wakey wakey

Bumps the golang-x-dependencies group with 3 updates in the / directory: [golang.org/x/crypto](https://github.com/golang/crypto), [golang.org/x/net](https://github.com/golang/net) and [golang.org/x/sync](https://github.com/golang/sync).


Updates `golang.org/x/crypto` from 0.53.0 to 0.54.0
- [Commits](https://github.com/golang/crypto/compare/v0.53.0...v0.54.0)

Updates `golang.org/x/net` from 0.56.0 to 0.57.0
- [Commits](https://github.com/golang/net/compare/v0.56.0...v0.57.0)

Updates `golang.org/x/sync` from 0.21.0 to 0.22.0
- [Commits](https://github.com/golang/sync/compare/v0.21.0...v0.22.0)

Updates `golang.org/x/sys` from 0.46.0 to 0.47.0
- [Commits](https://github.com/golang/sys/compare/v0.46.0...v0.47.0)

Updates `golang.org/x/term` from 0.44.0 to 0.45.0
- [Commits](https://github.com/golang/term/compare/v0.44.0...v0.45.0)

---
updated-dependencies:
- dependency-name: golang.org/x/crypto
  dependency-version: 0.54.0
  dependency-type: direct:production
  update-type: version-update:semver-minor
  dependency-group: golang-x-dependencies
- dependency-name: golang.org/x/net
  dependency-version: 0.57.0
  dependency-type: direct:production
  update-type: version-update:semver-minor
  dependency-group: golang-x-dependencies
- dependency-name: golang.org/x/sync
  dependency-version: 0.22.0
  dependency-type: direct:production
  update-type: version-update:semver-minor
  dependency-group: golang-x-dependencies
- dependency-name: golang.org/x/sys
  dependency-version: 0.47.0
  dependency-type: direct:production
  update-type: version-update:semver-minor
  dependency-group: golang-x-dependencies
- dependency-name: golang.org/x/term
  dependency-version: 0.45.0
  dependency-type: direct:production
  update-type: version-update:semver-minor
  dependency-group: golang-x-dependencies
...

Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>

.github/actions/code-sign/action.yml

@@ -25,9 +25,9 @@ inputs:
     required: false
     default: "code-signer"
   key-prefix:
-    description: "S3 key prefix the caller is authorized to write under"
+    description: "S3 key prefix to write under; defaults to code-signing/<owner>/<repo> of the calling repo"
     required: false
-    default: "code-signing/slackhq/nebula"
+    default: ""
 
 runs:
   using: composite
@@ -57,6 +57,9 @@ runs:
         KEY_PREFIX: ${{ inputs.key-prefix }}
       run: |
         set -eu
+        # Default the prefix to this repo so the S3 key attributes the sign correctly.
+        # nebula-nightly runs this same action but writes under its own repo's prefix.
+        KEY_PREFIX="${KEY_PREFIX:-code-signing/$GITHUB_REPOSITORY}"
         RUN="${GITHUB_RUN_ID}-${GITHUB_RUN_ATTEMPT}"
 
         find "$SIGN_PATH" -name '*.exe' -print | while read -r path

.github/workflows/release.yml

@@ -12,7 +12,7 @@ jobs:
     steps:
       - uses: actions/checkout@v7
 
-      - uses: actions/setup-go@v6
+      - uses: actions/setup-go@v7
         with:
           go-version: '1.25'
           check-latest: true
@@ -38,7 +38,7 @@ jobs:
     steps:
       - uses: actions/checkout@v7
 
-      - uses: actions/setup-go@v6
+      - uses: actions/setup-go@v7
         with:
           go-version: '1.25'
           check-latest: true
@@ -78,7 +78,7 @@ jobs:
     steps:
       - uses: actions/checkout@v7
 
-      - uses: actions/setup-go@v6
+      - uses: actions/setup-go@v7
         with:
           go-version: '1.25'
           check-latest: true

.github/workflows/smoke-extra.yml

@@ -32,7 +32,7 @@ jobs:
 
     - uses: actions/checkout@v7
 
-    - uses: actions/setup-go@v6
+    - uses: actions/setup-go@v7
       with:
         go-version: '1.25'
         check-latest: true
@@ -64,7 +64,7 @@ jobs:
 
     - uses: actions/checkout@v7
 
-    - uses: actions/setup-go@v6
+    - uses: actions/setup-go@v7
       with:
         go-version: '1.25'
         check-latest: true
@@ -90,7 +90,7 @@ jobs:
 
     - uses: actions/checkout@v7
 
-    - uses: actions/setup-go@v6
+    - uses: actions/setup-go@v7
       with:
         go-version: '1.25'
         check-latest: true

.github/workflows/smoke.yml

@@ -20,7 +20,7 @@ jobs:
 
     - uses: actions/checkout@v7
 
-    - uses: actions/setup-go@v6
+    - uses: actions/setup-go@v7
       with:
         go-version: '1.25'
         check-latest: true

.github/workflows/test.yml

@@ -20,7 +20,7 @@ jobs:
 
     - uses: actions/checkout@v7
 
-    - uses: actions/setup-go@v6
+    - uses: actions/setup-go@v7
       with:
         go-version: '1.25'
         check-latest: true
@@ -80,7 +80,7 @@ jobs:
 
     - uses: actions/checkout@v7
 
-    - uses: actions/setup-go@v6
+    - uses: actions/setup-go@v7
       with:
         go-version: '1.25'
         check-latest: true
@@ -125,7 +125,7 @@ jobs:
 
     - uses: actions/checkout@v7
 
-    - uses: actions/setup-go@v6
+    - uses: actions/setup-go@v7
       with:
         go-version: '1.25'
         check-latest: true

cert/pem.go

@@ -148,6 +148,9 @@ func MarshalSigningPublicKeyToPEM(curve Curve, b []byte) []byte {
 	}
 }
 
+// UnmarshalPublicKeyFromPEM will try to unmarshal the first pem block in a byte array, returning any non
+// consumed data or an error on failure. Only key-agreement (ECDH) public key banners are accepted.
+// Use UnmarshalSigningPublicKeyFromPEM for Ed25519/ECDSA banners.
 func UnmarshalPublicKeyFromPEM(b []byte) ([]byte, []byte, Curve, error) {
 	k, r := pem.Decode(b)
 	if k == nil {
@@ -156,10 +159,10 @@ func UnmarshalPublicKeyFromPEM(b []byte) ([]byte, []byte, Curve, error) {
 	var expectedLen int
 	var curve Curve
 	switch k.Type {
-	case X25519PublicKeyBanner, Ed25519PublicKeyBanner:
+	case X25519PublicKeyBanner:
 		expectedLen = 32
 		curve = Curve_CURVE25519
-	case P256PublicKeyBanner, ECDSAP256PublicKeyBanner:
+	case P256PublicKeyBanner:
 		// Uncompressed
 		expectedLen = 65
 		curve = Curve_P256
@@ -172,6 +175,33 @@ func UnmarshalPublicKeyFromPEM(b []byte) ([]byte, []byte, Curve, error) {
 	return k.Bytes, r, curve, nil
 }
 
+// UnmarshalSigningPublicKeyFromPEM will try to unmarshal the first pem block in a byte array, returning any non
+// consumed data or an error on failure. Only Ed25519/ECDSA public key banners are accepted.
+// Use UnmarshalPublicKeyFromPEM for X25519/P256 (ECDH) banners.
+func UnmarshalSigningPublicKeyFromPEM(b []byte) ([]byte, []byte, Curve, error) {
+	k, r := pem.Decode(b)
+	if k == nil {
+		return nil, r, 0, fmt.Errorf("input did not contain a valid PEM encoded block")
+	}
+	var expectedLen int
+	var curve Curve
+	switch k.Type {
+	case Ed25519PublicKeyBanner:
+		expectedLen = 32
+		curve = Curve_CURVE25519
+	case ECDSAP256PublicKeyBanner:
+		// Uncompressed
+		expectedLen = 65
+		curve = Curve_P256
+	default:
+		return nil, r, 0, fmt.Errorf("bytes did not contain a proper Ed25519/ECDSA public key banner")
+	}
+	if len(k.Bytes) != expectedLen {
+		return nil, r, 0, fmt.Errorf("key was not %d bytes, is invalid %s public key", expectedLen, curve)
+	}
+	return k.Bytes, r, curve, nil
+}
+
 func MarshalPrivateKeyToPEM(curve Curve, b []byte) []byte {
 	switch curve {
 	case Curve_CURVE25519:

cert/pem_test.go

@@ -255,60 +255,6 @@ AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA=
 func TestUnmarshalPublicKeyFromPEM(t *testing.T) {
 	t.Parallel()
 	pubKey := []byte(`# A good key
------BEGIN NEBULA ED25519 PUBLIC KEY-----
-AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA=
------END NEBULA ED25519 PUBLIC KEY-----
-`)
-	shortKey := []byte(`# A short key
------BEGIN NEBULA ED25519 PUBLIC KEY-----
-AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA==
------END NEBULA ED25519 PUBLIC KEY-----
-`)
-	invalidBanner := []byte(`# Invalid banner
------BEGIN NOT A NEBULA PUBLIC KEY-----
-AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA=
------END NOT A NEBULA PUBLIC KEY-----
-`)
-	invalidPem := []byte(`# Not a valid PEM format
--BEGIN NEBULA ED25519 PUBLIC KEY-----
-AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA=
--END NEBULA ED25519 PUBLIC KEY-----`)
-
-	keyBundle := appendByteSlices(pubKey, shortKey, invalidBanner, invalidPem)
-
-	// Success test case
-	k, rest, curve, err := UnmarshalPublicKeyFromPEM(keyBundle)
-	assert.Len(t, k, 32)
-	assert.Equal(t, Curve_CURVE25519, curve)
-	require.NoError(t, err)
-	assert.Equal(t, rest, appendByteSlices(shortKey, invalidBanner, invalidPem))
-
-	// Fail due to short key
-	k, rest, curve, err = UnmarshalPublicKeyFromPEM(rest)
-	assert.Nil(t, k)
-	assert.Equal(t, Curve_CURVE25519, curve)
-	assert.Equal(t, rest, appendByteSlices(invalidBanner, invalidPem))
-	require.EqualError(t, err, "key was not 32 bytes, is invalid CURVE25519 public key")
-
-	// Fail due to invalid banner
-	k, rest, curve, err = UnmarshalPublicKeyFromPEM(rest)
-	assert.Nil(t, k)
-	assert.Equal(t, Curve_CURVE25519, curve)
-	require.EqualError(t, err, "bytes did not contain a proper public key banner")
-	assert.Equal(t, rest, invalidPem)
-
-	// Fail due to invalid PEM format, because
-	// it's missing the requisite pre-encapsulation boundary.
-	k, rest, curve, err = UnmarshalPublicKeyFromPEM(rest)
-	assert.Nil(t, k)
-	assert.Equal(t, Curve_CURVE25519, curve)
-	assert.Equal(t, rest, invalidPem)
-	require.EqualError(t, err, "input did not contain a valid PEM encoded block")
-}
-
-func TestUnmarshalX25519PublicKey(t *testing.T) {
-	t.Parallel()
-	pubKey := []byte(`# A good key
 -----BEGIN NEBULA X25519 PUBLIC KEY-----
 AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA=
 -----END NEBULA X25519 PUBLIC KEY-----
@@ -319,7 +265,7 @@ AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
 AAAAAAAAAAAAAAAAAAAAAAA=
 -----END NEBULA P256 PUBLIC KEY-----
 `)
-	oldPubP256Key := []byte(`# A good key
+	signingKey := []byte(`# A signing key has the wrong scope for this function
 -----BEGIN NEBULA ECDSA P256 PUBLIC KEY-----
 AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
 AAAAAAAAAAAAAAAAAAAAAAA=
@@ -340,44 +286,118 @@ AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA=
 AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA=
 -END NEBULA X25519 PUBLIC KEY-----`)
 
-	keyBundle := appendByteSlices(pubKey, pubP256Key, oldPubP256Key, shortKey, invalidBanner, invalidPem)
+	keyBundle := appendByteSlices(pubKey, pubP256Key, signingKey, shortKey, invalidBanner, invalidPem)
 
-	// Success test case
+	// X25519 key
 	k, rest, curve, err := UnmarshalPublicKeyFromPEM(keyBundle)
 	assert.Len(t, k, 32)
 	require.NoError(t, err)
-	assert.Equal(t, rest, appendByteSlices(pubP256Key, oldPubP256Key, shortKey, invalidBanner, invalidPem))
+	assert.Equal(t, rest, appendByteSlices(pubP256Key, signingKey, shortKey, invalidBanner, invalidPem))
 	assert.Equal(t, Curve_CURVE25519, curve)
 
-	// Success test case
+	// P256 key
 	k, rest, curve, err = UnmarshalPublicKeyFromPEM(rest)
 	assert.Len(t, k, 65)
 	require.NoError(t, err)
-	assert.Equal(t, rest, appendByteSlices(oldPubP256Key, shortKey, invalidBanner, invalidPem))
+	assert.Equal(t, rest, appendByteSlices(signingKey, shortKey, invalidBanner, invalidPem))
 	assert.Equal(t, Curve_P256, curve)
 
-	// Success test case
+	// Reject a signing public key (Ed25519/ECDSA banner)
-	k, rest, curve, err = UnmarshalPublicKeyFromPEM(rest)
+	k, rest, _, err = UnmarshalPublicKeyFromPEM(rest)
-	assert.Len(t, k, 65)
+	assert.Nil(t, k)
-	require.NoError(t, err)
 	assert.Equal(t, rest, appendByteSlices(shortKey, invalidBanner, invalidPem))
-	assert.Equal(t, Curve_P256, curve)
+	require.EqualError(t, err, "bytes did not contain a proper public key banner")
 
 	// Fail due to short key
-	k, rest, curve, err = UnmarshalPublicKeyFromPEM(rest)
+	k, rest, _, err = UnmarshalPublicKeyFromPEM(rest)
 	assert.Nil(t, k)
 	assert.Equal(t, rest, appendByteSlices(invalidBanner, invalidPem))
 	require.EqualError(t, err, "key was not 32 bytes, is invalid CURVE25519 public key")
 
 	// Fail due to invalid banner
-	k, rest, curve, err = UnmarshalPublicKeyFromPEM(rest)
+	k, rest, _, err = UnmarshalPublicKeyFromPEM(rest)
 	assert.Nil(t, k)
 	require.EqualError(t, err, "bytes did not contain a proper public key banner")
 	assert.Equal(t, rest, invalidPem)
 
 	// Fail due to invalid PEM format, because
 	// it's missing the requisite pre-encapsulation boundary.
-	k, rest, curve, err = UnmarshalPublicKeyFromPEM(rest)
+	k, rest, _, err = UnmarshalPublicKeyFromPEM(rest)
+	assert.Nil(t, k)
+	assert.Equal(t, rest, invalidPem)
+	require.EqualError(t, err, "input did not contain a valid PEM encoded block")
+}
+
+func TestUnmarshalSigningPublicKeyFromPEM(t *testing.T) {
+	t.Parallel()
+	pubKey := []byte(`# A good key
+-----BEGIN NEBULA ED25519 PUBLIC KEY-----
+AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA=
+-----END NEBULA ED25519 PUBLIC KEY-----
+`)
+	pubP256Key := []byte(`# A good key
+-----BEGIN NEBULA ECDSA P256 PUBLIC KEY-----
+AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
+AAAAAAAAAAAAAAAAAAAAAAA=
+-----END NEBULA ECDSA P256 PUBLIC KEY-----
+`)
+	ecdhKey := []byte(`# A key-agreement key has the wrong scope for this function
+-----BEGIN NEBULA X25519 PUBLIC KEY-----
+AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA=
+-----END NEBULA X25519 PUBLIC KEY-----
+`)
+	shortKey := []byte(`# A short key
+-----BEGIN NEBULA ED25519 PUBLIC KEY-----
+AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA==
+-----END NEBULA ED25519 PUBLIC KEY-----
+`)
+	invalidBanner := []byte(`# Invalid banner
+-----BEGIN NOT A NEBULA PUBLIC KEY-----
+AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA=
+-----END NOT A NEBULA PUBLIC KEY-----
+`)
+	invalidPem := []byte(`# Not a valid PEM format
+-BEGIN NEBULA ED25519 PUBLIC KEY-----
+AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA=
+-END NEBULA ED25519 PUBLIC KEY-----`)
+
+	keyBundle := appendByteSlices(pubKey, pubP256Key, ecdhKey, shortKey, invalidBanner, invalidPem)
+
+	// Ed25519 key
+	k, rest, curve, err := UnmarshalSigningPublicKeyFromPEM(keyBundle)
+	assert.Len(t, k, 32)
+	require.NoError(t, err)
+	assert.Equal(t, rest, appendByteSlices(pubP256Key, ecdhKey, shortKey, invalidBanner, invalidPem))
+	assert.Equal(t, Curve_CURVE25519, curve)
+
+	// ECDSA P256 key
+	k, rest, curve, err = UnmarshalSigningPublicKeyFromPEM(rest)
+	assert.Len(t, k, 65)
+	require.NoError(t, err)
+	assert.Equal(t, rest, appendByteSlices(ecdhKey, shortKey, invalidBanner, invalidPem))
+	assert.Equal(t, Curve_P256, curve)
+
+	// Reject a key-agreement public key (X25519/P256 banner)
+	k, rest, _, err = UnmarshalSigningPublicKeyFromPEM(rest)
+	assert.Nil(t, k)
+	assert.Equal(t, rest, appendByteSlices(shortKey, invalidBanner, invalidPem))
+	require.EqualError(t, err, "bytes did not contain a proper Ed25519/ECDSA public key banner")
+
+	// Fail due to short key
+	k, rest, _, err = UnmarshalSigningPublicKeyFromPEM(rest)
+	assert.Nil(t, k)
+	assert.Equal(t, rest, appendByteSlices(invalidBanner, invalidPem))
+	require.EqualError(t, err, "key was not 32 bytes, is invalid CURVE25519 public key")
+
+	// Fail due to invalid banner
+	k, rest, _, err = UnmarshalSigningPublicKeyFromPEM(rest)
+	assert.Nil(t, k)
+	require.EqualError(t, err, "bytes did not contain a proper Ed25519/ECDSA public key banner")
+	assert.Equal(t, rest, invalidPem)
+
+	// Fail due to invalid PEM format, because
+	// it's missing the requisite pre-encapsulation boundary.
+	k, rest, _, err = UnmarshalSigningPublicKeyFromPEM(rest)
 	assert.Nil(t, k)
 	assert.Equal(t, rest, invalidPem)
 	require.EqualError(t, err, "input did not contain a valid PEM encoded block")

cmd/nebula-service/main.go

@@ -53,7 +53,12 @@ func main() {
 	l := logging.NewLogger(os.Stdout)
 
 	if *serviceFlag != "" {
-		if err := doService(configPath, configTest, Build, serviceFlag); err != nil {
+		if *configTest {
+			fmt.Println("-test is not supported with -service, run the config test without -service")
+			os.Exit(1)
+		}
+
+		if err := doService(configPath, Build, serviceFlag); err != nil {
 			l.Error("Service command failed", "error", err)
 			os.Exit(1)
 		}
@@ -93,15 +98,14 @@ func main() {
 	}
 
 	if !*configTest {
-		wait, err := ctrl.Start()
+		if err := ctrl.Start(); err != nil {
-		if err != nil {
 			util.LogWithContextIfNeeded("Error while running", err, l)
 			os.Exit(1)
 		}
 
 		go ctrl.ShutdownBlock()
 
-		if err := wait(); err != nil {
+		if err := ctrl.Wait(); err != nil {
 			l.Error("Nebula stopped due to fatal error", "error", err)
 			os.Exit(2)
 		}

cmd/nebula-service/service.go

@@ -3,6 +3,7 @@ package main
 import (
 	"fmt"
 	"log"
+	"os"
 
 	"github.com/kardianos/service"
 	"github.com/slackhq/nebula"
@@ -14,7 +15,6 @@ var logger service.Logger
 
 type program struct {
 	configPath *string
-	configTest *bool
 	build      string
 	control    *nebula.Control
 }
@@ -40,22 +40,41 @@ func (p *program) Start(s service.Service) error {
 		}
 	})
 
-	p.control, err = nebula.Main(c, *p.configTest, Build, l, nil)
+	p.control, err = nebula.Main(c, false, Build, l, nil)
 	if err != nil {
 		return err
 	}
 
-	p.control.Start()
+	if err := p.control.Start(); err != nil {
+		return err
+	}
+
+	// Nebula can stop itself on a fatal packet reader error, make sure to log it if it happens.
+	go func() {
+		if err := p.control.Wait(); err != nil {
+			logger.Error(fmt.Sprintf("Nebula stopped due to fatal error: %v", err))
+			os.Exit(2)
+		}
+	}()
+
 	return nil
 }
 
 func (p *program) Stop(s service.Service) error {
 	logger.Info("Nebula service stopping.")
+	if p.control == nil {
+		return nil
+	}
+
 	p.control.Stop()
+
+	// block until nebula has fully drained before reporting stopped.
+	// error logging is handled by Start.
+	_ = p.control.Wait()
 	return nil
 }
 
-func doService(configPath *string, configTest *bool, build string, serviceFlag *string) error {
+func doService(configPath *string, build string, serviceFlag *string) error {
 	if *configPath == "" {
 		p, err := config.DefaultPath()
 		if err != nil {
@@ -73,7 +92,6 @@ func doService(configPath *string, configTest *bool, build string, serviceFlag *
 
 	prg := &program{
 		configPath: configPath,
-		configTest: configTest,
 		build:      build,
 	}
 
@@ -105,8 +123,9 @@ func doService(configPath *string, configTest *bool, build string, serviceFlag *
 	switch *serviceFlag {
 	case "run":
 		if err := s.Run(); err != nil {
-			// Route any errors to the system logger
+			// Route any errors to the system logger and report the failure
 			logger.Error(err)
+			return err
 		}
 	default:
 		if err := service.Control(s, *serviceFlag); err != nil {

cmd/nebula/close_on_timer_test.go

@@ -0,0 +1,96 @@
+//go:build linux && !android && !e2e_testing
+
+package main
+
+import (
+	"fmt"
+	"net/netip"
+	"os"
+	"path/filepath"
+	"runtime"
+	"testing"
+	"time"
+
+	"github.com/slackhq/nebula"
+	"github.com/slackhq/nebula/cert"
+	cert_test "github.com/slackhq/nebula/cert_test"
+	"github.com/slackhq/nebula/config"
+	"github.com/slackhq/nebula/test"
+	"github.com/stretchr/testify/require"
+)
+
+// TestControlStopClosesOnTimer reproduces the dnclient lifecycle: nebula runs as
+// a library, and on a config update dnclient calls Stop() in-process to tear the
+// old instance down before starting a new one. This boots a real nebula (real
+// blocking UDP sockets, tun disabled), lets it run, then Stop()s it on a timer
+// and asserts it actually closes. If the reader goroutines parked in recvmmsg
+// don't wake on Close(), Wait() blocks forever and this fails with a goroutine
+// dump instead of relying on a process signal to unstick them.
+func TestControlStopClosesOnTimer(t *testing.T) {
+	l := test.NewLogger()
+	dir := t.TempDir()
+
+	before := time.Now().Add(-time.Hour)
+	after := time.Now().Add(time.Hour)
+	ca, _, caKey, caPEM := cert_test.NewTestCaCert(cert.Version2, cert.Curve_CURVE25519, before, after, nil, nil, nil)
+	networks := []netip.Prefix{netip.MustParsePrefix("10.0.0.1/24")}
+	_, _, keyPEM, certPEM := cert_test.NewTestCert(cert.Version2, cert.Curve_CURVE25519, ca, caKey, "close-on-timer", before, after, networks, nil, nil)
+
+	caPath := filepath.Join(dir, "ca.pem")
+	certPath := filepath.Join(dir, "cert.pem")
+	keyPath := filepath.Join(dir, "key.pem")
+	require.NoError(t, os.WriteFile(caPath, caPEM, 0o600))
+	require.NoError(t, os.WriteFile(certPath, certPEM, 0o600))
+	require.NoError(t, os.WriteFile(keyPath, keyPEM, 0o600))
+
+	// tun disabled so no device/root is needed; routines: 2 so we exercise the
+	// multi-socket (SO_REUSEPORT) teardown, which is where dnclient runs.
+	configBody := fmt.Sprintf(`
+pki:
+  ca: %s
+  cert: %s
+  key: %s
+listen:
+  host: 127.0.0.1
+  port: 0
+tun:
+  disabled: true
+firewall:
+  outbound:
+    - port: any
+      proto: any
+      host: any
+  inbound:
+    - port: any
+      proto: any
+      host: any
+routines: 2
+`, caPath, certPath, keyPath)
+	require.NoError(t, os.WriteFile(filepath.Join(dir, "config.yml"), []byte(configBody), 0o600))
+
+	c := config.NewC(l)
+	require.NoError(t, c.Load(dir))
+
+	ctrl, err := nebula.Main(c, false, "close-on-timer", l, nil)
+	require.NoError(t, err)
+	require.NoError(t, ctrl.Start())
+
+	// Run like a live nebula, then close on a timer, exactly as dnclient does.
+	<-time.NewTimer(5 * time.Second).C
+
+	stopped := make(chan struct{})
+	go func() {
+		ctrl.Stop() // closes the udp sockets (shutdown(2)) and the tun
+		ctrl.Wait() // blocks until every reader goroutine has returned
+		close(stopped)
+	}()
+
+	select {
+	case <-stopped:
+		t.Log("nebula closed cleanly on timer")
+	case <-time.After(10 * time.Second):
+		buf := make([]byte, 1<<20)
+		n := runtime.Stack(buf, true)
+		t.Fatalf("nebula did NOT close within 10s of Stop(): a blocking reader never woke\n%s", buf[:n])
+	}
+}

cmd/nebula/main.go

@@ -84,8 +84,7 @@ func main() {
 	}
 
 	if !*configTest {
-		wait, err := ctrl.Start()
+		if err := ctrl.Start(); err != nil {
-		if err != nil {
 			util.LogWithContextIfNeeded("Error while running", err, l)
 			os.Exit(1)
 		}
@@ -93,7 +92,7 @@ func main() {
 		go ctrl.ShutdownBlock()
 		notifyReady(l)
 
