byte, mtu)
@@ -81,6 +80,7 @@ func Test_NewConnectionManagerTest(t *testing.T) {
 	}
 	hostinfo.ConnectionState = &ConnectionState{
 		myCert: &dummyCert{version: cert.Version1},
+		H:      &noise.HandshakeState{},
 	}
 	nc.hostMap.unlockedAddHostInfo(hostinfo, ifce)
 
@@ -130,7 +130,7 @@ func Test_NewConnectionManagerTest2(t *testing.T) {
 		initiatingVersion: cert.Version1,
 		privateKey:        []byte{},
 		v1Cert:            &dummyCert{version: cert.Version1},
-		v1Credential:      nil,
+		v1HandshakeBytes:  []byte{},
 	}
 
 	lh := newTestLighthouse()
@@ -151,8 +151,6 @@ func Test_NewConnectionManagerTest2(t *testing.T) {
 	punchy := NewPunchyFromConfig(test.NewLogger(), conf)
 	nc := newConnectionManagerFromConfig(test.NewLogger(), conf, hostMap, punchy)
 	nc.intf = ifce
-	ifce.pmtudManager = newPMTUDManagerFromConfig(test.NewLogger(), conf, ifce.inside)
-	ifce.pmtudManager.intf = ifce
 	p := []byte("")
 	nb := make([]byte, 12, 12)
 	out := make([]byte, mtu)
@@ -165,6 +163,7 @@ func Test_NewConnectionManagerTest2(t *testing.T) {
 	}
 	hostinfo.ConnectionState = &ConnectionState{
 		myCert: &dummyCert{version: cert.Version1},
+		H:      &noise.HandshakeState{},
 	}
 	nc.hostMap.unlockedAddHostInfo(hostinfo, ifce)
 
@@ -216,7 +215,7 @@ func Test_NewConnectionManager_DisconnectInactive(t *testing.T) {
 		initiatingVersion: cert.Version1,
 		privateKey:        []byte{},
 		v1Cert:            &dummyCert{version: cert.Version1},
-		v1Credential:      nil,
+		v1HandshakeBytes:  []byte{},
 	}
 
 	lh := newTestLighthouse()
@@ -250,6 +249,7 @@ func Test_NewConnectionManager_DisconnectInactive(t *testing.T) {
 	}
 	hostinfo.ConnectionState = &ConnectionState{
 		myCert: &dummyCert{version: cert.Version1},
+		H:      &noise.HandshakeState{},
 	}
 	nc.hostMap.unlockedAddHostInfo(hostinfo, ifce)
 
@@ -342,7 +342,7 @@ func Test_NewConnectionManagerTest_DisconnectInvalid(t *testing.T) {
 	cs := &CertState{
 		privateKey:       []byte{},
 		v1Cert:           &dummyCert{},
-		v1Credential: nil,
+		v1HandshakeBytes: []byte{},
 	}
 
 	lh := newTestLighthouse()
@@ -365,8 +365,6 @@ func Test_NewConnectionManagerTest_DisconnectInvalid(t *testing.T) {
 	punchy := NewPunchyFromConfig(test.NewLogger(), conf)
 	nc := newConnectionManagerFromConfig(test.NewLogger(), conf, hostMap, punchy)
 	nc.intf = ifce
-	ifce.pmtudManager = newPMTUDManagerFromConfig(test.NewLogger(), conf, ifce.inside)
-	ifce.pmtudManager.intf = ifce
 	ifce.connectionManager = nc
 
 	hostinfo := &HostInfo{
@@ -374,6 +372,7 @@ func Test_NewConnectionManagerTest_DisconnectInvalid(t *testing.T) {
 		ConnectionState: &ConnectionState{
 			myCert:   &dummyCert{},
 			peerCert: cachedPeerCert,
+			H:        &noise.HandshakeState{},
 		},
 	}
 	nc.hostMap.unlockedAddHostInfo(hostinfo, ifce)

connection_state.go

@@ -1,19 +1,23 @@
 package nebula
 
 import (
+	"crypto/rand"
 	"encoding/json"
+	"fmt"
 	"sync"
 	"sync/atomic"
 
+	"github.com/flynn/noise"
 	"github.com/slackhq/nebula/cert"
-	"github.com/slackhq/nebula/handshake"
+	"github.com/slackhq/nebula/noiseutil"
 )
 
-const ReplayWindow = 1024
+const ReplayWindow = 1024 //todo I've started seeing out-of-window messages in testing?
 
 type ConnectionState struct {
 	eKey           *NebulaCipherState
 	dKey           *NebulaCipherState
+	H              *noise.HandshakeState
 	myCert         cert.Certificate
 	peerCert       *cert.CachedCertificate
 	initiator      bool
@@ -22,24 +26,55 @@ type ConnectionState struct {
 	writeLock      sync.Mutex
 }
 
-// newConnectionStateFromResult builds a fully-populated ConnectionState from a
+func NewConnectionState(cs *CertState, crt cert.Certificate, initiator bool, pattern noise.HandshakePattern) (*ConnectionState, error) {
-// completed handshake.Result. It seeds messageCounter and the replay window so
+	var dhFunc noise.DHFunc
-// that the post-handshake message indices already used on the wire don't count
+	switch crt.Curve() {
-// as missed traffic in the data plane.
+	case cert.Curve_CURVE25519:
-func newConnectionStateFromResult(r *handshake.Result) *ConnectionState {
+		dhFunc = noise.DH25519
+	case cert.Curve_P256:
+		if cs.pkcs11Backed {
+			dhFunc = noiseutil.DHP256PKCS11
+		} else {
+			dhFunc = noiseutil.DHP256
+		}
+	default:
+		return nil, fmt.Errorf("invalid curve: %s", crt.Curve())
+	}
+
+	var ncs noise.CipherSuite
+	if cs.cipher == "chachapoly" {
+		ncs = noise.NewCipherSuite(dhFunc, noise.CipherChaChaPoly, noise.HashSHA256)
+	} else {
+		ncs = noise.NewCipherSuite(dhFunc, noiseutil.CipherAESGCM, noise.HashSHA256)
+	}
+
+	static := noise.DHKey{Private: cs.privateKey, Public: crt.PublicKey()}
+	hs, err := noise.NewHandshakeState(noise.Config{
+		CipherSuite:   ncs,
+		Random:        rand.Reader,
+		Pattern:       pattern,
+		Initiator:     initiator,
+		StaticKeypair: static,
+		//NOTE: These should come from CertState (pki.go) when we finally implement it
+		PresharedKey:          []byte{},
+		PresharedKeyPlacement: 0,
+	})
+	if err != nil {
+		return nil, fmt.Errorf("NewConnectionState: %s", err)
+	}
+
+	// The queue and ready params prevent a counter race that would happen when
+	// sending stored packets and simultaneously accepting new traffic.
 	ci := &ConnectionState{
-		myCert:    r.MyCert,
+		H:         hs,
-		initiator: r.Initiator,
+		initiator: initiator,
-		peerCert:  r.RemoteCert,
-		eKey:      NewNebulaCipherState(r.EKey),
-		dKey:      NewNebulaCipherState(r.DKey),
 		window:    NewBits(ReplayWindow),
+		myCert:    crt,
 	}
-	ci.messageCounter.Add(r.MessageIndex)
+	// always start the counter from 2, as packet 1 and packet 2 are handshake packets.
-	for i := uint64(1); i <= r.MessageIndex; i++ {
+	ci.messageCounter.Add(2)
-		ci.window.Update(nil, i)
+
-	}
+	return ci, nil
-	return ci
 }
 
 func (cs *ConnectionState) MarshalJSON() ([]byte, error) {

connection_state_test.go

@@ -1,114 +0,0 @@
-package nebula
-
-import (
-	"net/netip"
-	"testing"
-	"time"
-
-	"github.com/flynn/noise"
-	"github.com/slackhq/nebula/cert"
-	ct "github.com/slackhq/nebula/cert_test"
-	"github.com/slackhq/nebula/handshake"
-	"github.com/slackhq/nebula/header"
-	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
-)
-
-// runTestHandshake runs a complete IX handshake between two freshly-built
-// peers and returns the initiator and responder Results. Used to produce
-// real cipher states for tests that need to exercise post-handshake glue.
-func runTestHandshake(t *testing.T) (initR, respR *handshake.Result) {
-	t.Helper()
-
-	ca, _, caKey, _ := ct.NewTestCaCert(
-		cert.Version2, cert.Curve_CURVE25519, time.Time{}, time.Time{}, nil, nil, nil,
-	)
-	caPool := ct.NewTestCAPool(ca)
-
-	makeCreds := func(name string, networks []netip.Prefix) handshake.GetCredentialFunc {
-		c, _, rawKey, _ := ct.NewTestCert(
-			cert.Version2, cert.Curve_CURVE25519, ca, caKey,
-			name, ca.NotBefore(), ca.NotAfter(), networks, nil, nil,
-		)
-		priv, _, _, err := cert.UnmarshalPrivateKeyFromPEM(rawKey)
-		require.NoError(t, err)
-		hsBytes, err := c.MarshalForHandshakes()
-		require.NoError(t, err)
-		ncs := noise.NewCipherSuite(noise.DH25519, noise.CipherChaChaPoly, noise.HashSHA256)
-		cred := handshake.NewCredential(c, hsBytes, priv, ncs)
-		return func(v cert.Version) *handshake.Credential {
-			if v == cert.Version2 {
-				return cred
-			}
-			return nil
-		}
-	}
-
-	verifier := func(c cert.Certificate) (*cert.CachedCertificate, error) {
-		return caPool.VerifyCertificate(time.Now(), c)
-	}
-
-	initCreds := makeCreds("initiator", []netip.Prefix{netip.MustParsePrefix("10.0.0.1/24")})
-	respCreds := makeCreds("responder", []netip.Prefix{netip.MustParsePrefix("10.0.0.2/24")})
-
-	initM, err := handshake.NewMachine(
-		cert.Version2, initCreds, verifier,
-		func() (uint32, error) { return 1000, nil },
-		true, header.HandshakeIXPSK0,
-	)
-	require.NoError(t, err)
-
-	respM, err := handshake.NewMachine(
-		cert.Version2, respCreds, verifier,
-		func() (uint32, error) { return 2000, nil },
-		false, header.HandshakeIXPSK0,
-	)
-	require.NoError(t, err)
-
-	msg1, err := initM.Initiate(nil)
-	require.NoError(t, err)
-
-	resp, respR, err := respM.ProcessPacket(nil, msg1)
-	require.NoError(t, err)
-	require.NotNil(t, respR)
-
-	_, initR, err = initM.ProcessPacket(nil, resp)
-	require.NoError(t, err)
-	require.NotNil(t, initR)
-
-	return initR, respR
-}
-
-func TestNewConnectionStateFromResult(t *testing.T) {
-	initR, respR := runTestHandshake(t)
-
-	t.Run("initiator", func(t *testing.T) {
-		ci := newConnectionStateFromResult(initR)
-		assert.True(t, ci.initiator)
-		assert.Equal(t, initR.MyCert, ci.myCert)
-		assert.Equal(t, initR.RemoteCert, ci.peerCert)
-		assert.NotNil(t, ci.eKey)
-		assert.NotNil(t, ci.dKey)
-
-		// IX has 2 handshake messages; the next data-plane send is counter=3.
-		assert.Equal(t, uint64(2), ci.messageCounter.Load(),
-			"messageCounter must equal Result.MessageIndex so the next send is N+1")
-
-		// Both handshake counters must be marked seen so they don't appear lost.
-		// Check returns false if an index has already been recorded.
-		assert.False(t, ci.window.Check(nil, 1), "counter 1 must already be seen")
-		assert.False(t, ci.window.Check(nil, 2), "counter 2 must already be seen")
-		// Counter 3 is the next data-plane message and must NOT be pre-marked.
-		assert.True(t, ci.window.Check(nil, 3), "counter 3 must not be pre-seeded")
-	})
-
-	t.Run("responder", func(t *testing.T) {
-		ci := newConnectionStateFromResult(respR)
-		assert.False(t, ci.initiator)
-		assert.Equal(t, respR.MyCert, ci.myCert)
-		assert.Equal(t, respR.RemoteCert, ci.peerCert)
-		assert.NotNil(t, ci.eKey)
-		assert.NotNil(t, ci.dKey)
-		assert.Equal(t, uint64(2), ci.messageCounter.Load())
-	})
-}

control.go

@@ -54,7 +54,6 @@ type Control struct {
 	dnsStart               func()
 	lighthouseStart        func()
 	connectionManagerStart func(context.Context)
-	pmtudManagerStart      func(context.Context)
 }
 
 type ControlHostInfo struct {
@@ -108,9 +107,6 @@ func (c *Control) Start() (func() error, error) {
 	if c.connectionManagerStart != nil {
 		go c.connectionManagerStart(c.ctx)
 	}
-	if c.pmtudManagerStart != nil {
-		go c.pmtudManagerStart(c.ctx)
-	}
 	if c.lighthouseStart != nil {
 		c.lighthouseStart()
 	}

e2e/handshake_manager_test.go

@@ -28,7 +28,6 @@ func makeHandshakePacket(from, to netip.AddrPort, subtype header.MessageSubType,
 }
 
 func TestHandshakeRetransmitDuplicate(t *testing.T) {
-	t.Parallel()
 	// Verify the responder correctly handles receiving the same msg1 multiple times
 	// (retransmission). The duplicate goes through CheckAndComplete -> ErrAlreadySeen
 	// and the cached response is resent.
@@ -79,7 +78,6 @@ func TestHandshakeRetransmitDuplicate(t *testing.T) {
 }
 
 func TestHandshakeTruncatedPacketRecovery(t *testing.T) {
-	t.Parallel()
 	// Verify that a truncated handshake packet is ignored and the real
 	// packet can still complete the handshake.
 
@@ -128,7 +126,6 @@ func TestHandshakeTruncatedPacketRecovery(t *testing.T) {
 }
 
 func TestHandshakeOrphanedMsg2Dropped(t *testing.T) {
-	t.Parallel()
 	// A msg2 arriving with no matching pending index should be silently dropped
 	// with no response sent and no state changes.
 
@@ -171,7 +168,6 @@ func TestHandshakeOrphanedMsg2Dropped(t *testing.T) {
 }
 
 func TestHandshakeUnknownMessageCounter(t *testing.T) {
-	t.Parallel()
 	// A handshake packet with an unexpected message counter should be silently
 	// dropped with no side effects and no UDP response.
 
@@ -203,7 +199,6 @@ func TestHandshakeUnknownMessageCounter(t *testing.T) {
 }
 
 func TestHandshakeUnknownSubtype(t *testing.T) {
-	t.Parallel()
 	// A handshake packet with an unknown subtype should be silently dropped.
 
 	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version1, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
@@ -229,7 +224,6 @@ func TestHandshakeUnknownSubtype(t *testing.T) {
 }
 
 func TestHandshakeLateResponse(t *testing.T) {
-	t.Parallel()
 	// After a handshake times out, a late response should be silently ignored
 	// with no new tunnels created.
 
@@ -279,7 +273,6 @@ func TestHandshakeLateResponse(t *testing.T) {
 }
 
 func TestHandshakeSelfConnectionRejected(t *testing.T) {
-	t.Parallel()
 	// Verify that a node rejects a handshake containing its own VPN IP in the
 	// peer cert. We do this by sending the initiator's own msg1 back to itself.
 
@@ -328,7 +321,6 @@ func TestHandshakeSelfConnectionRejected(t *testing.T) {
 }
 
 func TestHandshakeMessageCounter0Dropped(t *testing.T) {
-	t.Parallel()
 	// MessageCounter=0 is not a valid handshake message and should be dropped.
 
 	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version1, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
@@ -349,7 +341,6 @@ func TestHandshakeMessageCounter0Dropped(t *testing.T) {
 }
 
 func TestHandshakeRemoteAllowList(t *testing.T) {
-	t.Parallel()
 	// Verify that a handshake from a blocked underlay IP is dropped with no
 	// response and no state changes. Then verify the same packet from an
 	// allowed IP succeeds.
@@ -408,7 +399,6 @@ func TestHandshakeRemoteAllowList(t *testing.T) {
 }
 
 func TestHandshakeAlreadySeenPreferredRemote(t *testing.T) {
-	t.Parallel()
 	// When a duplicate msg1 arrives via ErrAlreadySeen, verify the tunnel
 	// remains functional and hostmap index count is stable.
 
@@ -455,7 +445,6 @@ func TestHandshakeAlreadySeenPreferredRemote(t *testing.T) {
 }
 
 func TestHandshakeWrongResponderPacketStore(t *testing.T) {
-	t.Parallel()
 	// Verify that when the wrong host responds, the cached packets are
 	// transferred to the new handshake, the evil tunnel is closed, evil's
 	// address is blocked, and the correct tunnel is eventually established.
@@ -519,7 +508,6 @@ func TestHandshakeWrongResponderPacketStore(t *testing.T) {
 }
 
 func TestHandshakeRelayComplete(t *testing.T) {
-	t.Parallel()
 	// Verify that a relay handshake completes correctly and relay state is
 	// properly maintained on all three nodes.
 

e2e/handshakes_test.go

@@ -84,7 +84,6 @@ func BenchmarkHotPathRelay(b *testing.B) {
 }
 
 func TestGoodHandshake(t *testing.T) {
-	t.Parallel()
 	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version1, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
 	myControl, myVpnIpNet, myUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "me", "10.128.0.1/24", nil)
 	theirControl, theirVpnIpNet, theirUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "them", "10.128.0.2/24", nil)
@@ -135,7 +134,6 @@ func TestGoodHandshake(t *testing.T) {
 }
 
 func TestGoodHandshakeNoOverlap(t *testing.T) {
-	t.Parallel()
 	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version2, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
 	myControl, myVpnIpNet, myUdpAddr, _ := newSimpleServer(cert.Version2, ca, caKey, "me", "10.128.0.1/24", nil)
 	theirControl, theirVpnIpNet, theirUdpAddr, _ := newSimpleServer(cert.Version2, ca, caKey, "them", "2001::69/24", nil) //look ma, cross-stack!
@@ -171,7 +169,6 @@ func TestGoodHandshakeNoOverlap(t *testing.T) {
 }
 
 func TestWrongResponderHandshake(t *testing.T) {
-	t.Parallel()
 	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version1, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
 
 	myControl, myVpnIpNet, myUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "me", "10.128.0.100/24", nil)
@@ -248,7 +245,6 @@ func TestWrongResponderHandshake(t *testing.T) {
 }
 
 func TestWrongResponderHandshakeStaticHostMap(t *testing.T) {
-	t.Parallel()
 	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version1, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
 
 	theirControl, theirVpnIpNet, theirUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "them", "10.128.0.99/24", nil)
@@ -331,7 +327,6 @@ func TestWrongResponderHandshakeStaticHostMap(t *testing.T) {
 }
 
 func TestStage1Race(t *testing.T) {
-	t.Parallel()
 	// This tests ensures that two hosts handshaking with each other at the same time will allow traffic to flow
 	// But will eventually collapse down to a single tunnel
 
@@ -412,7 +407,6 @@ func TestStage1Race(t *testing.T) {
 }
 
 func TestUncleanShutdownRaceLoser(t *testing.T) {
-	t.Parallel()
 	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version1, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
 	myControl, myVpnIpNet, myUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "me  ", "10.128.0.1/24", nil)
 	theirControl, theirVpnIpNet, theirUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "them", "10.128.0.2/24", nil)
@@ -462,7 +456,6 @@ func TestUncleanShutdownRaceLoser(t *testing.T) {
 }
 
 func TestUncleanShutdownRaceWinner(t *testing.T) {
-	t.Parallel()
 	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version1, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
 	myControl, myVpnIpNet, myUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "me  ", "10.128.0.1/24", nil)
 	theirControl, theirVpnIpNet, theirUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "them", "10.128.0.2/24", nil)
@@ -514,7 +507,6 @@ func TestUncleanShutdownRaceWinner(t *testing.T) {
 }
 
 func TestRelays(t *testing.T) {
-	t.Parallel()
 	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version1, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
 	myControl, myVpnIpNet, _, _ := newSimpleServer(cert.Version1, ca, caKey, "me     ", "10.128.0.1/24", m{"relay": m{"use_relays": true}})
 	relayControl, relayVpnIpNet, relayUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "relay  ", "10.128.0.128/24", m{"relay": m{"am_relay": true}})
@@ -544,7 +536,6 @@ func TestRelays(t *testing.T) {
 }
 
 func TestRelaysDontCareAboutIps(t *testing.T) {
-	t.Parallel()
 	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version2, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
 	myControl, myVpnIpNet, _, _ := newSimpleServer(cert.Version2, ca, caKey, "me     ", "10.128.0.1/24", m{"relay": m{"use_relays": true}})
 	relayControl, relayVpnIpNet, relayUdpAddr, _ := newSimpleServer(cert.Version2, ca, caKey, "relay  ", "2001::9999/24", m{"relay": m{"am_relay": true}})
@@ -574,7 +565,6 @@ func TestRelaysDontCareAboutIps(t *testing.T) {
 }
 
 func TestReestablishRelays(t *testing.T) {
-	t.Parallel()
 	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version1, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
 	myControl, myVpnIpNet, _, _ := newSimpleServer(cert.Version1, ca, caKey, "me     ", "10.128.0.1/24", m{"relay": m{"use_relays": true}})
 	relayControl, relayVpnIpNet, relayUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "relay  ", "10.128.0.128/24", m{"relay": m{"am_relay": true}})
@@ -706,7 +696,6 @@ func TestReestablishRelays(t *testing.T) {
 }
 
 func TestStage1RaceRelays(t *testing.T) {
-	t.Parallel()
 	//NOTE: this is a race between me and relay resulting in a full tunnel from me to them via relay
 	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version1, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
 	myControl, myVpnIpNet, myUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "me     ", "10.128.0.1/24", m{"relay": m{"use_relays": true}})
@@ -754,7 +743,6 @@ func TestStage1RaceRelays(t *testing.T) {
 }
 
 func TestStage1RaceRelays2(t *testing.T) {
-	t.Parallel()
 	//NOTE: this is a race between me and relay resulting in a full tunnel from me to them via relay
 	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version1, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
 	myControl, myVpnIpNet, myUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "me     ", "10.128.0.1/24", m{"relay": m{"use_relays": true}})
@@ -831,7 +819,6 @@ func TestStage1RaceRelays2(t *testing.T) {
 }
 
 func TestRehandshakingRelays(t *testing.T) {
-	t.Parallel()
 	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version1, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
 	myControl, myVpnIpNet, _, _ := newSimpleServer(cert.Version1, ca, caKey, "me     ", "10.128.0.1/24", m{"relay": m{"use_relays": true}})
 	relayControl, relayVpnIpNet, relayUdpAddr, relayConfig := newSimpleServer(cert.Version1, ca, caKey, "relay  ", "10.128.0.128/24", m{"relay": m{"am_relay": true}})
@@ -935,7 +922,6 @@ func TestRehandshakingRelays(t *testing.T) {
 }
 
 func TestRehandshakingRelaysPrimary(t *testing.T) {
-	t.Parallel()
 	// This test is the same as TestRehandshakingRelays but one of the terminal types is a primary swap winner
 	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version1, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
 	myControl, myVpnIpNet, _, _ := newSimpleServer(cert.Version1, ca, caKey, "me     ", "10.128.0.128/24", m{"relay": m{"use_relays": true}})
@@ -1040,7 +1026,6 @@ func TestRehandshakingRelaysPrimary(t *testing.T) {
 }
 
 func TestRehandshaking(t *testing.T) {
-	t.Parallel()
 	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version1, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
 	myControl, myVpnIpNet, myUdpAddr, myConfig := newSimpleServer(cert.Version1, ca, caKey, "me  ", "10.128.0.2/24", nil)
 	theirControl, theirVpnIpNet, theirUdpAddr, theirConfig := newSimpleServer(cert.Version1, ca, caKey, "them", "10.128.0.1/24", nil)
@@ -1136,7 +1121,6 @@ func TestRehandshaking(t *testing.T) {
 }
 
 func TestRehandshakingLoser(t *testing.T) {
-	t.Parallel()
 	// The purpose of this test is that the race loser renews their certificate and rehandshakes. The final tunnel
 	// Should be the one with the new certificate
 	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version1, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
@@ -1235,7 +1219,6 @@ func TestRehandshakingLoser(t *testing.T) {
 }
 
 func TestRaceRegression(t *testing.T) {
-	t.Parallel()
 	// This test forces stage 1, stage 2, stage 1 to be received by me from them
 	// We had a bug where we were not finding the duplicate handshake and responding to the final stage 1 which
 	// caused a cross-linked hostinfo
@@ -1296,7 +1279,6 @@ func TestRaceRegression(t *testing.T) {
 }
 
 func TestV2NonPrimaryWithLighthouse(t *testing.T) {
-	t.Parallel()
 	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version2, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
 	lhControl, lhVpnIpNet, lhUdpAddr, _ := newSimpleServer(cert.Version2, ca, caKey, "lh  ", "10.128.0.1/24, ff::1/64", m{"lighthouse": m{"am_lighthouse": true}})
 
@@ -1337,7 +1319,6 @@ func TestV2NonPrimaryWithLighthouse(t *testing.T) {
 }
 
 func TestV2NonPrimaryWithOffNetLighthouse(t *testing.T) {
-	t.Parallel()
 	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version2, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
 	lhControl, lhVpnIpNet, lhUdpAddr, _ := newSimpleServer(cert.Version2, ca, caKey, "lh  ", "2001::1/64", m{"lighthouse": m{"am_lighthouse": true}})
 
@@ -1378,7 +1359,6 @@ func TestV2NonPrimaryWithOffNetLighthouse(t *testing.T) {
 }
 
 func TestLighthouseUpdateOnReload(t *testing.T) {
-	t.Parallel()
 	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version2, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
 
 	// Create the lighthouse
@@ -1454,7 +1434,6 @@ func TestLighthouseUpdateOnReload(t *testing.T) {
 }
 
 func TestGoodHandshakeUnsafeDest(t *testing.T) {
-	t.Parallel()
 	unsafePrefix := "192.168.6.0/24"
 	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version2, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
 	theirControl, theirVpnIpNet, theirUdpAddr, _ := newSimpleServerWithUdpAndUnsafeNetworks(cert.Version2, ca, caKey, "spooky", "10.128.0.2/24", netip.MustParseAddrPort("10.64.0.2:4242"), unsafePrefix, nil)

e2e/leak_test.go

@@ -1,51 +0,0 @@
-//go:build e2e_testing
-// +build e2e_testing
-
-package e2e
-
-import (
-	"testing"
-	"time"
-
-	"github.com/slackhq/nebula/cert"
-	"github.com/slackhq/nebula/cert_test"
-	"github.com/slackhq/nebula/e2e/router"
-	"go.uber.org/goleak"
-)
-
-// TestNoGoroutineLeaks brings up two nebula instances, completes a tunnel,
-// stops both, and asserts no goroutines leak past the shutdown. goleak's
-// retry mechanism gives the wg.Wait()-driven goroutines a moment to drain
-// before failing the assertion.
-//
-// IgnoreCurrent is necessary in the parallelized suite: other tests can
-// leave goroutines mid-shutdown when this one runs (Stop is async, the
-// wg.Wait() drain is not blocking on test return). We're checking that
-// *this* test's setup tears down cleanly, not that the whole suite is
-// idle at this moment. Intentionally NOT t.Parallel()'d for the same
-// reason — concurrent test goroutines would always show up.
-func TestNoGoroutineLeaks(t *testing.T) {
-	defer goleak.VerifyNone(t, goleak.IgnoreCurrent())
-
-	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version1, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
-	myControl, myVpnIpNet, myUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "me", "10.128.0.1/24", nil)
-	theirControl, theirVpnIpNet, theirUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "them", "10.128.0.2/24", nil)
-
-	myControl.InjectLightHouseAddr(theirVpnIpNet[0].Addr(), theirUdpAddr)
-	theirControl.InjectLightHouseAddr(myVpnIpNet[0].Addr(), myUdpAddr)
-
-	myControl.Start()
-	theirControl.Start()
-
-	r := router.NewR(t, myControl, theirControl)
-	assertTunnel(t, myVpnIpNet[0].Addr(), theirVpnIpNet[0].Addr(), myControl, theirControl, r)
-
-	myControl.Stop()
-	theirControl.Stop()
-	r.RenderFlow()
-
-	// Settle period: Stop() is non-blocking; the wg-driven goroutines need
-	// a moment to drain. goleak retries internally too, but a short explicit
-	// settle reduces flakes when the suite is busy.
-	time.Sleep(50 * time.Millisecond)
-}

e2e/tunnels_test.go

@@ -19,7 +19,6 @@ import (
 )
 
 func TestDropInactiveTunnels(t *testing.T) {
-	t.Parallel()
 	// 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{})
@@ -64,7 +63,6 @@ func TestDropInactiveTunnels(t *testing.T) {
 }
 
 func TestCertUpgrade(t *testing.T) {
-	t.Parallel()
 	// 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{})
@@ -159,7 +157,6 @@ func TestCertUpgrade(t *testing.T) {
 }
 
 func TestCertDowngrade(t *testing.T) {
-	t.Parallel()
 	// 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{})
@@ -258,7 +255,6 @@ func TestCertDowngrade(t *testing.T) {
 }
 
 func TestCertMismatchCorrection(t *testing.T) {
-	t.Parallel()
 	// 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{})
@@ -326,7 +322,6 @@ func TestCertMismatchCorrection(t *testing.T) {
 }
 
 func TestCrossStackRelaysWork(t *testing.T) {
-	t.Parallel()
 	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version2, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
 	myControl, myVpnIpNet, _, _ := newSimpleServer(cert.Version2, ca, caKey, "me     ", "10.128.0.1/24,fc00::1/64", m{"relay": m{"use_relays": true}})
 	relayControl, relayVpnIpNet, relayUdpAddr, _ := newSimpleServer(cert.Version2, ca, caKey, "relay  ", "10.128.0.128/24,fc00::128/64", m{"relay": m{"am_relay": true}})
@@ -374,7 +369,6 @@ func TestCrossStackRelaysWork(t *testing.T) {
 }
 
 func TestCloseTunnelAuthenticated(t *testing.T) {
-	t.Parallel()
 	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version1, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
 	myControl, myVpnIpNet, myUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "me", "10.128.0.1/24", m{"tunnels": m{"drop_inactive": true, "inactivity_timeout": "5s"}})
 	theirControl, theirVpnIpNet, theirUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "them", "10.128.0.2/24", m{"tunnels": m{"drop_inactive": true, "inactivity_timeout": "10m"}})

firewall_test.go

@@ -1033,7 +1033,7 @@ func TestNewFirewallFromConfig(t *testing.T) {
 	l := test.NewLogger()
 	// Test a bad rule definition
 	c := &dummyCert{}
-	cs, err := newCertState(cert.Version2, nil, c, false, cert.Curve_CURVE25519, nil, "aes")
+	cs, err := newCertState(cert.Version2, nil, c, false, cert.Curve_CURVE25519, nil)
 	require.NoError(t, err)
 
 	conf := config.NewC(test.NewLogger())

go.mod

@@ -22,7 +22,6 @@ require (
 	github.com/stefanberger/go-pkcs11uri v0.0.0-20230803200340-78284954bff6
 	github.com/stretchr/testify v1.11.1
 	github.com/vishvananda/netlink v1.3.1
-	go.uber.org/goleak v1.3.0
 	go.yaml.in/yaml/v3 v3.0.4
 	golang.org/x/crypto v0.50.0
 	golang.org/x/exp v0.0.0-20230725093048-515e97ebf090

handshake/credential.go

@@ -1,57 +0,0 @@
-package handshake
-
-import (
-	"crypto/rand"
-
-	"github.com/flynn/noise"
-	"github.com/slackhq/nebula/cert"
-)
-
-// Credential holds everything needed to participate in a handshake
-// at a given cert version. Version and Curve are read from Cert; the public
-// half of the static keypair likewise comes from Cert.PublicKey().
-type Credential struct {
-	Cert        cert.Certificate  // the certificate
-	Bytes       []byte            // pre-marshaled certificate bytes
-	privateKey  []byte            // static private key (public half lives in Cert)
-	cipherSuite noise.CipherSuite // pre-built cipher suite (DH + cipher + hash)
-}
-
-// NewCredential creates a Credential with all material needed for handshake
-// participation. The cipherSuite should be pre-built by the caller with the
-// appropriate DH function, cipher, and hash.
-func NewCredential(
-	c cert.Certificate,
-	hsBytes []byte,
-	privateKey []byte,
-	cipherSuite noise.CipherSuite,
-) *Credential {
-	return &Credential{
-		Cert:        c,
-		Bytes:       hsBytes,
-		privateKey:  privateKey,
-		cipherSuite: cipherSuite,
-	}
-}
-
-// buildHandshakeState creates a noise.HandshakeState from this credential.
-func (hc *Credential) buildHandshakeState(initiator bool, pattern noise.HandshakePattern) (*noise.HandshakeState, error) {
-	return noise.NewHandshakeState(noise.Config{
-		CipherSuite:           hc.cipherSuite,
-		Random:                rand.Reader,
-		Pattern:               pattern,
-		Initiator:             initiator,
-		StaticKeypair:         noise.DHKey{Private: hc.privateKey, Public: hc.Cert.PublicKey()},
-		PresharedKey:          []byte{},
-		PresharedKeyPlacement: 0,
-	})
-}
-
-// GetCredentialFunc returns the handshake credential for the given version,
-// or nil if that version is not available.
-//
-// Implementations must return credentials drawn from a snapshot stable for
-// the lifetime of any single Machine. The Machine may call this multiple
-// times during a handshake (e.g. when negotiating to the peer's version)
-// and assumes the underlying static keypair is consistent across calls.
-type GetCredentialFunc func(v cert.Version) *Credential

handshake/errors.go

@@ -1,21 +0,0 @@
-package handshake
-
-import "errors"
-
-var (
-	ErrInitiateOnResponder     = errors.New("initiate called on responder")
-	ErrInitiateAlreadyCalled   = errors.New("initiate already called")
-	ErrInitiateNotCalled       = errors.New("initiate must be called before ProcessPacket for initiators")
-	ErrPacketTooShort          = errors.New("packet too short")
-	ErrPublicKeyMismatch       = errors.New("public key mismatch between certificate and handshake")
-	ErrIncompleteHandshake     = errors.New("handshake completed without receiving required content")
-	ErrMachineFailed           = errors.New("handshake machine has failed")
-	ErrUnknownSubtype          = errors.New("unknown handshake subtype")
-	ErrMissingContent          = errors.New("expected handshake content but message was empty")
-	ErrUnexpectedContent       = errors.New("received unexpected handshake content")
-	ErrIndexAllocation         = errors.New("failed to allocate local index")
-	ErrNoCredential            = errors.New("no handshake credential available for cert version")
-	ErrAsymmetricCipherKeys    = errors.New("noise produced only one cipher key")
-	ErrMultiMessageUnsupported = errors.New("multi-message handshake patterns are not yet supported by the manager")
-	ErrSubtypeMismatch         = errors.New("packet subtype does not match handshake machine subtype")
-)

handshake/handshake.proto

@@ -1,29 +0,0 @@
-// This file documents the wire format the nebula handshake speaks. It is
-// not run through protoc; the encoder/decoder in payload.go is hand-written
-// against this shape directly to keep the parser narrow and panic-free.
-//
-// Any change to the wire format must be reflected here, and adding a new
-// field requires updating MarshalPayload / unmarshalPayloadDetails together
-// with the field-uniqueness and wire-type checks in those functions.
-
-syntax = "proto3";
-package nebula.handshake;
-
-message NebulaHandshake {
-  NebulaHandshakeDetails Details = 1;
-  bytes Hmac = 2;
-}
-
-message NebulaHandshakeDetails {
-  bytes Cert = 1;
-  uint32 InitiatorIndex = 2;
-  uint32 ResponderIndex = 3;
-  // Cookie was reserved for an anti-DoS mechanism that was never
-  // implemented. No released version of nebula has ever populated it; the
-  // hand-written parser silently skips it on read.
-  uint64 Cookie = 4 [deprecated = true];
-  uint64 Time = 5;
-  uint32 CertVersion = 8;
-  // reserved for WIP multiport
-  reserved 6, 7;
-}

handshake/helpers_test.go

@@ -1,116 +0,0 @@
-package handshake
-
-import (
-	"net/netip"
-	"testing"
-	"time"
-
-	"github.com/flynn/noise"
-	"github.com/slackhq/nebula/cert"
-	ct "github.com/slackhq/nebula/cert_test"
-	"github.com/slackhq/nebula/header"
-	"github.com/stretchr/testify/require"
-)
-
-// testCertState holds cert material for a test peer.
-type testCertState struct {
-	version cert.Version
-	creds   map[cert.Version]*Credential
-}
-
-func (s *testCertState) getCredential(v cert.Version) *Credential {
-	return s.creds[v]
-}
-
-func newTestCertState(
-	t *testing.T, ca cert.Certificate, caKey []byte, name string, networks []netip.Prefix,
-) *testCertState {
-	return newTestCertStateWithCipher(t, ca, caKey, name, networks, noise.CipherChaChaPoly)
-}
-
-func newTestCertStateWithCipher(
-	t *testing.T, ca cert.Certificate, caKey []byte, name string, networks []netip.Prefix,
-	cipher noise.CipherFunc,
-) *testCertState {
-	t.Helper()
-	c, _, rawPrivKey, _ := ct.NewTestCert(
-		cert.Version2, cert.Curve_CURVE25519, ca, caKey,
-		name, ca.NotBefore(), ca.NotAfter(), networks, nil, nil,
-	)
-
-	priv, _, _, err := cert.UnmarshalPrivateKeyFromPEM(rawPrivKey)
-	require.NoError(t, err)
-
-	hsBytes, err := c.MarshalForHandshakes()
-	require.NoError(t, err)
-
-	ncs := noise.NewCipherSuite(noise.DH25519, cipher, noise.HashSHA256)
-	return &testCertState{
-		version: cert.Version2,
-		creds: map[cert.Version]*Credential{
-			cert.Version2: NewCredential(c, hsBytes, priv, ncs),
-		},
-	}
-}
-
-func testVerifier(pool *cert.CAPool) CertVerifier {
-	return func(c cert.Certificate) (*cert.CachedCertificate, error) {
-		return pool.VerifyCertificate(time.Now(), c)
-	}
-}
-
-func newTestMachine(
-	t *testing.T,
-	cs *testCertState,
-	verifier CertVerifier,
-	initiator bool,
-	localIndex uint32,
-) *Machine {
-	t.Helper()
-	m, err := NewMachine(
-		cs.version, cs.getCredential,
-		verifier, func() (uint32, error) { return localIndex, nil },
-		initiator, header.HandshakeIXPSK0,
-	)
-	require.NoError(t, err)
-	return m
-}
-
-func initiateHandshake(
-	t *testing.T,
-	initCS *testCertState, initVerifier CertVerifier,
-	respCS *testCertState, respVerifier CertVerifier,
-) (initM, respM *Machine, respResult *Result, resp []byte, err error) {
-	t.Helper()
-	initM = newTestMachine(t, initCS, initVerifier, true, 100)
-	msg1, merr := initM.Initiate(nil)
-	require.NoError(t, merr)
-
-	respM = newTestMachine(t, respCS, respVerifier, false, 200)
-	resp, respResult, err = respM.ProcessPacket(nil, msg1)
-	return
-}
-
-func doFullHandshake(
-	t *testing.T, initCS, respCS *testCertState, caPool *cert.CAPool,
-) (initResult, respResult *Result) {
-	t.Helper()
-	v := testVerifier(caPool)
-
-	initM := newTestMachine(t, initCS, v, true, 1000)
-	respM := newTestMachine(t, respCS, v, false, 2000)
-
-	msg1, err := initM.Initiate(nil)
-	require.NoError(t, err)
-
-	resp, respResult, err := respM.ProcessPacket(nil, msg1)
-	require.NoError(t, err)
-	require.NotNil(t, respResult)
-	require.NotEmpty(t, resp)
-
-	_, initResult, err = initM.ProcessPacket(nil, resp)
-	require.NoError(t, err)
-	require.NotNil(t, initResult)
-
-	return initResult, respResult
-}

handshake/machine.go

@@ -1,444 +0,0 @@
-package handshake
-
-import (
-	"bytes"
-	"fmt"
-	"slices"
-	"time"
-
-	"github.com/flynn/noise"
-	"github.com/slackhq/nebula/cert"
-	"github.com/slackhq/nebula/header"
-)
-
-// IndexAllocator is called by the Machine to allocate a local index for the
-// handshake. It is called at most once, when the first outgoing message that
-// carries a payload is built.
-//
-// Implementations MUST NOT return 0. Zero is reserved as a sentinel meaning
-// "no index assigned" on the wire and in the payload-presence checks. If an
-// allocator ever returned 0, a legitimate handshake's payload could be
-// indistinguishable from an empty one and would be rejected.
-type IndexAllocator func() (uint32, error)
-
-// CertVerifier is called by the Machine after reconstructing the peer's
-// certificate from the handshake. The verifier performs all validation
-// (CA trust, expiry, policy checks, allow lists).
-type CertVerifier func(cert.Certificate) (*cert.CachedCertificate, error)
-
-// Result contains the results of a successful handshake.
-// Returned by ProcessPacket when the handshake is complete.
-type Result struct {
-	EKey          *noise.CipherState
-	DKey          *noise.CipherState
-	MyCert        cert.Certificate
-	RemoteCert    *cert.CachedCertificate
-	RemoteIndex   uint32
-	LocalIndex    uint32
-	HandshakeTime uint64
-	MessageIndex  uint64 // number of messages exchanged during the handshake
-	Initiator     bool
-}
-
-// Machine drives a Noise handshake through N messages. It handles Noise
-// protocol operations, certificate reconstruction, and payload encoding.
-// Certificate validation is delegated to the caller via CertVerifier.
-//
-// A Machine is not safe for concurrent use. The caller must ensure that
-// Initiate and ProcessPacket are not called concurrently.
-//
-// Error contract: when ProcessPacket or Initiate returns an error, callers
-// must check Failed() to decide what to do next. If Failed() is false the
-// underlying noise state was not advanced (the packet was rejected before
-// ReadMessage took effect, or the rejection is non-fatal like a stale
-// retransmit) and the Machine can accept another packet. If Failed() is
-// true the Machine is unrecoverable and the caller must abandon it.
-type Machine struct {
-	hs             *noise.HandshakeState
-	getCred        GetCredentialFunc
-	allocIndex     IndexAllocator
-	verifier       CertVerifier
-	result         *Result
-	msgs           []msgFlags
-	myVersion      cert.Version
-	subtype        header.MessageSubType
-	indexAllocated bool
-	remoteCertSet  bool
-	payloadSet     bool
-	failed         bool
-}
-
-// NewMachine creates a handshake state machine. The subtype determines both
-// the noise pattern and the per-message content layout. The credential for
-// `version` is fetched via getCred and used to seed the noise.HandshakeState.
-// IndexAllocator is called lazily when the first outgoing payload is built.
-func NewMachine(
-	version cert.Version,
-	getCred GetCredentialFunc,
-	verifier CertVerifier,
-	allocIndex IndexAllocator,
-	initiator bool,
-	subtype header.MessageSubType,
-) (*Machine, error) {
-	info, err := subtypeInfoFor(subtype)
-	if err != nil {
-		return nil, err
-	}
-
-	cred := getCred(version)
-	if cred == nil {
-		return nil, fmt.Errorf("%w: %v", ErrNoCredential, version)
-	}
-
-	hs, err := cred.buildHandshakeState(initiator, info.pattern)
-	if err != nil {
-		return nil, fmt.Errorf("build noise state: %w", err)
-	}
-
-	return &Machine{
-		hs:         hs,
-		subtype:    subtype,
-		msgs:       info.msgs,
-		getCred:    getCred,
-		allocIndex: allocIndex,
-		verifier:   verifier,
-		myVersion:  version,
-		result: &Result{
-			Initiator: initiator,
-		},
-	}, nil
-}
-
-// Failed returns true if the Machine is in an unrecoverable state.
-func (m *Machine) Failed() bool {
-	return m.failed
-}
-
-// Subtype returns the handshake subtype this Machine was built for.
-func (m *Machine) Subtype() header.MessageSubType {
-	return m.subtype
-}
-
-// MessageIndex returns the noise handshake message index, which equals the
-// wire counter of the most recently sent or received message.
-func (m *Machine) MessageIndex() int {
-	return m.hs.MessageIndex()
-}
-
-// requireComplete checks that both a peer cert and payload have been received.
-// Marks the machine as failed if not.
-func (m *Machine) requireComplete() error {
-	if !m.payloadSet || !m.remoteCertSet {
-		m.failed = true
-		return ErrIncompleteHandshake
-	}
-	return nil
-}
-
-// myMsgFlags returns the flags for the current outgoing message.
-func (m *Machine) myMsgFlags() msgFlags {
-	idx := m.hs.MessageIndex()
-	if idx < len(m.msgs) {
-		return m.msgs[idx]
-	}
-	return msgFlags{}
-}
-
-// peerMsgFlags returns the flags for the message we just read.
-func (m *Machine) peerMsgFlags() msgFlags {
-	idx := m.hs.MessageIndex() - 1
-	if idx >= 0 && idx < len(m.msgs) {
-		return m.msgs[idx]
-	}
-	return msgFlags{}
-}
-
-// Initiate produces the first handshake message. Only valid for initiators,
-// and must be called exactly once before ProcessPacket.
-//
-// out is a destination buffer the message is appended to and returned. Pass
-// nil to allocate fresh, or pass a re-used buffer sliced to length 0 (e.g.
-// buf[:0]) with sufficient capacity to avoid allocation.
-//
-// An error return may not indicate a fatal condition, check Failed() to
-// determine if the Machine can still be used.
-func (m *Machine) Initiate(out []byte) ([]byte, error) {
-	if m.failed {
-		return nil, ErrMachineFailed
-	}
-	if !m.result.Initiator {
-		m.failed = true
-		return nil, ErrInitiateOnResponder
-	}
-	if m.hs.MessageIndex() != 0 {
-		m.failed = true
-		return nil, ErrInitiateAlreadyCalled
-	}
-
-	// At MessageIndex=0 with RemoteIndex still zero, buildResponse produces
-	// header counter 1 and remote index 0, which is what the initial message needs.
-	out, _, _, err := m.buildResponse(out)
-	if err != nil {
-		m.failed = true
-		return nil, err
-	}
-	return out, nil
-}
-
-// ProcessPacket handles an incoming handshake message. It advances the Noise
-// state, validates the peer certificate via the verifier, and optionally
-// produces a response.
-//
-// out is a destination buffer the response is appended to and returned. Pass
-// nil to allocate fresh, or pass a re-used buffer sliced to length 0 (e.g.
-// buf[:0]) with sufficient capacity to avoid allocation. The returned slice
-// is nil when no outgoing message is produced (handshake complete on this
-// side, or final message of a multi-message pattern).
-//
-// Returns a non-nil Result when the handshake is complete.
-// An error return may not indicate a fatal condition, check Failed() to
-// determine if the Machine can still be used.
-func (m *Machine) ProcessPacket(out, packet []byte) ([]byte, *Result, error) {
-	if m.failed {
-		return nil, nil, ErrMachineFailed
-	}
-	if len(packet) < header.Len {
-		return nil, nil, ErrPacketTooShort
-	}
-	// Reject packets whose subtype doesn't match the one this Machine was
-	// built for. A pending handshake that suddenly receives a different
-	// subtype on its index is either a stray packet that matched by chance
-	// or a peer protocol violation; drop it without failing the Machine so
-	// the legitimate retransmit can still complete.
-	if header.MessageSubType(packet[1]) != m.subtype {
-		return nil, nil, ErrSubtypeMismatch
-	}
-	if m.result.Initiator && m.hs.MessageIndex() == 0 {
-		m.failed = true
-		return nil, nil, ErrInitiateNotCalled
-	}
-
-	// The (eKey, dKey) ordering here is correct for IX, where the initiator
-	// completes the handshake by reading the responder's stage-2 message.
-	// noise returns (cs1, cs2) where cs1 is the initiator->responder cipher.
-	// For 3-message patterns where a responder finishes by reading the final
-	// message, this ordering would be wrong; revisit when XX/pqIX lands.
-	msg, eKey, dKey, err := m.hs.ReadMessage(nil, packet[header.Len:])
-	if err != nil {
-		// Noise ReadMessage failed. The noise library checkpoints and rolls back
-		// on failure, so the Machine is still alive. The caller can retry with
-		// a different packet.
-		return nil, nil, fmt.Errorf("noise ReadMessage: %w", err)
-	}
-
-	// From here on, noise state has advanced. Any error is fatal.
-	flags := m.peerMsgFlags()
-
-	if err := m.processPayload(msg, flags); err != nil {
-		return nil, nil, err
-	}
-
-	// If ReadMessage derived keys, the handshake is complete. Noise should
-	// always produce both keys together; asymmetry is a protocol invariant
-	// violation.
-	if eKey != nil || dKey != nil {
-		if eKey == nil || dKey == nil {
-			m.failed = true
-			return nil, nil, ErrAsymmetricCipherKeys
-		}
-		if err := m.requireComplete(); err != nil {
-			return nil, nil, err
-		}
-		return nil, m.completed(eKey, dKey), nil
-	}
-
-	// ReadMessage didn't complete, produce the next outgoing message
-	out, dk, ek, err := m.buildResponse(out)
-	if err != nil {
-		m.failed = true
-		return nil, nil, err
-	}
-
-	if ek != nil || dk != nil {
-		if ek == nil || dk == nil {
-			m.failed = true
-			return nil, nil, ErrAsymmetricCipherKeys
-		}
-		if err := m.requireComplete(); err != nil {
-			return nil, nil, err
-		}
-		return out, m.completed(ek, dk), nil
-	}
-
-	return out, nil, nil
-}
-
-func (m *Machine) completed(eKey, dKey *noise.CipherState) *Result {
-	m.result.EKey = eKey
-	m.result.DKey = dKey
-	m.result.MessageIndex = uint64(m.hs.MessageIndex())
-	return m.result
-}
-
-func (m *Machine) processPayload(msg []byte, flags msgFlags) error {
-	if len(msg) == 0 {
-		if flags.expectsPayload || flags.expectsCert {
-			m.failed = true
-			return ErrMissingContent
-		}
-		return nil
-	}
-
-	payload, err := UnmarshalPayload(msg)
-	if err != nil {
-		m.failed = true
-		return fmt.Errorf("unmarshal handshake: %w", err)
-	}
-
-	// Assert the payload contains exactly what we expect
-	hasPayloadData := payload.InitiatorIndex != 0 || payload.ResponderIndex != 0 || payload.Time != 0
-	if hasPayloadData != flags.expectsPayload {
-		m.failed = true
-		return ErrUnexpectedContent
-	}
-
-	hasCertData := len(payload.Cert) > 0
-	if hasCertData != flags.expectsCert {
-		m.failed = true
-		return ErrUnexpectedContent
-	}
-
-	// Process payload
-	if flags.expectsPayload {
-		if m.result.Initiator {
-			m.result.RemoteIndex = payload.ResponderIndex
-		} else {
-			m.result.RemoteIndex = payload.InitiatorIndex
-		}
-		m.result.HandshakeTime = payload.Time
-		m.payloadSet = true
-	}
-
-	// Process certificate
-	if flags.expectsCert {
-		if err := m.validateCert(payload); err != nil {
-			return err
-		}
-	}
-
-	return nil
-}
-
-func (m *Machine) validateCert(payload Payload) error {
-	cred := m.getCred(m.myVersion)
-	if cred == nil {
-		m.failed = true
-		return fmt.Errorf("%w: %v", ErrNoCredential, m.myVersion)
-	}
-	rc, err := cert.Recombine(
-		cert.Version(payload.CertVersion),
-		payload.Cert,
-		m.hs.PeerStatic(),
-		cred.Cert.Curve(),
-	)
-	if err != nil {
-		m.failed = true
-		return fmt.Errorf("recombine cert: %w", err)
-	}
-
-	if !bytes.Equal(rc.PublicKey(), m.hs.PeerStatic()) {
-		m.failed = true
-		return ErrPublicKeyMismatch
-	}
-
-	// Version negotiation, if the peer sent a different version and we have it, switch
-	if rc.Version() != m.myVersion {
-		if m.getCred(rc.Version()) != nil {
-			m.myVersion = rc.Version()
-		}
-	}
-
-	verified, err := m.verifier(rc)
-	if err != nil {
-		m.failed = true
-		return fmt.Errorf("verify cert: %w", err)
-	}
-
-	m.result.RemoteCert = verified
-	m.remoteCertSet = true
-	return nil
-}
-
-func (m *Machine) marshalOutgoing(flags msgFlags) ([]byte, error) {
-	if !flags.expectsPayload && !flags.expectsCert {
-		return nil, nil
-	}
-
-	var p Payload
-	if flags.expectsPayload {
-		if !m.indexAllocated {
-			index, err := m.allocIndex()
-			if err != nil {
-				return nil, fmt.Errorf("%w: %w", ErrIndexAllocation, err)
-			}
-			m.result.LocalIndex = index
-			m.indexAllocated = true
-		}
-
-		if m.result.Initiator {
-			p.InitiatorIndex = m.result.LocalIndex
-		} else {
-			p.ResponderIndex = m.result.LocalIndex
-			p.InitiatorIndex = m.result.RemoteIndex
-		}
-		p.Time = uint64(time.Now().UnixNano())
-	}
-	if flags.expectsCert {
-		cred := m.getCred(m.myVersion)
-		if cred == nil {
-			return nil, fmt.Errorf("%w: %v", ErrNoCredential, m.myVersion)
-		}
-		p.Cert = cred.Bytes
-		p.CertVersion = uint32(cred.Cert.Version())
-		m.result.MyCert = cred.Cert
-	}
-
-	return MarshalPayload(nil, p), nil
-}
-
-func (m *Machine) buildResponse(out []byte) ([]byte, *noise.CipherState, *noise.CipherState, error) {
-	flags := m.myMsgFlags()
-	hsBytes, err := m.marshalOutgoing(flags)
-	if err != nil {
-		return nil, nil, nil, err
-	}
-
-	// Extend out by header.Len to make room for the header. slices.Grow is a
-	// no-op when the cap is already sufficient (the zero-copy case where the
-	// caller passed a pre-sized buffer). header.Encode overwrites the new
-	// bytes, so they don't need to be zeroed.
-	start := len(out)
-	out = slices.Grow(out, header.Len)[:start+header.Len]
-	header.Encode(
-		out[start:],
-		header.Version, header.Handshake, m.subtype,
-		m.result.RemoteIndex,
-		uint64(m.hs.MessageIndex()+1),
-	)
-
-	// noise.WriteMessage appends the encrypted handshake message to out,
-	// reusing capacity when present.
-	//
-	// The (dKey, eKey) ordering here is correct for IX, where the responder
-	// completes the handshake by writing the stage-2 message. noise returns
-	// (cs1, cs2) where cs1 is the initiator->responder cipher (which is the
-	// responder's decrypt key). For 3-message patterns where an initiator
-	// finishes by writing the final message, this ordering would be wrong;
-	// revisit when XX/pqIX lands.
-	out, dKey, eKey, err := m.hs.WriteMessage(out, hsBytes)
-	if err != nil {
-		return nil, nil, nil, fmt.Errorf("noise WriteMessage: %w", err)
-	}
-
-	return out, dKey, eKey, nil
-}