-		if err := wait(); err != nil {
+		if err := ctrl.Wait(); err != nil {
 			l.Error("Nebula stopped due to fatal error", "error", err)
 			os.Exit(2)
 		}

connection_manager.go

@@ -44,6 +44,11 @@ type connectionManager struct {
 	inactivityTimeout       atomic.Int64
 	dropInactive            atomic.Bool
 
+	// Wake-from-sleep handling, sampled once per tick in Start
+	wakeDetector       *wakeDetector
+	clearOnWake        atomic.Bool
+	wakeClearThreshold atomic.Int64
+
 	l *slog.Logger
 }
 
@@ -54,6 +59,7 @@ func newConnectionManagerFromConfig(l *slog.Logger, c *config.C, hm *HostMap, p
 		punchy:        p,
 		relayUsed:     make(map[uint32]struct{}),
 		relayUsedLock: &sync.RWMutex{},
+		wakeDetector:  newWakeDetector(),
 	}
 
 	cm.reload(c, true)
@@ -98,12 +104,38 @@ func (cm *connectionManager) reload(c *config.C, initial bool) {
 			)
 		}
 	}
+
+	if initial || c.HasChanged("tunnels.clear_on_wake") {
+		old := cm.clearOnWake.Load()
+		cm.clearOnWake.Store(c.GetBool("tunnels.clear_on_wake", true))
+		if !initial {
+			cm.l.Info("Clear on wake setting has changed",
+				"oldBool", old,
+				"newBool", cm.clearOnWake.Load(),
+			)
+		}
+	}
+
+	if initial || c.HasChanged("tunnels.wake_clear_threshold") {
+		old := cm.getWakeClearThreshold()
+		cm.wakeClearThreshold.Store((int64)(c.GetDuration("tunnels.wake_clear_threshold", 30*time.Second)))
+		if !initial {
+			cm.l.Info("Wake clear threshold has changed",
+				"oldDuration", old,
+				"newDuration", cm.getWakeClearThreshold(),
+			)
+		}
+	}
 }
 
 func (cm *connectionManager) getInactivityTimeout() time.Duration {
 	return (time.Duration)(cm.inactivityTimeout.Load())
 }
 
+func (cm *connectionManager) getWakeClearThreshold() time.Duration {
+	return (time.Duration)(cm.wakeClearThreshold.Load())
+}
+
 func (cm *connectionManager) In(h *HostInfo) {
 	h.in.Store(true)
 }
@@ -136,6 +168,73 @@ func (cm *connectionManager) getAndResetTrafficCheck(h *HostInfo, now time.Time)
 	return in, out
 }
 
+// checkWake runs once per tick and clears every tunnel when the machine has just returned from system sleep.
+// Tunnels rarely survive a suspend: our NAT mappings have expired and our address has usually changed, so every
+// established hostinfo is a corpse that will eat 15-20s of traffic checks before the wheel declares it dead.
+// Clearing now means the first packet after wake starts a fresh handshake immediately.
+//
+// The suspend itself costs nothing here: the ticker driving us is frozen with the rest of the process and this
+// fires within one tick of resume.
+func (cm *connectionManager) checkWake() {
+	slept, ok := cm.wakeDetector.Sample()
+	if !ok || slept == 0 {
+		return
+	}
+
+	// The clock pair is read non-atomically, so scheduling jitter shows up as tiny sub-millisecond "sleeps".
+	// Keep the floor well above that so a zero/nonsense threshold can't clear tunnels on every tick.
+	threshold := max(cm.getWakeClearThreshold(), time.Second)
+
+	if slept < threshold {
+		// Short suspends (lid closed and quickly reopened) often come back before NAT state expires; those
+		// tunnels may well be alive, leave them to the normal traffic checks.
+		if slept >= time.Second {
+			cm.l.Debug("Woke from sleep below the clear threshold, leaving tunnels alone",
+				"sleptFor", slept,
+				"threshold", threshold,
+			)
+		}
+		return
+	}
+
+	if !cm.clearOnWake.Load() {
+		cm.l.Info("Woke from sleep, tunnels.clear_on_wake is disabled so tunnels are left to the normal traffic checks", "sleptFor", slept)
+		return
+	}
+
+	closed := cm.clearAllTunnels()
+	cm.l.Info("Woke from sleep, cleared tunnels", "sleptFor", slept, "tunnelsCleared", closed)
+
+	// Our public address almost certainly changed; get it to the lighthouses as soon as possible so peers can
+	// find us again. The update rides over a fresh lighthouse handshake. If the network isn't back up yet these
+	// sends fail harmlessly and the periodic update worker retries within lighthouse.interval.
+	cm.intf.lightHouse.TriggerUpdate()
+}
+
+// clearAllTunnels closes every tunnel in the hostmap locally, without notifying the remotes. It is the wake-from-
+// sleep counterpart to Control.CloseAllTunnels: after a suspend the remotes stopped hearing from us long ago, and
+// close packets fired into a network that may not even be up yet are wasted, so we only tear down our own state
+// and let the next packet to each host start a fresh handshake.
+func (cm *connectionManager) clearAllTunnels() int {
+	cm.hostMap.RLock()
+	hostinfos := make([]*HostInfo, 0, len(cm.hostMap.Indexes))
+	for _, h := range cm.hostMap.Indexes {
+		hostinfos = append(hostinfos, h)
+	}
+	cm.hostMap.RUnlock()
+
+	for _, h := range hostinfos {
+		cm.intf.closeTunnel(h)
+	}
+
+	// With every tunnel gone no relay can be in use, drop the usage tracking wholesale.
+	cm.relayUsedLock.Lock()
+	clear(cm.relayUsed)
+	cm.relayUsedLock.Unlock()
+
+	return len(hostinfos)
+}
+
 func (cm *connectionManager) Start(ctx context.Context) {
 	clockSource := time.NewTicker(cm.trafficTimer.t.tickDuration)
 	defer clockSource.Stop()
@@ -150,6 +249,7 @@ func (cm *connectionManager) Start(ctx context.Context) {
 			return
 
 		case now := <-clockSource.C:
+			cm.checkWake()
 			cm.trafficTimer.Advance(now)
 			for {
 				localIndex, has := cm.trafficTimer.Purge()

connection_manager_test.go

@@ -501,3 +501,85 @@ func (d *dummyCert) MarshalJSON() ([]byte, error) {
 func (d *dummyCert) Copy() cert.Certificate {
 	return d
 }
+
+func TestConnectionManager_WakeClear(t *testing.T) {
+	l := test.NewLogger()
+	localrange := netip.MustParsePrefix("10.1.1.1/24")
+	vpnIp := netip.MustParseAddr("172.1.1.2")
+	preferredRanges := []netip.Prefix{localrange}
+
+	// Very incomplete mock objects
+	hostMap := newHostMap(l)
+	hostMap.preferredRanges.Store(&preferredRanges)
+
+	cs := &CertState{
+		initiatingVersion: cert.Version1,
+		privateKey:        []byte{},
+		v1Cert:            &dummyCert{version: cert.Version1},
+		v1Credential:      nil,
+	}
+
+	lh := newTestLighthouse()
+	ifce := &Interface{
+		hostMap:          hostMap,
+		inside:           &overlaytest.NoopTun{},
+		outside:          &udp.NoopConn{},
+		firewall:         &Firewall{},
+		lightHouse:       lh,
+		pki:              &PKI{},
+		handshakeManager: NewHandshakeManager(l, hostMap, lh, &udp.NoopConn{}, defaultHandshakeConfig),
+		l:                l,
+	}
+	ifce.pki.cs.Store(cs)
+
+	// Create manager
+	conf := config.NewC(test.NewLogger())
+	punchy := NewPunchyFromConfig(test.NewLogger(), conf, nil)
+	nc := newConnectionManagerFromConfig(test.NewLogger(), conf, hostMap, punchy)
+	nc.intf = ifce
+
+	// Drive the wake detector from a fake clock pair
+	suspended := time.Duration(0)
+	nc.wakeDetector = &wakeDetector{read: func() (time.Duration, bool) { return suspended, true }}
+	nc.checkWake() // primes the baseline
+
+	addTunnel := func(localIndex uint32) *HostInfo {
+		hostinfo := &HostInfo{
+			vpnAddrs:      []netip.Addr{vpnIp},
+			localIndexId:  localIndex,
+			remoteIndexId: 9901,
+		}
+		hostinfo.ConnectionState = &ConnectionState{
+			myCert: &dummyCert{version: cert.Version1},
+		}
+		nc.hostMap.unlockedAddHostInfo(hostinfo, ifce)
+		return hostinfo
+	}
+
+	addTunnel(1099)
+	nc.RelayUsed(5000)
+
+	// No suspend, nothing happens
+	nc.checkWake()
+	assert.Contains(t, nc.hostMap.Indexes, uint32(1099))
+
+	// A suspend below the threshold leaves tunnels alone
+	suspended += 5 * time.Second
+	nc.checkWake()
+	assert.Contains(t, nc.hostMap.Indexes, uint32(1099))
+
+	// A suspend past the threshold clears everything, including relay usage tracking
+	suspended += time.Hour
+	nc.checkWake()
+	assert.Empty(t, nc.hostMap.Indexes)
+	assert.Empty(t, nc.hostMap.Hosts)
+	assert.Empty(t, nc.relayUsed)
+
+	// With clear_on_wake disabled the tunnels survive a long suspend
+	addTunnel(1100)
+	nc.clearOnWake.Store(false)
+	suspended += time.Hour
+	nc.checkWake()
+	assert.Contains(t, nc.hostMap.Indexes, uint32(1100))
+	assert.Contains(t, nc.hostMap.Hosts, vpnIp)
+}

connection_state.go

@@ -2,11 +2,13 @@ package nebula
 
 import (
 	"encoding/json"
+	"log/slog"
 	"sync"
 	"sync/atomic"
 
 	"github.com/slackhq/nebula/cert"
 	"github.com/slackhq/nebula/handshake"
+	"github.com/slackhq/nebula/header"
 	"github.com/slackhq/nebula/noiseutil"
 )
 
@@ -20,6 +22,7 @@ type ConnectionState struct {
 	initiator      bool
 	messageCounter atomic.Uint64
 	window         *Bits
+	decryptLock    sync.Mutex
 	writeLock      sync.Mutex
 }
 
@@ -54,3 +57,52 @@ func (cs *ConnectionState) MarshalJSON() ([]byte, error) {
 func (cs *ConnectionState) Curve() cert.Curve {
 	return cs.myCert.Curve()
 }
+
+func (cs *ConnectionState) Decrypt(l *slog.Logger, messageCounter uint64, out []byte, packet []byte, nb []byte) ([]byte, error) {
+	var err error
+	cs.decryptLock.Lock()
+	result := cs.window.Check(l, messageCounter)
+	cs.decryptLock.Unlock()
+	if !result {
+		return nil, ErrAlreadySeen
+	}
+
+	out, err = cs.dKey.DecryptDanger(out, packet[:header.Len], packet[header.Len:], messageCounter, nb)
+	if err != nil {
+		return nil, err
+	}
+
+	cs.decryptLock.Lock()
+	result = cs.window.Update(l, messageCounter)
+	cs.decryptLock.Unlock()
+	if !result {
+		return nil, ErrAlreadySeen
+	}
+	return out, nil
+}
+
+// VerifyRelay verifies AEAD protected (but not encrypted) relay frames. packet must be length-checked by the caller.
+func (cs *ConnectionState) VerifyRelay(l *slog.Logger, messageCounter uint64, packet []byte, nb []byte) error {
+	cs.decryptLock.Lock()
+	result := cs.window.Check(l, messageCounter)
+	cs.decryptLock.Unlock()
+	if !result {
+		return ErrAlreadySeen
+	}
+
+	signedPayload := packet[:len(packet)-cs.dKey.Overhead()]
+	signatureValue := packet[len(packet)-cs.dKey.Overhead():]
+	_, err := cs.dKey.DecryptDanger(nil, signedPayload, signatureValue, messageCounter, nb)
+	if err != nil {
+		return err
+	}
+
+	cs.decryptLock.Lock()
+	result = cs.window.Update(l, messageCounter)
+	cs.decryptLock.Unlock()
+	if !result {
+		return ErrAlreadySeen
+	}
+
+	return nil
+}

control.go

@@ -69,29 +69,29 @@ type ControlHostInfo struct {
 }
 
 // Start actually runs nebula, this is a nonblocking call.
-// The returned function blocks until nebula has fully stopped and returns the
+// Use Wait to block until nebula has fully stopped and to learn whether a fatal reader error caused the shutdown.
-// first fatal reader error (if any). A nil error means nebula shut down
+func (c *Control) Start() error {
-// gracefully; a non-nil error means a reader hit an unexpected failure that
-// triggered the shutdown.
-func (c *Control) Start() (func() error, error) {
 	c.stateLock.Lock()
 	defer c.stateLock.Unlock()
 	switch c.state {
 	case StateReady:
 		//yay!
 	case StateStopped, StateStopping:
-		return nil, ErrAlreadyStopped
+		return ErrAlreadyStopped
 	case StateStarted:
-		return nil, ErrAlreadyStarted
+		return ErrAlreadyStarted
 	default:
-		return nil, ErrUnknownState
+		return ErrUnknownState
 	}
 
 	// Activate the interface
 	err := c.f.activate()
 	if err != nil {
+		// Cancel before Close so a caller returning from Wait always observes a dead Context
+		c.cancel()
+		_ = c.f.Close()
 		c.state = StateStopped
-		return nil, err
+		return err
 	}
 
 	// Call all the delayed funcs that waited patiently for the interface to be created.
@@ -114,13 +114,9 @@ func (c *Control) Start() (func() error, error) {
 	c.f.triggerShutdown = c.Stop
 
 	// Start reading packets.
-	out, err := c.f.run()
+	c.f.run()
-	if err != nil {
-		c.state = StateStopped
-		return nil, err
-	}
 	c.state = StateStarted
-	return out, nil
+	return nil
 }
 
 func (c *Control) State() RunState {
@@ -133,10 +129,26 @@ func (c *Control) Context() context.Context {
 	return c.ctx
 }
 
-// Stop is a non-blocking call that signals nebula to close all tunnels and shut down
+// Stop tears nebula down, closing all tunnels and releasing everything it holds.
+// Use Wait to block until the shutdown has completed.
+// A Control that has been stopped cannot be started again, Start will return ErrAlreadyStopped.
 func (c *Control) Stop() {
 	c.stateLock.Lock()
-	if c.state != StateStarted {
+	switch c.state {
+	case StateStarted:
+		// Fall through to the full teardown below
+
+	case StateReady:
+		// Never started
+		c.cancel()
+		c.state = StateStopped
+		if err := c.f.Close(); err != nil {
+			c.l.Error("Close interface failed", "error", err)
+		}
+		c.stateLock.Unlock()
+		return
+
+	default:
 		c.stateLock.Unlock()
 		// We are stopping or stopped already
 		return
@@ -145,19 +157,26 @@ func (c *Control) Stop() {
 	c.state = StateStopping
 	c.stateLock.Unlock()
 
-	// Stop the handshakeManager (and other services), to prevent new tunnels from
+	// Closing tunnels can be slow with a large hostmap, don't hold the lock for it
-	// being created while we're shutting them all down.
 	c.cancel()
-
 	c.CloseAllTunnels(false)
+
+	c.stateLock.Lock()
+	c.state = StateStopped
 	if err := c.f.Close(); err != nil {
 		c.l.Error("Close interface failed", "error", err)
 	}
-	c.stateLock.Lock()
-	c.state = StateStopped
 	c.stateLock.Unlock()
 }
 
+// Wait blocks until nebula has fully stopped, either via Stop or an internal fatal error,
+// and returns the first fatal packet reader error if there was one.
+// It is safe to call from multiple goroutines and at any point in the lifecycle,
+// but a Wait on a Control that is never started and never stopped will block forever.
+func (c *Control) Wait() error {
+	return c.f.wait()
+}
+
 // ShutdownBlock will listen for and block on term and interrupt signals, calling Control.Stop() once signalled
 func (c *Control) ShutdownBlock() {
 	sigChan := make(chan os.Signal, 1)
@@ -170,8 +189,15 @@ func (c *Control) ShutdownBlock() {
 	c.Stop()
 }
 
-// RebindUDPServer asks the UDP listener to rebind it's listener. Mainly used on mobile clients when interfaces change
+// RebindUDPServer asks the UDP listener to rebind it's listener. Mainly used on mobile clients when interfaces change.
 func (c *Control) RebindUDPServer() {
+	c.stateLock.Lock()
+	defer c.stateLock.Unlock()
+
+	if c.state != StateStarted {
+		return
+	}
+
 	_ = c.f.outside.Rebind()
 
 	// Trigger a lighthouse update, useful for mobile clients that should have an update interval of 0
@@ -305,7 +331,7 @@ func (c *Control) CloseAllTunnels(excludeLighthouses bool) (closed int) {
 
 		c.l.Debug("Sending close tunnel message",
 			"vpnAddrs", h.vpnAddrs,
-			"udpAddr", h.remote,
+			"udpAddr", h.GetRemote(),
 		)
 		closed++
 	}
@@ -350,7 +376,7 @@ func copyHostInfo(h *HostInfo, preferredRanges []netip.Prefix) ControlHostInfo {
 		RemoteAddrs:            h.remotes.CopyAddrs(preferredRanges),
 		CurrentRelaysToMe:      h.relayState.CopyRelayIps(),
 		CurrentRelaysThroughMe: h.relayState.CopyRelayForIps(),
-		CurrentRemote:          h.remote,
+		CurrentRemote:          h.GetRemote(),
 	}
 
 	for i, a := range h.vpnAddrs {

control_lifecycle_test.go

@@ -0,0 +1,292 @@
+package nebula
+
+import (
+	"context"
+	"errors"
+	"io"
+	"net/netip"
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/gaissmai/bart"
+	"github.com/slackhq/nebula/config"
+	"github.com/slackhq/nebula/routing"
+	"github.com/slackhq/nebula/test"
+	"github.com/slackhq/nebula/udp"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+type fakeDevice struct {
+	closeOnce sync.Once
+	closedCh  chan struct{}
+	closed    bool
+}
+
+func newFakeDevice() *fakeDevice {
+	return &fakeDevice{closedCh: make(chan struct{})}
+}
+
+// Read blocks until Close like a real tun with no traffic, then reports EOF
+// the same way a closed device does
+func (d *fakeDevice) Read(p []byte) (int, error) {
+	<-d.closedCh
+	return 0, io.EOF
+}
+
+func (d *fakeDevice) Write(p []byte) (int, error) { return len(p), nil }
+
+func (d *fakeDevice) Close() error {
+	d.closeOnce.Do(func() {
+		d.closed = true
+		close(d.closedCh)
+	})
+	return nil
+}
+
+func (d *fakeDevice) Activate() error                       { return nil }
+func (d *fakeDevice) Networks() []netip.Prefix              { return nil }
+func (d *fakeDevice) Name() string                          { return "fake" }
+func (d *fakeDevice) RoutesFor(netip.Addr) routing.Gateways { return nil }
+func (d *fakeDevice) SupportsMultiqueue() bool              { return false }
+func (d *fakeDevice) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+	return nil, errors.New("unsupported")
+}
+
+// newReadyControl hand-builds the minimum Control that Main would have
+// produced right before Start, including the construction token NewInterface
+// takes so waiters block until Close releases the resources
+func newReadyControl(t *testing.T) (*Control, *fakeDevice, *fakeConn) {
+	l := test.NewLogger()
+	dev := newFakeDevice()
+	conn := &fakeConn{}
+	ctx, cancel := context.WithCancel(context.Background())
+
+	myVpnNet := netip.MustParsePrefix("10.128.0.1/16")
+	nt := new(bart.Lite)
+	nt.Insert(myVpnNet)
+	cs := &CertState{
+		myVpnNetworks:      []netip.Prefix{myVpnNet},
+		myVpnNetworksTable: nt,
+	}
+	lh, err := NewLightHouseFromConfig(ctx, l, config.NewC(l), cs, nil, nil)
+	require.NoError(t, err)
+
+	f := &Interface{
+		ctx:        ctx,
+		inside:     dev,
+		outside:    conn,
+		writers:    []udp.Conn{conn},
+		readers:    make([]io.ReadWriteCloser, 1),
+		routines:   1,
+		hostMap:    newHostMap(l),
+		lightHouse: lh,
+		l:          l,
+	}
+	f.wg.Add(1)
+
+	return &Control{
+		state:  StateReady,
+		f:      f,
+		l:      l,
+		ctx:    ctx,
+		cancel: cancel,
+	}, dev, conn
+}
+
+func TestControl_StopBeforeStart(t *testing.T) {
+	c, dev, conn := newReadyControl(t)
+
+	// A Stop on a never started control must release everything Main acquired
+	c.Stop()
+	assert.Equal(t, StateStopped, c.State())
+	assert.True(t, dev.closed, "the tun device should have been closed")
+	assert.True(t, conn.closed, "the udp socket should have been closed")
+	require.ErrorIs(t, c.ctx.Err(), context.Canceled, "the service context should have been cancelled")
+
+	// Wait must return promptly now that the resources are released
+	require.NoError(t, c.Wait())
+
+	// A stopped control can never be started
+	require.ErrorIs(t, c.Start(), ErrAlreadyStopped)
+
+	// A second Stop is a harmless no-op
+	c.Stop()
+	assert.Equal(t, StateStopped, c.State())
+	require.NoError(t, c.Wait())
+}
+
+func TestControl_WaitBlocksUntilStop(t *testing.T) {
+	c, _, _ := newReadyControl(t)
+
+	done := make(chan error, 1)
+	go func() { done <- c.Wait() }()
+
+	select {
+	case <-done:
+		t.Fatal("Wait returned before Stop")
+	case <-time.After(50 * time.Millisecond):
+	}
+
+	c.Stop()
+	select {
+	case err := <-done:
+		require.NoError(t, err)
+	case <-time.After(time.Second):
+		t.Fatal("Wait did not return after Stop")
+	}
+}
+
+type fakeConn struct {
+	closed  bool
+	rebinds int
+}
+
+func (c *fakeConn) Rebind() error                            { c.rebinds++; return nil }
+func (c *fakeConn) LocalAddr() (netip.AddrPort, error)       { return netip.AddrPort{}, nil }
+func (c *fakeConn) ListenOut(_ udp.EncReader) error          { return nil }
+func (c *fakeConn) WriteTo(_ []byte, _ netip.AddrPort) error { return nil }
+func (c *fakeConn) ReloadConfig(_ *config.C)                 {}
+func (c *fakeConn) SupportsMultipleReaders() bool            { return true }
+func (c *fakeConn) Close() error                             { c.closed = true; return nil }
+
+type multiqueueDevice struct {
+	*fakeDevice
+}
+
+func (d *multiqueueDevice) SupportsMultiqueue() bool { return true }
+
+func TestControl_StartMultiqueueFailureReleases(t *testing.T) {
+	dev := &multiqueueDevice{fakeDevice: newFakeDevice()}
+	conn := &fakeConn{}
+	ctx, cancel := context.WithCancel(context.Background())
+	f := &Interface{
+		ctx:      ctx,
+		inside:   dev,
+		outside:  conn,
+		writers:  []udp.Conn{conn},
+		readers:  make([]io.ReadWriteCloser, 2),
+		routines: 2,
+		l:        test.NewLogger(),
+	}
+	f.wg.Add(1)
+
+	c := &Control{
+		state:  StateReady,
+		f:      f,
+		l:      test.NewLogger(),
+		ctx:    ctx,
+		cancel: cancel,
+	}
+
+	// The second reader fails to open, everything must be released
+	require.Error(t, c.Start())
+	assert.Equal(t, StateStopped, c.State())
+	assert.True(t, dev.closed, "the tun device should have been closed")
+	assert.True(t, conn.closed, "the udp socket should have been closed")
+	require.ErrorIs(t, c.ctx.Err(), context.Canceled)
+
+	// And Wait must not hang on the construction token
+	require.NoError(t, c.Wait())
+}
+
+func TestInterface_CloseIsIdempotent(t *testing.T) {
+	dev := newFakeDevice()
+	f := &Interface{
+		inside: dev,
+		l:      test.NewLogger(),
+	}
+	f.wg.Add(1)
+
+	require.NoError(t, f.Close())
+	assert.True(t, dev.closed)
+
+	// A second Close must not double release the wg token or the device
+	require.NoError(t, f.Close())
+	require.NoError(t, f.wait())
+}
+
+func TestControl_FatalErrorReportsThroughWait(t *testing.T) {
+	c, dev, conn := newReadyControl(t)
+
+	// Mirror what Start wires up, without needing real packet readers
+	c.f.triggerShutdown = c.Stop
+	c.state = StateStarted
+
+	boom := errors.New("boom")
+	c.f.onFatal(boom)
+
+	require.ErrorIs(t, c.Wait(), boom)
+	assert.Equal(t, StateStopped, c.State())
+	assert.True(t, dev.closed)
+	assert.True(t, conn.closed)
+
+	// A second fatal error must not fire the shutdown again or replace the first
+	c.f.onFatal(errors.New("later"))
+	require.ErrorIs(t, c.Wait(), boom)
+
+	// Wait stays factual, a Stop after the death does not mask the error
+	c.Stop()
+	require.ErrorIs(t, c.Wait(), boom)
+}
+
+func TestControl_ConcurrentStopAndStart(t *testing.T) {
+	c, _, _ := newReadyControl(t)
+
+	var wg sync.WaitGroup
+	for i := 0; i < 2; i++ {
+		wg.Go(func() { c.Stop() })
+	}
+	wg.Go(func() { _ = c.Start() })
+	wg.Go(func() {
+		_ = c.Wait()
+		// A returned Wait must always observe the final state, no matter how
+		// the race resolved
+		assert.Equal(t, StateStopped, c.State())
+	})
+	wg.Wait()
+
+	// However the race resolves, the control must end fully stopped with no
+	// panic and Wait must observe the final state
+	require.NoError(t, c.Wait())
+	assert.Equal(t, StateStopped, c.State())
+	require.ErrorIs(t, c.Start(), ErrAlreadyStopped)
+}
+
+func TestControl_StartStopLifecycle(t *testing.T) {
+	c, dev, conn := newReadyControl(t)
+
+	require.NoError(t, c.Start())
+	assert.Equal(t, StateStarted, c.State())
+	require.ErrorIs(t, c.Start(), ErrAlreadyStarted)
+
+	// Stop must unpark the reader blocked in the device and release everything
+	c.Stop()
+	assert.Equal(t, StateStopped, c.State())
+	assert.True(t, dev.closed, "the tun device should have been closed")
+	assert.True(t, conn.closed, "the udp socket should have been closed")
+	require.ErrorIs(t, c.ctx.Err(), context.Canceled)
+
+	// The reader drained off a closed device, that is not a fatal error
+	require.NoError(t, c.Wait())
+	require.ErrorIs(t, c.Start(), ErrAlreadyStopped)
+}
+
+func TestControl_RebindIsGatedByState(t *testing.T) {
+	c, _, conn := newReadyControl(t)
+
+	// A rebind before Start reaches nothing, the interface is not up
+	c.RebindUDPServer()
+	assert.Equal(t, 0, conn.rebinds, "rebind before start must be a no-op")
+
+	require.NoError(t, c.Start())
+	c.RebindUDPServer()
+	assert.Equal(t, 1, conn.rebinds, "rebind while started must reach the conn")
+
+	// A rebind racing a completed stop must not touch the closed conn
+	c.Stop()
+	require.NoError(t, c.Wait())
+	c.RebindUDPServer()
+	assert.Equal(t, 1, conn.rebinds, "rebind after stop must be a no-op")
+}

control_test.go

@@ -1,6 +1,8 @@
 package nebula
 
 import (
+	"bytes"
+	"log/slog"
 	"net"
 	"net/netip"
 	"reflect"
@@ -9,6 +11,7 @@ import (
 	"github.com/slackhq/nebula/cert"
 	"github.com/slackhq/nebula/test"
 	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
 )
 
 func TestControl_GetHostInfoByVpnIp(t *testing.T) {
@@ -42,8 +45,7 @@ func TestControl_GetHostInfoByVpnIp(t *testing.T) {
 	assert.True(t, ok)
 
 	crt := &dummyCert{}
-	hm.unlockedAddHostInfo(&HostInfo{
+	hi := &HostInfo{
-		remote:  remote1,
 		remotes: remotes,
 		ConnectionState: &ConnectionState{
 			peerCert: &cert.CachedCertificate{Certificate: crt},
@@ -56,13 +58,14 @@ func TestControl_GetHostInfoByVpnIp(t *testing.T) {
 			relayForByAddr: map[netip.Addr]*Relay{},
 			relayForByIdx:  map[uint32]*Relay{},
 		},
-	}, &Interface{})
+	}
+	hi.remote.Store(&remote1)
+	hm.unlockedAddHostInfo(hi, &Interface{})
 
 	vpnIp2, ok := netip.AddrFromSlice(ipNet2.IP)
 	assert.True(t, ok)
 
-	hm.unlockedAddHostInfo(&HostInfo{
+	hi2 := &HostInfo{
-		remote:  remote1,
 		remotes: remotes,
 		ConnectionState: &ConnectionState{
 			peerCert: nil,
@@ -75,7 +78,9 @@ func TestControl_GetHostInfoByVpnIp(t *testing.T) {
 			relayForByAddr: map[netip.Addr]*Relay{},
 			relayForByIdx:  map[uint32]*Relay{},
 		},
-	}, &Interface{})
+	}
+	hi2.remote.Store(&remote1)
+	hm.unlockedAddHostInfo(hi2, &Interface{})
 
 	c := Control{
 		state: StateReady,
@@ -119,3 +124,153 @@ func assertFields(t *testing.T, expected []string, actualStruct any) {
 
 	assert.Equal(t, expected, fields)
 }
+
+// alwaysAllowV4/V6 are check funcs that accept every entry (including nil pointers),
+// letting us inject a nil *V4AddrPort/*V6AddrPort into a RemoteList's reported cache
+// the same way a malformed proto message off the wire could.
+func alwaysAllowV4(netip.Addr, *V4AddrPort) bool { return true }
+func alwaysAllowV6(netip.Addr, *V6AddrPort) bool { return true }
+
+// TestGetRelays_SkipsNilRelayAddrs proves GetRelays tolerates nil entries in the
+// RelayVpnAddrs proto slice (which protoAddrToNetAddr would nil-deref on) and still
+// returns the valid relays, including the legacy OldRelayVpnAddrs.
+func TestGetRelays_SkipsNilRelayAddrs(t *testing.T) {
+	good := netip.MustParseAddr("10.0.0.9")
+
+	d := &NebulaMetaDetails{
+		OldRelayVpnAddrs: []uint32{0x0a000001}, // 10.0.0.1
+		RelayVpnAddrs: []*Addr{
+			nil,
+			netAddrToProtoAddr(good),
+			nil,
+		},
+	}
+
+	var relays []netip.Addr
+	require.NotPanics(t, func() { relays = d.GetRelays() })
+
+	assert.Equal(t, []netip.Addr{
+		netip.MustParseAddr("10.0.0.1"),
+		good,
+	}, relays)
+}
+
+// TestGetRelays_AllNil ensures an all-nil RelayVpnAddrs slice yields no relays and no panic.
+func TestGetRelays_AllNil(t *testing.T) {
+	d := &NebulaMetaDetails{RelayVpnAddrs: []*Addr{nil, nil}}
+	var relays []netip.Addr
+	require.NotPanics(t, func() { relays = d.GetRelays() })
+	assert.Empty(t, relays)
+}
+
+// TestRemoteList_CopyCache_SkipsNilReported proves CopyCache skips nil reported
+// pointers (v4 and v6) instead of nil-dereferencing them in protoV*AddrPortToNetAddrPort.
+func TestRemoteList_CopyCache_SkipsNilReported(t *testing.T) {
+	owner := netip.MustParseAddr("10.0.0.1")
+	rl := NewRemoteList([]netip.Addr{owner}, nil)
+
+	rl.unlockedSetV4(owner, owner, []*V4AddrPort{
+		nil,
+		newIp4AndPortFromString("1.2.3.4:5"),
+		nil,
+	}, alwaysAllowV4)
+
+	rl.unlockedSetV6(owner, owner, []*V6AddrPort{
+		nil,
+		newIp6AndPortFromString("[1::1]:6"),
+		nil,
+	}, alwaysAllowV6)
+
+	var cm *CacheMap
+	require.NotPanics(t, func() { cm = rl.CopyCache() })
+
+	c := (*cm)[owner.String()]
+	require.NotNil(t, c)
+	assert.ElementsMatch(t, []netip.AddrPort{
+		netip.MustParseAddrPort("1.2.3.4:5"),
+		netip.MustParseAddrPort("[1::1]:6"),
+	}, c.Reported)
+}
+
+// TestRemoteList_Rebuild_SkipsNilReported drives unlockedCollect (via Rebuild) with
+// nil reported entries and confirms only the valid addresses survive, with no panic.
+func TestRemoteList_Rebuild_SkipsNilReported(t *testing.T) {
+	owner := netip.MustParseAddr("10.0.0.1")
+	rl := NewRemoteList([]netip.Addr{owner}, nil)
+
+	rl.unlockedSetV4(owner, owner, []*V4AddrPort{
+		nil,
+		newIp4AndPortFromString("1.2.3.4:5"),
+	}, alwaysAllowV4)
+	rl.unlockedSetV6(owner, owner, []*V6AddrPort{
+		newIp6AndPortFromString("[1::1]:6"),
+		nil,
+	}, alwaysAllowV6)
+
+	require.NotPanics(t, func() { rl.Rebuild([]netip.Prefix{}) })
+
+	assert.ElementsMatch(t, []netip.AddrPort{
+		netip.MustParseAddrPort("1.2.3.4:5"),
+		netip.MustParseAddrPort("[1::1]:6"),
+	}, rl.addrs)
+}
+
+// newRelayControl marshals a NebulaControl the way it arrives on the wire so we can feed
+// it through HandleControlMsg's unmarshal + validate path.
+func newRelayControl(t *testing.T, typ NebulaControl_MessageType, from, to *Addr) []byte {
+	t.Helper()
+	msg := &NebulaControl{
+		Type:          typ,
+		RelayFromAddr: from,
+		RelayToAddr:   to,
+	}
+	b, err := msg.Marshal()
+	require.NoError(t, err)
+	return b
+}
+
+// TestRelayManager_HandleControlMsg_NilRelayAddrs verifies the validation block added to
+// HandleControlMsg: CreateRelay{Request,Response} carrying a nil RelayFromAddr or
+// RelayToAddr are dropped with a debug log rather than nil-dereferencing downstream.
+func TestRelayManager_HandleControlMsg_NilRelayAddrs(t *testing.T) {
+	good := netAddrToProtoAddr(netip.MustParseAddr("10.0.0.9"))
+
+	cases := []struct {
+		name    string
+		typ     NebulaControl_MessageType
+		from    *Addr
+		to      *Addr
+		wantLog string // debug substring expected, "" == expect no drop log
+	}{
+		{"request nil from", NebulaControl_CreateRelayRequest, nil, good, "nil RelayFromAddr"},
+		{"request nil to", NebulaControl_CreateRelayRequest, good, nil, "nil RelayToAddr"},
+		{"request both nil", NebulaControl_CreateRelayRequest, nil, nil, "nil RelayFromAddr"},
+		{"response nil from", NebulaControl_CreateRelayResponse, nil, good, "nil RelayFromAddr"},
+		{"response nil to", NebulaControl_CreateRelayResponse, good, nil, "nil RelayToAddr"},
+		// A non-relay control type is not subject to the relay-addr validation and must
+		// pass through it untouched (the final switch simply no-ops on it).
+		{"unrelated type nil addrs", NebulaControl_None, nil, nil, ""},
+	}
+
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			var buf bytes.Buffer
+			l := test.NewLoggerWithOutputAndLevel(&buf, slog.LevelDebug)
+			rm := &relayManager{l: l, hostmap: newHostMap(l)}
+			rm.useRelays.Store(true)
+
+			f := &Interface{l: l}
+			h := &HostInfo{vpnAddrs: []netip.Addr{netip.MustParseAddr("10.0.0.2")}, localIndexId: 1}
+
+			d := newRelayControl(t, tc.typ, tc.from, tc.to)
+
+			require.NotPanics(t, func() { rm.HandleControlMsg(h, d, f) })
+
+			if tc.wantLog == "" {
+				assert.NotContains(t, buf.String(), "nil Relay")
+			} else {
+				assert.Contains(t, buf.String(), tc.wantLog)
+			}
+		})
+	}
+}

control_tester.go

@@ -125,6 +125,14 @@ func (c *Control) GetHostmap() *HostMap {
 	return c.f.hostMap
 }
 
+// GetHostmapIndexCount returns the number of entries in the main hostmap Indexes table, holding
+// the hostmap read lock so tests can poll it while connection manager churns tunnels.
+func (c *Control) GetHostmapIndexCount() int {
+	c.f.hostMap.RLock()
+	defer c.f.hostMap.RUnlock()
+	return len(c.f.hostMap.Indexes)
+}
+
 func (c *Control) GetF() *Interface {
 	return c.f
 }

dns_server.go

@@ -97,8 +97,7 @@ func (d *dnsServer) reload(c *config.C, initial bool) error {
 	newAddr := getDnsServerAddr(c)
 
 	d.serverMu.Lock()
-	running := d.server
+	running := d.server != nil
-	runningStarted := d.started
 	sameAddr := d.addr == newAddr
 	d.addr = newAddr
 	d.enabled.Store(enabled)
@@ -112,7 +111,7 @@ func (d *dnsServer) reload(c *config.C, initial bool) error {
 	}
 
 	if !enabled {
-		if running != nil {
+		if running {
 			d.Stop()
 		}
 		// Drop any records that accumulated while enabled; a later re-enable
@@ -121,12 +120,12 @@ func (d *dnsServer) reload(c *config.C, initial bool) error {
 		return nil
 	}
 
-	if running == nil {
+	if !running {
 		// Was disabled (or never started); bring it up now.
 		go d.Start()
 	} else if !sameAddr {
-		d.shutdownServer(running, runningStarted, "reload")
+		// Stop clears the slot before shutting down, otherwise the Start below can find the dying server and refuse
-		// Old Start goroutine has now exited; bring up a fresh listener on the new address.
+		d.Stop()
 		go d.Start()
 	}
 
@@ -162,7 +161,9 @@ func (d *dnsServer) Start() {
 
 	started := make(chan struct{})
 	d.serverMu.Lock()
-	if d.ctx.Err() != nil {
+	// Re-check enabled under the lock, a disable that raced our check above snapshots the slot under it too.
+	// Two reloads in quick succession can both spawn a Start, the loser would orphan the live listener past Stop
+	if d.ctx.Err() != nil || d.server != nil || !d.enabled.Load() {
 		d.serverMu.Unlock()
 		return
 	}
@@ -200,6 +201,14 @@ func (d *dnsServer) Start() {
 		close(started)
 	}
 
+	// Release our slot, unless a reload already replaced us, so a dead listener can't block a future Start
+	d.serverMu.Lock()
+	if d.server == server {
+		d.server = nil
+		d.started = nil
+	}
+	d.serverMu.Unlock()
+
 	if err != nil {
 		d.l.Warn("Failed to run the DNS responder", "error", err)
 	}

dns_server_test.go

@@ -194,14 +194,51 @@ func TestDnsServer_reload_initial_serveDnsWithoutLighthouse(t *testing.T) {
 }
 
 func TestDnsServer_reload_sameAddr_noOp(t *testing.T) {
+	port := freeUDPPort(t)
 	ds, c := newTestDnsServer(t)
-	setDnsConfig(c, "127.0.0.1", "0", true, true)
+	setDnsConfig(c, "127.0.0.1", port, true, true)
-
 	require.NoError(t, ds.reload(c, true))
-	// No server running yet, no addr change. Reload should not spawn anything.
+
+	go ds.Start()
+	waitForBind(t, ds)
+
+	ds.serverMu.Lock()
+	before := ds.server
+	ds.serverMu.Unlock()
+	require.NotNil(t, before)
+
+	// Same address, so the running listener must be left alone rather than rebuilt under live queries
 	require.NoError(t, ds.reload(c, false))
 	assert.True(t, ds.enabled.Load())
-	assert.Nil(t, ds.server)
+
+	ds.serverMu.Lock()
+	after := ds.server
+	ds.serverMu.Unlock()
+	assert.Same(t, before, after, "a same-address reload must not restart the listener")
+
+	ds.Stop()
+}
+
+// The branch the old sameAddr test was accidentally hitting: enabled with nothing running means reload starts it.
+func TestDnsServer_reload_whenNotRunning_starts(t *testing.T) {
+	port := freeUDPPort(t)
+	ds, c := newTestDnsServer(t)
+	setDnsConfig(c, "127.0.0.1", port, true, true)
+
+	// initial only records config, it never starts anything
+	require.NoError(t, ds.reload(c, true))
+	ds.serverMu.Lock()
+	assert.Nil(t, ds.server, "the initial reload must not start a listener")
+	ds.serverMu.Unlock()
+
+	require.NoError(t, ds.reload(c, false))
+	waitForBind(t, ds)
+
+	ds.serverMu.Lock()
+	assert.NotNil(t, ds.server, "a reload with nothing running should bring DNS up")
+	ds.serverMu.Unlock()
+
+	ds.Stop()
 }
 
 func TestDnsServer_StartStop_lifecycle(t *testing.T) {
@@ -427,3 +464,168 @@ func waitFor(t *testing.T, cond func() bool) {
 	}
 	t.Fatal("timed out waiting for condition")
 }
+
+// Two reloads in quick succession, or a HUP before Control.Start, can race two Starts at the same listener.
+func TestDnsServer_Start_isIdempotent(t *testing.T) {
+	port := freeUDPPort(t)
+	ds, c := newTestDnsServer(t)
+	setDnsConfig(c, "127.0.0.1", port, true, true)
+	require.NoError(t, ds.reload(c, true))
+
+	go ds.Start()
+	waitForBind(t, ds)
+
+	ds.serverMu.Lock()
+	first := ds.server
+	ds.serverMu.Unlock()
+	require.NotNil(t, first)
+
+	// If the second Start replaces the tracked server, Stop kills the wrong one and the port leaks
+	done := make(chan struct{})
+	go func() {
+		ds.Start()
+		close(done)
+	}()
+	select {
+	case <-done:
+	case <-time.After(time.Second * 5):
+		t.Fatal("second Start never returned")
+	}
+
+	ds.serverMu.Lock()
+	second := ds.server
+	ds.serverMu.Unlock()
+	assert.Same(t, first, second, "a second Start must not replace the running server")
+
+	// The real proof, after Stop the port must actually be free
+	ds.Stop()
+	waitFor(t, func() bool {
+		pc, err := net.ListenPacket("udp", "127.0.0.1:"+port)
+		if err != nil {
+			return false
+		}
+		_ = pc.Close()
+		return true
+	})
+}
+
+// An address change must actually end up listening on the new port. Start's guard refuses when a server is already
+// installed, so reload has to clear the slot before shutting the old one down.
+func TestDnsServer_reload_addrChange_restarts(t *testing.T) {
+	first := freeUDPPort(t)
+	second := freeUDPPort(t)
+
+	ds, c := newTestDnsServer(t)
+	setDnsConfig(c, "127.0.0.1", first, true, true)
+	require.NoError(t, ds.reload(c, true))
+
+	go ds.Start()
+	waitForBind(t, ds)
+
+	// Cycle a few times, the failure this guards against depends on which goroutine wins serverMu
+	for i := range 8 {
+		want := second
+		if i%2 == 1 {
+			want = first
+		}
+		setDnsConfig(c, "127.0.0.1", want, true, true)
+		require.NoError(t, ds.reload(c, false))
+		waitForBind(t, ds)
+
+		ds.serverMu.Lock()
+		srv := ds.server
+		ds.serverMu.Unlock()
+		require.NotNil(t, srv, "reload left DNS down instead of restarting it")
+		require.Equal(t, "127.0.0.1:"+want, srv.Addr, "reload should be serving the new address")
+	}
+
+	// Land back on second so the port assertions below are meaningful
+	setDnsConfig(c, "127.0.0.1", second, true, true)
+	require.NoError(t, ds.reload(c, false))
+	waitForBind(t, ds)
+
+	// The old port must be released and the new one actually held
+	waitFor(t, func() bool {
+		pc, err := net.ListenPacket("udp", "127.0.0.1:"+first)
+		if err != nil {
+			return false
+		}
+		_ = pc.Close()
+		return true
+	})
+	_, err := net.ListenPacket("udp", "127.0.0.1:"+second)
+	require.Error(t, err, "the new address should be bound by the DNS responder")
+
+	ds.Stop()
+}
+
+// A listener that dies on its own must release the slot, or a later same-addr reload sees it as running and no-ops.
+func TestDnsServer_Start_bindFailure_releasesSlot(t *testing.T) {
+	port := freeUDPPort(t)
+	blocker, err := net.ListenPacket("udp", "127.0.0.1:"+port)
+	require.NoError(t, err)
+
+	ds, c := newTestDnsServer(t)
+	setDnsConfig(c, "127.0.0.1", port, true, true)
+	require.NoError(t, ds.reload(c, true))
+
+	ds.Start() // returns once the bind fails
+
+	ds.serverMu.Lock()
+	assert.Nil(t, ds.server, "a listener that failed to bind must not stay parked in the slot")
+	ds.serverMu.Unlock()
+
+	// With the slot released, a reload can retry once the port frees up
+	require.NoError(t, blocker.Close())
+	require.NoError(t, ds.reload(c, false))
+	waitForBind(t, ds)
+
+	ds.serverMu.Lock()
+	assert.NotNil(t, ds.server, "a same-addr reload should retry after a failed bind")
+	ds.serverMu.Unlock()
+
+	ds.Stop()
+}
+
+// A disable that lands while Start is between its unlocked check and the guard must not leave a listener behind.
+func TestDnsServer_Start_refusesWhenDisabledUnderLock(t *testing.T) {
+	port := freeUDPPort(t)
+	ds, c := newTestDnsServer(t)
+	setDnsConfig(c, "127.0.0.1", port, true, true)
+	require.NoError(t, ds.reload(c, true))
+	require.True(t, ds.enabled.Load())
+
+	// Holding serverMu parks Start on the lock, the only way to land the disable in that window on purpose
+	ds.serverMu.Lock()
+
+	done := make(chan struct{})
+	go func() {
+		ds.Start()
+		close(done)
+	}()
+
+	select {
+	case <-done:
+		ds.serverMu.Unlock()
+		t.Fatal("Start returned early, the test never exercised the window")
+	case <-time.After(time.Millisecond * 100):
+	}
+
+	// The disable reload's critical section. It sees nothing running, so it never calls Stop.
+	ds.enabled.Store(false)
+	ds.serverMu.Unlock()
+
+	select {
+	case <-done:
+	case <-time.After(time.Second * 5):
+		t.Fatal("Start never returned")
+	}
+
+	ds.serverMu.Lock()
+	assert.Nil(t, ds.server, "Start must not install a listener a disable already cancelled")
+	ds.serverMu.Unlock()
+
+	pc, err := net.ListenPacket("udp", "127.0.0.1:"+port)
+	require.NoError(t, err, "an orphaned listener is still holding the port")
+	_ = pc.Close()
+}

e2e/echo_test.go

@@ -0,0 +1,85 @@
+//go:build e2e_testing
+// +build e2e_testing
+
+package e2e
+
+import (
+	"testing"
+	"time"
+
+	"github.com/slackhq/nebula"
+	"github.com/slackhq/nebula/cert"
+	"github.com/slackhq/nebula/cert_test"
+	"github.com/slackhq/nebula/e2e/router"
+	"github.com/slackhq/nebula/header"
+	"github.com/slackhq/nebula/udp"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func assertTestRequestEchoed(t *testing.T, cipher string) {
+	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version1, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
+	over := m{"cipher": cipher}
+	a, aNet, aUdp, _ := newSimpleServer(cert.Version1, ca, caKey, "a", "10.128.0.1/24", over)
+	b, bNet, bUdp, _ := newSimpleServer(cert.Version1, ca, caKey, "b", "10.128.0.2/24", over)
+
+	a.InjectLightHouseAddr(bNet[0].Addr(), bUdp)
+	b.InjectLightHouseAddr(aNet[0].Addr(), aUdp)
+	a.Start()
+	b.Start()
+	t.Cleanup(func() { a.Stop(); b.Stop() })
+	r := router.NewR(t, a, b)
+	defer r.RenderFlow()
+
+	assertTunnel(t, aNet[0].Addr(), bNet[0].Addr(), a, b, r)
+	drainUDPTx(a)
+	drainUDPTx(b)
+
+	payload := []byte("a test payload well over sixteen bytes long, wow it's so very long long long!")
+	require.Greater(t, len(payload), header.Len)
+	a.GetF().SendMessageToVpnAddr(header.Test, header.TestRequest, bNet[0].Addr(), payload, make([]byte, 12, 12), make([]byte, udp.MTU))
+
+	// Deliver A's request to B; B must echo a reply back
+	b.InjectUDPPacket(a.GetFromUDP(true))
+	reply := nextUDPTxOfType(t, b, header.Test, header.TestReply, 2*time.Second)
+
+	assert.Equal(t, aUdp, reply.To, "the reply must go back to the requester")
+	// header + echoed payload + 16-byte AEAD tag: proves the whole payload
+	// round-tripped rather than being dropped or truncated.
+	assert.Equal(t, header.Len+len(payload)+16, len(reply.Data), "the full payload must be echoed back")
+}
+
+func TestTestRequestEchoesLongPayloadAES(t *testing.T) {
+	assertTestRequestEchoed(t, "aes")
+}
+
+func TestTestRequestEchoesLongPayloadChaChaPoly(t *testing.T) {
+	assertTestRequestEchoed(t, "chachapoly")
+}
+
+// drainUDPTx empties a control's UDP tx queue without blocking.
+func drainUDPTx(c *nebula.Control) {
+	for c.GetFromUDP(false) != nil {
+	}
+}
+
+// nextUDPTxOfType returns the next packet a control transmits whose nebula
+// header matches (wantType, wantSub), skipping unrelated packets.
+// It fails the test if none arrives within the timeout.
+func nextUDPTxOfType(t *testing.T, c *nebula.Control, wantType header.MessageType, wantSub header.MessageSubType, within time.Duration) *udp.Packet {
+	t.Helper()
+	ch := c.GetUDPTxChan()
+	timeout := time.After(within)
+	for {
+		select {
+		case p := <-ch:
+			var h header.H
+			if err := h.Parse(p.Data); err == nil && h.Type == wantType && h.Subtype == wantSub {
+				return p
+			}
+		case <-timeout:
+			t.Fatalf("timed out waiting for a %v/%v packet on the udp tx queue", wantType, wantSub)
+			return nil
+		}
+	}
+}

e2e/handshakes_test.go

@@ -405,7 +405,7 @@ func TestStage1Race(t *testing.T) {
 
 	r.Log("Spin until connection manager tears down a tunnel")
 