handshake/machine_test.go

@@ -1,662 +0,0 @@
-package handshake
-
-import (
-	"net/netip"
-	"testing"
-	"time"
-
-	"github.com/flynn/noise"
-	"github.com/slackhq/nebula/cert"
-	ct "github.com/slackhq/nebula/cert_test"
-	"github.com/slackhq/nebula/header"
-	"github.com/slackhq/nebula/noiseutil"
-	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
-)
-
-func TestMachineIXHappyPath(t *testing.T) {
-	ca, _, caKey, _ := ct.NewTestCaCert(
-		cert.Version2, cert.Curve_CURVE25519, time.Time{}, time.Time{}, nil, nil, nil,
-	)
-	caPool := ct.NewTestCAPool(ca)
-
-	initCS := newTestCertState(t, ca, caKey, "initiator", []netip.Prefix{netip.MustParsePrefix("10.0.0.1/24")})
-	respCS := newTestCertState(t, ca, caKey, "responder", []netip.Prefix{netip.MustParsePrefix("10.0.0.2/24")})
-
-	initR, respR := doFullHandshake(t, initCS, respCS, caPool)
-
-	assert.Equal(t, "responder", initR.RemoteCert.Certificate.Name())
-	assert.Equal(t, "initiator", respR.RemoteCert.Certificate.Name())
-
-	assert.Equal(t, uint32(1000), initR.LocalIndex)
-	assert.Equal(t, uint32(2000), initR.RemoteIndex)
-	assert.Equal(t, uint32(2000), respR.LocalIndex)
-	assert.Equal(t, uint32(1000), respR.RemoteIndex)
-
-	assert.Equal(t, uint64(2), initR.MessageIndex, "IX has 2 messages")
-	assert.Equal(t, uint64(2), respR.MessageIndex, "IX has 2 messages")
-
-	ct1, err := initR.EKey.Encrypt(nil, nil, []byte("hello"))
-	require.NoError(t, err)
-	pt1, err := respR.DKey.Decrypt(nil, nil, ct1)
-	require.NoError(t, err)
-	assert.Equal(t, []byte("hello"), pt1)
-
-	ct2, err := respR.EKey.Encrypt(nil, nil, []byte("world"))
-	require.NoError(t, err)
-	pt2, err := initR.DKey.Decrypt(nil, nil, ct2)
-	require.NoError(t, err)
-	assert.Equal(t, []byte("world"), pt2)
-}
-
-func TestMachineInitiateErrors(t *testing.T) {
-	ca, _, caKey, _ := ct.NewTestCaCert(
-		cert.Version2, cert.Curve_CURVE25519, time.Time{}, time.Time{}, nil, nil, nil,
-	)
-	caPool := ct.NewTestCAPool(ca)
-	cs := newTestCertState(t, ca, caKey, "test", []netip.Prefix{netip.MustParsePrefix("10.0.0.1/24")})
-	v := testVerifier(caPool)
-
-	t.Run("initiate on responder", func(t *testing.T) {
-		m := newTestMachine(t, cs, v, false, 100)
-		_, err := m.Initiate(nil)
-		require.ErrorIs(t, err, ErrInitiateOnResponder)
-		assert.True(t, m.Failed())
-	})
-
-	t.Run("initiate called twice", func(t *testing.T) {
-		m := newTestMachine(t, cs, v, true, 100)
-		_, err := m.Initiate(nil)
-		require.NoError(t, err)
-		_, err = m.Initiate(nil)
-		require.ErrorIs(t, err, ErrInitiateAlreadyCalled)
-		assert.True(t, m.Failed())
-	})
-
-	t.Run("process packet before initiate on initiator", func(t *testing.T) {
-		m := newTestMachine(t, cs, v, true, 100)
-		_, _, err := m.ProcessPacket(nil, make([]byte, 100))
-		require.ErrorIs(t, err, ErrInitiateNotCalled)
-		assert.True(t, m.Failed())
-	})
-
-	t.Run("calling failed machine", func(t *testing.T) {
-		m := newTestMachine(t, cs, v, false, 100)
-		_, err := m.Initiate(nil) // fails: responder
-		require.Error(t, err)
-		_, err = m.Initiate(nil) // fails: already failed
-		require.ErrorIs(t, err, ErrMachineFailed)
-	})
-}
-
-func TestMachineProcessPacketErrors(t *testing.T) {
-	ca, _, caKey, _ := ct.NewTestCaCert(
-		cert.Version2, cert.Curve_CURVE25519, time.Time{}, time.Time{}, nil, nil, nil,
-	)
-	caPool := ct.NewTestCAPool(ca)
-	cs := newTestCertState(t, ca, caKey, "test", []netip.Prefix{netip.MustParsePrefix("10.0.0.1/24")})
-	v := testVerifier(caPool)
-
-	t.Run("packet too short", func(t *testing.T) {
-		m := newTestMachine(t, cs, v, false, 100)
-		_, _, err := m.ProcessPacket(nil, []byte{1, 2, 3})
-		require.ErrorIs(t, err, ErrPacketTooShort)
-		assert.False(t, m.Failed(), "short packet should not kill machine")
-	})
-
-	t.Run("noise decryption failure is recoverable", func(t *testing.T) {
-		initCS := newTestCertState(t, ca, caKey, "init", []netip.Prefix{netip.MustParsePrefix("10.0.0.1/24")})
-		initM := newTestMachine(t, initCS, v, true, 100)
-		msg1, err := initM.Initiate(nil)
-		require.NoError(t, err)
-
-		respM := newTestMachine(t, cs, v, false, 200)
-		resp, _, err := respM.ProcessPacket(nil, msg1)
-		require.NoError(t, err)
-
-		corrupted := make([]byte, len(resp))
-		copy(corrupted, resp)
-		for i := header.Len; i < len(corrupted); i++ {
-			corrupted[i] ^= 0xff
-		}
-		_, _, err = initM.ProcessPacket(nil, corrupted)
-		require.Error(t, err)
-		assert.False(t, initM.Failed(), "noise failure should be recoverable")
-
-		// And the machine should still complete a real handshake afterward.
-		_, result, err := initM.ProcessPacket(nil, resp)
-		require.NoError(t, err)
-		require.NotNil(t, result, "initiator should complete on the legitimate response")
-	})
-
-	t.Run("invalid cert is fatal", func(t *testing.T) {
-		otherCA, _, otherCAKey, _ := ct.NewTestCaCert(
-			cert.Version2, cert.Curve_CURVE25519, time.Time{}, time.Time{}, nil, nil, nil,
-		)
-		otherCS := newTestCertState(t, otherCA, otherCAKey, "other", []netip.Prefix{netip.MustParsePrefix("10.0.0.2/24")})
-
-		initM := newTestMachine(t, otherCS, testVerifier(ct.NewTestCAPool(otherCA)), true, 100)
-		msg1, err := initM.Initiate(nil)
-		require.NoError(t, err)
-
-		respM := newTestMachine(t, cs, v, false, 200)
-		_, _, err = respM.ProcessPacket(nil, msg1)
-		require.Error(t, err)
-		assert.True(t, respM.Failed(), "cert validation failure should kill machine")
-	})
-
-	t.Run("subtype mismatch is recoverable", func(t *testing.T) {
-		initCS := newTestCertState(t, ca, caKey, "init", []netip.Prefix{netip.MustParsePrefix("10.0.0.1/24")})
-		initM := newTestMachine(t, initCS, v, true, 100)
-		msg1, err := initM.Initiate(nil)
-		require.NoError(t, err)
-
-		// Mutate the subtype byte (offset 1 in the header) to a value the
-		// responder Machine wasn't built for.
-		bad := make([]byte, len(msg1))
-		copy(bad, msg1)
-		bad[1] = 0xff
-
-		respM := newTestMachine(t, cs, v, false, 200)
-		_, _, err = respM.ProcessPacket(nil, bad)
-		require.ErrorIs(t, err, ErrSubtypeMismatch)
-		assert.False(t, respM.Failed(), "subtype mismatch should not kill the machine")
-
-		// And the machine should still complete a real handshake afterward.
-		resp, result, err := respM.ProcessPacket(nil, msg1)
-		require.NoError(t, err)
-		require.NotNil(t, result, "responder should complete on the legitimate stage-1 packet")
-		assert.NotEmpty(t, resp, "responder should produce a stage-2 reply")
-	})
-}
-
-// TestMachineProcessPayload exercises processPayload's internal validation
-// directly. Most of these failure modes can't be reached black-box once the
-// subtype check at the top of ProcessPacket gates external callers, so we
-// drive them by hand here for coverage.
-func TestMachineProcessPayload(t *testing.T) {
-	ca, _, caKey, _ := ct.NewTestCaCert(
-		cert.Version2, cert.Curve_CURVE25519, time.Time{}, time.Time{}, nil, nil, nil,
-	)
-	caPool := ct.NewTestCAPool(ca)
-	cs := newTestCertState(t, ca, caKey, "test", []netip.Prefix{netip.MustParsePrefix("10.0.0.1/24")})
-	v := testVerifier(caPool)
-
-	t.Run("empty message with expects fails", func(t *testing.T) {
-		m := newTestMachine(t, cs, v, false, 100)
-		err := m.processPayload(nil, msgFlags{expectsPayload: true, expectsCert: true})
-		require.ErrorIs(t, err, ErrMissingContent)
-		assert.True(t, m.Failed())
-	})
-
-	t.Run("empty message with no expects passes", func(t *testing.T) {
-		m := newTestMachine(t, cs, v, false, 100)
-		err := m.processPayload(nil, msgFlags{})
-		require.NoError(t, err)
-		assert.False(t, m.Failed())
-	})
-
-	t.Run("malformed protobuf is fatal", func(t *testing.T) {
-		m := newTestMachine(t, cs, v, false, 100)
-		err := m.processPayload([]byte{0xff, 0xff, 0xff}, msgFlags{expectsPayload: true, expectsCert: true})
-		require.Error(t, err)
-		assert.True(t, m.Failed())
-	})
-
-	t.Run("unexpected payload data is fatal", func(t *testing.T) {
-		m := newTestMachine(t, cs, v, false, 100)
-		// A payload with index data when none was expected.
-		bytes := MarshalPayload(nil, Payload{InitiatorIndex: 42, Time: 1})
-		err := m.processPayload(bytes, msgFlags{expectsPayload: false, expectsCert: false})
-		require.ErrorIs(t, err, ErrUnexpectedContent)
-		assert.True(t, m.Failed())
-	})
-
-	t.Run("unexpected cert data is fatal", func(t *testing.T) {
-		m := newTestMachine(t, cs, v, false, 100)
-		// A payload with cert when none was expected.
-		bytes := MarshalPayload(nil, Payload{Cert: []byte{1, 2, 3}, CertVersion: 2})
-		err := m.processPayload(bytes, msgFlags{expectsPayload: false, expectsCert: false})
-		require.ErrorIs(t, err, ErrUnexpectedContent)
-		assert.True(t, m.Failed())
-	})
-
-	t.Run("missing payload data when expected is fatal", func(t *testing.T) {
-		m := newTestMachine(t, cs, v, false, 100)
-		// Cert present, but no index/time fields.
-		bytes := MarshalPayload(nil, Payload{Cert: []byte{1, 2, 3}, CertVersion: 2})
-		err := m.processPayload(bytes, msgFlags{expectsPayload: true, expectsCert: true})
-		require.ErrorIs(t, err, ErrUnexpectedContent)
-		assert.True(t, m.Failed())
-	})
-}
-
-// TestMachineRequireComplete checks the fail-on-incomplete-handshake path
-// directly. Like processPayload above this isn't reachable from a normal IX
-// flow, so we drive it by hand.
-func TestMachineRequireComplete(t *testing.T) {
-	ca, _, caKey, _ := ct.NewTestCaCert(
-		cert.Version2, cert.Curve_CURVE25519, time.Time{}, time.Time{}, nil, nil, nil,
-	)
-	caPool := ct.NewTestCAPool(ca)
-	cs := newTestCertState(t, ca, caKey, "test", []netip.Prefix{netip.MustParsePrefix("10.0.0.1/24")})
-	v := testVerifier(caPool)
-
-	t.Run("missing both fails", func(t *testing.T) {
-		m := newTestMachine(t, cs, v, false, 100)
-		err := m.requireComplete()
-		require.ErrorIs(t, err, ErrIncompleteHandshake)
-		assert.True(t, m.Failed())
-	})
-
-	t.Run("payload only fails", func(t *testing.T) {
-		m := newTestMachine(t, cs, v, false, 100)
-		m.payloadSet = true
-		err := m.requireComplete()
-		require.ErrorIs(t, err, ErrIncompleteHandshake)
-		assert.True(t, m.Failed())
-	})
-
-	t.Run("cert only fails", func(t *testing.T) {
-		m := newTestMachine(t, cs, v, false, 100)
-		m.remoteCertSet = true
-		err := m.requireComplete()
-		require.ErrorIs(t, err, ErrIncompleteHandshake)
-		assert.True(t, m.Failed())
-	})
-
-	t.Run("both set passes", func(t *testing.T) {
-		m := newTestMachine(t, cs, v, false, 100)
-		m.payloadSet = true
-		m.remoteCertSet = true
-		err := m.requireComplete()
-		require.NoError(t, err)
-		assert.False(t, m.Failed())
-	})
-}
-
-func TestMachineAESCipher(t *testing.T) {
-	ca, _, caKey, _ := ct.NewTestCaCert(
-		cert.Version2, cert.Curve_CURVE25519, time.Time{}, time.Time{}, nil, nil, nil,
-	)
-	caPool := ct.NewTestCAPool(ca)
-
-	initCS := newTestCertStateWithCipher(
-		t, ca, caKey, "init",
-		[]netip.Prefix{netip.MustParsePrefix("10.0.0.1/24")},
-		noiseutil.CipherAESGCM,
-	)
-	respCS := newTestCertStateWithCipher(
-		t, ca, caKey, "resp",
-		[]netip.Prefix{netip.MustParsePrefix("10.0.0.2/24")},
-		noiseutil.CipherAESGCM,
-	)
-
-	initR, respR := doFullHandshake(t, initCS, respCS, caPool)
-
-	ct1, err := initR.EKey.Encrypt(nil, nil, []byte("works"))
-	require.NoError(t, err)
-	pt1, err := respR.DKey.Decrypt(nil, nil, ct1)
-	require.NoError(t, err)
-	assert.Equal(t, []byte("works"), pt1)
-
-	ct2, err := respR.EKey.Encrypt(nil, nil, []byte("back"))
-	require.NoError(t, err)
-	pt2, err := initR.DKey.Decrypt(nil, nil, ct2)
-	require.NoError(t, err)
-	assert.Equal(t, []byte("back"), pt2)
-}
-
-func TestResultFields(t *testing.T) {
-	ca, _, caKey, _ := ct.NewTestCaCert(
-		cert.Version2, cert.Curve_CURVE25519, time.Time{}, time.Time{}, nil, nil, nil,
-	)
-	caPool := ct.NewTestCAPool(ca)
-	initCS := newTestCertState(t, ca, caKey, "init", []netip.Prefix{netip.MustParsePrefix("10.0.0.1/24")})
-	respCS := newTestCertState(t, ca, caKey, "resp", []netip.Prefix{netip.MustParsePrefix("10.0.0.2/24")})
-
-	initR, respR := doFullHandshake(t, initCS, respCS, caPool)
-
-	assert.True(t, initR.Initiator)
-	assert.False(t, respR.Initiator)
-	assert.NotZero(t, initR.HandshakeTime)
-	assert.NotZero(t, respR.HandshakeTime)
-	assert.NotNil(t, initR.RemoteCert)
-	assert.NotNil(t, respR.RemoteCert)
-}
-
-func TestMachineBufferReuse(t *testing.T) {
-	ca, _, caKey, _ := ct.NewTestCaCert(
-		cert.Version2, cert.Curve_CURVE25519, time.Time{}, time.Time{}, nil, nil, nil,
-	)
-	caPool := ct.NewTestCAPool(ca)
-	initCS := newTestCertState(t, ca, caKey, "init", []netip.Prefix{netip.MustParsePrefix("10.0.0.1/24")})
-	respCS := newTestCertState(t, ca, caKey, "resp", []netip.Prefix{netip.MustParsePrefix("10.0.0.2/24")})
-	v := testVerifier(caPool)
-
-	initM := newTestMachine(t, initCS, v, true, 1000)
-	respM := newTestMachine(t, respCS, v, false, 2000)
-
-	msg1, err := initM.Initiate(nil)
-	require.NoError(t, err)
-
-	t.Run("response writes into provided buffer", func(t *testing.T) {
-		buf := make([]byte, 0, 4096)
-		resp, result, err := respM.ProcessPacket(buf, msg1)
-		require.NoError(t, err)
-		require.NotNil(t, result)
-
-		assert.NotEmpty(t, resp, "response should have content")
-		assert.Equal(t, &buf[:1][0], &resp[:1][0],
-			"response should reuse the provided buffer's backing array")
-	})
-
-	t.Run("initiate writes into provided buffer", func(t *testing.T) {
-		initM2 := newTestMachine(t, initCS, v, true, 3000)
-		buf := make([]byte, 0, 4096)
-		msg, err := initM2.Initiate(buf)
-		require.NoError(t, err)
-
-		assert.NotEmpty(t, msg, "initiate should have content")
-		assert.Equal(t, &buf[:1][0], &msg[:1][0],
-			"initiate should reuse the provided buffer's backing array")
-	})
-
-	t.Run("nil out still works", func(t *testing.T) {
-		initM2 := newTestMachine(t, initCS, v, true, 4000)
-		respM2 := newTestMachine(t, respCS, v, false, 5000)
-
-		msg1, err := initM2.Initiate(nil)
-		require.NoError(t, err)
-
-		resp, _, err := respM2.ProcessPacket(nil, msg1)
-		require.NoError(t, err)
-
-		out, result, err := initM2.ProcessPacket(nil, resp)
-		require.NoError(t, err)
-		assert.NotNil(t, result)
-		assert.Nil(t, out, "initiator should have no response for IX msg2")
-	})
-}
-
-func TestMachineMsgIndexTracking(t *testing.T) {
-	ca, _, caKey, _ := ct.NewTestCaCert(
-		cert.Version2, cert.Curve_CURVE25519, time.Time{}, time.Time{}, nil, nil, nil,
-	)
-	caPool := ct.NewTestCAPool(ca)
-	initCS := newTestCertState(t, ca, caKey, "init", []netip.Prefix{netip.MustParsePrefix("10.0.0.1/24")})
-	respCS := newTestCertState(t, ca, caKey, "resp", []netip.Prefix{netip.MustParsePrefix("10.0.0.2/24")})
-	v := testVerifier(caPool)
-
-	initM := newTestMachine(t, initCS, v, true, 100)
-	respM := newTestMachine(t, respCS, v, false, 200)
-
-	msg1, err := initM.Initiate(nil)
-	require.NoError(t, err)
-
-	resp1, result1, err := respM.ProcessPacket(nil, msg1)
-	require.NoError(t, err)
-	assert.NotNil(t, result1)
-
-	_, result2, err := initM.ProcessPacket(nil, resp1)
-	require.NoError(t, err)
-	assert.NotNil(t, result2)
-}
-
-func TestMachineThreeMessagePattern(t *testing.T) {
-	registerTestXXInfo(t)
-
-	// Use HandshakeXX (3 messages) to verify the Machine handles multi-message
-	// patterns correctly. XX flow:
-	//   msg1 (I->R): [E]           - payload only, no cert
-	//   msg2 (R->I): [E, ee, S, es] - payload + cert
-	//   msg3 (I->R): [S, se]       - cert only (no payload, not first two)
-
-	ca, _, caKey, _ := ct.NewTestCaCert(
-		cert.Version2, cert.Curve_CURVE25519, time.Time{}, time.Time{}, nil, nil, nil,
-	)
-	caPool := ct.NewTestCAPool(ca)
-	v := testVerifier(caPool)
-
-	initCS := newTestCertState(t, ca, caKey, "init", []netip.Prefix{netip.MustParsePrefix("10.0.0.1/24")})
-	respCS := newTestCertState(t, ca, caKey, "resp", []netip.Prefix{netip.MustParsePrefix("10.0.0.2/24")})
-
-	initM, err := NewMachine(
-		cert.Version2,
-		initCS.getCredential, v,
-		func() (uint32, error) { return 1000, nil },
-		true, header.HandshakeXXPSK0,
-	)
-	require.NoError(t, err)
-
-	respM, err := NewMachine(
-		cert.Version2,
-		respCS.getCredential, v,
-		func() (uint32, error) { return 2000, nil },
-		false, header.HandshakeXXPSK0,
-	)
-	require.NoError(t, err)
-
-	// msg1: initiator -> responder (E only, no cert)
-	msg1, err := initM.Initiate(nil)
-	require.NoError(t, err)
-	assert.NotEmpty(t, msg1)
-
-	// Responder processes msg1, should not complete yet, should produce msg2
-	msg2, result, err := respM.ProcessPacket(nil, msg1)
-	require.NoError(t, err)
-	assert.Nil(t, result, "XX should not complete on msg1")
-	assert.NotEmpty(t, msg2, "responder should produce msg2")
-
-	// Initiator processes msg2: gets responder's cert, produces msg3, and
-	// completes (WriteMessage for msg3 derives keys)
-	msg3, initResult, err := initM.ProcessPacket(nil, msg2)
-	require.NoError(t, err)
-	require.NotNil(t, initResult, "XX initiator should complete after reading msg2 and writing msg3")
-	assert.NotEmpty(t, msg3, "initiator should produce msg3")
-	assert.Equal(t, "resp", initResult.RemoteCert.Certificate.Name())
-
-	// Responder processes msg3: gets initiator's cert and completes
-	_, respResult, err := respM.ProcessPacket(nil, msg3)
-	require.NoError(t, err)
-	require.NotNil(t, respResult, "XX responder should complete on msg3")
-	assert.Equal(t, "init", respResult.RemoteCert.Certificate.Name())
-
-	assert.Equal(t, uint64(3), initResult.MessageIndex, "XX has 3 messages")
-	assert.Equal(t, uint64(3), respResult.MessageIndex, "XX has 3 messages")
-
-	// Verify keys work
-	ct1, err := initResult.EKey.Encrypt(nil, nil, []byte("three messages"))
-	require.NoError(t, err)
-	pt1, err := respResult.DKey.Decrypt(nil, nil, ct1)
-	require.NoError(t, err)
-	assert.Equal(t, []byte("three messages"), pt1)
-}
-
-// NOTE: ErrIncompleteHandshake is tested implicitly. It can't be triggered with
-// IX since the cert is always in the payload. A 3-message pattern test (HybridIX)
-// should exercise the case where cert arrives in msg3 and verify that completing
-// without it fails.
-
-func TestMachineExpiredCert(t *testing.T) {
-	ca, _, caKey, _ := ct.NewTestCaCert(
-		cert.Version2, cert.Curve_CURVE25519,
-		time.Now().Add(-24*time.Hour), time.Now().Add(24*time.Hour),
-		nil, nil, nil,
-	)
-	caPool := ct.NewTestCAPool(ca)
-
-	expCert, _, expKeyPEM, _ := ct.NewTestCert(
-		cert.Version2, cert.Curve_CURVE25519, ca, caKey,
-		"expired", time.Now().Add(-2*time.Hour), time.Now().Add(-1*time.Hour),
-		[]netip.Prefix{netip.MustParsePrefix("10.0.0.1/24")}, nil, nil,
-	)
-	expKey, _, _, err := cert.UnmarshalPrivateKeyFromPEM(expKeyPEM)
-	require.NoError(t, err)
-	expHsBytes, err := expCert.MarshalForHandshakes()
-	require.NoError(t, err)
-	ncs := noise.NewCipherSuite(noise.DH25519, noise.CipherChaChaPoly, noise.HashSHA256)
-
-	expiredCS := &testCertState{
-		version: cert.Version2,
-		creds: map[cert.Version]*Credential{
-			cert.Version2: NewCredential(expCert, expHsBytes, expKey, ncs),
-		},
-	}
-
-	respCS := newTestCertState(
-		t, ca, caKey, "responder",
-		[]netip.Prefix{netip.MustParsePrefix("10.0.0.2/24")},
-	)
-
-	_, respM, _, _, err := initiateHandshake(
-		t, expiredCS, testVerifier(caPool),
-		respCS, testVerifier(caPool),
-	)
-	require.ErrorContains(t, err, "verify cert")
-	assert.True(t, respM.Failed())
-}
-
-func TestMachineNoCertNetworks(t *testing.T) {
-	ca, _, caKey, _ := ct.NewTestCaCert(
-		cert.Version2, cert.Curve_CURVE25519, time.Time{}, time.Time{}, nil, nil, nil,
-	)
-	caPool := ct.NewTestCAPool(ca)
-
-	caHsBytes, err := ca.MarshalForHandshakes()
-	require.NoError(t, err)
-	ncs := noise.NewCipherSuite(noise.DH25519, noise.CipherChaChaPoly, noise.HashSHA256)
-
-	noNetCS := &testCertState{
-		version: cert.Version2,
-		creds: map[cert.Version]*Credential{
-			cert.Version2: NewCredential(ca, caHsBytes, caKey, ncs),
-		},
-	}
-
-	respCS := newTestCertState(
-		t, ca, caKey, "responder",
-		[]netip.Prefix{netip.MustParsePrefix("10.0.0.2/24")},
-	)
-
-	_, respM, _, _, err := initiateHandshake(
-		t, noNetCS, testVerifier(caPool),
-		respCS, testVerifier(caPool),
-	)
-	require.Error(t, err)
-	assert.True(t, respM.Failed())
-}
-
-func TestMachineDifferentCAs(t *testing.T) {
-	ca1, _, caKey1, _ := ct.NewTestCaCert(
-		cert.Version2, cert.Curve_CURVE25519, time.Time{}, time.Time{}, nil, nil, nil,
-	)
-	ca2, _, caKey2, _ := ct.NewTestCaCert(
-		cert.Version2, cert.Curve_CURVE25519, time.Time{}, time.Time{}, nil, nil, nil,
-	)
-
-	initCS := newTestCertState(
-		t, ca1, caKey1, "init",
-		[]netip.Prefix{netip.MustParsePrefix("10.0.0.1/24")},
-	)
-	respCS := newTestCertState(
-		t, ca2, caKey2, "resp",
-		[]netip.Prefix{netip.MustParsePrefix("10.0.0.2/24")},
-	)
-
-	_, respM, _, _, err := initiateHandshake(
-		t, initCS, testVerifier(ct.NewTestCAPool(ca1)),
-		respCS, testVerifier(ct.NewTestCAPool(ca2)),
-	)
-	require.ErrorContains(t, err, "verify cert")
-	assert.True(t, respM.Failed())
-}
-
-func TestMachineVersionNegotiation(t *testing.T) {
-	ca1, _, caKey1, _ := ct.NewTestCaCert(
-		cert.Version1, cert.Curve_CURVE25519, time.Time{}, time.Time{}, nil, nil, nil,
-	)
-	ca2, _, caKey2, _ := ct.NewTestCaCert(
-		cert.Version2, cert.Curve_CURVE25519, time.Time{}, time.Time{}, nil, nil, nil,
-	)
-	caPool := ct.NewTestCAPool(ca1, ca2)
-
-	makeMultiVersionResp := func(t *testing.T) *testCertState {
-		t.Helper()
-		respCertV1, _, respKeyPEM, _ := ct.NewTestCert(
-			cert.Version1, cert.Curve_CURVE25519, ca1, caKey1, "resp",
-			ca1.NotBefore(), ca1.NotAfter(),
-			[]netip.Prefix{netip.MustParsePrefix("10.0.0.2/24")}, nil, nil,
-		)
-		respKey, _, _, _ := cert.UnmarshalPrivateKeyFromPEM(respKeyPEM)
-		respCertV2, _ := ct.NewTestCertDifferentVersion(respCertV1, cert.Version2, ca2, caKey2)
-		respHsV1, _ := respCertV1.MarshalForHandshakes()
-		respHsV2, _ := respCertV2.MarshalForHandshakes()
-		ncs := noise.NewCipherSuite(noise.DH25519, noise.CipherChaChaPoly, noise.HashSHA256)
-		return &testCertState{
-			version: cert.Version1,
-			creds: map[cert.Version]*Credential{
-				cert.Version1: NewCredential(respCertV1, respHsV1, respKey, ncs),
-				cert.Version2: NewCredential(respCertV2, respHsV2, respKey, ncs),
-			},
-		}
-	}
-
-	t.Run("responder matches initiator version", func(t *testing.T) {
-		initCS := newTestCertState(
-			t, ca2, caKey2, "init",
-			[]netip.Prefix{netip.MustParsePrefix("10.0.0.1/24")},
-		)
-		respCS := makeMultiVersionResp(t)
-		v := testVerifier(caPool)
-
-		initM, _, respResult, resp, err := initiateHandshake(
-			t, initCS, v,
-			respCS, v,
-		)
-		require.NoError(t, err)
-		require.NotNil(t, respResult)
-
-		assert.Equal(t, cert.Version2, respResult.MyCert.Version(),
-			"responder should negotiate to initiator's version")
-
-		_, initResult, err := initM.ProcessPacket(nil, resp)
-		require.NoError(t, err)
-		require.NotNil(t, initResult)
-		assert.Equal(t, cert.Version2, initResult.RemoteCert.Certificate.Version(),
-			"initiator should see V2 cert from responder")
-	})
-
-	t.Run("responder keeps version when no match available", func(t *testing.T) {
-		initCS := newTestCertState(
-			t, ca2, caKey2, "init",
-			[]netip.Prefix{netip.MustParsePrefix("10.0.0.1/24")},
-		)
-
-		respCert, _, respKeyPEM, _ := ct.NewTestCert(
-			cert.Version1, cert.Curve_CURVE25519, ca1, caKey1, "resp",
-			ca1.NotBefore(), ca1.NotAfter(),
-			[]netip.Prefix{netip.MustParsePrefix("10.0.0.2/24")}, nil, nil,
-		)
-		respKey, _, _, _ := cert.UnmarshalPrivateKeyFromPEM(respKeyPEM)
-		respHs, _ := respCert.MarshalForHandshakes()
-		ncs := noise.NewCipherSuite(noise.DH25519, noise.CipherChaChaPoly, noise.HashSHA256)
-		respCS := &testCertState{
-			version: cert.Version1,
-			creds: map[cert.Version]*Credential{
-				cert.Version1: NewCredential(respCert, respHs, respKey, ncs),
-			},
-		}
-
-		v := testVerifier(caPool)
-		_, _, respResult, _, err := initiateHandshake(
-			t, initCS, v,
-			respCS, v,
-		)
-		require.NoError(t, err)
-		require.NotNil(t, respResult)
-
-		assert.Equal(t, cert.Version1, respResult.MyCert.Version(),
-			"responder should keep V1 when V2 not available")
-	})
-}

handshake/patterns.go

@@ -1,54 +0,0 @@
-package handshake
-
-import (
-	"fmt"
-
-	"github.com/flynn/noise"
-	"github.com/slackhq/nebula/header"
-)
-
-// msgFlags tracks what application data a handshake message carries.
-type msgFlags struct {
-	expectsPayload bool // message carries indexes and time
-	expectsCert    bool // message carries the certificate
-}
-
-// subtypeInfo bundles the noise pattern with the per-message flags for a
-// given handshake subtype.
-type subtypeInfo struct {
-	pattern noise.HandshakePattern
-	msgs    []msgFlags
-}
-
-// subtypeInfos defines the noise pattern and message content layout for each
-// handshake subtype.
-var subtypeInfos = map[header.MessageSubType]subtypeInfo{
-	// IX: 2 messages, both carry payload and cert
-	header.HandshakeIXPSK0: {
-		pattern: noise.HandshakeIX,
-		msgs: []msgFlags{
-			{expectsPayload: true, expectsCert: true},
-			{expectsPayload: true, expectsCert: true},
-		},
-	},
-
-	// XX: 3 messages
-	//   msg1 (I->R): payload only
-	//   msg2 (R->I): payload + cert
-	//   msg3 (I->R): cert only
-	//header.HandshakeXXPSK0: {
-	//	pattern: noise.HandshakeXX,
-	//	msgs: []msgFlags{
-	//		{expectsPayload: true, expectsCert: false},
-	//		{expectsPayload: true, expectsCert: true},
-	//		{expectsPayload: false, expectsCert: true},
-	//	},
-	//},
-}
-
-func subtypeInfoFor(subtype header.MessageSubType) (subtypeInfo, error) {
-	if info, ok := subtypeInfos[subtype]; ok {
-		return info, nil
-	}
-	return subtypeInfo{}, fmt.Errorf("%w: %d", ErrUnknownSubtype, subtype)
-}

handshake/patterns_test.go

@@ -1,63 +0,0 @@
-package handshake
-
-import (
-	"testing"
-
-	"github.com/flynn/noise"
-	"github.com/slackhq/nebula/header"
-	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
-)
-
-func TestSubtypeInfo(t *testing.T) {
-	t.Run("IX", func(t *testing.T) {
-		info, err := subtypeInfoFor(header.HandshakeIXPSK0)
-		require.NoError(t, err)
-		assert.Equal(t, noise.HandshakeIX.Name, info.pattern.Name)
-		require.Len(t, info.msgs, 2)
-		// msg1: payload + cert
-		assert.True(t, info.msgs[0].expectsPayload)
-		assert.True(t, info.msgs[0].expectsCert)
-		// msg2: payload + cert
-		assert.True(t, info.msgs[1].expectsPayload)
-		assert.True(t, info.msgs[1].expectsCert)
-	})
-
-	t.Run("XX", func(t *testing.T) {
-		registerTestXXInfo(t)
-		info, err := subtypeInfoFor(header.HandshakeXXPSK0)
-		require.NoError(t, err)
-		assert.Equal(t, noise.HandshakeXX.Name, info.pattern.Name)
-		require.Len(t, info.msgs, 3)
-		// msg1: payload only
-		assert.True(t, info.msgs[0].expectsPayload)
-		assert.False(t, info.msgs[0].expectsCert)
-		// msg2: payload + cert
-		assert.True(t, info.msgs[1].expectsPayload)
-		assert.True(t, info.msgs[1].expectsCert)
-		// msg3: cert only
-		assert.False(t, info.msgs[2].expectsPayload)
-		assert.True(t, info.msgs[2].expectsCert)
-	})
-
-	t.Run("unknown subtype returns error", func(t *testing.T) {
-		_, err := subtypeInfoFor(99)
-		require.ErrorIs(t, err, ErrUnknownSubtype)
-	})
-}
-
-// registerTestXXInfo temporarily registers XX subtype info for testing.
-func registerTestXXInfo(t *testing.T) {
-	t.Helper()
-	subtypeInfos[header.HandshakeXXPSK0] = subtypeInfo{
-		pattern: noise.HandshakeXX,
-		msgs: []msgFlags{
-			{expectsPayload: true, expectsCert: false},
-			{expectsPayload: true, expectsCert: true},
-			{expectsPayload: false, expectsCert: true},
-		},
-	}
-	t.Cleanup(func() {
-		delete(subtypeInfos, header.HandshakeXXPSK0)
-	})
-}

handshake/payload.go

@@ -1,173 +0,0 @@
-package handshake
-
-import (
-	"errors"
-	"math"
-
-	"google.golang.org/protobuf/encoding/protowire"
-)
-
-var (
-	errInvalidHandshakeMessage = errors.New("invalid handshake message")
-	errInvalidHandshakeDetails = errors.New("invalid handshake details")
-)
-
-// Payload represents the decoded fields of a handshake message.
-// Wire format is protobuf-compatible with NebulaHandshake{Details: NebulaHandshakeDetails{...}}.
-type Payload struct {
-	Cert           []byte
-	InitiatorIndex uint32
-	ResponderIndex uint32
-	Time           uint64
-	CertVersion    uint32
-}
-
-// Proto field numbers for NebulaHandshakeDetails
-const (
-	fieldCert           = 1 // bytes
-	fieldInitiatorIndex = 2 // uint32
-	fieldResponderIndex = 3 // uint32
-	fieldTime           = 5 // uint64
-	fieldCertVersion    = 8 // uint32
-)
-
-// MarshalPayload encodes a handshake payload in protobuf wire format compatible
-// with NebulaHandshake{Details: NebulaHandshakeDetails{...}}.
-// Returns out (which may be nil), with the marshalled Payload appended to it.
-func MarshalPayload(out []byte, p Payload) []byte {
-	var details []byte
-
-	if len(p.Cert) > 0 {
-		details = protowire.AppendTag(details, fieldCert, protowire.BytesType)
-		details = protowire.AppendBytes(details, p.Cert)
-	}
-	if p.InitiatorIndex != 0 {
-		details = protowire.AppendTag(details, fieldInitiatorIndex, protowire.VarintType)
-		details = protowire.AppendVarint(details, uint64(p.InitiatorIndex))
-	}
-	if p.ResponderIndex != 0 {
-		details = protowire.AppendTag(details, fieldResponderIndex, protowire.VarintType)
-		details = protowire.AppendVarint(details, uint64(p.ResponderIndex))
-	}
-	if p.Time != 0 {
-		details = protowire.AppendTag(details, fieldTime, protowire.VarintType)
-		details = protowire.AppendVarint(details, p.Time)
-	}
-	if p.CertVersion != 0 {
-		details = protowire.AppendTag(details, fieldCertVersion, protowire.VarintType)
-		details = protowire.AppendVarint(details, uint64(p.CertVersion))
-	}
-
-	out = protowire.AppendTag(out, 1, protowire.BytesType)
-	out = protowire.AppendBytes(out, details)
-
-	return out
-}
-
-// UnmarshalPayload decodes a protobuf-encoded NebulaHandshake message.
-func UnmarshalPayload(b []byte) (Payload, error) {
-	var p Payload
-
-	for len(b) > 0 {
-		num, typ, n := protowire.ConsumeTag(b)
-		if n < 0 {
-			return p, errInvalidHandshakeMessage
-		}
-		b = b[n:]
-
-		switch {
-		case num == 1 && typ == protowire.BytesType:
-			details, n := protowire.ConsumeBytes(b)
-			if n < 0 {
-				return p, errInvalidHandshakeMessage
-			}
-			b = b[n:]
-			if err := unmarshalPayloadDetails(&p, details); err != nil {
-				return p, err
-			}
-		default:
-			n := protowire.ConsumeFieldValue(num, typ, b)
-			if n < 0 {
-				return p, errInvalidHandshakeMessage
-			}
-			b = b[n:]
-		}
-	}
-
-	return p, nil
-}
-
-func unmarshalPayloadDetails(p *Payload, b []byte) error {
-	for len(b) > 0 {
-		num, typ, n := protowire.ConsumeTag(b)
-		if n < 0 {
-			return errInvalidHandshakeDetails
-		}
-		b = b[n:]
-
-		// For known field numbers, reject any non-matching wire type as a
-		// hard error rather than silently skipping. The caller will catch
-		// missing-field cases downstream, but a wire-type mismatch on a tag
-		// we know is a peer protocol violation worth flagging here.
-		// Repeated occurrences of a singular field follow proto3 last-wins.
-		switch num {
-		case fieldCert:
-			if typ != protowire.BytesType {
-				return errInvalidHandshakeDetails
-			}
-			v, n := protowire.ConsumeBytes(b)
-			if n < 0 {
-				return errInvalidHandshakeDetails
-			}
-			p.Cert = append([]byte(nil), v...)
-			b = b[n:]
-		case fieldInitiatorIndex:
-			if typ != protowire.VarintType {
-				return errInvalidHandshakeDetails
-			}
-			v, n := protowire.ConsumeVarint(b)
-			if n < 0 || v > math.MaxUint32 {
-				return errInvalidHandshakeDetails
-			}
-			p.InitiatorIndex = uint32(v)
-			b = b[n:]
-		case fieldResponderIndex:
-			if typ != protowire.VarintType {
-				return errInvalidHandshakeDetails
-			}
-			v, n := protowire.ConsumeVarint(b)
-			if n < 0 || v > math.MaxUint32 {
-				return errInvalidHandshakeDetails
-			}
-			p.ResponderIndex = uint32(v)
-			b = b[n:]
-		case fieldTime:
-			if typ != protowire.VarintType {
-				return errInvalidHandshakeDetails
-			}
-			v, n := protowire.ConsumeVarint(b)
-			if n < 0 {
-				return errInvalidHandshakeDetails
-			}
-			p.Time = v
-			b = b[n:]
-		case fieldCertVersion:
-			if typ != protowire.VarintType {
-				return errInvalidHandshakeDetails
-			}
-			v, n := protowire.ConsumeVarint(b)
-			if n < 0 || v > math.MaxUint32 {
-				return errInvalidHandshakeDetails
-			}
-			p.CertVersion = uint32(v)
-			b = b[n:]
-		default:
-			n := protowire.ConsumeFieldValue(num, typ, b)
-			if n < 0 {
-				return errInvalidHandshakeDetails
-			}
-			b = b[n:]
-		}
-	}
-	return nil
-}

handshake/payload_test.go

@@ -1,361 +0,0 @@
-package handshake
-
-import (
-	"bytes"
-	"math"
-	"testing"
-
-	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
-	"google.golang.org/protobuf/encoding/protowire"
-)
-
-func TestPayloadRoundTrip(t *testing.T) {
-	t.Run("all fields set", func(t *testing.T) {
-		data := MarshalPayload(nil, Payload{
-			Cert:           []byte("test-cert-bytes"),
-			CertVersion:    2,
-			InitiatorIndex: 12345,
-			ResponderIndex: 67890,
-			Time:           1234567890,
-		})
-
-		got, err := UnmarshalPayload(data)
-		require.NoError(t, err)
-
-		assert.Equal(t, []byte("test-cert-bytes"), got.Cert)
-		assert.Equal(t, uint32(12345), got.InitiatorIndex)
-		assert.Equal(t, uint32(67890), got.ResponderIndex)
-		assert.Equal(t, uint64(1234567890), got.Time)
-		assert.Equal(t, uint32(2), got.CertVersion)
-	})
-
-	t.Run("minimal fields", func(t *testing.T) {
-		data := MarshalPayload(nil, Payload{InitiatorIndex: 1})
-
-		got, err := UnmarshalPayload(data)
-		require.NoError(t, err)
-
-		assert.Equal(t, uint32(1), got.InitiatorIndex)
-		assert.Equal(t, uint32(0), got.ResponderIndex)
-		assert.Equal(t, uint64(0), got.Time)
-		assert.Nil(t, got.Cert)
-	})
-
-	t.Run("empty payload", func(t *testing.T) {
-		data := MarshalPayload(nil, Payload{})
-
-		got, err := UnmarshalPayload(data)
-		require.NoError(t, err)
-
-		assert.Equal(t, uint32(0), got.InitiatorIndex)
-	})
-
-	t.Run("large cert bytes", func(t *testing.T) {
-		bigCert := make([]byte, 4096)
-		for i := range bigCert {
-			bigCert[i] = byte(i % 256)
-		}
-
-		data := MarshalPayload(nil, Payload{
-			Cert:           bigCert,
-			CertVersion:    2,
-			InitiatorIndex: 999,
-		})
-
-		got, err := UnmarshalPayload(data)
-		require.NoError(t, err)
-
-		assert.Equal(t, bigCert, got.Cert)
-		assert.Equal(t, uint32(999), got.InitiatorIndex)
-	})
-
-	t.Run("append to existing buffer", func(t *testing.T) {
-		prefix := []byte("prefix")
-		data := MarshalPayload(prefix, Payload{InitiatorIndex: 42})
-
-		assert.Equal(t, []byte("prefix"), data[:6])
-
-		got, err := UnmarshalPayload(data[6:])
-		require.NoError(t, err)
-		assert.Equal(t, uint32(42), got.InitiatorIndex)
-	})
-}
-
-func TestPayloadUnknownFields(t *testing.T) {
-	t.Run("unknown field in outer message is skipped", func(t *testing.T) {
-		// Marshal a normal payload then append an unknown field (field 99, varint)
-		data := MarshalPayload(nil, Payload{InitiatorIndex: 42})
-		data = protowire.AppendTag(data, 99, protowire.VarintType)
-		data = protowire.AppendVarint(data, 12345)
-
-		got, err := UnmarshalPayload(data)
-		require.NoError(t, err)
-		assert.Equal(t, uint32(42), got.InitiatorIndex)
-	})
-
-	t.Run("unknown field in details is skipped", func(t *testing.T) {
-		// Build details with a known field + unknown field
-		var details []byte
-		details = protowire.AppendTag(details, fieldInitiatorIndex, protowire.VarintType)
-		details = protowire.AppendVarint(details, 77)
-		// Unknown field 50, varint
-		details = protowire.AppendTag(details, 50, protowire.VarintType)
-		details = protowire.AppendVarint(details, 9999)
-		// Another known field after the unknown one
-		details = protowire.AppendTag(details, fieldResponderIndex, protowire.VarintType)
-		details = protowire.AppendVarint(details, 88)
-
-		// Wrap in outer message
-		var data []byte
-		data = protowire.AppendTag(data, 1, protowire.BytesType)
-		data = protowire.AppendBytes(data, details)
-
-		got, err := UnmarshalPayload(data)
-		require.NoError(t, err)
-		assert.Equal(t, uint32(77), got.InitiatorIndex)
-		assert.Equal(t, uint32(88), got.ResponderIndex)
-	})
-
-	t.Run("reserved fields 6 and 7 are skipped", func(t *testing.T) {
-		// Fields 6 and 7 are reserved in the proto definition
-		var details []byte
-		details = protowire.AppendTag(details, fieldInitiatorIndex, protowire.VarintType)
-		details = protowire.AppendVarint(details, 100)
-		details = protowire.AppendTag(details, 6, protowire.VarintType)
-		details = protowire.AppendVarint(details, 1)
-		details = protowire.AppendTag(details, 7, protowire.VarintType)
-		details = protowire.AppendVarint(details, 2)
-
-		var data []byte
-		data = protowire.AppendTag(data, 1, protowire.BytesType)
-		data = protowire.AppendBytes(data, details)
-
-		got, err := UnmarshalPayload(data)
-		require.NoError(t, err)
-		assert.Equal(t, uint32(100), got.InitiatorIndex)
-	})
-}
-
-func TestPayloadBytesConsumed(t *testing.T) {
-	t.Run("all bytes consumed on valid input", func(t *testing.T) {
-		original := Payload{
-			Cert:           []byte("cert"),
-			CertVersion:    2,
-			InitiatorIndex: 100,
-			ResponderIndex: 200,
-			Time:           999,
-		}
-		data := MarshalPayload(nil, original)
-
-		got, err := UnmarshalPayload(data)
-		require.NoError(t, err)
-
-		// Re-marshal and compare — proves we consumed and reproduced all fields
-		remarshaled := MarshalPayload(nil, got)
-		assert.Equal(t, data, remarshaled)
-	})
-}
-
-// wrapDetails wraps raw detail bytes in the outer NebulaHandshake envelope
-// so UnmarshalPayload can reach unmarshalPayloadDetails.
-func wrapDetails(details []byte) []byte {
-	var out []byte
-	out = protowire.AppendTag(out, 1, protowire.BytesType)
-	out = protowire.AppendBytes(out, details)
-	return out
-}
-
-func TestPayloadUnmarshalErrors(t *testing.T) {
-	t.Run("nil input", func(t *testing.T) {
-		got, err := UnmarshalPayload(nil)
-		require.NoError(t, err)
-		assert.Equal(t, uint32(0), got.InitiatorIndex)
-	})
-
-	t.Run("truncated outer tag", func(t *testing.T) {
-		_, err := UnmarshalPayload([]byte{0x80})
-		assert.Error(t, err)
-	})
-
-	t.Run("truncated outer details field", func(t *testing.T) {
-		_, err := UnmarshalPayload([]byte{0x0a, 0x64, 0x01, 0x02, 0x03, 0x04, 0x05})
-		assert.Error(t, err)
-	})
-
-	t.Run("truncated outer unknown field", func(t *testing.T) {
-		// Valid tag for unknown field 99 varint, but no value follows
-		var data []byte
-		data = protowire.AppendTag(data, 99, protowire.VarintType)
-		_, err := UnmarshalPayload(data)
-		assert.Error(t, err)
-	})
-
-	t.Run("truncated details tag", func(t *testing.T) {
-		_, err := UnmarshalPayload(wrapDetails([]byte{0x80}))
-		assert.Error(t, err)
-	})
-
-	t.Run("truncated cert bytes", func(t *testing.T) {
-		// Field 1 (cert), bytes type, length 10 but only 2 bytes
-		var details []byte
-		details = protowire.AppendTag(details, fieldCert, protowire.BytesType)
-		details = append(details, 0x0a, 0x01, 0x02) // length 10, only 2 bytes
-		_, err := UnmarshalPayload(wrapDetails(details))
-		assert.Error(t, err)
-	})
-
-	t.Run("truncated initiator index varint", func(t *testing.T) {
-		var details []byte
-		details = protowire.AppendTag(details, fieldInitiatorIndex, protowire.VarintType)
-		details = append(details, 0x80) // incomplete varint
-		_, err := UnmarshalPayload(wrapDetails(details))
-		assert.Error(t, err)
-	})
-
-	t.Run("truncated responder index varint", func(t *testing.T) {
-		var details []byte
-		details = protowire.AppendTag(details, fieldResponderIndex, protowire.VarintType)
-		details = append(details, 0x80)
-		_, err := UnmarshalPayload(wrapDetails(details))
-		assert.Error(t, err)
-	})
-
-	t.Run("truncated time varint", func(t *testing.T) {
-		var details []byte
-		details = protowire.AppendTag(details, fieldTime, protowire.VarintType)
-		details = append(details, 0x80)
-		_, err := UnmarshalPayload(wrapDetails(details))
-		assert.Error(t, err)
-	})
-
-	t.Run("truncated cert version varint", func(t *testing.T) {
-		var details []byte
-		details = protowire.AppendTag(details, fieldCertVersion, protowire.VarintType)
-		details = append(details, 0x80)
-		_, err := UnmarshalPayload(wrapDetails(details))
-		assert.Error(t, err)
-	})
-
-	t.Run("truncated unknown field in details", func(t *testing.T) {
-		var details []byte
-		details = protowire.AppendTag(details, 50, protowire.VarintType)
-		details = append(details, 0x80) // incomplete varint
-		_, err := UnmarshalPayload(wrapDetails(details))
-		assert.Error(t, err)
-	})
-
-	t.Run("cert with wrong wire type rejected", func(t *testing.T) {
-		// fieldCert as Varint instead of Bytes.
-		var details []byte
-		details = protowire.AppendTag(details, fieldCert, protowire.VarintType)
-		details = protowire.AppendVarint(details, 42)
-		_, err := UnmarshalPayload(wrapDetails(details))
-		assert.Error(t, err)
-	})
-
-	t.Run("initiator index with wrong wire type rejected", func(t *testing.T) {
-		// fieldInitiatorIndex as Bytes instead of Varint.
-		var details []byte
-		details = protowire.AppendTag(details, fieldInitiatorIndex, protowire.BytesType)
-		details = protowire.AppendBytes(details, []byte{1, 2, 3})
-		_, err := UnmarshalPayload(wrapDetails(details))
-		assert.Error(t, err)
-	})
-
-	t.Run("time with wrong wire type rejected", func(t *testing.T) {
-		var details []byte
-		details = protowire.AppendTag(details, fieldTime, protowire.BytesType)
-		details = protowire.AppendBytes(details, []byte{1, 2, 3})
-		_, err := UnmarshalPayload(wrapDetails(details))
-		assert.Error(t, err)
-	})
-
-	t.Run("cert version with wrong wire type rejected", func(t *testing.T) {
-		var details []byte
-		details = protowire.AppendTag(details, fieldCertVersion, protowire.BytesType)
-		details = protowire.AppendBytes(details, []byte{1, 2, 3})
-		_, err := UnmarshalPayload(wrapDetails(details))
-		assert.Error(t, err)
-	})
-
-	t.Run("repeated singular field follows proto3 last-wins", func(t *testing.T) {
-		// Per proto3, multiple instances of a singular field are accepted and
-		// the last value wins. We keep this behavior so that peers using
-		// alternative encoders aren't rejected.
-		var details []byte
-		details = protowire.AppendTag(details, fieldInitiatorIndex, protowire.VarintType)
-		details = protowire.AppendVarint(details, 1)
-		details = protowire.AppendTag(details, fieldInitiatorIndex, protowire.VarintType)
-		details = protowire.AppendVarint(details, 42)
-		got, err := UnmarshalPayload(wrapDetails(details))
-		require.NoError(t, err)
-		assert.Equal(t, uint32(42), got.InitiatorIndex)
-	})
-
-	t.Run("initiator index varint overflow rejected", func(t *testing.T) {
-		var details []byte
-		details = protowire.AppendTag(details, fieldInitiatorIndex, protowire.VarintType)
-		details = protowire.AppendVarint(details, math.MaxUint32+1)
-		_, err := UnmarshalPayload(wrapDetails(details))
-		assert.Error(t, err)
-	})
-
-	t.Run("cert version varint overflow rejected", func(t *testing.T) {
-		var details []byte
-		details = protowire.AppendTag(details, fieldCertVersion, protowire.VarintType)
-		details = protowire.AppendVarint(details, math.MaxUint32+1)
-		_, err := UnmarshalPayload(wrapDetails(details))
-		assert.Error(t, err)
-	})
-
-}
-
-// FuzzPayload feeds arbitrary bytes through UnmarshalPayload to confirm it
-// never panics, and for any input that parses cleanly, that re-marshal +
-// re-parse is a fix-point. Inputs come from an authenticated peer (post-
-// noise-decrypt), so the threat model is "valid peer behaving arbitrarily,"
-// not "unauthenticated injection."
-func FuzzPayload(f *testing.F) {
-	// Seed corpus with a handful of known-good shapes.
-	f.Add(MarshalPayload(nil, Payload{}))
-	f.Add(MarshalPayload(nil, Payload{Cert: []byte{1, 2, 3}, CertVersion: 2}))
-	f.Add(MarshalPayload(nil, Payload{InitiatorIndex: 42, Time: 1}))
-	f.Add(MarshalPayload(nil, Payload{
-		Cert:           []byte("seed-cert"),
-		InitiatorIndex: 1,
-		ResponderIndex: 2,
-		Time:           3,
-		CertVersion:    2,
-	}))
-	f.Add([]byte{})
-	f.Add([]byte{0xff})
-
-	f.Fuzz(func(t *testing.T, data []byte) {
-		p1, err := UnmarshalPayload(data)
-		if err != nil {
-			return
-		}
-
-		// For any input that parses, re-marshaling and re-parsing must
-		// yield an equivalent Payload. This catches dispatch bugs (e.g.
-		// emitting a field on marshal that we don't accept on parse) and
-		// any non-idempotent parsing behavior.
-		b2 := MarshalPayload(nil, p1)
-		p2, err := UnmarshalPayload(b2)
-		if err != nil {
-			t.Fatalf("re-parse of self-marshaled payload failed: %v\nintermediate: %x\n", err, b2)
-		}
-		if !payloadsEqual(p1, p2) {
-			t.Fatalf("re-marshal not idempotent\nfirst:  %+v\nsecond: %+v", p1, p2)
-		}
-	})
-}
-
-func payloadsEqual(a, b Payload) bool {
-	return bytes.Equal(a.Cert, b.Cert) &&
-		a.InitiatorIndex == b.InitiatorIndex &&
-		a.ResponderIndex == b.ResponderIndex &&
-		a.Time == b.Time &&
-		a.CertVersion == b.CertVersion
-}

handshake_ix.go

@@ -0,0 +1,813 @@
+package nebula
+
+import (
+	"bytes"
+	"context"
+	"log/slog"
+	"net/netip"
+	"time"
+
+	"github.com/flynn/noise"
+	"github.com/slackhq/nebula/cert"
+	"github.com/slackhq/nebula/header"
+)
+
+// NOISE IX Handshakes
+
+// This function constructs a handshake packet, but does not actually send it
+// Sending is done by the handshake manager
+func ixHandshakeStage0(f *Interface, hh *HandshakeHostInfo) bool {
+	err := f.handshakeManager.allocateIndex(hh)
+	if err != nil {
+		f.l.Error("Failed to generate index",
+			"error", err,
+			"vpnAddrs", hh.hostinfo.vpnAddrs,
+			"handshake", m{"stage": 0, "style": "ix_psk0"},
+		)
+		return false
+	}
+
+	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 {
+		f.l.Error("Unable to handshake with host because no certificate is available",
+			"vpnAddrs", hh.hostinfo.vpnAddrs,
+			"handshake", m{"stage": 0, "style": "ix_psk0"},
+			"certVersion", v,
+		)
+		return false
+	}
+
+	crtHs := cs.getHandshakeBytes(v)
+	if crtHs == nil {
+		f.l.Error("Unable to handshake with host because no certificate handshake bytes is available",
+			"vpnAddrs", hh.hostinfo.vpnAddrs,
+			"handshake", m{"stage": 0, "style": "ix_psk0"},
+			"certVersion", v,
+		)
+		return false
+	}
+
+	ci, err := NewConnectionState(cs, crt, true, noise.HandshakeIX)
+	if err != nil {
+		f.l.Error("Failed to create connection state",
+			"error", err,
+			"vpnAddrs", hh.hostinfo.vpnAddrs,
+			"handshake", m{"stage": 0, "style": "ix_psk0"},
+			"certVersion", v,
+		)
+		return false
+	}
+	hh.hostinfo.ConnectionState = ci
+
+	hs := &NebulaHandshake{
+		Details: &NebulaHandshakeDetails{
+			InitiatorIndex: hh.hostinfo.localIndexId,
+			Time:           uint64(time.Now().UnixNano()),
+			Cert:           crtHs,
+			CertVersion:    uint32(v),
+		},
+	}
+
+	hsBytes, err := hs.Marshal()
+	if err != nil {
+		f.l.Error("Failed to marshal handshake message",
+			"error", err,
+			"vpnAddrs", hh.hostinfo.vpnAddrs,
+			"certVersion", v,
+			"handshake", m{"stage": 0, "style": "ix_psk0"},
+		)
+		return false
+	}
+
+	h := header.Encode(make([]byte, header.Len), header.Version, header.Handshake, header.HandshakeIXPSK0, 0, 1)
+
+	msg, _, _, err := ci.H.WriteMessage(h, hsBytes)
+	if err != nil {
+		f.l.Error("Failed to call noise.WriteMessage",
+			"error", err,
+			"vpnAddrs", hh.hostinfo.vpnAddrs,
+			"handshake", m{"stage": 0, "style": "ix_psk0"},
+		)
+		return false
+	}
+
+	// We are sending handshake packet 1, so we don't expect to receive
+	// handshake packet 1 from the responder
+	ci.window.Update(f.l, 1)
+
+	hh.hostinfo.HandshakePacket[0] = msg
+	hh.ready = true
+	return true
+}
+
+func ixHandshakeStage1(f *Interface, via ViaSender, packet []byte, h *header.H) {
+	cs := f.pki.getCertState()
+	crt := cs.GetDefaultCertificate()
+	if crt == nil {
+		f.l.Error("Unable to handshake with host because no certificate is available",
+			"from", via,
+			"handshake", m{"stage": 0, "style": "ix_psk0"},
+			"certVersion", cs.initiatingVersion,
+		)
+		return
+	}
+
+	ci, err := NewConnectionState(cs, crt, false, noise.HandshakeIX)
+	if err != nil {
+		f.l.Error("Failed to create connection state",
+			"error", err,
+			"from", via,
+			"handshake", m{"stage": 1, "style": "ix_psk0"},
+		)
+		return
+	}
+
+	// Mark packet 1 as seen so it doesn't show up as missed
+	ci.window.Update(f.l, 1)
+
+	msg, _, _, err := ci.H.ReadMessage(nil, packet[header.Len:])
+	if err != nil {
+		f.l.Error("Failed to call noise.ReadMessage",
+			"error", err,
+			"from", via,
+			"handshake", m{"stage": 1, "style": "ix_psk0"},
+		)
+		return
+	}
+
+	hs := &NebulaHandshake{}
+	err = hs.Unmarshal(msg)
+	if err != nil || hs.Details == nil {
+		f.l.Error("Failed unmarshal handshake message",
+			"error", err,
+			"from", via,
+			"handshake", m{"stage": 1, "style": "ix_psk0"},
+		)
+		return
+	}
+
+	rc, err := cert.Recombine(cert.Version(hs.Details.CertVersion), hs.Details.Cert, ci.H.PeerStatic(), ci.Curve())
+	if err != nil {
+		f.l.Info("Handshake did not contain a certificate",
+			"error", err,
+			"from", via,
+			"handshake", m{"stage": 1, "style": "ix_psk0"},
+		)
+		return
+	}
+
+	remoteCert, err := f.pki.GetCAPool().VerifyCertificate(time.Now(), rc)
+	if err != nil {
+		fp, fperr := rc.Fingerprint()
+		if fperr != nil {
+			fp = "<error generating certificate fingerprint>"
+		}
+
+		attrs := []slog.Attr{
+			slog.Any("error", err),
+			slog.Any("from", via),
+			slog.Any("handshake", m{"stage": 1, "style": "ix_psk0"}),
+			slog.Any("certVpnNetworks", rc.Networks()),
+			slog.String("certFingerprint", fp),
+		}
+		if f.l.Enabled(context.Background(), slog.LevelDebug) {
+			attrs = append(attrs, slog.Any("cert", rc))
+		}
+
+		// LogAttrs is intentional: attrs is a pre-built []slog.Attr slice that
+		// callers grow conditionally, which has no pair-form equivalent.
+		//nolint:sloglint
+		f.l.LogAttrs(context.Background(), slog.LevelInfo, "Invalid certificate from host", attrs...)
+		return
+	}
+
+	if !bytes.Equal(remoteCert.Certificate.PublicKey(), ci.H.PeerStatic()) {
+		f.l.Info("public key mismatch between certificate and handshake",
+			"from", via,
+			"handshake", m{"stage": 1, "style": "ix_psk0"},
+			"cert", remoteCert,
+		)
+		return
+	}
+
+	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.Enabled(context.Background(), slog.LevelDebug) {
+				f.l.Debug("Might be unable to handshake with host due to missing certificate version",
+					"error", err,
+					"from", via,
+					"handshake", m{"stage": 1, "style": "ix_psk0"},
+					"cert", remoteCert,
+				)
+			}
+		} else {
+			// Record the certificate we are actually using
+			ci.myCert = myCertOtherVersion
+		}
+	}
+
+	if len(remoteCert.Certificate.Networks()) == 0 {
+		f.l.Info("No networks in certificate",
+			"error", err,
+			"from", via,
+			"cert", remoteCert,
+			"handshake", m{"stage": 1, "style": "ix_psk0"},
+		)
+		return
+	}
+
+	certName := remoteCert.Certificate.Name()
+	certVersion := remoteCert.Certificate.Version()
+	fingerprint := remoteCert.Fingerprint
+	issuer := remoteCert.Certificate.Issuer()
+	vpnNetworks := remoteCert.Certificate.Networks()
+
+	anyVpnAddrsInCommon := false
+	vpnAddrs := make([]netip.Addr, len(vpnNetworks))
+	for i, network := range vpnNetworks {
+		if f.myVpnAddrsTable.Contains(network.Addr()) {
+			f.l.Error("Refusing to handshake with myself",
+				"vpnNetworks", vpnNetworks,
+				"from", via,
+				"certName", certName,
+				"certVersion", certVersion,
+				"fingerprint", fingerprint,
+				"issuer", issuer,
+				"handshake", m{"stage": 1, "style": "ix_psk0"},
+			)
+			return
+		}
+		vpnAddrs[i] = network.Addr()
+		if f.myVpnNetworksTable.Contains(network.Addr()) {
+			anyVpnAddrsInCommon = true
+		}
+	}
+
+	if !via.IsRelayed {
+		// We only want to apply the remote allow list for direct tunnels here
+		if !f.lightHouse.GetRemoteAllowList().AllowAll(vpnAddrs, via.UdpAddr.Addr()) {
+			if f.l.Enabled(context.Background(), slog.LevelDebug) {
+				f.l.Debug("lighthouse.remote_allow_list denied incoming handshake",
+					"vpnAddrs", vpnAddrs,
+					"from", via,
+				)
+			}
+			return
+		}
+	}
+
+	myIndex, err := generateIndex(f.l)
+	if err != nil {
+		f.l.Error("Failed to generate index",
+			"error", err,
+			"vpnAddrs", vpnAddrs,
+			"from", via,
+			"certName", certName,
+			"certVersion", certVersion,
+			"fingerprint", fingerprint,
+			"issuer", issuer,
+			"handshake", m{"stage": 1, "style": "ix_psk0"},
+		)
+		return
+	}
+
+	hostinfo := &HostInfo{
+		ConnectionState:   ci,
+		localIndexId:      myIndex,
+		remoteIndexId:     hs.Details.InitiatorIndex,
+		vpnAddrs:          vpnAddrs,
+		HandshakePacket:   make(map[uint8][]byte, 0),
+		lastHandshakeTime: hs.Details.Time,
+		relayState: RelayState{
+			relays:         nil,
+			relayForByAddr: map[netip.Addr]*Relay{},
+			relayForByIdx:  map[uint32]*Relay{},
+		},
+	}
+
+	msgRxL := f.l.With(
+		"vpnAddrs", vpnAddrs,
+		"from", via,
+		"certName", certName,
+		"certVersion", certVersion,
+		"fingerprint", fingerprint,
+		"issuer", issuer,
+		"initiatorIndex", hs.Details.InitiatorIndex,
+		"responderIndex", hs.Details.ResponderIndex,
+		"remoteIndex", h.RemoteIndex,
+		"handshake", m{"stage": 1, "style": "ix_psk0"},
+	)
+
+	if anyVpnAddrsInCommon {
+		msgRxL.Info("Handshake message received")
+	} else {
+		//todo warn if not lighthouse or relay?
+		msgRxL.Info("Handshake message received, but no vpnNetworks in common.")
+	}
+
+	hs.Details.ResponderIndex = myIndex
+	hs.Details.Cert = cs.getHandshakeBytes(ci.myCert.Version())
+	if hs.Details.Cert == nil {
+		msgRxL.Error("Unable to handshake with host because no certificate handshake bytes is available",
+			"myCertVersion", ci.myCert.Version(),
+		)
+		return
+	}
+
+	hs.Details.CertVersion = uint32(ci.myCert.Version())
+	// Update the time in case their clock is way off from ours
+	hs.Details.Time = uint64(time.Now().UnixNano())
+
+	hsBytes, err := hs.Marshal()
+	if err != nil {
+		f.l.Error("Failed to marshal handshake message",
+			"error", err,
+			"vpnAddrs", hostinfo.vpnAddrs,
+			"from", via,
+			"certName", certName,
+			"certVersion", certVersion,
+			"fingerprint", fingerprint,
+			"issuer", issuer,
+			"handshake", m{"stage": 1, "style": "ix_psk0"},
+		)
+		return
+	}
+
+	nh := header.Encode(make([]byte, header.Len), header.Version, header.Handshake, header.HandshakeIXPSK0, hs.Details.InitiatorIndex, 2)
+	msg, dKey, eKey, err := ci.H.WriteMessage(nh, hsBytes)
+	if err != nil {
+		f.l.Error("Failed to call noise.WriteMessage",
+			"error", err,
+			"vpnAddrs", hostinfo.vpnAddrs,
+			"from", via,
+			"certName", certName,
+			"certVersion", certVersion,
+			"fingerprint", fingerprint,
+			"issuer", issuer,
+			"handshake", m{"stage": 1, "style": "ix_psk0"},
+		)
+		return
+	} else if dKey == nil || eKey == nil {
+		f.l.Error("Noise did not arrive at a key",
+			"vpnAddrs", hostinfo.vpnAddrs,
+			"from", via,
+			"certName", certName,
+			"certVersion", certVersion,
+			"fingerprint", fingerprint,
+			"issuer", issuer,
+			"handshake", m{"stage": 1, "style": "ix_psk0"},
+		)
+		return
+	}
+
+	hostinfo.HandshakePacket[0] = make([]byte, len(packet[header.Len:]))
+	copy(hostinfo.HandshakePacket[0], packet[header.Len:])
+
+	// Regardless of whether you are the sender or receiver, you should arrive here
+	// and complete standing up the connection.
+	hostinfo.HandshakePacket[2] = make([]byte, len(msg))
+	copy(hostinfo.HandshakePacket[2], msg)
+
+	// We are sending handshake packet 2, so we don't expect to receive
+	// handshake packet 2 from the initiator.
+	ci.window.Update(f.l, 2)
+
+	ci.peerCert = remoteCert
+	ci.dKey = NewNebulaCipherState(dKey)
+	ci.eKey = NewNebulaCipherState(eKey)
+
+	hostinfo.remotes = f.lightHouse.QueryCache(vpnAddrs)
+	if !via.IsRelayed {
+		hostinfo.SetRemote(via.UdpAddr)
+	}
+	hostinfo.buildNetworks(f.myVpnNetworksTable, remoteCert.Certificate)
+
+	existing, err := f.handshakeManager.CheckAndComplete(hostinfo, 0, f)
+	if err != nil {
+		switch err {
+		case ErrAlreadySeen:
+			// Update remote if preferred
+			if existing.SetRemoteIfPreferred(f.hostMap, via) {
+				// Send a test packet to ensure the other side has also switched to
+				// the preferred remote
+				f.SendMessageToVpnAddr(header.Test, header.TestRequest, vpnAddrs[0], []byte(""), make([]byte, 12, 12), make([]byte, mtu))
+			}
+
+			msg = existing.HandshakePacket[2]
+			f.messageMetrics.Tx(header.Handshake, header.MessageSubType(msg[1]), 1)
+			if !via.IsRelayed {
+				err := f.outside.WriteTo(msg, via.UdpAddr)
+				if err != nil {
+					f.l.Error("Failed to send handshake message",
+						"vpnAddrs", existing.vpnAddrs,
+						"from", via,
+						"handshake", m{"stage": 2, "style": "ix_psk0"},
+						"cached", true,
+						"error", err,
+					)
+				} else {
+					f.l.Info("Handshake message sent",
+						"vpnAddrs", existing.vpnAddrs,
+						"from", via,
+						"handshake", m{"stage": 2, "style": "ix_psk0"},
+						"cached", true,
+					)
+				}
+				return
+			} else {
+				if via.relay == nil {
+					f.l.Error("Handshake send failed: both addr and via.relay are nil.")
+					return
+				}
+				hostinfo.relayState.InsertRelayTo(via.relayHI.vpnAddrs[0])
+				f.SendVia(via.relayHI, via.relay, msg, make([]byte, 12), make([]byte, mtu), false)
+				f.l.Info("Handshake message sent",
+					"vpnAddrs", existing.vpnAddrs,
+					"relay", via.relayHI.vpnAddrs[0],
+					"handshake", m{"stage": 2, "style": "ix_psk0"},
+					"cached", true,
+				)
+				return
+			}
+		case ErrExistingHostInfo:
+			// This means there was an existing tunnel and this handshake was older than the one we are currently based on
+			f.l.Info("Handshake too old",
+				"vpnAddrs", vpnAddrs,
+				"from", via,
+				"certName", certName,
+				"certVersion", certVersion,
+				"oldHandshakeTime", existing.lastHandshakeTime,
+				"newHandshakeTime", hostinfo.lastHandshakeTime,
+				"fingerprint", fingerprint,
+				"issuer", issuer,
+				"initiatorIndex", hs.Details.InitiatorIndex,
+				"responderIndex", hs.Details.ResponderIndex,
+				"remoteIndex", h.RemoteIndex,
+				"handshake", m{"stage": 1, "style": "ix_psk0"},
+			)
+
+			// Send a test packet to trigger an authenticated tunnel test, this should suss out any lingering tunnel issues
+			f.SendMessageToVpnAddr(header.Test, header.TestRequest, vpnAddrs[0], []byte(""), make([]byte, 12, 12), make([]byte, mtu))
+			return
+		case ErrLocalIndexCollision:
+			// This means we failed to insert because of collision on localIndexId. Just let the next handshake packet retry
+			f.l.Error("Failed to add HostInfo due to localIndex collision",
+				"vpnAddrs", vpnAddrs,
+				"from", via,
+				"certName", certName,
+				"certVersion", certVersion,
+				"fingerprint", fingerprint,
+				"issuer", issuer,
+				"initiatorIndex", hs.Details.InitiatorIndex,
+				"responderIndex", hs.Details.ResponderIndex,
+				"remoteIndex", h.RemoteIndex,
+				"handshake", m{"stage": 1, "style": "ix_psk0"},
+				"localIndex", hostinfo.localIndexId,
+				"collision", existing.vpnAddrs,
+			)
+			return
+		default:
+			// Shouldn't happen, but just in case someone adds a new error type to CheckAndComplete
+			// And we forget to update it here
+			f.l.Error("Failed to add HostInfo to HostMap",
+				"error", err,
+				"vpnAddrs", vpnAddrs,
+				"from", via,
+				"certName", certName,
+				"certVersion", certVersion,
+				"fingerprint", fingerprint,
+				"issuer", issuer,
+				"initiatorIndex", hs.Details.InitiatorIndex,
+				"responderIndex", hs.Details.ResponderIndex,
+				"remoteIndex", h.RemoteIndex,
+				"handshake", m{"stage": 1, "style": "ix_psk0"},
+			)
+			return
+		}
+	}
+
+	// Do the send
+	f.messageMetrics.Tx(header.Handshake, header.MessageSubType(msg[1]), 1)
+	if !via.IsRelayed {
+		err = f.outside.WriteTo(msg, via.UdpAddr)
+		log := f.l.With(
+			"vpnAddrs", vpnAddrs,
+			"from", via,
+			"certName", certName,
+			"certVersion", certVersion,
+			"fingerprint", fingerprint,
+			"issuer", issuer,
+			"initiatorIndex", hs.Details.InitiatorIndex,
+			"responderIndex", hs.Details.ResponderIndex,
+			"remoteIndex", h.RemoteIndex,
+			"handshake", m{"stage": 2, "style": "ix_psk0"},
+		)
+		if err != nil {
+			log.Error("Failed to send handshake", "error", err)
+		} else {
+			log.Info("Handshake message sent")
+		}
+	} else {
+		if via.relay == nil {
+			f.l.Error("Handshake send failed: both addr and via.relay are nil.")
+			return
+		}
+		hostinfo.relayState.InsertRelayTo(via.relayHI.vpnAddrs[0])
+		// I successfully received a handshake. Just in case I marked this tunnel as 'Disestablished', ensure
+		// it's correctly marked as working.
+		via.relayHI.relayState.UpdateRelayForByIdxState(via.remoteIdx, Established)
+		f.SendVia(via.relayHI, via.relay, msg, make([]byte, 12), make([]byte, mtu), false)
+		f.l.Info("Handshake message sent",
+			"vpnAddrs", vpnAddrs,
+			"relay", via.relayHI.vpnAddrs[0],
+			"certName", certName,
+			"certVersion", certVersion,
+			"fingerprint", fingerprint,
+			"issuer", issuer,
+			"initiatorIndex", hs.Details.InitiatorIndex,
+			"responderIndex", hs.Details.ResponderIndex,
+			"remoteIndex", h.RemoteIndex,
+			"handshake", m{"stage": 2, "style": "ix_psk0"},
+		)
+	}
+
+	f.connectionManager.AddTrafficWatch(hostinfo)
+
+	hostinfo.remotes.RefreshFromHandshake(vpnAddrs)
+
+	// Don't wait for UpdateWorker
+	if f.lightHouse.IsAnyLighthouseAddr(vpnAddrs) {
+		f.lightHouse.TriggerUpdate()
+	}
+
+	return
+}
+
+func ixHandshakeStage2(f *Interface, via ViaSender, hh *HandshakeHostInfo, packet []byte, h *header.H) bool {
+	if hh == nil {
+		// Nothing here to tear down, got a bogus stage 2 packet
+		return true
+	}
+
+	hh.Lock()
+	defer hh.Unlock()
+
+	hostinfo := hh.hostinfo
+	if !via.IsRelayed {
+		// The vpnAddr we know about is the one we tried to handshake with, use it to apply the remote allow list.
+		if !f.lightHouse.GetRemoteAllowList().AllowAll(hostinfo.vpnAddrs, via.UdpAddr.Addr()) {
+			if f.l.Enabled(context.Background(), slog.LevelDebug) {
+				f.l.Debug("lighthouse.remote_allow_list denied incoming handshake",
+					"vpnAddrs", hostinfo.vpnAddrs,
+					"from", via,
+				)
+			}
+			return false
+		}
+	}
+
+	ci := hostinfo.ConnectionState
+	msg, eKey, dKey, err := ci.H.ReadMessage(nil, packet[header.Len:])
+	if err != nil {
+		f.l.Error("Failed to call noise.ReadMessage",
+			"error", err,
+			"vpnAddrs", hostinfo.vpnAddrs,
+			"from", via,
+			"handshake", m{"stage": 2, "style": "ix_psk0"},
+			"header", h,
+		)
+
+		// We don't want to tear down the connection on a bad ReadMessage because it could be an attacker trying
+		// to DOS us. Every other error condition after should to allow a possible good handshake to complete in the
+		// near future
+		return false
+	} else if dKey == nil || eKey == nil {
+		f.l.Error("Noise did not arrive at a key",
+			"vpnAddrs", hostinfo.vpnAddrs,
+			"from", via,
+			"handshake", m{"stage": 2, "style": "ix_psk0"},
+		)
+
+		// This should be impossible in IX but just in case, if we get here then there is no chance to recover
+		// the handshake state machine. Tear it down
+		return true
+	}
+
+	hs := &NebulaHandshake{}
+	err = hs.Unmarshal(msg)
+	if err != nil || hs.Details == nil {
+		f.l.Error("Failed unmarshal handshake message",
+			"error", err,
+			"vpnAddrs", hostinfo.vpnAddrs,
+			"from", via,
+			"handshake", m{"stage": 2, "style": "ix_psk0"},
+		)
+
+		// The handshake state machine is complete, if things break now there is no chance to recover. Tear down and start again
+		return true
+	}
+
+	rc, err := cert.Recombine(cert.Version(hs.Details.CertVersion), hs.Details.Cert, ci.H.PeerStatic(), ci.Curve())
+	if err != nil {
+		f.l.Info("Handshake did not contain a certificate",
+			"error", err,
+			"from", via,
+			"vpnAddrs", hostinfo.vpnAddrs,
+			"handshake", m{"stage": 2, "style": "ix_psk0"},
+		)
+		return true
+	}
+
+	remoteCert, err := f.pki.GetCAPool().VerifyCertificate(time.Now(), rc)
+	if err != nil {
+		fp, err := rc.Fingerprint()
+		if err != nil {
+			fp = "<error generating certificate fingerprint>"
+		}
+
+		attrs := []slog.Attr{
+			slog.Any("error", err),
+			slog.Any("from", via),
+			slog.Any("vpnAddrs", hostinfo.vpnAddrs),
+			slog.Any("handshake", m{"stage": 2, "style": "ix_psk0"}),
+			slog.String("certFingerprint", fp),
+			slog.Any("certVpnNetworks", rc.Networks()),
+		}
+		if f.l.Enabled(context.Background(), slog.LevelDebug) {
+			attrs = append(attrs, slog.Any("cert", rc))
+		}
+
+		// LogAttrs is intentional: attrs is a pre-built []slog.Attr slice that
+		// callers grow conditionally, which has no pair-form equivalent.
+		//nolint:sloglint
+		f.l.LogAttrs(context.Background(), slog.LevelInfo, "Invalid certificate from host", attrs...)
+		return true
+	}
+	if !bytes.Equal(remoteCert.Certificate.PublicKey(), ci.H.PeerStatic()) {
+		f.l.Info("public key mismatch between certificate and handshake",
+			"from", via,
+			"handshake", m{"stage": 2, "style": "ix_psk0"},
+			"cert", remoteCert,
+		)
+		return true
+	}
+
+	if len(remoteCert.Certificate.Networks()) == 0 {
+		f.l.Info("No networks in certificate",
+			"error", err,
+			"from", via,
+			"vpnAddrs", hostinfo.vpnAddrs,
+			"cert", remoteCert,
+			"handshake", m{"stage": 2, "style": "ix_psk0"},
+		)
+		return true
+	}
+
+	vpnNetworks := remoteCert.Certificate.Networks()
+	certName := remoteCert.Certificate.Name()
+	certVersion := remoteCert.Certificate.Version()
+	fingerprint := remoteCert.Fingerprint
+	issuer := remoteCert.Certificate.Issuer()
+
+	hostinfo.remoteIndexId = hs.Details.ResponderIndex
+	hostinfo.lastHandshakeTime = hs.Details.Time
+
+	// Store their cert and our symmetric keys
+	ci.peerCert = remoteCert
+	ci.dKey = NewNebulaCipherState(dKey)
+	ci.eKey = NewNebulaCipherState(eKey)
+
+	// Make sure the current udpAddr being used is set for responding
+	if !via.IsRelayed {
+		hostinfo.SetRemote(via.UdpAddr)
+	} else {
+		hostinfo.relayState.InsertRelayTo(via.relayHI.vpnAddrs[0])
+	}
+
+	correctHostResponded := false
+	anyVpnAddrsInCommon := false
+	vpnAddrs := make([]netip.Addr, len(vpnNetworks))
+	for i, network := range vpnNetworks {
+		vpnAddrs[i] = network.Addr()
+		if f.myVpnNetworksTable.Contains(network.Addr()) {
+			anyVpnAddrsInCommon = true
+		}
+		if hostinfo.vpnAddrs[0] == network.Addr() {
+			// todo is it more correct to see if any of hostinfo.vpnAddrs are in the cert? it should have len==1, but one day it might not?
+			correctHostResponded = true
+		}
+	}
+
+	// Ensure the right host responded
+	if !correctHostResponded {
+		f.l.Info("Incorrect host responded to handshake",
+			"intendedVpnAddrs", hostinfo.vpnAddrs,
+			"haveVpnNetworks", vpnNetworks,
+			"from", via,
+			"certName", certName,
+			"certVersion", certVersion,
+			"handshake", m{"stage": 2, "style": "ix_psk0"},
+		)
+
+		// Release our old handshake from pending, it should not continue
+		f.handshakeManager.DeleteHostInfo(hostinfo)
+
+		// Create a new hostinfo/handshake for the intended vpn ip
+		//TODO is hostinfo.vpnAddrs[0] always the address to use?
+		f.handshakeManager.StartHandshake(hostinfo.vpnAddrs[0], func(newHH *HandshakeHostInfo) {
+			// Block the current used address
+			newHH.hostinfo.remotes = hostinfo.remotes
+			newHH.hostinfo.remotes.BlockRemote(via)
+
+			f.l.Info("Blocked addresses for handshakes",
+				"blockedUdpAddrs", newHH.hostinfo.remotes.CopyBlockedRemotes(),
+				"vpnNetworks", vpnNetworks,
+				"remotes", newHH.hostinfo.remotes.CopyAddrs(f.hostMap.GetPreferredRanges()),
+			)
+
+			// Swap the packet store to benefit the original intended recipient
+			newHH.packetStore = hh.packetStore
+			hh.packetStore = []*cachedPacket{}
+
+			// Finally, put the correct vpn addrs in the host info, tell them to close the tunnel, and return true to tear down
+			hostinfo.vpnAddrs = vpnAddrs
+			f.sendCloseTunnel(hostinfo)
+		})
+
+		return true
+	}
+
+	// Mark packet 2 as seen so it doesn't show up as missed
+	ci.window.Update(f.l, 2)
+
+	duration := time.Since(hh.startTime).Nanoseconds()
+	msgRxL := f.l.With(
+		"vpnAddrs", vpnAddrs,
+		"from", via,
+		"certName", certName,
+		"certVersion", certVersion,
+		"fingerprint", fingerprint,
+		"issuer", issuer,
+		"initiatorIndex", hs.Details.InitiatorIndex,
+		"responderIndex", hs.Details.ResponderIndex,
+		"remoteIndex", h.RemoteIndex,
+		"handshake", m{"stage": 2, "style": "ix_psk0"},
+		"durationNs", duration,
+		"sentCachedPackets", len(hh.packetStore),
+	)
+	if anyVpnAddrsInCommon {
+		msgRxL.Info("Handshake message received")
+	} else {
+		//todo warn if not lighthouse or relay?
+		msgRxL.Info("Handshake message received, but no vpnNetworks in common.")
+	}
+
+	// Build up the radix for the firewall if we have subnets in the cert
+	hostinfo.vpnAddrs = vpnAddrs
+	hostinfo.buildNetworks(f.myVpnNetworksTable, remoteCert.Certificate)
+
+	// Complete our handshake and update metrics, this will replace any existing tunnels for the vpnAddrs here
+	f.handshakeManager.Complete(hostinfo, f)
+	f.connectionManager.AddTrafficWatch(hostinfo)
+
+	if f.l.Enabled(context.Background(), slog.LevelDebug) {
+		hostinfo.logger(f.l).Debug("Sending stored packets",
+			"count", len(hh.packetStore),
+		)
+	}
+
+	if len(hh.packetStore) > 0 {
+		nb := make([]byte, 12, 12)
+		out := make([]byte, mtu)
+		for _, cp := range hh.packetStore {
+			cp.callback(cp.messageType, cp.messageSubType, hostinfo, cp.packet, nb, out)
+		}
+		f.cachedPacketMetrics.sent.Inc(int64(len(hh.packetStore)))
+	}
+
+	hostinfo.remotes.RefreshFromHandshake(vpnAddrs)
+	f.metricHandshakes.Update(duration)
+
+	// Don't wait for UpdateWorker
+	if f.lightHouse.IsAnyLighthouseAddr(vpnAddrs) {
+		f.lightHouse.TriggerUpdate()
+	}
+
+	return false
+}

handshake_manager.go

@@ -14,7 +14,6 @@ import (
 
 	"github.com/rcrowley/go-metrics"
 	"github.com/slackhq/nebula/cert"
-	"github.com/slackhq/nebula/handshake"
 	"github.com/slackhq/nebula/header"
 	"github.com/slackhq/nebula/udp"
 )
@@ -23,18 +22,7 @@ const (
 	DefaultHandshakeTryInterval   = time.Millisecond * 100
 	DefaultHandshakeRetries       = 10
 	DefaultHandshakeTriggerBuffer = 64
-
+	DefaultUseRelays              = true
-	// maxCachedPackets is how many unsent packets we'll buffer per pending
-	// handshake before dropping further ones.
-	maxCachedPackets = 100
-
-	// HandshakePacket map keys mirror the IX protocol stage convention:
-	//   stage 0 = the initiator's first message (and what the responder
-	//             receives, stripped of header)
-	//   stage 2 = the responder's reply
-	// Other handshake patterns will need new keys when added.
-	handshakePacketStage0 uint8 = 0
-	handshakePacketStage2 uint8 = 2
 )
 
 var (
@@ -42,6 +30,7 @@ var (
 		tryInterval:   DefaultHandshakeTryInterval,
 		retries:       DefaultHandshakeRetries,
 		triggerBuffer: DefaultHandshakeTriggerBuffer,
+		useRelays:     DefaultUseRelays,
 	}
 )
 
@@ -49,6 +38,7 @@ type HandshakeConfig struct {
 	tryInterval   time.Duration
 	retries       int64
 	triggerBuffer int
+	useRelays     bool
 
 	messageMetrics *MessageMetrics
 }
@@ -86,11 +76,10 @@ type HandshakeHostInfo struct {
 	packetStore               []*cachedPacket  // A set of packets to be transmitted once the handshake completes
 
 	hostinfo *HostInfo
-	machine  *handshake.Machine // The handshake state machine, set during stage 0 (initiator) or beginHandshake (responder multi-message)
 }
 
 func (hh *HandshakeHostInfo) cachePacket(l *slog.Logger, t header.MessageType, st header.MessageSubType, packet []byte, f packetCallback, m *cachedPacketMetrics) {
-	if len(hh.packetStore) < maxCachedPackets {
+	if len(hh.packetStore) < 100 {
 		tempPacket := make([]byte, len(packet))
 		copy(tempPacket, packet)
 
@@ -148,18 +137,6 @@ func (hm *HandshakeManager) Run(ctx context.Context) {
 }
 
 func (hm *HandshakeManager) HandleIncoming(via ViaSender, packet []byte, h *header.H) {
-	// Gate on known handshake subtypes. Unknown subtypes (or future ones we
-	// don't yet support) are dropped here rather than silently routed through
-	// the IX path. Add a case when introducing a new pattern.
-	switch h.Subtype {
-	case header.HandshakeIXPSK0:
-		// supported
-	default:
-		hm.l.Debug("dropping handshake with unsupported subtype",
-			"from", via, "subtype", h.Subtype)
-		return
-	}
-
 	// First remote allow list check before we know the vpnIp
 	if !via.IsRelayed {
 		if !hm.lightHouse.GetRemoteAllowList().AllowUnknownVpnAddr(via.UdpAddr.Addr()) {
@@ -168,27 +145,19 @@ func (hm *HandshakeManager) HandleIncoming(via ViaSender, packet []byte, h *head
 		}
 	}
 
-	// First message of a new handshake. The wire format requires RemoteIndex
+	switch h.Subtype {
-	// to be zero here (the initiator has no responder index to fill in yet),
+	case header.HandshakeIXPSK0:
-	// and generateIndex never allocates 0, so any non-zero RemoteIndex on a
+		switch h.MessageCounter {
-	// stage-1 packet is malformed or someone probing for an index collision.
+		case 1:
-	// Drop without paying the cost of running noise on a pending Machine.
+			ixHandshakeStage1(hm.f, via, packet, h)
-	if h.MessageCounter == 1 {
-		if h.RemoteIndex != 0 {
-			hm.l.Debug("dropping stage-1 handshake with non-zero RemoteIndex",
-				"from", via, "remoteIndex", h.RemoteIndex)
-			return
-		}
-		hm.beginHandshake(via, packet, h)
-		return
-	}
 
-	// Continuation message must match a pending handshake by index.
+		case 2:
-	// Anything else is an orphaned packet (e.g., late retransmit after
+			newHostinfo := hm.queryIndex(h.RemoteIndex)
-	// timeout) and is dropped.
+			tearDown := ixHandshakeStage2(hm.f, via, newHostinfo, packet, h)
-	if hh := hm.queryIndex(h.RemoteIndex); hh != nil {
+			if tearDown && newHostinfo != nil {
-		hm.continueHandshake(via, hh, packet)
+				hm.DeleteHostInfo(newHostinfo.hostinfo)
-		return
+			}
+		}
 	}
 }
 
@@ -214,22 +183,13 @@ func (hm *HandshakeManager) handleOutbound(vpnIp netip.Addr, lighthouseTriggered
 	hostinfo := hh.hostinfo
 	// If we are out of time, clean up
 	if hh.counter >= hm.config.retries {
-		fields := []any{
+		hh.hostinfo.logger(hm.l).Info("Handshake timed out",
 			"udpAddrs", hh.hostinfo.remotes.CopyAddrs(hm.mainHostMap.GetPreferredRanges()),
 			"initiatorIndex", hh.hostinfo.localIndexId,
 			"remoteIndex", hh.hostinfo.remoteIndexId,
+			"handshake", m{"stage": 1, "style": "ix_psk0"},
 			"durationNs", time.Since(hh.startTime).Nanoseconds(),
-		}
+		)
-		// hh.machine can be nil here if buildStage0Packet never succeeded
-		// (e.g., no certificate available). In that case there's no useful
-		// handshake metadata to log.
-		if hh.machine != nil {
-			fields = append(fields, "handshake", m{
-				"stage": uint64(hh.machine.MessageIndex()),
-				"style": header.SubTypeName(header.Handshake, hh.machine.Subtype()),
-			})
-		}
-		hh.hostinfo.logger(hm.l).Info("Handshake timed out", fields...)
 		hm.metricTimedOut.Inc(1)
 		hm.DeleteHostInfo(hostinfo)
 		return
@@ -240,25 +200,12 @@ func (hm *HandshakeManager) handleOutbound(vpnIp netip.Addr, lighthouseTriggered
 
 	// Check if we have a handshake packet to transmit yet
 	if !hh.ready {
-		if !hm.buildStage0Packet(hh) {
+		if !ixHandshakeStage0(hm.f, hh) {
 			hm.OutboundHandshakeTimer.Add(vpnIp, hm.config.tryInterval*time.Duration(hh.counter))
 			return
 		}
 	}
 
-	// TODO: this hardcodes "always retransmit stage 0", which is correct for
-	// IX (the initiator only ever sends one packet, msg1) but wrong the
-	// moment a 3+ message pattern lands. The retry loop should resend the
-	// most recent outgoing message, not always stage 0. That implies
-	// HandshakeHostInfo tracking a single "currentOutbound" packet (bytes +
-	// header metadata) that gets replaced as the handshake progresses,
-	// instead of indexing into HandshakePacket.
-	stage0 := hostinfo.HandshakePacket[handshakePacketStage0]
-	hsFields := m{
-		"stage": uint64(hh.machine.MessageIndex()),
-		"style": header.SubTypeName(header.Handshake, hh.machine.Subtype()),
-	}
-
 	// Get a remotes object if we don't already have one.
 	// This is mainly to protect us as this should never be the case
 	// NB ^ This comment doesn't jive. It's how the thing gets initialized.
@@ -292,13 +239,13 @@ func (hm *HandshakeManager) handleOutbound(vpnIp netip.Addr, lighthouseTriggered
 	// Send the handshake to all known ips, stage 2 takes care of assigning the hostinfo.remote based on the first to reply
 	var sentTo []netip.AddrPort
 	hostinfo.remotes.ForEach(hm.mainHostMap.GetPreferredRanges(), func(addr netip.AddrPort, _ bool) {
-		hm.messageMetrics.Tx(header.Handshake, hh.machine.Subtype(), 1)
+		hm.messageMetrics.Tx(header.Handshake, header.MessageSubType(hostinfo.HandshakePacket[0][1]), 1)
-		err := hm.outside.WriteTo(stage0, addr)
+		err := hm.outside.WriteTo(hostinfo.HandshakePacket[0], addr)
 		if err != nil {
 			hostinfo.logger(hm.l).Error("Failed to send handshake message",
 				"udpAddr", addr,
 				"initiatorIndex", hostinfo.localIndexId,
-				"handshake", hsFields,
+				"handshake", m{"stage": 1, "style": "ix_psk0"},
 				"error", err,
 			)
 
@@ -313,17 +260,156 @@ func (hm *HandshakeManager) handleOutbound(vpnIp netip.Addr, lighthouseTriggered
 		hostinfo.logger(hm.l).Info("Handshake message sent",
 			"udpAddrs", sentTo,
 			"initiatorIndex", hostinfo.localIndexId,
-			"handshake", hsFields,
+			"handshake", m{"stage": 1, "style": "ix_psk0"},
 		)
 	} else if hm.l.Enabled(context.Background(), slog.LevelDebug) {
 		hostinfo.logger(hm.l).Debug("Handshake message sent",
 			"udpAddrs", sentTo,
 			"initiatorIndex", hostinfo.localIndexId,
-			"handshake", hsFields,
+			"handshake", m{"stage": 1, "style": "ix_psk0"},
 		)
 	}
 
-	hm.f.relayManager.StartRelays(hm.f, vpnIp, hostinfo, stage0)
+	if hm.config.useRelays && len(hostinfo.remotes.relays) > 0 {
+		hostinfo.logger(hm.l).Info("Attempt to relay through hosts", "relays", hostinfo.remotes.relays)
+		// Send a RelayRequest to all known Relay IP's
+		for _, relay := range hostinfo.remotes.relays {
+			// Don't relay through the host I'm trying to connect to
+			if relay == vpnIp {
+				continue
+			}
+
+			// Don't relay to myself
+			if hm.f.myVpnAddrsTable.Contains(relay) {
+				continue
+			}
+
+			relayHostInfo := hm.mainHostMap.QueryVpnAddr(relay)
+			if relayHostInfo == nil || !relayHostInfo.remote.IsValid() {
+				hostinfo.logger(hm.l).Info("Establish tunnel to relay target", "relay", relay.String())
+				hm.f.Handshake(relay)
+				continue
+			}
+			// Check the relay HostInfo to see if we already established a relay through
+			existingRelay, ok := relayHostInfo.relayState.QueryRelayForByIp(vpnIp)
+			if !ok {
+				// No relays exist or requested yet.
+				if relayHostInfo.remote.IsValid() {
+					idx, err := AddRelay(hm.l, relayHostInfo, hm.mainHostMap, vpnIp, nil, TerminalType, Requested)
+					if err != nil {
+						hostinfo.logger(hm.l).Info("Failed to add relay to hostmap", "relay", relay.String(), "error", err)
+					}
+
+					m := NebulaControl{
+						Type:                NebulaControl_CreateRelayRequest,
+						InitiatorRelayIndex: idx,
+					}
+
+					switch relayHostInfo.GetCert().Certificate.Version() {
+					case cert.Version1:
+						if !hm.f.myVpnAddrs[0].Is4() {
+							hostinfo.logger(hm.l).Error("can not establish v1 relay with a v6 network because the relay is not running a current nebula version")
+							continue
+						}
+
+						if !vpnIp.Is4() {
+							hostinfo.logger(hm.l).Error("can not establish v1 relay with a v6 remote network because the relay is not running a current nebula version")
+							continue
+						}
+
+						b := hm.f.myVpnAddrs[0].As4()
+						m.OldRelayFromAddr = binary.BigEndian.Uint32(b[:])
+						b = vpnIp.As4()
+						m.OldRelayToAddr = binary.BigEndian.Uint32(b[:])
+					case cert.Version2:
+						m.RelayFromAddr = netAddrToProtoAddr(hm.f.myVpnAddrs[0])
+						m.RelayToAddr = netAddrToProtoAddr(vpnIp)
+					default:
+						hostinfo.logger(hm.l).Error("Unknown certificate version found while creating relay")
+						continue
+					}
+
+					msg, err := m.Marshal()
+					if err != nil {
+						hostinfo.logger(hm.l).Error("Failed to marshal Control message to create relay", "error", err)
+					} else {
+						hm.f.SendMessageToHostInfo(header.Control, 0, relayHostInfo, msg, make([]byte, 12), make([]byte, mtu))
+						hm.l.Info("send CreateRelayRequest",
+							"relayFrom", hm.f.myVpnAddrs[0],
+							"relayTo", vpnIp,
+							"initiatorRelayIndex", idx,
+							"relay", relay,
+						)
+					}
+				}
+				continue
+			}
+
+			switch existingRelay.State {
+			case Established:
+				hostinfo.logger(hm.l).Info("Send handshake via relay", "relay", relay.String())
+				hm.f.SendVia(relayHostInfo, existingRelay, hostinfo.HandshakePacket[0], make([]byte, 12), make([]byte, mtu), false)
+			case Disestablished:
+				// Mark this relay as 'requested'
+				relayHostInfo.relayState.UpdateRelayForByIpState(vpnIp, Requested)
+				fallthrough
+			case Requested:
+				hostinfo.logger(hm.l).Info("Re-send CreateRelay request", "relay", relay.String())
+				// Re-send the CreateRelay request, in case the previous one was lost.
+				m := NebulaControl{
+					Type:                NebulaControl_CreateRelayRequest,
+					InitiatorRelayIndex: existingRelay.LocalIndex,
+				}
+
+				switch relayHostInfo.GetCert().Certificate.Version() {
+				case cert.Version1:
+					if !hm.f.myVpnAddrs[0].Is4() {
+						hostinfo.logger(hm.l).Error("can not establish v1 relay with a v6 network because the relay is not running a current nebula version")
+						continue
+					}
+
+					if !vpnIp.Is4() {
+						hostinfo.logger(hm.l).Error("can not establish v1 relay with a v6 remote network because the relay is not running a current nebula version")
+						continue
+					}
+
+					b := hm.f.myVpnAddrs[0].As4()
+					m.OldRelayFromAddr = binary.BigEndian.Uint32(b[:])
+					b = vpnIp.As4()
+					m.OldRelayToAddr = binary.BigEndian.Uint32(b[:])
+				case cert.Version2:
+					m.RelayFromAddr = netAddrToProtoAddr(hm.f.myVpnAddrs[0])
+					m.RelayToAddr = netAddrToProtoAddr(vpnIp)
+				default:
+					hostinfo.logger(hm.l).Error("Unknown certificate version found while creating relay")
+					continue
+				}
+				msg, err := m.Marshal()
+				if err != nil {
+					hostinfo.logger(hm.l).Error("Failed to marshal Control message to create relay", "error", err)
+				} else {
+					// This must send over the hostinfo, not over hm.Hosts[ip]
+					hm.f.SendMessageToHostInfo(header.Control, 0, relayHostInfo, msg, make([]byte, 12), make([]byte, mtu))
+					hm.l.Info("send CreateRelayRequest",
+						"relayFrom", hm.f.myVpnAddrs[0],
+						"relayTo", vpnIp,
+						"initiatorRelayIndex", existingRelay.LocalIndex,
+						"relay", relay,
+					)
+				}
+			case PeerRequested:
+				// PeerRequested only occurs in Forwarding relays, not Terminal relays, and this is a Terminal relay case.
+				fallthrough
+			default:
+				hostinfo.logger(hm.l).Error("Relay unexpected state",
+					"vpnIp", vpnIp,
+					"state", existingRelay.State,
+					"relay", relay,
+				)
+
+			}
+		}
+	}
 
 	// If a lighthouse triggered this attempt then we are still in the timer wheel and do not need to re-add
 	if !lighthouseTriggered {
@@ -501,7 +587,7 @@ func (hm *HandshakeManager) Complete(hostinfo *HostInfo, f *Interface) {
 // allocateIndex generates a unique localIndexId for this HostInfo
 // and adds it to the pendingHostMap. Will error if we are unable to generate
 // a unique localIndexId
-func (hm *HandshakeManager) allocateIndex(hh *HandshakeHostInfo) (uint32, error) {
+func (hm *HandshakeManager) allocateIndex(hh *HandshakeHostInfo) error {
 	hm.mainHostMap.RLock()
 	defer hm.mainHostMap.RUnlock()
 	hm.Lock()
@@ -510,7 +596,7 @@ func (hm *HandshakeManager) allocateIndex(hh *HandshakeHostInfo) (uint32, error)
 	for range 32 {
 		index, err := generateIndex(hm.l)
 		if err != nil {
-			return 0, err
+			return err
 		}
 
 		_, inPending := hm.indexes[index]
@@ -519,11 +605,11 @@ func (hm *HandshakeManager) allocateIndex(hh *HandshakeHostInfo) (uint32, error)
 		if !inMain && !inPending {
 			hh.hostinfo.localIndexId = index
 			hm.indexes[index] = hh
-			return index, nil
+			return nil
 		}
 	}
 
-	return 0, errors.New("failed to generate unique localIndexId")
+	return errors.New("failed to generate unique localIndexId")
 }
 
 func (hm *HandshakeManager) DeleteHostInfo(hostinfo *HostInfo) {
@@ -642,524 +728,3 @@ func generateIndex(l *slog.Logger) (uint32, error) {
 func hsTimeout(tries int64, interval time.Duration) time.Duration {
 	return time.Duration(tries / 2 * ((2 * int64(interval)) + (tries-1)*int64(interval)))
 }
-
-// buildStage0Packet creates the initial handshake packet for the initiator.
-func (hm *HandshakeManager) buildStage0Packet(hh *HandshakeHostInfo) bool {
-	cs := hm.f.pki.getCertState()
-	v := cs.DefaultVersion()
-	if hh.initiatingVersionOverride != cert.VersionPre1 {
-		v = hh.initiatingVersionOverride
-	} else if v < cert.Version2 {
-		for _, a := range hh.hostinfo.vpnAddrs {
-			if a.Is6() {
-				v = cert.Version2
-				break
-			}
-		}
-	}
-
-	cred := cs.GetCredential(v)
-	if cred == nil {
-		hm.f.l.Error("Unable to handshake with host because no certificate is available",
-			"vpnAddrs", hh.hostinfo.vpnAddrs, "certVersion", v)
-		return false
-	}
-
-	machine, err := handshake.NewMachine(
-		v, cs.GetCredential,
-		hm.certVerifier(), func() (uint32, error) { return hm.allocateIndex(hh) },
-		true, header.HandshakeIXPSK0,
-	)
-	if err != nil {
-		hm.f.l.Error("Failed to create handshake machine",
-			"vpnAddrs", hh.hostinfo.vpnAddrs, "error", err)
-		return false
-	}
-
-	msg, err := machine.Initiate(nil)
-	if err != nil {
-		hm.f.l.Error("Failed to initiate handshake",
-			"vpnAddrs", hh.hostinfo.vpnAddrs, "error", err)
-		return false
-	}
-
-	// hostinfo.ConnectionState stays nil until the handshake completes in
-	// continueHandshake. Pre-completion control surfaces guard with nil
-	// checks; the data plane never observes a pending hostinfo.
-	hh.hostinfo.HandshakePacket[handshakePacketStage0] = msg
-	hh.machine = machine
-	hh.ready = true
-	return true
-}
-
-// beginHandshake handles an incoming handshake packet that doesn't match any
-// existing pending handshake. It creates a new responder Machine and processes
-// the first message.
-func (hm *HandshakeManager) beginHandshake(via ViaSender, packet []byte, h *header.H) {
-	f := hm.f
-	cs := f.pki.getCertState()
-
-	v := cs.DefaultVersion()
-	if cs.GetCredential(v) == nil {
-		f.l.Error("Unable to handshake with host because no certificate is available",
-			"from", via, "certVersion", v)
-		return
-	}
-
-	machine, err := handshake.NewMachine(
-		v, cs.GetCredential,
-		hm.certVerifier(), func() (uint32, error) { return generateIndex(f.l) },
-		false, header.HandshakeIXPSK0,
-	)
-	if err != nil {
-		f.l.Error("Failed to create handshake machine", "from", via, "error", err)
-		return
-	}
-
-	response, result, err := machine.ProcessPacket(nil, packet)
-	if err != nil {
-		f.l.Error("Failed to process handshake packet", "from", via, "error", err)
-		return
-	}
-
-	if result == nil {
-		// Multi-message pattern: the responder Machine would need to be
-		// registered in hm.indexes so a future inbound packet finds it via
-		// continueHandshake. The current manager doesn't do that yet, so
-		// fail loudly rather than silently dropping the in-flight handshake.
-		// TODO: support multi-message responder flows (XX, pqIX, etc.).
-		// See also the IX-shaped cipher key assignment in handshake.Machine.
-		f.l.Error("multi-message handshake responder is not supported",
-			"from", via, "error", handshake.ErrMultiMessageUnsupported)
-		return
-	}
-
-	remoteCert := result.RemoteCert
-	if remoteCert == nil {
-		f.l.Error("Handshake did not produce a peer certificate", "from", via)
-		return
-	}
-
-	// Validate peer identity
-	vpnAddrs, anyVpnAddrsInCommon, ok := hm.validatePeerCert(via, remoteCert)
-	if !ok {
-		return
-	}
-
-	hostinfo := &HostInfo{
-		ConnectionState:   newConnectionStateFromResult(result),
-		localIndexId:      result.LocalIndex,
-		remoteIndexId:     result.RemoteIndex,
-		vpnAddrs:          vpnAddrs,
-		HandshakePacket:   make(map[uint8][]byte, 0),
-		lastHandshakeTime: result.HandshakeTime,
-		relayState: RelayState{
-			relays:         nil,
-			relayForByAddr: map[netip.Addr]*Relay{},
-			relayForByIdx:  map[uint32]*Relay{},
-		},
-	}
-
-	msg := "Handshake message received"
-	if !anyVpnAddrsInCommon {
-		msg = "Handshake message received, but no vpnNetworks in common."
-	}
-	f.l.Info(msg,
-		"vpnAddrs", vpnAddrs,
-		"from", via,
-		"certName", remoteCert.Certificate.Name(),
-		"certVersion", remoteCert.Certificate.Version(),
-		"fingerprint", remoteCert.Fingerprint,
-		"issuer", remoteCert.Certificate.Issuer(),
-		"initiatorIndex", result.RemoteIndex,
-		"responderIndex", result.LocalIndex,
-		"handshake", m{"stage": uint64(machine.MessageIndex()), "style": header.SubTypeName(header.Handshake, machine.Subtype())},
-	)
-
-	// packet aliases the listener's incoming buffer, so this copy must stay.
-	hostinfo.HandshakePacket[handshakePacketStage0] = make([]byte, len(packet[header.Len:]))
-	copy(hostinfo.HandshakePacket[handshakePacketStage0], packet[header.Len:])
-
-	// response was freshly allocated by ProcessPacket; safe to retain directly.
-	if response != nil {
-		hostinfo.HandshakePacket[handshakePacketStage2] = response
-	}
-
-	hostinfo.remotes = f.lightHouse.QueryCache(vpnAddrs)
-	if !via.IsRelayed {
-		hostinfo.SetRemote(via.UdpAddr)
-	}
-	hostinfo.buildNetworks(f.myVpnNetworksTable, remoteCert.Certificate)
-
-	existing, err := hm.CheckAndComplete(hostinfo, handshakePacketStage0, f)
-	if err != nil {
-		hm.handleCheckAndCompleteError(err, existing, hostinfo, via)
-		return
-	}
-
-	hm.sendHandshakeResponse(via, response, hostinfo, false)
-	f.connectionManager.AddTrafficWatch(hostinfo)
-	hostinfo.remotes.RefreshFromHandshake(vpnAddrs)
-
-	// Don't wait for UpdateWorker
-	if f.lightHouse.IsAnyLighthouseAddr(vpnAddrs) {
-		f.lightHouse.TriggerUpdate()
-	}
-}
-
-// continueHandshake feeds an incoming packet to an existing pending handshake Machine.
-func (hm *HandshakeManager) continueHandshake(via ViaSender, hh *HandshakeHostInfo, packet []byte) {
-	f := hm.f
-
-	hh.Lock()
-	defer hh.Unlock()
-
-	// Re-verify hh is still tracked. Between queryIndex returning and us taking
-	// hh.Lock, handleOutbound may have timed out and deleted it. Once we hold
-	// hh.Lock no other deleter can race our index: handleOutbound also takes
-	// hh.Lock first, and handleRecvError targets a main-hostmap entry with a
-	// different localIndexId.
-	hm.RLock()
-	cur, ok := hm.indexes[hh.hostinfo.localIndexId]
-	hm.RUnlock()
-	if !ok || cur != hh {
-		return
-	}
-
-	hostinfo := hh.hostinfo
-	if !via.IsRelayed {
-		if !f.lightHouse.GetRemoteAllowList().AllowAll(hostinfo.vpnAddrs, via.UdpAddr.Addr()) {
-			f.l.Debug("lighthouse.remote_allow_list denied incoming handshake",
-				"vpnAddrs", hostinfo.vpnAddrs, "from", via)
-			return
-		}
-	}
-
-	machine := hh.machine
-	if machine == nil {
-		f.l.Error("No handshake machine available for continuation",
-			"vpnAddrs", hostinfo.vpnAddrs, "from", via)
-		hm.DeleteHostInfo(hostinfo)
-		return
-	}
-
-	response, result, err := machine.ProcessPacket(nil, packet)
-	if err != nil {
-		// Recoverable errors are routine noise, log at Debug. Fatal errors get a Warn.
-		if machine.Failed() {
-			f.l.Warn("Failed to process handshake packet, abandoning",
-				"vpnAddrs", hostinfo.vpnAddrs, "from", via, "error", err)
-			hm.DeleteHostInfo(hostinfo)
-		} else {
-			f.l.Debug("Failed to process handshake packet",
-				"vpnAddrs", hostinfo.vpnAddrs, "from", via, "error", err)
-		}
-		return
-	}
-
-	if response != nil {
-		hm.sendHandshakeResponse(via, response, hostinfo, false)
-	}
-
-	if result == nil {
-		return
-	}
-
-	// Handshake complete; build the ConnectionState now that we have keys and a verified peer cert.
-	hostinfo.ConnectionState = newConnectionStateFromResult(result)
-
-	remoteCert := result.RemoteCert
-	if remoteCert == nil {
-		f.l.Error("Handshake completed without peer certificate",
-			"vpnAddrs", hostinfo.vpnAddrs, "from", via)
-		hm.DeleteHostInfo(hostinfo)
-		return
-	}
-
-	vpnNetworks := remoteCert.Certificate.Networks()
-	hostinfo.remoteIndexId = result.RemoteIndex
-	hostinfo.lastHandshakeTime = result.HandshakeTime
-
-	if !via.IsRelayed {
-		hostinfo.SetRemote(via.UdpAddr)
-	} else {
-		hostinfo.relayState.InsertRelayTo(via.relayHI.vpnAddrs[0])
-	}
-
-	// Verify correct host responded (initiator check)
-	vpnAddrs := make([]netip.Addr, len(vpnNetworks))
-	correctHostResponded := false
-	anyVpnAddrsInCommon := false
-	for i, network := range vpnNetworks {
-		// inside.go drops self-routed packets at the firewall stage, but we'd
-		// rather not let a self-handshake complete in the first place: it
-		// wastes a hostmap slot, suppresses no log, and obscures routing
-		// misconfig. Explicit refusal here mirrors the responder-side check
-		// in validatePeerCert.
-		if f.myVpnAddrsTable.Contains(network.Addr()) {
-			f.l.Error("Refusing to handshake with myself",
-				"vpnNetworks", vpnNetworks,
-				"from", via,
-				"certName", remoteCert.Certificate.Name(),
-				"certVersion", remoteCert.Certificate.Version(),
-				"fingerprint", remoteCert.Fingerprint,
-				"issuer", remoteCert.Certificate.Issuer(),
-				"handshake", m{"stage": uint64(machine.MessageIndex()), "style": header.SubTypeName(header.Handshake, machine.Subtype())},
-			)
-			hm.DeleteHostInfo(hostinfo)
-			return
-		}
-		vpnAddrs[i] = network.Addr()
-		if hostinfo.vpnAddrs[0] == network.Addr() {
-			correctHostResponded = true
-		}
-		if f.myVpnNetworksTable.Contains(network.Addr()) {
-			anyVpnAddrsInCommon = true
-		}
-	}
-
-	if !correctHostResponded {
-		f.l.Info("Incorrect host responded to handshake",
-			"intendedVpnAddrs", hostinfo.vpnAddrs,
-			"haveVpnNetworks", vpnNetworks,
-			"from", via,
-			"certName", remoteCert.Certificate.Name(),
-			"certVersion", remoteCert.Certificate.Version(),
-			"fingerprint", remoteCert.Fingerprint,
-			"issuer", remoteCert.Certificate.Issuer(),
-			"handshake", m{"stage": uint64(machine.MessageIndex()), "style": header.SubTypeName(header.Handshake, machine.Subtype())},
-		)
-
-		hm.DeleteHostInfo(hostinfo)
-		hm.StartHandshake(hostinfo.vpnAddrs[0], func(newHH *HandshakeHostInfo) {
-			newHH.hostinfo.remotes = hostinfo.remotes
-			newHH.hostinfo.remotes.BlockRemote(via)
-			newHH.packetStore = hh.packetStore
-			hh.packetStore = []*cachedPacket{}
-			hostinfo.vpnAddrs = vpnAddrs
-			f.sendCloseTunnel(hostinfo)
-		})
-		return
-	}
-
-	duration := time.Since(hh.startTime).Nanoseconds()
-	msg := "Handshake message received"
-	if !anyVpnAddrsInCommon {
-		msg = "Handshake message received, but no vpnNetworks in common."
-	}
-	f.l.Info(msg,
-		"vpnAddrs", vpnAddrs,
-		"from", via,
-		"certName", remoteCert.Certificate.Name(),
-		"certVersion", remoteCert.Certificate.Version(),
-		"fingerprint", remoteCert.Fingerprint,
-		"issuer", remoteCert.Certificate.Issuer(),
-		"initiatorIndex", result.LocalIndex,
-		"responderIndex", result.RemoteIndex,
-		"handshake", m{"stage": uint64(machine.MessageIndex()), "style": header.SubTypeName(header.Handshake, machine.Subtype())},
-		"durationNs", duration,
-		"sentCachedPackets", len(hh.packetStore),
-	)
-
-	hostinfo.vpnAddrs = vpnAddrs
-	hostinfo.buildNetworks(f.myVpnNetworksTable, remoteCert.Certificate)
-
-	hm.Complete(hostinfo, f)
-	f.connectionManager.AddTrafficWatch(hostinfo)
-
-	if len(hh.packetStore) > 0 {
-		if f.l.Enabled(context.Background(), slog.LevelDebug) {
-			hostinfo.logger(f.l).Debug("Sending stored packets", "count", len(hh.packetStore))
-		}
-		nb := make([]byte, 12, 12)
-		out := make([]byte, mtu)
-		for _, cp := range hh.packetStore {
-			cp.callback(cp.messageType, cp.messageSubType, hostinfo, cp.packet, nb, out)
-		}
-		f.cachedPacketMetrics.sent.Inc(int64(len(hh.packetStore)))
-	}
-
-	hostinfo.remotes.RefreshFromHandshake(vpnAddrs)
-	f.metricHandshakes.Update(duration)
-
-	// Don't wait for UpdateWorker
-	if f.lightHouse.IsAnyLighthouseAddr(vpnAddrs) {
-		f.lightHouse.TriggerUpdate()
-	}
-}
-
-// validatePeerCert checks the peer certificate for self-connection and remote allow list.
-// Returns the VPN addrs, whether any of them fall within one of our own VPN
-// networks, and true if valid; false if rejected.
-func (hm *HandshakeManager) validatePeerCert(via ViaSender, remoteCert *cert.CachedCertificate) ([]netip.Addr, bool, bool) {
-	f := hm.f
-	vpnNetworks := remoteCert.Certificate.Networks()
-
-	// The cert package rejects host certs with no networks at parse time, so
-	// reaching this state would mean an invariant was bypassed elsewhere.
-	// Refuse explicitly so downstream code (which indexes vpnAddrs[0]) can't
-	// panic if that invariant ever changes.
-	if len(vpnNetworks) == 0 {
-		f.l.Info("No networks in certificate",
-			"from", via, "cert", remoteCert)
-		return nil, false, false
-	}
-
-	vpnAddrs := make([]netip.Addr, len(vpnNetworks))
-	anyVpnAddrsInCommon := false
-
-	for i, network := range vpnNetworks {
-		if f.myVpnAddrsTable.Contains(network.Addr()) {
-			f.l.Error("Refusing to handshake with myself",
-				"vpnNetworks", vpnNetworks,
-				"from", via,
-				"certName", remoteCert.Certificate.Name(),
-				"certVersion", remoteCert.Certificate.Version(),
-				"fingerprint", remoteCert.Fingerprint,
-				"issuer", remoteCert.Certificate.Issuer(),
-			)
-			return nil, false, false
-		}
-		vpnAddrs[i] = network.Addr()
-		if f.myVpnNetworksTable.Contains(network.Addr()) {
-			anyVpnAddrsInCommon = true
-		}
-	}
-
-	if !via.IsRelayed {
-		if !f.lightHouse.GetRemoteAllowList().AllowAll(vpnAddrs, via.UdpAddr.Addr()) {
-			f.l.Debug("lighthouse.remote_allow_list denied incoming handshake",
-				"vpnAddrs", vpnAddrs, "from", via)
-			return nil, false, false
-		}
-	}
-
-	return vpnAddrs, anyVpnAddrsInCommon, true
-}
-
-// sendHandshakeResponse sends a handshake response via the appropriate transport.
-// cached is true when msg is a stored response being retransmitted because
-// the peer's stage-1 retransmit landed (the ErrAlreadySeen path); false on a
-// fresh response.
-func (hm *HandshakeManager) sendHandshakeResponse(via ViaSender, msg []byte, hostinfo *HostInfo, cached bool) {
-	if msg == nil {
-		return
-	}
-
-	f := hm.f
-	f.messageMetrics.Tx(header.Handshake, header.MessageSubType(msg[1]), 1)
-
-	// Common log fields. peerCert may be nil during intermediate
-	// multi-message flows (handshake hasn't completed yet); skip the cert
-	// block if so.
-	logFields := []any{
-		"vpnAddrs", hostinfo.vpnAddrs,
-		"handshake", m{"stage": uint64(2), "style": header.SubTypeName(header.Handshake, header.HandshakeIXPSK0)},
-		"cached", cached,
-		"initiatorIndex", hostinfo.remoteIndexId,
-		"responderIndex", hostinfo.localIndexId,
-	}
-	if peerCert := hostinfo.ConnectionState.peerCert; peerCert != nil {
-		logFields = append(logFields,
-			"certName", peerCert.Certificate.Name(),
-			"certVersion", peerCert.Certificate.Version(),
-			"fingerprint", peerCert.Fingerprint,
-			"issuer", peerCert.Certificate.Issuer(),
-		)
-	}
-
-	if !via.IsRelayed {
-		fields := append(logFields, "from", via)
-		err := f.outside.WriteTo(msg, via.UdpAddr)
-		if err != nil {
-			f.l.Error("Failed to send handshake message", append(fields, "error", err)...)
-		} else {
-			f.l.Info("Handshake message sent", fields...)
-		}
-	} else {
-		if via.relay == nil {
-			f.l.Error("Handshake send failed: both addr and via.relay are nil.")
-			return
-		}
-		hostinfo.relayState.InsertRelayTo(via.relayHI.vpnAddrs[0])
-		// We received a valid handshake on this relay, so make sure the relay
-		// state reflects that, in case it had been marked Disestablished.
-		via.relayHI.relayState.UpdateRelayForByIdxState(via.remoteIdx, Established)
-		f.SendVia(via.relayHI, via.relay, msg, make([]byte, 12), make([]byte, mtu), false)
-		f.l.Info("Handshake message sent", append(logFields, "relay", via.relayHI.vpnAddrs[0])...)
-	}
-}
-
-// handleCheckAndCompleteError handles errors from CheckAndComplete.
-// This only fires from the responder-side beginHandshake path, after the
-// peer cert has been validated and ConnectionState populated, so peerCert
-// is always non-nil for the cases that log it.
-func (hm *HandshakeManager) handleCheckAndCompleteError(err error, existing, hostinfo *HostInfo, via ViaSender) {
-	f := hm.f
-	peerCert := hostinfo.ConnectionState.peerCert
-	hsFields := m{"stage": uint64(1), "style": header.SubTypeName(header.Handshake, header.HandshakeIXPSK0)}
-
-	switch err {
-	case ErrAlreadySeen:
-		if existing.SetRemoteIfPreferred(f.hostMap, via) {
-			f.SendMessageToVpnAddr(header.Test, header.TestRequest, hostinfo.vpnAddrs[0], []byte(""), make([]byte, 12, 12), make([]byte, mtu))
-		}
-		// Resend the original response. The peer is committed to that response's
-		// ephemeral keys; a freshly-built one would have different keys and break
-		// the tunnel even though both sides "completed" the handshake.
-		if msg := existing.HandshakePacket[handshakePacketStage2]; msg != nil {
-			hm.sendHandshakeResponse(via, msg, existing, true)
-		}
-
-	case ErrExistingHostInfo:
-		f.l.Info("Handshake too old",
-			"vpnAddrs", hostinfo.vpnAddrs,
-			"from", via,
-			"certName", peerCert.Certificate.Name(),
-			"certVersion", peerCert.Certificate.Version(),
-			"fingerprint", peerCert.Fingerprint,
-			"issuer", peerCert.Certificate.Issuer(),
-			"oldHandshakeTime", existing.lastHandshakeTime,
-			"newHandshakeTime", hostinfo.lastHandshakeTime,
-			"initiatorIndex", hostinfo.remoteIndexId,
-			"responderIndex", hostinfo.localIndexId,
-			"handshake", hsFields,
-		)
-		f.SendMessageToVpnAddr(header.Test, header.TestRequest, hostinfo.vpnAddrs[0], []byte(""), make([]byte, 12, 12), make([]byte, mtu))
-
-	case ErrLocalIndexCollision:
-		f.l.Error("Failed to add HostInfo due to localIndex collision",
-			"vpnAddrs", hostinfo.vpnAddrs,
-			"from", via,
-			"certName", peerCert.Certificate.Name(),
-			"certVersion", peerCert.Certificate.Version(),
-			"fingerprint", peerCert.Fingerprint,
-			"issuer", peerCert.Certificate.Issuer(),
-			"localIndex", hostinfo.localIndexId,
-			"initiatorIndex", hostinfo.remoteIndexId,
-			"responderIndex", hostinfo.localIndexId,
-			"handshake", hsFields,
-		)
-
-	default:
-		f.l.Error("Failed to add HostInfo to HostMap",
-			"vpnAddrs", hostinfo.vpnAddrs,
-			"from", via,
-			"error", err,
-			"certName", peerCert.Certificate.Name(),
-			"certVersion", peerCert.Certificate.Version(),
-			"fingerprint", peerCert.Fingerprint,
-			"issuer", peerCert.Certificate.Issuer(),
-			"initiatorIndex", hostinfo.remoteIndexId,
-			"responderIndex", hostinfo.localIndexId,
-			"handshake", hsFields,
-		)
-	}
-}
-
-// certVerifier returns a CertVerifier that validates certs against the current CA pool.
-func (hm *HandshakeManager) certVerifier() handshake.CertVerifier {
-	return func(c cert.Certificate) (*cert.CachedCertificate, error) {
-		return hm.f.pki.GetCAPool().VerifyCertificate(time.Now(), c)
-	}
-}

handshake_manager_test.go

@@ -5,7 +5,6 @@ import (
 	"testing"
 	"time"
 