-	for len(myControl.GetHostmap().Indexes)+len(theirControl.GetHostmap().Indexes) > 2 {
+	for myControl.GetHostmapIndexCount()+theirControl.GetHostmapIndexCount() > 2 {
 		assertTunnel(t, myVpnIpNet[0].Addr(), theirVpnIpNet[0].Addr(), myControl, theirControl, r)
 		t.Log("Connection manager hasn't ticked yet")
 		time.Sleep(time.Second)
@@ -453,9 +453,11 @@ func TestUncleanShutdownRaceLoser(t *testing.T) {
 
 	r.Log("Nuke my hostmap")
 	myHostmap := myControl.GetHostmap()
+	myHostmap.Lock()
 	myHostmap.Hosts = map[netip.Addr]*nebula.HostInfo{}
 	myHostmap.Indexes = map[uint32]*nebula.HostInfo{}
 	myHostmap.RemoteIndexes = map[uint32]*nebula.HostInfo{}
+	myHostmap.Unlock()
 
 	myControl.InjectTunPacket(BuildTunUDPPacket(theirVpnIpNet[0].Addr(), 80, myVpnIpNet[0].Addr(), 80, []byte("Hi from me again")))
 	p = r.RouteForAllUntilTxTun(theirControl)
@@ -465,10 +467,10 @@ func TestUncleanShutdownRaceLoser(t *testing.T) {
 	assertTunnel(t, myVpnIpNet[0].Addr(), theirVpnIpNet[0].Addr(), myControl, theirControl, r)
 
 	r.Log("Wait for the dead index to go away")
-	start := len(theirControl.GetHostmap().Indexes)
+	start := theirControl.GetHostmapIndexCount()
 	for {
 		assertTunnel(t, myVpnIpNet[0].Addr(), theirVpnIpNet[0].Addr(), myControl, theirControl, r)
-		if len(theirControl.GetHostmap().Indexes) < start {
+		if theirControl.GetHostmapIndexCount() < start {
 			break
 		}
 		time.Sleep(time.Second)
@@ -504,9 +506,11 @@ func TestUncleanShutdownRaceWinner(t *testing.T) {
 
 	r.Log("Nuke my hostmap")
 	theirHostmap := theirControl.GetHostmap()
+	theirHostmap.Lock()
 	theirHostmap.Hosts = map[netip.Addr]*nebula.HostInfo{}
 	theirHostmap.Indexes = map[uint32]*nebula.HostInfo{}
 	theirHostmap.RemoteIndexes = map[uint32]*nebula.HostInfo{}
+	theirHostmap.Unlock()
 
 	theirControl.InjectTunPacket(BuildTunUDPPacket(myVpnIpNet[0].Addr(), 80, theirVpnIpNet[0].Addr(), 80, []byte("Hi from them again")))
 	p = r.RouteForAllUntilTxTun(myControl)
@@ -517,10 +521,10 @@ func TestUncleanShutdownRaceWinner(t *testing.T) {
 	assertTunnel(t, myVpnIpNet[0].Addr(), theirVpnIpNet[0].Addr(), myControl, theirControl, r)
 
 	r.Log("Wait for the dead index to go away")
-	start := len(myControl.GetHostmap().Indexes)
+	start := myControl.GetHostmapIndexCount()
 	for {
 		assertTunnel(t, myVpnIpNet[0].Addr(), theirVpnIpNet[0].Addr(), myControl, theirControl, r)
-		if len(myControl.GetHostmap().Indexes) < start {
+		if myControl.GetHostmapIndexCount() < start {
 			break
 		}
 		time.Sleep(time.Second)
@@ -628,10 +632,10 @@ func TestReestablishRelays(t *testing.T) {
 	r.Log("Close the tunnel")
 	relayControl.CloseTunnel(theirVpnIpNet[0].Addr(), true)
 
-	start := len(myControl.GetHostmap().Indexes)
+	start := myControl.GetHostmapIndexCount()
-	curIndexes := len(myControl.GetHostmap().Indexes)
+	curIndexes := myControl.GetHostmapIndexCount()
 	for curIndexes >= start {
-		curIndexes = len(myControl.GetHostmap().Indexes)
+		curIndexes = myControl.GetHostmapIndexCount()
 		r.Logf("Wait for the dead index to go away:start=%v indexes, current=%v indexes", start, curIndexes)
 		myControl.InjectTunPacket(BuildTunUDPPacket(theirVpnIpNet[0].Addr(), 80, myVpnIpNet[0].Addr(), 80, []byte("Hi from me should fail")))
 
@@ -721,6 +725,70 @@ func TestReestablishRelays(t *testing.T) {
 
 }
 
+func TestRelayHandshakeOverDisestablishedEntry(t *testing.T) {
+	t.Parallel()
+	// If them tears down the tunnel while me keeps Established relay state, me's next
+	// handshake flows through the relay with no fresh CreateRelayRequest and lands on
+	// them's Disestablished terminal relay entry. them must re-establish that entry, or
+	// its first transmit deletes its only relay and the tunnel is born transmit-dead:
+	// them can receive but every send is silently dropped.
+	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}})
+	theirControl, theirVpnIpNet, theirUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "them   ", "10.128.0.2/24", m{"relay": m{"use_relays": true}})
+
+	// Teach my how to get to the relay and that their can be reached via the relay
+	myControl.InjectLightHouseAddr(relayVpnIpNet[0].Addr(), relayUdpAddr)
+	myControl.InjectRelays(theirVpnIpNet[0].Addr(), []netip.Addr{relayVpnIpNet[0].Addr()})
+	relayControl.InjectLightHouseAddr(theirVpnIpNet[0].Addr(), theirUdpAddr)
+
+	// Build a router so we don't have to reason who gets which packet
+	r := router.NewR(t, myControl, relayControl, theirControl)
+	defer r.RenderFlow()
+
+	// Start the servers
+	myControl.Start()
+	relayControl.Start()
+	theirControl.Start()
+
+	t.Log("Trigger a handshake from me to them via the relay")
+	myControl.InjectTunPacket(BuildTunUDPPacket(theirVpnIpNet[0].Addr(), 80, myVpnIpNet[0].Addr(), 80, []byte("Hi from me")))
+
+	p := r.RouteForAllUntilTxTun(theirControl)
+	assertUdpPacket(t, []byte("Hi from me"), p, myVpnIpNet[0].Addr(), theirVpnIpNet[0].Addr(), 80, 80)
+	oldIdx := myControl.GetHostInfoByVpnAddr(theirVpnIpNet[0].Addr(), false).LocalIndex
+
+	t.Log("Close the tunnel on them only, marking their relay entry Disestablished")
+	theirControl.CloseTunnel(myVpnIpNet[0].Addr(), true)
+
+	t.Log("Re-handshake from me, riding the still-Established relay state")
+	myControl.ReHandshake(theirVpnIpNet[0].Addr())
+	for {
+		h := myControl.GetHostInfoByVpnAddr(theirVpnIpNet[0].Addr(), false)
+		if h != nil && h.LocalIndex != oldIdx && h.RemoteIndex != 0 {
+			break
+		}
+		r.RouteForAllExitFunc(func(*udp.Packet, *nebula.Control) router.ExitType {
+			return router.RouteAndExit
+		})
+	}
+
+	hAtThem := theirControl.GetHostInfoByVpnAddr(myVpnIpNet[0].Addr(), false)
+	require.NotNil(t, hAtThem, "them should have completed the relayed handshake")
+	require.Equal(t, []netip.Addr{relayVpnIpNet[0].Addr()}, hAtThem.CurrentRelaysToMe, "them should know a relay for the new tunnel")
+
+	t.Log("Send from them to me; their only relay entry must survive the transmit")
+	theirControl.InjectTunPacket(BuildTunUDPPacket(myVpnIpNet[0].Addr(), 80, theirVpnIpNet[0].Addr(), 80, []byte("Hi from them")))
+	require.Never(t, func() bool {
+		h := theirControl.GetHostInfoByVpnAddr(myVpnIpNet[0].Addr(), false)
+		return h == nil || len(h.CurrentRelaysToMe) == 0
+	}, time.Second, 10*time.Millisecond, "them deleted its only relay entry; the tunnel is permanently transmit-dead")
+
+	p = r.RouteForAllUntilTxTun(myControl)
+	assertUdpPacket(t, []byte("Hi from them"), p, theirVpnIpNet[0].Addr(), myVpnIpNet[0].Addr(), 80, 80)
+	r.RenderHostmaps("Final hostmaps", myControl, relayControl, theirControl)
+}
+
 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
@@ -819,18 +887,18 @@ func TestStage1RaceRelays2(t *testing.T) {
 
 	t.Log("Wait until we remove extra tunnels")
 	t.Logf("Waiting for hostinfos to be removed... myControl=%d theirControl=%d relayControl=%d",
-		len(myControl.GetHostmap().Indexes),
+		myControl.GetHostmapIndexCount(),
-		len(theirControl.GetHostmap().Indexes),
+		theirControl.GetHostmapIndexCount(),
-		len(relayControl.GetHostmap().Indexes),
+		relayControl.GetHostmapIndexCount(),
 	)
-	hostInfos := len(myControl.GetHostmap().Indexes) + len(theirControl.GetHostmap().Indexes) + len(relayControl.GetHostmap().Indexes)
+	hostInfos := myControl.GetHostmapIndexCount() + theirControl.GetHostmapIndexCount() + relayControl.GetHostmapIndexCount()
 	retries := 60
 	for hostInfos > 6 && retries > 0 {
-		hostInfos = len(myControl.GetHostmap().Indexes) + len(theirControl.GetHostmap().Indexes) + len(relayControl.GetHostmap().Indexes)
+		hostInfos = myControl.GetHostmapIndexCount() + theirControl.GetHostmapIndexCount() + relayControl.GetHostmapIndexCount()
 		t.Logf("Waiting for hostinfos to be removed... myControl=%d theirControl=%d relayControl=%d",
-			len(myControl.GetHostmap().Indexes),
+			myControl.GetHostmapIndexCount(),
-			len(theirControl.GetHostmap().Indexes),
+			theirControl.GetHostmapIndexCount(),
-			len(relayControl.GetHostmap().Indexes),
+			relayControl.GetHostmapIndexCount(),
 		)
 		assertTunnel(t, myVpnIpNet[0].Addr(), theirVpnIpNet[0].Addr(), myControl, theirControl, r)
 		t.Log("Connection manager hasn't ticked yet")
@@ -924,24 +992,24 @@ func TestRehandshakingRelays(t *testing.T) {
 	assertTunnel(t, theirVpnIpNet[0].Addr(), myVpnIpNet[0].Addr(), theirControl, myControl, r)
 	r.RenderHostmaps("working hostmaps", myControl, relayControl, theirControl)
 	// We should have two hostinfos on all sides
-	for len(myControl.GetHostmap().Indexes) != 2 {
+	for myControl.GetHostmapIndexCount() != 2 {
-		t.Logf("Waiting for myControl hostinfos (%v != 2) to get cleaned up from lack of use...", len(myControl.GetHostmap().Indexes))
+		t.Logf("Waiting for myControl hostinfos (%v != 2) to get cleaned up from lack of use...", myControl.GetHostmapIndexCount())
 		r.Log("Assert the relay tunnel still works")
 		assertTunnel(t, theirVpnIpNet[0].Addr(), myVpnIpNet[0].Addr(), theirControl, myControl, r)
 		r.Log("yupitdoes")
 		time.Sleep(time.Second)
 	}
 	t.Logf("myControl hostinfos got cleaned up!")
-	for len(theirControl.GetHostmap().Indexes) != 2 {
+	for theirControl.GetHostmapIndexCount() != 2 {
-		t.Logf("Waiting for theirControl hostinfos (%v != 2) to get cleaned up from lack of use...", len(theirControl.GetHostmap().Indexes))
+		t.Logf("Waiting for theirControl hostinfos (%v != 2) to get cleaned up from lack of use...", theirControl.GetHostmapIndexCount())
 		r.Log("Assert the relay tunnel still works")
 		assertTunnel(t, theirVpnIpNet[0].Addr(), myVpnIpNet[0].Addr(), theirControl, myControl, r)
 		r.Log("yupitdoes")
 		time.Sleep(time.Second)
 	}
 	t.Logf("theirControl hostinfos got cleaned up!")
-	for len(relayControl.GetHostmap().Indexes) != 2 {
+	for relayControl.GetHostmapIndexCount() != 2 {
-		t.Logf("Waiting for relayControl hostinfos (%v != 2) to get cleaned up from lack of use...", len(relayControl.GetHostmap().Indexes))
+		t.Logf("Waiting for relayControl hostinfos (%v != 2) to get cleaned up from lack of use...", relayControl.GetHostmapIndexCount())
 		r.Log("Assert the relay tunnel still works")
 		assertTunnel(t, theirVpnIpNet[0].Addr(), myVpnIpNet[0].Addr(), theirControl, myControl, r)
 		r.Log("yupitdoes")
@@ -1029,24 +1097,24 @@ func TestRehandshakingRelaysPrimary(t *testing.T) {
 	assertTunnel(t, theirVpnIpNet[0].Addr(), myVpnIpNet[0].Addr(), theirControl, myControl, r)
 	r.RenderHostmaps("working hostmaps", myControl, relayControl, theirControl)
 	// We should have two hostinfos on all sides
-	for len(myControl.GetHostmap().Indexes) != 2 {
+	for myControl.GetHostmapIndexCount() != 2 {
-		t.Logf("Waiting for myControl hostinfos (%v != 2) to get cleaned up from lack of use...", len(myControl.GetHostmap().Indexes))
+		t.Logf("Waiting for myControl hostinfos (%v != 2) to get cleaned up from lack of use...", myControl.GetHostmapIndexCount())
 		r.Log("Assert the relay tunnel still works")
 		assertTunnel(t, theirVpnIpNet[0].Addr(), myVpnIpNet[0].Addr(), theirControl, myControl, r)
 		r.Log("yupitdoes")
 		time.Sleep(time.Second)
 	}
 	t.Logf("myControl hostinfos got cleaned up!")
-	for len(theirControl.GetHostmap().Indexes) != 2 {
+	for theirControl.GetHostmapIndexCount() != 2 {
-		t.Logf("Waiting for theirControl hostinfos (%v != 2) to get cleaned up from lack of use...", len(theirControl.GetHostmap().Indexes))
+		t.Logf("Waiting for theirControl hostinfos (%v != 2) to get cleaned up from lack of use...", theirControl.GetHostmapIndexCount())
 		r.Log("Assert the relay tunnel still works")
 		assertTunnel(t, theirVpnIpNet[0].Addr(), myVpnIpNet[0].Addr(), theirControl, myControl, r)
 		r.Log("yupitdoes")
 		time.Sleep(time.Second)
 	}
 	t.Logf("theirControl hostinfos got cleaned up!")
-	for len(relayControl.GetHostmap().Indexes) != 2 {
+	for relayControl.GetHostmapIndexCount() != 2 {
-		t.Logf("Waiting for relayControl hostinfos (%v != 2) to get cleaned up from lack of use...", len(relayControl.GetHostmap().Indexes))
+		t.Logf("Waiting for relayControl hostinfos (%v != 2) to get cleaned up from lack of use...", relayControl.GetHostmapIndexCount())
 		r.Log("Assert the relay tunnel still works")
 		assertTunnel(t, theirVpnIpNet[0].Addr(), myVpnIpNet[0].Addr(), theirControl, myControl, r)
 		r.Log("yupitdoes")
@@ -1123,7 +1191,7 @@ func TestRehandshaking(t *testing.T) {
 	theirConfig.ReloadConfigString(string(rc))
 
 	r.Log("Spin until there is only 1 tunnel")
-	for len(myControl.GetHostmap().Indexes)+len(theirControl.GetHostmap().Indexes) > 2 {
+	for myControl.GetHostmapIndexCount()+theirControl.GetHostmapIndexCount() > 2 {
 		assertTunnel(t, myVpnIpNet[0].Addr(), theirVpnIpNet[0].Addr(), myControl, theirControl, r)
 		t.Log("Connection manager hasn't ticked yet")
 		time.Sleep(time.Second)
@@ -1223,7 +1291,7 @@ func TestRehandshakingLoser(t *testing.T) {
 	myConfig.ReloadConfigString(string(rc))
 
 	r.Log("Spin until there is only 1 tunnel")
-	for len(myControl.GetHostmap().Indexes)+len(theirControl.GetHostmap().Indexes) > 2 {
+	for myControl.GetHostmapIndexCount()+theirControl.GetHostmapIndexCount() > 2 {
 		assertTunnel(t, myVpnIpNet[0].Addr(), theirVpnIpNet[0].Addr(), myControl, theirControl, r)
 		t.Log("Connection manager hasn't ticked yet")
 		time.Sleep(time.Second)
@@ -1535,3 +1603,78 @@ func TestGoodHandshakeUnsafeDest(t *testing.T) {
 	myControl.Stop()
 	theirControl.Stop()
 }
+
+func TestMultiVpnAddrDeletePrimaryKeepsSecondAddr(t *testing.T) {
+	t.Parallel()
+	// Regression for the hostmap multi-vpnAddr delete bug. A dual-stack (v4+v6) V2-cert peer that
+	// handshakes twice at once ends up with two hostinfos linked in the shared next/prev chain, with the
+	// primary owning both addresses. Deleting that primary (e.g. connection manager dropping it, a
+	// CloseTunnel, a collision) must promote the surviving sibling for EVERY address. The pre-fix code
+	// unlinked the chain once per address, so it promoted the sibling for the first address and orphaned
+	// the second: the peer stayed reachable at its v4 addr but not its v6 addr despite a live tunnel.
+
+	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,fd00::1/64", nil)
+	theirControl, theirVpnIpNet, theirUdpAddr, _ := newSimpleServer(cert.Version2, ca, caKey, "them", "10.128.0.2/24,fd00::2/64", nil)
+
+	// This bug only exists for peers carrying more than one vpn address
+	require.Len(t, theirVpnIpNet, 2)
+	theirV4 := theirVpnIpNet[0].Addr()
+	theirV6 := theirVpnIpNet[1].Addr()
+
+	// Put their info in our lighthouse and vice versa
+	myControl.InjectLightHouseAddr(theirV4, theirUdpAddr)
+	theirControl.InjectLightHouseAddr(myVpnIpNet[0].Addr(), myUdpAddr)
+
+	// Build a router so we don't have to reason who gets which packet
+	r := router.NewR(t, myControl, theirControl)
+	defer r.RenderFlow()
+
+	myControl.Start()
+	theirControl.Start()
+
+	// Race a handshake so both of us build a hostinfo for the other, leaving my hostmap with a single
+	// host (them) backed by two linked hostinfos, just like TestStage1Race.
+	myControl.InjectTunPacket(BuildTunUDPPacket(theirV4, 80, myVpnIpNet[0].Addr(), 80, []byte("Hi from me")))
+	theirControl.InjectTunPacket(BuildTunUDPPacket(myVpnIpNet[0].Addr(), 80, theirV4, 80, []byte("Hi from them")))
+
+	myHsForThem := myControl.GetFromUDP(true)
+	theirHsForMe := theirControl.GetFromUDP(true)
+
+	r.InjectUDPPacket(theirControl, myControl, theirHsForMe)
+	r.InjectUDPPacket(myControl, theirControl, myHsForThem)
+
+	r.RouteForAllUntilTxTun(theirControl)
+	r.RouteForAllUntilTxTun(myControl)
+
+	r.RenderHostmaps("Racing hostmaps", myControl, theirControl)
+
+	// Two hostinfos for them means the shared next/prev chain has a sibling to promote. The Hosts map has
+	// one entry per vpn address (two, for dual stack), so the index count is what tells us there are two
+	// hostinfos.
+	require.Len(t, myControl.ListHostmapIndexes(false), 2)
+
+	// The primary owns both of their addresses
+	primaryV4 := myControl.GetHostInfoByVpnAddr(theirV4, false)
+	primaryV6 := myControl.GetHostInfoByVpnAddr(theirV6, false)
+	require.NotNil(t, primaryV4)
+	require.NotNil(t, primaryV6)
+	require.Equal(t, primaryV4.LocalIndex, primaryV6.LocalIndex, "both addrs should point at the same primary")
+
+	// Delete the primary tunnel. localOnly so we don't perturb their side, we only care about my hostmap.
+	require.True(t, myControl.CloseTunnel(theirV4, true))
+
+	// The surviving sibling must still serve BOTH addresses.
+	survivorV4 := myControl.GetHostInfoByVpnAddr(theirV4, false)
+	survivorV6 := myControl.GetHostInfoByVpnAddr(theirV6, false)
+	require.NotNil(t, survivorV4, "v4 addr should still resolve to the surviving tunnel")
+	// Pre-fix this is nil: the second address was orphaned when the primary was deleted.
+	require.NotNil(t, survivorV6, "v6 addr was orphaned after deleting the primary (multi-vpnAddr delete bug)")
+	assert.Equal(t, survivorV4.LocalIndex, survivorV6.LocalIndex, "both addrs should promote to the same survivor")
+	assert.NotEqual(t, primaryV4.LocalIndex, survivorV4.LocalIndex, "a different hostinfo should now be primary")
+
+	r.RenderHostmaps("Final hostmaps", myControl, theirControl)
+
+	myControl.Stop()
+	theirControl.Stop()
+}

e2e/tunnels_test.go

@@ -43,8 +43,8 @@ func TestDropInactiveTunnels(t *testing.T) {
 	r.Log("Go inactive and wait for the tunnels to get dropped")
 	waitStart := time.Now()
 	for {
-		myIndexes := len(myControl.GetHostmap().Indexes)
+		myIndexes := myControl.GetHostmapIndexCount()
-		theirIndexes := len(theirControl.GetHostmap().Indexes)
+		theirIndexes := theirControl.GetHostmapIndexCount()
 		if myIndexes == 0 && theirIndexes == 0 {
 			break
 		}
@@ -493,8 +493,8 @@ func TestCloseTunnelAuthenticated(t *testing.T) {
 
 	waitStart := time.Now()
 	for {
-		myIndexes := len(myControl.GetHostmap().Indexes)
+		myIndexes := myControl.GetHostmapIndexCount()
-		theirIndexes := len(theirControl.GetHostmap().Indexes)
+		theirIndexes := theirControl.GetHostmapIndexCount()
 		if myIndexes == 0 && theirIndexes == 0 {
 			break
 		}
@@ -548,8 +548,8 @@ func TestCloseTunnelAuthenticated(t *testing.T) {
 	r.Log("Injected bogus close tunnel. Let's see!")
 	waitStart = time.Now()
 	for {
-		myIndexes := len(myControl.GetHostmap().Indexes)
+		myIndexes := myControl.GetHostmapIndexCount()
-		theirIndexes := len(theirControl.GetHostmap().Indexes)
+		theirIndexes := theirControl.GetHostmapIndexCount()
 		if myIndexes == 0 {
 			t.Fatal("myIndexes should not be 0")
 		}

examples/config.yml

@@ -390,6 +390,20 @@ logging:
   # This setting is reloadable
   #inactivity_timeout: 10m
 