-	"github.com/gaissmai/bart"
 	"github.com/slackhq/nebula/cert"
 	"github.com/slackhq/nebula/header"
 	"github.com/slackhq/nebula/test"
@@ -28,7 +27,7 @@ func Test_NewHandshakeManagerVpnIp(t *testing.T) {
 		initiatingVersion: cert.Version1,
 		privateKey:        []byte{},
 		v1Cert:            &dummyCert{version: cert.Version1},
-		v1Credential:      nil,
+		v1HandshakeBytes:  []byte{},
 	}
 
 	blah := NewHandshakeManager(l, mainHM, lh, &udp.NoopConn{}, defaultHandshakeConfig)
@@ -101,137 +100,3 @@ func (mw *mockEncWriter) GetHostInfo(_ netip.Addr) *HostInfo {
 func (mw *mockEncWriter) GetCertState() *CertState {
 	return &CertState{initiatingVersion: cert.Version2}
 }
-
-func TestValidatePeerCert(t *testing.T) {
-	l := test.NewLogger()
-
-	myNetwork := netip.MustParsePrefix("10.0.0.1/24")
-	myAddrTable := new(bart.Lite)
-	myAddrTable.Insert(netip.PrefixFrom(myNetwork.Addr(), myNetwork.Addr().BitLen()))
-	myNetTable := new(bart.Lite)
-	myNetTable.Insert(myNetwork.Masked())
-
-	newHM := func() *HandshakeManager {
-		hm := NewHandshakeManager(l, newHostMap(l), newTestLighthouse(), &udp.NoopConn{}, defaultHandshakeConfig)
-		hm.f = &Interface{
-			handshakeManager:   hm,
-			pki:                &PKI{},
-			l:                  l,
-			myVpnAddrsTable:    myAddrTable,
-			myVpnNetworksTable: myNetTable,
-			lightHouse:         hm.lightHouse,
-		}
-		return hm
-	}
-
-	cached := func(networks ...netip.Prefix) *cert.CachedCertificate {
-		return &cert.CachedCertificate{
-			Certificate: &dummyCert{name: "peer", networks: networks},
-		}
-	}
-
-	via := ViaSender{
-		UdpAddr:   netip.MustParseAddrPort("198.51.100.7:4242"),
-		IsRelayed: true, // skip the remote allow list (covered separately)
-	}
-
-	t.Run("addr inside our networks sets anyVpnAddrsInCommon", func(t *testing.T) {
-		hm := newHM()
-		// 10.0.0.2 falls inside our 10.0.0.0/24
-		addrs, common, ok := hm.validatePeerCert(via, cached(netip.MustParsePrefix("10.0.0.2/24")))
-		assert.True(t, ok)
-		assert.True(t, common)
-		assert.Equal(t, []netip.Addr{netip.MustParseAddr("10.0.0.2")}, addrs)
-	})
-
-	t.Run("addr outside our networks leaves anyVpnAddrsInCommon false", func(t *testing.T) {
-		hm := newHM()
-		addrs, common, ok := hm.validatePeerCert(via, cached(netip.MustParsePrefix("192.168.1.5/24")))
-		assert.True(t, ok)
-		assert.False(t, common)
-		assert.Equal(t, []netip.Addr{netip.MustParseAddr("192.168.1.5")}, addrs)
-	})
-
-	t.Run("any matching network is enough", func(t *testing.T) {
-		hm := newHM()
-		addrs, common, ok := hm.validatePeerCert(via, cached(
-			netip.MustParsePrefix("192.168.1.5/24"),
-			netip.MustParsePrefix("10.0.0.42/24"),
-		))
-		assert.True(t, ok)
-		assert.True(t, common)
-		assert.Len(t, addrs, 2)
-	})
-
-	t.Run("self-handshake is rejected", func(t *testing.T) {
-		hm := newHM()
-		// 10.0.0.1 is in myVpnAddrsTable
-		addrs, common, ok := hm.validatePeerCert(via, cached(netip.MustParsePrefix("10.0.0.1/24")))
-		assert.False(t, ok)
-		assert.False(t, common)
-		assert.Nil(t, addrs)
-	})
-
-	t.Run("cert with no networks is rejected", func(t *testing.T) {
-		hm := newHM()
-		addrs, common, ok := hm.validatePeerCert(via, cached())
-		assert.False(t, ok)
-		assert.False(t, common)
-		assert.Nil(t, addrs)
-	})
-}
-
-func TestHandleIncomingDispatch(t *testing.T) {
-	l := test.NewLogger()
-
-	newHM := func() *HandshakeManager {
-		hm := NewHandshakeManager(l, newHostMap(l), newTestLighthouse(), &udp.NoopConn{}, defaultHandshakeConfig)
-		hm.f = &Interface{
-			handshakeManager: hm,
-			pki:              &PKI{},
-			l:                l,
-		}
-		return hm
-	}
-
-	via := ViaSender{
-		UdpAddr:   netip.MustParseAddrPort("198.51.100.7:4242"),
-		IsRelayed: true, // bypass remote allow list
-	}
-
-	// A packet body of zero length is fine for these tests: dispatch is
-	// gated on header fields, and we assert that we never reach noise/cert
-	// processing for any of the malformed shapes here.
-	pkt := make([]byte, header.Len)
-
-	t.Run("unsupported subtype dropped", func(t *testing.T) {
-		hm := newHM()
-		h := &header.H{Type: header.Handshake, Subtype: header.MessageSubType(99), MessageCounter: 1}
-		hm.HandleIncoming(via, pkt, h)
-		assert.Empty(t, hm.indexes, "no pending handshake should be created")
-	})
-
-	t.Run("stage-1 with non-zero RemoteIndex dropped", func(t *testing.T) {
-		hm := newHM()
-		h := &header.H{
-			Type:           header.Handshake,
-			Subtype:        header.HandshakeIXPSK0,
-			RemoteIndex:    0xdeadbeef,
-			MessageCounter: 1,
-		}
-		hm.HandleIncoming(via, pkt, h)
-		assert.Empty(t, hm.indexes, "spoofed stage-1 must not create a pending machine")
-	})
-
-	t.Run("continuation with no matching pending index dropped", func(t *testing.T) {
-		hm := newHM()
-		h := &header.H{
-			Type:           header.Handshake,
-			Subtype:        header.HandshakeIXPSK0,
-			RemoteIndex:    0xcafef00d,
-			MessageCounter: 2,
-		}
-		hm.HandleIncoming(via, pkt, h)
-		assert.Empty(t, hm.indexes, "orphan stage-2 must not create state")
-	})
-}

header/header.go

@@ -57,8 +57,6 @@ const (
 const (
 	TestRequest MessageSubType = 0
 	TestReply   MessageSubType = 1
-	MTUDProbeRequest MessageSubType = 2
-	MTUDProbeReply   MessageSubType = 3
 )
 
 const (
@@ -71,8 +69,6 @@ var ErrHeaderTooShort = errors.New("header is too short")
 var subTypeTestMap = map[MessageSubType]string{
 	TestRequest: "testRequest",
 	TestReply:   "testReply",
-	MTUDProbeRequest: "mtudProbeRequest",
-	MTUDProbeReply:   "mtudProbeReply",
 }
 
 var subTypeNoneMap = map[MessageSubType]string{0: "none"}

inside.go

@@ -9,10 +9,11 @@ import (
 	"github.com/slackhq/nebula/header"
 	"github.com/slackhq/nebula/iputil"
 	"github.com/slackhq/nebula/noiseutil"
+	"github.com/slackhq/nebula/overlay/batch"
 	"github.com/slackhq/nebula/routing"
 )
 
-func (f *Interface) consumeInsidePacket(packet []byte, fwPacket *firewall.Packet, nb, out []byte, q int, localCache firewall.ConntrackCache) {
+func (f *Interface) consumeInsidePacket(packet []byte, fwPacket *firewall.Packet, nb []byte, sendBatch batch.TxBatcher, rejectBuf []byte, q int, localCache firewall.ConntrackCache) {
 	err := newPacket(packet, false, fwPacket)
 	if err != nil {
 		if f.l.Enabled(context.Background(), slog.LevelDebug) {
@@ -57,7 +58,7 @@ func (f *Interface) consumeInsidePacket(packet []byte, fwPacket *firewall.Packet
 	})
 
 	if hostinfo == nil {
-		f.rejectInside(packet, out, q)
+		f.rejectInside(packet, rejectBuf, q)
 		if f.l.Enabled(context.Background(), slog.LevelDebug) {
 			f.l.Debug("dropping outbound packet, vpnAddr not in our vpn networks or in unsafe networks",
 				"vpnAddr", fwPacket.RemoteAddr,
@@ -73,10 +74,9 @@ func (f *Interface) consumeInsidePacket(packet []byte, fwPacket *firewall.Packet
 
 	dropReason := f.firewall.Drop(*fwPacket, false, hostinfo, f.pki.GetCAPool(), localCache)
 	if dropReason == nil {
-		f.sendNoMetrics(header.Message, 0, hostinfo.ConnectionState, hostinfo, netip.AddrPort{}, packet, nb, out, q)
+		f.sendInsideMessage(hostinfo, packet, nb, sendBatch, rejectBuf, q)
-
 	} else {
-		f.rejectInside(packet, out, q)
+		f.rejectInside(packet, rejectBuf, q)
 		if f.l.Enabled(context.Background(), slog.LevelDebug) {
 			hostinfo.logger(f.l).Debug("dropping outbound packet",
 				"fwPacket", fwPacket,
@@ -86,6 +86,67 @@ func (f *Interface) consumeInsidePacket(packet []byte, fwPacket *firewall.Packet
 	}
 }
 
+// sendInsideMessage encrypts a firewall-approved inside packet into the
+// caller's batch slot for later sendmmsg flush. When hostinfo.remote is not
+// valid we fall through to the relay slow path via the unbatched sendNoMetrics
+// so relay behavior is unchanged.
+func (f *Interface) sendInsideMessage(hostinfo *HostInfo, p, nb []byte, sendBatch batch.TxBatcher, rejectBuf []byte, q int) {
+	ci := hostinfo.ConnectionState
+	if ci.eKey == nil {
+		return
+	}
+
+	if !hostinfo.remote.IsValid() {
+		// Slow path: relay fallback. Reuse rejectBuf as the ciphertext
+		// scratch; sendNoMetrics arranges header space for SendVia.
+		f.sendNoMetrics(header.Message, 0, ci, hostinfo, netip.AddrPort{}, p, nb, rejectBuf, q)
+		return
+	}
+
+	scratch := sendBatch.Next()
+	if scratch == nil {
+		// Batch full: bypass batching and send this packet directly so we
+		// never drop traffic on over-subscribed iterations.
+		f.sendNoMetrics(header.Message, 0, ci, hostinfo, netip.AddrPort{}, p, nb, rejectBuf, q)
+		return
+	}
+
+	if noiseutil.EncryptLockNeeded {
+		ci.writeLock.Lock()
+	}
+	c := ci.messageCounter.Add(1)
+
+	out := header.Encode(scratch, header.Version, header.Message, 0, hostinfo.remoteIndexId, c)
+	f.connectionManager.Out(hostinfo)
+
+	if hostinfo.lastRebindCount != f.rebindCount {
+		//NOTE: there is an update hole if a tunnel isn't used and exactly 256 rebinds occur before the tunnel is
+		// finally used again. This tunnel would eventually be torn down and recreated if this action didn't help.
+		f.lightHouse.QueryServer(hostinfo.vpnAddrs[0])
+		hostinfo.lastRebindCount = f.rebindCount
+		if f.l.Enabled(context.Background(), slog.LevelDebug) {
+			hostinfo.logger(f.l).Debug("Lighthouse update triggered for punch due to rebind counter",
+				"vpnAddrs", hostinfo.vpnAddrs,
+			)
+		}
+	}
+
+	out, err := ci.eKey.EncryptDanger(out, out, p, c, nb)
+	if noiseutil.EncryptLockNeeded {
+		ci.writeLock.Unlock()
+	}
+	if err != nil {
+		hostinfo.logger(f.l).Error("Failed to encrypt outgoing packet",
+			"error", err,
+			"udpAddr", hostinfo.remote,
+			"counter", c,
+		)
+		return
+	}
+
+	sendBatch.Commit(len(out), hostinfo.remote)
+}
+
 func (f *Interface) rejectInside(packet []byte, out []byte, q int) {
 	if !f.firewall.InSendReject {
 		return

interface.go

@@ -4,7 +4,6 @@ import (
 	"context"
 	"errors"
 	"fmt"
-	"io"
 	"log/slog"
 	"net/netip"
 	"sync"
@@ -18,6 +17,8 @@ import (
 	"github.com/slackhq/nebula/firewall"
 	"github.com/slackhq/nebula/header"
 	"github.com/slackhq/nebula/overlay"
+	"github.com/slackhq/nebula/overlay/batch"
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/udp"
 )
 
@@ -34,7 +35,6 @@ type InterfaceConfig struct {
 	HandshakeManager   *HandshakeManager
 	lightHouse         *LightHouse
 	connectionManager  *connectionManager
-	pmtudManager       *pmtudManager
 	DropLocalBroadcast bool
 	DropMulticast      bool
 	routines           int
@@ -58,7 +58,6 @@ type Interface struct {
 	pki                   *PKI
 	firewall              *Firewall
 	connectionManager     *connectionManager
-	pmtudManager          *pmtudManager
 	handshakeManager      *HandshakeManager
 	dnsServer             *dnsServer
 	createTime            time.Time
@@ -90,7 +89,11 @@ type Interface struct {
 
 	ctx     context.Context
 	writers []udp.Conn
-	readers []io.ReadWriteCloser
+	readers []tio.Queue
+	// batchers is one per tun queue, wrapping readers[i].
+	// decryptToTun sends plaintext into the batch.RxBatcher;
+	// listenOut calls its Flush at the end of each UDP recvmmsg batch.
+	batchers []batch.RxBatcher
 	wg       sync.WaitGroup
 
 	// fatalErr holds the first unexpected reader error that caused shutdown.
@@ -189,7 +192,8 @@ func NewInterface(ctx context.Context, c *InterfaceConfig) (*Interface, error) {
 		routines:              c.routines,
 		version:               c.version,
 		writers:               make([]udp.Conn, c.routines),
-		readers:               make([]io.ReadWriteCloser, c.routines),
+		readers:               make([]tio.Queue, c.routines),
+		batchers:              make([]batch.RxBatcher, c.routines),
 		myVpnNetworks:         cs.myVpnNetworks,
 		myVpnNetworksTable:    cs.myVpnNetworksTable,
 		myVpnAddrs:            cs.myVpnAddrs,
@@ -197,7 +201,6 @@ func NewInterface(ctx context.Context, c *InterfaceConfig) (*Interface, error) {
 		myBroadcastAddrsTable: cs.myVpnBroadcastAddrsTable,
 		relayManager:          c.relayManager,
 		connectionManager:     c.connectionManager,
-		pmtudManager:          c.pmtudManager,
 		conntrackCacheTimeout: c.ConntrackCacheTimeout,
 
 		metricHandshakes: metrics.GetOrRegisterHistogram("handshakes", nil, metrics.NewExpDecaySample(1028, 0.015)),
@@ -215,7 +218,6 @@ func NewInterface(ctx context.Context, c *InterfaceConfig) (*Interface, error) {
 	ifce.reQueryWait.Store(int64(c.reQueryWait))
 
 	ifce.connectionManager.intf = ifce
-	ifce.pmtudManager.intf = ifce
 
 	return ifce, nil
 }
@@ -249,15 +251,16 @@ func (f *Interface) activate() error {
 	metrics.GetOrRegisterGauge("routines", nil).Update(int64(f.routines))
 
 	// Prepare n tun queues
-	var reader io.ReadWriteCloser = f.inside
 	for i := 0; i < f.routines; i++ {
 		if i > 0 {
-			reader, err = f.inside.NewMultiQueueReader()
+			if err = f.inside.NewMultiQueueReader(); err != nil {
-			if err != nil {
 				return err
 			}
 		}
-		f.readers[i] = reader
+	}
+	f.readers = f.inside.Readers()
+	for i := range f.readers {
+		f.batchers[i] = batch.NewTCPCoalescer(f.readers[i])
 	}
 
 	f.wg.Add(1) // for us to wait on Close() to return
@@ -315,14 +318,24 @@ func (f *Interface) listenOut(i int) {
 
 	ctCache := firewall.NewConntrackCacheTicker(f.ctx, f.l, f.conntrackCacheTimeout)
 	lhh := f.lightHouse.NewRequestHandler()
-	plaintext := make([]byte, udp.MTU)
 	h := &header.H{}
 	fwPacket := &firewall.Packet{}
 	nb := make([]byte, 12, 12)
 
-	err := li.ListenOut(func(fromUdpAddr netip.AddrPort, payload []byte) {
+	coalescer := f.batchers[i]
+
+	listener := func(fromUdpAddr netip.AddrPort, payload []byte) {
+		plaintext := f.batchers[i].Reserve(len(payload))
 		f.readOutsidePackets(ViaSender{UdpAddr: fromUdpAddr}, plaintext[:0], payload, h, fwPacket, lhh, nb, i, ctCache.Get())
-	})
+	}
+
+	flusher := func() {
+		if err := coalescer.Flush(); err != nil {
+			f.l.Error("Failed to flush tun coalescer", "error", err)
+		}
+	}
+
+	err := li.ListenOut(listener, flusher)
 
 	if err != nil && !f.closed.Load() {
 		f.l.Error("Error while reading inbound packet, closing", "error", err)
@@ -332,16 +345,16 @@ func (f *Interface) listenOut(i int) {
 	f.l.Debug("underlay reader is done", "reader", i)
 }
 
-func (f *Interface) listenIn(reader io.ReadWriteCloser, i int) {
+func (f *Interface) listenIn(reader tio.Queue, i int) {
-	packet := make([]byte, mtu)
+	rejectBuf := make([]byte, mtu)
-	out := make([]byte, mtu)
+	sb := batch.NewSendBatch(batch.SendBatchCap, udp.MTU+32)
 	fwPacket := &firewall.Packet{}
 	nb := make([]byte, 12, 12)
 
 	conntrackCache := firewall.NewConntrackCacheTicker(f.ctx, f.l, f.conntrackCacheTimeout)
 
 	for {
-		n, err := reader.Read(packet)
+		pkts, err := reader.Read()
 		if err != nil {
 			if !f.closed.Load() {
 				f.l.Error("Error while reading outbound packet, closing", "error", err, "reader", i)
@@ -350,12 +363,29 @@ func (f *Interface) listenIn(reader io.ReadWriteCloser, i int) {
 			break
 		}
 
-		f.consumeInsidePacket(packet[:n], fwPacket, nb, out, i, conntrackCache.Get())
+		sb.Reset()
+		for _, pkt := range pkts {
+			if sb.Len() >= sb.Cap() {
+				f.flushBatch(sb, i)
+				sb.Reset()
+			}
+			f.consumeInsidePacket(pkt, fwPacket, nb, sb, rejectBuf, i, conntrackCache.Get())
+		}
+		if sb.Len() > 0 {
+			f.flushBatch(sb, i)
+		}
 	}
 
 	f.l.Debug("overlay reader is done", "reader", i)
 }
 
+func (f *Interface) flushBatch(sb batch.TxBatcher, q int) {
+	bufs, dsts := sb.Get()
+	if err := f.writers[q].WriteBatch(bufs, dsts); err != nil {
+		f.l.Error("Failed to write outgoing batch", "error", err, "writer", q)
+	}
+}
+
 func (f *Interface) RegisterConfigChangeCallbacks(c *config.C) {
 	c.RegisterReloadCallback(f.reloadFirewall)
 	c.RegisterReloadCallback(f.reloadSendRecvError)

main.go

@@ -172,7 +172,6 @@ func Main(c *config.C, configTest bool, buildVersion string, l *slog.Logger, dev
 	hostMap := NewHostMapFromConfig(l, c)
 	punchy := NewPunchyFromConfig(l, c)
 	connManager := newConnectionManagerFromConfig(l, c, hostMap, punchy)
-	pmtudMgr := newPMTUDManagerFromConfig(l, c, tun)
 	lightHouse, err := NewLightHouseFromConfig(ctx, l, c, pki.getCertState(), udpConns[0], punchy)
 	if err != nil {
 		return nil, util.ContextualizeIfNeeded("Failed to initialize lighthouse handler", err)
@@ -185,10 +184,14 @@ func Main(c *config.C, configTest bool, buildVersion string, l *slog.Logger, dev
 		messageMetrics = newMessageMetricsOnlyRecvError()
 	}
 
+	useRelays := c.GetBool("relay.use_relays", DefaultUseRelays) && !c.GetBool("relay.am_relay", false)
+
 	handshakeConfig := HandshakeConfig{
 		tryInterval:   c.GetDuration("handshakes.try_interval", DefaultHandshakeTryInterval),
 		retries:       int64(c.GetInt("handshakes.retries", DefaultHandshakeRetries)),
 		triggerBuffer: c.GetInt("handshakes.trigger_buffer", DefaultHandshakeTriggerBuffer),
+		useRelays:     useRelays,
+
 		messageMetrics: messageMetrics,
 	}
 
@@ -209,7 +212,6 @@ func Main(c *config.C, configTest bool, buildVersion string, l *slog.Logger, dev
 		DnsServer:             ds,
 		HandshakeManager:      handshakeManager,
 		connectionManager:     connManager,
-		pmtudManager:          pmtudMgr,
 		lightHouse:            lightHouse,
 		tryPromoteEvery:       c.GetUint32("counters.try_promote", defaultPromoteEvery),
 		reQueryEvery:          c.GetUint32("counters.requery_every_packets", defaultReQueryEvery),
@@ -268,7 +270,6 @@ func Main(c *config.C, configTest bool, buildVersion string, l *slog.Logger, dev
 		dnsStart:               ds.Start,
 		lighthouseStart:        lightHouse.StartUpdateWorker,
 		connectionManagerStart: connManager.Start,
-		pmtudManagerStart:      pmtudMgr.Start,
 	}, nil
 }
 

nebula.pb.go

@@ -124,7 +124,7 @@ func (x NebulaControl_MessageType) String() string {
 }
 
 func (NebulaControl_MessageType) EnumDescriptor() ([]byte, []int) {
-	return fileDescriptor_2d65afa7693df5ef, []int{6, 0}
+	return fileDescriptor_2d65afa7693df5ef, []int{8, 0}
 }
 
 type NebulaMeta struct {
@@ -489,6 +489,142 @@ func (m *NebulaPing) GetTime() uint64 {
 	return 0
 }
 
+type NebulaHandshake struct {
+	Details *NebulaHandshakeDetails `protobuf:"bytes,1,opt,name=Details,proto3" json:"Details,omitempty"`
+	Hmac    []byte                  `protobuf:"bytes,2,opt,name=Hmac,proto3" json:"Hmac,omitempty"`
+}
+
+func (m *NebulaHandshake) Reset()         { *m = NebulaHandshake{} }
+func (m *NebulaHandshake) String() string { return proto.CompactTextString(m) }
+func (*NebulaHandshake) ProtoMessage()    {}
+func (*NebulaHandshake) Descriptor() ([]byte, []int) {
+	return fileDescriptor_2d65afa7693df5ef, []int{6}
+}
+func (m *NebulaHandshake) XXX_Unmarshal(b []byte) error {
+	return m.Unmarshal(b)
+}
+func (m *NebulaHandshake) XXX_Marshal(b []byte, deterministic bool) ([]byte, error) {
+	if deterministic {
+		return xxx_messageInfo_NebulaHandshake.Marshal(b, m, deterministic)
+	} else {
+		b = b[:cap(b)]
+		n, err := m.MarshalToSizedBuffer(b)
+		if err != nil {
+			return nil, err
+		}
+		return b[:n], nil
+	}
+}
+func (m *NebulaHandshake) XXX_Merge(src proto.Message) {
+	xxx_messageInfo_NebulaHandshake.Merge(m, src)
+}
+func (m *NebulaHandshake) XXX_Size() int {
+	return m.Size()
+}
+func (m *NebulaHandshake) XXX_DiscardUnknown() {
+	xxx_messageInfo_NebulaHandshake.DiscardUnknown(m)
+}
+
+var xxx_messageInfo_NebulaHandshake proto.InternalMessageInfo
+
+func (m *NebulaHandshake) GetDetails() *NebulaHandshakeDetails {
+	if m != nil {
+		return m.Details
+	}
+	return nil
+}
+
+func (m *NebulaHandshake) GetHmac() []byte {
+	if m != nil {
+		return m.Hmac
+	}
+	return nil
+}
+
+type NebulaHandshakeDetails struct {
+	Cert           []byte `protobuf:"bytes,1,opt,name=Cert,proto3" json:"Cert,omitempty"`
+	InitiatorIndex uint32 `protobuf:"varint,2,opt,name=InitiatorIndex,proto3" json:"InitiatorIndex,omitempty"`
+	ResponderIndex uint32 `protobuf:"varint,3,opt,name=ResponderIndex,proto3" json:"ResponderIndex,omitempty"`
+	Cookie         uint64 `protobuf:"varint,4,opt,name=Cookie,proto3" json:"Cookie,omitempty"`
+	Time           uint64 `protobuf:"varint,5,opt,name=Time,proto3" json:"Time,omitempty"`
+	CertVersion    uint32 `protobuf:"varint,8,opt,name=CertVersion,proto3" json:"CertVersion,omitempty"`
+}
+
+func (m *NebulaHandshakeDetails) Reset()         { *m = NebulaHandshakeDetails{} }
+func (m *NebulaHandshakeDetails) String() string { return proto.CompactTextString(m) }
+func (*NebulaHandshakeDetails) ProtoMessage()    {}
+func (*NebulaHandshakeDetails) Descriptor() ([]byte, []int) {
+	return fileDescriptor_2d65afa7693df5ef, []int{7}
+}
+func (m *NebulaHandshakeDetails) XXX_Unmarshal(b []byte) error {
+	return m.Unmarshal(b)
+}
+func (m *NebulaHandshakeDetails) XXX_Marshal(b []byte, deterministic bool) ([]byte, error) {
+	if deterministic {
+		return xxx_messageInfo_NebulaHandshakeDetails.Marshal(b, m, deterministic)
+	} else {
+		b = b[:cap(b)]
+		n, err := m.MarshalToSizedBuffer(b)
+		if err != nil {
+			return nil, err
+		}
+		return b[:n], nil
+	}
+}
+func (m *NebulaHandshakeDetails) XXX_Merge(src proto.Message) {
+	xxx_messageInfo_NebulaHandshakeDetails.Merge(m, src)
+}
+func (m *NebulaHandshakeDetails) XXX_Size() int {
+	return m.Size()
+}
+func (m *NebulaHandshakeDetails) XXX_DiscardUnknown() {
+	xxx_messageInfo_NebulaHandshakeDetails.DiscardUnknown(m)
+}
+
+var xxx_messageInfo_NebulaHandshakeDetails proto.InternalMessageInfo
+
+func (m *NebulaHandshakeDetails) GetCert() []byte {
+	if m != nil {
+		return m.Cert
+	}
+	return nil
+}
+
+func (m *NebulaHandshakeDetails) GetInitiatorIndex() uint32 {
+	if m != nil {
+		return m.InitiatorIndex
+	}
+	return 0
+}
+
+func (m *NebulaHandshakeDetails) GetResponderIndex() uint32 {
+	if m != nil {
+		return m.ResponderIndex
+	}
+	return 0
+}
+
+func (m *NebulaHandshakeDetails) GetCookie() uint64 {
+	if m != nil {
+		return m.Cookie
+	}
+	return 0
+}
+
+func (m *NebulaHandshakeDetails) GetTime() uint64 {
+	if m != nil {
+		return m.Time
+	}
+	return 0
+}
+
+func (m *NebulaHandshakeDetails) GetCertVersion() uint32 {
+	if m != nil {
+		return m.CertVersion
+	}
+	return 0
+}
+
 type NebulaControl struct {
 	Type                NebulaControl_MessageType `protobuf:"varint,1,opt,name=Type,proto3,enum=nebula.NebulaControl_MessageType" json:"Type,omitempty"`
 	InitiatorRelayIndex uint32                    `protobuf:"varint,2,opt,name=InitiatorRelayIndex,proto3" json:"InitiatorRelayIndex,omitempty"`
@@ -503,7 +639,7 @@ func (m *NebulaControl) Reset()         { *m = NebulaControl{} }
 func (m *NebulaControl) String() string { return proto.CompactTextString(m) }
 func (*NebulaControl) ProtoMessage()    {}
 func (*NebulaControl) Descriptor() ([]byte, []int) {
-	return fileDescriptor_2d65afa7693df5ef, []int{6}
+	return fileDescriptor_2d65afa7693df5ef, []int{8}
 }
 func (m *NebulaControl) XXX_Unmarshal(b []byte) error {
 	return m.Unmarshal(b)
@@ -593,55 +729,65 @@ func init() {
 	proto.RegisterType((*V4AddrPort)(nil), "nebula.V4AddrPort")
 	proto.RegisterType((*V6AddrPort)(nil), "nebula.V6AddrPort")
 	proto.RegisterType((*NebulaPing)(nil), "nebula.NebulaPing")
+	proto.RegisterType((*NebulaHandshake)(nil), "nebula.NebulaHandshake")
+	proto.RegisterType((*NebulaHandshakeDetails)(nil), "nebula.NebulaHandshakeDetails")
 	proto.RegisterType((*NebulaControl)(nil), "nebula.NebulaControl")
 }
 
 func init() { proto.RegisterFile("nebula.proto", fileDescriptor_2d65afa7693df5ef) }
 
 var fileDescriptor_2d65afa7693df5ef = []byte{
-	// 665 bytes of a gzipped FileDescriptorProto
+	// 785 bytes of a gzipped FileDescriptorProto
-	0x1f, 0x8b, 0x08, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, 0xff, 0x84, 0x54, 0xcd, 0x6e, 0xd3, 0x5c,
+	0x1f, 0x8b, 0x08, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, 0xff, 0x84, 0x55, 0xcd, 0x6e, 0xeb, 0x44,
-	0x10, 0x8d, 0x1d, 0x27, 0x69, 0x27, 0x4d, 0x3e, 0x7f, 0x53, 0x51, 0x12, 0x24, 0xac, 0xe0, 0x45,
+	0x14, 0x8e, 0x1d, 0x27, 0x4e, 0x4f, 0x7e, 0xae, 0x39, 0x15, 0xc1, 0x41, 0x22, 0x0a, 0x5e, 0x54,
-	0x55, 0xb1, 0x48, 0x51, 0x5a, 0xba, 0xa6, 0x2d, 0x42, 0xa9, 0xd4, 0x9f, 0x70, 0x55, 0x8a, 0xc4,
+	0x57, 0x2c, 0x72, 0x51, 0x5a, 0xae, 0x58, 0x72, 0x1b, 0x84, 0xd2, 0xaa, 0x3f, 0x61, 0x54, 0x8a,
-	0xce, 0xb5, 0x2f, 0x8d, 0x55, 0xc7, 0x37, 0xb5, 0x6f, 0x50, 0xf3, 0x16, 0x3c, 0x0c, 0x0f, 0x01,
+	0xc4, 0x06, 0xb9, 0xf6, 0xd0, 0x58, 0x71, 0x3c, 0xa9, 0x3d, 0x41, 0xcd, 0x5b, 0xf0, 0x30, 0x3c,
-	0xbb, 0x2e, 0x59, 0xa2, 0x66, 0xc9, 0x92, 0x17, 0x40, 0xf7, 0xfa, 0xbf, 0x31, 0xb0, 0xbb, 0x33,
+	0x04, 0xec, 0xba, 0x42, 0x2c, 0x51, 0xbb, 0x64, 0xc9, 0x0b, 0xa0, 0x19, 0xff, 0x27, 0x86, 0xbb,
-	0xe7, 0x9c, 0x99, 0xc9, 0xc9, 0x8c, 0x61, 0xcd, 0xa7, 0x97, 0x33, 0xcf, 0xea, 0x4f, 0x03, 0xc6,
+	0x9b, 0x73, 0xbe, 0xef, 0x3b, 0x73, 0xe6, 0xf3, 0x9c, 0x31, 0x74, 0x02, 0x7a, 0xb7, 0xf1, 0xed,
-	0x19, 0xd6, 0xa3, 0xc8, 0xfc, 0xa9, 0x02, 0x9c, 0xca, 0xe7, 0x09, 0xe5, 0x16, 0x0e, 0x40, 0x3b,
+	0xf1, 0x3a, 0x64, 0x9c, 0x61, 0x33, 0x8e, 0xac, 0xbf, 0x55, 0x80, 0x2b, 0xb9, 0xbc, 0xa4, 0xdc,
-	0x9f, 0x4f, 0x69, 0x47, 0xe9, 0x29, 0x5b, 0xed, 0x81, 0xd1, 0x8f, 0x35, 0x19, 0xa3, 0x7f, 0x42,
+	0xc6, 0x09, 0x68, 0x37, 0xdb, 0x35, 0x35, 0x95, 0x91, 0xf2, 0xba, 0x37, 0x19, 0x8e, 0x13, 0x4d,
-	0xc3, 0xd0, 0xba, 0xa2, 0x82, 0x45, 0x24, 0x17, 0x77, 0xa0, 0xf1, 0x9a, 0x72, 0xcb, 0xf5, 0xc2,
+	0xce, 0x18, 0x5f, 0xd2, 0x28, 0xb2, 0xef, 0xa9, 0x60, 0x11, 0xc9, 0xc5, 0x63, 0xd0, 0xbf, 0xa6,
-	0x8e, 0xda, 0x53, 0xb6, 0x9a, 0x83, 0xee, 0xb2, 0x2c, 0x26, 0x90, 0x84, 0x69, 0xfe, 0x52, 0xa0,
+	0xdc, 0xf6, 0xfc, 0xc8, 0x54, 0x47, 0xca, 0xeb, 0xf6, 0x64, 0xb0, 0x2f, 0x4b, 0x08, 0x24, 0x65,
-	0x99, 0x2b, 0x85, 0x2b, 0xa0, 0x9d, 0x32, 0x9f, 0xea, 0x15, 0x6c, 0xc1, 0xea, 0x90, 0x85, 0xfc,
+	0x5a, 0xff, 0x28, 0xd0, 0x2e, 0x94, 0xc2, 0x16, 0x68, 0x57, 0x2c, 0xa0, 0x46, 0x0d, 0xbb, 0x70,
-	0xed, 0x8c, 0x06, 0x73, 0x5d, 0x41, 0x84, 0x76, 0x1a, 0x12, 0x3a, 0xf5, 0xe6, 0xba, 0x8a, 0x4f,
+	0x30, 0x63, 0x11, 0xff, 0x76, 0x43, 0xc3, 0xad, 0xa1, 0x20, 0x42, 0x2f, 0x0b, 0x09, 0x5d, 0xfb,
-	0x60, 0x43, 0xe4, 0xde, 0x4d, 0x1d, 0x8b, 0xd3, 0x53, 0xc6, 0xdd, 0x8f, 0xae, 0x6d, 0x71, 0x97,
+	0x5b, 0x43, 0xc5, 0x8f, 0xa1, 0x2f, 0x72, 0xdf, 0xad, 0x5d, 0x9b, 0xd3, 0x2b, 0xc6, 0xbd, 0x9f,
-	0xf9, 0x7a, 0x15, 0xbb, 0xf0, 0x48, 0x60, 0x27, 0xec, 0x13, 0x75, 0x0a, 0x90, 0x96, 0x40, 0xa3,
+	0x3c, 0xc7, 0xe6, 0x1e, 0x0b, 0x8c, 0x3a, 0x0e, 0xe0, 0x43, 0x81, 0x5d, 0xb2, 0x9f, 0xa9, 0x5b,
-	0x99, 0x6f, 0x8f, 0x0b, 0x50, 0x0d, 0xdb, 0x00, 0x02, 0x7a, 0x3f, 0x66, 0xd6, 0xc4, 0xd5, 0xeb,
+	0x82, 0xb4, 0x14, 0x9a, 0x6f, 0x02, 0x67, 0x51, 0x82, 0x1a, 0xd8, 0x03, 0x10, 0xd0, 0xf7, 0x0b,
-	0xb8, 0x0e, 0xff, 0x65, 0x71, 0xd4, 0xb6, 0x21, 0x26, 0x1b, 0x59, 0x7c, 0x7c, 0x38, 0xa6, 0xf6,
+	0x66, 0xaf, 0x3c, 0xa3, 0x89, 0x87, 0xf0, 0x2a, 0x8f, 0xe3, 0x6d, 0x75, 0xd1, 0xd9, 0xdc, 0xe6,
-	0xb5, 0xbe, 0x22, 0x26, 0x4b, 0xc3, 0x88, 0xb2, 0x8a, 0x4f, 0xa1, 0x5b, 0x3e, 0xd9, 0xbe, 0x7d,
+	0x8b, 0xe9, 0x82, 0x3a, 0x4b, 0xa3, 0x25, 0x3a, 0xcb, 0xc2, 0x98, 0x72, 0x80, 0x9f, 0xc0, 0xa0,
-	0xad, 0x83, 0xf9, 0x4d, 0x85, 0xff, 0x97, 0x4c, 0x41, 0x13, 0xe0, 0xcc, 0x73, 0x2e, 0xa6, 0xfe,
+	0xba, 0xb3, 0x77, 0xce, 0xd2, 0x00, 0xeb, 0x77, 0x15, 0x3e, 0xd8, 0x33, 0x05, 0x2d, 0x80, 0x6b,
-	0xbe, 0xe3, 0x04, 0xd2, 0xfa, 0xd6, 0x81, 0xda, 0x51, 0x48, 0x2e, 0x8b, 0x9b, 0xd0, 0x48, 0x08,
+	0xdf, 0xbd, 0x5d, 0x07, 0xef, 0x5c, 0x37, 0x94, 0xd6, 0x77, 0x4f, 0x55, 0x53, 0x21, 0x85, 0x2c,
-	0x75, 0x69, 0xf2, 0x5a, 0x62, 0xb2, 0xc8, 0x91, 0x04, 0xc4, 0x3e, 0xe8, 0x67, 0x9e, 0x43, 0xa8,
+	0x1e, 0x81, 0x9e, 0x12, 0x9a, 0xd2, 0xe4, 0x4e, 0x6a, 0xb2, 0xc8, 0x91, 0x14, 0xc4, 0x31, 0x18,
-	0x67, 0xcd, 0xe3, 0x54, 0xd8, 0xa9, 0xf5, 0xaa, 0x71, 0xc5, 0x25, 0x0c, 0x07, 0xd0, 0x2a, 0x92,
+	0xd7, 0xbe, 0x4b, 0xa8, 0x6f, 0x6f, 0x93, 0x54, 0x64, 0x36, 0x46, 0xf5, 0xa4, 0xe2, 0x1e, 0x86,
-	0x1b, 0xbd, 0xea, 0x52, 0xf5, 0x22, 0x05, 0x77, 0xa1, 0x79, 0xb1, 0x2b, 0x9e, 0x23, 0x16, 0x70,
+	0x13, 0xe8, 0x96, 0xc9, 0xfa, 0xa8, 0xbe, 0x57, 0xbd, 0x4c, 0xc1, 0x13, 0x68, 0xdf, 0x9e, 0x88,
-	0xf1, 0xa7, 0x0b, 0x05, 0x26, 0x8a, 0x0c, 0x22, 0x79, 0x9a, 0x54, 0xed, 0x65, 0x2a, 0xed, 0x81,
+	0xe5, 0x9c, 0x85, 0x5c, 0x7c, 0x74, 0xa1, 0xc0, 0x54, 0x91, 0x43, 0xa4, 0x48, 0x93, 0xaa, 0xb7,
-	0x6a, 0x2f, 0xa7, 0xca, 0x68, 0xd8, 0x81, 0x86, 0xcd, 0x66, 0x3e, 0xa7, 0x41, 0xa7, 0x2a, 0x8c,
+	0xb9, 0x4a, 0xdb, 0x51, 0xbd, 0x2d, 0xa8, 0x72, 0x1a, 0x9a, 0xa0, 0x3b, 0x6c, 0x13, 0x70, 0x1a,
-	0x21, 0x49, 0x68, 0x6e, 0x82, 0x26, 0x7f, 0x71, 0x1b, 0xd4, 0xa1, 0x2b, 0x5d, 0xd3, 0x88, 0x3a,
+	0x9a, 0x75, 0x61, 0x0c, 0x49, 0x43, 0xeb, 0x08, 0x34, 0x79, 0xe2, 0x1e, 0xa8, 0x33, 0x4f, 0xba,
-	0x74, 0x45, 0x7c, 0xcc, 0xe4, 0x26, 0x6a, 0x44, 0x3d, 0x66, 0xe6, 0x2e, 0x40, 0x36, 0x06, 0x62,
+	0xa6, 0x11, 0x75, 0xe6, 0x89, 0xf8, 0x82, 0xc9, 0x9b, 0xa8, 0x11, 0xf5, 0x82, 0x59, 0x27, 0x00,
-	0xa4, 0x8a, 0x5c, 0x26, 0x51, 0x05, 0x04, 0x4d, 0x60, 0x52, 0xd3, 0x22, 0xf2, 0x6d, 0xbe, 0x02,
+	0x79, 0x1b, 0x88, 0xb1, 0x2a, 0x76, 0x99, 0xc4, 0x15, 0x10, 0x34, 0x81, 0x49, 0x4d, 0x97, 0xc8,
-	0xc8, 0xc6, 0xf8, 0x57, 0x8f, 0xb4, 0x42, 0x35, 0x57, 0xe1, 0x36, 0x39, 0xac, 0x91, 0xeb, 0x5f,
+	0xb5, 0xf5, 0x15, 0x40, 0xde, 0xc6, 0xfb, 0xf6, 0xc8, 0x2a, 0xd4, 0x0b, 0x15, 0x1e, 0xd3, 0xc1,
-	0xfd, 0xfd, 0xb0, 0x04, 0xa3, 0xe4, 0xb0, 0x10, 0xb4, 0x73, 0x77, 0x42, 0xe3, 0x3e, 0xf2, 0x6d,
+	0x9a, 0x7b, 0xc1, 0xfd, 0xff, 0x0f, 0x96, 0x60, 0x54, 0x0c, 0x16, 0x82, 0x76, 0xe3, 0xad, 0x68,
-	0x9a, 0x4b, 0x67, 0x23, 0xc4, 0x7a, 0x05, 0x57, 0xa1, 0x16, 0x2d, 0xa1, 0x62, 0x7e, 0xa9, 0x42,
+	0xb2, 0x8f, 0x5c, 0x5b, 0xd6, 0xde, 0xd8, 0x08, 0xb1, 0x51, 0xc3, 0x03, 0x68, 0xc4, 0x97, 0x50,
-	0x2b, 0x2a, 0x7c, 0xc8, 0x7c, 0x1e, 0x30, 0x0f, 0x5f, 0x16, 0xba, 0x3f, 0x2b, 0x76, 0x8f, 0x49,
+	0xb1, 0x7e, 0x84, 0x57, 0x71, 0xdd, 0x99, 0x1d, 0xb8, 0xd1, 0xc2, 0x5e, 0x52, 0xfc, 0x32, 0x9f,
-	0x25, 0x03, 0xbc, 0x80, 0xf5, 0x23, 0xdf, 0xe5, 0xae, 0xc5, 0x59, 0x20, 0x57, 0xe0, 0xc8, 0x77,
+	0x51, 0x45, 0x5e, 0x9f, 0x9d, 0x0e, 0x32, 0xe6, 0xee, 0xa0, 0x8a, 0x26, 0x66, 0x2b, 0xdb, 0x91,
-	0xe8, 0x6d, 0xec, 0x53, 0x19, 0x24, 0x14, 0x84, 0x86, 0x53, 0xe6, 0x3b, 0x34, 0xaf, 0x88, 0x7c,
+	0x4d, 0x74, 0x88, 0x5c, 0x5b, 0x7f, 0x28, 0xd0, 0xaf, 0xd6, 0x09, 0xfa, 0x94, 0x86, 0x5c, 0xee,
-	0x29, 0x83, 0xf0, 0x39, 0xb4, 0x93, 0xa5, 0x3c, 0x67, 0xf2, 0xaf, 0xd1, 0xd2, 0x03, 0x78, 0x80,
+	0xd2, 0x21, 0x72, 0x8d, 0x47, 0xd0, 0x3b, 0x0b, 0x3c, 0xee, 0xd9, 0x9c, 0x85, 0x67, 0x81, 0x4b,
-	0xe4, 0x97, 0xfb, 0x4d, 0xc0, 0x26, 0x92, 0x5d, 0x4b, 0xd9, 0x4b, 0x18, 0xf6, 0xa1, 0x99, 0x2f,
+	0x1f, 0x13, 0xa7, 0x77, 0xb2, 0x82, 0x47, 0x68, 0xb4, 0x66, 0x81, 0x4b, 0x13, 0x5e, 0xec, 0xe7,
-	0x5c, 0x76, 0x38, 0x79, 0x42, 0x7a, 0x0c, 0x69, 0xf1, 0x46, 0x89, 0xa2, 0x48, 0x31, 0x87, 0x7f,
+	0x4e, 0x16, 0xfb, 0xd0, 0x9c, 0x32, 0xb6, 0xf4, 0xa8, 0xa9, 0x49, 0x67, 0x92, 0x28, 0xf3, 0xab,
-	0xfa, 0x8e, 0x6d, 0x00, 0x1e, 0x06, 0xd4, 0xe2, 0x54, 0xf2, 0x09, 0xbd, 0x99, 0xd1, 0x90, 0xeb,
+	0x91, 0xfb, 0x85, 0x23, 0x68, 0x8b, 0x1e, 0x6e, 0x69, 0x18, 0x79, 0x2c, 0x30, 0x5b, 0xb2, 0x60,
-	0x0a, 0x3e, 0x86, 0xf5, 0x42, 0x5e, 0x58, 0x12, 0x52, 0x5d, 0x3d, 0xd8, 0xf9, 0x7a, 0x6f, 0x28,
+	0x31, 0x75, 0xae, 0xb5, 0x9a, 0x86, 0x7e, 0xae, 0xb5, 0x74, 0xa3, 0x65, 0xfd, 0x5a, 0x87, 0x6e,
-	0x77, 0xf7, 0x86, 0xf2, 0xe3, 0xde, 0x50, 0x3e, 0x2f, 0x8c, 0xca, 0xdd, 0xc2, 0xa8, 0x7c, 0x5f,
+	0x7c, 0xb0, 0x29, 0x0b, 0x78, 0xc8, 0x7c, 0xfc, 0xa2, 0xf4, 0xdd, 0x3e, 0x2d, 0xbb, 0x96, 0x90,
-	0x18, 0x95, 0x0f, 0xdd, 0x2b, 0x97, 0x8f, 0x67, 0x97, 0x7d, 0x9b, 0x4d, 0xb6, 0x43, 0xcf, 0xb2,
+	0x2a, 0x3e, 0xdd, 0xe7, 0x70, 0x98, 0x1d, 0x4e, 0x0e, 0x4f, 0xf1, 0xdc, 0x55, 0x90, 0x50, 0x64,
-	0xaf, 0xc7, 0x37, 0xdb, 0xd1, 0x48, 0x97, 0x75, 0xf9, 0x39, 0xdf, 0xf9, 0x1d, 0x00, 0x00, 0xff,
+	0xc7, 0x2c, 0x28, 0x62, 0x07, 0xaa, 0x20, 0xfc, 0x0c, 0x7a, 0xe9, 0x38, 0xdf, 0x30, 0x79, 0xa9,
-	0xff, 0x51, 0x0a, 0xe3, 0xd7, 0xde, 0x05, 0x00, 0x00,
+	0xb5, 0xec, 0xe9, 0xd8, 0x41, 0x8a, 0xcf, 0xc2, 0x37, 0x21, 0x5b, 0x49, 0x76, 0x23, 0x63, 0xef,
+	0x61, 0x38, 0x86, 0x76, 0xb1, 0x70, 0xd5, 0x93, 0x53, 0x24, 0x64, 0xcf, 0x48, 0x56, 0x5c, 0xaf,
+	0x50, 0x94, 0x29, 0xd6, 0xec, 0xbf, 0xfe, 0x00, 0x7d, 0xc0, 0x69, 0x48, 0x6d, 0x4e, 0x25, 0x9f,
+	0xd0, 0x87, 0x0d, 0x8d, 0xb8, 0xa1, 0xe0, 0x47, 0x70, 0x58, 0xca, 0x0b, 0x4b, 0x22, 0x6a, 0xa8,
+	0xa7, 0xc7, 0xbf, 0x3d, 0x0f, 0x95, 0xa7, 0xe7, 0xa1, 0xf2, 0xd7, 0xf3, 0x50, 0xf9, 0xe5, 0x65,
+	0x58, 0x7b, 0x7a, 0x19, 0xd6, 0xfe, 0x7c, 0x19, 0xd6, 0x7e, 0x18, 0xdc, 0x7b, 0x7c, 0xb1, 0xb9,
+	0x1b, 0x3b, 0x6c, 0xf5, 0x26, 0xf2, 0x6d, 0x67, 0xb9, 0x78, 0x78, 0x13, 0xb7, 0x74, 0xd7, 0x94,
+	0x3f, 0xc2, 0xe3, 0x7f, 0x03, 0x00, 0x00, 0xff, 0xff, 0xea, 0x6f, 0xbc, 0x50, 0x18, 0x07, 0x00,
+	0x00,
 }
 
 func (m *NebulaMeta) Marshal() (dAtA []byte, err error) {
@@ -926,6 +1072,103 @@ func (m *NebulaPing) MarshalToSizedBuffer(dAtA []byte) (int, error) {
 	return len(dAtA) - i, nil
 }
 
+func (m *NebulaHandshake) Marshal() (dAtA []byte, err error) {
+	size := m.Size()
+	dAtA = make([]byte, size)
+	n, err := m.MarshalToSizedBuffer(dAtA[:size])
+	if err != nil {
+		return nil, err
+	}
+	return dAtA[:n], nil
+}
+
+func (m *NebulaHandshake) MarshalTo(dAtA []byte) (int, error) {
+	size := m.Size()
+	return m.MarshalToSizedBuffer(dAtA[:size])
+}
+
+func (m *NebulaHandshake) MarshalToSizedBuffer(dAtA []byte) (int, error) {
+	i := len(dAtA)
+	_ = i
+	var l int
+	_ = l
+	if len(m.Hmac) > 0 {
+		i -= len(m.Hmac)
+		copy(dAtA[i:], m.Hmac)
+		i = encodeVarintNebula(dAtA, i, uint64(len(m.Hmac)))
+		i--
+		dAtA[i] = 0x12
+	}
+	if m.Details != nil {
+		{
+			size, err := m.Details.MarshalToSizedBuffer(dAtA[:i])
+			if err != nil {
+				return 0, err
+			}
+			i -= size
+			i = encodeVarintNebula(dAtA, i, uint64(size))
+		}
+		i--
+		dAtA[i] = 0xa
+	}
+	return len(dAtA) - i, nil
+}
+
+func (m *NebulaHandshakeDetails) Marshal() (dAtA []byte, err error) {
+	size := m.Size()
+	dAtA = make([]byte, size)
+	n, err := m.MarshalToSizedBuffer(dAtA[:size])
+	if err != nil {
+		return nil, err
+	}
+	return dAtA[:n], nil
+}
+
+func (m *NebulaHandshakeDetails) MarshalTo(dAtA []byte) (int, error) {
+	size := m.Size()
+	return m.MarshalToSizedBuffer(dAtA[:size])
+}
+
+func (m *NebulaHandshakeDetails) MarshalToSizedBuffer(dAtA []byte) (int, error) {
+	i := len(dAtA)
+	_ = i
+	var l int
+	_ = l
+	if m.CertVersion != 0 {
+		i = encodeVarintNebula(dAtA, i, uint64(m.CertVersion))
+		i--
+		dAtA[i] = 0x40
+	}
+	if m.Time != 0 {
+		i = encodeVarintNebula(dAtA, i, uint64(m.Time))
+		i--
+		dAtA[i] = 0x28
+	}
+	if m.Cookie != 0 {
+		i = encodeVarintNebula(dAtA, i, uint64(m.Cookie))
+		i--
+		dAtA[i] = 0x20
+	}
+	if m.ResponderIndex != 0 {
+		i = encodeVarintNebula(dAtA, i, uint64(m.ResponderIndex))
+		i--
+		dAtA[i] = 0x18
+	}
+	if m.InitiatorIndex != 0 {
+		i = encodeVarintNebula(dAtA, i, uint64(m.InitiatorIndex))
+		i--
+		dAtA[i] = 0x10
+	}
+	if len(m.Cert) > 0 {
+		i -= len(m.Cert)
+		copy(dAtA[i:], m.Cert)
+		i = encodeVarintNebula(dAtA, i, uint64(len(m.Cert)))
+		i--
+		dAtA[i] = 0xa
+	}
+	return len(dAtA) - i, nil
+}
+
 func (m *NebulaControl) Marshal() (dAtA []byte, err error) {
 	size := m.Size()
 	dAtA = make([]byte, size)
@@ -1132,6 +1375,51 @@ func (m *NebulaPing) Size() (n int) {
 	return n
 }
 
+func (m *NebulaHandshake) Size() (n int) {
+	if m == nil {
+		return 0
+	}
+	var l int
+	_ = l
+	if m.Details != nil {
+		l = m.Details.Size()
+		n += 1 + l + sovNebula(uint64(l))
+	}
+	l = len(m.Hmac)
+	if l > 0 {
+		n += 1 + l + sovNebula(uint64(l))
+	}
+	return n
+}
+
+func (m *NebulaHandshakeDetails) Size() (n int) {
+	if m == nil {
+		return 0
+	}
+	var l int
+	_ = l
+	l = len(m.Cert)
+	if l > 0 {
+		n += 1 + l + sovNebula(uint64(l))
+	}
+	if m.InitiatorIndex != 0 {
+		n += 1 + sovNebula(uint64(m.InitiatorIndex))
+	}
+	if m.ResponderIndex != 0 {
+		n += 1 + sovNebula(uint64(m.ResponderIndex))
+	}
+	if m.Cookie != 0 {
+		n += 1 + sovNebula(uint64(m.Cookie))
+	}
+	if m.Time != 0 {
+		n += 1 + sovNebula(uint64(m.Time))
+	}
+	if m.CertVersion != 0 {
+		n += 1 + sovNebula(uint64(m.CertVersion))
+	}
+	return n
+}
+
 func (m *NebulaControl) Size() (n int) {
 	if m == nil {
 		return 0
@@ -1948,6 +2236,305 @@ func (m *NebulaPing) Unmarshal(dAtA []byte) error {
 	}
 	return nil
 }
+func (m *NebulaHandshake) Unmarshal(dAtA []byte) error {
+	l := len(dAtA)
+	iNdEx := 0
+	for iNdEx < l {
+		preIndex := iNdEx
+		var wire uint64
+		for shift := uint(0); ; shift += 7 {
+			if shift >= 64 {
+				return ErrIntOverflowNebula
+			}
+			if iNdEx >= l {
+				return io.ErrUnexpectedEOF
+			}
+			b := dAtA[iNdEx]
+			iNdEx++
+			wire |= uint64(b&0x7F) << shift
+			if b < 0x80 {
+				break
+			}
+		}
+		fieldNum := int32(wire >> 3)
+		wireType := int(wire & 0x7)
+		if wireType == 4 {
+			return fmt.Errorf("proto: NebulaHandshake: wiretype end group for non-group")
+		}
+		if fieldNum <= 0 {
+			return fmt.Errorf("proto: NebulaHandshake: illegal tag %d (wire type %d)", fieldNum, wire)
+		}
+		switch fieldNum {
+		case 1:
+			if wireType != 2 {
+				return fmt.Errorf("proto: wrong wireType = %d for field Details", wireType)
+			}
+			var msglen int
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowNebula
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				msglen |= int(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+			if msglen < 0 {
+				return ErrInvalidLengthNebula
+			}
+			postIndex := iNdEx + msglen
+			if postIndex < 0 {
+				return ErrInvalidLengthNebula
+			}
+			if postIndex > l {
+				return io.ErrUnexpectedEOF
+			}
+			if m.Details == nil {
+				m.Details = &NebulaHandshakeDetails{}
+			}
+			if err := m.Details.Unmarshal(dAtA[iNdEx:postIndex]); err != nil {
+				return err
+			}
+			iNdEx = postIndex
+		case 2:
+			if wireType != 2 {
+				return fmt.Errorf("proto: wrong wireType = %d for field Hmac", wireType)
+			}
+			var byteLen int
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowNebula
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				byteLen |= int(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+			if byteLen < 0 {
+				return ErrInvalidLengthNebula
+			}
+			postIndex := iNdEx + byteLen
+			if postIndex < 0 {
+				return ErrInvalidLengthNebula
+			}
+			if postIndex > l {
+				return io.ErrUnexpectedEOF
+			}
+			m.Hmac = append(m.Hmac[:0], dAtA[iNdEx:postIndex]...)
+			if m.Hmac == nil {
+				m.Hmac = []byte{}
+			}
+			iNdEx = postIndex
+		default:
+			iNdEx = preIndex
+			skippy, err := skipNebula(dAtA[iNdEx:])
+			if err != nil {
+				return err
+			}
+			if (skippy < 0) || (iNdEx+skippy) < 0 {
+				return ErrInvalidLengthNebula
+			}
+			if (iNdEx + skippy) > l {
+				return io.ErrUnexpectedEOF
+			}
+			iNdEx += skippy
+		}
+	}
+
+	if iNdEx > l {
+		return io.ErrUnexpectedEOF
+	}
+	return nil
+}
+func (m *NebulaHandshakeDetails) Unmarshal(dAtA []byte) error {
+	l := len(dAtA)
+	iNdEx := 0
+	for iNdEx < l {
+		preIndex := iNdEx
+		var wire uint64
+		for shift := uint(0); ; shift += 7 {
+			if shift >= 64 {
+				return ErrIntOverflowNebula
+			}
+			if iNdEx >= l {
+				return io.ErrUnexpectedEOF
+			}
+			b := dAtA[iNdEx]
+			iNdEx++
+			wire |= uint64(b&0x7F) << shift
+			if b < 0x80 {
+				break
+			}
+		}
+		fieldNum := int32(wire >> 3)
+		wireType := int(wire & 0x7)
+		if wireType == 4 {
+			return fmt.Errorf("proto: NebulaHandshakeDetails: wiretype end group for non-group")
+		}
+		if fieldNum <= 0 {
+			return fmt.Errorf("proto: NebulaHandshakeDetails: illegal tag %d (wire type %d)", fieldNum, wire)
+		}
+		switch fieldNum {
+		case 1:
+			if wireType != 2 {
+				return fmt.Errorf("proto: wrong wireType = %d for field Cert", wireType)
+			}
+			var byteLen int
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowNebula
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				byteLen |= int(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+			if byteLen < 0 {
+				return ErrInvalidLengthNebula
+			}
+			postIndex := iNdEx + byteLen
+			if postIndex < 0 {
+				return ErrInvalidLengthNebula
+			}
+			if postIndex > l {
+				return io.ErrUnexpectedEOF
+			}
+			m.Cert = append(m.Cert[:0], dAtA[iNdEx:postIndex]...)
+			if m.Cert == nil {
+				m.Cert = []byte{}
+			}
+			iNdEx = postIndex
+		case 2:
+			if wireType != 0 {
+				return fmt.Errorf("proto: wrong wireType = %d for field InitiatorIndex", wireType)
+			}
+			m.InitiatorIndex = 0
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowNebula
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				m.InitiatorIndex |= uint32(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+		case 3:
+			if wireType != 0 {
+				return fmt.Errorf("proto: wrong wireType = %d for field ResponderIndex", wireType)
+			}
+			m.ResponderIndex = 0
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowNebula
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				m.ResponderIndex |= uint32(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+		case 4:
+			if wireType != 0 {
+				return fmt.Errorf("proto: wrong wireType = %d for field Cookie", wireType)
+			}
+			m.Cookie = 0
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowNebula
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				m.Cookie |= uint64(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+		case 5:
+			if wireType != 0 {
+				return fmt.Errorf("proto: wrong wireType = %d for field Time", wireType)
+			}
+			m.Time = 0
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowNebula
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				m.Time |= uint64(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+		case 8:
+			if wireType != 0 {
+				return fmt.Errorf("proto: wrong wireType = %d for field CertVersion", wireType)
+			}
+			m.CertVersion = 0
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowNebula
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				m.CertVersion |= uint32(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+		default:
+			iNdEx = preIndex
+			skippy, err := skipNebula(dAtA[iNdEx:])
+			if err != nil {
+				return err
+			}
+			if (skippy < 0) || (iNdEx+skippy) < 0 {
+				return ErrInvalidLengthNebula
+			}
+			if (iNdEx + skippy) > l {
+				return io.ErrUnexpectedEOF
+			}
+			iNdEx += skippy
+		}
+	}
+
+	if iNdEx > l {
+		return io.ErrUnexpectedEOF
+	}
+	return nil
+}
 func (m *NebulaControl) Unmarshal(dAtA []byte) error {
 	l := len(dAtA)
 	iNdEx := 0

nebula.proto

@@ -60,9 +60,21 @@ message NebulaPing {
 	uint64 Time = 2;
 }
 
-// NebulaHandshake / NebulaHandshakeDetails moved to
+message NebulaHandshake {
-// handshake/handshake.proto. The handshake package speaks that wire format
+  NebulaHandshakeDetails Details = 1;
-// directly via a hand-written encoder/decoder.
+  bytes Hmac = 2;
+}
+
+message NebulaHandshakeDetails {
+  bytes Cert = 1;
+  uint32 InitiatorIndex = 2;
+  uint32 ResponderIndex = 3;
+  uint64 Cookie = 4;
+  uint64 Time = 5;
+  uint32 CertVersion = 8;
+  // reserved for WIP multiport
+  reserved 6, 7;
+}
 
 message NebulaControl {
   enum MessageType {

outside.go

@@ -183,20 +183,11 @@ func (f *Interface) readOutsidePackets(via ViaSender, out []byte, packet []byte,
 			return
 		}
 