+  # clear_on_wake controls whether all tunnels are immediately torn down (locally, without notifying the remotes)
+  # when the machine detects it has just woken from system sleep. Tunnels rarely survive a suspend: NAT mappings
+  # expire and the machine's address usually changes, so waiting for the normal liveness checks costs 15-20 seconds
+  # of black-holed traffic per tunnel after wake. Clearing them means the first packet after wake starts a fresh
+  # handshake right away.
+  # This setting is reloadable
+  #clear_on_wake: true
+
+  # wake_clear_threshold is the minimum time the machine must have been suspended for clear_on_wake to act.
+  # Suspends shorter than this often come back before NAT state expires, so those tunnels may still be alive and
+  # are left to the normal liveness checks. Values below 1s are treated as 1s.
+  # This setting is reloadable
+  #wake_clear_threshold: 30s
+
 # Nebula security group configuration
 firewall:
   # Action to take when a packet is not allowed by the firewall rules.

examples/service_scripts/nebula.service

@@ -8,6 +8,15 @@ Before=sshd.service
 Type=notify
 NotifyAccess=main
 SyslogIdentifier=nebula
+
+# Uncomment to run as an unprivileged user with only CAP_NET_ADMIN. Requires a
+# nebula user that owns the config directory. Add CAP_NET_BIND_SERVICE to both
+# lines if any listener (lighthouse DNS, listen.port, stats, sshd) binds <1024.
+#User=nebula
+#Group=nebula
+#CapabilityBoundingSet=CAP_NET_ADMIN
+#AmbientCapabilities=CAP_NET_ADMIN
+
 ExecReload=/bin/kill -HUP $MAINPID
 ExecStart=/usr/local/bin/nebula -config /etc/nebula/config.yml
 Restart=always

firewall.go

@@ -44,8 +44,8 @@ type Firewall struct {
 	InRules  *FirewallTable
 	OutRules *FirewallTable
 
-	InSendReject  bool
+	InboundSendReject  bool
-	OutSendReject bool
+	OutboundSendReject bool
 
 	//TODO: we should have many more options for TCP, an option for ICMP, and mimic the kernel a bit better
 	// https://www.kernel.org/doc/Documentation/networking/nf_conntrack-sysctl.txt
@@ -216,23 +216,23 @@ func NewFirewallFromConfig(l *slog.Logger, cs *CertState, c *config.C) (*Firewal
 	inboundAction := c.GetString("firewall.inbound_action", "drop")
 	switch inboundAction {
 	case "reject":
-		fw.InSendReject = true
+		fw.InboundSendReject = true
 	case "drop":
-		fw.InSendReject = false
+		fw.InboundSendReject = false
 	default:
 		l.Warn("invalid firewall.inbound_action, defaulting to `drop`", "action", inboundAction)
-		fw.InSendReject = false
+		fw.InboundSendReject = false
 	}
 
 	outboundAction := c.GetString("firewall.outbound_action", "drop")
 	switch outboundAction {
 	case "reject":
-		fw.OutSendReject = true
+		fw.OutboundSendReject = true
 	case "drop":
-		fw.OutSendReject = false
+		fw.OutboundSendReject = false
 	default:
 		l.Warn("invalid firewall.outbound_action, defaulting to `drop`", "action", outboundAction)
-		fw.OutSendReject = false
+		fw.OutboundSendReject = false
 	}
 
 	err := AddFirewallRulesFromConfig(l, false, c, fw)
@@ -423,11 +423,6 @@ var ErrNoMatchingRule = errors.New("no matching rule in firewall table")
 // Drop returns an error if the packet should be dropped, explaining why. It
 // returns nil if the packet should not be dropped.
 func (f *Firewall) Drop(fp firewall.Packet, incoming bool, h *HostInfo, caPool *cert.CAPool, localCache firewall.ConntrackCache) error {
-	// Check if we spoke to this tuple, if we did then allow this packet
-	if f.inConns(fp, h, caPool, localCache) {
-		return nil
-	}
-
 	// Make sure remote address matches nebula certificate, and determine how to treat it
 	if h.networks == nil {
 		// Simple case: Certificate has one address and no unsafe networks
@@ -461,6 +456,11 @@ func (f *Firewall) Drop(fp firewall.Packet, incoming bool, h *HostInfo, caPool *
 		return ErrInvalidLocalIP
 	}
 
+	// Check if we spoke to this tuple, if we did then allow this packet
+	if f.inConns(fp, h, caPool, localCache) {
+		return nil
+	}
+
 	table := f.OutRules
 	if incoming {
 		table = f.InRules

firewall_test.go

@@ -916,6 +916,159 @@ func TestFirewall_DropIPSpoofing(t *testing.T) {
 	assert.Equal(t, fw.Drop(p, true, &h1, cp, nil), ErrInvalidRemoteIP)
 }
 
+func TestFirewall_ConntrackSourceSpoofingAcrossPeers(t *testing.T) {
+	l := test.NewLoggerWithOutput(&bytes.Buffer{})
+
+	myVpnNetworksTable := new(bart.Lite)
+	myVpnNetworksTable.Insert(netip.MustParsePrefix("192.0.2.1/24"))
+
+	owner := &dummyCert{
+		name:     "owner",
+		networks: []netip.Prefix{netip.MustParsePrefix("192.0.2.1/24")},
+	}
+
+	victim := &cert.CachedCertificate{
+		Certificate: &dummyCert{
+			name:     "victim",
+			networks: []netip.Prefix{netip.MustParsePrefix("192.0.2.2/24")},
+		},
+	}
+	victimHI := HostInfo{
+		ConnectionState: &ConnectionState{peerCert: victim},
+		vpnAddrs:        []netip.Addr{netip.MustParseAddr("192.0.2.2")},
+	}
+	victimHI.buildNetworks(myVpnNetworksTable, victim.Certificate)
+
+	attacker := &cert.CachedCertificate{
+		Certificate: &dummyCert{
+			name:     "attacker",
+			networks: []netip.Prefix{netip.MustParsePrefix("192.0.2.3/24")},
+		},
+	}
+	attackerHI := HostInfo{
+		ConnectionState: &ConnectionState{peerCert: attacker},
+		vpnAddrs:        []netip.Addr{netip.MustParseAddr("192.0.2.3")},
+	}
+	attackerHI.buildNetworks(myVpnNetworksTable, attacker.Certificate)
+
+	fw := NewFirewall(l, time.Second, time.Minute, time.Hour, owner)
+	// Allow any inbound traffic that passes the cert / source-IP checks.
+	require.NoError(t, fw.AddRule(true, firewall.ProtoAny, 0, 0, []string{"any"}, "", "", "", "", ""))
+	cp := cert.NewCAPool()
+
+	flow := firewall.Packet{
+		LocalAddr:  netip.MustParseAddr("192.0.2.1"),
+		RemoteAddr: netip.MustParseAddr("192.0.2.2"),
+		LocalPort:  443,
+		RemotePort: 55000,
+		Protocol:   firewall.ProtoUDP,
+	}
+
+	require.NoError(t, fw.Drop(flow, true, &victimHI, cp, nil),
+		"victim's own traffic from its own overlay IP must be allowed")
+
+	unseen := flow
+	unseen.RemotePort = 55001
+	assert.Equal(t, ErrInvalidRemoteIP, fw.Drop(unseen, true, &attackerHI, cp, nil),
+		"sanity: attacker forging victim's source IP must be rejected when no conntrack entry exists")
+
+	got := fw.Drop(flow, true, &attackerHI, cp, nil)
+	t.Logf("attacker replaying victim's 4-tuple: Drop returned %v (nil == packet ALLOWED == spoof succeeded)", got)
+	assert.Equal(t, ErrInvalidRemoteIP, got,
+		"SECURITY: attacker spoofed victim's overlay source IP (192.0.2.2) by reusing an existing conntrack 4-tuple; Drop returned %v instead of rejecting", got)
+}
+
+// BenchmarkFirewallDropConntrackHit measures Drop on an already-established flow
+// (a conntrack hit). This is the fast path that the source-IP<->cert binding
+// reordering adds work to, so it quantifies the cost of moving the address checks
+// ahead of the conntrack lookup. Cases:
+//   - simple:  peer cert has one address, no unsafe networks (h.networks == nil),
+//     so the remote-address check is a single netip.Addr compare.
+//   - complex: peer cert has unsafe networks (h.networks populated), so the
+//     remote-address check is a BART lookup.
+//   - noCache/localCache: whether a per-batch ConntrackCache is supplied, which in
+//     the original code let the fast path skip straight past the address checks.
+func BenchmarkFirewallDropConntrackHit(b *testing.B) {
+	l := test.NewLoggerWithOutput(&bytes.Buffer{})
+
+	myVpnNetworksTable := new(bart.Lite)
+	myVpnNetworksTable.Insert(netip.MustParsePrefix("192.0.2.1/24"))
+
+	owner := &dummyCert{
+		name:     "owner",
+		networks: []netip.Prefix{netip.MustParsePrefix("192.0.2.1/24")},
+	}
+
+	simpleCert := &cert.CachedCertificate{
+		Certificate: &dummyCert{
+			name:     "simple",
+			networks: []netip.Prefix{netip.MustParsePrefix("192.0.2.2/24")},
+		},
+	}
+	simpleHost := &HostInfo{
+		ConnectionState: &ConnectionState{peerCert: simpleCert},
+		vpnAddrs:        []netip.Addr{netip.MustParseAddr("192.0.2.2")},
+	}
+	simpleHost.buildNetworks(myVpnNetworksTable, simpleCert.Certificate)
+
+	complexCert := &cert.CachedCertificate{
+		Certificate: &dummyCert{
+			name:           "complex",
+			networks:       []netip.Prefix{netip.MustParsePrefix("192.0.2.2/24")},
+			unsafeNetworks: []netip.Prefix{netip.MustParsePrefix("198.51.100.0/24")},
+		},
+	}
+	complexHost := &HostInfo{
+		ConnectionState: &ConnectionState{peerCert: complexCert},
+		vpnAddrs:        []netip.Addr{netip.MustParseAddr("192.0.2.2")},
+	}
+	complexHost.buildNetworks(myVpnNetworksTable, complexCert.Certificate)
+
+	cp := cert.NewCAPool()
+
+	flow := firewall.Packet{
+		LocalAddr:  netip.MustParseAddr("192.0.2.1"),
+		RemoteAddr: netip.MustParseAddr("192.0.2.2"),
+		LocalPort:  443,
+		RemotePort: 55000,
+		Protocol:   firewall.ProtoUDP,
+	}
+
+	cases := []struct {
+		name     string
+		host     *HostInfo
+		useCache bool
+	}{
+		{"simple/noCache", simpleHost, false},
+		{"simple/localCache", simpleHost, true},
+		{"complex/noCache", complexHost, false},
+		{"complex/localCache", complexHost, true},
+	}
+
+	for _, tc := range cases {
+		b.Run(tc.name, func(b *testing.B) {
+			fw := NewFirewall(l, time.Second, time.Minute, time.Hour, owner)
+			require.NoError(b, fw.AddRule(true, firewall.ProtoAny, 0, 0, []string{"any"}, "", "", "", "", ""))
+
+			// Establish the conntrack entry so every benchmarked Drop is a hit.
+			require.NoError(b, fw.Drop(flow, true, tc.host, cp, nil))
+
+			var cache firewall.ConntrackCache
+			if tc.useCache {
+				cache = firewall.ConntrackCache{}
+			}
+
+			b.ReportAllocs()
+			b.ResetTimer()
+			for i := 0; i < b.N; i++ {
+				if err := fw.Drop(flow, true, tc.host, cp, cache); err != nil {
+					b.Fatal(err)
+				}
+			}
+		})
+	}
+}
+
 func BenchmarkLookup(b *testing.B) {
 	ml := func(m map[string]struct{}, a [][]string) {
 		for n := 0; n < b.N; n++ {

go.mod

@@ -12,7 +12,7 @@ require (
 	github.com/gaissmai/bart v0.28.0
 	github.com/gogo/protobuf v1.3.2
 	github.com/google/gopacket v1.1.19
-	github.com/kardianos/service v1.2.4
+	github.com/kardianos/service v1.3.0
 	github.com/miekg/dns v1.1.72
 	github.com/miekg/pkcs11 v1.1.2
 	github.com/nbrownus/go-metrics-prometheus v0.0.0-20210712211119-974a6260965f
@@ -24,15 +24,15 @@ require (
 	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 v0.53.0
+	golang.org/x/crypto v0.54.0
 	golang.org/x/exp v0.0.0-20230725093048-515e97ebf090
-	golang.org/x/net v0.56.0
+	golang.org/x/net v0.57.0
-	golang.org/x/sync v0.21.0
+	golang.org/x/sync v0.22.0
-	golang.org/x/sys v0.46.0
+	golang.org/x/sys v0.47.0
-	golang.org/x/term v0.44.0
+	golang.org/x/term v0.45.0
 	golang.zx2c4.com/wintun v0.0.0-20230126152724-0fa3db229ce2
 	golang.zx2c4.com/wireguard v0.0.0-20230325221338-052af4a8072b
-	golang.zx2c4.com/wireguard/windows v0.6.1
+	golang.zx2c4.com/wireguard/windows v1.0.1
 	google.golang.org/protobuf v1.36.11
 	gopkg.in/yaml.v3 v3.0.1
 	gvisor.dev/gvisor v0.0.0-20240423190808-9d7a357edefe
@@ -50,7 +50,7 @@ require (
 	github.com/prometheus/procfs v0.16.1 // indirect
 	github.com/vishvananda/netns v0.0.5 // indirect
 	go.yaml.in/yaml/v2 v2.4.2 // indirect
-	golang.org/x/mod v0.34.0 // indirect
+	golang.org/x/mod v0.36.0 // indirect
 	golang.org/x/time v0.5.0 // indirect
-	golang.org/x/tools v0.43.0 // indirect
+	golang.org/x/tools v0.45.0 // indirect
 )

go.sum

@@ -66,8 +66,8 @@ github.com/json-iterator/go v1.1.10/go.mod h1:KdQUCv79m/52Kvf8AW2vK1V8akMuk1QjK/
 github.com/json-iterator/go v1.1.11/go.mod h1:KdQUCv79m/52Kvf8AW2vK1V8akMuk1QjK/uOdHXbAo4=
 github.com/julienschmidt/httprouter v1.2.0/go.mod h1:SYymIcj16QtmaHHD7aYtjjsJG7VTCxuUUipMqKk8s4w=
 github.com/julienschmidt/httprouter v1.3.0/go.mod h1:JR6WtHb+2LUe8TCKY3cZOxFyyO8IZAc4RVcycCCAKdM=
-github.com/kardianos/service v1.2.4 h1:XNlGtZOYNx2u91urOdg/Kfmc+gfmuIo1Dd3rEi2OgBk=
+github.com/kardianos/service v1.3.0 h1:/LGy+xPP2TM+GLTiCZ2di7cy0Jd/qrawlTUfqKYFdTI=
-github.com/kardianos/service v1.2.4/go.mod h1:E4V9ufUuY82F7Ztlu1eN9VXWIQxg8NoLQlmFe0MtrXc=
+github.com/kardianos/service v1.3.0/go.mod h1:E4V9ufUuY82F7Ztlu1eN9VXWIQxg8NoLQlmFe0MtrXc=
 github.com/kisielk/errcheck v1.5.0/go.mod h1:pFxgyoBC7bSaBwPgfKdkLd5X25qrDl4LWUI2bnpBCr8=
 github.com/kisielk/gotool v1.0.0/go.mod h1:XhKaO+MFFWcvkIS/tQcRk01m1F5IRFswLeQ+oQHNcck=
 github.com/klauspost/compress v1.18.0 h1:c/Cqfb0r+Yi+JtIEq73FWXVkRonBlf0CRNYc8Zttxdo=
@@ -162,16 +162,16 @@ golang.org/x/crypto v0.0.0-20190308221718-c2843e01d9a2/go.mod h1:djNgcEr1/C05ACk
 golang.org/x/crypto v0.0.0-20191011191535-87dc89f01550/go.mod h1:yigFU9vqHzYiE8UmvKecakEJjdnWj3jj499lnFckfCI=
 golang.org/x/crypto v0.0.0-20200622213623-75b288015ac9/go.mod h1:LzIPMQfyMNhhGPhUkYOs5KpL4U8rLKemX1yGLhDgUto=
 golang.org/x/crypto v0.0.0-20210322153248-0c34fe9e7dc2/go.mod h1:T9bdIzuCu7OtxOm1hfPfRQxPLYneinmdGuTeoZ9dtd4=
-golang.org/x/crypto v0.53.0 h1:QZ4Muo8THX6CizN2vPPd5fBGHyogrdK9fG4wLPFUsto=
+golang.org/x/crypto v0.54.0 h1:YLIA59K4fiNzHzjnZt2tUJQjQtUWfWbeHBqKtk3eScw=
-golang.org/x/crypto v0.53.0/go.mod h1:DNLU434OwVakk9PzuwV8w62mAJpRJL3vsgcfp4Qnsio=
+golang.org/x/crypto v0.54.0/go.mod h1:KWL8ny2AZdGR2cWmzeHrp2azQPGogOv+HeQaVEXC2dk=
 golang.org/x/exp v0.0.0-20230725093048-515e97ebf090 h1:Di6/M8l0O2lCLc6VVRWhgCiApHV8MnQurBnFSHsQtNY=
 golang.org/x/exp v0.0.0-20230725093048-515e97ebf090/go.mod h1:FXUEEKJgO7OQYeo8N01OfiKP8RXMtf6e8aTskBGqWdc=
 golang.org/x/lint v0.0.0-20200302205851-738671d3881b/go.mod h1:3xt1FjdF8hUf6vQPIChWIBhFzV8gjjsPE/fR3IyQdNY=
 golang.org/x/mod v0.1.1-0.20191105210325-c90efee705ee/go.mod h1:QqPTAvyqsEbceGzBzNggFXnrqF1CaUcvgkdR5Ot7KZg=
 golang.org/x/mod v0.2.0/go.mod h1:s0Qsj1ACt9ePp/hMypM3fl4fZqREWJwdYDEqhRiZZUA=
 golang.org/x/mod v0.3.0/go.mod h1:s0Qsj1ACt9ePp/hMypM3fl4fZqREWJwdYDEqhRiZZUA=
-golang.org/x/mod v0.34.0 h1:xIHgNUUnW6sYkcM5Jleh05DvLOtwc6RitGHbDk4akRI=
+golang.org/x/mod v0.36.0 h1:JJjpVx6myfUsUdAzZuOSTTmRE0PfZeNWzzvKrP7amb4=
-golang.org/x/mod v0.34.0/go.mod h1:ykgH52iCZe79kzLLMhyCUzhMci+nQj+0XkbXpNYtVjY=
+golang.org/x/mod v0.36.0/go.mod h1:moc6ELqsWcOw5Ef3xVprK5ul/MvtVvkIXLziUOICjUQ=
 golang.org/x/net v0.0.0-20180724234803-3673e40ba225/go.mod h1:mL1N/T3taQHkDXs73rZJwtUhF3w3ftmwwsq0BUmARs4=
 golang.org/x/net v0.0.0-20181114220301-adae6a3d119a/go.mod h1:mL1N/T3taQHkDXs73rZJwtUhF3w3ftmwwsq0BUmARs4=
 golang.org/x/net v0.0.0-20190108225652-1e06a53dbb7e/go.mod h1:mL1N/T3taQHkDXs73rZJwtUhF3w3ftmwwsq0BUmARs4=
@@ -182,8 +182,8 @@ golang.org/x/net v0.0.0-20200226121028-0de0cce0169b/go.mod h1:z5CRVTTTmAJ677TzLL
 golang.org/x/net v0.0.0-20200625001655-4c5254603344/go.mod h1:/O7V0waA8r7cgGh81Ro3o1hOxt32SMVPicZroKQ2sZA=
 golang.org/x/net v0.0.0-20201021035429-f5854403a974/go.mod h1:sp8m0HH+o8qH0wwXwYZr8TS3Oi6o0r6Gce1SSxlDquU=
 golang.org/x/net v0.0.0-20210226172049-e18ecbb05110/go.mod h1:m0MpNAwzfU5UDzcl9v0D8zg8gWTRqZa9RBIspLL5mdg=
-golang.org/x/net v0.56.0 h1:Rw8j/hFzGvJUZwNBXnAtf5sVDVt+65SK2C7IxCxZt5o=
+golang.org/x/net v0.57.0 h1:K5+3DljvIuDG9/Jv9rvyMywYNFCQ9RSUY6OOTTkT+tE=
-golang.org/x/net v0.56.0/go.mod h1:D3Ku6r+V6JROoZK144D2XfMHFcMq/0zSfLelVTCFKec=
+golang.org/x/net v0.57.0/go.mod h1:KpXc8iv+r3XplLAG/f7Jsf9RPszJzdR0f58q9vGOuEU=
 golang.org/x/oauth2 v0.0.0-20190226205417-e64efc72b421/go.mod h1:gOpvHmFTYa4IltrdGE7lF6nIHvwfUNPOp7c8zoXwtLw=
 golang.org/x/sync v0.0.0-20181108010431-42b317875d0f/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sync v0.0.0-20181221193216-37e7f081c4d4/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
@@ -191,8 +191,8 @@ golang.org/x/sync v0.0.0-20190423024810-112230192c58/go.mod h1:RxMgew5VJxzue5/jJ
 golang.org/x/sync v0.0.0-20190911185100-cd5d95a43a6e/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sync v0.0.0-20201020160332-67f06af15bc9/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sync v0.0.0-20201207232520-09787c993a3a/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
-golang.org/x/sync v0.21.0 h1:HLII4xRRTtCRkxYp4HNFF0Js/Og6q2i++KXbg0gHCwM=
+golang.org/x/sync v0.22.0 h1:SZjpbeLmrCk4xhRSZFNZW5gFUeCeFgjekvI/+gfScek=
-golang.org/x/sync v0.21.0/go.mod h1:9xrNwdLfx4jkKbNva9FpL6vEN7evnE43NNNJQ2LF3+0=
+golang.org/x/sync v0.22.0/go.mod h1:9xrNwdLfx4jkKbNva9FpL6vEN7evnE43NNNJQ2LF3+0=
 golang.org/x/sys v0.0.0-20180905080454-ebe1bf3edb33/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY=
 golang.org/x/sys v0.0.0-20181116152217-5ac8a444bdc5/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY=
 golang.org/x/sys v0.0.0-20190215142949-d0b11bdaac8a/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY=
@@ -208,11 +208,11 @@ golang.org/x/sys v0.0.0-20210124154548-22da62e12c0c/go.mod h1:h1NjWce9XRLGQEsW7w
 golang.org/x/sys v0.0.0-20210603081109-ebe580a85c40/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.2.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.10.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
-golang.org/x/sys v0.46.0 h1:noSf2Fq6F8DBgS+LysIkx7rIExoNHJsxOAtPp4rthXw=
+golang.org/x/sys v0.47.0 h1:o7XGOvZQCADBQQ4Y7VNq2dRWQR7JmOUW8Kxx4ZsNgWs=
-golang.org/x/sys v0.46.0/go.mod h1:4GL1E5IUh+htKOUEOaiffhrAeqysfVGipDYzABqnCmw=
+golang.org/x/sys v0.47.0/go.mod h1:4GL1E5IUh+htKOUEOaiffhrAeqysfVGipDYzABqnCmw=
 golang.org/x/term v0.0.0-20201126162022-7de9c90e9dd1/go.mod h1:bj7SfCRtBDWHUb9snDiAeCFNEtKQo2Wmx5Cou7ajbmo=
-golang.org/x/term v0.44.0 h1:0rLvDRCtNj0gZkyIXhCyOb2OAzEhLVqc4B+hrsBhrmc=
+golang.org/x/term v0.45.0 h1:NwWyBmoJCbfTHpxrWoZ9C6/VxOf7ic219I8xZZFdrf0=
-golang.org/x/term v0.44.0/go.mod h1:7ze4MdzUzLXpSAoFP1H0bOI9aXDqveSvatT5vKcFh2Y=
+golang.org/x/term v0.45.0/go.mod h1:9aqxs0blBcrm/n0L9QW0aRVD+ktan8ssZromtqJC43w=
 golang.org/x/text v0.3.0/go.mod h1:NqM8EUOU14njkJ3fqMW+pc6Ldnwhi/IjpwHt7yyuwOQ=
 golang.org/x/text v0.3.2/go.mod h1:bEr9sfX3Q8Zfm5fL9x+3itogRgK3+ptLWKqgva+5dAk=
 golang.org/x/text v0.3.3/go.mod h1:5Zoc/QRtKVWzQhOtBMvqHzDpF6irO9z98xDceosuGiQ=
@@ -223,8 +223,8 @@ golang.org/x/tools v0.0.0-20191119224855-298f0cb1881e/go.mod h1:b+2E5dAYhXwXZwtn
 golang.org/x/tools v0.0.0-20200130002326-2f3ba24bd6e7/go.mod h1:TB2adYChydJhpapKDTa4BR/hXlZSLoq2Wpct/0txZ28=
 golang.org/x/tools v0.0.0-20200619180055-7c47624df98f/go.mod h1:EkVYQZoAsY45+roYkvgYkIh4xh/qjgUK9TdY2XT94GE=
 golang.org/x/tools v0.0.0-20210106214847-113979e3529a/go.mod h1:emZCQorbCU4vsT4fOWvOPXz4eW1wZW4PmDk9uLelYpA=
-golang.org/x/tools v0.43.0 h1:12BdW9CeB3Z+J/I/wj34VMl8X+fEXBxVR90JeMX5E7s=
+golang.org/x/tools v0.45.0 h1:18qN3FAooORvApf5XjCXgsuayZOEtXf6JK18I3+ONa8=
-golang.org/x/tools v0.43.0/go.mod h1:uHkMso649BX2cZK6+RpuIPXS3ho2hZo4FVwfoy1vIk0=
+golang.org/x/tools v0.45.0/go.mod h1:LuUGqqaXcXMEFEruIVJVm5mgDD8vww/z/SR1gQ4uE/0=
 golang.org/x/xerrors v0.0.0-20190717185122-a985d3407aa7/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=
 golang.org/x/xerrors v0.0.0-20191011141410-1b5146add898/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=
 golang.org/x/xerrors v0.0.0-20191204190536-9bdfabe68543/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=
@@ -233,8 +233,8 @@ golang.zx2c4.com/wintun v0.0.0-20230126152724-0fa3db229ce2 h1:B82qJJgjvYKsXS9jeu
 golang.zx2c4.com/wintun v0.0.0-20230126152724-0fa3db229ce2/go.mod h1:deeaetjYA+DHMHg+sMSMI58GrEteJUUzzw7en6TJQcI=
 golang.zx2c4.com/wireguard v0.0.0-20230325221338-052af4a8072b h1:J1CaxgLerRR5lgx3wnr6L04cJFbWoceSK9JWBdglINo=
 golang.zx2c4.com/wireguard v0.0.0-20230325221338-052af4a8072b/go.mod h1:tqur9LnfstdR9ep2LaJT4lFUl0EjlHtge+gAjmsHUG4=
-golang.zx2c4.com/wireguard/windows v0.6.1 h1:XMaKojH1Hs/raMrmnir4n35nTvzvWj7NmSYzHn2F4qU=
+golang.zx2c4.com/wireguard/windows v1.0.1 h1:eOxiDVbywPC+ZQqvdCK7x+ZwWXKbYv50TtH8ysFIbw8=
-golang.zx2c4.com/wireguard/windows v0.6.1/go.mod h1:04aqInu5GYuTFvMuDw/rKBAF7mHrltW/3rekpfbbZDM=
+golang.zx2c4.com/wireguard/windows v1.0.1/go.mod h1:+fbT3FFdX4zzYDLwJh5+HPEcNN/3HyNdzhNSVsQM+zs=
 google.golang.org/appengine v1.4.0/go.mod h1:xpcJRLb0r/rnEns0DIKYYv+WjYCduHsrkT7/EB5XEv4=
 google.golang.org/protobuf v0.0.0-20200109180630-ec00e32a8dfd/go.mod h1:DFci5gLYBciE7Vtevhsrf46CRTquxDuWsQurQQe4oz8=
 google.golang.org/protobuf v0.0.0-20200221191635-4d8936d0db64/go.mod h1:kwYJMbMJ01Woi6D6+Kah6886xMZcty6N08ah7+eCXa0=

handshake_manager.go

@@ -430,14 +430,11 @@ func (hm *HandshakeManager) CheckAndComplete(hostinfo *HostInfo, handshakePacket
 	// Check if we already have a tunnel with this vpn ip
 	existingHostInfo, found := hm.mainHostMap.Hosts[hostinfo.vpnAddrs[0]]
 	if found && existingHostInfo != nil {
-		testHostInfo := existingHostInfo
+		// Is it just a delayed handshake packet? Check every hostinfo we hold for this address.
-		for testHostInfo != nil {
+		for _, testHostInfo := range hm.mainHostMap.unlockedGetHostList(hostinfo.vpnAddrs[0]) {
-			// Is it just a delayed handshake packet?
 			if bytes.Equal(hostinfo.HandshakePacket[handshakePacket], testHostInfo.HandshakePacket[handshakePacket]) {
 				return testHostInfo, ErrAlreadySeen
 			}
-
-			testHostInfo = testHostInfo.next
 		}
 
 		// Is this a newer handshake?
@@ -532,7 +529,9 @@ func (hm *HandshakeManager) DeleteHostInfo(hostinfo *HostInfo) {
 
 func (hm *HandshakeManager) unlockedDeleteHostInfo(hostinfo *HostInfo) {
 	for _, addr := range hostinfo.vpnAddrs {
-		delete(hm.vpnIps, addr)
+		if cur, ok := hm.vpnIps[addr]; ok && cur.hostinfo == hostinfo {
+			delete(hm.vpnIps, addr)
+		}
 	}
 
 	if len(hm.vpnIps) == 0 {
@@ -1080,7 +1079,7 @@ func (hm *HandshakeManager) sendHandshakeResponse(via ViaSender, msg []byte, hos
 		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)
+		via.relayHI.relayState.UpdateRelayForByIdxState(via.relay.LocalIndex, 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])...)
 	}

hostmap.go

@@ -56,11 +56,20 @@ type Relay struct {
 }
 
 type HostMap struct {
-	sync.RWMutex    //Because we concurrently read and write to our maps
+	sync.RWMutex  //Because we concurrently read and write to our maps
-	Indexes         map[uint32]*HostInfo
+	Indexes       map[uint32]*HostInfo
-	Relays          map[uint32]*HostInfo // Maps a Relay IDX to a Relay HostInfo object
+	Relays        map[uint32]*HostInfo // Maps a Relay IDX to a Relay HostInfo object
-	RemoteIndexes   map[uint32]*HostInfo
+	RemoteIndexes map[uint32]*HostInfo
+	// Hosts maps a vpn address to its primary hostinfo, one entry per address we hold a tunnel
+	// for. moreHosts only has an entry while an address is held by 2 or more hostinfos and stores
+	// the full most-recent-first list; moreHosts[a][0] is always the same hostinfo as Hosts[a].
+	// Each address gets its own independent list, so a hostinfo owning multiple addresses can
+	// never corrupt another address's ordering the way the old shared next/prev chain could.
+	// Entries in moreHosts are only ever written by unlockedSetHostsForAddr; Hosts is written
+	// directly only in the single-hostinfo fast paths where moreHosts is known to have no entry,
+	// and unlockedDeleteHostInfo swaps either map for a fresh one when it fully drains.
 	Hosts           map[netip.Addr]*HostInfo
+	moreHosts       map[netip.Addr][]*HostInfo
 	preferredRanges atomic.Pointer[[]netip.Prefix]
 	l               *slog.Logger
 }
@@ -229,7 +238,7 @@ const (
 )
 
 type HostInfo struct {
-	remote          netip.AddrPort
+	remote          atomic.Pointer[netip.AddrPort]
 	remotes         *RemoteList
 	promoteCounter  atomic.Uint32
 	ConnectionState *ConnectionState
@@ -266,10 +275,6 @@ type HostInfo struct {
 	lastRoam       time.Time
 	lastRoamRemote netip.AddrPort
 
-	// Used to track other hostinfos for this vpn ip since only 1 can be primary
-	// Synchronised via hostmap lock and not the hostinfo lock.
-	next, prev *HostInfo
-
 	//TODO: in, out, and others might benefit from being an atomic.Int32. We could collapse connectionManager pendingDeletion, relayUsed, and in/out into this 1 thing
 	in, out, pendingDeletion atomic.Bool
 
@@ -282,7 +287,6 @@ type HostInfo struct {
 type ViaSender struct {
 	UdpAddr   netip.AddrPort
 	relayHI   *HostInfo // relayHI is the host info object of the relay
-	remoteIdx uint32    // remoteIdx is the index included in the header of the received packet
 	relay     *Relay    // relay contains the rest of the relay information, including the PeerIP of the host trying to communicate with us.
 	IsRelayed bool      // IsRelayed is true if the packet was sent through a relay
 }
@@ -334,6 +338,7 @@ func newHostMap(l *slog.Logger) *HostMap {
 		Relays:        map[uint32]*HostInfo{},
 		RemoteIndexes: map[uint32]*HostInfo{},
 		Hosts:         map[netip.Addr]*HostInfo{},
+		moreHosts:     map[netip.Addr][]*HostInfo{},
 		l:             l,
 	}
 }
@@ -382,13 +387,55 @@ func (hm *HostMap) EmitStats() {
 	metrics.GetOrRegisterGauge("hostmap.main.relayIndexes", nil).Update(int64(relaysLen))
 }
 
-// DeleteHostInfo will fully unlink the hostinfo and return true if it was the final hostinfo for this vpn ip
+// unlockedSetHostsForAddr stores the per-address hostinfo list (list[0] is the primary). An empty
+// list removes the address. This is the one place Hosts and moreHosts are written together, keep
+// it that way. Callers must hold the write lock.
+func (hm *HostMap) unlockedSetHostsForAddr(addr netip.Addr, list []*HostInfo) {
+	if len(list) == 0 {
+		delete(hm.Hosts, addr)
+		delete(hm.moreHosts, addr)
+		return
+	}
+	hm.Hosts[addr] = list[0]
+	if len(list) > 1 {
+		hm.moreHosts[addr] = list
+	} else {
+		delete(hm.moreHosts, addr)
+	}
+}
+
+// unlockedGetHostList returns every hostinfo holding addr, primary first, or nil if we have no
+// tunnel for addr. The common single-hostinfo case builds a fresh one element list, so keep this
+// off the packet hot path; the primary is a direct Hosts read. Callers must hold the lock (read
+// or write).
+func (hm *HostMap) unlockedGetHostList(addr netip.Addr) []*HostInfo {
+	if list, ok := hm.moreHosts[addr]; ok {
+		return list
+	}
+	if h, ok := hm.Hosts[addr]; ok {
+		return []*HostInfo{h}
+	}
+	return nil
+}
+
+// removeHostInfo returns list with hi removed (order preserved), or list unchanged if hi is
+// absent. It deletes in place: every mutator holds the hostmap write lock and no reader ever
+// retains a slice across a mutation (readers iterate under RLock), so there is no snapshot to
+// invalidate.
+func removeHostInfo(list []*HostInfo, hi *HostInfo) []*HostInfo {
+	idx := slices.Index(list, hi)
+	if idx < 0 {
+		return list
+	}
+	return slices.Delete(list, idx, idx+1)
+}
+
+// DeleteHostInfo will fully unlink the hostinfo and return true if no other hostinfo still holds
+// any of its vpn addrs, meaning we no longer have a tunnel to the peer
 func (hm *HostMap) DeleteHostInfo(hostinfo *HostInfo) bool {
 	// Delete the host itself, ensuring it's not modified anymore
 	hm.Lock()
-	// If we have a previous or next hostinfo then we are not the last one for this vpn ip
+	final := hm.unlockedDeleteHostInfo(hostinfo)
-	final := (hostinfo.next == nil && hostinfo.prev == nil)
-	hm.unlockedDeleteHostInfo(hostinfo)
 	hm.Unlock()
 
 	return final
@@ -400,85 +447,66 @@ func (hm *HostMap) MakePrimary(hostinfo *HostInfo) {
 	hm.unlockedMakePrimary(hostinfo)
 }
 
-func (hm *HostMap) unlockedMakePrimary(hostinfo *HostInfo) {
+// unlockedMakePrimary reports whether hostinfo is (now) the primary for each of its addresses,
-	// Get the current primary, if it exists
+// false only when it is no longer in the hostmap at all.
-	oldHostinfo := hm.Hosts[hostinfo.vpnAddrs[0]]
+func (hm *HostMap) unlockedMakePrimary(hostinfo *HostInfo) bool {
-
+	// A hostinfo that is no longer in the hostmap must not be re-inserted here. Callers can race
-	// Every address in the hostinfo gets elevated to primary
+	// tunnel teardown, deciding to promote under the read lock and only taking the write lock
-	for _, vpnAddr := range hostinfo.vpnAddrs {
+	// after a delete fully unlinked the hostinfo (connection manager swapPrimary, AddRelay). Every
-		//NOTE: It is possible that we leave a dangling hostinfo here but connection manager works on
+	// live hostinfo is registered in Indexes by unlockedAddHostInfo, so this is a membership test.
-		// indexes so it should be fine.
+	if hm.Indexes[hostinfo.localIndexId] != hostinfo {
-		hm.Hosts[vpnAddr] = hostinfo
+		return false
 	}
 
-	// If we are already primary then we won't bother re-linking
+	// Move hostinfo to the front (primary) of each of its address lists. The lists are
-	if oldHostinfo == hostinfo {
+	// independent per address, so this can never leave a dangling entry the way promoting
-		return
+	// against a single shared chain could.
-	}
-
-	// Unlink this hostinfo
-	if hostinfo.prev != nil {
-		hostinfo.prev.next = hostinfo.next
-	}
-	if hostinfo.next != nil {
-		hostinfo.next.prev = hostinfo.prev
-	}
-
-	// If there wasn't a previous primary then clear out any links
-	if oldHostinfo == nil {
-		hostinfo.next = nil
-		hostinfo.prev = nil
-		return
-	}
-
-	// Relink the hostinfo as primary
-	hostinfo.next = oldHostinfo
-	oldHostinfo.prev = hostinfo
-	hostinfo.prev = nil
-}
-
-func (hm *HostMap) unlockedDeleteHostInfo(hostinfo *HostInfo) {
 	for _, addr := range hostinfo.vpnAddrs {
-		h := hm.Hosts[addr]
+		if hm.Hosts[addr] == hostinfo {
-		for h != nil {
+			// Already primary for this address, the list is already in the right order
-			if h == hostinfo {
+			continue
-				hm.unlockedInnerDeleteHostInfo(h, addr)
-			}
-			h = h.next
 		}
+		list := removeHostInfo(hm.unlockedGetHostList(addr), hostinfo)
+		list = append([]*HostInfo{hostinfo}, list...)
+		hm.unlockedSetHostsForAddr(addr, list)
 	}
+	return true
 }
 
-func (hm *HostMap) unlockedInnerDeleteHostInfo(hostinfo *HostInfo, addr netip.Addr) {
+// unlockedDeleteHostInfo removes hostinfo from every one of its address lists and from the index
-	primary, ok := hm.Hosts[addr]
+// maps. It returns true if this was the last hostinfo for all of its addresses (we no longer have
-	isLastHostinfo := hostinfo.next == nil && hostinfo.prev == nil
+// any tunnel to the peer), which the caller uses to decide whether to clear learned lighthouse
-	if ok && primary == hostinfo {
+// state and disestablish relays.
-		// The vpn addr pointer points to the same hostinfo as the local index id, we can remove it
+func (hm *HostMap) unlockedDeleteHostInfo(hostinfo *HostInfo) bool {
-		delete(hm.Hosts, addr)
+	// Remove this hostinfo from each of its address lists. The lists are independent, so a
-		if len(hm.Hosts) == 0 {
+	// sibling is never promoted to an address it does not own and no other list is touched.
-			hm.Hosts = map[netip.Addr]*HostInfo{}
+	final := true
-		}
+	for _, addr := range hostinfo.vpnAddrs {
-
+		if list, ok := hm.moreHosts[addr]; ok {
-		if hostinfo.next != nil {
+			list = removeHostInfo(list, hostinfo)
-			// We had more than 1 hostinfo at this vpn addr, promote the next in the list to primary
+			hm.unlockedSetHostsForAddr(addr, list)
-			hm.Hosts[addr] = hostinfo.next
+			if len(list) > 0 {
-			// It is primary, there is no previous hostinfo now
+				final = false
-			hostinfo.next.prev = nil
+			}
-		}
+		} else if existing, ok := hm.Hosts[addr]; ok {
-
+			if existing == hostinfo {
-	} else {
+				// Common case, the only hostinfo for this address. moreHosts has no entry to clean up.
-		// Relink if we were in the middle of multiple hostinfos for this vpn addr
+				delete(hm.Hosts, addr)
-		if hostinfo.prev != nil {
+			} else {
-			hostinfo.prev.next = hostinfo.next
+				// We don't hold this address but another hostinfo does, we still have a tunnel to the peer
-		}
+				final = false
-
+			}
-		if hostinfo.next != nil {
-			hostinfo.next.prev = hostinfo.prev
 		}
 	}
 
-	hostinfo.next = nil
+	// Go maps never shrink their buckets, replace fully drained maps so a node that churned
-	hostinfo.prev = nil
+	// through a large peer count gives the memory back. Same idiom as the index maps below.
+	if len(hm.Hosts) == 0 {
+		hm.Hosts = map[netip.Addr]*HostInfo{}
+	}
+	if len(hm.moreHosts) == 0 {
+		hm.moreHosts = map[netip.Addr][]*HostInfo{}
+	}
 
 	// The remote index uses index ids outside our control so lets make sure we are only removing
 	// the remote index pointer here if it points to the hostinfo we are deleting
@@ -502,7 +530,7 @@ func (hm *HostMap) unlockedInnerDeleteHostInfo(hostinfo *HostInfo, addr netip.Ad
 		)
 	}
 
-	if isLastHostinfo {
+	if final {
 		// I have lost connectivity to my peers. My relay tunnel is likely broken. Mark the next
 		// hops as 'Requested' so that new relay tunnels are created in the future.
 		hm.unlockedDisestablishVpnAddrRelayFor(hostinfo)
@@ -511,6 +539,8 @@ func (hm *HostMap) unlockedInnerDeleteHostInfo(hostinfo *HostInfo, addr netip.Ad
 	for _, localRelayIdx := range hostinfo.relayState.CopyRelayForIdxs() {
 		delete(hm.Relays, localRelayIdx)
 	}
+
+	return final
 }
 
 func (hm *HostMap) QueryIndex(index uint32) *HostInfo {
@@ -554,19 +584,30 @@ func (hm *HostMap) QueryVpnAddrsRelayFor(targetIps []netip.Addr, relayHostIp net
 	hm.RLock()
 	defer hm.RUnlock()
 
+	// This runs per relayed packet, so check the primary with a single map probe and only consult
+	// moreHosts when the primary can't relay for us.
 	h, ok := hm.Hosts[relayHostIp]
 	if !ok {
 		return nil, nil, errors.New("unable to find host")
 	}
 
-	for h != nil {
+	for _, targetIp := range targetIps {
-		for _, targetIp := range targetIps {
+		r, ok := h.relayState.QueryRelayForByIp(targetIp)
-			r, ok := h.relayState.QueryRelayForByIp(targetIp)
+		if ok && r.State == Established {
-			if ok && r.State == Established {
+			return h, r, nil
-				return h, r, nil
+		}
+	}
+
+	if list, ok := hm.moreHosts[relayHostIp]; ok {
+		// list[0] is the primary we already checked
+		for _, h := range list[1:] {
+			for _, targetIp := range targetIps {
+				r, ok := h.relayState.QueryRelayForByIp(targetIp)
+				if ok && r.State == Established {
+					return h, r, nil
+				}
 			}
 		}
-		h = h.next
 	}
 
 	return nil, nil, errors.New("unable to find host with relay")
@@ -574,20 +615,14 @@ func (hm *HostMap) QueryVpnAddrsRelayFor(targetIps []netip.Addr, relayHostIp net
 
 func (hm *HostMap) unlockedDisestablishVpnAddrRelayFor(hi *HostInfo) {
 	for _, relayHostIp := range hi.relayState.CopyRelayIps() {
-		if h, ok := hm.Hosts[relayHostIp]; ok {
+		for _, h := range hm.unlockedGetHostList(relayHostIp) {
-			for h != nil {
+			h.relayState.UpdateRelayForByIpState(hi.vpnAddrs[0], Disestablished)
-				h.relayState.UpdateRelayForByIpState(hi.vpnAddrs[0], Disestablished)
-				h = h.next
-			}
 		}
 	}
 	for _, rs := range hi.relayState.CopyAllRelayFor() {
 		if rs.Type == ForwardingType {
-			if h, ok := hm.Hosts[rs.PeerAddr]; ok {
+			for _, h := range hm.unlockedGetHostList(rs.PeerAddr) {
-				for h != nil {
+				h.relayState.UpdateRelayForByIpState(hi.vpnAddrs[0], Disestablished)
-					h.relayState.UpdateRelayForByIpState(hi.vpnAddrs[0], Disestablished)
-					h = h.next
-				}
 			}
 		}
 	}
@@ -637,22 +672,27 @@ func (hm *HostMap) unlockedAddHostInfo(hostinfo *HostInfo, f *Interface) {
 }
 
 func (hm *HostMap) unlockedInnerAddHostInfo(vpnAddr netip.Addr, hostinfo *HostInfo, f *Interface) {
-	existing := hm.Hosts[vpnAddr]
+	existing, ok := hm.Hosts[vpnAddr]
-	hm.Hosts[vpnAddr] = hostinfo
+	if !ok {
-
+		// Common case, the first hostinfo for this address. moreHosts stays empty.
-	if existing != nil && existing != hostinfo {
+		hm.Hosts[vpnAddr] = hostinfo
-		hostinfo.next = existing
+		return
-		existing.prev = hostinfo
 	}
 
-	i := 1
+	// The new hostinfo becomes the primary for this address. Remove any stale copy of it first so
-	check := hostinfo
+	// we never hold a duplicate, then prepend.
-	for check != nil {
+	list, ok := hm.moreHosts[vpnAddr]
-		if i > MaxHostInfosPerVpnIp {
+	if !ok {
-			hm.unlockedDeleteHostInfo(check)
+		list = []*HostInfo{existing}
-		}
+	}
-		check = check.next
+	list = removeHostInfo(list, hostinfo)
-		i++
+	list = append([]*HostInfo{hostinfo}, list...)
+	hm.unlockedSetHostsForAddr(vpnAddr, list)
+
+	// Enforce the per-address cap by fully retiring the oldest hostinfo once we exceed it.
+	// Deleting it removes it from all of its addresses and the index maps, matching prior behavior.
+	if len(list) > MaxHostInfosPerVpnIp {
+		hm.unlockedDeleteHostInfo(list[len(list)-1])
 	}
 }
 
@@ -684,7 +724,7 @@ func (hm *HostMap) ForEachIndex(f controlEach) {
 func (i *HostInfo) TryPromoteBest(preferredRanges []netip.Prefix, ifce *Interface) {
 	c := i.promoteCounter.Add(1)
 	if c%ifce.tryPromoteEvery.Load() == 0 {
-		remote := i.remote
+		remote := i.GetRemote()
 
 		// return early if we are already on a preferred remote
 		if remote.IsValid() {
@@ -726,11 +766,18 @@ func (i *HostInfo) GetCert() *cert.CachedCertificate {
 	return nil
 }
 
+func (i *HostInfo) GetRemote() netip.AddrPort {
+	if p := i.remote.Load(); p != nil {
+		return *p
+	}
+	return netip.AddrPort{}
+}
+
 // TODO: Maybe use ViaSender here?
 func (i *HostInfo) SetRemote(remote netip.AddrPort) {
 	// We copy here because we likely got this remote from a source that reuses the object
-	if i.remote != remote {
+	if i.GetRemote() != remote {
-		i.remote = remote
+		i.remote.Store(&remote)
 		i.remotes.LearnRemote(i.vpnAddrs[0], remote)
 	}
 }
@@ -742,7 +789,7 @@ func (i *HostInfo) SetRemoteIfPreferred(hm *HostMap, via ViaSender) bool {
 		return false
 	}
 
-	currentRemote := i.remote
+	currentRemote := i.GetRemote()
 	if !currentRemote.IsValid() {
 		i.SetRemote(via.UdpAddr)
 		return true

hostmap_test.go

@@ -2,6 +2,7 @@ package nebula
 
 import (
 	"net/netip"
+	"slices"
 	"testing"
 
 	"github.com/slackhq/nebula/config"
@@ -10,78 +11,84 @@ import (
 	"github.com/stretchr/testify/require"
 )
 