-		switch h.Subtype {
+		if h.Subtype == header.TestRequest {
-		case header.TestRequest:
 			// This testRequest might be from TryPromoteBest, so we should roam
 			// to the new IP address before responding
 			f.handleHostRoaming(hostinfo, via)
 			f.send(header.Test, header.TestReply, ci, hostinfo, d, nb, out)
-		case header.MTUDProbeRequest:
-			// Reply with just the 8-byte ack header so the reverse path doesn't have to
-			// carry the full probe size; we only verify the forward direction.
-			if len(d) >= 8 {
-				f.send(header.Test, header.MTUDProbeReply, ci, hostinfo, d[:8], nb, out)
-			}
-		case header.MTUDProbeReply:
-			f.pmtudManager.HandleReply(hostinfo.localIndexId, d)
 		}
 
 		// Fallthrough to the bottom to record incoming traffic
@@ -266,7 +257,6 @@ func (f *Interface) readOutsidePackets(via ViaSender, out []byte, packet []byte,
 
 // closeTunnel closes a tunnel locally, it does not send a closeTunnel packet to the remote
 func (f *Interface) closeTunnel(hostInfo *HostInfo) {
-	f.pmtudManager.OnTunnelDown(hostInfo)
 	final := f.hostMap.DeleteHostInfo(hostInfo)
 	if final {
 		// We no longer have any tunnels with this vpn addr, clear learned lighthouse state to lower memory usage
@@ -306,7 +296,6 @@ func (f *Interface) handleHostRoaming(hostinfo *HostInfo, via ViaSender) {
 		hostinfo.lastRoam = time.Now()
 		hostinfo.lastRoamRemote = hostinfo.remote
 		hostinfo.SetRemote(via.UdpAddr)
-		f.pmtudManager.OnRoam(hostinfo)
 	}
 
 }
@@ -583,7 +572,7 @@ func (f *Interface) decryptToTun(hostinfo *HostInfo, messageCounter uint64, out
 	}
 
 	f.connectionManager.In(hostinfo)
-	_, err = f.readers[q].Write(out)
+	err = f.batchers[q].Commit(out)
 	if err != nil {
 		f.l.Error("Failed to write to tun", "error", err)
 	}

overlay/batch/batch.go

@@ -0,0 +1,33 @@
+package batch
+
+import "net/netip"
+
+type RxBatcher interface {
+	// Reserve creates a pkt to borrow
+	Reserve(sz int) []byte
+	// Commit borrows pkt. The caller must keep pkt valid until the next Flush
+	Commit(pkt []byte) error
+	// Flush emits every queued packet in arrival order. Returns the
+	// first error observed; keeps draining so one bad packet doesn't hold up
+	// the rest. After Flush returns, borrowed payload slices may be recycled.
+	Flush() error
+}
+
+type TxBatcher interface {
+	// Next returns a zero-length slice with slotCap capacity over the next unused
+	// slot's backing bytes. The caller writes into the returned slice and then
+	// calls Commit with the final length and destination. Next returns nil when
+	// the batch is full.
+	Next() []byte
+	// Commit records the slot just returned by Next as a packet of length n
+	// destined for dst.
+	Commit(n int, dst netip.AddrPort)
+	// Reset clears committed slots; backing storage is retained for reuse.
+	Reset()
+	// Len returns the number of committed packets.
+	Len() int
+	// Cap returns the maximum number of slots in the batch.
+	Cap() int
+	// Get returns the buffers needed to send the batch
+	Get() ([][]byte, []netip.AddrPort)
+}

overlay/batch/passthrough.go

@@ -0,0 +1,57 @@
+package batch
+
+import (
+	"io"
+
+	"github.com/slackhq/nebula/udp"
+)
+
+// Passthrough is a RxBatcher that doesn't batch anything, it just accumulates and then sends packets.
+type Passthrough struct {
+	out     io.Writer
+	slots   [][]byte
+	backing []byte
+	cursor  int
+}
+
+func NewPassthrough(w io.Writer) *Passthrough {
+	const baseNumSlots = 128
+	return &Passthrough{
+		out:     w,
+		slots:   make([][]byte, 0, baseNumSlots),
+		backing: make([]byte, 0, baseNumSlots*udp.MTU),
+	}
+}
+
+func (p *Passthrough) Reserve(sz int) []byte {
+	if len(p.backing)+sz > cap(p.backing) {
+		// Grow: allocate a fresh backing. Already-committed slices still
+		// reference the old array and remain valid until Flush drops them.
+		newCap := max(cap(p.backing)*2, sz)
+		p.backing = make([]byte, 0, newCap)
+	}
+	start := len(p.backing)
+	p.backing = p.backing[:start+sz]
+	return p.backing[start : start+sz : start+sz] //return zero length, sz-cap slice
+}
+
+func (p *Passthrough) Commit(pkt []byte) error {
+	p.slots = append(p.slots, pkt)
+	return nil
+}
+
+func (p *Passthrough) Flush() error {
+	var firstErr error
+	for _, s := range p.slots {
+		_, err := p.out.Write(s)
+		if err != nil && firstErr == nil {
+			firstErr = err
+		}
+	}
+	for i := range p.slots {
+		p.slots[i] = nil
+	}
+	p.slots = p.slots[:0]
+	p.backing = p.backing[:0]
+	return firstErr
+}

overlay/batch/tcp_coalesce.go

@@ -0,0 +1,506 @@
+package batch
+
+import (
+	"bytes"
+	"encoding/binary"
+	"io"
+
+	"github.com/slackhq/nebula/overlay/tio"
+)
+
+// ipProtoTCP is the IANA protocol number for TCP. Hardcoded instead of
+// reaching for golang.org/x/sys/unix — that package doesn't define the
+// constant on Windows, which would break cross-compiles even though this
+// file runs unchanged on every platform.
+const ipProtoTCP = 6
+
+// tcpCoalesceBufSize caps total bytes per superpacket. Mirrors the kernel's
+// sk_gso_max_size of ~64KiB; anything beyond this would be rejected anyway.
+const tcpCoalesceBufSize = 65535
+
+// tcpCoalesceMaxSegs caps how many segments we'll coalesce into a single
+// superpacket. Keeping this well below the kernel's TSO ceiling bounds
+// latency.
+const tcpCoalesceMaxSegs = 64
+
+// tcpCoalesceHdrCap is the scratch space we copy a seed's IP+TCP header
+// into. IPv6 (40) + TCP with full options (60) = 100 bytes.
+const tcpCoalesceHdrCap = 100
+
+// initialSlots is the starting capacity of the slot pool. One flow per
+// packet is the worst case so this matches a typical UDP recvmmsg batch.
+const initialSlots = 64
+
+// flowKey identifies a TCP flow by {src, dst, sport, dport, family}.
+// Comparable, so linear scans over the slot list stay tight.
+type flowKey struct {
+	src, dst     [16]byte
+	sport, dport uint16
+	isV6         bool
+}
+
+// coalesceSlot is one entry in the coalescer's ordered event queue. When
+// passthrough is true the slot holds a single borrowed packet that must be
+// emitted verbatim (non-TCP, non-admissible TCP, or oversize seed). When
+// passthrough is false the slot is an in-progress coalesced superpacket:
+// hdrBuf is a mutable copy of the seed's IP+TCP header (we patch total
+// length and pseudo-header partial at flush), and payIovs are *borrowed*
+// slices from the caller's plaintext buffers — no payload is ever copied.
+// The caller (listenOut) must keep those buffers alive until Flush.
+type coalesceSlot struct {
+	passthrough bool
+	rawPkt      []byte // borrowed when passthrough
+
+	fk       flowKey
+	hdrBuf   [tcpCoalesceHdrCap]byte
+	hdrLen   int
+	ipHdrLen int
+	isV6     bool
+	gsoSize  int
+	numSeg   int
+	totalPay int
+	nextSeq  uint32
+	// psh closes the chain: set when the last-accepted segment had PSH or
+	// was sub-gsoSize. No further appends after that.
+	psh     bool
+	payIovs [][]byte
+}
+
+// TCPCoalescer accumulates adjacent in-flow TCP data segments across
+// multiple concurrent flows and emits each flow's run as a single TSO
+// superpacket via tio.GSOWriter. All output — coalesced or not — is
+// deferred until Flush so arrival order is preserved on the wire. Owns
+// no locks; one coalescer per TUN write queue.
+type TCPCoalescer struct {
+	plainW io.Writer
+	gsoW   tio.GSOWriter // nil when the queue doesn't support TSO
+
+	// slots is the ordered event queue. Flush walks it once and emits each
+	// entry as either a WriteGSO (coalesced) or a plainW.Write (passthrough).
+	slots []*coalesceSlot
+	// openSlots maps a flow key to its most recent non-sealed slot, so new
+	// segments can extend an in-progress superpacket in O(1). Slots are
+	// removed from this map when they close (PSH or short-last-segment),
+	// when a non-admissible packet for that flow arrives, or in Flush.
+	openSlots map[flowKey]*coalesceSlot
+	pool      []*coalesceSlot // free list for reuse
+
+	backing []byte
+}
+
+func NewTCPCoalescer(w io.Writer) *TCPCoalescer {
+	c := &TCPCoalescer{
+		plainW:    w,
+		slots:     make([]*coalesceSlot, 0, initialSlots),
+		openSlots: make(map[flowKey]*coalesceSlot, initialSlots),
+		pool:      make([]*coalesceSlot, 0, initialSlots),
+		backing:   make([]byte, 0, initialSlots*65535),
+	}
+	if gw, ok := w.(tio.GSOWriter); ok && gw.GSOSupported() {
+		c.gsoW = gw
+	}
+	return c
+}
+
+// parsedTCP holds the fields extracted from a single parse so later steps
+// (admission, slot lookup, canAppend) don't re-walk the header.
+type parsedTCP struct {
+	fk        flowKey
+	ipHdrLen  int
+	tcpHdrLen int
+	hdrLen    int
+	payLen    int
+	seq       uint32
+	flags     byte
+}
+
+// parseTCPBase extracts the flow key and IP/TCP offsets for any TCP packet,
+// regardless of whether it's admissible for coalescing. Returns ok=false
+// for non-TCP or malformed input. Accepts IPv4 (no options, no fragmentation)
+// and IPv6 (no extension headers).
+func parseTCPBase(pkt []byte) (parsedTCP, bool) {
+	var p parsedTCP
+	if len(pkt) < 20 {
+		return p, false
+	}
+	v := pkt[0] >> 4
+	switch v {
+	case 4:
+		ihl := int(pkt[0]&0x0f) * 4
+		if ihl != 20 {
+			return p, false
+		}
+		if pkt[9] != ipProtoTCP {
+			return p, false
+		}
+		// Reject actual fragmentation (MF or non-zero frag offset).
+		if binary.BigEndian.Uint16(pkt[6:8])&0x3fff != 0 {
+			return p, false
+		}
+		totalLen := int(binary.BigEndian.Uint16(pkt[2:4]))
+		if totalLen > len(pkt) || totalLen < ihl {
+			return p, false
+		}
+		p.ipHdrLen = 20
+		p.fk.isV6 = false
+		copy(p.fk.src[:4], pkt[12:16])
+		copy(p.fk.dst[:4], pkt[16:20])
+		pkt = pkt[:totalLen]
+	case 6:
+		if len(pkt) < 40 {
+			return p, false
+		}
+		if pkt[6] != ipProtoTCP {
+			return p, false
+		}
+		payloadLen := int(binary.BigEndian.Uint16(pkt[4:6]))
+		if 40+payloadLen > len(pkt) {
+			return p, false
+		}
+		p.ipHdrLen = 40
+		p.fk.isV6 = true
+		copy(p.fk.src[:], pkt[8:24])
+		copy(p.fk.dst[:], pkt[24:40])
+		pkt = pkt[:40+payloadLen]
+	default:
+		return p, false
+	}
+
+	if len(pkt) < p.ipHdrLen+20 {
+		return p, false
+	}
+	tcpOff := int(pkt[p.ipHdrLen+12]>>4) * 4
+	if tcpOff < 20 || tcpOff > 60 {
+		return p, false
+	}
+	if len(pkt) < p.ipHdrLen+tcpOff {
+		return p, false
+	}
+	p.tcpHdrLen = tcpOff
+	p.hdrLen = p.ipHdrLen + tcpOff
+	p.payLen = len(pkt) - p.hdrLen
+	p.seq = binary.BigEndian.Uint32(pkt[p.ipHdrLen+4 : p.ipHdrLen+8])
+	p.flags = pkt[p.ipHdrLen+13]
+	p.fk.sport = binary.BigEndian.Uint16(pkt[p.ipHdrLen : p.ipHdrLen+2])
+	p.fk.dport = binary.BigEndian.Uint16(pkt[p.ipHdrLen+2 : p.ipHdrLen+4])
+	return p, true
+}
+
+// coalesceable reports whether a parsed TCP segment is eligible for
+// coalescing. Accepts only ACK or ACK|PSH with a non-empty payload.
+func (p parsedTCP) coalesceable() bool {
+	const ack = 0x10
+	const psh = 0x08
+	if p.flags&^(ack|psh) != 0 || p.flags&ack == 0 {
+		return false
+	}
+	return p.payLen > 0
+}
+
+func (c *TCPCoalescer) Reserve(sz int) []byte {
+	if len(c.backing)+sz > cap(c.backing) {
+		// Grow: allocate a fresh backing. Already-committed slices still
+		// reference the old array and remain valid until Flush drops them.
+		newCap := max(cap(c.backing)*2, sz)
+		c.backing = make([]byte, 0, newCap)
+	}
+	start := len(c.backing)
+	c.backing = c.backing[:start+sz]
+	return c.backing[start : start+sz : start+sz] //return zero length, sz-cap slice
+}
+
+// Commit borrows pkt. The caller must keep pkt valid until the next Flush,
+// whether or not the packet was coalesced — passthrough (non-admissible)
+// packets are queued and written at Flush time, not synchronously.
+func (c *TCPCoalescer) Commit(pkt []byte) error {
+	if c.gsoW == nil {
+		c.addPassthrough(pkt)
+		return nil
+	}
+
+	info, ok := parseTCPBase(pkt)
+	if !ok {
+		// Non-TCP or malformed — can't possibly collide with an open flow.
+		c.addPassthrough(pkt)
+		return nil
+	}
+	if !info.coalesceable() {
+		// TCP but not admissible (SYN/FIN/RST/URG/CWR/ECE or zero-payload).
+		// Seal this flow's open slot so later in-flow packets don't extend
+		// it and accidentally reorder past this passthrough.
+		delete(c.openSlots, info.fk)
+		c.addPassthrough(pkt)
+		return nil
+	}
+
+	if open := c.openSlots[info.fk]; open != nil {
+		if c.canAppend(open, pkt, info) {
+			c.appendPayload(open, pkt, info)
+			if open.psh {
+				delete(c.openSlots, info.fk)
+			}
+			return nil
+		}
+		// Can't extend — seal it and fall through to seed a fresh slot.
+		delete(c.openSlots, info.fk)
+	}
+	c.seed(pkt, info)
+	return nil
+}
+
+// Flush emits every queued event in arrival order. Coalesced slots go out
+// via WriteGSO; passthrough slots go out via plainW.Write. Returns the
+// first error observed; keeps draining so one bad packet doesn't hold up
+// the rest. After Flush returns, borrowed payload slices may be recycled.
+func (c *TCPCoalescer) Flush() error {
+	var first error
+	for _, s := range c.slots {
+		var err error
+		if s.passthrough {
+			_, err = c.plainW.Write(s.rawPkt)
+		} else {
+			err = c.flushSlot(s)
+		}
+		if err != nil && first == nil {
+			first = err
+		}
+		c.release(s)
+	}
+	for i := range c.slots {
+		c.slots[i] = nil
+	}
+	c.slots = c.slots[:0]
+	for k := range c.openSlots {
+		delete(c.openSlots, k)
+	}
+
+	c.backing = c.backing[:0]
+	return first
+}
+
+func (c *TCPCoalescer) addPassthrough(pkt []byte) {
+	s := c.take()
+	s.passthrough = true
+	s.rawPkt = pkt
+	c.slots = append(c.slots, s)
+}
+
+func (c *TCPCoalescer) seed(pkt []byte, info parsedTCP) {
+	if info.hdrLen > tcpCoalesceHdrCap || info.hdrLen+info.payLen > tcpCoalesceBufSize {
+		// Pathological shape — can't fit our scratch, emit as-is.
+		c.addPassthrough(pkt)
+		return
+	}
+	s := c.take()
+	s.passthrough = false
+	s.rawPkt = nil
+	copy(s.hdrBuf[:], pkt[:info.hdrLen])
+	s.hdrLen = info.hdrLen
+	s.ipHdrLen = info.ipHdrLen
+	s.isV6 = info.fk.isV6
+	s.fk = info.fk
+	s.gsoSize = info.payLen
+	s.numSeg = 1
+	s.totalPay = info.payLen
+	s.nextSeq = info.seq + uint32(info.payLen)
+	s.psh = info.flags&0x08 != 0
+	s.payIovs = append(s.payIovs[:0], pkt[info.hdrLen:info.hdrLen+info.payLen])
+	c.slots = append(c.slots, s)
+	if !s.psh {
+		c.openSlots[info.fk] = s
+	}
+}
+
+// canAppend reports whether info's packet extends the slot's seed: same
+// header shape and stable contents, adjacent seq, not oversized, chain not
+// closed.
+func (c *TCPCoalescer) canAppend(s *coalesceSlot, pkt []byte, info parsedTCP) bool {
+	if s.psh {
+		return false
+	}
+	if info.hdrLen != s.hdrLen {
+		return false
+	}
+	if info.seq != s.nextSeq {
+		return false
+	}
+	if s.numSeg >= tcpCoalesceMaxSegs {
+		return false
+	}
+	if info.payLen > s.gsoSize {
+		return false
+	}
+	if s.hdrLen+s.totalPay+info.payLen > tcpCoalesceBufSize {
+		return false
+	}
+	if !headersMatch(s.hdrBuf[:s.hdrLen], pkt[:info.hdrLen], s.isV6, s.ipHdrLen) {
+		return false
+	}
+	return true
+}
+
+func (c *TCPCoalescer) appendPayload(s *coalesceSlot, pkt []byte, info parsedTCP) {
+	s.payIovs = append(s.payIovs, pkt[info.hdrLen:info.hdrLen+info.payLen])
+	s.numSeg++
+	s.totalPay += info.payLen
+	s.nextSeq = info.seq + uint32(info.payLen)
+	if info.flags&0x08 != 0 {
+		// Propagate PSH into the seed header so kernel TSO sets it on the
+		// last segment. Without this the sender's push signal is dropped.
+		s.hdrBuf[s.ipHdrLen+13] |= 0x08
+	}
+	if info.payLen < s.gsoSize || info.flags&0x08 != 0 {
+		s.psh = true
+	}
+}
+
+func (c *TCPCoalescer) take() *coalesceSlot {
+	if n := len(c.pool); n > 0 {
+		s := c.pool[n-1]
+		c.pool[n-1] = nil
+		c.pool = c.pool[:n-1]
+		return s
+	}
+	return &coalesceSlot{}
+}
+
+func (c *TCPCoalescer) release(s *coalesceSlot) {
+	s.passthrough = false
+	s.rawPkt = nil
+	for i := range s.payIovs {
+		s.payIovs[i] = nil
+	}
+	s.payIovs = s.payIovs[:0]
+	s.numSeg = 0
+	s.totalPay = 0
+	s.psh = false
+	c.pool = append(c.pool, s)
+}
+
+// flushSlot patches the header and calls WriteGSO. Does not remove the
+// slot from c.slots.
+func (c *TCPCoalescer) flushSlot(s *coalesceSlot) error {
+	total := s.hdrLen + s.totalPay
+	l4Len := total - s.ipHdrLen
+	hdr := s.hdrBuf[:s.hdrLen]
+
+	if s.isV6 {
+		binary.BigEndian.PutUint16(hdr[4:6], uint16(l4Len))
+	} else {
+		binary.BigEndian.PutUint16(hdr[2:4], uint16(total))
+		hdr[10] = 0
+		hdr[11] = 0
+		binary.BigEndian.PutUint16(hdr[10:12], ipv4HdrChecksum(hdr[:s.ipHdrLen]))
+	}
+
+	var psum uint32
+	if s.isV6 {
+		psum = pseudoSumIPv6(hdr[8:24], hdr[24:40], ipProtoTCP, l4Len)
+	} else {
+		psum = pseudoSumIPv4(hdr[12:16], hdr[16:20], ipProtoTCP, l4Len)
+	}
+	tcsum := s.ipHdrLen + 16
+	binary.BigEndian.PutUint16(hdr[tcsum:tcsum+2], foldOnceNoInvert(psum))
+
+	return c.gsoW.WriteGSO(hdr[:s.ipHdrLen], hdr[s.ipHdrLen:], s.payIovs)
+}
+
+// headersMatch compares two IP+TCP header prefixes for byte-for-byte
+// equality on every field that must be identical across coalesced
+// segments. Size/IPID/IPCsum/seq/flags/tcpCsum are masked out.
+func headersMatch(a, b []byte, isV6 bool, ipHdrLen int) bool {
+	if len(a) != len(b) {
+		return false
+	}
+	if isV6 {
+		// IPv6: bytes [0:4] = version/TC/flow-label, [6:8] = next_hdr/hop,
+		// [8:40] = src+dst. Skip [4:6] payload length.
+		if !bytes.Equal(a[0:4], b[0:4]) {
+			return false
+		}
+		if !bytes.Equal(a[6:40], b[6:40]) {
+			return false
+		}
+	} else {
+		// IPv4: [0:2] version/IHL/TOS, [6:10] flags/fragoff/TTL/proto,
+		// [12:20] src+dst. Skip [2:4] total len, [4:6] id, [10:12] csum.
+		if !bytes.Equal(a[0:2], b[0:2]) {
+			return false
+		}
+		if !bytes.Equal(a[6:10], b[6:10]) {
+			return false
+		}
+		if !bytes.Equal(a[12:20], b[12:20]) {
+			return false
+		}
+	}
+	// TCP: compare [0:4] ports, [8:13] ack+dataoff, [14:16] window,
+	// [18:tcpHdrLen] options (incl. urgent).
+	tcp := ipHdrLen
+	if !bytes.Equal(a[tcp:tcp+4], b[tcp:tcp+4]) {
+		return false
+	}
+	if !bytes.Equal(a[tcp+8:tcp+13], b[tcp+8:tcp+13]) {
+		return false
+	}
+	if !bytes.Equal(a[tcp+14:tcp+16], b[tcp+14:tcp+16]) {
+		return false
+	}
+	if !bytes.Equal(a[tcp+18:], b[tcp+18:]) {
+		return false
+	}
+	return true
+}
+
+// ipv4HdrChecksum computes the IPv4 header checksum over hdr (which must
+// already have its checksum field zeroed) and returns the folded/inverted
+// 16-bit value to store.
+func ipv4HdrChecksum(hdr []byte) uint16 {
+	var sum uint32
+	for i := 0; i+1 < len(hdr); i += 2 {
+		sum += uint32(binary.BigEndian.Uint16(hdr[i : i+2]))
+	}
+	if len(hdr)%2 == 1 {
+		sum += uint32(hdr[len(hdr)-1]) << 8
+	}
+	for sum>>16 != 0 {
+		sum = (sum & 0xffff) + (sum >> 16)
+	}
+	return ^uint16(sum)
+}
+
+// pseudoSumIPv4 / pseudoSumIPv6 build the TCP pseudo-header partial sum
+// expected by the virtio NEEDS_CSUM kernel path: the 32-bit accumulator
+// before folding.
+func pseudoSumIPv4(src, dst []byte, proto byte, l4Len int) uint32 {
+	var sum uint32
+	sum += uint32(binary.BigEndian.Uint16(src[0:2]))
+	sum += uint32(binary.BigEndian.Uint16(src[2:4]))
+	sum += uint32(binary.BigEndian.Uint16(dst[0:2]))
+	sum += uint32(binary.BigEndian.Uint16(dst[2:4]))
+	sum += uint32(proto)
+	sum += uint32(l4Len)
+	return sum
+}
+
+func pseudoSumIPv6(src, dst []byte, proto byte, l4Len int) uint32 {
+	var sum uint32
+	for i := 0; i < 16; i += 2 {
+		sum += uint32(binary.BigEndian.Uint16(src[i : i+2]))
+		sum += uint32(binary.BigEndian.Uint16(dst[i : i+2]))
+	}
+	sum += uint32(l4Len >> 16)
+	sum += uint32(l4Len & 0xffff)
+	sum += uint32(proto)
+	return sum
+}
+
+// foldOnceNoInvert folds the 32-bit accumulator to 16 bits and returns it
+// unchanged (no one's complement). This is what virtio NEEDS_CSUM wants in
+// the L4 checksum field — the kernel will add the payload sum and invert.
+func foldOnceNoInvert(sum uint32) uint16 {
+	for sum>>16 != 0 {
+		sum = (sum & 0xffff) + (sum >> 16)
+	}
+	return uint16(sum)
+}

overlay/batch/tcp_coalesce_test.go

@@ -0,0 +1,618 @@
+package batch
+
+import (
+	"encoding/binary"
+	"testing"
+)
+
+// fakeTunWriter records plain Writes and WriteGSO calls without touching a
+// real TUN fd. WriteGSO records the IP header, transport header, and
+// borrowed payload fragments separately so tests can inspect each.
+type fakeTunWriter struct {
+	gsoEnabled bool
+	writes     [][]byte
+	gsoWrites  []fakeGSOWrite
+}
+
+// fakeGSOWrite captures one WriteGSO call. hdr is the concatenation of the
+// IP and transport headers (in that order), gsoSize / isV6 / csumStart are
+// derived from the call so existing assertions keep working unchanged.
+type fakeGSOWrite struct {
+	hdr       []byte
+	pays      [][]byte
+	gsoSize   uint16
+	isV6      bool
+	csumStart uint16
+}
+
+// total returns hdrLen + sum of pay lens.
+func (g fakeGSOWrite) total() int {
+	n := len(g.hdr)
+	for _, p := range g.pays {
+		n += len(p)
+	}
+	return n
+}
+
+// payLen sums the pays.
+func (g fakeGSOWrite) payLen() int {
+	var n int
+	for _, p := range g.pays {
+		n += len(p)
+	}
+	return n
+}
+
+func (w *fakeTunWriter) Write(p []byte) (int, error) {
+	buf := make([]byte, len(p))
+	copy(buf, p)
+	w.writes = append(w.writes, buf)
+	return len(p), nil
+}
+
+func (w *fakeTunWriter) WriteGSO(hdr []byte, transportHdr []byte, pays [][]byte) error {
+	hcopy := make([]byte, len(hdr)+len(transportHdr))
+	copy(hcopy, hdr)
+	copy(hcopy[len(hdr):], transportHdr)
+	paysCopy := make([][]byte, len(pays))
+	for i, p := range pays {
+		pc := make([]byte, len(p))
+		copy(pc, p)
+		paysCopy[i] = pc
+	}
+	var gsoSize uint16
+	if len(pays) > 1 {
+		gsoSize = uint16(len(pays[0]))
+	}
+	isV6 := len(hdr) > 0 && hdr[0]>>4 == 6
+	w.gsoWrites = append(w.gsoWrites, fakeGSOWrite{
+		hdr:       hcopy,
+		pays:      paysCopy,
+		gsoSize:   gsoSize,
+		isV6:      isV6,
+		csumStart: uint16(len(hdr)),
+	})
+	return nil
+}
+
+func (w *fakeTunWriter) GSOSupported() bool { return w.gsoEnabled }
+
+// buildTCPv4 constructs a minimal IPv4+TCP packet with the given payload,
+// seq, and flags. Assumes no IP options and a 20-byte TCP header.
+func buildTCPv4(seq uint32, flags byte, payload []byte) []byte {
+	return buildTCPv4Ports(1000, 2000, seq, flags, payload)
+}
+
+// buildTCPv4Ports is buildTCPv4 with caller-specified ports so tests can
+// build distinct flows.
+func buildTCPv4Ports(sport, dport uint16, seq uint32, flags byte, payload []byte) []byte {
+	const ipHdrLen = 20
+	const tcpHdrLen = 20
+	total := ipHdrLen + tcpHdrLen + len(payload)
+	pkt := make([]byte, total)
+
+	pkt[0] = 0x45
+	pkt[1] = 0x00
+	binary.BigEndian.PutUint16(pkt[2:4], uint16(total))
+	binary.BigEndian.PutUint16(pkt[4:6], 0)
+	binary.BigEndian.PutUint16(pkt[6:8], 0x4000)
+	pkt[8] = 64
+	pkt[9] = ipProtoTCP
+	copy(pkt[12:16], []byte{10, 0, 0, 1})
+	copy(pkt[16:20], []byte{10, 0, 0, 2})
+
+	binary.BigEndian.PutUint16(pkt[20:22], sport)
+	binary.BigEndian.PutUint16(pkt[22:24], dport)
+	binary.BigEndian.PutUint32(pkt[24:28], seq)
+	binary.BigEndian.PutUint32(pkt[28:32], 12345)
+	pkt[32] = 0x50
+	pkt[33] = flags
+	binary.BigEndian.PutUint16(pkt[34:36], 0xffff)
+
+	copy(pkt[40:], payload)
+	return pkt
+}
+
+const (
+	tcpAck    = 0x10
+	tcpPsh    = 0x08
+	tcpSyn    = 0x02
+	tcpFin    = 0x01
+	tcpAckPsh = tcpAck | tcpPsh
+)
+
+func TestCoalescerPassthroughWhenGSOUnavailable(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: false}
+	c := NewTCPCoalescer(w)
+	pkt := buildTCPv4(1000, tcpAck, []byte("hello"))
+	if err := c.Commit(pkt); err != nil {
+		t.Fatal(err)
+	}
+	// No sync write — passthrough is deferred to Flush.
+	if len(w.writes) != 0 || len(w.gsoWrites) != 0 {
+		t.Fatalf("no Add-time writes: got writes=%d gso=%d", len(w.writes), len(w.gsoWrites))
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	if len(w.writes) != 1 || len(w.gsoWrites) != 0 {
+		t.Fatalf("want single plain write, got writes=%d gso=%d", len(w.writes), len(w.gsoWrites))
+	}
+}
+
+func TestCoalescerNonTCPPassthrough(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := NewTCPCoalescer(w)
+	pkt := make([]byte, 28)
+	pkt[0] = 0x45
+	binary.BigEndian.PutUint16(pkt[2:4], 28)
+	pkt[9] = 1
+	copy(pkt[12:16], []byte{10, 0, 0, 1})
+	copy(pkt[16:20], []byte{10, 0, 0, 2})
+	if err := c.Commit(pkt); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	if len(w.writes) != 1 || len(w.gsoWrites) != 0 {
+		t.Fatalf("ICMP should pass through unchanged")
+	}
+}
+
+func TestCoalescerSeedThenFlushAlone(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := NewTCPCoalescer(w)
+	pkt := buildTCPv4(1000, tcpAck, make([]byte, 1000))
+	if err := c.Commit(pkt); err != nil {
+		t.Fatal(err)
+	}
+	if len(w.writes) != 0 || len(w.gsoWrites) != 0 {
+		t.Fatalf("unexpected output before flush")
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	// Single-segment flush now goes through WriteGSO with GSO_NONE
+	// (virtio NEEDS_CSUM lets the kernel fill in the L4 csum).
+	if len(w.gsoWrites) != 1 || len(w.writes) != 0 {
+		t.Fatalf("single-seg flush: writes=%d gso=%d", len(w.writes), len(w.gsoWrites))
+	}
+	g := w.gsoWrites[0]
+	if g.total() != 40+1000 {
+		t.Errorf("super total=%d want %d", g.total(), 40+1000)
+	}
+	if g.payLen() != 1000 {
+		t.Errorf("payLen=%d want 1000", g.payLen())
+	}
+}
+
+func TestCoalescerCoalescesAdjacentACKs(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := NewTCPCoalescer(w)
+	pay := make([]byte, 1200)
+	if err := c.Commit(buildTCPv4(1000, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Commit(buildTCPv4(2200, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Commit(buildTCPv4(3400, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	if len(w.gsoWrites) != 1 {
+		t.Fatalf("want 1 gso write, got %d (plain=%d)", len(w.gsoWrites), len(w.writes))
+	}
+	g := w.gsoWrites[0]
+	if g.gsoSize != 1200 {
+		t.Errorf("gsoSize=%d want 1200", g.gsoSize)
+	}
+	if len(g.hdr) != 40 {
+		t.Errorf("hdrLen=%d want 40", len(g.hdr))
+	}
+	if g.csumStart != 20 {
+		t.Errorf("csumStart=%d want 20", g.csumStart)
+	}
+	if len(g.pays) != 3 {
+		t.Errorf("pay count=%d want 3", len(g.pays))
+	}
+	if g.total() != 40+3*1200 {
+		t.Errorf("superpacket len=%d want %d", g.total(), 40+3*1200)
+	}
+	if tot := binary.BigEndian.Uint16(g.hdr[2:4]); int(tot) != g.total() {
+		t.Errorf("ip total_length=%d want %d", tot, g.total())
+	}
+}
+
+func TestCoalescerRejectsSeqGap(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := NewTCPCoalescer(w)
+	pay := make([]byte, 1200)
+	if err := c.Commit(buildTCPv4(1000, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Commit(buildTCPv4(3000, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	// Each packet flushes as its own single-segment WriteGSO now.
+	if len(w.gsoWrites) != 2 || len(w.writes) != 0 {
+		t.Fatalf("seq gap: want 2 gso writes got writes=%d gso=%d", len(w.writes), len(w.gsoWrites))
+	}
+}
+
+func TestCoalescerRejectsFlagMismatch(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := NewTCPCoalescer(w)
+	pay := make([]byte, 1200)
+	if err := c.Commit(buildTCPv4(1000, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	// SYN|ACK is non-admissible. Must flush matching flow's slot (gso)
+	// and then plain-write the SYN packet itself.
+	syn := buildTCPv4(2200, tcpSyn|tcpAck, pay)
+	if err := c.Commit(syn); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	if len(w.writes) != 1 || len(w.gsoWrites) != 1 {
+		t.Fatalf("flag mismatch: want 1 plain + 1 gso, got writes=%d gso=%d", len(w.writes), len(w.gsoWrites))
+	}
+}
+
+func TestCoalescerRejectsFIN(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := NewTCPCoalescer(w)
+	fin := buildTCPv4(1000, tcpAck|tcpFin, []byte("x"))
+	if err := c.Commit(fin); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	// FIN isn't admissible — passthrough as plain, no slot, no gso.
+	if len(w.writes) != 1 || len(w.gsoWrites) != 0 {
+		t.Fatalf("FIN should be passthrough, got writes=%d gso=%d", len(w.writes), len(w.gsoWrites))
+	}
+}
+
+func TestCoalescerShortLastSegmentClosesChain(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := NewTCPCoalescer(w)
+	full := make([]byte, 1200)
+	half := make([]byte, 500)
+	if err := c.Commit(buildTCPv4(1000, tcpAck, full)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Commit(buildTCPv4(2200, tcpAck, half)); err != nil {
+		t.Fatal(err)
+	}
+	// Chain now closed; next packet seeds a new slot on the same flow
+	// after flushing the old one.
+	if err := c.Commit(buildTCPv4(2700, tcpAck, full)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	// Expect two gso writes: the first two packets coalesced, then the
+	// third flushed alone (single-seg via GSO_NONE).
+	if len(w.gsoWrites) != 2 {
+		t.Fatalf("want 2 gso writes got %d", len(w.gsoWrites))
+	}
+	if len(w.writes) != 0 {
+		t.Fatalf("want 0 plain writes got %d", len(w.writes))
+	}
+	if w.gsoWrites[0].gsoSize != 1200 {
+		t.Errorf("gsoSize=%d want 1200", w.gsoWrites[0].gsoSize)
+	}
+	if got, want := w.gsoWrites[0].total(), 40+1200+500; got != want {
+		t.Errorf("super len=%d want %d", got, want)
+	}
+}
+
+func TestCoalescerPSHFinalizesChain(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := NewTCPCoalescer(w)
+	pay := make([]byte, 1200)
+	if err := c.Commit(buildTCPv4(1000, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Commit(buildTCPv4(2200, tcpAckPsh, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Commit(buildTCPv4(3400, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	// First two coalesce; the third seeds a fresh slot that flushes alone.
+	if len(w.gsoWrites) != 2 {
+		t.Fatalf("want 2 gso writes got %d", len(w.gsoWrites))
+	}
+	if len(w.writes) != 0 {
+		t.Fatalf("want 0 plain writes got %d", len(w.writes))
+	}
+}
+
+// TestCoalescerPropagatesPSHFromAppended ensures that when an appended
+// segment carries PSH (or is short, sealing the chain), the PSH bit ends
+// up in the emitted superpacket's TCP flags. The kernel TSO path keeps
+// PSH only on the last segment iff the input header has it set; if the
+// coalescer drops it the sender's push signal never reaches the receiver.
+func TestCoalescerPropagatesPSHFromAppended(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := NewTCPCoalescer(w)
+	pay := make([]byte, 1200)
+	// Seed has no PSH; second segment carries PSH and seals the chain.
+	if err := c.Commit(buildTCPv4(1000, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Commit(buildTCPv4(2200, tcpAckPsh, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Flush(0); err != nil {
+		t.Fatal(err)
+	}
+	if len(w.gsoWrites) != 1 {
+		t.Fatalf("want 1 gso write got %d", len(w.gsoWrites))
+	}
+	g := w.gsoWrites[0]
+	const ipHdrLen = 20
+	flags := g.hdr[ipHdrLen+13]
+	if flags&tcpPsh == 0 {
+		t.Fatalf("PSH lost from coalesced superpacket: flags=0x%02x", flags)
+	}
+	if flags&tcpAck == 0 {
+		t.Fatalf("ACK missing from coalesced superpacket: flags=0x%02x", flags)
+	}
+}
+
+func TestCoalescerRejectsDifferentFlow(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := NewTCPCoalescer(w)
+	pay := make([]byte, 1200)
+	p1 := buildTCPv4(1000, tcpAck, pay)
+	p2 := buildTCPv4(2200, tcpAck, pay)
+	binary.BigEndian.PutUint16(p2[20:22], 9999)
+	if err := c.Commit(p1); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Commit(p2); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	// Two independent flows, each flushes its own single-segment WriteGSO.
+	if len(w.gsoWrites) != 2 || len(w.writes) != 0 {
+		t.Fatalf("diff flow: want 2 gso writes got writes=%d gso=%d", len(w.writes), len(w.gsoWrites))
+	}
+}
+
+func TestCoalescerRejectsIPOptions(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := NewTCPCoalescer(w)
+	pay := make([]byte, 500)
+	pkt := buildTCPv4(1000, tcpAck, pay)
+	// Bump IHL to 6 to simulate 4 bytes of IP options. Don't actually add
+	// bytes — parser should bail before it matters.
+	pkt[0] = 0x46
+	if err := c.Commit(pkt); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	// Non-admissible parse → passthrough as plain.
+	if len(w.writes) != 1 || len(w.gsoWrites) != 0 {
+		t.Fatalf("IP options should passthrough, got writes=%d gso=%d", len(w.writes), len(w.gsoWrites))
+	}
+}
+
+func TestCoalescerCapBySegments(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := NewTCPCoalescer(w)
+	pay := make([]byte, 512)
+	seq := uint32(1000)
+	for i := 0; i < tcpCoalesceMaxSegs+5; i++ {
+		if err := c.Commit(buildTCPv4(seq, tcpAck, pay)); err != nil {
+			t.Fatal(err)
+		}
+		seq += uint32(len(pay))
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	for _, g := range w.gsoWrites {
+		segs := len(g.pays)
+		if segs > tcpCoalesceMaxSegs {
+			t.Fatalf("super exceeded seg cap: %d > %d", segs, tcpCoalesceMaxSegs)
+		}
+	}
+}
+
+// TestCoalescerMultipleFlowsInSameBatch proves two interleaved bulk TCP
+// flows coalesce independently in a single Flush.
+func TestCoalescerMultipleFlowsInSameBatch(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := NewTCPCoalescer(w)
+	pay := make([]byte, 1200)
+
+	// Flow A: sport 1000. Flow B: sport 3000.
+	if err := c.Commit(buildTCPv4Ports(1000, 2000, 100, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Commit(buildTCPv4Ports(3000, 2000, 500, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Commit(buildTCPv4Ports(1000, 2000, 1300, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Commit(buildTCPv4Ports(3000, 2000, 1700, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Commit(buildTCPv4Ports(1000, 2000, 2500, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Commit(buildTCPv4Ports(3000, 2000, 2900, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+
+	if len(w.gsoWrites) != 2 {
+		t.Fatalf("want 2 gso writes (one per flow), got %d", len(w.gsoWrites))
+	}
+	if len(w.writes) != 0 {
+		t.Fatalf("want no plain writes, got %d", len(w.writes))
+	}
+	// Each superpacket should carry 3 segments.
+	for i, g := range w.gsoWrites {
+		if len(g.pays) != 3 {
+			t.Errorf("gso[%d]: segs=%d want 3", i, len(g.pays))
+		}
+		if g.gsoSize != 1200 {
+			t.Errorf("gso[%d]: gsoSize=%d want 1200", i, g.gsoSize)
+		}
+	}
+	// Verify each superpacket carries the source port it was seeded with.
+	seenSports := map[uint16]bool{}
+	for _, g := range w.gsoWrites {
+		sp := binary.BigEndian.Uint16(g.hdr[20:22])
+		seenSports[sp] = true
+	}
+	if !seenSports[1000] || !seenSports[3000] {
+		t.Errorf("expected superpackets for sports 1000 and 3000, got %v", seenSports)
+	}
+}
+
+// TestCoalescerPreservesArrivalOrder confirms that with passthrough and
+// coalesced events both queued, Flush emits them in Add order rather than
+// writing passthrough packets synchronously.
+func TestCoalescerPreservesArrivalOrder(t *testing.T) {
+	w := &orderedFakeWriter{gsoEnabled: true}
+	c := NewTCPCoalescer(w)
+	// Sequence: coalesceable TCP, ICMP (passthrough), coalesceable TCP on
+	// a different flow. Expected emit order: gso(X), plain(ICMP), gso(Y).
+	pay := make([]byte, 1200)
+	if err := c.Commit(buildTCPv4Ports(1000, 2000, 100, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	icmp := make([]byte, 28)
+	icmp[0] = 0x45
+	binary.BigEndian.PutUint16(icmp[2:4], 28)
+	icmp[9] = 1
+	copy(icmp[12:16], []byte{10, 0, 0, 1})
+	copy(icmp[16:20], []byte{10, 0, 0, 3})
+	if err := c.Commit(icmp); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Commit(buildTCPv4Ports(3000, 2000, 500, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	// Nothing should have hit the writer synchronously.
+	if len(w.events) != 0 {
+		t.Fatalf("Add emitted events synchronously: %v", w.events)
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	if got, want := w.events, []string{"gso", "plain", "gso"}; !stringSliceEq(got, want) {
+		t.Fatalf("flush order=%v want %v", got, want)
+	}
+}
+
+// orderedFakeWriter records only the sequence of call types so tests can
+// assert arrival order without inspecting bytes.
+type orderedFakeWriter struct {
+	gsoEnabled bool
+	events     []string
+}
+
+func (w *orderedFakeWriter) Write(p []byte) (int, error) {
+	w.events = append(w.events, "plain")
+	return len(p), nil
+}
+
+func (w *orderedFakeWriter) WriteGSO(hdr []byte, transportHdr []byte, pays [][]byte) error {
+	w.events = append(w.events, "gso")
+	return nil
+}
+
+func (w *orderedFakeWriter) GSOSupported() bool { return w.gsoEnabled }
+
+func stringSliceEq(a, b []string) bool {
+	if len(a) != len(b) {
+		return false
+	}
+	for i := range a {
+		if a[i] != b[i] {
+			return false
+		}
+	}
+	return true
+}
+
+// TestCoalescerInterleavedFlowsPreserveOrdering checks that a non-admissible
+// packet (SYN) mid-flow only flushes its own flow, not others.
+func TestCoalescerInterleavedFlowsPreserveOrdering(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := NewTCPCoalescer(w)
+	pay := make([]byte, 1200)
+
+	// Flow A two segments.
+	if err := c.Commit(buildTCPv4Ports(1000, 2000, 100, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Commit(buildTCPv4Ports(1000, 2000, 1300, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	// Flow B two segments.
+	if err := c.Commit(buildTCPv4Ports(3000, 2000, 500, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Commit(buildTCPv4Ports(3000, 2000, 1700, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	// Flow A SYN (non-admissible) — must flush only flow A's slot.
+	syn := buildTCPv4Ports(1000, 2000, 9999, tcpSyn|tcpAck, pay)
+	if err := c.Commit(syn); err != nil {
+		t.Fatal(err)
+	}
+	// Flow B continues — should still be coalesced with its seed.
+	if err := c.Commit(buildTCPv4Ports(3000, 2000, 2900, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+
+	// Expected:
+	//   - 1 gso for flow A (first 2 segments)
+	//   - 1 plain for flow A SYN
+	//   - 1 gso for flow B (3 segments)
+	if len(w.gsoWrites) != 2 {
+		t.Fatalf("want 2 gso writes, got %d", len(w.gsoWrites))
+	}
+	if len(w.writes) != 1 {
+		t.Fatalf("want 1 plain write (SYN), got %d", len(w.writes))
+	}
+	// Find the 3-segment gso (flow B) and the 2-segment gso (flow A).
+	var segCounts []int
+	for _, g := range w.gsoWrites {
+		segCounts = append(segCounts, len(g.pays))
+	}
+	if !(segCounts[0] == 2 && segCounts[1] == 3) && !(segCounts[0] == 3 && segCounts[1] == 2) {
+		t.Errorf("unexpected segment counts: %v (want 2 and 3)", segCounts)
+	}
+}

overlay/batch/tx_batch.go

@@ -0,0 +1,61 @@
+package batch
+
+import "net/netip"
+
+const SendBatchCap = 128
+
+// SendBatch accumulates encrypted UDP packets for potential TX offloading.
+// One SendBatch is owned by each listenIn goroutine; no locking is needed.
+// The backing storage holds up to batchCap packets of slotCap bytes each;
+// bufs and dsts are parallel slices of committed slots.
+type SendBatch struct {
+	bufs     [][]byte
+	dsts     []netip.AddrPort
+	backing  []byte
+	slotCap  int
+	batchCap int
+	nextSlot int
+}
+
+func NewSendBatch(batchCap, slotCap int) *SendBatch {
+	return &SendBatch{
+		bufs:     make([][]byte, 0, batchCap),
+		dsts:     make([]netip.AddrPort, 0, batchCap),
+		backing:  make([]byte, batchCap*slotCap),
+		slotCap:  slotCap,
+		batchCap: batchCap,
+	}
+}
+
+func (b *SendBatch) Next() []byte {
+	if b.nextSlot >= b.batchCap {
+		return nil
+	}
+	start := b.nextSlot * b.slotCap
+	return b.backing[start : start : start+b.slotCap] //set len to 0 but cap to slotCap
+}
+
+func (b *SendBatch) Commit(n int, dst netip.AddrPort) {
+	start := b.nextSlot * b.slotCap
+	b.bufs = append(b.bufs, b.backing[start:start+n])
+	b.dsts = append(b.dsts, dst)
+	b.nextSlot++
+}
+
+func (b *SendBatch) Reset() {
+	b.bufs = b.bufs[:0]
+	b.dsts = b.dsts[:0]
+	b.nextSlot = 0
+}
+
+func (b *SendBatch) Len() int {
+	return len(b.bufs)
+}
+
+func (b *SendBatch) Cap() int {
+	return b.batchCap
+}
+
+func (b *SendBatch) Get() ([][]byte, []netip.AddrPort) {
+	return b.bufs, b.dsts
+}

overlay/batch/tx_batch_test.go

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

overlay/device.go

@@ -4,24 +4,21 @@ import (
 	"io"
 	"net/netip"
 
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 )
 
+// defaultBatchBufSize is the per-Queue scratch size for Read on backends
+// that don't do TSO segmentation. 65535 covers any single IP packet.
+const defaultBatchBufSize = 65535
+
 type Device interface {
-	io.ReadWriteCloser
+	io.Closer
 	Activate() error
 	Networks() []netip.Prefix
 	Name() string
 	RoutesFor(netip.Addr) routing.Gateways
-	SupportsMultiqueue() bool
+	SupportsMultiqueue() bool //todo remove?
-	NewMultiQueueReader() (io.ReadWriteCloser, error)
+	NewMultiQueueReader() error
-	// SupportsPerPeerMTU reports whether SetPeerMTU is implemented for real on
+	Readers() []tio.Queue
-	// this platform. PMTUD requires this; the manager will refuse to enable when
-	// false even if the operator set tun.max_mtu, because a discovered MTU we
-	// can't actually install does the operator no good.
-	SupportsPerPeerMTU() bool
-	// SetPeerMTU installs a per-peer MTU on the routing table so the kernel will
-	// surface PTB / EMSGSIZE for inside packets to that peer that would exceed mtu.
-	// Pass mtu=0 to remove the override and let the device default apply.
-	SetPeerMTU(addr netip.Addr, mtu int) error
 }

overlay/overlaytest/noop.go

@@ -4,9 +4,9 @@ package overlaytest
 
 import (
 	"errors"
-	"io"
 	"net/netip"
 
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 )
 
@@ -31,28 +31,24 @@ func (NoopTun) Name() string {
 	return "noop"
 }
 
-func (NoopTun) Read([]byte) (int, error) {
+func (NoopTun) Read() ([][]byte, error) {
-	return 0, nil
+	return nil, nil
 }
 
 func (NoopTun) Write([]byte) (int, error) {
 	return 0, nil
 }
 
-func (NoopTun) SupportsPerPeerMTU() bool {
-	return false
-}
-
-func (NoopTun) SetPeerMTU(addr netip.Addr, mtu int) error {
-	return nil
-}
-
 func (NoopTun) SupportsMultiqueue() bool {
 	return false
 }
 