+// chainIds returns the localIndexIds of the hostinfos holding addr, primary (index 0) first. It
+// also validates the Hosts/moreHosts sync contract on every call so a mutation that broke it
+// fails fast.
+func chainIds(t *testing.T, hm *HostMap, addr netip.Addr) []uint32 {
+	t.Helper()
+	assertHostMapInvariants(t, hm)
+	list := hm.unlockedGetHostList(addr)
+	ids := make([]uint32, len(list))
+	for i, h := range list {
+		ids[i] = h.localIndexId
+	}
+	return ids
+}
+
+// assertHostMapInvariants checks the Hosts/moreHosts contract: moreHosts only holds addresses
+// with 2 or more hostinfos, its first entry is always the primary in Hosts, lists never hold
+// duplicates, every hostinfo in a list owns the address and is registered in Indexes, and every
+// indexed hostinfo is reachable through each of its addresses.
+func assertHostMapInvariants(t *testing.T, hm *HostMap) {
+	t.Helper()
+	for addr, list := range hm.moreHosts {
+		require.GreaterOrEqualf(t, len(list), 2, "moreHosts[%s] must hold at least 2 hostinfos", addr)
+		require.Samef(t, hm.Hosts[addr], list[0], "moreHosts[%s][0] must match the primary in Hosts", addr)
+		seen := map[*HostInfo]bool{}
+		for _, h := range list {
+			require.NotNilf(t, h, "moreHosts[%s] must never hold a nil hostinfo", addr)
+			require.Falsef(t, seen[h], "moreHosts[%s] holds hostinfo %d twice", addr, h.localIndexId)
+			seen[h] = true
+			require.Samef(t, hm.Indexes[h.localIndexId], h, "moreHosts[%s] member %d is not registered in Indexes", addr, h.localIndexId)
+			require.Truef(t, slices.Contains(h.vpnAddrs, addr), "moreHosts[%s] member %d does not own the address", addr, h.localIndexId)
+		}
+	}
+	for addr, h := range hm.Hosts {
+		require.NotNilf(t, h, "Hosts[%s] must never be nil", addr)
+		require.Samef(t, hm.Indexes[h.localIndexId], h, "Hosts[%s] primary %d is not registered in Indexes", addr, h.localIndexId)
+		require.Truef(t, slices.Contains(h.vpnAddrs, addr), "Hosts[%s] primary (index %d) does not own the address", addr, h.localIndexId)
+	}
+	for idx, h := range hm.Indexes {
+		require.Equalf(t, idx, h.localIndexId, "Indexes[%d] holds hostinfo with localIndexId %d", idx, h.localIndexId)
+		for _, va := range h.vpnAddrs {
+			require.Truef(t, slices.Contains(hm.unlockedGetHostList(va), h), "indexed hostinfo %d is missing from the list for %s", idx, va)
+		}
+	}
+}
+
 func TestHostMap_MakePrimary(t *testing.T) {
 	l := test.NewLogger()
 	hm := newHostMap(l)
 
 	f := &Interface{}
+	a := netip.MustParseAddr("0.0.0.1")
 
-	h1 := &HostInfo{vpnAddrs: []netip.Addr{netip.MustParseAddr("0.0.0.1")}, localIndexId: 1}
+	h1 := &HostInfo{vpnAddrs: []netip.Addr{a}, localIndexId: 1}
-	h2 := &HostInfo{vpnAddrs: []netip.Addr{netip.MustParseAddr("0.0.0.1")}, localIndexId: 2}
+	h2 := &HostInfo{vpnAddrs: []netip.Addr{a}, localIndexId: 2}
-	h3 := &HostInfo{vpnAddrs: []netip.Addr{netip.MustParseAddr("0.0.0.1")}, localIndexId: 3}
+	h3 := &HostInfo{vpnAddrs: []netip.Addr{a}, localIndexId: 3}
-	h4 := &HostInfo{vpnAddrs: []netip.Addr{netip.MustParseAddr("0.0.0.1")}, localIndexId: 4}
+	h4 := &HostInfo{vpnAddrs: []netip.Addr{a}, localIndexId: 4}
 
 	hm.unlockedAddHostInfo(h4, f)
 	hm.unlockedAddHostInfo(h3, f)
 	hm.unlockedAddHostInfo(h2, f)
 	hm.unlockedAddHostInfo(h1, f)
 
-	// Make sure we go h1 -> h2 -> h3 -> h4
+	// Most-recently-added is primary: h1, h2, h3, h4
-	prim := hm.QueryVpnAddr(netip.MustParseAddr("0.0.0.1"))
+	assert.Equal(t, []uint32{1, 2, 3, 4}, chainIds(t, hm, a))
-	assert.Equal(t, h1.localIndexId, prim.localIndexId)
+	assert.Equal(t, h1, hm.QueryVpnAddr(a))
-	assert.Equal(t, h2.localIndexId, prim.next.localIndexId)
-	assert.Nil(t, prim.prev)
-	assert.Equal(t, h1.localIndexId, h2.prev.localIndexId)
-	assert.Equal(t, h3.localIndexId, h2.next.localIndexId)
-	assert.Equal(t, h2.localIndexId, h3.prev.localIndexId)
-	assert.Equal(t, h4.localIndexId, h3.next.localIndexId)
-	assert.Equal(t, h3.localIndexId, h4.prev.localIndexId)
-	assert.Nil(t, h4.next)
 
-	// Swap h3/middle to primary
+	// Swap the middle to primary: h3, h1, h2, h4
 	hm.MakePrimary(h3)
+	assert.Equal(t, []uint32{3, 1, 2, 4}, chainIds(t, hm, a))
+	assert.Equal(t, h3, hm.QueryVpnAddr(a))
 
-	// Make sure we go h3 -> h1 -> h2 -> h4
+	// Swap the tail to primary: h4, h3, h1, h2
-	prim = hm.QueryVpnAddr(netip.MustParseAddr("0.0.0.1"))
-	assert.Equal(t, h3.localIndexId, prim.localIndexId)
-	assert.Equal(t, h1.localIndexId, prim.next.localIndexId)
-	assert.Nil(t, prim.prev)
-	assert.Equal(t, h2.localIndexId, h1.next.localIndexId)
-	assert.Equal(t, h3.localIndexId, h1.prev.localIndexId)
-	assert.Equal(t, h4.localIndexId, h2.next.localIndexId)
-	assert.Equal(t, h1.localIndexId, h2.prev.localIndexId)
-	assert.Equal(t, h2.localIndexId, h4.prev.localIndexId)
-	assert.Nil(t, h4.next)
-
-	// Swap h4/tail to primary
 	hm.MakePrimary(h4)
+	assert.Equal(t, []uint32{4, 3, 1, 2}, chainIds(t, hm, a))
 
-	// Make sure we go h4 -> h3 -> h1 -> h2
+	// Swapping the current primary again is a no-op
-	prim = hm.QueryVpnAddr(netip.MustParseAddr("0.0.0.1"))
-	assert.Equal(t, h4.localIndexId, prim.localIndexId)
-	assert.Equal(t, h3.localIndexId, prim.next.localIndexId)
-	assert.Nil(t, prim.prev)
-	assert.Equal(t, h1.localIndexId, h3.next.localIndexId)
-	assert.Equal(t, h4.localIndexId, h3.prev.localIndexId)
-	assert.Equal(t, h2.localIndexId, h1.next.localIndexId)
-	assert.Equal(t, h3.localIndexId, h1.prev.localIndexId)
-	assert.Equal(t, h1.localIndexId, h2.prev.localIndexId)
-	assert.Nil(t, h2.next)
-
-	// Swap h4 again should be no-op
 	hm.MakePrimary(h4)
-
+	assert.Equal(t, []uint32{4, 3, 1, 2}, chainIds(t, hm, a))
-	// Make sure we go h4 -> h3 -> h1 -> h2
-	prim = hm.QueryVpnAddr(netip.MustParseAddr("0.0.0.1"))
-	assert.Equal(t, h4.localIndexId, prim.localIndexId)
-	assert.Equal(t, h3.localIndexId, prim.next.localIndexId)
-	assert.Nil(t, prim.prev)
-	assert.Equal(t, h1.localIndexId, h3.next.localIndexId)
-	assert.Equal(t, h4.localIndexId, h3.prev.localIndexId)
-	assert.Equal(t, h2.localIndexId, h1.next.localIndexId)
-	assert.Equal(t, h3.localIndexId, h1.prev.localIndexId)
-	assert.Equal(t, h1.localIndexId, h2.prev.localIndexId)
-	assert.Nil(t, h2.next)
 }
 
 func TestHostMap_DeleteHostInfo(t *testing.T) {
@@ -89,13 +96,14 @@ func TestHostMap_DeleteHostInfo(t *testing.T) {
 	hm := newHostMap(l)
 
 	f := &Interface{}
+	a := netip.MustParseAddr("0.0.0.1")
 
-	h1 := &HostInfo{vpnAddrs: []netip.Addr{netip.MustParseAddr("0.0.0.1")}, localIndexId: 1}
+	h1 := &HostInfo{vpnAddrs: []netip.Addr{a}, localIndexId: 1}
-	h2 := &HostInfo{vpnAddrs: []netip.Addr{netip.MustParseAddr("0.0.0.1")}, localIndexId: 2}
+	h2 := &HostInfo{vpnAddrs: []netip.Addr{a}, localIndexId: 2}
-	h3 := &HostInfo{vpnAddrs: []netip.Addr{netip.MustParseAddr("0.0.0.1")}, localIndexId: 3}
+	h3 := &HostInfo{vpnAddrs: []netip.Addr{a}, localIndexId: 3}
-	h4 := &HostInfo{vpnAddrs: []netip.Addr{netip.MustParseAddr("0.0.0.1")}, localIndexId: 4}
+	h4 := &HostInfo{vpnAddrs: []netip.Addr{a}, localIndexId: 4}
-	h5 := &HostInfo{vpnAddrs: []netip.Addr{netip.MustParseAddr("0.0.0.1")}, localIndexId: 5}
+	h5 := &HostInfo{vpnAddrs: []netip.Addr{a}, localIndexId: 5}
-	h6 := &HostInfo{vpnAddrs: []netip.Addr{netip.MustParseAddr("0.0.0.1")}, localIndexId: 6}
+	h6 := &HostInfo{vpnAddrs: []netip.Addr{a}, localIndexId: 6}
 
 	hm.unlockedAddHostInfo(h6, f)
 	hm.unlockedAddHostInfo(h5, f)
@@ -104,94 +112,243 @@ func TestHostMap_DeleteHostInfo(t *testing.T) {
 	hm.unlockedAddHostInfo(h2, f)
 	hm.unlockedAddHostInfo(h1, f)
 
-	// h6 should be deleted
+	// h6 is evicted by the MaxHostInfosPerVpnIp cap; the rest are newest-first.
-	assert.Nil(t, h6.next)
+	assert.Nil(t, hm.QueryIndex(h6.localIndexId))
-	assert.Nil(t, h6.prev)
+	assert.Equal(t, []uint32{1, 2, 3, 4, 5}, chainIds(t, hm, a))
-	h := hm.QueryIndex(h6.localIndexId)
-	assert.Nil(t, h)
 
-	// Make sure we go h1 -> h2 -> h3 -> h4 -> h5
+	// Delete primary; not final since siblings remain.
-	prim := hm.QueryVpnAddr(netip.MustParseAddr("0.0.0.1"))
+	assert.False(t, hm.DeleteHostInfo(h1))
-	assert.Equal(t, h1.localIndexId, prim.localIndexId)
+	assert.Equal(t, []uint32{2, 3, 4, 5}, chainIds(t, hm, a))
-	assert.Equal(t, h2.localIndexId, prim.next.localIndexId)
-	assert.Nil(t, prim.prev)
-	assert.Equal(t, h1.localIndexId, h2.prev.localIndexId)
-	assert.Equal(t, h3.localIndexId, h2.next.localIndexId)
-	assert.Equal(t, h2.localIndexId, h3.prev.localIndexId)
-	assert.Equal(t, h4.localIndexId, h3.next.localIndexId)
-	assert.Equal(t, h3.localIndexId, h4.prev.localIndexId)
-	assert.Equal(t, h5.localIndexId, h4.next.localIndexId)
-	assert.Equal(t, h4.localIndexId, h5.prev.localIndexId)
-	assert.Nil(t, h5.next)
 
-	// Delete primary
+	// Deleting the same hostinfo again must not report final while siblings remain and must not
-	hm.DeleteHostInfo(h1)
+	// disturb the list. The old chain code got this wrong: the first delete nil'd next/prev, so a
-	assert.Nil(t, h1.prev)
+	// second delete looked final and wiped lighthouse state out from under the live sibling.
-	assert.Nil(t, h1.next)
+	assert.False(t, hm.DeleteHostInfo(h1))
+	assert.Equal(t, []uint32{2, 3, 4, 5}, chainIds(t, hm, a))
 
-	// Make sure we go h2 -> h3 -> h4 -> h5
+	// Delete a middle node.
-	prim = hm.QueryVpnAddr(netip.MustParseAddr("0.0.0.1"))
+	assert.False(t, hm.DeleteHostInfo(h3))
-	assert.Equal(t, h2.localIndexId, prim.localIndexId)
+	assert.Equal(t, []uint32{2, 4, 5}, chainIds(t, hm, a))
-	assert.Equal(t, h3.localIndexId, prim.next.localIndexId)
-	assert.Nil(t, prim.prev)
-	assert.Equal(t, h3.localIndexId, h2.next.localIndexId)
-	assert.Equal(t, h2.localIndexId, h3.prev.localIndexId)
-	assert.Equal(t, h4.localIndexId, h3.next.localIndexId)
-	assert.Equal(t, h3.localIndexId, h4.prev.localIndexId)
-	assert.Equal(t, h5.localIndexId, h4.next.localIndexId)
-	assert.Equal(t, h4.localIndexId, h5.prev.localIndexId)
-	assert.Nil(t, h5.next)
 
-	// Delete in the middle
+	// Delete the tail.
-	hm.DeleteHostInfo(h3)
+	assert.False(t, hm.DeleteHostInfo(h5))
-	assert.Nil(t, h3.prev)
+	assert.Equal(t, []uint32{2, 4}, chainIds(t, hm, a))
-	assert.Nil(t, h3.next)
 
-	// Make sure we go h2 -> h4 -> h5
+	// Delete the head; h4 remains and becomes primary.
-	prim = hm.QueryVpnAddr(netip.MustParseAddr("0.0.0.1"))
+	assert.False(t, hm.DeleteHostInfo(h2))
-	assert.Equal(t, h2.localIndexId, prim.localIndexId)
+	assert.Equal(t, []uint32{4}, chainIds(t, hm, a))
-	assert.Equal(t, h4.localIndexId, prim.next.localIndexId)
+	assert.Equal(t, h4, hm.QueryVpnAddr(a))
-	assert.Nil(t, prim.prev)
-	assert.Equal(t, h4.localIndexId, h2.next.localIndexId)
-	assert.Equal(t, h2.localIndexId, h4.prev.localIndexId)
-	assert.Equal(t, h5.localIndexId, h4.next.localIndexId)
-	assert.Equal(t, h4.localIndexId, h5.prev.localIndexId)
-	assert.Nil(t, h5.next)
 
-	// Delete the tail
+	// Delete the only remaining item; final is true and the address is gone.
-	hm.DeleteHostInfo(h5)
+	assert.True(t, hm.DeleteHostInfo(h4))
-	assert.Nil(t, h5.prev)
+	assert.Empty(t, chainIds(t, hm, a))
-	assert.Nil(t, h5.next)
+	assert.Nil(t, hm.QueryVpnAddr(a))
 
-	// Make sure we go h2 -> h4
+	// Deleting an already-gone hostinfo is still final; nothing holds the address anymore.
-	prim = hm.QueryVpnAddr(netip.MustParseAddr("0.0.0.1"))
+	assert.True(t, hm.DeleteHostInfo(h4))
-	assert.Equal(t, h2.localIndexId, prim.localIndexId)
+	assert.Empty(t, chainIds(t, hm, a))
-	assert.Equal(t, h4.localIndexId, prim.next.localIndexId)
+}
-	assert.Nil(t, prim.prev)
-	assert.Equal(t, h4.localIndexId, h2.next.localIndexId)
-	assert.Equal(t, h2.localIndexId, h4.prev.localIndexId)
-	assert.Nil(t, h4.next)
 
-	// Delete the head
+// TestHostMap_MakePrimary_DeletedHostInfo covers promoting a hostinfo that lost a race with
-	hm.DeleteHostInfo(h2)
+// tunnel teardown: swapPrimary and AddRelay decide to promote while holding a stale pointer and
-	assert.Nil(t, h2.prev)
+// only take the write lock after a delete fully unlinked the hostinfo. MakePrimary must be a
-	assert.Nil(t, h2.next)
+// no-op, not a resurrection that installs an unmanaged primary.
+func TestHostMap_MakePrimary_DeletedHostInfo(t *testing.T) {
+	l := test.NewLogger()
+	hm := newHostMap(l)
+	f := &Interface{}
+	a := netip.MustParseAddr("0.0.0.1")
 
-	// Make sure we only have h4
+	h1 := &HostInfo{vpnAddrs: []netip.Addr{a}, localIndexId: 1}
-	prim = hm.QueryVpnAddr(netip.MustParseAddr("0.0.0.1"))
+	h2 := &HostInfo{vpnAddrs: []netip.Addr{a}, localIndexId: 2}
-	assert.Equal(t, h4.localIndexId, prim.localIndexId)
+	hm.unlockedAddHostInfo(h1, f)
-	assert.Nil(t, prim.prev)
+	hm.unlockedAddHostInfo(h2, f)
-	assert.Nil(t, prim.next)
-	assert.Nil(t, h4.next)
 
-	// Delete the only item
+	// h1 is fully deleted while another goroutine still holds a pointer to it.
-	hm.DeleteHostInfo(h4)
+	assert.False(t, hm.DeleteHostInfo(h1))
-	assert.Nil(t, h4.prev)
+	assert.Equal(t, []uint32{2}, chainIds(t, hm, a))
-	assert.Nil(t, h4.next)
 
-	// Make sure we have nil
+	// The stale promote must not bring it back.
-	prim = hm.QueryVpnAddr(netip.MustParseAddr("0.0.0.1"))
+	hm.MakePrimary(h1)
-	assert.Nil(t, prim)
+	assert.Equal(t, []uint32{2}, chainIds(t, hm, a))
+	assert.Equal(t, h2, hm.QueryVpnAddr(a))
+	assert.Nil(t, hm.QueryIndex(h1.localIndexId))
+}
+
+// TestHostMap_QueryVpnAddrsRelayFor_NonPrimary makes sure a relay established on an older
+// hostinfo is still found after a newer tunnel without relay state takes primary for the same
+// address. The lookup checks the primary first and falls back to the rest of the list.
+func TestHostMap_QueryVpnAddrsRelayFor_NonPrimary(t *testing.T) {
+	l := test.NewLogger()
+	hm := newHostMap(l)
+	f := &Interface{}
+	relayAddr := netip.MustParseAddr("0.0.0.9")
+	target := netip.MustParseAddr("0.0.0.1")
+
+	older := &HostInfo{
+		vpnAddrs:     []netip.Addr{relayAddr},
+		localIndexId: 1,
+		relayState: RelayState{
+			relayForByAddr: map[netip.Addr]*Relay{},
+			relayForByIdx:  map[uint32]*Relay{},
+		},
+	}
+	older.relayState.InsertRelay(target, 100, &Relay{Type: ForwardingType, State: Established, LocalIndex: 100, PeerAddr: target})
+	hm.unlockedAddHostInfo(older, f)
+
+	// The relay is found on the primary.
+	h, r, err := hm.QueryVpnAddrsRelayFor([]netip.Addr{target}, relayAddr)
+	require.NoError(t, err)
+	assert.Equal(t, older, h)
+	assert.Equal(t, uint32(100), r.LocalIndex)
+
+	// A re-handshake with no relay state takes primary; the established relay on the older
+	// hostinfo must still be found through the fallback.
+	newer := &HostInfo{vpnAddrs: []netip.Addr{relayAddr}, localIndexId: 2}
+	hm.unlockedAddHostInfo(newer, f)
+	assert.Equal(t, []uint32{2, 1}, chainIds(t, hm, relayAddr))
+
+	h, r, err = hm.QueryVpnAddrsRelayFor([]netip.Addr{target}, relayAddr)
+	require.NoError(t, err)
+	assert.Equal(t, older, h)
+	assert.Equal(t, uint32(100), r.LocalIndex)
+
+	// No hostinfo at all is a plain miss.
+	_, _, err = hm.QueryVpnAddrsRelayFor([]netip.Addr{target}, netip.MustParseAddr("0.0.0.42"))
+	require.Error(t, err)
+}
+
+// TestHostMap_DeleteHostInfo_MultipleVpnAddrs exercises the case where a hostinfo carries more than one
+// vpnAddr and shares its next/prev chain with a live sibling. Deleting the head must not corrupt the
+// sibling: every address the sibling owns has to keep pointing at it. The pre-fix code unlinked the shared
+// chain once per vpnAddr, so on the first address it nil'd next/prev, and on the second address the node
+// looked already-detached: it dropped the map entry instead of promoting the sibling (and tripped the
+// isLastHostinfo relay teardown). See unlockedDeleteHostInfo.
+func TestHostMap_DeleteHostInfo_MultipleVpnAddrs(t *testing.T) {
+	l := test.NewLogger()
+	hm := newHostMap(l)
+
+	f := &Interface{}
+
+	a := netip.MustParseAddr("0.0.0.1")
+	b := netip.MustParseAddr("0.0.0.2")
+
+	// Two tunnels for the same peer, each reachable at both a and b.
+	other := &HostInfo{vpnAddrs: []netip.Addr{a, b}, localIndexId: 1}
+	head := &HostInfo{vpnAddrs: []netip.Addr{a, b}, localIndexId: 2}
+
+	hm.unlockedAddHostInfo(other, f)
+	hm.unlockedAddHostInfo(head, f)
+
+	// head is primary for both addresses, other is next in each address's list.
+	assert.Equal(t, head, hm.QueryVpnAddr(a))
+	assert.Equal(t, head, hm.QueryVpnAddr(b))
+	assert.Equal(t, []uint32{2, 1}, chainIds(t, hm, a))
+	assert.Equal(t, []uint32{2, 1}, chainIds(t, hm, b))
+
+	// Delete the head. other is still live, so it must become primary for BOTH addresses.
+	assert.False(t, hm.DeleteHostInfo(head))
+	assert.Equal(t, other, hm.QueryVpnAddr(a))
+	assert.Equal(t, other, hm.QueryVpnAddr(b))
+	assert.Equal(t, []uint32{1}, chainIds(t, hm, a))
+	assert.Equal(t, []uint32{1}, chainIds(t, hm, b))
+
+	// head is fully removed from the index map.
+	assert.Nil(t, hm.QueryIndex(head.localIndexId))
+}
+
+// TestHostMap_DeleteHostInfo_DivergentVpnAddrs covers chained hostinfos for the same peer whose
+// vpnAddrs sets differ (a re-handshake cert added a second address). Deleting the superset node
+// must not promote a sibling to an address it does not own.
+func TestHostMap_DeleteHostInfo_DivergentVpnAddrs(t *testing.T) {
+	l := test.NewLogger()
+	hm := newHostMap(l)
+	f := &Interface{}
+	a := netip.MustParseAddr("0.0.0.1")
+	b := netip.MustParseAddr("0.0.0.2")
+
+	// sub owns only a; super (a newer handshake) owns a and b.
+	sub := &HostInfo{vpnAddrs: []netip.Addr{a}, localIndexId: 1}
+	super := &HostInfo{vpnAddrs: []netip.Addr{a, b}, localIndexId: 2}
+	hm.unlockedAddHostInfo(sub, f)
+	hm.unlockedAddHostInfo(super, f)
+
+	assert.Equal(t, []uint32{2, 1}, chainIds(t, hm, a))
+	assert.Equal(t, []uint32{2}, chainIds(t, hm, b))
+
+	// Delete super: a promotes to sub (which owns it); b has no remaining owner and must be
+	// removed, not dangled at sub (which does not own b).
+	assert.False(t, hm.DeleteHostInfo(super))
+	assert.Equal(t, []uint32{1}, chainIds(t, hm, a))
+	assert.Empty(t, chainIds(t, hm, b))
+	assert.Equal(t, sub, hm.QueryVpnAddr(a))
+	assert.Nil(t, hm.QueryVpnAddr(b))
+	assert.Nil(t, hm.QueryIndex(super.localIndexId))
+
+	// Deleting sub cleans up fully.
+	assert.True(t, hm.DeleteHostInfo(sub))
+	assert.Nil(t, hm.QueryVpnAddr(a))
+	assertHostMapInvariants(t, hm)
+}
+
+// TestHostMap_AddDivergentOverlap covers a new hostinfo claiming addresses currently owned by two
+// DIFFERENT hostinfos. The old single shared next/prev chain overwrote a pointer and orphaned one
+// of them (in Indexes but unreachable via its address); independent per-address lists cannot.
+func TestHostMap_AddDivergentOverlap(t *testing.T) {
+	l := test.NewLogger()
+	hm := newHostMap(l)
+	f := &Interface{}
+	a := netip.MustParseAddr("0.0.0.1")
+	b := netip.MustParseAddr("0.0.0.2")
+
+	hiA := &HostInfo{vpnAddrs: []netip.Addr{a}, localIndexId: 1}
+	hiP := &HostInfo{vpnAddrs: []netip.Addr{b}, localIndexId: 2}
+	hm.unlockedAddHostInfo(hiA, f)
+	hm.unlockedAddHostInfo(hiP, f)
+
+	hiB := &HostInfo{vpnAddrs: []netip.Addr{a, b}, localIndexId: 3}
+	hm.unlockedAddHostInfo(hiB, f)
+
+	assert.Equal(t, []uint32{3, 1}, chainIds(t, hm, a))
+	assert.Equal(t, []uint32{3, 2}, chainIds(t, hm, b))
+	// hiA is still reachable via its address (not orphaned) and still indexed.
+	assert.Contains(t, chainIds(t, hm, a), hiA.localIndexId)
+	assert.NotNil(t, hm.QueryIndex(hiA.localIndexId))
+}
+
+// TestHostMap_MaxHostInfosPerVpnIp_MultipleVpnAddrs verifies the MaxHostInfosPerVpnIp overflow prune
+// (unlockedInnerAddHostInfo calls unlockedDeleteHostInfo on the oldest node once the chain is too long)
+// still behaves when hostinfos carry more than one vpnAddr. The pruned node is always the tail, so it is
+// primary for none of the addresses, and both address chains must stay consistent afterwards.
+func TestHostMap_MaxHostInfosPerVpnIp_MultipleVpnAddrs(t *testing.T) {
+	l := test.NewLogger()
+	hm := newHostMap(l)
+
+	f := &Interface{}
+
+	a := netip.MustParseAddr("0.0.0.1")
+	b := netip.MustParseAddr("0.0.0.2")
+
+	// Add one more than the cap, newest last so it becomes head. Every hostinfo owns both a and b.
+	hostinfos := make([]*HostInfo, 0, MaxHostInfosPerVpnIp+1)
+	for i := 0; i <= MaxHostInfosPerVpnIp; i++ {
+		hostinfos = append(hostinfos, &HostInfo{vpnAddrs: []netip.Addr{a, b}, localIndexId: uint32(i + 1)})
+	}
+	// Add oldest first (highest index in our slice) so the very first one added is the overflow victim.
+	for i := len(hostinfos) - 1; i >= 0; i-- {
+		hm.unlockedAddHostInfo(hostinfos[i], f)
+	}
+
+	oldest := hostinfos[len(hostinfos)-1]
+
+	// The oldest hostinfo was pruned from both lists and the index map.
+	assert.Nil(t, hm.QueryIndex(oldest.localIndexId))
+
+	// Both addresses hold exactly MaxHostInfosPerVpnIp survivors in the same order; oldest is absent.
+	require.Len(t, chainIds(t, hm, a), MaxHostInfosPerVpnIp)
+	assert.Equal(t, chainIds(t, hm, a), chainIds(t, hm, b), "both addresses must list the same survivors in the same order")
+	assert.NotContains(t, chainIds(t, hm, a), oldest.localIndexId)
+	assert.Equal(t, hm.QueryVpnAddr(a), hm.QueryVpnAddr(b))
 }
 
 func TestHostMap_reload(t *testing.T) {

inside.go

@@ -87,7 +87,7 @@ func (f *Interface) consumeInsidePacket(packet []byte, fwPacket *firewall.Packet
 }
 
 func (f *Interface) rejectInside(packet []byte, out []byte, q int) {
-	if !f.firewall.InSendReject {
+	if !f.firewall.OutboundSendReject {
 		return
 	}
 