-func (NoopTun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+func (NoopTun) NewMultiQueueReader() error {
-	return nil, errors.New("unsupported")
+	return errors.New("unsupported")
+}
+
+func (NoopTun) Readers() []tio.Queue {
+	return []tio.Queue{NoopTun{}}
 }
 
 func (NoopTun) Close() error {

overlay/tio/container_gso_linux.go

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

overlay/tio/container_poll_linux.go

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

overlay/tio/tio.go

@@ -0,0 +1,67 @@
+package tio
+
+import (
+	"io"
+)
+
+// defaultBatchBufSize is the per-Queue scratch size for Read on backends
+// that don't do TSO segmentation. 65535 covers any single IP packet.
+const defaultBatchBufSize = 65535
+
+// Container holds one or many Queue objects and helps close them in an orderly way
+type Container interface {
+	io.Closer
+	Queues() []Queue
+
+	// Add takes a tun fd, adds it to the container, and prepares it for use as a Queue
+	Add(fd int) error
+
+	io.Closer
+}
+
+// Queue is a readable/writable Poll queue. One Queue is driven by a single
+// read goroutine plus concurrent writers (see Write / WriteReject below).
+type Queue interface {
+	io.Closer
+
+	// Read returns one or more packets. The returned slices are borrowed
+	// from the Queue's internal buffer and are only valid until the next
+	// Read or Close on this Queue - callers must encrypt or copy each
+	// slice before the next call. Not safe for concurrent Reads.
+	Read() ([][]byte, error)
+
+	// Write emits a single packet on the plaintext (outside→inside)
+	// delivery path. Not safe for concurrent Writes.
+	Write(p []byte) (int, error)
+}
+
+// GSOWriter is implemented by Queues that can emit a TCP TSO superpacket
+// assembled from a header prefix plus one or more borrowed payload
+// fragments, in a single vectored write (writev with a leading
+// virtio_net_hdr). This lets the coalescer avoid copying payload bytes
+// between the caller's decrypt buffer and the TUN. Backends without GSO
+// support return false from GSOSupported and coalescing is skipped.
+//
+// hdr contains the IPv4/IPv6 + TCP header prefix (mutable - callers will
+// have filled in total length and pseudo-header partial). pays are
+// non-overlapping payload fragments whose concatenation is the full
+// superpacket payload; they are read-only from the writer's perspective
+// and must remain valid until the call returns. gsoSize is the MSS:
+// every segment except possibly the last is exactly that many bytes.
+// csumStart is the byte offset where the TCP header begins within hdr.
+//
+// # TODO fold into Queue
+//
+// hdr's TCP checksum field must already hold the pseudo-header partial
+// sum (single-fold, not inverted), per virtio NEEDS_CSUM semantics.
+type GSOWriter interface {
+	// WriteGSO emits a TCP TSO superpacket in a single writev. hdr is the
+	// IPv4/IPv6 + TCP header prefix (already finalized — total length, IP csum,
+	// and TCP pseudo-header partial set by the caller). pays are payload
+	// fragments whose concatenation forms the full coalesced payload; each
+	// slice is read-only and must stay valid until return.
+	// every segment in pays except possibly the last is exactly the same size.
+	// csumStart is the byte offset where the TCP header begins within hdr.
+	WriteGSO(hdr []byte, transportHdr []byte, pays [][]byte) error
+	GSOSupported() bool
+}

overlay/tio/tio_gso_linux.go

@@ -0,0 +1,353 @@
+package tio
+
+import (
+	"fmt"
+	"io"
+	"os"
+	"sync"
+	"sync/atomic"
+	"syscall"
+	"unsafe"
+
+	"golang.org/x/sys/unix"
+)
+
+// Space for segmented output. Worst case is many small segments, each paying
+// an IP+TCP header. Should be a multiple of 64KiB.
+// const tunSegBufSize = 0xffff * 8 TODO larger? config?
+const tunSegBufSize = 131072
+
+// tunSegBufCap is the total size we allocate for the per-reader segment
+// buffer. It is sized as one worst-case TSO superpacket (tunSegBufSize) plus
+// the same again as drain headroom so a Read wake can accumulate
+// additional packets after an initial big read without overflowing.
+const tunSegBufCap = tunSegBufSize * 2
+
+// tunDrainCap caps how many packets a single Read will accumulate via
+// the post-wake drain loop. Sized to soak up a burst of small ACKs while
+// bounding how much work a single caller holds before handing off.
+const tunDrainCap = 64 //256
+
+// gsoInitialPayIovs is the starting capacity (in payload fragments) of
+// Offload.gsoIovs. Sized to cover the default coalesce segment cap without
+// any reallocations.
+const gsoInitialPayIovs = 66
+
+// validVnetHdr is the 10-byte virtio_net_hdr we prepend to every non-GSO TUN
+// write. Only flag set is VIRTIO_NET_HDR_F_DATA_VALID, which marks the skb
+// CHECKSUM_UNNECESSARY so the receiving network stack skips L4 checks
+// verification. All packets that reach the plain Write paths
+// already carry a valid L4 checksum (either supplied by a remote peer whose
+// ciphertext we AEAD-authenticated, or produced by finishChecksum during TSO
+// segmentation, or built locally by CreateRejectPacket), so trusting them is
+// safe.
+var validVnetHdr = [virtioNetHdrLen]byte{unix.VIRTIO_NET_HDR_F_DATA_VALID}
+
+// Offload wraps a TUN file descriptor with poll-based reads. The FD provided will be changed to non-blocking.
+// A shared eventfd allows Close to wake all readers blocked in poll.
+type Offload struct {
+	fd         int
+	shutdownFd int
+	readPoll   [2]unix.PollFd
+	writePoll  [2]unix.PollFd
+	writeLock  sync.Mutex //there's more than one potential write source per-routine, so we need this to protect writePoll
+	closed     atomic.Bool
+	readBuf    []byte   // scratch for a single raw read (virtio hdr + superpacket)
+	segBuf     []byte   // backing store for segmented output
+	segOff     int      // cursor into segBuf for the current Read drain
+	pending    [][]byte // segments returned from the most recent Read
+
+	// gsoHdrBuf is a per-queue 10-byte scratch for the virtio_net_hdr emitted
+	// by WriteGSO. Separate from validVnetHdr so a concurrent non-GSO Write on
+	// another queue never observes a half-written header.
+	gsoHdrBuf [virtioNetHdrLen]byte
+	// gsoIovs is the writev iovec scratch for WriteGSO. Sized to hold the
+	// virtio header + IP/TCP header + up to gsoInitialPayIovs payload
+	// fragments; grown on demand if a coalescer pushes more.
+	gsoIovs []unix.Iovec
+}
+
+func newOffload(fd int, shutdownFd int) (*Offload, error) {
+	if err := unix.SetNonblock(fd, true); err != nil {
+		return nil, fmt.Errorf("failed to set tun fd non-blocking: %w", err)
+	}
+
+	out := &Offload{
+		fd:         fd,
+		shutdownFd: shutdownFd,
+		closed:     atomic.Bool{},
+		readBuf:    make([]byte, virtioNetHdrLen+tunReadBufSize),
+		readPoll: [2]unix.PollFd{
+			{Fd: int32(fd), Events: unix.POLLIN},
+			{Fd: int32(shutdownFd), Events: unix.POLLIN},
+		},
+		writePoll: [2]unix.PollFd{
+			{Fd: int32(fd), Events: unix.POLLOUT},
+			{Fd: int32(shutdownFd), Events: unix.POLLIN},
+		},
+		writeLock: sync.Mutex{},
+
+		segBuf:  make([]byte, tunSegBufCap),
+		gsoIovs: make([]unix.Iovec, 2, 2+gsoInitialPayIovs),
+	}
+
+	out.gsoIovs[0].Base = &out.gsoHdrBuf[0]
+	out.gsoIovs[0].SetLen(virtioNetHdrLen)
+
+	return out, nil
+}
+
+func (r *Offload) blockOnRead() error {
+	const problemFlags = unix.POLLHUP | unix.POLLNVAL | unix.POLLERR
+	var err error
+	for {
+		_, err = unix.Poll(r.readPoll[:], -1)
+		if err != unix.EINTR {
+			break
+		}
+	}
+	//always reset these!
+	tunEvents := r.readPoll[0].Revents
+	shutdownEvents := r.readPoll[1].Revents
+	r.readPoll[0].Revents = 0
+	r.readPoll[1].Revents = 0
+	//do the err check before trusting the potentially bogus bits we just got
+	if err != nil {
+		return err
+	}
+	if shutdownEvents&(unix.POLLIN|problemFlags) != 0 {
+		return os.ErrClosed
+	} else if tunEvents&problemFlags != 0 {
+		return os.ErrClosed
+	}
+	return nil
+}
+
+func (r *Offload) blockOnWrite() error {
+	const problemFlags = unix.POLLHUP | unix.POLLNVAL | unix.POLLERR
+	var err error
+	for {
+		_, err = unix.Poll(r.writePoll[:], -1)
+		if err != unix.EINTR {
+			break
+		}
+	}
+	//always reset these!
+	r.writeLock.Lock()
+	tunEvents := r.writePoll[0].Revents
+	shutdownEvents := r.writePoll[1].Revents
+	r.writePoll[0].Revents = 0
+	r.writePoll[1].Revents = 0
+	r.writeLock.Unlock()
+	//do the err check before trusting the potentially bogus bits we just got
+	if err != nil {
+		return err
+	}
+	if shutdownEvents&(unix.POLLIN|problemFlags) != 0 {
+		return os.ErrClosed
+	} else if tunEvents&problemFlags != 0 {
+		return os.ErrClosed
+	}
+	return nil
+}
+
+func (r *Offload) readRaw(buf []byte) (int, error) {
+	for {
+		if n, err := unix.Read(r.fd, buf); err == nil {
+			return n, nil
+		} else if err == unix.EAGAIN {
+			if err = r.blockOnRead(); err != nil {
+				return 0, err
+			}
+			continue
+		} else if err == unix.EINTR {
+			continue
+		} else if err == unix.EBADF {
+			return 0, os.ErrClosed
+		} else {
+			return 0, err
+		}
+	}
+}
+
+// Read reads one or more superpackets from the tun and returns the
+// resulting packets. The first read blocks via poll; once the fd is known
+// readable we drain additional packets non-blocking until the kernel queue
+// is empty (EAGAIN), we've collected tunDrainCap packets, or we're out of
+// segBuf headroom. This amortizes the poll wake over bursts of small
+// packets (e.g. TCP ACKs). Slices point into the Offload's internal buffers
+// and are only valid until the next Read or Close on this Queue.
+func (r *Offload) Read() ([][]byte, error) {
+	r.pending = r.pending[:0]
+	r.segOff = 0
+
+	// Initial (blocking) read. Retry on decode errors so a single bad
+	// packet does not stall the reader.
+	for {
+		n, err := r.readRaw(r.readBuf)
+		if err != nil {
+			return nil, err
+		}
+		if err := r.decodeRead(n); err != nil {
+			// Drop and read again — a bad packet should not kill the reader.
+			continue
+		}
+		break
+	}
+
+	// Drain: non-blocking reads until the kernel queue is empty, the drain
+	// cap is reached, or segBuf no longer has room for another worst-case
+	// superpacket.
+	for len(r.pending) < tunDrainCap && tunSegBufCap-r.segOff >= tunSegBufSize {
+		n, err := unix.Read(r.fd, r.readBuf)
+		if err != nil {
+			// EAGAIN / EINTR / anything else: stop draining. We already
+			// have a valid batch from the first read.
+			break
+		}
+		if n <= 0 {
+			break
+		}
+		if err := r.decodeRead(n); err != nil {
+			// Drop this packet and stop the drain; we'd rather hand off
+			// what we have than keep spinning here.
+			break
+		}
+	}
+
+	return r.pending, nil
+}
+
+// decodeRead decodes the virtio header plus payload in r.readBuf[:n], appends
+// the segments to r.pending, and advances r.segOff by the total scratch used.
+func (r *Offload) decodeRead(n int) error {
+	if n < virtioNetHdrLen {
+		return fmt.Errorf("short tun read: %d < %d", n, virtioNetHdrLen)
+	}
+	var hdr VirtioNetHdr
+	hdr.decode(r.readBuf[:virtioNetHdrLen])
+	before := len(r.pending)
+	if err := segmentInto(r.readBuf[virtioNetHdrLen:n], hdr, &r.pending, r.segBuf[r.segOff:]); err != nil {
+		return err
+	}
+	for k := before; k < len(r.pending); k++ {
+		r.segOff += len(r.pending[k])
+	}
+	return nil
+}
+
+func (r *Offload) Write(buf []byte) (int, error) {
+	iovs := [2]unix.Iovec{
+		{Base: &validVnetHdr[0]},
+		{Base: &buf[0]},
+	}
+	iovs[0].SetLen(virtioNetHdrLen)
+	iovs[1].SetLen(len(buf))
+	return r.writeWithScratch(buf, &iovs)
+}
+
+func (r *Offload) writeWithScratch(buf []byte, iovs *[2]unix.Iovec) (int, error) {
+	if len(buf) == 0 {
+		return 0, nil
+	}
+	// Point the payload iovec at the caller's buffer. iovs[0] is pre-wired
+	// to validVnetHdr during Offload construction so we don't rebuild it here.
+	iovs[1].Base = &buf[0]
+	iovs[1].SetLen(len(buf))
+	return r.rawWrite(unsafe.Slice(&iovs[0], len(iovs)))
+}
+
+func (r *Offload) rawWrite(iovs []unix.Iovec) (int, error) {
+	for {
+		n, _, errno := syscall.Syscall(unix.SYS_WRITEV, uintptr(r.fd), uintptr(unsafe.Pointer(&iovs[0])), uintptr(len(iovs)))
+		if errno == 0 {
+			if int(n) < virtioNetHdrLen {
+				return 0, io.ErrShortWrite
+			}
+			return int(n) - virtioNetHdrLen, nil
+		}
+		if errno == unix.EAGAIN {
+			if err := r.blockOnWrite(); err != nil {
+				return 0, err
+			}
+			continue
+		}
+		if errno == unix.EINTR {
+			continue
+		}
+		if errno == unix.EBADF {
+			return 0, os.ErrClosed
+		}
+		return 0, errno
+	}
+}
+
+// GSOSupported reports whether this queue was opened with IFF_VNET_HDR and
+// can accept WriteGSO. When false, callers should fall back to per-segment
+// Write calls.
+func (r *Offload) GSOSupported() bool { return true }
+
+func (r *Offload) WriteGSO(hdr []byte, transportHdr []byte, pays [][]byte) error {
+	if len(hdr) == 0 || len(pays) == 0 || len(transportHdr) == 0 {
+		return nil
+	}
+	vhdr := VirtioNetHdr{
+		Flags:      unix.VIRTIO_NET_HDR_F_NEEDS_CSUM,
+		HdrLen:     uint16(len(hdr) + len(transportHdr)),
+		GSOSize:    uint16(len(pays[0])),
+		CsumStart:  uint16(len(hdr)),
+		CsumOffset: 16, // TCP checksum field lives 16 bytes into the TCP header
+	}
+	if len(pays) > 1 {
+		ipVer := hdr[0] >> 4
+		if ipVer == 6 {
+			vhdr.GSOType = unix.VIRTIO_NET_HDR_GSO_TCPV6
+		} else if ipVer == 4 {
+			vhdr.GSOType = unix.VIRTIO_NET_HDR_GSO_TCPV4
+		} else {
+			vhdr.GSOType = unix.VIRTIO_NET_HDR_GSO_NONE
+			vhdr.GSOSize = 0
+		}
+	} else {
+		vhdr.GSOType = unix.VIRTIO_NET_HDR_GSO_NONE
+		vhdr.GSOSize = 0
+	}
+	vhdr.encode(r.gsoHdrBuf[:])
+
+	// Build the iovec array: [virtio_hdr, hdr, transportHdr, pays...]. r.gsoIovs[0] is
+	// wired to gsoHdrBuf at construction and never changes.
+	need := 3 + len(pays)
+	if cap(r.gsoIovs) < need {
+		grown := make([]unix.Iovec, need)
+		grown[0] = r.gsoIovs[0]
+		r.gsoIovs = grown
+	} else {
+		r.gsoIovs = r.gsoIovs[:need]
+	}
+	r.gsoIovs[1].Base = &hdr[0]
+	r.gsoIovs[1].SetLen(len(hdr))
+	r.gsoIovs[2].Base = &transportHdr[0]
+	r.gsoIovs[2].SetLen(len(transportHdr))
+	for i, p := range pays {
+		r.gsoIovs[3+i].Base = &p[0]
+		r.gsoIovs[3+i].SetLen(len(p))
+	}
+
+	_, err := r.rawWrite(r.gsoIovs)
+	return err
+}
+
+func (r *Offload) Close() error {
+	if r.closed.Swap(true) {
+		return nil
+	}
+
+	//shutdownFd is owned by the container, so we should not close it
+	var err error
+	if r.fd >= 0 {
+		err = unix.Close(r.fd)
+		r.fd = -1
+	}
+
+	return err
+}

overlay/tio/tio_poll_linux.go

@@ -0,0 +1,164 @@
+package tio
+
+import (
+	"fmt"
+	"os"
+	"sync/atomic"
+
+	"golang.org/x/sys/unix"
+)
+
+// Maximum size we accept for a single read from a TUN with IFF_VNET_HDR. A
+// TSO superpacket can be up to 64KiB of payload plus a single L2/L3/L4 header
+// prefix plus the virtio header.
+const tunReadBufSize = 65535
+
+type Poll struct {
+	fd int
+
+	readPoll  [2]unix.PollFd
+	writePoll [2]unix.PollFd
+	closed    atomic.Bool
+
+	readBuf  []byte
+	batchRet [1][]byte
+}
+
+func newPoll(fd int, shutdownFd int) (*Poll, error) {
+	if err := unix.SetNonblock(fd, true); err != nil {
+		_ = unix.Close(fd)
+		return nil, fmt.Errorf("failed to set Poll device as nonblocking: %w", err)
+	}
+
+	out := &Poll{
+		fd:      fd,
+		readBuf: make([]byte, tunReadBufSize),
+		readPoll: [2]unix.PollFd{
+			{Fd: int32(fd), Events: unix.POLLIN},
+			{Fd: int32(shutdownFd), Events: unix.POLLIN},
+		},
+		writePoll: [2]unix.PollFd{
+			{Fd: int32(fd), Events: unix.POLLOUT},
+			{Fd: int32(shutdownFd), Events: unix.POLLIN},
+		},
+	}
+	return out, nil
+}
+
+// blockOnRead waits until the Poll fd is readable or shutdown has been signaled.
+// Returns os.ErrClosed if Close was called.
+func (t *Poll) blockOnRead() error {
+	const problemFlags = unix.POLLHUP | unix.POLLNVAL | unix.POLLERR
+	var err error
+	for {
+		_, err = unix.Poll(t.readPoll[:], -1)
+		if err != unix.EINTR {
+			break
+		}
+	}
+	tunEvents := t.readPoll[0].Revents
+	shutdownEvents := t.readPoll[1].Revents
+	t.readPoll[0].Revents = 0
+	t.readPoll[1].Revents = 0
+	if err != nil {
+		return err
+	}
+	if shutdownEvents&(unix.POLLIN|problemFlags) != 0 {
+		return os.ErrClosed
+	}
+	if tunEvents&problemFlags != 0 {
+		return os.ErrClosed
+	}
+	return nil
+}
+
+func (t *Poll) blockOnWrite() error {
+	const problemFlags = unix.POLLHUP | unix.POLLNVAL | unix.POLLERR
+	var err error
+	for {
+		_, err = unix.Poll(t.writePoll[:], -1)
+		if err != unix.EINTR {
+			break
+		}
+	}
+	tunEvents := t.writePoll[0].Revents
+	shutdownEvents := t.writePoll[1].Revents
+	t.writePoll[0].Revents = 0
+	t.writePoll[1].Revents = 0
+	if err != nil {
+		return err
+	}
+	if shutdownEvents&(unix.POLLIN|problemFlags) != 0 {
+		return os.ErrClosed
+	}
+	if tunEvents&problemFlags != 0 {
+		return os.ErrClosed
+	}
+	return nil
+}
+
+func (t *Poll) Read() ([][]byte, error) {
+	n, err := t.readOne(t.readBuf)
+	if err != nil {
+		return nil, err
+	}
+	t.batchRet[0] = t.readBuf[:n]
+	return t.batchRet[:], nil
+}
+
+func (t *Poll) readOne(to []byte) (int, error) {
+	for {
+		n, errno := unix.Read(t.fd, to)
+		if errno == nil {
+			return n, nil
+		}
+		switch errno {
+		case unix.EAGAIN:
+			if err := t.blockOnRead(); err != nil {
+				return 0, err
+			}
+		case unix.EINTR:
+			// retry
+		case unix.EBADF:
+			return 0, os.ErrClosed
+		default:
+			return 0, errno
+		}
+	}
+}
+
+// Write is only valid for single threaded use
+func (t *Poll) Write(from []byte) (int, error) {
+	for {
+		n, errno := unix.Write(t.fd, from)
+		if errno == nil {
+			return n, nil
+		}
+		switch errno {
+		case unix.EAGAIN:
+			if err := t.blockOnWrite(); err != nil {
+				return 0, err
+			}
+		case unix.EINTR:
+			// retry
+		case unix.EBADF:
+			return 0, os.ErrClosed
+		default:
+			return 0, errno
+		}
+	}
+}
+
+func (t *Poll) Close() error {
+	if t.closed.Swap(true) {
+		return nil
+	}
+	//shutdownFd is owned by the container, so we should not close it
+	var err error
+	if t.fd >= 0 {
+		err = unix.Close(t.fd)
+		t.fd = -1
+	}
+
+	return err
+}

overlay/tio/tun_file_linux_test.go

@@ -0,0 +1,82 @@
+//go:build linux && !android && !e2e_testing
+// +build linux,!android,!e2e_testing
+
+package tio
+
+import (
+	"errors"
+	"os"
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/require"
+	"golang.org/x/sys/unix"
+)
+
+// newReadPipe returns a read fd. The matching write fd is registered for cleanup.
+// The caller takes ownership of the read fd (pass it to newOffload / newFriend).
+func newReadPipe(t *testing.T) int {
+	t.Helper()
+	var fds [2]int
+	if err := unix.Pipe2(fds[:], unix.O_CLOEXEC); err != nil {
+		t.Fatalf("pipe2: %v", err)
+	}
+	t.Cleanup(func() { _ = unix.Close(fds[1]) })
+	return fds[0]
+}
+
+func TestPoll_WakeForShutdown_WakesFriends(t *testing.T) {
+	pipe1 := newReadPipe(t)
+	pipe2 := newReadPipe(t)
+	parent, err := NewPollContainer()
+	require.NoError(t, err)
+	require.NoError(t, parent.Add(pipe1))
+	require.NoError(t, parent.Add(pipe2))
+	t.Cleanup(func() {
+		_ = unix.Close(pipe1)
+		_ = unix.Close(pipe2)
+	})
+
+	readers := parent.Queues()
+	errs := make([]error, len(readers))
+	var wg sync.WaitGroup
+	for i, r := range readers {
+		wg.Add(1)
+		go func(i int, r Queue) {
+			defer wg.Done()
+			_, errs[i] = r.Read()
+		}(i, r)
+	}
+
+	time.Sleep(50 * time.Millisecond)
+
+	if err := parent.Close(); err != nil {
+		t.Fatalf("Close: %v", err)
+	}
+
+	done := make(chan struct{})
+	go func() { wg.Wait(); close(done) }()
+	select {
+	case <-done:
+	case <-time.After(2 * time.Second):
+		t.Fatal("readers did not wake")
+	}
+
+	for i, err := range errs {
+		if !errors.Is(err, os.ErrClosed) {
+			t.Errorf("reader %d: expected os.ErrClosed, got %v", i, err)
+		}
+	}
+}
+
+func TestPoll_Close_Idempotent(t *testing.T) {
+	tf, err := newPoll(newReadPipe(t), 1)
+	require.NoError(t, err)
+	if err := tf.Close(); err != nil {
+		t.Fatalf("first Close: %v", err)
+	}
+	if err := tf.Close(); err != nil {
+		t.Fatalf("second Close should be a no-op, got %v", err)
+	}
+}

overlay/tio/tun_linux_offload.go

@@ -0,0 +1,331 @@
+//go:build linux && !android && !e2e_testing
+// +build linux,!android,!e2e_testing
+
+package tio
+
+import (
+	"encoding/binary"
+	"errors"
+	"fmt"
+
+	"golang.org/x/sys/unix"
+	"gvisor.dev/gvisor/pkg/tcpip/checksum"
+)
+
+// Protocol header size bounds used to validate / cap kernel-supplied offsets.
+const (
+	ipv4HeaderMinLen = 20 // IHL=5, no options
+	ipv4HeaderMaxLen = 60 // IHL=15, max options
+	ipv6FixedLen     = 40 // IPv6 base header; extensions would extend this
+	tcpHeaderMinLen  = 20 // data-offset=5, no options
+	tcpHeaderMaxLen  = 60 // data-offset=15, max options
+)
+
+// Byte offsets inside an IPv4 header.
+const (
+	ipv4TotalLenOff = 2
+	ipv4IDOff       = 4
+	ipv4ChecksumOff = 10
+	ipv4SrcOff      = 12
+	ipv4AddrsEnd    = 20 // end of dst address (ipv4SrcOff + 2*4)
+)
+
+// Byte offsets inside an IPv6 header.
+const (
+	ipv6PayloadLenOff = 4
+	ipv6SrcOff        = 8
+	ipv6AddrsEnd      = 40 // end of dst address (ipv6SrcOff + 2*16)
+)
+
+// Byte offsets inside a TCP header (relative to its start, i.e. csumStart).
+const (
+	tcpSeqOff      = 4
+	tcpDataOffOff  = 12 // upper nibble is header len in 32-bit words
+	tcpFlagsOff    = 13
+	tcpChecksumOff = 16
+)
+
+// tcpFinPshMask is cleared on every segment except the last of a TSO burst.
+const tcpFinPshMask = 0x09 // FIN(0x01) | PSH(0x08)
+
+func checkVirtioValid(pkt []byte, hdr VirtioNetHdr) error {
+	// When RSC_INFO is set the csum_start/csum_offset fields are repurposed to
+	// carry coalescing info rather than checksum offsets. A TUN writing via
+	// IFF_VNET_HDR should never emit this, but if it did we would silently
+	// miscompute the segment checksums — refuse the packet instead.
+	if hdr.Flags&unix.VIRTIO_NET_HDR_F_RSC_INFO != 0 {
+		return fmt.Errorf("virtio RSC_INFO flag not supported on TUN reads")
+	}
+	if len(pkt) < ipv4HeaderMinLen {
+		return fmt.Errorf("packet too short")
+	}
+	ipVersion := pkt[0] >> 4
+	switch hdr.GSOType {
+	case unix.VIRTIO_NET_HDR_GSO_TCPV4:
+		if ipVersion != 4 {
+			return fmt.Errorf("invalid IP version %d for GSO type %d", ipVersion, hdr.GSOType)
+		}
+	case unix.VIRTIO_NET_HDR_GSO_TCPV6:
+		if ipVersion != 6 {
+			return fmt.Errorf("invalid IP version %d for GSO type %d", ipVersion, hdr.GSOType)
+		}
+	default:
+		if !(ipVersion == 6 || ipVersion == 4) {
+			return fmt.Errorf("invalid IP version %d for GSO type %d", ipVersion, hdr.GSOType)
+		}
+	}
+
+	return nil
+}
+
+func handleGSONone(pkt []byte, hdr VirtioNetHdr, out *[][]byte, scratch []byte) error {
+	if len(pkt) > len(scratch) {
+		return fmt.Errorf("packet larger than segment buffer: %d > %d", len(pkt), len(scratch))
+	}
+	copy(scratch, pkt)
+	seg := scratch[:len(pkt)]
+	if hdr.Flags&unix.VIRTIO_NET_HDR_F_NEEDS_CSUM != 0 {
+		if err := finishChecksum(seg, hdr); err != nil {
+			return err
+		}
+	}
+	*out = append(*out, seg)
+	return nil
+}
+
+func correctHdrLen(pkt []byte, hdr *VirtioNetHdr) error {
+	// Thank you wireguard-go for documenting these edge-cases
+	// Don't trust hdr.hdrLen from the kernel as it can be equal to the length
+	// of the entire first packet when the kernel is handling it as part of a
+	// FORWARD path. Instead, parse the transport header length and add it onto
+	// csumStart, which is synonymous for IP header length.
+	const tcpDataOffset = 12
+
+	if hdr.GSOType == unix.VIRTIO_NET_HDR_GSO_UDP_L4 {
+		hdr.HdrLen = hdr.CsumStart + 8
+	} else {
+		if len(pkt) <= int(hdr.CsumStart+tcpDataOffset) {
+			return errors.New("packet is too short")
+		}
+
+		tcpHLen := uint16(pkt[hdr.CsumStart+tcpDataOffset] >> 4 * 4)
+		if tcpHLen < 20 || tcpHLen > 60 {
+			// A TCP header must be between 20 and 60 bytes in length.
+			return fmt.Errorf("tcp header len is invalid: %d", tcpHLen)
+		}
+		hdr.HdrLen = hdr.CsumStart + tcpHLen
+	}
+
+	if len(pkt) < int(hdr.HdrLen) {
+		return fmt.Errorf("length of packet (%d) < virtioNetHdr.HdrLen (%d)", len(pkt), hdr.HdrLen)
+	}
+
+	if hdr.HdrLen < hdr.CsumStart {
+		return fmt.Errorf("virtioNetHdr.HdrLen (%d) < virtioNetHdr.CsumStart (%d)", hdr.HdrLen, hdr.CsumStart)
+	}
+	cSumAt := int(hdr.CsumStart + hdr.CsumStart)
+	if cSumAt+1 >= len(pkt) {
+		return fmt.Errorf("end of checksum offset (%d) exceeds packet length (%d)", cSumAt+1, len(pkt))
+	}
+	return nil
+}
+
+// segmentInto splits a TUN-side packet described by hdr into one or more
+// IP packets, each appended to *out as a slice of scratch. scratch must be
+// sized to hold every segment (including replicated headers).
+func segmentInto(pkt []byte, hdr VirtioNetHdr, out *[][]byte, scratch []byte) error {
+	if err := checkVirtioValid(pkt, hdr); err != nil {
+		return err
+	}
+	if hdr.GSOType == unix.VIRTIO_NET_HDR_GSO_NONE {
+		return handleGSONone(pkt, hdr, out, scratch)
+	}
+	if err := correctHdrLen(pkt, &hdr); err != nil {
+		return err
+	}
+	switch hdr.GSOType {
+	case unix.VIRTIO_NET_HDR_GSO_TCPV4, unix.VIRTIO_NET_HDR_GSO_TCPV6:
+		return segmentTCP(pkt, hdr, out, scratch)
+
+	default:
+		return fmt.Errorf("unsupported virtio gso type: %d", hdr.GSOType)
+	}
+}
+
+// finishChecksum computes the L4 checksum for a non-GSO packet that the kernel
+// handed us with NEEDS_CSUM set. csum_start / csum_offset point at the 16-bit
+// checksum field; we zero it, fold a full sum (the field was pre-loaded with
+// the pseudo-header partial sum by the kernel), and store the result.
+func finishChecksum(seg []byte, hdr VirtioNetHdr) error {
+	cs := int(hdr.CsumStart)
+	co := int(hdr.CsumOffset)
+	if cs+co+2 > len(seg) {
+		return fmt.Errorf("csum offsets out of range: start=%d offset=%d len=%d", cs, co, len(seg))
+	}
+	// The kernel stores a partial pseudo-header sum at [cs+co:]; sum over the
+	// L4 region starting at cs, folding the prior partial in as the seed.
+	partial := binary.BigEndian.Uint16(seg[cs+co : cs+co+2])
+	seg[cs+co] = 0
+	seg[cs+co+1] = 0
+	binary.BigEndian.PutUint16(seg[cs+co:cs+co+2], ^checksum.Checksum(seg[cs:], partial))
+	return nil
+}
+
+// segmentTCP software-segments a TSO superpacket into one IP packet per MSS
+// chunk. The caller guarantees hdr.GSOType is TCPV4 or TCPV6.
+//
+// Hot-path shape: the per-segment loop only sums the payload chunk. The TCP
+// header, the IPv4 header, and the pseudo-header src/dst/proto contributions
+// are each summed once up front — every segment reuses those three pre-folded
+// uint32 values and combines them with small per-segment deltas (seq, flags,
+// tcpLen, ip_id, total_len) that are cheap to fold in.
+func segmentTCP(pkt []byte, hdr VirtioNetHdr, out *[][]byte, scratch []byte) error {
+	if hdr.GSOSize == 0 {
+		return fmt.Errorf("gso_size is zero")
+	}
+	if hdr.CsumStart == 0 {
+		return fmt.Errorf("csum_start is zero")
+	}
+
+	isV4 := hdr.GSOType == unix.VIRTIO_NET_HDR_GSO_TCPV4
+	headerLen := int(hdr.HdrLen) // already corrected by the caller
+	csumStart := int(hdr.CsumStart)
+
+	tcpHdrLen := int(pkt[csumStart+tcpDataOffOff]>>4) * 4
+
+	payload := pkt[headerLen:]
+	payLen := len(payload)
+	gsoSize := int(hdr.GSOSize)
+	numSeg := (payLen + gsoSize - 1) / gsoSize
+	if numSeg == 0 {
+		numSeg = 1
+	}
+
+	need := numSeg*headerLen + payLen
+	if need > len(scratch) {
+		return fmt.Errorf("scratch too small for %d segments: need %d have %d", numSeg, need, len(scratch))
+	}
+
+	origSeq := binary.BigEndian.Uint32(pkt[csumStart+tcpSeqOff : csumStart+tcpSeqOff+4])
+	origFlags := pkt[csumStart+tcpFlagsOff]
+
+	// Precompute the TCP header sum with seq/flags/csum zeroed. Copy onto
+	// the stack, zero the per-segment-varying fields, sum once.
+	var tmp [tcpHeaderMaxLen]byte
+	copy(tmp[:tcpHdrLen], pkt[csumStart:headerLen])
+	tmp[tcpSeqOff], tmp[tcpSeqOff+1], tmp[tcpSeqOff+2], tmp[tcpSeqOff+3] = 0, 0, 0, 0
+	tmp[tcpFlagsOff] = 0
+	tmp[tcpChecksumOff], tmp[tcpChecksumOff+1] = 0, 0
+	baseTcpHdrSum := uint32(checksum.Checksum(tmp[:tcpHdrLen], 0))
+
+	// Pseudo-header src+dst+proto contribution (tcpLen varies per segment).
+	var baseProtoSum uint32
+	if isV4 {
+		baseProtoSum = uint32(checksum.Checksum(pkt[ipv4SrcOff:ipv4AddrsEnd], 0))
+	} else {
+		baseProtoSum = uint32(checksum.Checksum(pkt[ipv6SrcOff:ipv6AddrsEnd], 0))
+	}
+	baseProtoSum += uint32(unix.IPPROTO_TCP)
+
+	// Precompute IPv4 header sum with total_len/id/csum zeroed.
+	var origIPID uint16
+	var ihl int
+	var baseIPHdrSum uint32
+	if isV4 {
+		origIPID = binary.BigEndian.Uint16(pkt[ipv4IDOff : ipv4IDOff+2])
+		ihl = int(pkt[0]&0x0f) * 4
+		if ihl < ipv4HeaderMinLen || ihl > csumStart {
+			return fmt.Errorf("bad IPv4 IHL: %d", ihl)
+		}
+		var ipTmp [ipv4HeaderMaxLen]byte
+		copy(ipTmp[:ihl], pkt[:ihl])
+		ipTmp[ipv4TotalLenOff], ipTmp[ipv4TotalLenOff+1] = 0, 0
+		ipTmp[ipv4IDOff], ipTmp[ipv4IDOff+1] = 0, 0
+		ipTmp[ipv4ChecksumOff], ipTmp[ipv4ChecksumOff+1] = 0, 0
+		baseIPHdrSum = uint32(checksum.Checksum(ipTmp[:ihl], 0))
+	}
+
+	off := 0
+	for i := 0; i < numSeg; i++ {
+		segStart := i * gsoSize
+		segEnd := segStart + gsoSize
+		if segEnd > payLen {
+			segEnd = payLen
+		}
+		segPayLen := segEnd - segStart
+
+		copy(scratch[off:], pkt[:headerLen])
+		copy(scratch[off+headerLen:], payload[segStart:segEnd])
+		seg := scratch[off : off+headerLen+segPayLen]
+		off += headerLen + segPayLen
+
+		segSeq := origSeq + uint32(segStart)
+		segFlags := origFlags
+		if i != numSeg-1 {
+			segFlags = origFlags &^ tcpFinPshMask
+		}
+		totalLen := headerLen + segPayLen
+
+		// Patch IP header and write the v4 header checksum from the precomputed base.
+		if isV4 {
+			segID := origIPID + uint16(i)
+			binary.BigEndian.PutUint16(seg[ipv4TotalLenOff:ipv4TotalLenOff+2], uint16(totalLen))
+			binary.BigEndian.PutUint16(seg[ipv4IDOff:ipv4IDOff+2], segID)
+			ipSum := baseIPHdrSum + uint32(totalLen) + uint32(segID)
+			binary.BigEndian.PutUint16(seg[ipv4ChecksumOff:ipv4ChecksumOff+2], foldComplement(ipSum))
+		} else {
+			// IPv6 payload length excludes the fixed header but includes any
+			// extension headers between [ipv6FixedLen:csumStart].
+			binary.BigEndian.PutUint16(seg[ipv6PayloadLenOff:ipv6PayloadLenOff+2], uint16(headerLen-ipv6FixedLen+segPayLen))
+		}
+
+		// Patch TCP header.
+		binary.BigEndian.PutUint32(seg[csumStart+tcpSeqOff:csumStart+tcpSeqOff+4], segSeq)
+		seg[csumStart+tcpFlagsOff] = segFlags
+		// (csum is written below; its prior contents in `seg` don't affect the
+		// computation since we never sum over the segment's own header.)
+
+		tcpLen := tcpHdrLen + segPayLen
+		paySum := uint32(checksum.Checksum(payload[segStart:segEnd], 0))
+
+		// Combine pre-folded uint32s into a wider accumulator, then fold. Using
+		// uint64 guards against overflow when segSeq's high bits set.
+		wide := uint64(baseTcpHdrSum) + uint64(paySum) + uint64(baseProtoSum)
+		wide += uint64(segSeq) + uint64(segFlags) + uint64(tcpLen)
+		wide = (wide & 0xffffffff) + (wide >> 32)
+		wide = (wide & 0xffffffff) + (wide >> 32)
+		binary.BigEndian.PutUint16(seg[csumStart+tcpChecksumOff:csumStart+tcpChecksumOff+2], foldComplement(uint32(wide)))
+
+		*out = append(*out, seg)
+	}
+
+	return nil
+}
+
+// foldComplement folds a 32-bit one's-complement partial sum to 16 bits and
+// complements it, yielding the on-wire Internet checksum value.
+func foldComplement(sum uint32) uint16 {
+	sum = (sum & 0xffff) + (sum >> 16)
+	sum = (sum & 0xffff) + (sum >> 16)
+	return ^uint16(sum)
+}
+
+// pseudoHeaderIPv4 returns the folded pseudo-header sum used to verify a TCP
+// segment's checksum in tests. src/dst are 4 bytes each.
+func pseudoHeaderIPv4(src, dst []byte, proto byte, tcpLen int) uint16 {
+	s := uint32(checksum.Checksum(src, 0)) + uint32(checksum.Checksum(dst, 0))
+	s += uint32(proto) + uint32(tcpLen)
+	s = (s & 0xffff) + (s >> 16)
+	s = (s & 0xffff) + (s >> 16)
+	return uint16(s)
+}
+
+// pseudoHeaderIPv6 returns the folded pseudo-header sum used to verify a TCP
+// segment's checksum in tests. src/dst are 16 bytes each.
+func pseudoHeaderIPv6(src, dst []byte, proto byte, tcpLen int) uint16 {
+	s := uint32(checksum.Checksum(src, 0)) + uint32(checksum.Checksum(dst, 0))
+	s += uint32(tcpLen>>16) + uint32(tcpLen&0xffff) + uint32(proto)
+	s = (s & 0xffff) + (s >> 16)
+	s = (s & 0xffff) + (s >> 16)
+	return uint16(s)
+}

overlay/tio/tun_linux_offload_test.go

@@ -0,0 +1,328 @@
+//go:build linux && !android && !e2e_testing
+// +build linux,!android,!e2e_testing
+
+package tio
+
+import (
+	"encoding/binary"
+	"os"
+	"testing"
+
+	"golang.org/x/sys/unix"
+	"gvisor.dev/gvisor/pkg/tcpip/checksum"
+)
+
+// verifyChecksum confirms that the one's-complement sum across `b`, seeded
+// with a folded pseudo-header sum, equals all-ones (valid).
+func verifyChecksum(b []byte, pseudo uint16) bool {
+	return checksum.Checksum(b, pseudo) == 0xffff
+}
+
+// buildTSOv4 builds a synthetic IPv4/TCP TSO superpacket with a payload of
+// `payLen` bytes split at `mss`.
+func buildTSOv4(t *testing.T, payLen, mss int) ([]byte, VirtioNetHdr) {
+	t.Helper()
+	const ipLen = 20
+	const tcpLen = 20
+	pkt := make([]byte, ipLen+tcpLen+payLen)
+
+	// IPv4 header
+	pkt[0] = 0x45 // version 4, IHL 5
+	// total length is meaningless for TSO but set it anyway
+	binary.BigEndian.PutUint16(pkt[2:4], uint16(ipLen+tcpLen+payLen))
+	binary.BigEndian.PutUint16(pkt[4:6], 0x4242) // original ID
+	pkt[8] = 64                                  // TTL
+	pkt[9] = unix.IPPROTO_TCP
+	copy(pkt[12:16], []byte{10, 0, 0, 1}) // src
+	copy(pkt[16:20], []byte{10, 0, 0, 2}) // dst
+
+	// TCP header
+	binary.BigEndian.PutUint16(pkt[20:22], 12345) // sport
+	binary.BigEndian.PutUint16(pkt[22:24], 80)    // dport
+	binary.BigEndian.PutUint32(pkt[24:28], 10000) // seq
+	binary.BigEndian.PutUint32(pkt[28:32], 20000) // ack
+	pkt[32] = 0x50                                // data offset 5 words
+	pkt[33] = 0x18                                // ACK | PSH
+	binary.BigEndian.PutUint16(pkt[34:36], 65535) // window
+
+	// payload
+	for i := 0; i < payLen; i++ {
+		pkt[ipLen+tcpLen+i] = byte(i & 0xff)
+	}
+
+	return pkt, VirtioNetHdr{
+		Flags:      unix.VIRTIO_NET_HDR_F_NEEDS_CSUM,
+		GSOType:    unix.VIRTIO_NET_HDR_GSO_TCPV4,
+		HdrLen:     uint16(ipLen + tcpLen),
+		GSOSize:    uint16(mss),
+		CsumStart:  uint16(ipLen),
+		CsumOffset: 16,
+	}
+}
+
+func TestSegmentTCPv4(t *testing.T) {
+	const mss = 100
+	const numSeg = 3
+	pkt, hdr := buildTSOv4(t, mss*numSeg, mss)
+
+	scratch := make([]byte, tunSegBufSize)
+	var out [][]byte
+	if err := segmentTCP(pkt, hdr, &out, scratch); err != nil {
+		t.Fatalf("segmentTCP: %v", err)
+	}
+	if len(out) != numSeg {
+		t.Fatalf("expected %d segments, got %d", numSeg, len(out))
+	}
+
+	for i, seg := range out {
+		if len(seg) != 40+mss {
+			t.Errorf("seg %d: unexpected len %d", i, len(seg))
+		}
+		totalLen := binary.BigEndian.Uint16(seg[2:4])
+		if totalLen != uint16(40+mss) {
+			t.Errorf("seg %d: total_len=%d want %d", i, totalLen, 40+mss)
+		}
+		id := binary.BigEndian.Uint16(seg[4:6])
+		if id != 0x4242+uint16(i) {
+			t.Errorf("seg %d: ip id=%#x want %#x", i, id, 0x4242+uint16(i))
+		}
+		seq := binary.BigEndian.Uint32(seg[24:28])
+		wantSeq := uint32(10000 + i*mss)
+		if seq != wantSeq {
+			t.Errorf("seg %d: seq=%d want %d", i, seq, wantSeq)
+		}
+		flags := seg[33]
+		wantFlags := byte(0x10) // ACK only, PSH cleared
+		if i == numSeg-1 {
+			wantFlags = 0x18 // ACK | PSH preserved on last
+		}
+		if flags != wantFlags {
+			t.Errorf("seg %d: flags=%#x want %#x", i, flags, wantFlags)
+		}
+		// IPv4 header checksum must verify against itself.
+		if !verifyChecksum(seg[:20], 0) {
+			t.Errorf("seg %d: bad IPv4 header checksum", i)
+		}
+		// TCP checksum must verify against the pseudo-header.
+		psum := pseudoHeaderIPv4(seg[12:16], seg[16:20], unix.IPPROTO_TCP, 20+mss)
+		if !verifyChecksum(seg[20:], psum) {
+			t.Errorf("seg %d: bad TCP checksum", i)
+		}
+	}
+}
+
+func TestSegmentTCPv4OddTail(t *testing.T) {
+	// Payload of 250 bytes with MSS 100 → segments of 100, 100, 50.
+	pkt, hdr := buildTSOv4(t, 250, 100)
+	scratch := make([]byte, tunSegBufSize)
+	var out [][]byte
+	if err := segmentTCP(pkt, hdr, &out, scratch); err != nil {
+		t.Fatalf("segmentTCP: %v", err)
+	}
+	if len(out) != 3 {
+		t.Fatalf("want 3 segments, got %d", len(out))
+	}
+	wantPayLens := []int{100, 100, 50}
+	for i, seg := range out {
+		if len(seg)-40 != wantPayLens[i] {
+			t.Errorf("seg %d: pay len %d want %d", i, len(seg)-40, wantPayLens[i])
+		}
+		if !verifyChecksum(seg[:20], 0) {
+			t.Errorf("seg %d: bad IPv4 header checksum", i)
+		}
+		psum := pseudoHeaderIPv4(seg[12:16], seg[16:20], unix.IPPROTO_TCP, 20+wantPayLens[i])
+		if !verifyChecksum(seg[20:], psum) {
+			t.Errorf("seg %d: bad TCP checksum", i)
+		}
+	}
+}
+
+func TestSegmentTCPv6(t *testing.T) {
+	const ipLen = 40
+	const tcpLen = 20
+	const mss = 120
+	const numSeg = 2
+	payLen := mss * numSeg
+	pkt := make([]byte, ipLen+tcpLen+payLen)
+
+	// IPv6 header
+	pkt[0] = 0x60 // version 6
+	binary.BigEndian.PutUint16(pkt[4:6], uint16(tcpLen+payLen))
+	pkt[6] = unix.IPPROTO_TCP
+	pkt[7] = 64
+	// src/dst fe80::1 / fe80::2
+	pkt[8] = 0xfe
+	pkt[9] = 0x80
+	pkt[23] = 1
+	pkt[24] = 0xfe
+	pkt[25] = 0x80
+	pkt[39] = 2
+
+	// TCP header
+	binary.BigEndian.PutUint16(pkt[40:42], 12345)
+	binary.BigEndian.PutUint16(pkt[42:44], 80)
+	binary.BigEndian.PutUint32(pkt[44:48], 7)
+	binary.BigEndian.PutUint32(pkt[48:52], 99)
+	pkt[52] = 0x50
+	pkt[53] = 0x19 // FIN | ACK | PSH — exercise FIN clearing too
+	binary.BigEndian.PutUint16(pkt[54:56], 65535)
+
+	for i := 0; i < payLen; i++ {
+		pkt[ipLen+tcpLen+i] = byte(i)
+	}
+
+	hdr := VirtioNetHdr{
+		Flags:      unix.VIRTIO_NET_HDR_F_NEEDS_CSUM,
+		GSOType:    unix.VIRTIO_NET_HDR_GSO_TCPV6,
+		HdrLen:     uint16(ipLen + tcpLen),
+		GSOSize:    uint16(mss),
+		CsumStart:  uint16(ipLen),
+		CsumOffset: 16,
+	}
+
+	scratch := make([]byte, tunSegBufSize)
+	var out [][]byte
+	if err := segmentTCP(pkt, hdr, &out, scratch); err != nil {
+		t.Fatalf("segmentTCP: %v", err)
+	}
+	if len(out) != numSeg {
+		t.Fatalf("want %d segments, got %d", numSeg, len(out))
+	}
+
+	for i, seg := range out {
+		if len(seg) != ipLen+tcpLen+mss {
+			t.Errorf("seg %d: len %d want %d", i, len(seg), ipLen+tcpLen+mss)
+		}
+		pl := binary.BigEndian.Uint16(seg[4:6])
+		if pl != uint16(tcpLen+mss) {
+			t.Errorf("seg %d: payload_length=%d want %d", i, pl, tcpLen+mss)
+		}
+		seq := binary.BigEndian.Uint32(seg[44:48])
+		if seq != uint32(7+i*mss) {
+			t.Errorf("seg %d: seq=%d want %d", i, seq, 7+i*mss)
+		}
+		flags := seg[53]
+		// Original flags = 0x19 (FIN|ACK|PSH). FIN(0x01)+PSH(0x08) should be
+		// cleared on all but the last; ACK(0x10) always preserved.
+		wantFlags := byte(0x10)
+		if i == numSeg-1 {
+			wantFlags = 0x19
+		}
+		if flags != wantFlags {
+			t.Errorf("seg %d: flags=%#x want %#x", i, flags, wantFlags)
+		}
+		psum := pseudoHeaderIPv6(seg[8:24], seg[24:40], unix.IPPROTO_TCP, tcpLen+mss)
+		if !verifyChecksum(seg[ipLen:], psum) {
+			t.Errorf("seg %d: bad TCP checksum", i)
+		}
+	}
+}
+
+func TestSegmentGSONonePassesThrough(t *testing.T) {
+	pkt, hdr := buildTSOv4(t, 100, 100)
+	hdr.GSOType = unix.VIRTIO_NET_HDR_GSO_NONE
+	hdr.Flags = 0 // no NEEDS_CSUM, leave packet untouched
+
+	scratch := make([]byte, tunSegBufSize)
+	var out [][]byte
+	if err := segmentInto(pkt, hdr, &out, scratch); err != nil {
+		t.Fatalf("segmentInto: %v", err)
+	}
+	if len(out) != 1 {
+		t.Fatalf("want 1 segment, got %d", len(out))
+	}
+	if len(out[0]) != len(pkt) {
+		t.Fatalf("unexpected length: %d vs %d", len(out[0]), len(pkt))
+	}
+}
+
+func TestSegmentRejectsUDP(t *testing.T) {
+	hdr := VirtioNetHdr{GSOType: unix.VIRTIO_NET_HDR_GSO_UDP}
+	var out [][]byte
+	if err := segmentInto(nil, hdr, &out, nil); err == nil {
+		t.Fatalf("expected rejection for UDP GSO")
+	}
+}
+
+func BenchmarkSegmentTCPv4(b *testing.B) {
+	sizes := []struct {
+		name   string
+		payLen int
+		mss    int
+	}{
+		{"64KiB_MSS1460", 65000, 1460},
+		{"16KiB_MSS1460", 16384, 1460},
+		{"4KiB_MSS1460", 4096, 1460},
+	}
+	for _, sz := range sizes {
+		b.Run(sz.name, func(b *testing.B) {
+			const ipLen = 20
+			const tcpLen = 20
+			pkt := make([]byte, ipLen+tcpLen+sz.payLen)
+			pkt[0] = 0x45
+			binary.BigEndian.PutUint16(pkt[2:4], uint16(ipLen+tcpLen+sz.payLen))
+			binary.BigEndian.PutUint16(pkt[4:6], 0x4242)
+			pkt[8] = 64
+			pkt[9] = unix.IPPROTO_TCP
+			copy(pkt[12:16], []byte{10, 0, 0, 1})
+			copy(pkt[16:20], []byte{10, 0, 0, 2})
+			binary.BigEndian.PutUint16(pkt[20:22], 12345)
+			binary.BigEndian.PutUint16(pkt[22:24], 80)
+			binary.BigEndian.PutUint32(pkt[24:28], 10000)
+			binary.BigEndian.PutUint32(pkt[28:32], 20000)
+			pkt[32] = 0x50
+			pkt[33] = 0x18
+			binary.BigEndian.PutUint16(pkt[34:36], 65535)
+			for i := 0; i < sz.payLen; i++ {
+				pkt[ipLen+tcpLen+i] = byte(i)
+			}
+			hdr := VirtioNetHdr{
+				Flags:      unix.VIRTIO_NET_HDR_F_NEEDS_CSUM,
+				GSOType:    unix.VIRTIO_NET_HDR_GSO_TCPV4,
+				HdrLen:     uint16(ipLen + tcpLen),
+				GSOSize:    uint16(sz.mss),
+				CsumStart:  uint16(ipLen),
+				CsumOffset: 16,
+			}
+
+			scratch := make([]byte, tunSegBufSize)
+			out := make([][]byte, 0, 64)
+
+			b.SetBytes(int64(len(pkt)))
+			b.ResetTimer()
+			for i := 0; i < b.N; i++ {
+				out = out[:0]
+				if err := segmentTCP(pkt, hdr, &out, scratch); err != nil {
+					b.Fatal(err)
+				}
+			}
+		})
+	}
+}
+
+// TestTunFileWriteVnetHdrNoAlloc verifies the IFF_VNET_HDR fast-path write is
+// allocation-free. We write to /dev/null so every call succeeds synchronously.
+func TestTunFileWriteVnetHdrNoAlloc(t *testing.T) {
+	fd, err := unix.Open("/dev/null", os.O_WRONLY, 0)
+	if err != nil {
+		t.Fatalf("open /dev/null: %v", err)
+	}
+	t.Cleanup(func() { _ = unix.Close(fd) })
+
+	tf := &Offload{fd: fd}
+
+	payload := make([]byte, 1400)
+	// Warm up (first call may trigger one-time internal allocations elsewhere).
+	if _, err := tf.Write(payload); err != nil {
+		t.Fatalf("Write: %v", err)
+	}
+
+	allocs := testing.AllocsPerRun(1000, func() {
+		if _, err := tf.Write(payload); err != nil {
+			t.Fatalf("Write: %v", err)
+		}
+	})
+	if allocs != 0 {
+		t.Fatalf("Write allocated %.1f times per call, want 0", allocs)
+	}
+}

overlay/tio/vnethdr_linux.go

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

overlay/tun_android.go

@@ -13,17 +13,38 @@ import (
 
 	"github.com/gaissmai/bart"
 	"github.com/slackhq/nebula/config"
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 	"github.com/slackhq/nebula/util"
 )
 
 type tun struct {
-	io.ReadWriteCloser
+	rwc         io.ReadWriteCloser
 	fd          int
 	vpnNetworks []netip.Prefix
 	Routes      atomic.Pointer[[]Route]
 	routeTree   atomic.Pointer[bart.Table[routing.Gateways]]
 	l           *slog.Logger
+
+	readBuf  []byte
+	batchRet [1][]byte
+}
+
+func (t *tun) Read() ([][]byte, error) {
+	n, err := t.rwc.Read(t.readBuf)
+	if err != nil {
+		return nil, err
+	}
+	t.batchRet[0] = t.readBuf[:n]
+	return t.batchRet[:], nil
+}
+
+func (t *tun) Write(p []byte) (int, error) {
+	return t.rwc.Write(p)
+}
+
+func (t *tun) Close() error {
+	return t.rwc.Close()
 }
 
 func newTunFromFd(c *config.C, l *slog.Logger, deviceFd int, vpnNetworks []netip.Prefix) (*tun, error) {
@@ -32,10 +53,11 @@ func newTunFromFd(c *config.C, l *slog.Logger, deviceFd int, vpnNetworks []netip
 	file := os.NewFile(uintptr(deviceFd), "/dev/net/tun")
 
 	t := &tun{
-		ReadWriteCloser: file,
+		rwc:         file,
 		fd:          deviceFd,
 		vpnNetworks: vpnNetworks,
 		l:           l,
+		readBuf:     make([]byte, defaultBatchBufSize),
 	}
 
 	err := t.reload(c, true)
@@ -62,7 +84,7 @@ func (t *tun) RoutesFor(ip netip.Addr) routing.Gateways {
 	return r
 }
 
-func (t tun) Activate() error {
+func (t *tun) Activate() error {
 	return nil
 }
 
@@ -95,18 +117,14 @@ func (t *tun) Name() string {
 	return "android"
 }
 
-func (t *tun) SupportsPerPeerMTU() bool {
-	return false
-}
-
-func (t *tun) SetPeerMTU(addr netip.Addr, mtu int) error {
-	return nil
-}
-
 func (t *tun) SupportsMultiqueue() bool {
 	return false
 }
 
-func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+func (t *tun) NewMultiQueueReader() error {
-	return nil, fmt.Errorf("TODO: multiqueue not implemented for android")
+	return fmt.Errorf("TODO: multiqueue not implemented for android")
+}
+
+func (t *tun) Readers() []tio.Queue {
+	return []tio.Queue{t}
 }

overlay/tun_darwin.go

@@ -16,6 +16,7 @@ import (
 
 	"github.com/gaissmai/bart"
 	"github.com/slackhq/nebula/config"
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 	"github.com/slackhq/nebula/util"
 	netroute "golang.org/x/net/route"
@@ -23,7 +24,7 @@ import (
 )
 
 type tun struct {
-	io.ReadWriteCloser
+	rwc         io.ReadWriteCloser
 	Device      string
 	vpnNetworks []netip.Prefix
 	DefaultMTU  int
@@ -34,6 +35,9 @@ type tun struct {
 
 	// cache out buffer since we need to prepend 4 bytes for tun metadata
 	out []byte
+
+	readBuf  []byte
+	batchRet [1][]byte
 }
 
 type ifReq struct {
@@ -124,11 +128,12 @@ func newTun(c *config.C, l *slog.Logger, vpnNetworks []netip.Prefix, _ bool) (*t
 	}
 
 	t := &tun{
-		ReadWriteCloser: os.NewFile(uintptr(fd), ""),
+		rwc:         os.NewFile(uintptr(fd), ""),
 		Device:      name,
 		vpnNetworks: vpnNetworks,
 		DefaultMTU:  c.GetInt("tun.mtu", DefaultMTU),
 		l:           l,
+		readBuf:     make([]byte, defaultBatchBufSize),
 	}
 
 	err = t.reload(c, true)
@@ -158,8 +163,8 @@ func newTunFromFd(_ *config.C, _ *slog.Logger, _ int, _ []netip.Prefix) (*tun, e
 }
 
 func (t *tun) Close() error {
-	if t.ReadWriteCloser != nil {
+	if t.rwc != nil {
-		return t.ReadWriteCloser.Close()
+		return t.rwc.Close()
 	}
 	return nil
 }
@@ -502,15 +507,24 @@ func delRoute(prefix netip.Prefix, gateway netroute.Addr) error {
 	return nil
 }
 
-func (t *tun) Read(to []byte) (int, error) {
+func (t *tun) readOne(to []byte) (int, error) {
 	buf := make([]byte, len(to)+4)
 
-	n, err := t.ReadWriteCloser.Read(buf)
+	n, err := t.rwc.Read(buf)
 
 	copy(to, buf[4:])
 	return n - 4, err
 }
 
+func (t *tun) Read() ([][]byte, error) {
+	n, err := t.readOne(t.readBuf)
+	if err != nil {
+		return nil, err
+	}
+	t.batchRet[0] = t.readBuf[:n]
+	return t.batchRet[:], nil
+}
+
 // Write is only valid for single threaded use
 func (t *tun) Write(from []byte) (int, error) {
 	buf := t.out
@@ -536,7 +550,7 @@ func (t *tun) Write(from []byte) (int, error) {
 
 	copy(buf[4:], from)
 
-	n, err := t.ReadWriteCloser.Write(buf)
+	n, err := t.rwc.Write(buf)
 	return n - 4, err
 }
 
@@ -548,18 +562,14 @@ func (t *tun) Name() string {
 	return t.Device
 }
 
-func (t *tun) SupportsPerPeerMTU() bool {
-	return false
-}
-
-func (t *tun) SetPeerMTU(addr netip.Addr, mtu int) error {
-	return nil
-}
-
 func (t *tun) SupportsMultiqueue() bool {
 	return false
 }
 
-func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+func (t *tun) NewMultiQueueReader() error {
-	return nil, fmt.Errorf("TODO: multiqueue not implemented for darwin")
+	return fmt.Errorf("TODO: multiqueue not implemented for darwin")
+}
+
+func (t *tun) Readers() []tio.Queue {
+	return []tio.Queue{t}
 }

overlay/tun_disabled.go

@@ -10,6 +10,7 @@ import (
 
 	"github.com/rcrowley/go-metrics"
 	"github.com/slackhq/nebula/iputil"
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 )
 
@@ -21,6 +22,24 @@ type disabledTun struct {
 	tx         metrics.Counter
 	rx         metrics.Counter
 	l          *slog.Logger
+	numReaders int
+
+	batchRet [1][]byte
+}
+
+func (t *disabledTun) Read() ([][]byte, error) {
+	r, ok := <-t.read
+	if !ok {
+		return nil, io.EOF
+	}
+
+	t.tx.Inc(1)
+	if t.l.Enabled(context.Background(), slog.LevelDebug) {
+		t.l.Debug("Write payload", "raw", prettyPacket(r))
+	}
+
+	t.batchRet[0] = r
+	return t.batchRet[:], nil
 }
 
 func newDisabledTun(vpnNetworks []netip.Prefix, queueLen int, metricsEnabled bool, l *slog.Logger) *disabledTun {
@@ -28,6 +47,7 @@ func newDisabledTun(vpnNetworks []netip.Prefix, queueLen int, metricsEnabled boo
 		vpnNetworks: vpnNetworks,
 		read:        make(chan []byte, queueLen),
 		l:           l,
+		numReaders:  1,
 	}
 
 	if metricsEnabled {
@@ -57,24 +77,6 @@ func (*disabledTun) Name() string {
 	return "disabled"
 }
 
-func (t *disabledTun) Read(b []byte) (int, error) {
-	r, ok := <-t.read
-	if !ok {
-		return 0, io.EOF
-	}
-
-	if len(r) > len(b) {
-		return 0, fmt.Errorf("packet larger than mtu: %d > %d bytes", len(r), len(b))
-	}
-
-	t.tx.Inc(1)
-	if t.l.Enabled(context.Background(), slog.LevelDebug) {
-		t.l.Debug("Write payload", "raw", prettyPacket(r))
-	}
-
-	return copy(b, r), nil
-}
-
 func (t *disabledTun) handleICMPEchoRequest(b []byte) bool {
 	out := make([]byte, len(b))
 	out = iputil.CreateICMPEchoResponse(b, out)
@@ -106,20 +108,21 @@ func (t *disabledTun) Write(b []byte) (int, error) {
 	return len(b), nil
 }
 
-func (t *disabledTun) SupportsPerPeerMTU() bool {
-	return false
-}
-
-func (t *disabledTun) SetPeerMTU(addr netip.Addr, mtu int) error {
-	return nil
-}
-
 func (t *disabledTun) SupportsMultiqueue() bool {
 	return true
 }
 
-func (t *disabledTun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+func (t *disabledTun) NewMultiQueueReader() error {
-	return t, nil
+	t.numReaders++
+	return nil
+}
+
+func (t *disabledTun) Readers() []tio.Queue {
+	out := make([]tio.Queue, t.numReaders)
+	for i := range t.numReaders {
+		out[i] = t
+	}
+	return out
 }
 
 func (t *disabledTun) Close() error {

overlay/tun_file_linux_test.go

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

overlay/tun_freebsd.go

@@ -7,7 +7,6 @@ import (
 	"bytes"
 	"errors"
 	"fmt"
-	"io"
 	"io/fs"
 	"log/slog"
 	"net/netip"
@@ -20,7 +19,7 @@ import (
 	"github.com/gaissmai/bart"
 
 	"github.com/slackhq/nebula/config"
-
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 	"github.com/slackhq/nebula/util"
 	netroute "golang.org/x/net/route"
@@ -103,6 +102,9 @@ type tun struct {
 	readPoll  [2]unix.PollFd
 	writePoll [2]unix.PollFd
 	closed    atomic.Bool
+
+	readBuf  []byte
+	batchRet [1][]byte
 }
 
 // blockOnRead waits until the tun fd is readable or shutdown has been signaled.
@@ -157,7 +159,16 @@ func (t *tun) blockOnWrite() error {
 	return nil
 }
 
-func (t *tun) Read(to []byte) (int, error) {
+func (t *tun) Read() ([][]byte, error) {
+	n, err := t.readOne(t.readBuf)
+	if err != nil {
+		return nil, err
+	}
+	t.batchRet[0] = t.readBuf[:n]
+	return t.batchRet[:], nil
+}
+
+func (t *tun) readOne(to []byte) (int, error) {
 	// first 4 bytes is protocol family, in network byte order
 	var head [4]byte
 	iovecs := [2]syscall.Iovec{
@@ -375,6 +386,7 @@ func newTun(c *config.C, l *slog.Logger, vpnNetworks []netip.Prefix, _ bool) (*t
 		MTU:         c.GetInt("tun.mtu", DefaultMTU),
 		l:           l,
 		fd:          fd,
+		readBuf:     make([]byte, defaultBatchBufSize),
 		shutdownR:   shutdownR,
 		shutdownW:   shutdownW,
 		readPoll: [2]unix.PollFd{
@@ -561,20 +573,12 @@ func (t *tun) Name() string {
 	return t.Device
 }
 
-func (t *tun) SupportsPerPeerMTU() bool {
-	return false
-}
-
-func (t *tun) SetPeerMTU(addr netip.Addr, mtu int) error {
-	return nil
-}
-
 func (t *tun) SupportsMultiqueue() bool {
 	return false
 }
 
-func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+func (t *tun) NewMultiQueueReader() error {
-	return nil, fmt.Errorf("TODO: multiqueue not implemented for freebsd")
+	return fmt.Errorf("TODO: multiqueue not implemented for freebsd")
 }
 
 func (t *tun) addRoutes(logErrors bool) error {
@@ -601,6 +605,10 @@ func (t *tun) addRoutes(logErrors bool) error {
 	return nil
 }
 
+func (t *tun) Readers() []tio.Queue {
+	return []tio.Queue{t}
+}
+
 func (t *tun) removeRoutes(routes []Route) error {
 	for _, r := range routes {
 		if !r.Install {

overlay/tun_ios.go

@@ -16,16 +16,37 @@ import (
 
 	"github.com/gaissmai/bart"
 	"github.com/slackhq/nebula/config"
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 	"github.com/slackhq/nebula/util"
 )
 
 type tun struct {
-	io.ReadWriteCloser
+	rwc         io.ReadWriteCloser
 	vpnNetworks []netip.Prefix
 	Routes      atomic.Pointer[[]Route]
 	routeTree   atomic.Pointer[bart.Table[routing.Gateways]]
 	l           *slog.Logger
+
+	readBuf  []byte
+	batchRet [1][]byte
+}
+
+func (t *tun) Read() ([][]byte, error) {
+	n, err := t.rwc.Read(t.readBuf)
+	if err != nil {
+		return nil, err
+	}
+	t.batchRet[0] = t.readBuf[:n]
+	return t.batchRet[:], nil
+}
+
+func (t *tun) Write(p []byte) (int, error) {
+	return t.rwc.Write(p)
+}
+
+func (t *tun) Close() error {
+	return t.rwc.Close()
 }
 
 func newTun(_ *config.C, _ *slog.Logger, _ []netip.Prefix, _ bool) (*tun, error) {
@@ -36,8 +57,9 @@ func newTunFromFd(c *config.C, l *slog.Logger, deviceFd int, vpnNetworks []netip
 	file := os.NewFile(uintptr(deviceFd), "/dev/tun")
 	t := &tun{
 		vpnNetworks: vpnNetworks,
-		ReadWriteCloser: &tunReadCloser{f: file},
+		rwc:         &tunReadCloser{f: file},
 		l:           l,
+		readBuf:     make([]byte, defaultBatchBufSize),
 	}
 
 	err := t.reload(c, true)
@@ -151,18 +173,14 @@ func (t *tun) Name() string {
 	return "iOS"
 }
 
-func (t *tun) SupportsPerPeerMTU() bool {
-	return false
-}
-
-func (t *tun) SetPeerMTU(addr netip.Addr, mtu int) error {
-	return nil
-}
-
 func (t *tun) SupportsMultiqueue() bool {
 	return false
 }
 
-func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+func (t *tun) NewMultiQueueReader() error {
-	return nil, fmt.Errorf("TODO: multiqueue not implemented for ios")
+	return fmt.Errorf("TODO: multiqueue not implemented for ios")
+}
+
+func (t *tun) Readers() []tio.Queue {
+	return []tio.Queue{t}
 }

overlay/tun_linux.go

@@ -4,9 +4,7 @@
 package overlay
 
 import (
-	"encoding/binary"
 	"fmt"
-	"io"
 	"log/slog"
 	"net"
 	"net/netip"
@@ -19,180 +17,15 @@ import (
 
 	"github.com/gaissmai/bart"
 	"github.com/slackhq/nebula/config"
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 	"github.com/slackhq/nebula/util"
 	"github.com/vishvananda/netlink"
 	"golang.org/x/sys/unix"
 )
 
-// tunFile wraps a TUN file descriptor with poll-based reads. The FD provided will be changed to non-blocking.
-// A shared eventfd allows Close to wake all readers blocked in poll.
-type tunFile struct {
-	fd         int
-	shutdownFd int
-	lastOne    bool
-	readPoll   [2]unix.PollFd
-	writePoll  [2]unix.PollFd
-	closed     bool
-}
-
-// newFriend makes a tunFile for a MultiQueueReader that copies the shutdown eventfd from the parent tun
-func (r *tunFile) newFriend(fd int) (*tunFile, error) {
-	if err := unix.SetNonblock(fd, true); err != nil {
-		return nil, fmt.Errorf("failed to set tun fd non-blocking: %w", err)
-	}
-	return &tunFile{
-		fd:         fd,
-		shutdownFd: r.shutdownFd,
-		readPoll: [2]unix.PollFd{
-			{Fd: int32(fd), Events: unix.POLLIN},
-			{Fd: int32(r.shutdownFd), Events: unix.POLLIN},
-		},
-		writePoll: [2]unix.PollFd{
-			{Fd: int32(fd), Events: unix.POLLOUT},
-			{Fd: int32(r.shutdownFd), Events: unix.POLLIN},
-		},
-	}, nil
-}
-
-func newTunFd(fd int) (*tunFile, error) {
-	if err := unix.SetNonblock(fd, true); err != nil {
-		return nil, fmt.Errorf("failed to set tun fd non-blocking: %w", err)
-	}
-
-	shutdownFd, err := unix.Eventfd(0, unix.EFD_NONBLOCK|unix.EFD_CLOEXEC)
-	if err != nil {
-		return nil, fmt.Errorf("failed to create eventfd: %w", err)
-	}
-
-	out := &tunFile{
-		fd:         fd,
-		shutdownFd: shutdownFd,
-		lastOne:    true,
-		readPoll: [2]unix.PollFd{
-			{Fd: int32(fd), Events: unix.POLLIN},
-			{Fd: int32(shutdownFd), Events: unix.POLLIN},
-		},
-		writePoll: [2]unix.PollFd{
-			{Fd: int32(fd), Events: unix.POLLOUT},
-			{Fd: int32(shutdownFd), Events: unix.POLLIN},
-		},
-	}
-
-	return out, nil
-}
-
-func (r *tunFile) blockOnRead() error {
-	const problemFlags = unix.POLLHUP | unix.POLLNVAL | unix.POLLERR
-	var err error
-	for {
-		_, err = unix.Poll(r.readPoll[:], -1)
-		if err != unix.EINTR {
-			break
-		}
-	}
-	//always reset these!
-	tunEvents := r.readPoll[0].Revents
-	shutdownEvents := r.readPoll[1].Revents
-	r.readPoll[0].Revents = 0
-	r.readPoll[1].Revents = 0
-	//do the err check before trusting the potentially bogus bits we just got
-	if err != nil {
-		return err
-	}
-	if shutdownEvents&(unix.POLLIN|problemFlags) != 0 {
-		return os.ErrClosed
-	} else if tunEvents&problemFlags != 0 {
-		return os.ErrClosed
-	}
-	return nil
-}
-
-func (r *tunFile) blockOnWrite() error {
-	const problemFlags = unix.POLLHUP | unix.POLLNVAL | unix.POLLERR
-	var err error
-	for {
-		_, err = unix.Poll(r.writePoll[:], -1)
-		if err != unix.EINTR {
-			break
-		}
-	}
-	//always reset these!
-	tunEvents := r.writePoll[0].Revents
-	shutdownEvents := r.writePoll[1].Revents
-	r.writePoll[0].Revents = 0
-	r.writePoll[1].Revents = 0
-	//do the err check before trusting the potentially bogus bits we just got
-	if err != nil {
-		return err
-	}
-	if shutdownEvents&(unix.POLLIN|problemFlags) != 0 {
-		return os.ErrClosed
-	} else if tunEvents&problemFlags != 0 {
-		return os.ErrClosed
-	}
-	return nil
-}
-
-func (r *tunFile) Read(buf []byte) (int, error) {
-	for {
-		if n, err := unix.Read(r.fd, buf); err == nil {
-			return n, nil
-		} else if err == unix.EAGAIN {
-			if err = r.blockOnRead(); err != nil {
-				return 0, err
-			}
-			continue
-		} else if err == unix.EINTR {
-			continue
-		} else if err == unix.EBADF {
-			return 0, os.ErrClosed
-		} else {
-			return 0, err
-		}
-	}
-}
-
-func (r *tunFile) Write(buf []byte) (int, error) {
-	for {
-		if n, err := unix.Write(r.fd, buf); err == nil {
-			return n, nil
-		} else if err == unix.EAGAIN {
-			if err = r.blockOnWrite(); err != nil {
-				return 0, err
-			}
-			continue
-		} else if err == unix.EINTR {
-			continue
-		} else if err == unix.EBADF {
-			return 0, os.ErrClosed
-		} else {
-			return 0, err
-		}
-	}
-}
-
-func (r *tunFile) wakeForShutdown() error {
-	var buf [8]byte
-	binary.NativeEndian.PutUint64(buf[:], 1)
-	_, err := unix.Write(int(r.readPoll[1].Fd), buf[:])
-	return err
-}
-
-func (r *tunFile) Close() error {
-	if r.closed { // avoid closing more than once. Technically a fd could get re-used, which would be a problem
-		return nil
-	}
-	r.closed = true
-	if r.lastOne {
-		_ = unix.Close(r.shutdownFd)
-	}
-	return unix.Close(r.fd)
-}
-
 type tun struct {
-	*tunFile
+	readers     tio.Container
-	readers     []*tunFile
 	closeLock   sync.Mutex
 	Device      string
 	vpnNetworks []netip.Prefix
@@ -201,6 +34,7 @@ type tun struct {
 	TXQueueLen  int
 	deviceIndex int
 	ioctlFd     uintptr
+	vnetHdr     bool
 
 	Routes                    atomic.Pointer[[]Route]
 	routeTree                 atomic.Pointer[bart.Table[routing.Gateways]]
@@ -239,7 +73,9 @@ type ifreqQLEN struct {
 }
 
 func newTunFromFd(c *config.C, l *slog.Logger, deviceFd int, vpnNetworks []netip.Prefix) (*tun, error) {
-	t, err := newTunGeneric(c, l, deviceFd, vpnNetworks)
+	// We don't know what flags the caller opened this fd with and can't turn
+	// on IFF_VNET_HDR after TUNSETIFF, so skip offload on inherited fds.
+	t, err := newTunGeneric(c, l, deviceFd, false, vpnNetworks)
 	if err != nil {
 		return nil, err
 	}
@@ -249,46 +85,83 @@ func newTunFromFd(c *config.C, l *slog.Logger, deviceFd int, vpnNetworks []netip
 	return t, nil
 }
 
-func newTun(c *config.C, l *slog.Logger, vpnNetworks []netip.Prefix, multiqueue bool) (*tun, error) {
+// openTunDev opens /dev/net/tun, creating the device node first if it's
+// missing (docker containers occasionally omit it).
+func openTunDev() (int, error) {
 	fd, err := unix.Open("/dev/net/tun", os.O_RDWR, 0)
-	if err != nil {
+	if err == nil {
-		// If /dev/net/tun doesn't exist, try to create it (will happen in docker)
+		return fd, nil
-		if os.IsNotExist(err) {
-			err = os.MkdirAll("/dev/net", 0755)
-			if err != nil {
-				return nil, fmt.Errorf("/dev/net/tun doesn't exist, failed to mkdir -p /dev/net: %w", err)
 	}
-			err = unix.Mknod("/dev/net/tun", unix.S_IFCHR|0600, int(unix.Mkdev(10, 200)))
+	if !os.IsNotExist(err) {
-			if err != nil {
+		return -1, err
-				return nil, fmt.Errorf("failed to create /dev/net/tun: %w", err)
+	}
+	if err = os.MkdirAll("/dev/net", 0755); err != nil {
+		return -1, fmt.Errorf("/dev/net/tun doesn't exist, failed to mkdir -p /dev/net: %w", err)
+	}
+	if err = unix.Mknod("/dev/net/tun", unix.S_IFCHR|0600, int(unix.Mkdev(10, 200))); err != nil {
+		return -1, fmt.Errorf("failed to create /dev/net/tun: %w", err)
 	}
-
 	fd, err = unix.Open("/dev/net/tun", os.O_RDWR, 0)
 	if err != nil {
-				return nil, fmt.Errorf("created /dev/net/tun, but still failed: %w", err)
+		return -1, fmt.Errorf("created /dev/net/tun, but still failed: %w", err)
-			}
-		} else {
-			return nil, err
 	}
+	return fd, nil
 }
 
+// tunSetIff runs TUNSETIFF with the given flags and returns the kernel-chosen
+// device name on success.
+func tunSetIff(fd int, name string, flags uint16) (string, error) {
 	var req ifReq
-	req.Flags = uint16(unix.IFF_TUN | unix.IFF_NO_PI)
+	req.Flags = flags
+	copy(req.Name[:], name)
+	if err := ioctl(uintptr(fd), uintptr(unix.TUNSETIFF), uintptr(unsafe.Pointer(&req))); err != nil {
+		return "", err
+	}
+	return strings.Trim(string(req.Name[:]), "\x00"), nil
+}
+
+// tsoOffloadFlags are the TUN_F_* bits we ask the kernel to enable when a
+// TSO-capable TUN is available. CSUM is required as a prerequisite for TSO.
+const tsoOffloadFlags = unix.TUN_F_CSUM | unix.TUN_F_TSO4 | unix.TUN_F_TSO6
+
+func newTun(c *config.C, l *slog.Logger, vpnNetworks []netip.Prefix, multiqueue bool) (*tun, error) {
+	baseFlags := uint16(unix.IFF_TUN | unix.IFF_NO_PI)
 	if multiqueue {
-		req.Flags |= unix.IFF_MULTI_QUEUE
+		baseFlags |= unix.IFF_MULTI_QUEUE
 	}
 	nameStr := c.GetString("tun.dev", "")
-	copy(req.Name[:], nameStr)
-	if err = ioctl(uintptr(fd), uintptr(unix.TUNSETIFF), uintptr(unsafe.Pointer(&req))); err != nil {
-		_ = unix.Close(fd)
-		return nil, &NameError{
-			Name:       nameStr,
-			Underlying: err,
-		}
-	}
-	name := strings.Trim(string(req.Name[:]), "\x00")
 
-	t, err := newTunGeneric(c, l, fd, vpnNetworks)
+	// First try to open with IFF_VNET_HDR + TUNSETOFFLOAD so we can receive
+	// TSO superpackets. If either step fails (older kernel, unprivileged
+	// container, etc.) we close and fall back to a plain TUN.
+	fd, err := openTunDev()
+	if err != nil {
+		return nil, err
+	}
+	vnetHdr := true
+	name, err := tunSetIff(fd, nameStr, baseFlags|unix.IFF_VNET_HDR)
+	if err != nil {
+		_ = unix.Close(fd)
+		vnetHdr = false
+	} else if err = ioctl(uintptr(fd), unix.TUNSETOFFLOAD, uintptr(tsoOffloadFlags)); err != nil {
+		l.Warn("Failed to enable TUN offload (TSO); proceeding without virtio headers", "error", err)
+		_ = unix.Close(fd)
+		vnetHdr = false
+	}
+
+	if !vnetHdr {
+		fd, err = openTunDev()
+		if err != nil {
+			return nil, err
+		}
+		name, err = tunSetIff(fd, nameStr, baseFlags)
+		if err != nil {
+			_ = unix.Close(fd)
+			return nil, &NameError{Name: nameStr, Underlying: err}
+		}
+	}
+
+	t, err := newTunGeneric(c, l, fd, vnetHdr, vpnNetworks)
 	if err != nil {
 		return nil, err
 	}
@@ -299,16 +172,29 @@ func newTun(c *config.C, l *slog.Logger, vpnNetworks []netip.Prefix, multiqueue
 }
 
 // newTunGeneric does all the stuff common to different tun initialization paths. It will close your files on error.
-func newTunGeneric(c *config.C, l *slog.Logger, fd int, vpnNetworks []netip.Prefix) (*tun, error) {
+func newTunGeneric(c *config.C, l *slog.Logger, fd int, vnetHdr bool, vpnNetworks []netip.Prefix) (*tun, error) {
-	tfd, err := newTunFd(fd)
+	var container tio.Container
+	var err error
+	if vnetHdr {
+		container, err = tio.NewOffloadContainer()
+	} else {
+		container, err = tio.NewPollContainer()
+	}
+
 	if err != nil {
 		_ = unix.Close(fd)
 		return nil, err
 	}
+	err = container.Add(fd)
+	if err != nil {
+		_ = unix.Close(fd)
+		return nil, err
+	}
+
 	t := &tun{
-		tunFile:                   tfd,
+		readers:                   container,
-		readers:                   []*tunFile{tfd},
 		closeLock:                 sync.Mutex{},
+		vnetHdr:                   vnetHdr,
 		vpnNetworks:               vpnNetworks,
 		TXQueueLen:                c.GetInt("tun.tx_queue", 500),
 		useSystemRoutes:           c.GetBool("tun.use_system_route_table", false),
@@ -368,13 +254,6 @@ func (t *tun) reload(c *config.C, initial bool) error {
 		}
 	}
 
-	// tun.max_mtu raises the device MTU above tun.mtu so PMTUD has headroom to
-	// install per-peer routes between tun.mtu (floor) and tun.max_mtu (ceiling).
-	// When unset (default 0) the device MTU is unchanged from existing behavior.
-	if pmtudCeiling := c.GetInt("tun.max_mtu", 0); pmtudCeiling > newMaxMTU {
-		newMaxMTU = pmtudCeiling
-	}
-
 	t.MaxMTU = newMaxMTU
 	t.DefaultMTU = newDefaultMTU
 
@@ -417,32 +296,38 @@ func (t *tun) SupportsMultiqueue() bool {
 	return true
 }
 
-func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+func (t *tun) NewMultiQueueReader() error {
 	t.closeLock.Lock()
 	defer t.closeLock.Unlock()
 
 	fd, err := unix.Open("/dev/net/tun", os.O_RDWR, 0)
 	if err != nil {
-		return nil, err
+		return err
 	}
 
-	var req ifReq
+	flags := uint16(unix.IFF_TUN | unix.IFF_NO_PI | unix.IFF_MULTI_QUEUE)
-	req.Flags = uint16(unix.IFF_TUN | unix.IFF_NO_PI | unix.IFF_MULTI_QUEUE)
+	if t.vnetHdr {
-	copy(req.Name[:], t.Device)
+		flags |= unix.IFF_VNET_HDR
-	if err = ioctl(uintptr(fd), uintptr(unix.TUNSETIFF), uintptr(unsafe.Pointer(&req))); err != nil {
+	}
+	if _, err = tunSetIff(fd, t.Device, flags); err != nil {
 		_ = unix.Close(fd)
-		return nil, err
+		return err
 	}
 
-	out, err := t.tunFile.newFriend(fd)
+	if t.vnetHdr {
+		if err = ioctl(uintptr(fd), unix.TUNSETOFFLOAD, uintptr(tsoOffloadFlags)); err != nil {
+			_ = unix.Close(fd)
+			return fmt.Errorf("failed to enable offload on multiqueue tun fd: %w", err)
+		}
+	}
+
+	err = t.readers.Add(fd)
 	if err != nil {
 		_ = unix.Close(fd)
-		return nil, err
+		return err
 	}
 
-	t.readers = append(t.readers, out)
+	return nil
-
-	return out, nil
 }
 
 func (t *tun) RoutesFor(ip netip.Addr) routing.Gateways {
@@ -603,7 +488,7 @@ func (t *tun) setDefaultRoute(cidr netip.Prefix) error {
 		LinkIndex: t.deviceIndex,
 		Dst:       dr,
 		MTU:       t.DefaultMTU,
-		AdvMSS:    t.advMSS(Route{Cidr: cidr}),
+		AdvMSS:    t.advMSS(Route{}),
 		Scope:     unix.RT_SCOPE_LINK,
 		Src:       net.IP(cidr.Addr().AsSlice()),
 		Protocol:  unix.RTPROT_KERNEL,
@@ -712,69 +597,18 @@ func (t *tun) Name() string {
 	return t.Device
 }
 
-func (t *tun) SupportsPerPeerMTU() bool {
-	return true
-}
-
-// SetPeerMTU installs a host route (/32 for an IPv4 vpn address, /128 for an IPv6
-// vpn address) to addr through this tun device with the given MTU. This causes
-// the kernel to reject (or surface PTB to apps for) inside packets to addr that
-// would exceed mtu. Pass mtu=0 to remove the override and let the per-vpn-network
-// route apply again. PoC: assumes addr is reachable directly via this device.
-func (t *tun) SetPeerMTU(addr netip.Addr, mtu int) error {
-	bits := addr.BitLen()
-	prefix := netip.PrefixFrom(addr, bits)
-
-	dr := &net.IPNet{
-		IP:   addr.AsSlice(),
-		Mask: net.CIDRMask(bits, bits),
-	}
-
-	if mtu == 0 {
-		nr := netlink.Route{
-			LinkIndex: t.deviceIndex,
-			Dst:       dr,
-			Scope:     unix.RT_SCOPE_LINK,
-		}
-		if err := netlink.RouteDel(&nr); err != nil {
-			return fmt.Errorf("failed to remove per-peer mtu route %v: %w", prefix, err)
-		}
-		return nil
-	}
-
-	nr := netlink.Route{
-		LinkIndex: t.deviceIndex,
-		Dst:       dr,
-		MTU:       mtu,
-		AdvMSS:    t.advMSS(Route{Cidr: prefix, MTU: mtu}),
-		Scope:     unix.RT_SCOPE_LINK,
-	}
-	if err := netlink.RouteReplace(&nr); err != nil {
-		return fmt.Errorf("failed to set per-peer mtu route %v mtu=%d: %w", prefix, mtu, err)
-	}
-	return nil
-}
-
 func (t *tun) advMSS(r Route) int {
 	mtu := r.MTU
 	if r.MTU == 0 {
 		mtu = t.DefaultMTU
 	}
 
-	// We only need to set advmss if the route MTU does not match the device MTU.
+	// We only need to set advmss if the route MTU does not match the device MTU
-	if mtu == t.MaxMTU {
+	if mtu != t.MaxMTU {
-		return 0
-	}
-
-	// MSS = MTU - (IP header + TCP header). TCP is always 20 bytes; IP is 20 for
-	// v4 and 40 for v6. r.Cidr is the route destination so it tells us which
-	// family this route is in. If Cidr is unset (empty Route) we default to v4.
-	addr := r.Cidr.Addr()
-	if addr.Is6() && !addr.Is4In6() {
-		return mtu - 60
-	}
 		return mtu - 40
 	}
+	return 0
+}
 
 func (t *tun) watchRoutes() {
 	rch := make(chan netlink.RouteUpdate)
@@ -927,6 +761,10 @@ func (t *tun) updateRoutes(r netlink.RouteUpdate) {
 	t.routeTree.Store(newTree)
 }
 
+func (t *tun) Readers() []tio.Queue {
+	return t.readers.Queues()
+}
+
 func (t *tun) Close() error {
 	t.closeLock.Lock()
 	defer t.closeLock.Unlock()
@@ -936,32 +774,10 @@ func (t *tun) Close() error {
 		t.routeChan = nil
 	}
 
-	// Signal all readers blocked in poll to wake up and exit
-	_ = t.tunFile.wakeForShutdown()
-
 	if t.ioctlFd > 0 {
 		_ = unix.Close(int(t.ioctlFd))
 		t.ioctlFd = 0
 	}
 
-	for i := range t.readers {
+	return t.readers.Close()
-		if i == 0 {
-			continue //we want to close the zeroth reader last
-		}
-		err := t.readers[i].Close()
-		if err != nil {
-			t.l.Error("error closing tun reader", "reader", i, "error", err)
-		} else {
-			t.l.Info("closed tun reader", "reader", i)
-		}
-	}
-
-	//this is t.readers[0] too
-	err := t.tunFile.Close()
-	if err != nil {
-		t.l.Error("error closing tun reader", "reader", 0, "error", err)
-	} else {
-		t.l.Info("closed tun reader", "reader", 0)
-	}
-	return err
 }

overlay/tun_linux_test.go

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

overlay/tun_netbsd.go

@@ -6,7 +6,6 @@ package overlay
 import (
 	"errors"
 	"fmt"
-	"io"
 	"log/slog"
 	"net/netip"
 	"os"
@@ -17,6 +16,7 @@ import (
 
 	"github.com/gaissmai/bart"
 	"github.com/slackhq/nebula/config"
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 	"github.com/slackhq/nebula/util"
 	netroute "golang.org/x/net/route"
@@ -66,6 +66,22 @@ type tun struct {
 	l           *slog.Logger
 	f           *os.File
 	fd          int
+
+	readBuf  []byte
+	batchRet [1][]byte
+}
+
+func (t *tun) Read() ([][]byte, error) {
+	n, err := t.readOne(t.readBuf)
+	if err != nil {
+		return nil, err
+	}
+	t.batchRet[0] = t.readBuf[:n]
+	return t.batchRet[:], nil
+}
+
+func (t *tun) Readers() []tio.Queue {
+	return []tio.Queue{t}
 }
 
 var deviceNameRE = regexp.MustCompile(`^tun[0-9]+$`)
@@ -102,6 +118,7 @@ func newTun(c *config.C, l *slog.Logger, vpnNetworks []netip.Prefix, _ bool) (*t
 		vpnNetworks: vpnNetworks,
 		MTU:         c.GetInt("tun.mtu", DefaultMTU),
 		l:           l,
+		readBuf:     make([]byte, defaultBatchBufSize),
 	}
 
 	err = t.reload(c, true)
@@ -141,7 +158,7 @@ func (t *tun) Close() error {
 	return nil
 }
 
-func (t *tun) Read(to []byte) (int, error) {
+func (t *tun) readOne(to []byte) (int, error) {
 	rc, err := t.f.SyscallConn()
 	if err != nil {
 		return 0, fmt.Errorf("failed to get syscall conn for tun: %w", err)
@@ -390,20 +407,12 @@ func (t *tun) Name() string {
 	return t.Device
 }
 
-func (t *tun) SupportsPerPeerMTU() bool {
-	return false
-}
-
-func (t *tun) SetPeerMTU(addr netip.Addr, mtu int) error {
-	return nil
-}
-
 func (t *tun) SupportsMultiqueue() bool {
 	return false
 }
 
-func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+func (t *tun) NewMultiQueueReader() error {
-	return nil, fmt.Errorf("TODO: multiqueue not implemented for netbsd")
+	return fmt.Errorf("TODO: multiqueue not implemented for netbsd")
 }
 
 func (t *tun) addRoutes(logErrors bool) error {

overlay/tun_openbsd.go

@@ -6,7 +6,6 @@ package overlay
 import (
 	"errors"
 	"fmt"
-	"io"
 	"log/slog"
 	"net/netip"
 	"os"
@@ -17,6 +16,7 @@ import (
 
 	"github.com/gaissmai/bart"
 	"github.com/slackhq/nebula/config"
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 	"github.com/slackhq/nebula/util"
 	netroute "golang.org/x/net/route"
@@ -59,6 +59,18 @@ type tun struct {
 	fd          int
 	// cache out buffer since we need to prepend 4 bytes for tun metadata
 	out []byte
+
+	readBuf  []byte
+	batchRet [1][]byte
+}
+
+func (t *tun) Read() ([][]byte, error) {
+	n, err := t.readOne(t.readBuf)
+	if err != nil {
+		return nil, err
+	}
+	t.batchRet[0] = t.readBuf[:n]
+	return t.batchRet[:], nil
 }
 
 var deviceNameRE = regexp.MustCompile(`^tun[0-9]+$`)
@@ -95,6 +107,7 @@ func newTun(c *config.C, l *slog.Logger, vpnNetworks []netip.Prefix, _ bool) (*t
 		vpnNetworks: vpnNetworks,
 		MTU:         c.GetInt("tun.mtu", DefaultMTU),
 		l:           l,
+		readBuf:     make([]byte, defaultBatchBufSize),
 	}
 
 	err = t.reload(c, true)
@@ -124,7 +137,7 @@ func (t *tun) Close() error {
 	return nil
 }
 
-func (t *tun) Read(to []byte) (int, error) {
+func (t *tun) readOne(to []byte) (int, error) {
 	buf := make([]byte, len(to)+4)
 
 	n, err := t.f.Read(buf)
@@ -310,20 +323,12 @@ func (t *tun) Name() string {
 	return t.Device
 }
 
-func (t *tun) SupportsPerPeerMTU() bool {
-	return false
-}
-
-func (t *tun) SetPeerMTU(addr netip.Addr, mtu int) error {
-	return nil
-}
-
 func (t *tun) SupportsMultiqueue() bool {
 	return false
 }
 
-func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+func (t *tun) NewMultiQueueReader() error {
-	return nil, fmt.Errorf("TODO: multiqueue not implemented for openbsd")
+	return fmt.Errorf("TODO: multiqueue not implemented for openbsd")
 }
 
 func (t *tun) addRoutes(logErrors bool) error {
@@ -374,6 +379,10 @@ func (t *tun) deviceBytes() (o [16]byte) {
 	return
 }
 
+func (t *tun) Readers() []tio.Queue {
+	return []tio.Queue{t}
+}
+
 func addRoute(prefix netip.Prefix, gateways []netip.Prefix) error {
 	sock, err := unix.Socket(unix.AF_ROUTE, unix.SOCK_RAW, unix.AF_UNSPEC)
 	if err != nil {

overlay/tun_tester.go

@@ -14,6 +14,7 @@ import (
 
 	"github.com/gaissmai/bart"
 	"github.com/slackhq/nebula/config"
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 )
 
@@ -27,6 +28,17 @@ type TestTun struct {
 	closed    atomic.Bool
 	rxPackets chan []byte // Packets to receive into nebula
 	TxPackets chan []byte // Packets transmitted outside by nebula
+
+	batchRet [1][]byte
+}
+
+func (t *TestTun) Read() ([][]byte, error) {
+	p, ok := <-t.rxPackets
+	if !ok {
+		return nil, os.ErrClosed
+	}
+	t.batchRet[0] = p
+	return t.batchRet[:], nil
 }
 
 func newTun(c *config.C, l *slog.Logger, vpnNetworks []netip.Prefix, _ bool) (*TestTun, error) {
@@ -105,14 +117,6 @@ func (t *TestTun) Name() string {
 	return t.Device
 }
 
-func (t *TestTun) SupportsPerPeerMTU() bool {
-	return false
-}
-
-func (t *TestTun) SetPeerMTU(addr netip.Addr, mtu int) error {
-	return nil
-}
-
 func (t *TestTun) Write(b []byte) (n int, err error) {
 	if t.closed.Load() {
 		return 0, io.ErrClosedPipe
@@ -132,19 +136,14 @@ func (t *TestTun) Close() error {
 	return nil
 }
 
-func (t *TestTun) Read(b []byte) (int, error) {
+func (t *TestTun) Readers() []tio.Queue {
-	p, ok := <-t.rxPackets
+	return []tio.Queue{t}
-	if !ok {
-		return 0, os.ErrClosed
-	}
-	copy(b, p)
-	return len(p), nil
 }
 
 func (t *TestTun) SupportsMultiqueue() bool {
 	return false
 }
 
-func (t *TestTun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+func (t *TestTun) NewMultiQueueReader() error {
-	return nil, fmt.Errorf("TODO: multiqueue not implemented")
+	return fmt.Errorf("TODO: multiqueue not implemented")
 }

overlay/tun_windows.go

@@ -6,7 +6,6 @@ package overlay
 import (
 	"crypto"
 	"fmt"
-	"io"
 	"log/slog"
 	"net/netip"
 	"os"
@@ -18,6 +17,7 @@ import (
 
 	"github.com/gaissmai/bart"
 	"github.com/slackhq/nebula/config"
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 	"github.com/slackhq/nebula/util"
 	"github.com/slackhq/nebula/wintun"
@@ -36,6 +36,18 @@ type winTun struct {
 	l           *slog.Logger
 
 	tun *wintun.NativeTun
+
+	readBuf  []byte
+	batchRet [1][]byte
+}
+
+func (t *winTun) Read() ([][]byte, error) {
+	n, err := t.tun.Read(t.readBuf, 0)
+	if err != nil {
+		return nil, err
+	}
+	t.batchRet[0] = t.readBuf[:n]
+	return t.batchRet[:], nil
 }
 
 func newTunFromFd(_ *config.C, _ *slog.Logger, _ int, _ []netip.Prefix) (Device, error) {
@@ -55,6 +67,7 @@ func newTun(c *config.C, l *slog.Logger, vpnNetworks []netip.Prefix, _ bool) (*w
 	}
 
 	t := &winTun{
+		readBuf:     make([]byte, defaultBatchBufSize),
 		Device:      deviceName,
 		vpnNetworks: vpnNetworks,
 		MTU:         c.GetInt("tun.mtu", DefaultMTU),
@@ -229,18 +242,6 @@ func (t *winTun) Name() string {
 	return t.Device
 }
 
-func (t *winTun) SupportsPerPeerMTU() bool {
-	return false
-}
-
-func (t *winTun) SetPeerMTU(addr netip.Addr, mtu int) error {
-	return nil
-}
-
-func (t *winTun) Read(b []byte) (int, error) {
-	return t.tun.Read(b, 0)
-}
-
 func (t *winTun) Write(b []byte) (int, error) {
 	return t.tun.Write(b, 0)
 }
@@ -249,8 +250,12 @@ func (t *winTun) SupportsMultiqueue() bool {
 	return false
 }
 
-func (t *winTun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+func (t *winTun) NewMultiQueueReader() error {
-	return nil, fmt.Errorf("TODO: multiqueue not implemented for windows")
+	return fmt.Errorf("TODO: multiqueue not implemented for windows")
+}
+
+func (t *winTun) Readers() []tio.Queue {
+	return []tio.Queue{t}
 }
 
 func (t *winTun) Close() error {

overlay/user.go

@@ -6,6 +6,7 @@ import (
 	"net/netip"
 
 	"github.com/slackhq/nebula/config"
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 )
 
@@ -23,17 +24,34 @@ func NewUserDevice(vpnNetworks []netip.Prefix) (Device, error) {
 		outboundWriter: ow,
 		inboundReader:  ir,
 		inboundWriter:  iw,
+		numReaders:     1,
 	}, nil
 }
 
 type UserDevice struct {
 	vpnNetworks []netip.Prefix
+	numReaders  int
 
 	outboundReader *io.PipeReader
 	outboundWriter *io.PipeWriter
 
 	inboundReader *io.PipeReader
 	inboundWriter *io.PipeWriter
+
+	readBuf  []byte
+	batchRet [1][]byte
+}
+
+func (d *UserDevice) Read() ([][]byte, error) {
+	if d.readBuf == nil {
+		d.readBuf = make([]byte, defaultBatchBufSize)
+	}
+	n, err := d.outboundReader.Read(d.readBuf)
+	if err != nil {
+		return nil, err
+	}
+	d.batchRet[0] = d.readBuf[:n]
+	return d.batchRet[:], nil
 }
 
 func (d *UserDevice) Activate() error {
@@ -46,32 +64,31 @@ func (d *UserDevice) RoutesFor(ip netip.Addr) routing.Gateways {
 	return routing.Gateways{routing.NewGateway(ip, 1)}
 }
 
-func (d *UserDevice) SupportsPerPeerMTU() bool {
-	return false
-}
-
-func (d *UserDevice) SetPeerMTU(addr netip.Addr, mtu int) error {
-	return nil
-}
-
 func (d *UserDevice) SupportsMultiqueue() bool {
 	return true
 }
 
-func (d *UserDevice) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+func (d *UserDevice) NewMultiQueueReader() error {
-	return d, nil
+	d.numReaders++
+	return nil
+}
+
+func (d *UserDevice) Readers() []tio.Queue {
+	out := make([]tio.Queue, d.numReaders)
+	for i := range d.numReaders {
+		out[i] = d
+	}
+	return out
 }
 
 func (d *UserDevice) Pipe() (*io.PipeReader, *io.PipeWriter) {
 	return d.inboundReader, d.outboundWriter
 }
 
-func (d *UserDevice) Read(p []byte) (n int, err error) {
-	return d.outboundReader.Read(p)
-}
 func (d *UserDevice) Write(p []byte) (n int, err error) {
 	return d.inboundWriter.Write(p)
 }
+
 func (d *UserDevice) Close() error {
 	d.inboundWriter.Close()
 	d.outboundWriter.Close()

pki.go

@@ -15,12 +15,9 @@ import (
 	"sync/atomic"
 	"time"
 
-	"github.com/flynn/noise"
 	"github.com/gaissmai/bart"
 	"github.com/slackhq/nebula/cert"
 	"github.com/slackhq/nebula/config"
-	"github.com/slackhq/nebula/handshake"
-	"github.com/slackhq/nebula/noiseutil"
 	"github.com/slackhq/nebula/util"
 )
 
@@ -32,10 +29,10 @@ type PKI struct {
 
 type CertState struct {
 	v1Cert           cert.Certificate
-	v1Credential *handshake.Credential
+	v1HandshakeBytes []byte
 
 	v2Cert           cert.Certificate
-	v2Credential *handshake.Credential
+	v2HandshakeBytes []byte
 
 	initiatingVersion cert.Version
 	privateKey        []byte
@@ -95,35 +92,13 @@ func (p *PKI) reload(c *config.C, initial bool) error {
 }
 
 func (p *PKI) reloadCerts(c *config.C, initial bool) *util.ContextualError {
-	var cipher string
+	newState, err := newCertStateFromConfig(c)
-	var currentState *CertState
-	if initial {
-		cipher = c.GetString("cipher", "aes")
-		//TODO: this sucks and we should make it not a global
-		switch cipher {
-		case "aes":
-			noiseEndianness = binary.BigEndian
-		case "chachapoly":
-			noiseEndianness = binary.LittleEndian
-		default:
-			return util.NewContextualError(
-				"unknown cipher",
-				m{"cipher": cipher},
-				nil,
-			)
-		}
-	} else {
-		// Cipher cant be hot swapped so just leave it at what it was before
-		currentState = p.cs.Load()
-		cipher = currentState.cipher
-	}
-
-	newState, err := newCertStateFromConfig(c, cipher)
 	if err != nil {
 		return util.NewContextualError("Could not load client cert", nil, err)
 	}
 
-	if currentState != nil {
+	if !initial {
+		currentState := p.cs.Load()
 		if newState.v1Cert != nil {
 			if currentState.v1Cert == nil {
 				//adding certs is fine, actually. Networks-in-common confirmed in newCertState().
@@ -183,6 +158,25 @@ func (p *PKI) reloadCerts(c *config.C, initial bool) *util.ContextualError {
 				)
 			}
 		}
+
+		// Cipher cant be hot swapped so just leave it at what it was before
+		newState.cipher = currentState.cipher
+
+	} else {
+		newState.cipher = c.GetString("cipher", "aes")
+		//TODO: this sucks and we should make it not a global
+		switch newState.cipher {
+		case "aes":
+			noiseEndianness = binary.BigEndian
+		case "chachapoly":
+			noiseEndianness = binary.LittleEndian
+		default:
+			return util.NewContextualError(
+				"unknown cipher",
+				m{"cipher": newState.cipher},
+				nil,
+			)
+		}
 	}
 
 	p.cs.Store(newState)
@@ -214,20 +208,6 @@ func (cs *CertState) GetDefaultCertificate() cert.Certificate {
 	return c
 }
 
-// DefaultVersion returns the preferred cert version for initiating handshakes.
-func (cs *CertState) DefaultVersion() cert.Version { return cs.initiatingVersion }
-
-// GetCredential returns the pre-computed handshake credential for the given version, or nil.
-func (cs *CertState) GetCredential(v cert.Version) *handshake.Credential {
-	switch v {
-	case cert.Version1:
-		return cs.v1Credential
-	case cert.Version2:
-		return cs.v2Credential
-	}
-	return nil
-}
-
 func (cs *CertState) getCertificate(v cert.Version) cert.Certificate {
 	switch v {
 	case cert.Version1:
@@ -239,25 +219,17 @@ func (cs *CertState) getCertificate(v cert.Version) cert.Certificate {
 	return nil
 }
 
-func newCipherSuite(curve cert.Curve, pkcs11backed bool, cipher string) (noise.CipherSuite, error) {
+// getHandshakeBytes returns the cached bytes to be used in a handshake message for the requested version.
-	var dhFunc noise.DHFunc
+// Callers must check if the return []byte is nil.
-	switch curve {
+func (cs *CertState) getHandshakeBytes(v cert.Version) []byte {
-	case cert.Curve_CURVE25519:
+	switch v {
-		dhFunc = noise.DH25519
+	case cert.Version1:
-	case cert.Curve_P256:
+		return cs.v1HandshakeBytes
-		if pkcs11backed {
+	case cert.Version2:
-			dhFunc = noiseutil.DHP256PKCS11
+		return cs.v2HandshakeBytes
-		} else {
-			dhFunc = noiseutil.DHP256
-		}
 	default:
-		return nil, fmt.Errorf("unsupported curve: %s", curve)
+		return nil
 	}
-
-	if cipher == "chachapoly" {
-		return noise.NewCipherSuite(dhFunc, noise.CipherChaChaPoly, noise.HashSHA256), nil
-	}
-	return noise.NewCipherSuite(dhFunc, noiseutil.CipherAESGCM, noise.HashSHA256), nil
 }
 
 func (cs *CertState) String() string {
@@ -289,7 +261,7 @@ func (cs *CertState) MarshalJSON() ([]byte, error) {
 	return json.Marshal(msg)
 }
 
-func newCertStateFromConfig(c *config.C, cipher string) (*CertState, error) {
+func newCertStateFromConfig(c *config.C) (*CertState, error) {
 	var err error
 
 	privPathOrPEM := c.GetString("pki.key", "")
@@ -373,14 +345,13 @@ func newCertStateFromConfig(c *config.C, cipher string) (*CertState, error) {
 		return nil, fmt.Errorf("unknown pki.initiating_version: %v", rawInitiatingVersion)
 	}
 
-	return newCertState(initiatingVersion, v1, v2, isPkcs11, curve, rawKey, cipher)
+	return newCertState(initiatingVersion, v1, v2, isPkcs11, curve, rawKey)
 }
 
-func newCertState(dv cert.Version, v1, v2 cert.Certificate, pkcs11backed bool, privateKeyCurve cert.Curve, privateKey []byte, cipher string) (*CertState, error) {
+func newCertState(dv cert.Version, v1, v2 cert.Certificate, pkcs11backed bool, privateKeyCurve cert.Curve, privateKey []byte) (*CertState, error) {
 	cs := CertState{
 		privateKey:               privateKey,
 		pkcs11Backed:             pkcs11backed,
-		cipher:                   cipher,
 		myVpnNetworksTable:       new(bart.Lite),
 		myVpnAddrsTable:          new(bart.Lite),
 		myVpnBroadcastAddrsTable: new(bart.Lite),
@@ -413,14 +384,10 @@ func newCertState(dv cert.Version, v1, v2 cert.Certificate, pkcs11backed bool, p
 
 		v1hs, err := v1.MarshalForHandshakes()
 		if err != nil {
-			return nil, fmt.Errorf("error marshalling v1 certificate for handshake: %w", err)
+			return nil, fmt.Errorf("error marshalling certificate for handshake: %w", err)
-		}
-		ncs, err := newCipherSuite(v1.Curve(), pkcs11backed, cipher)
-		if err != nil {
-			return nil, err
 		}
 		cs.v1Cert = v1
-		cs.v1Credential = handshake.NewCredential(v1, v1hs, privateKey, ncs)
+		cs.v1HandshakeBytes = v1hs
 
 		if cs.initiatingVersion == 0 {
 			cs.initiatingVersion = cert.Version1
@@ -438,14 +405,10 @@ func newCertState(dv cert.Version, v1, v2 cert.Certificate, pkcs11backed bool, p
 
 		v2hs, err := v2.MarshalForHandshakes()
 		if err != nil {
-			return nil, fmt.Errorf("error marshalling v2 certificate for handshake: %w", err)
+			return nil, fmt.Errorf("error marshalling certificate for handshake: %w", err)
-		}
-		ncs, err := newCipherSuite(v2.Curve(), pkcs11backed, cipher)
-		if err != nil {
-			return nil, err
 		}
 		cs.v2Cert = v2
-		cs.v2Credential = handshake.NewCredential(v2, v2hs, privateKey, ncs)
+		cs.v2HandshakeBytes = v2hs
 
 		if cs.initiatingVersion == 0 {
 			cs.initiatingVersion = cert.Version2

pmtud_manager.go

@@ -1,623 +0,0 @@
-package nebula
-
-import (
-	"context"
-	"encoding/binary"
-	"log/slog"
-	"math/rand/v2"
-	"net/netip"
-	"sync"
-	"sync/atomic"
-	"time"
-
-	"github.com/slackhq/nebula/config"
-	"github.com/slackhq/nebula/header"
-	"github.com/slackhq/nebula/overlay"
-)
-
-// PMTUD PoC: discover the path MTU per-tunnel via authenticated probes that ride
-// the existing crypto session. We follow RFC 8899 PLPMTUD: a binary search
-// between a known-good floor and a configured ceiling, with N consecutive probe
-// losses at a size treated as "doesn't fit." Confirmed PMTU is pushed to the
-// overlay device, which on Linux installs a per-host route with the discovered
-// MTU. The kernel then surfaces EMSGSIZE / PTB to apps writing to the tun.
-//
-// Probe payload format (request):
-//
-//	[magic uint32 BE][probeID uint32 BE][padding 0x00...]
-//
-// Reply is a small ack with the same magic and probeID and no padding. We do not
-// verify the reverse-path MTU; only the forward direction matters for the
-// receiver's MTU on the inside.
-
-const (
-	pmtudMagic uint32 = 0x504D5544 // 'P' 'M' 'U' 'D'
-	pmtudFloor        = 1280       // IPv6 minimum payload, also a safe internet MTU floor
-
-	// pmtudConverged is the bytes-tolerance for stopping the search.
-	pmtudConverged = 8
-
-	// pmtudMaxLoss matches RFC 8899 MAX_PROBES (default 3).
-	pmtudMaxLoss = 3
-
-	// pmtudProbeInterval is the time between probe ticks during the search phase.
-	// Once a peer converges the wheel stops ticking it; re-validation is driven
-	// by connection_manager via MaybeProbeAsTest at its natural test cadence.
-	pmtudProbeInterval = 500 * time.Millisecond
-
-	// pmtudWheelMax is the wheel's maximum supported scheduling duration. We
-	// only ever schedule at pmtudProbeInterval today, but the wheel needs a
-	// max greater than its tick to allocate its slot ring sensibly.
-	pmtudWheelMax = 5 * time.Second
-
-	// pmtudOverheadPessimistic assumes IPv6 underlay + relay framing:
-	//   IPv6(40) + UDP(8) + outer nebula(16) + outer AEAD tag(16)
-	// + inner nebula(16) + inner AEAD tag(16) = 112 bytes.
-	// TODO: track underlay address family and per-peer relay state on the HostInfo
-	// so the manager can use the actual overhead for that tunnel and recover the
-	// 32 bytes we pessimistically give up on direct IPv6 paths and the 52 bytes on
-	// direct IPv4 paths.
-	pmtudOverheadPessimistic = 112
-
-	// pmtudUnsupportedAfter is the number of consecutive lost probes (across any
-	// sizes) without ever receiving a reply that we treat as evidence the peer
-	// does not understand the MTUDProbeRequest subtype (i.e. it's running an
-	// older nebula). After this many failures with everReplied=false we mark the
-	// peer pmtud-unsupported and stop scheduling probes. K is small enough that
-	// it fires before the binary search would naturally converge to floor (which
-	// would otherwise be ~30 wasted probes), but large enough to absorb a few
-	// transient probe losses on a path that's just starting to settle.
-	pmtudUnsupportedAfter = 5
-)
-
-// pmtudPeer tracks the binary-search state for one tunnel.
-type pmtudPeer struct {
-	mu       sync.Mutex
-	addr     netip.Addr
-	localIdx uint32
-
-	// low is the largest outer IP packet size we have a confirmed ack for.
-	// high is the smallest size we believe fails (the search ceiling to start).
-	low, high int
-
-	// inFlightSize is the outer IP packet size of the probe currently awaiting
-	// an ack. 0 means no probe in flight.
-	inFlightSize int
-	// inFlightID matches the probeID echoed in the reply.
-	inFlightID uint32
-	// losses counts consecutive failures at inFlightSize.
-	losses int
-
-	// firstProbe is true until we have sent the first probe of a search. The
-	// first probe targets the ceiling directly (RFC 8899 permits this Search
-	// Algorithm choice); operators who set tun.max_mtu typically have a path
-	// that supports it, so we converge in one probe in the common case.
-	firstProbe bool
-	// everReplied is true once we have ever received any MTUDProbeReply from
-	// this peer. Combined with consecutiveFailures, this lets us detect peers
-	// that don't understand the new subtype and stop probing them.
-	everReplied bool
-	// consecutiveFailures counts probes lost without an intervening reply.
-	// Resets to 0 on any successful reply.
-	consecutiveFailures int
-	// unsupported is set true once we conclude the peer doesn't speak PMTUD.
-	// The manager skips probes for unsupported peers.
-	unsupported bool
-
-	// converged means we have a confirmed PMTU and are in the slow re-validation phase.
-	converged bool
-	// applied is the inner MTU we last pushed to the overlay device (0 if never).
-	applied int
-}
-
-func (p *pmtudPeer) overhead() int {
-	// TODO: branch on actual underlay family + relay state for this peer.
-	return pmtudOverheadPessimistic
-}
-
-func (p *pmtudPeer) midpoint() int {
-	return (p.low + p.high) / 2
-}
-
-type pmtudManager struct {
-	intf   *Interface
-	device overlay.Device
-
-	// peers is keyed by HostInfo.localIndexId.
-	peers sync.Map // map[uint32]*pmtudPeer
-
-	wheel *LockingTimerWheel[uint32]
-
-	// floor is the always-safe inner MTU (= tun.mtu). Per-peer routes start here
-	// on tunnel-up so unprobed traffic is always small enough to fit. Stored as
-	// atomic int64 so reload can update it without coordinating with the readers
-	// in tick/HandleReply/OnTunnelUp.
-	floor atomic.Int64
-	// ceiling is the search ceiling expressed as an outer IP packet size, derived
-	// from tun.max_mtu (which is the kernel's device MTU on the tun) plus our
-	// pessimistic overhead. PMTUD will not probe larger than this.
-	ceiling atomic.Int64
-
-	enabled atomic.Bool
-
-	l *slog.Logger
-}
-
-func newPMTUDManagerFromConfig(l *slog.Logger, c *config.C, device overlay.Device) *pmtudManager {
-	m := &pmtudManager{
-		device: device,
-		wheel:  NewLockingTimerWheel[uint32](pmtudProbeInterval, pmtudWheelMax),
-		l:      l,
-	}
-	c.RegisterReloadCallback(func(c *config.C) { m.reload(c, false) })
-	m.reload(c, true)
-	return m
-}
-
-// reload applies tun.mtu / tun.max_mtu changes to the manager. On the initial
-// call (during construction) it just snapshots state; on a live reload it also
-// transitions in-flight peers to match the new bounds: clearing per-peer routes
-// when newly disabled, seeding peers from the hostmap and flipping DF on
-// outside sockets when newly enabled, and rebounding existing searches in
-// place when only the ceiling moved.
-func (m *pmtudManager) reload(c *config.C, initial bool) {
-	if !initial && !c.HasChanged("tun.mtu") && !c.HasChanged("tun.max_mtu") {
-		return
-	}
-
-	floor := c.GetInt("tun.mtu", overlay.DefaultMTU)
-	maxMTU := c.GetInt("tun.max_mtu", 0)
-
-	enable := maxMTU > floor && m.device.SupportsPerPeerMTU()
-	var ceiling int
-	if enable {
-		ceiling = maxMTU + pmtudOverheadPessimistic
-	}
-
-	if initial {
-		m.floor.Store(int64(floor))
-		m.ceiling.Store(int64(ceiling))
-		m.enabled.Store(enable)
-		switch {
-		case enable:
-			m.l.Info("pmtud enabled", "floor", floor, "ceiling", ceiling, "tun.max_mtu", maxMTU)
-		case maxMTU > floor:
-			m.l.Warn("pmtud disabled: this platform does not yet support per-peer MTU routes",
-				"tun.max_mtu", maxMTU)
-		}
-		return
-	}
-
-	wasEnabled := m.enabled.Load()
-	m.floor.Store(int64(floor))
-	m.ceiling.Store(int64(ceiling))
-	m.enabled.Store(enable)
-
-	switch {
-	case wasEnabled && !enable:
-		m.disableLive(floor, maxMTU)
-	case !wasEnabled && enable:
-		m.enableLive(floor, ceiling, maxMTU)
-	case wasEnabled && enable:
-		m.reboundLive(floor, ceiling, maxMTU)
-	}
-}
-
-// disableLive clears per-peer routes and drops all peer state. We do not
-// disable DF on the outside sockets; once on, it stays on for the life of the
-// process. Operators flipping pmtud off live get correct routing behavior; if
-// they want the historical no-DF behavior back they need to restart.
-func (m *pmtudManager) disableLive(floor, maxMTU int) {
-	m.peers.Range(func(k, v any) bool {
-		p := v.(*pmtudPeer)
-		p.mu.Lock()
-		applied := p.applied
-		addr := p.addr
-		p.applied = 0
-		p.mu.Unlock()
-		if applied != 0 {
-			if err := m.device.SetPeerMTU(addr, 0); err != nil {
-				m.l.Warn("pmtud: failed to clear per-peer mtu on disable", "addr", addr, "error", err)
-			}
-		}
-		m.peers.Delete(k)
-		return true
-	})
-	m.l.Info("pmtud disabled (tun.max_mtu <= tun.mtu)", "tun.mtu", floor, "tun.max_mtu", maxMTU)
-}
-
-// enableLive flips DF on every outside socket. We don't pre-seed existing
-// tunnels here; connection_manager's normal test cadence will eventually call
-// MaybeProbeAsTest for each peer, which seeds on miss and lets the wheel pick
-// up the search from there. New tunnels established after this point still
-// take the OnTunnelUp fast path.
-func (m *pmtudManager) enableLive(floor, ceiling, maxMTU int) {
-	m.enableDF()
-	m.l.Info("pmtud enabled", "floor", floor, "ceiling", ceiling, "tun.max_mtu", maxMTU)
-}
-
-// reboundLive resets each peer's search state to the new bounds. Peers whose
-// confirmed PMTU still fits under the new ceiling keep their applied route in
-// place during the new search; peers whose confirmed PMTU exceeds the new
-// ceiling get cleared back to floor and re-search from scratch. The unsupported
-// flag is preserved because peer software version doesn't change on reload.
-func (m *pmtudManager) reboundLive(floor, ceiling, maxMTU int) {
-	overhead := pmtudOverheadPessimistic
-	m.peers.Range(func(k, v any) bool {
-		p := v.(*pmtudPeer)
-		p.mu.Lock()
-		if p.applied > 0 && p.applied+overhead > ceiling {
-			if err := m.device.SetPeerMTU(p.addr, 0); err != nil {
-				m.l.Warn("pmtud: failed to clear per-peer mtu on rebound", "addr", p.addr, "error", err)
-			} else {
-				p.applied = 0
-			}
-		}
-		p.low = floor + overhead
-		p.high = ceiling
-		p.inFlightSize = 0
-		p.inFlightID = 0
-		p.losses = 0
-		p.firstProbe = !p.unsupported
-		p.converged = false
-		idx := p.localIdx
-		unsupported := p.unsupported
-		p.mu.Unlock()
-		if !unsupported {
-			m.wheel.Add(idx, pmtudProbeInterval)
-		}
-		return true
-	})
-	m.l.Info("pmtud reloaded", "floor", floor, "ceiling", ceiling, "tun.max_mtu", maxMTU)
-}
-
-// enableDF asks every outside socket to set the don't-fragment bit on outbound
-// packets. Idempotent: safe to call from both Start (initial enable) and from a
-// live reload that flips pmtud on.
-func (m *pmtudManager) enableDF() {
-	for i, w := range m.intf.writers {
-		if err := w.EnablePathMTUDiscovery(); err != nil {
-			m.l.Warn("pmtud: failed to enable path mtu discovery on outside socket; pmtud will not work correctly",
-				"writer", i, "error", err)
-		}
-	}
-}
-
-// Start runs the probe scheduler until ctx is done. The loop runs even when PMTUD
-// is disabled at startup so a hot reload can turn it on without restarting nebula.
-//
-// When PMTUD is enabled at startup we ask each outside socket to enable
-// path-MTU discovery (DF on every send). This is intentionally gated on the
-// feature being on so that operators who haven't opted in keep the historical
-// behavior where the kernel may fragment outbound nebula UDP packets. A live
-// reload from disabled to enabled will also flip DF on via enableLive; the
-// reverse direction does not turn DF off, so flipping pmtud back off live
-// keeps DF on until restart.
-func (m *pmtudManager) Start(ctx context.Context) {
-	if m.enabled.Load() {
-		m.enableDF()
-	}
-
-	ticker := time.NewTicker(m.wheel.t.tickDuration)
-	defer ticker.Stop()
-
-	for {
-		select {
-		case <-ctx.Done():
-			return
-		case now := <-ticker.C:
-			m.wheel.Advance(now)
-			for {
-				idx, has := m.wheel.Purge()
-				if !has {
-					break
-				}
-				m.tick(idx)
-			}
-		}
-	}
-}
-
-// OnTunnelUp is called when a HostInfo becomes traffic-watched. The kernel
-// already routes packets to this peer through the per-vpn-network route (mtu =
-// tun.mtu), so the floor is in effect implicitly. We just kick off the search
-// here; HandleReply will install a per-host /32 (or /128) route once a larger
-// size is confirmed.
-func (m *pmtudManager) OnTunnelUp(hi *HostInfo) {
-	if !m.enabled.Load() {
-		return
-	}
-	m.seedPeer(hi)
-}
-
-// seedPeer is the shared body of OnTunnelUp and the live-reload enable path.
-// LoadOrStore protects against double-seeding the same localIndexId from a
-// race between OnTunnelUp and a reload-driven hostmap walk.
-func (m *pmtudManager) seedPeer(hi *HostInfo) {
-	if hi == nil || len(hi.vpnAddrs) == 0 {
-		return
-	}
-	floor := int(m.floor.Load())
-	ceiling := int(m.ceiling.Load())
-	p := &pmtudPeer{
-		addr:       hi.vpnAddrs[0],
-		localIdx:   hi.localIndexId,
-		low:        floor + pmtudOverheadPessimistic,
-		high:       ceiling,
-		firstProbe: true,
-	}
-	if _, loaded := m.peers.LoadOrStore(hi.localIndexId, p); loaded {
-		return
-	}
-	m.wheel.Add(hi.localIndexId, pmtudProbeInterval)
-}
-
-// OnTunnelDown is called when a HostInfo is being torn down. Removes any per-host
-// MTU override so the device default applies again.
-func (m *pmtudManager) OnTunnelDown(hi *HostInfo) {
-	if hi == nil {
-		return
-	}
-	v, ok := m.peers.LoadAndDelete(hi.localIndexId)
-	if !ok {
-		return
-	}
-	p := v.(*pmtudPeer)
-	p.mu.Lock()
-	applied := p.applied
-	addr := p.addr
-	p.applied = 0
-	p.mu.Unlock()
-	if applied != 0 {
-		if err := m.device.SetPeerMTU(addr, 0); err != nil {
-			m.l.Warn("pmtud: failed to clear per-peer mtu", "addr", addr, "error", err)
-		}
-	}
-}
-
-// OnRoam is called when a HostInfo's remote underlay address changes. The path
-// MTU may now be different; drop the per-host route so the kernel falls back to
-// the per-vpn-network route (mtu = tun.mtu floor), then restart the search.
-// We do not reset the unsupported flag: peer software version doesn't change on
-// roam, so once we've decided a peer doesn't speak PMTUD we stay decided.
-func (m *pmtudManager) OnRoam(hi *HostInfo) {
-	if !m.enabled.Load() || hi == nil {
-		return
-	}
-	v, ok := m.peers.Load(hi.localIndexId)
-	if !ok {
-		return
-	}
-	p := v.(*pmtudPeer)
-	p.mu.Lock()
-	if p.unsupported {
-		p.mu.Unlock()
-		return
-	}
-	p.low = int(m.floor.Load()) + pmtudOverheadPessimistic
-	p.high = int(m.ceiling.Load())
-	p.inFlightSize = 0
-	p.inFlightID = 0
-	p.losses = 0
-	p.consecutiveFailures = 0
-	p.firstProbe = true
-	p.converged = false
-	if p.applied != 0 {
-		if err := m.device.SetPeerMTU(p.addr, 0); err != nil {
-			m.l.Warn("pmtud: failed to clear per-peer mtu on roam", "addr", p.addr, "error", err)
-		} else {
-			p.applied = 0
-		}
-	}
-	p.mu.Unlock()
-	m.wheel.Add(hi.localIndexId, pmtudProbeInterval)
-}
-
-// MaybeProbeAsTest is called by connection_manager when it would otherwise send
-// a TestRequest because a tunnel has gone silent. If we have a confirmed PMTU
-// for this peer that's larger than the floor, we send a probe at that size
-// instead. The reply confirms both liveness (consumed by connection_manager via
-// the existing inbound traffic accounting fallthrough in outside.go) and that
-// the confirmed PMTU still fits (consumed by HandleReply here). One synthetic
-// packet does the work of two.
-//
-// Returns true if a probe was sent. False means the caller should send a
-// regular TestRequest at the floor.
-//
-// On probe failure, connection_manager's existing pendingDeletion timeout will
-// tear the tunnel down. Heavy hammer, but correct: a re-handshake re-runs PMTUD
-// discovery against the now-shrunken path. A future EMSGSIZE-capture followup
-// can replace this with a soft-drop-and-research flow.
-func (m *pmtudManager) MaybeProbeAsTest(hi *HostInfo) bool {
-	if !m.enabled.Load() || hi == nil {
-		return false
-	}
-	v, ok := m.peers.Load(hi.localIndexId)
-	if !ok {
-		// Tunnel pre-dates the manager being aware of it (e.g. pmtud was just
-		// enabled live, or AddTrafficWatch fired before this call). Seed the
-		// peer so the wheel picks up the search; let connection_manager send
-		// its regular TestRequest this cycle.
-		m.seedPeer(hi)
-		return false
-	}
-	p := v.(*pmtudPeer)
-	p.mu.Lock()
-	if p.unsupported || p.applied == 0 {
-		p.mu.Unlock()
-		return false
-	}
-	overhead := p.overhead()
-	size := p.applied + overhead
-	id := rand.Uint32()
-	p.inFlightSize = size
-	p.inFlightID = id
-	p.mu.Unlock()
-
-	m.sendProbe(hi, size, id, overhead)
-	return true
-}
-
-// HandleReply consumes an MTUDProbeReply payload from the receive path.
-func (m *pmtudManager) HandleReply(localIdx uint32, payload []byte) {
-	if !m.enabled.Load() {
-		return
-	}
-	if len(payload) < 8 {
-		return
-	}
-	if binary.BigEndian.Uint32(payload[0:4]) != pmtudMagic {
-		return
-	}
-	id := binary.BigEndian.Uint32(payload[4:8])
-
-	v, ok := m.peers.Load(localIdx)
-	if !ok {
-		return
-	}
-	p := v.(*pmtudPeer)
-	p.mu.Lock()
-	defer p.mu.Unlock()
-
-	if p.inFlightSize == 0 || p.inFlightID != id {
-		return
-	}
-
-	confirmed := p.inFlightSize
-	p.low = confirmed
-	p.inFlightSize = 0
-	p.losses = 0
-	p.everReplied = true
-	p.consecutiveFailures = 0
-
-	innerMTU := confirmed - p.overhead()
-	// Only install a /32 override when it would actually raise the MTU above the
-	// per-vpn-network floor route. If the discovered MTU is <= floor, the /24
-	// already covers it; installing a /32 at floor would just create roam churn.
-	if innerMTU > int(m.floor.Load()) && p.applied != innerMTU {
-		if err := m.device.SetPeerMTU(p.addr, innerMTU); err != nil {
-			m.l.Warn("pmtud: failed to apply per-peer mtu", "addr", p.addr, "innerMTU", innerMTU, "error", err)
-		} else {
-			m.l.Info("pmtud probe confirmed",
-				"addr", p.addr,
-				"outerMTU", confirmed,
-				"innerMTU", innerMTU,
-				"low", p.low,
-				"high", p.high,
-			)
-			p.applied = innerMTU
-		}
-	}
-
-	if p.high-p.low <= pmtudConverged {
-		p.converged = true
-	} else {
-		p.converged = false
-	}
-}
-
-// tick handles one wheel firing for a single peer.
-func (m *pmtudManager) tick(localIdx uint32) {
-	v, ok := m.peers.Load(localIdx)
-	if !ok {
-		return
-	}
-	p := v.(*pmtudPeer)
-	p.mu.Lock()
-
-	if p.unsupported {
-		p.mu.Unlock()
-		return
-	}
-
-	// If a probe was outstanding, this tick is the loss timeout.
-	if p.inFlightSize != 0 {
-		p.losses++
-		p.consecutiveFailures++
-		if p.losses >= pmtudMaxLoss {
-			p.high = p.inFlightSize
-			p.inFlightSize = 0
-			p.losses = 0
-			if p.high-p.low <= pmtudConverged {
-				p.converged = true
-			}
-		}
-	}
-
-	// If we've never gotten a reply from this peer and we've burned through our
-	// failure budget, conclude the peer doesn't understand the MTUDProbeRequest
-	// subtype and stop scheduling probes for it.
-	if !p.everReplied && p.consecutiveFailures >= pmtudUnsupportedAfter {
-		p.unsupported = true
-		addr := p.addr
-		p.mu.Unlock()
-		m.l.Info("pmtud: peer not responding to probes, marking unsupported",
-			"addr", addr, "failures", pmtudUnsupportedAfter)
-		return
-	}
-
-	hi := m.intf.hostMap.QueryIndex(localIdx)
-	if hi == nil {
-		p.mu.Unlock()
-		m.peers.Delete(localIdx)
-		return
-	}
-
-	// Once a peer converges, the wheel stops scheduling for it. Re-validation
-	// (and the resulting black hole detection) is driven by connection_manager
-	// via MaybeProbeAsTest at its natural test cadence, so a converged peer
-	// has nothing for the wheel to do until OnRoam or a tunnel down/up cycle
-	// triggers a fresh search.
-	if p.converged {
-		p.mu.Unlock()
-		return
-	}
-
-	ceiling := int(m.ceiling.Load())
-	var size int
-	switch {
-	case p.firstProbe:
-		// Probe the ceiling directly. If the path supports it (the common case
-		// when an operator has explicitly configured tun.max_mtu), we converge
-		// in one round trip. If it fails, the standard binary search resumes
-		// on the next tick from the (low, ceiling) bounds.
-		size = ceiling
-		p.firstProbe = false
-	case p.losses > 0 && p.inFlightSize != 0:
-		size = p.inFlightSize
-	default:
-		size = p.midpoint()
-	}
-	if size < pmtudFloor {
-		size = pmtudFloor
-	}
-	if size > ceiling {
-		size = ceiling
-	}
-
-	id := rand.Uint32()
-	p.inFlightSize = size
-	p.inFlightID = id
-	overhead := p.overhead()
-	p.mu.Unlock()
-
-	m.sendProbe(hi, size, id, overhead)
-	m.wheel.Add(localIdx, pmtudProbeInterval)
-}
-
-// sendProbe builds an MTUDProbeRequest payload that will produce an outer IP
-// packet of approximately `outerSize` bytes, then sends it.
-func (m *pmtudManager) sendProbe(hi *HostInfo, outerSize int, id uint32, overhead int) {
-	payloadLen := outerSize - overhead
-	if payloadLen < 8 {
-		payloadLen = 8
-	}
-	p := make([]byte, payloadLen)
-	binary.BigEndian.PutUint32(p[0:4], pmtudMagic)
-	binary.BigEndian.PutUint32(p[4:8], id)
-	// remaining bytes are zero-padding
-
-	nb := make([]byte, 12)
-	out := make([]byte, outerSize+128) // headroom for header/tag/relay framing
-	m.intf.SendMessageToHostInfo(header.Test, header.MTUDProbeRequest, hi, p, nb, out)
-}

relay_manager.go

@@ -18,7 +18,6 @@ type relayManager struct {
 	l       *slog.Logger
 	hostmap *HostMap
 	amRelay atomic.Bool
-	useRelays atomic.Bool
 }
 
 func NewRelayManager(ctx context.Context, l *slog.Logger, hostmap *HostMap, c *config.C) *relayManager {
@@ -37,10 +36,8 @@ func NewRelayManager(ctx context.Context, l *slog.Logger, hostmap *HostMap, c *c
 }
 
 func (rm *relayManager) reload(c *config.C, initial bool) error {
-	if initial || c.HasChanged("relay.am_relay") || c.HasChanged("relay.use_relays") {
+	if initial || c.HasChanged("relay.am_relay") {
-		amRelay := c.GetBool("relay.am_relay", false)
+		rm.setAmRelay(c.GetBool("relay.am_relay", false))
-		rm.amRelay.Store(amRelay)
-		rm.useRelays.Store(c.GetBool("relay.use_relays", true) && !amRelay)
 	}
 	return nil
 }
@@ -49,157 +46,8 @@ func (rm *relayManager) GetAmRelay() bool {
 	return rm.amRelay.Load()
 }
 
-func (rm *relayManager) GetUseRelays() bool {
+func (rm *relayManager) setAmRelay(v bool) {
-	return rm.useRelays.Load()
+	rm.amRelay.Store(v)
-}
-
-// StartRelays drives the relay-establishment side of an outbound handshake attempt.
-// For each candidate relay it either kicks off a handshake to the relay, sends a CreateRelayRequest, retransmits
-// one that may have been lost, or, once the relay is Established, forwards the in-progress
-// stage 0 handshake packet for vpnIp through it.
-func (rm *relayManager) StartRelays(f *Interface, vpnIp netip.Addr, hostinfo *HostInfo, stage0 []byte) {
-	if !rm.GetUseRelays() || len(hostinfo.remotes.relays) == 0 {
-		return
-	}
-
-	hostinfo.logger(rm.l).Info("Attempt to relay through hosts", "relays", hostinfo.remotes.relays)
-	// Send a RelayRequest to all known Relay IP's
-	for _, relay := range hostinfo.remotes.relays {
-		// Don't relay through the host I'm trying to connect to
-		if relay == vpnIp {
-			continue
-		}
-
-		// Don't relay to myself
-		if f.myVpnAddrsTable.Contains(relay) {
-			continue
-		}
-
-		relayHostInfo := rm.hostmap.QueryVpnAddr(relay)
-		if relayHostInfo == nil || !relayHostInfo.remote.IsValid() {
-			hostinfo.logger(rm.l).Info("Establish tunnel to relay target", "relay", relay.String())
-			f.Handshake(relay)
-			continue
-		}
-		// Check the relay HostInfo to see if we already established a relay through
-		existingRelay, ok := relayHostInfo.relayState.QueryRelayForByIp(vpnIp)
-		if !ok {
-			// No relays exist or requested yet.
-			if relayHostInfo.remote.IsValid() {
-				idx, err := AddRelay(rm.l, relayHostInfo, rm.hostmap, vpnIp, nil, TerminalType, Requested)
-				if err != nil {
-					hostinfo.logger(rm.l).Info("Failed to add relay to hostmap", "relay", relay.String(), "error", err)
-				}
-
-				m := NebulaControl{
-					Type:                NebulaControl_CreateRelayRequest,
-					InitiatorRelayIndex: idx,
-				}
-
-				switch relayHostInfo.GetCert().Certificate.Version() {
-				case cert.Version1:
-					if !f.myVpnAddrs[0].Is4() {
-						hostinfo.logger(rm.l).Error("can not establish v1 relay with a v6 network because the relay is not running a current nebula version")
-						continue
-					}
-
-					if !vpnIp.Is4() {
-						hostinfo.logger(rm.l).Error("can not establish v1 relay with a v6 remote network because the relay is not running a current nebula version")
-						continue
-					}
-
-					b := f.myVpnAddrs[0].As4()
-					m.OldRelayFromAddr = binary.BigEndian.Uint32(b[:])
-					b = vpnIp.As4()
-					m.OldRelayToAddr = binary.BigEndian.Uint32(b[:])
-				case cert.Version2:
-					m.RelayFromAddr = netAddrToProtoAddr(f.myVpnAddrs[0])
-					m.RelayToAddr = netAddrToProtoAddr(vpnIp)
-				default:
-					hostinfo.logger(rm.l).Error("Unknown certificate version found while creating relay")
-					continue
-				}
-
-				msg, err := m.Marshal()
-				if err != nil {
-					hostinfo.logger(rm.l).Error("Failed to marshal Control message to create relay", "error", err)
-				} else {
-					f.SendMessageToHostInfo(header.Control, 0, relayHostInfo, msg, make([]byte, 12), make([]byte, mtu))
-					rm.l.Info("send CreateRelayRequest",
-						"relayFrom", f.myVpnAddrs[0],
-						"relayTo", vpnIp,
-						"initiatorRelayIndex", idx,
-						"relay", relay,
-					)
-				}
-			}
-			continue
-		}
-
-		switch existingRelay.State {
-		case Established:
-			hostinfo.logger(rm.l).Info("Send handshake via relay", "relay", relay.String())
-			f.SendVia(relayHostInfo, existingRelay, stage0, make([]byte, 12), make([]byte, mtu), false)
-		case Disestablished:
-			// Mark this relay as 'requested'
-			relayHostInfo.relayState.UpdateRelayForByIpState(vpnIp, Requested)
-			fallthrough
-		case Requested:
-			hostinfo.logger(rm.l).Info("Re-send CreateRelay request", "relay", relay.String())
-			// Re-send the CreateRelay request, in case the previous one was lost.
-			m := NebulaControl{
-				Type:                NebulaControl_CreateRelayRequest,
-				InitiatorRelayIndex: existingRelay.LocalIndex,
-			}
-
-			switch relayHostInfo.GetCert().Certificate.Version() {
-			case cert.Version1:
-				if !f.myVpnAddrs[0].Is4() {
-					hostinfo.logger(rm.l).Error("can not establish v1 relay with a v6 network because the relay is not running a current nebula version")
-					continue
-				}
-
-				if !vpnIp.Is4() {
-					hostinfo.logger(rm.l).Error("can not establish v1 relay with a v6 remote network because the relay is not running a current nebula version")
-					continue
-				}
-
-				b := f.myVpnAddrs[0].As4()
-				m.OldRelayFromAddr = binary.BigEndian.Uint32(b[:])
-				b = vpnIp.As4()
-				m.OldRelayToAddr = binary.BigEndian.Uint32(b[:])
-			case cert.Version2:
-				m.RelayFromAddr = netAddrToProtoAddr(f.myVpnAddrs[0])
-				m.RelayToAddr = netAddrToProtoAddr(vpnIp)
-			default:
-				hostinfo.logger(rm.l).Error("Unknown certificate version found while creating relay")
-				continue
-			}
-			msg, err := m.Marshal()
-			if err != nil {
-				hostinfo.logger(rm.l).Error("Failed to marshal Control message to create relay", "error", err)
-			} else {
-				// This must send over the hostinfo, not over hm.Hosts[ip]
-				f.SendMessageToHostInfo(header.Control, 0, relayHostInfo, msg, make([]byte, 12), make([]byte, mtu))
-				rm.l.Info("send CreateRelayRequest",
-					"relayFrom", f.myVpnAddrs[0],
-					"relayTo", vpnIp,
-					"initiatorRelayIndex", existingRelay.LocalIndex,
-					"relay", relay,
-				)
-			}
-		case PeerRequested:
-			// PeerRequested only occurs in Forwarding relays, not Terminal relays, and this is a Terminal relay case.
-			fallthrough
-		default:
-			hostinfo.logger(rm.l).Error("Relay unexpected state",
-				"vpnIp", vpnIp,
-				"state", existingRelay.State,
-				"relay", relay,
-			)
-
-		}
-	}
 }
 
 // AddRelay finds an available relay index on the hostmap, and associates the relay info with it.

udp/conn.go

@@ -8,6 +8,12 @@ import (
 
 const MTU = 9001
 
+// MaxWriteBatch is the largest batch any Conn.WriteBatch implementation is
+// required to accept. Callers SHOULD NOT pass more than this per call; Linux
+// backends preallocate sendmmsg scratch sized to this value, so exceeding it
+// only costs a chunked retry.
+const MaxWriteBatch = 128
+
 type EncReader func(
 	addr netip.AddrPort,
 	payload []byte,
@@ -16,16 +22,21 @@ type EncReader func(
 type Conn interface {
 	Rebind() error
 	LocalAddr() (netip.AddrPort, error)
-	ListenOut(r EncReader) error
+	// ListenOut invokes r for each received packet. On batch-capable
+	// backends (recvmmsg), flush is called after each batch is fully
+	// delivered — callers use it to flush per-batch accumulators such as
+	// TUN write coalescers. Single-packet backends call flush after each
+	// packet. flush must not be nil.
+	ListenOut(r EncReader, flush func()) error
 	WriteTo(b []byte, addr netip.AddrPort) error
+	// WriteBatch sends a contiguous batch of packets, each with its own
+	// destination. bufs and addrs must have the same length. Linux uses
+	// sendmmsg(2) for a single syscall; other backends fall back to a
+	// WriteTo loop. Returns on the first error; callers may observe a
+	// partial send if some packets went out before the error.
+	WriteBatch(bufs [][]byte, addrs []netip.AddrPort) error
 	ReloadConfig(c *config.C)
 	SupportsMultipleReaders() bool
-	// EnablePathMTUDiscovery sets the don't-fragment bit on outgoing packets for
-	// this socket. Called by the pmtud manager when PMTUD is enabled. A no-op on
-	// platforms that don't support it; nebula's default behavior (no DF, kernel
-	// fragmentation allowed) is preserved on those platforms and on this one when
-	// PMTUD is disabled.
-	EnablePathMTUDiscovery() error
 	Close() error
 }
 
@@ -37,7 +48,7 @@ func (NoopConn) Rebind() error {
 func (NoopConn) LocalAddr() (netip.AddrPort, error) {
 	return netip.AddrPort{}, nil
 }
-func (NoopConn) ListenOut(_ EncReader) error {
+func (NoopConn) ListenOut(_ EncReader, _ func()) error {
 	return nil
 }
 func (NoopConn) SupportsMultipleReaders() bool {
@@ -46,12 +57,12 @@ func (NoopConn) SupportsMultipleReaders() bool {
 func (NoopConn) WriteTo(_ []byte, _ netip.AddrPort) error {
 	return nil
 }
+func (NoopConn) WriteBatch(_ [][]byte, _ []netip.AddrPort) error {
+	return nil
+}
 func (NoopConn) ReloadConfig(_ *config.C) {
 	return
 }
-func (NoopConn) EnablePathMTUDiscovery() error {
-	return nil
-}
 func (NoopConn) Close() error {
 	return nil
 }

udp/udp_android.go

@@ -44,17 +44,3 @@ func NewListenConfig(multi bool) net.ListenConfig {
 func (u *GenericConn) Rebind() error {
 	return nil
 }
-
-// EnablePathMTUDiscovery sets the don't-fragment bit on outbound packets.
-// Android is Linux underneath, so we use IP_PMTUDISC_PROBE (kernel sets DF but
-// does not consume incoming ICMP frag-needed for its PMTU cache; the manager
-// drives discovery via authenticated probes).
-func (u *GenericConn) EnablePathMTUDiscovery() error {
-	v4 := u.isV4Socket()
-	return u.controlFD(func(fd uintptr) error {
-		if v4 {
-			return unix.SetsockoptInt(int(fd), unix.IPPROTO_IP, unix.IP_MTU_DISCOVER, unix.IP_PMTUDISC_PROBE)
-		}
-		return unix.SetsockoptInt(int(fd), unix.IPPROTO_IPV6, unix.IPV6_MTU_DISCOVER, unix.IPV6_PMTUDISC_PROBE)
-	})
-}

udp/udp_bsd.go

@@ -47,8 +47,3 @@ func NewListenConfig(multi bool) net.ListenConfig {
 func (u *GenericConn) Rebind() error {
 	return nil
 }
-
-// EnablePathMTUDiscovery is split into per-OS files: udp_freebsd.go handles
-// FreeBSD (which has both IP_DONTFRAG and IPV6_DONTFRAG in the unix package);
-// udp_openbsd.go handles OpenBSD (v6 only; the kernel doesn't expose a v4 DF
-// sockopt).

udp/udp_darwin.go

@@ -140,6 +140,15 @@ func (u *StdConn) WriteTo(b []byte, ap netip.AddrPort) error {
 	}
 }
 
+func (u *StdConn) WriteBatch(bufs [][]byte, addrs []netip.AddrPort) error {
+	for i, b := range bufs {
+		if err := u.WriteTo(b, addrs[i]); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
 func (u *StdConn) LocalAddr() (netip.AddrPort, error) {
 	a := u.UDPConn.LocalAddr()
 
@@ -165,7 +174,7 @@ func NewUDPStatsEmitter(udpConns []Conn) func() {
 	return func() {}
 }
 
-func (u *StdConn) ListenOut(r EncReader) error {
+func (u *StdConn) ListenOut(r EncReader, flush func()) error {
 	buffer := make([]byte, MTU)
 
 	for {
@@ -180,6 +189,7 @@ func (u *StdConn) ListenOut(r EncReader) error {
 		}
 
 		r(netip.AddrPortFrom(rua.Addr().Unmap(), rua.Port()), buffer[:n])
+		flush()
 	}
 }
 
@@ -187,17 +197,6 @@ func (u *StdConn) SupportsMultipleReaders() bool {
 	return false
 }
 
-// EnablePathMTUDiscovery sets the don't-fragment bit on every outbound packet.
-// On darwin we use IP_DONTFRAG (v4) / IPV6_DONTFRAG (v6). The kernel will return
-// EMSGSIZE for sends that exceed the local interface MTU; ICMP-driven PMTU
-// updates from upstream routers are processed by the kernel as usual.
-func (u *StdConn) EnablePathMTUDiscovery() error {
-	if u.isV4 {
-		return syscall.SetsockoptInt(int(u.sysFd), syscall.IPPROTO_IP, unix.IP_DONTFRAG, 1)
-	}
-	return syscall.SetsockoptInt(int(u.sysFd), syscall.IPPROTO_IPV6, unix.IPV6_DONTFRAG, 1)
-}
-
 func (u *StdConn) Rebind() error {
 	var err error
 	if u.isV4 {

udp/udp_freebsd.go

@@ -1,25 +0,0 @@
-//go:build freebsd && !e2e_testing
-// +build freebsd,!e2e_testing
-
-package udp
-
-import (
-	"golang.org/x/sys/unix"
-)
-
-// EnablePathMTUDiscovery sets the don't-fragment bit on outbound packets.
-// FreeBSD exposes IP_DONTFRAG (v4) and IPV6_DONTFRAG (v6) in golang.org/x/sys/unix.
-// Unlike Linux, BSDs don't have an explicit "don't consume incoming ICMP
-// frag-needed" knob for unconnected UDP sockets; the kernel's PMTU cache will
-// be updated from ICMP, which is benign for our usage (the cache only affects
-// what EMSGSIZE gets surfaced for; the manager drives its own discovery via
-// authenticated probes).
-func (u *GenericConn) EnablePathMTUDiscovery() error {
-	v4 := u.isV4Socket()
-	return u.controlFD(func(fd uintptr) error {
-		if v4 {
-			return unix.SetsockoptInt(int(fd), unix.IPPROTO_IP, unix.IP_DONTFRAG, 1)
-		}
-		return unix.SetsockoptInt(int(fd), unix.IPPROTO_IPV6, unix.IPV6_DONTFRAG, 1)
-	})
-}

udp/udp_generic.go

@@ -44,6 +44,15 @@ func (u *GenericConn) WriteTo(b []byte, addr netip.AddrPort) error {
 	return err
 }
 
+func (u *GenericConn) WriteBatch(bufs [][]byte, addrs []netip.AddrPort) error {
+	for i, b := range bufs {
+		if _, err := u.UDPConn.WriteToUDPAddrPort(b, addrs[i]); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
 func (u *GenericConn) LocalAddr() (netip.AddrPort, error) {
 	a := u.UDPConn.LocalAddr()
 
@@ -73,7 +82,7 @@ type rawMessage struct {
 	Len uint32
 }
 
-func (u *GenericConn) ListenOut(r EncReader) error {
+func (u *GenericConn) ListenOut(r EncReader, flush func()) error {
 	buffer := make([]byte, MTU)
 
 	var lastRecvErr time.Time
@@ -94,47 +103,10 @@ func (u *GenericConn) ListenOut(r EncReader) error {
 		}
 
 		r(netip.AddrPortFrom(rua.Addr().Unmap(), rua.Port()), buffer[:n])
+		flush()
 	}
 }
 
 func (u *GenericConn) SupportsMultipleReaders() bool {
 	return false
 }
-
-// EnablePathMTUDiscovery is implemented per-platform alongside Rebind, in
-// udp_android.go / udp_bsd.go / udp_netbsd.go / udp_windows.go.
-
-// controlFD invokes f with the underlying UDP socket file descriptor (or
-// handle, on Windows). Used by platform files for setsockopt calls that the
-// stdlib net.UDPConn does not expose directly.
-func (u *GenericConn) controlFD(f func(fd uintptr) error) error {
-	rc, err := u.UDPConn.SyscallConn()
-	if err != nil {
-		return err
-	}
-	var sockErr error
-	err = rc.Control(func(fd uintptr) {
-		sockErr = f(fd)
-	})
-	if err != nil {
-		return err
-	}
-	return sockErr
-}
-
-// isV4Socket reports whether the local bind address looks like an IPv4 socket.
-// Used by EnablePathMTUDiscovery to pick IPPROTO_IP vs IPPROTO_IPV6 socket
-// options. Assumes pure-v4 or pure-v6 sockets; a dual-stack v6 socket bound to
-// :: will be treated as v6 (correct: setting IPV6_DONTFRAG covers v4-mapped
-// traffic too on most stacks).
-func (u *GenericConn) isV4Socket() bool {
-	la := u.UDPConn.LocalAddr()
-	if la == nil {
-		return false
-	}
-	ua, ok := la.(*net.UDPAddr)
-	if !ok {
-		return false
-	}
-	return ua.IP.To4() != nil
-}

udp/udp_linux.go

@@ -24,6 +24,43 @@ type StdConn struct {
 	isV4    bool
 	l       *slog.Logger
 	batch   int
+
+	// sendmmsg scratch. Each queue has its own StdConn, so no locking is
+	// needed. Sized to MaxWriteBatch at construction; WriteBatch chunks
+	// larger inputs.
+	writeMsgs  []rawMessage
+	writeIovs  []iovec
+	writeNames [][]byte
+
+	// Per-entry UDP_SEGMENT cmsg scratch. writeCmsg is one contiguous slab
+	// of MaxWriteBatch * writeCmsgSpace bytes; each entry's cmsg header is
+	// pre-filled once in prepareWriteMessages. WriteBatch only rewrites the
+	// 2-byte gso_size payload (and toggles Hdr.Control on/off) per call.
+	writeCmsg      []byte
+	writeCmsgSpace int
+
+	// writeEntryEnd[e] is the bufs index *after* the last packet packed
+	// into mmsghdr entry e. Used to rewind `i` on partial sendmmsg success.
+	writeEntryEnd []int
+
+	// Preallocated closure + in/out slots for sendmmsg, so the hot path
+	// does not heap-allocate a fresh closure per call.
+	writeChunk int
+	writeSent  int
+	writeErrno syscall.Errno
+	writeFunc  func(fd uintptr) bool
+
+	// UDP GSO (sendmsg with UDP_SEGMENT cmsg) support. gsoSupported is
+	// probed once at socket creation. When true, WriteSegmented takes a
+	// single-syscall GSO path; otherwise it falls back to a WriteTo loop.
+	gsoSupported bool
+
+	// UDP GRO (recvmsg with UDP_GRO cmsg) support. groSupported is probed
+	// once at socket creation. When true, listenOutBatch allocates larger
+	// RX buffers and a per-entry cmsg slot so the kernel can coalesce
+	// consecutive same-flow datagrams into a single recvmmsg entry; the
+	// delivered cmsg carries the gso_size used to split them back apart.
+	groSupported bool
 }
 
 func setReusePort(network, address string, c syscall.RawConn) error {
@@ -70,22 +107,100 @@ func NewListener(l *slog.Logger, ip netip.Addr, port int, multi bool, batch int)
 	}
 	out.isV4 = af == unix.AF_INET
 
+	out.prepareWriteMessages(MaxWriteBatch)
+	out.writeFunc = out.sendmmsgRawWrite
+
+	out.prepareGSO()
+	// GRO delivers coalesced superpackets that need a cmsg to split back
+	// into segments. The single-packet RX path uses ReadFromUDPAddrPort
+	// and cannot see that cmsg, so only enable GRO for the batch path.
+	if batch > 1 {
+		out.prepareGRO()
+	}
+
 	return out, nil
 }
 
-// EnablePathMTUDiscovery sets IP_MTU_DISCOVER=IP_PMTUDISC_PROBE (IPV6 equivalent
+// prepareWriteMessages allocates one mmsghdr/iovec/sockaddr/cmsg scratch
-// for v6 sockets). This sets the don't-fragment bit on every outbound packet but
+// slot per sendmmsg entry. The iovec slab is sized to the same n so a
-// tells the kernel not to consume incoming ICMP frag-needed for its own PMTU
+// single entry can fan out to up to n iovecs (needed for UDP_SEGMENT runs
-// cache; we drive PMTU discovery from the application via authenticated probes
+// that coalesce consecutive bufs into one entry). Hdr.Iov / Hdr.Iovlen /
-// (RFC 8899). Called by the pmtud manager when PMTUD is enabled. Without this
+// Hdr.Control / Hdr.Controllen are wired per call since each entry can
-// call the socket retains nebula's historical behavior (no DF, kernel may
+// span a variable number of iovecs and may or may not carry a cmsg.
-// fragment), preserving compatibility with deployments that depend on UDP
+func (u *StdConn) prepareWriteMessages(n int) {
-// fragmentation.
+	u.writeMsgs = make([]rawMessage, n)
-func (u *StdConn) EnablePathMTUDiscovery() error {
+	u.writeIovs = make([]iovec, n)
-	if u.isV4 {
+	u.writeNames = make([][]byte, n)
-		return u.setSockOptIPInt(unix.IPPROTO_IP, unix.IP_MTU_DISCOVER, unix.IP_PMTUDISC_PROBE)
+	u.writeEntryEnd = make([]int, n)
+
+	u.writeCmsgSpace = unix.CmsgSpace(2)
+	u.writeCmsg = make([]byte, n*u.writeCmsgSpace)
+	for k := 0; k < n; k++ {
+		off := k * u.writeCmsgSpace
+		h := (*unix.Cmsghdr)(unsafe.Pointer(&u.writeCmsg[off]))
+		h.Level = unix.SOL_UDP
+		h.Type = unix.UDP_SEGMENT
+		setCmsgLen(h, unix.CmsgLen(2))
 	}
-	return u.setSockOptIPInt(unix.IPPROTO_IPV6, unix.IPV6_MTU_DISCOVER, unix.IPV6_PMTUDISC_PROBE)
+
+	for i := range u.writeMsgs {
+		u.writeNames[i] = make([]byte, unix.SizeofSockaddrInet6)
+		u.writeMsgs[i].Hdr.Name = &u.writeNames[i][0]
+	}
+}
+
+// maxGSOSegments caps the per-sendmsg GSO fan-out. Linux kernels have
+// historically capped UDP_MAX_SEGMENTS at 64; newer kernels raise it to 128
+// but we stay conservative so the same code works everywhere.
+const maxGSOSegments = 64
+
+// maxGSOBytes bounds the total payload per sendmsg() when UDP_SEGMENT is
+// set. The kernel stitches all iovecs into a single skb whose length the
+// UDP length field can represent, and also enforces sk_gso_max_size (which
+// on most devices is 65536). We use 65535 so ciphertext + headers always
+// fits, avoiding EMSGSIZE on large TSO superpackets.
+const maxGSOBytes = 65535
+
+// prepareGSO probes UDP_SEGMENT support
+func (u *StdConn) prepareGSO() {
+	var probeErr error
+	if err := u.rawConn.Control(func(fd uintptr) {
+		probeErr = unix.SetsockoptInt(int(fd), unix.IPPROTO_UDP, unix.UDP_SEGMENT, 0)
+	}); err != nil {
+		return
+	}
+	if probeErr != nil {
+		return
+	}
+	u.gsoSupported = true
+}
+
+// udpGROBufferSize sizes the per-entry recvmmsg buffer when UDP_GRO is on.
+// The kernel stitches a run of same-flow datagrams into a single skb whose
+// length is bounded by sk_gso_max_size (typically 65535); anything larger
+// would be MSG_TRUNCed. We use the maximum representable UDP length so a
+// full superpacket always lands intact.
+const udpGROBufferSize = 65535
+
+// udpGROCmsgPayload is the size of the UDP_GRO cmsg data delivered by the
+// kernel: a single int (gso_size in bytes). See udp_cmsg_recv() in
+// net/ipv4/udp.c.
+const udpGROCmsgPayload = 4
+
+// prepareGRO turns on UDP_GRO so the kernel coalesces consecutive same-flow
+// datagrams into one recvmmsg entry, with a cmsg carrying the gso_size used
+// to split them back apart on the application side.
+func (u *StdConn) prepareGRO() {
+	var probeErr error
+	if err := u.rawConn.Control(func(fd uintptr) {
+		probeErr = unix.SetsockoptInt(int(fd), unix.IPPROTO_UDP, unix.UDP_GRO, 1)
+	}); err != nil {
+		return
+	}
+	if probeErr != nil {
+		return
+	}
+	u.groSupported = true
 }
 
 func (u *StdConn) SupportsMultipleReaders() bool {
@@ -125,21 +240,6 @@ func (u *StdConn) setSockOptInt(opt int, n int) error {
 	return opErr
 }
 
-// setSockOptIPInt sets a socket option at a non-SOL_SOCKET level (e.g. IPPROTO_IP).
-func (u *StdConn) setSockOptIPInt(level, opt, n int) error {
-	if u.rawConn == nil {
-		return fmt.Errorf("no UDP connection")
-	}
-	var opErr error
-	err := u.rawConn.Control(func(fd uintptr) {
-		opErr = unix.SetsockoptInt(int(fd), level, opt, n)
-	})
-	if err != nil {
-		return err
-	}
-	return opErr
-}
-
 func (u *StdConn) SetRecvBuffer(n int) error {
 	return u.setSockOptInt(unix.SO_RCVBUFFORCE, n)
 }
@@ -201,7 +301,7 @@ func recvmmsg(fd uintptr, msgs []rawMessage) (int, bool, error) {
 	return int(n), true, nil
 }
 
-func (u *StdConn) listenOutSingle(r EncReader) error {
+func (u *StdConn) listenOutSingle(r EncReader, flush func()) error {
 	var err error
 	var n int
 	var from netip.AddrPort
@@ -214,15 +314,22 @@ func (u *StdConn) listenOutSingle(r EncReader) error {
 		}
 		from = netip.AddrPortFrom(from.Addr().Unmap(), from.Port())
 		r(from, buffer[:n])
+		flush()
 	}
 }
 
-func (u *StdConn) listenOutBatch(r EncReader) error {
+func (u *StdConn) listenOutBatch(r EncReader, flush func()) error {
 	var ip netip.Addr
 	var n int
 	var operr error
 
-	msgs, buffers, names := u.PrepareRawMessages(u.batch)
+	bufSize := MTU
+	cmsgSpace := 0
+	if u.groSupported {
+		bufSize = udpGROBufferSize
+		cmsgSpace = unix.CmsgSpace(udpGROCmsgPayload)
+	}
+	msgs, buffers, names, _ := u.PrepareRawMessages(u.batch, bufSize, cmsgSpace)
 
 	//reader needs to capture variables from this function, since it's used as a lambda with rawConn.Read
 	//defining it outside the loop so it gets re-used
@@ -232,6 +339,11 @@ func (u *StdConn) listenOutBatch(r EncReader) error {
 	}
 
 	for {
+		if cmsgSpace > 0 {
+			for i := range msgs {
+				setMsgControllen(&msgs[i].Hdr, cmsgSpace)
+			}
+		}
 		err := u.rawConn.Read(reader)
 		if err != nil {
 			return err
@@ -247,16 +359,72 @@ func (u *StdConn) listenOutBatch(r EncReader) error {
 			} else {
 				ip, _ = netip.AddrFromSlice(names[i][8:24])
 			}
-			r(netip.AddrPortFrom(ip.Unmap(), binary.BigEndian.Uint16(names[i][2:4])), buffers[i][:msgs[i].Len])
+			from := netip.AddrPortFrom(ip.Unmap(), binary.BigEndian.Uint16(names[i][2:4]))
+			payload := buffers[i][:msgs[i].Len]
+
+			segSize := 0
+			if u.groSupported {
+				segSize = parseUDPGRO(&msgs[i].Hdr)
 			}
+			if segSize <= 0 || segSize >= len(payload) {
+				// No coalescing happened (or a lone datagram).
+				r(from, payload)
+				continue
+			}
+			// GRO superpacket: the kernel guarantees every segment is
+			// exactly segSize bytes except for the final one, which may be
+			// short.
+			for off := 0; off < len(payload); off += segSize {
+				end := off + segSize
+				if end > len(payload) {
+					end = len(payload)
+				}
+				r(from, payload[off:end])
+			}
+		}
+		// End-of-batch: let callers (e.g. TUN write coalescer) flush any
+		// state they accumulated across this batch.
+		flush()
 	}
 }
 
-func (u *StdConn) ListenOut(r EncReader) error {
+// parseUDPGRO walks the control buffer on hdr looking for a SOL_UDP/UDP_GRO
+// cmsg and returns the gso_size (bytes per coalesced segment) it carries.
+// Returns 0 when no UDP_GRO cmsg is present, which is the normal case for
+// lone datagrams that the kernel did not coalesce.
+func parseUDPGRO(hdr *msghdr) int {
+	controllen := int(hdr.Controllen)
+	if controllen < unix.SizeofCmsghdr || hdr.Control == nil {
+		return 0
+	}
+	ctrl := unsafe.Slice(hdr.Control, controllen)
+	off := 0
+	for off+unix.SizeofCmsghdr <= len(ctrl) {
+		ch := (*unix.Cmsghdr)(unsafe.Pointer(&ctrl[off]))
+		clen := int(ch.Len)
+		if clen < unix.SizeofCmsghdr || off+clen > len(ctrl) {
+			return 0
+		}
+		if ch.Level == unix.SOL_UDP && ch.Type == unix.UDP_GRO {
+			dataOff := off + unix.CmsgLen(0)
+			if dataOff+udpGROCmsgPayload <= len(ctrl) {
+				return int(int32(binary.NativeEndian.Uint32(ctrl[dataOff : dataOff+udpGROCmsgPayload])))
+			}
+			return 0
+		}
+		// Advance by the aligned cmsg space. CmsgSpace(n) is the stride
+		// from one header to the next (len aligned up to the platform's
+		// cmsg alignment).
+		off += unix.CmsgSpace(clen - unix.CmsgLen(0))
+	}
+	return 0
+}
+
+func (u *StdConn) ListenOut(r EncReader, flush func()) error {
 	if u.batch == 1 {
-		return u.listenOutSingle(r)
+		return u.listenOutSingle(r, flush)
 	} else {
-		return u.listenOutBatch(r)
+		return u.listenOutBatch(r, flush)
 	}
 }
 
@@ -265,6 +433,209 @@ func (u *StdConn) WriteTo(b []byte, ip netip.AddrPort) error {
 	return err
 }
 
+// WriteBatch sends bufs via sendmmsg(2) using the preallocated scratch on
+// StdConn. Consecutive packets to the same destination with matching segment
+// sizes (all but possibly the last) are coalesced into a single mmsghdr entry
+// carrying a UDP_SEGMENT cmsg, so one syscall can mix runs of GSO superpackets
+// with plain one-off datagrams. Without GSO support every packet is its own
+// entry, matching the prior behaviour.
+//
+// Chunks larger than the scratch are processed across multiple syscalls. If
+// sendmmsg returns a fatal error before any entry is sent we fall back to
+// per-packet WriteTo for that chunk so the caller still gets best-effort
+// delivery.
+func (u *StdConn) WriteBatch(bufs [][]byte, addrs []netip.AddrPort) error {
+	if len(bufs) != len(addrs) {
+		return fmt.Errorf("WriteBatch: len(bufs)=%d != len(addrs)=%d", len(bufs), len(addrs))
+	}
+
+	i := 0
+	for i < len(bufs) {
+		baseI := i
+		entry := 0
+		iovIdx := 0
+
+		for entry < len(u.writeMsgs) && i < len(bufs) {
+			iovBudget := len(u.writeIovs) - iovIdx
+			if iovBudget < 1 {
+				break
+			}
+			runLen, segSize := u.planRun(bufs, addrs, i, iovBudget)
+			if runLen == 0 {
+				break
+			}
+
+			for k := 0; k < runLen; k++ {
+				b := bufs[i+k]
+				if len(b) == 0 {
+					u.writeIovs[iovIdx+k].Base = nil
+					setIovLen(&u.writeIovs[iovIdx+k], 0)
+				} else {
+					u.writeIovs[iovIdx+k].Base = &b[0]
+					setIovLen(&u.writeIovs[iovIdx+k], len(b))
+				}
+			}
+
+			nlen, err := writeSockaddr(u.writeNames[entry], addrs[i], u.isV4)
+			if err != nil {
+				return err
+			}
+
+			hdr := &u.writeMsgs[entry].Hdr
+			hdr.Iov = &u.writeIovs[iovIdx]
+			setMsgIovlen(hdr, runLen)
+			hdr.Namelen = uint32(nlen)
+
+			if runLen >= 2 {
+				off := entry * u.writeCmsgSpace
+				dataOff := off + unix.CmsgLen(0)
+				binary.NativeEndian.PutUint16(u.writeCmsg[dataOff:dataOff+2], uint16(segSize))
+				hdr.Control = &u.writeCmsg[off]
+				setMsgControllen(hdr, u.writeCmsgSpace)
+			} else {
+				hdr.Control = nil
+				setMsgControllen(hdr, 0)
+			}
+
+			i += runLen
+			iovIdx += runLen
+			u.writeEntryEnd[entry] = i
+			entry++
+		}
+
+		if entry == 0 {
+			return fmt.Errorf("sendmmsg: no progress")
+		}
+
+		sent, serr := u.sendmmsg(entry)
+		if serr != nil && sent <= 0 {
+			// Nothing went out for this chunk; fall back to WriteTo for each
+			// packet that was queued this iteration.
+			for k := baseI; k < i; k++ {
+				if werr := u.WriteTo(bufs[k], addrs[k]); werr != nil {
+					return werr
+				}
+			}
+			continue
+		}
+		if sent == 0 {
+			return fmt.Errorf("sendmmsg made no progress")
+		}
+		// Rewind i to the end of the last successfully sent entry. For a
+		// full-success send this leaves i unchanged; for a partial send it
+		// replays the remainder on the next outer-loop iteration.
+		i = u.writeEntryEnd[sent-1]
+	}
+	return nil
+}
+
+// planRun groups consecutive packets starting at `start` that can be sent as
+// a single UDP GSO superpacket (one sendmmsg entry with UDP_SEGMENT cmsg).
+// A run of length 1 means the entry carries no cmsg and the kernel treats
+// it as a plain datagram. Returns the run length and the per-segment size
+// (which equals len(bufs[start])). Without GSO support every call returns
+// runLen=1.
+func (u *StdConn) planRun(bufs [][]byte, addrs []netip.AddrPort, start, iovBudget int) (int, int) {
+	if start >= len(bufs) || iovBudget < 1 {
+		return 0, 0
+	}
+	segSize := len(bufs[start])
+	if !u.gsoSupported || segSize == 0 || segSize > maxGSOBytes {
+		return 1, segSize
+	}
+	dst := addrs[start]
+	maxLen := maxGSOSegments
+	if iovBudget < maxLen {
+		maxLen = iovBudget
+	}
+	runLen := 1
+	total := segSize
+	for runLen < maxLen && start+runLen < len(bufs) {
+		nextLen := len(bufs[start+runLen])
+		if nextLen == 0 || nextLen > segSize {
+			break
+		}
+		if addrs[start+runLen] != dst {
+			break
+		}
+		if total+nextLen > maxGSOBytes {
+			break
+		}
+		total += nextLen
+		runLen++
+		if nextLen < segSize {
+			// A short packet must be the last in the run.
+			break
+		}
+	}
+	return runLen, segSize
+}
+
+// sendmmsgRawWrite is the preallocated callback passed to rawConn.Write. It
+// reads its input (u.writeChunk) and writes its outputs (u.writeSent,
+// u.writeErrno) through StdConn fields so the closure itself does not
+// capture per-call locals and therefore does not heap-allocate.
+func (u *StdConn) sendmmsgRawWrite(fd uintptr) bool {
+	r1, _, errno := unix.Syscall6(
+		unix.SYS_SENDMMSG,
+		fd,
+		uintptr(unsafe.Pointer(&u.writeMsgs[0])),
+		uintptr(u.writeChunk),
+		0,
+		0,
+		0,
+	)
+	if errno == syscall.EAGAIN || errno == syscall.EWOULDBLOCK {
+		return false
+	}
+	u.writeSent = int(r1)
+	u.writeErrno = errno
+	return true
+}
+
+func (u *StdConn) sendmmsg(n int) (int, error) {
+	u.writeChunk = n
+	u.writeSent = 0
+	u.writeErrno = 0
+	if err := u.rawConn.Write(u.writeFunc); err != nil {
+		return u.writeSent, err
+	}
+	if u.writeErrno != 0 {
+		return u.writeSent, &net.OpError{Op: "sendmmsg", Err: u.writeErrno}
+	}
+	return u.writeSent, nil
+}
+
+// writeSockaddr encodes addr into buf (which must be at least
+// SizeofSockaddrInet6 bytes). Returns the number of bytes used. If isV4 is
+// true and addr is not a v4 (or v4-in-v6) address, returns an error.
+func writeSockaddr(buf []byte, addr netip.AddrPort, isV4 bool) (int, error) {
+	ap := addr.Addr().Unmap()
+	if isV4 {
+		if !ap.Is4() {
+			return 0, ErrInvalidIPv6RemoteForSocket
+		}
+		// struct sockaddr_in: { sa_family_t(2), in_port_t(2, BE), in_addr(4), zero(8) }
+		// sa_family is host endian.
+		binary.NativeEndian.PutUint16(buf[0:2], unix.AF_INET)
+		binary.BigEndian.PutUint16(buf[2:4], addr.Port())
+		ip4 := ap.As4()
+		copy(buf[4:8], ip4[:])
+		for j := 8; j < 16; j++ {
+			buf[j] = 0
+		}
+		return unix.SizeofSockaddrInet4, nil
+	}
+	// struct sockaddr_in6: { sa_family_t(2), in_port_t(2, BE), flowinfo(4), in6_addr(16), scope_id(4) }
+	binary.NativeEndian.PutUint16(buf[0:2], unix.AF_INET6)
+	binary.BigEndian.PutUint16(buf[2:4], addr.Port())
+	binary.NativeEndian.PutUint32(buf[4:8], 0)
+	ip6 := addr.Addr().As16()
+	copy(buf[8:24], ip6[:])
+	binary.NativeEndian.PutUint32(buf[24:28], 0)
+	return unix.SizeofSockaddrInet6, nil
+}
+
 func (u *StdConn) ReloadConfig(c *config.C) {
 	b := c.GetInt("listen.read_buffer", 0)
 	if b > 0 {

udp/udp_linux_32.go

@@ -30,13 +30,18 @@ type rawMessage struct {
 	Len uint32
 }
 
-func (u *StdConn) PrepareRawMessages(n int) ([]rawMessage, [][]byte, [][]byte) {
+func (u *StdConn) PrepareRawMessages(n, bufSize, cmsgSpace int) ([]rawMessage, [][]byte, [][]byte, []byte) {
 	msgs := make([]rawMessage, n)
 	buffers := make([][]byte, n)
 	names := make([][]byte, n)
 
+	var cmsgs []byte
+	if cmsgSpace > 0 {
+		cmsgs = make([]byte, n*cmsgSpace)
+	}
+
 	for i := range msgs {
-		buffers[i] = make([]byte, MTU)
+		buffers[i] = make([]byte, bufSize)
 		names[i] = make([]byte, unix.SizeofSockaddrInet6)
 
 		vs := []iovec{
@@ -48,7 +53,28 @@ func (u *StdConn) PrepareRawMessages(n int) ([]rawMessage, [][]byte, [][]byte) {
 
 		msgs[i].Hdr.Name = &names[i][0]
 		msgs[i].Hdr.Namelen = uint32(len(names[i]))
+
+		if cmsgSpace > 0 {
+			msgs[i].Hdr.Control = &cmsgs[i*cmsgSpace]
+			msgs[i].Hdr.Controllen = uint32(cmsgSpace)
+		}
 	}
 
-	return msgs, buffers, names
+	return msgs, buffers, names, cmsgs
+}
+
+func setIovLen(v *iovec, n int) {
+	v.Len = uint32(n)
+}
+
+func setMsgIovlen(m *msghdr, n int) {
+	m.Iovlen = uint32(n)
+}
+
+func setMsgControllen(m *msghdr, n int) {
+	m.Controllen = uint32(n)
+}
+
+func setCmsgLen(h *unix.Cmsghdr, n int) {
+	h.Len = uint32(n)
 }

udp/udp_linux_64.go

@@ -33,13 +33,18 @@ type rawMessage struct {
 	Pad0 [4]byte
 }
 
-func (u *StdConn) PrepareRawMessages(n int) ([]rawMessage, [][]byte, [][]byte) {
+func (u *StdConn) PrepareRawMessages(n, bufSize, cmsgSpace int) ([]rawMessage, [][]byte, [][]byte, []byte) {
 	msgs := make([]rawMessage, n)
 	buffers := make([][]byte, n)
 	names := make([][]byte, n)
 
+	var cmsgs []byte
+	if cmsgSpace > 0 {
+		cmsgs = make([]byte, n*cmsgSpace)
+	}
+
 	for i := range msgs {
-		buffers[i] = make([]byte, MTU)
+		buffers[i] = make([]byte, bufSize)
 		names[i] = make([]byte, unix.SizeofSockaddrInet6)
 
 		vs := []iovec{
@@ -51,7 +56,28 @@ func (u *StdConn) PrepareRawMessages(n int) ([]rawMessage, [][]byte, [][]byte) {
 
 		msgs[i].Hdr.Name = &names[i][0]
 		msgs[i].Hdr.Namelen = uint32(len(names[i]))
+
+		if cmsgSpace > 0 {
+			msgs[i].Hdr.Control = &cmsgs[i*cmsgSpace]
+			msgs[i].Hdr.Controllen = uint64(cmsgSpace)
+		}
 	}
 
-	return msgs, buffers, names
+	return msgs, buffers, names, cmsgs
+}
+
+func setIovLen(v *iovec, n int) {
+	v.Len = uint64(n)
+}
+
+func setMsgIovlen(m *msghdr, n int) {
+	m.Iovlen = uint64(n)
+}
+
+func setMsgControllen(m *msghdr, n int) {
+	m.Controllen = uint64(n)
+}
+
+func setCmsgLen(h *unix.Cmsghdr, n int) {
+	h.Len = uint64(n)
 }

udp/udp_netbsd.go

@@ -46,18 +46,3 @@ func NewListenConfig(multi bool) net.ListenConfig {
 func (u *GenericConn) Rebind() error {
 	return nil
 }
-
-// EnablePathMTUDiscovery sets the don't-fragment bit on outbound packets.
-// NetBSD exposes IPV6_DONTFRAG via golang.org/x/sys/unix but the kernel does
-// not provide a socket-level knob for setting DF on v4 UDP. The only IP-layer
-// constant exposed is IP_DF, which is the wire header flag, not a sockopt.
-// quic-go skips NetBSD for the same reason. So v4 sockets stay at nebula's
-// historical behavior (kernel may fragment); v6 gets DF.
-func (u *GenericConn) EnablePathMTUDiscovery() error {
-	if u.isV4Socket() {
-		return nil
-	}
-	return u.controlFD(func(fd uintptr) error {
-		return unix.SetsockoptInt(int(fd), unix.IPPROTO_IPV6, unix.IPV6_DONTFRAG, 1)
-	})
-}

udp/udp_openbsd.go

@@ -1,23 +0,0 @@
-//go:build openbsd && !e2e_testing
-// +build openbsd,!e2e_testing
-
-package udp
-
-import (
-	"golang.org/x/sys/unix"
-)
-
-// EnablePathMTUDiscovery sets the don't-fragment bit on outbound packets.
-// OpenBSD exposes IPV6_DONTFRAG via golang.org/x/sys/unix but the kernel does
-// not provide a socket-level knob for setting DF on v4 UDP. The only IP-layer
-// constant exposed is IP_DF, which is the wire header flag, not a sockopt.
-// quic-go skips OpenBSD for the same reason. So v4 sockets stay at nebula's
-// historical behavior (kernel may fragment); v6 gets DF.
-func (u *GenericConn) EnablePathMTUDiscovery() error {
-	if u.isV4Socket() {
-		return nil
-	}
-	return u.controlFD(func(fd uintptr) error {
-		return unix.SetsockoptInt(int(fd), unix.IPPROTO_IPV6, unix.IPV6_DONTFRAG, 1)
-	})
-}

udp/udp_rio_windows.go

@@ -140,7 +140,7 @@ func (u *RIOConn) bind(l *slog.Logger, sa windows.Sockaddr) error {
 	return nil
 }
 
-func (u *RIOConn) ListenOut(r EncReader) error {
+func (u *RIOConn) ListenOut(r EncReader, flush func()) error {
 	buffer := make([]byte, MTU)
 
 	var lastRecvErr time.Time
@@ -162,6 +162,7 @@ func (u *RIOConn) ListenOut(r EncReader) error {
 		}
 
 		r(netip.AddrPortFrom(netip.AddrFrom16(rua.Addr).Unmap(), (rua.Port>>8)|((rua.Port&0xff)<<8)), buffer[:n])
+		flush()
 	}
 }
 
@@ -316,6 +317,15 @@ func (u *RIOConn) WriteTo(buf []byte, ip netip.AddrPort) error {
 	return winrio.SendEx(u.rq, dataBuffer, 1, nil, addressBuffer, nil, nil, 0, 0)
 }
 
+func (u *RIOConn) WriteBatch(bufs [][]byte, addrs []netip.AddrPort) error {
+	for i, b := range bufs {
+		if err := u.WriteTo(b, addrs[i]); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
 func (u *RIOConn) LocalAddr() (netip.AddrPort, error) {
 	sa, err := windows.Getsockname(u.sock)
 	if err != nil {
@@ -335,12 +345,6 @@ func (u *RIOConn) Rebind() error {
 	return nil
 }
 
-// EnablePathMTUDiscovery is a no-op on Windows for now. PMTUD is Linux-only in
-// the initial PoC; Windows support would set IP_DONTFRAGMENT here.
-func (u *RIOConn) EnablePathMTUDiscovery() error {
-	return nil
-}
-
 func (u *RIOConn) ReloadConfig(*config.C) {}
 
 func (u *RIOConn) Close() error {

udp/udp_tester.go

@@ -122,13 +122,23 @@ func (u *TesterConn) WriteTo(b []byte, addr netip.AddrPort) error {
 	}
 }
 
-func (u *TesterConn) ListenOut(r EncReader) error {
+func (u *TesterConn) WriteBatch(bufs [][]byte, addrs []netip.AddrPort) error {
+	for i, b := range bufs {
+		if err := u.WriteTo(b, addrs[i]); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func (u *TesterConn) ListenOut(r EncReader, flush func()) error {
 	for {
 		select {
 		case <-u.done:
 			return os.ErrClosed
 		case p := <-u.RxPackets:
 			r(p.From, p.Data)
+			flush()
 		}
 	}
 }
@@ -152,10 +162,6 @@ func (u *TesterConn) Rebind() error {
 	return nil
 }
 
-func (u *TesterConn) EnablePathMTUDiscovery() error {
-	return nil
-}
-
 func (u *TesterConn) Close() error {
 	u.closeOnce.Do(func() {
 		close(u.done)

udp/udp_windows.go

@@ -9,8 +9,6 @@ import (
 	"net"
 	"net/netip"
 	"syscall"
-
-	"golang.org/x/sys/windows"
 )
 
 func NewListener(l *slog.Logger, ip netip.Addr, port int, multi bool, batch int) (Conn, error) {
@@ -46,27 +44,3 @@ func NewListenConfig(multi bool) net.ListenConfig {
 func (u *GenericConn) Rebind() error {
 	return nil
 }
-
-// Windows IP_DONTFRAGMENT and IPV6_DONTFRAG are not exposed in the
-// golang.org/x/sys/windows package. Defined locally per the values in
-// ws2ipdef.h / ws2tcpip.h. These are stable Win32 constants that have not
-// changed since at least Windows Vista.
-const (
-	winIPDontFragment = 14
-	winIPv6DontFrag   = 14
-)
-
-// EnablePathMTUDiscovery sets the don't-fragment bit on outbound packets.
-// Windows uses IP_DONTFRAGMENT (v4) and IPV6_DONTFRAG (v6) at IPPROTO_IP /
-// IPPROTO_IPV6 respectively. Note: this only enables DF on the GenericConn
-// fallback path. The RIO path (RIOConn) has its own EnablePathMTUDiscovery
-// in udp_rio_windows.go and is currently a no-op pending RIO-specific work.
-func (u *GenericConn) EnablePathMTUDiscovery() error {
-	v4 := u.isV4Socket()
-	return u.controlFD(func(fd uintptr) error {
-		if v4 {
-			return windows.SetsockoptInt(windows.Handle(fd), windows.IPPROTO_IP, winIPDontFragment, 1)
-		}
-		return windows.SetsockoptInt(windows.Handle(fd), windows.IPPROTO_IPV6, winIPv6DontFrag, 1)
-	})
-}