@@ -103,7 +103,7 @@ func (f *Interface) rejectInside(packet []byte, out []byte, q int) {
 }
 
 func (f *Interface) rejectOutside(packet []byte, ci *ConnectionState, hostinfo *HostInfo, nb, out []byte, q int) {
-	if !f.firewall.OutSendReject {
+	if !f.firewall.InboundSendReject {
 		return
 	}
 
@@ -333,7 +333,7 @@ func (f *Interface) SendVia(via *HostInfo,
 		via.logger(f.l).Info("Failed to EncryptDanger in sendVia", "error", err)
 		return
 	}
-	err = f.writers[0].WriteTo(out, via.remote)
+	err = f.writers[0].WriteTo(out, via.GetRemote())
 	if err != nil {
 		via.logger(f.l).Info("Failed to WriteTo in sendVia", "error", err)
 	}
@@ -344,7 +344,7 @@ func (f *Interface) sendNoMetrics(t header.MessageType, st header.MessageSubType
 	if ci.eKey == nil {
 		return
 	}
-	useRelay := !remote.IsValid() && !hostinfo.remote.IsValid()
+	useRelay := !remote.IsValid() && !hostinfo.GetRemote().IsValid()
 	fullOut := out
 
 	if useRelay {
@@ -403,8 +403,8 @@ func (f *Interface) sendNoMetrics(t header.MessageType, st header.MessageSubType
 				"udpAddr", remote,
 			)
 		}
-	} else if hostinfo.remote.IsValid() {
+	} else if hr := hostinfo.GetRemote(); hr.IsValid() {
-		err = f.writers[q].WriteTo(out, hostinfo.remote)
+		err = f.writers[q].WriteTo(out, hr)
 		if err != nil {
 			hostinfo.logger(f.l).Error("Failed to write outgoing packet",
 				"error", err,

interface.go

@@ -215,6 +215,9 @@ func NewInterface(ctx context.Context, c *InterfaceConfig) (*Interface, error) {
 
 	ifce.connectionManager.intf = ifce
 
+	// Held until Close so waiting on the interface blocks until the resources are actually released
+	ifce.wg.Add(1)
+
 	return ifce, nil
 }
 
@@ -258,17 +261,16 @@ func (f *Interface) activate() error {
 		f.readers[i] = reader
 	}
 
-	f.wg.Add(1) // for us to wait on Close() to return
+	// On error the caller owns the cleanup, Control.Start cancels the service context
+	// before releasing our resources so a waiter never observes a live context
 	if err = f.inside.Activate(); err != nil {
-		f.wg.Done()
-		f.inside.Close()
 		return err
 	}
 
 	return nil
 }
 
-func (f *Interface) run() (func() error, error) {
+func (f *Interface) run() {
 	// Launch n queues to read packets from udp
 	for i := 0; i < f.routines; i++ {
 		f.wg.Go(func() {
@@ -283,13 +285,14 @@ func (f *Interface) run() (func() error, error) {
 		})
 	}
 
-	return func() error {
+}
-		f.wg.Wait()
+
-		if e := f.fatalErr.Load(); e != nil {
+func (f *Interface) wait() error {
-			return *e
+	f.wg.Wait()
-		}
+	if e := f.fatalErr.Load(); e != nil {
-		return nil
+		return *e
-	}, nil
+	}
+	return nil
 }
 
 // onFatal stores the first fatal reader error, and calls triggerShutdown if it was the first one
@@ -322,7 +325,10 @@ func (f *Interface) listenOut(i int) {
 		f.readOutsidePackets(ViaSender{UdpAddr: fromUdpAddr}, plaintext[:0], payload, h, fwPacket, lhh, nb, i, ctCache.Get())
 	})
 
-	if err != nil && !f.closed.Load() {
+	// An error after teardown began is shutdown noise, the closed flag covers resources
+	// Close releases itself and the cancelled ctx covers ones torn down by their owners
+	// reacting to it, like the user device pipes
+	if err != nil && !f.closed.Load() && f.ctx.Err() == nil {
 		f.l.Error("Error while reading inbound packet, closing", "error", err)
 		f.onFatal(err)
 	}
@@ -341,7 +347,8 @@ func (f *Interface) listenIn(reader io.ReadWriteCloser, i int) {
 	for {
 		n, err := reader.Read(packet)
 		if err != nil {
-			if !f.closed.Load() {
+			// Same shutdown noise handling as listenOut
+			if !f.closed.Load() && f.ctx.Err() == nil {
 				f.l.Error("Error while reading outbound packet, closing", "error", err, "reader", i)
 				f.onFatal(err)
 			}
@@ -542,9 +549,15 @@ func (f *Interface) GetCertState() *CertState {
 	return f.pki.getCertState()
 }
 
+// Close releases the interface's resources: the udp sockets and the tun device.
+// It is idempotent and safe to call at any point in the lifecycle, including on an interface that never activated,
+// calls after the first return nil without doing anything.
 func (f *Interface) Close() error {
+	if !f.closed.CompareAndSwap(false, true) {
+		return nil
+	}
+
 	var errs []error
-	f.closed.Store(true)
 
 	// Release the udp readers
 	for i, u := range f.writers {
@@ -560,6 +573,8 @@ func (f *Interface) Close() error {
 	if closeErr != nil {
 		errs = append(errs, closeErr)
 	}
+
+	// Release the construction token so waiters know the resources are gone
 	f.wg.Done()
 	return errors.Join(errs...)
 }

iputil/packet.go

@@ -344,7 +344,7 @@ func ipv6FindUpperProtocol(packet []byte) (nextHeader uint8, offset int, isFragm
 				return nextHeader, offset, isFragment
 			}
 			nextHeader = packet[offset]
-			offset += int(packet[offset+1]+1) << 3
+			offset += (int(packet[offset+1]) + 1) << 3
 
 		case 44: // Fragment
 			if len(packet) < offset+8 {
@@ -361,7 +361,7 @@ func ipv6FindUpperProtocol(packet []byte) (nextHeader uint8, offset int, isFragm
 				return nextHeader, offset, isFragment
 			}
 			nextHeader = packet[offset]
-			offset += int(packet[offset+1]+2) << 2
+			offset += (int(packet[offset+1]) + 2) << 2
 
 		default:
 			return nextHeader, offset, isFragment

lighthouse.go

@@ -1418,6 +1418,9 @@ func (lhh *LightHouseHandler) handleHostPunchNotification(n *NebulaMeta, fromVpn
 
 	remoteAllowList := lhh.lh.GetRemoteAllowList()
 	for _, a := range n.Details.V4AddrPorts {
+		if a == nil {
+			continue
+		}
 		b := protoV4AddrPortToNetAddrPort(a)
 		if remoteAllowList.Allow(detailsVpnAddr, b.Addr()) {
 			lhh.lh.punchy.Schedule(b, detailsVpnAddr)
@@ -1425,6 +1428,9 @@ func (lhh *LightHouseHandler) handleHostPunchNotification(n *NebulaMeta, fromVpn
 	}
 
 	for _, a := range n.Details.V6AddrPorts {
+		if a == nil {
+			continue
+		}
 		b := protoV6AddrPortToNetAddrPort(a)
 		if remoteAllowList.Allow(detailsVpnAddr, b.Addr()) {
 			lhh.lh.punchy.Schedule(b, detailsVpnAddr)
@@ -1454,7 +1460,7 @@ func protoV6AddrPortToNetAddrPort(ap *V6AddrPort) netip.AddrPort {
 	b := [16]byte{}
 	binary.BigEndian.PutUint64(b[:8], ap.Hi)
 	binary.BigEndian.PutUint64(b[8:], ap.Lo)
-	return netip.AddrPortFrom(netip.AddrFrom16(b), uint16(ap.Port))
+	return netip.AddrPortFrom(netip.AddrFrom16(b).Unmap(), uint16(ap.Port))
 }
 
 func netAddrToProtoAddr(addr netip.Addr) *Addr {
@@ -1494,7 +1500,9 @@ func (d *NebulaMetaDetails) GetRelays() []netip.Addr {
 
 	if len(d.RelayVpnAddrs) > 0 {
 		for _, r := range d.RelayVpnAddrs {
-			relays = append(relays, protoAddrToNetAddr(r))
+			if r != nil {
+				relays = append(relays, protoAddrToNetAddr(r))
+			}
 		}
 	}
 	return relays

main.go

@@ -130,6 +130,17 @@ func Main(c *config.C, configTest bool, buildVersion string, l *slog.Logger, dev
 	udpConns := make([]udp.Conn, routines)
 	port := c.GetInt("listen.port", 0)
 
+	// Callers get no handle to these until the Control is returned, release them on any error.
+	defer func() {
+		if reterr != nil {
+			for _, u := range udpConns {
+				if u != nil {
+					_ = u.Close()
+				}
+			}
+		}
+	}()
+
 	if !configTest {
 		rawListenHost := c.GetString("listen.host", "0.0.0.0")
 		var listenHost netip.Addr

outside.go

@@ -102,27 +102,31 @@ func (f *Interface) readOutsidePackets(via ViaSender, out []byte, packet []byte,
 		return
 	}
 
+	if len(packet) < header.Len+hostinfo.ConnectionState.dKey.Overhead() {
+		f.messageMetrics.RxInvalid(1)
+		if f.l.Enabled(context.Background(), slog.LevelDebug) {
+			f.l.Debug("packet too small", "from", via, "length", len(packet))
+		}
+		return
+	}
+
 	// All remaining packets are encrypted
-	ci := hostinfo.ConnectionState
-	if !ci.window.Check(f.l, h.MessageCounter) {
-		return
-	}
-
-	// Relay packets are special
 	if isMessageRelay {
+		// Relay packets are special, this branch should always early-return
+		if err = hostinfo.ConnectionState.VerifyRelay(f.l, h.MessageCounter, packet, nb); err != nil {
+			if f.l.Enabled(context.Background(), slog.LevelDebug) {
+				hostinfo.logger(f.l).Debug("Failed to verify relay packet", "error", err, "from", via, "header", h)
+			}
+			return
+		}
 		f.handleOutsideRelayPacket(hostinfo, via, out, packet, h, fwPacket, lhf, nb, q, localCache)
-
 		return
 	}
 
-	out, err = f.decrypt(hostinfo, h.MessageCounter, out, packet, h, nb)
+	out, err = hostinfo.ConnectionState.Decrypt(f.l, h.MessageCounter, out, packet, nb)
 	if err != nil {
 		if f.l.Enabled(context.Background(), slog.LevelDebug) {
-			hostinfo.logger(f.l).Debug("Failed to decrypt packet",
+			hostinfo.logger(f.l).Debug("Failed to decrypt packet", "error", err, "from", via, "header", h)
-				"error", err,
-				"from", via,
-				"header", h,
-			)
 		}
 		return
 	}
@@ -150,7 +154,8 @@ func (f *Interface) readOutsidePackets(via ViaSender, out []byte, packet []byte,
 		case header.TestReply:
 			// No-op, useful for the Roaming and connectionManager side-effects above
 		case header.TestRequest:
-			f.send(header.Test, header.TestReply, ci, hostinfo, out, nb, out)
+			//recycle the input packet ciphertext as our output buffer
+			f.send(header.Test, header.TestReply, hostinfo.ConnectionState, hostinfo, out, nb, packet)
 		default:
 			hostinfo.logger(f.l).Error("IsValidSubType was true, but unexpected test subtype seen", "from", via, "header", h)
 			return
@@ -169,27 +174,8 @@ func (f *Interface) readOutsidePackets(via ViaSender, out []byte, packet []byte,
 }
 
 func (f *Interface) handleOutsideRelayPacket(hostinfo *HostInfo, via ViaSender, out []byte, packet []byte, h *header.H, fwPacket *firewall.Packet, lhf *LightHouseHandler, nb []byte, q int, localCache firewall.ConntrackCache) {
-	// The entire body is sent as AD, not encrypted.
+	// Successfully validated the thing. Get rid of the Relay header and the AEAD tag
-	// The packet consists of a 16-byte parsed Nebula header, Associated Data-protected payload, and a trailing 16-byte AEAD signature value.
+	signedPayload := packet[header.Len : len(packet)-hostinfo.ConnectionState.dKey.Overhead()]
-	// The packet is guaranteed to be at least 16 bytes at this point, b/c it got past the h.Parse() call above. If it's
-	// otherwise malformed (meaning, there is no trailing 16 byte AEAD value), then this will result in at worst a 0-length slice
-	// which will gracefully fail in the DecryptDanger call.
-	signedPayload := packet[:len(packet)-hostinfo.ConnectionState.dKey.Overhead()]
-	signatureValue := packet[len(packet)-hostinfo.ConnectionState.dKey.Overhead():]
-	var err error
-	out, err = hostinfo.ConnectionState.dKey.DecryptDanger(out, signedPayload, signatureValue, h.MessageCounter, nb)
-	if err != nil {
-		return
-	}
-	// Advance the replay window now that the frame is authenticated
-	if !hostinfo.ConnectionState.window.Update(f.l, h.MessageCounter) {
-		if f.l.Enabled(context.Background(), slog.LevelDebug) {
-			hostinfo.logger(f.l).Debug("dropping out of window relay packet", "header", h)
-		}
-		return
-	}
-	// Successfully validated the thing. Get rid of the Relay header.
-	signedPayload = signedPayload[header.Len:]
 	// Pull the Roaming parts up here, and return in all call paths.
 	f.handleHostRoaming(hostinfo, via)
 	// Track usage of both the HostInfo and the Relay for the received & authenticated packet
@@ -213,7 +199,6 @@ func (f *Interface) handleOutsideRelayPacket(hostinfo *HostInfo, via ViaSender,
 		via = ViaSender{
 			UdpAddr:   via.UdpAddr,
 			relayHI:   hostinfo,
-			remoteIdx: relay.RemoteIndex,
 			relay:     relay,
 			IsRelayed: true,
 		}
@@ -234,9 +219,10 @@ func (f *Interface) handleOutsideRelayPacket(hostinfo *HostInfo, via ViaSender,
 		if targetRelay.State == Established {
 			switch targetRelay.Type {
 			case ForwardingType:
-				// Forward this packet through the relay tunnel
+				// Forward this packet through the relay tunnel, rebuilding it in place.
-				// Find the target HostInfo
+				// Encode overwrites the old outer header, and the new AEAD tag lands where the old one was
-				f.SendVia(targetHI, targetRelay, signedPayload, nb, out, false)
+				fwdBuf := packet[:0:len(packet)] // Cap to len(packet) to protect memory from a larger parent buffer
+				f.SendVia(targetHI, targetRelay, signedPayload, nb, fwdBuf, true)
 			case TerminalType:
 				hostinfo.logger(f.l).Error("Unexpected Relay Type of Terminal")
 				return
@@ -276,7 +262,8 @@ func (f *Interface) sendCloseTunnel(h *HostInfo) {
 }
 
 func (f *Interface) handleHostRoaming(hostinfo *HostInfo, via ViaSender) {
-	if !via.IsRelayed && hostinfo.remote != via.UdpAddr {
+	curRemote := hostinfo.GetRemote()
+	if !via.IsRelayed && curRemote != via.UdpAddr {
 		if !f.lightHouse.GetRemoteAllowList().AllowAll(hostinfo.vpnAddrs, via.UdpAddr.Addr()) {
 			if f.l.Enabled(context.Background(), slog.LevelDebug) {
 				hostinfo.logger(f.l).Debug("lighthouse.remote_allow_list denied roaming", "newAddr", via.UdpAddr)
@@ -288,7 +275,7 @@ func (f *Interface) handleHostRoaming(hostinfo *HostInfo, via ViaSender) {
 			if f.l.Enabled(context.Background(), slog.LevelDebug) {
 				hostinfo.logger(f.l).Debug("Suppressing roam back to previous remote",
 					"suppressSeconds", RoamingSuppressSeconds,
-					"udpAddr", hostinfo.remote,
+					"udpAddr", curRemote,
 					"newAddr", via.UdpAddr,
 				)
 			}
@@ -296,11 +283,11 @@ func (f *Interface) handleHostRoaming(hostinfo *HostInfo, via ViaSender) {
 		}
 
 		hostinfo.logger(f.l).Info("Host roamed to new udp ip/port.",
-			"udpAddr", hostinfo.remote,
+			"udpAddr", curRemote,
 			"newAddr", via.UdpAddr,
 		)
 		hostinfo.lastRoam = time.Now()
-		hostinfo.lastRoamRemote = hostinfo.remote
+		hostinfo.lastRoamRemote = curRemote
 		hostinfo.SetRemote(via.UdpAddr)
 	}
 
@@ -420,16 +407,14 @@ func parseV6(data []byte, incoming bool, fp *firewall.Packet) error {
 			if dataLen <= offset+1 {
 				break
 			}
-
+			next = (int(data[offset+1]) + 2) << 2
-			next = int(data[offset+1]+2) << 2
 
 		default:
 			// Normal ipv6 header length processing
 			if dataLen <= offset+1 {
 				break
 			}
-
+			next = (int(data[offset+1]) + 1) << 3
-			next = int(data[offset+1]+1) << 3
 		}
 
 		if next <= 0 {
@@ -504,20 +489,6 @@ func parseV4(data []byte, incoming bool, fp *firewall.Packet) error {
 	return nil
 }
 
-func (f *Interface) decrypt(hostinfo *HostInfo, mc uint64, out []byte, packet []byte, h *header.H, nb []byte) ([]byte, error) {
-	var err error
-	out, err = hostinfo.ConnectionState.dKey.DecryptDanger(out, packet[:header.Len], packet[header.Len:], mc, nb)
-	if err != nil {
-		return nil, err
-	}
-
-	if !hostinfo.ConnectionState.window.Update(f.l, mc) {
-		return nil, ErrOutOfWindow
-	}
-
-	return out, nil
-}
-
 func (f *Interface) handleOutsideMessagePacket(hostinfo *HostInfo, out []byte, packet []byte, fwPacket *firewall.Packet, nb []byte, q int, localCache firewall.ConntrackCache) {
 	err := newPacket(out, true, fwPacket)
 	if err != nil {
@@ -589,10 +560,11 @@ func (f *Interface) handleRecvError(addr netip.AddrPort, h *header.H) {
 		return
 	}
 
-	if hostinfo.remote.IsValid() && hostinfo.remote != addr {
+	hr := hostinfo.GetRemote()
+	if hr.IsValid() && hr != addr {
 		f.l.Info("Someone spoofing recv_errors?",
 			"addr", addr,
-			"hostinfoRemote", hostinfo.remote,
+			"hostinfoRemote", hr,
 		)
 		return
 	}

outside_test.go

@@ -640,3 +640,38 @@ func serializeAH(ah *layers.IPSecAH) []byte {
 
 	return buf.Bytes()
 }
+
+// Test_newPacket_v6ExtHeaderOverflow is a regression test for the IPv6 extension-header
+// length uint8 overflow in parseV6. A Destination-Options header with HdrExtLen=255 spans
+// (255+1)*8 = 2048 bytes, so the real transport header sits at offset 2088. Before the fix
+// the advance was computed in uint8 and wrapped to 0 (then clamped to 8), so the firewall
+// read the transport header ~2KB too early from attacker-controlled option bytes while the
+// host OS parses the real header, a firewall port/proto bypass. The fix makes parseV6 land
+// on the same offset the host does.
+func Test_newPacket_v6ExtHeaderOverflow(t *testing.T) {
+	p := &firewall.Packet{}
+
+	const (
+		hdrLen      = 40              // IPv6 header
+		extLen      = 2048            // (255+1)*8, the true Destination-Options header size
+		realTCPAt   = hdrLen + extLen // 2088, where the host reads the transport header
+		forgedTCPAt = hdrLen + 8      // 48, where the pre-fix wrapped+clamped walk landed
+	)
+
+	pkt := make([]byte, realTCPAt+4)
+	pkt[0] = 0x60                                   // version 6
+	pkt[6] = byte(layers.IPProtocolIPv6Destination) // NextHeader -> Destination Options
+	pkt[40] = byte(firewall.ProtoTCP)               // Dest-Options NextHeader -> TCP
+	pkt[41] = 255                                   // HdrExtLen = 255
+
+	// Forged transport header at the pre-fix (wrong) offset: dst port 443.
+	binary.BigEndian.PutUint16(pkt[forgedTCPAt+2:forgedTCPAt+4], 443)
+	// Real transport header at the offset the host actually uses: dst port 22.
+	binary.BigEndian.PutUint16(pkt[realTCPAt+2:realTCPAt+4], 22)
+
+	require.NoError(t, newPacket(pkt, true, p))
+	assert.Equal(t, uint8(firewall.ProtoTCP), p.Protocol)
+	// LocalPort is the destination port for incoming traffic. It must be the real port (22)
+	// the host delivers to, not the forged 443 at the overflowed offset.
+	assert.Equal(t, uint16(22), p.LocalPort, "firewall must parse the real transport header, not the overflowed offset")
+}

overlay/tun_android.go

@@ -40,6 +40,7 @@ func newTunFromFd(c *config.C, l *slog.Logger, deviceFd int, vpnNetworks []netip
 
 	err := t.reload(c, true)
 	if err != nil {
+		_ = file.Close()
 		return nil, err
 	}
 

overlay/tun_darwin.go

@@ -23,7 +23,7 @@ import (
 )
 
 type tun struct {
-	io.ReadWriteCloser
+	f           *os.File
 	Device      string
 	vpnNetworks []netip.Prefix
 	DefaultMTU  int
@@ -31,9 +31,6 @@ type tun struct {
 	routeTree   atomic.Pointer[bart.Table[routing.Gateways]]
 	linkAddr    *netroute.LinkAddr
 	l           *slog.Logger
-
-	// cache out buffer since we need to prepend 4 bytes for tun metadata
-	out []byte
 }
 
 type ifReq struct {
@@ -124,11 +121,11 @@ func newTun(c *config.C, l *slog.Logger, vpnNetworks []netip.Prefix, _ bool) (*t
 	}
 
 	t := &tun{
-		ReadWriteCloser: os.NewFile(uintptr(fd), ""),
+		f:           os.NewFile(uintptr(fd), ""),
-		Device:          name,
+		Device:      name,
-		vpnNetworks:     vpnNetworks,
+		vpnNetworks: vpnNetworks,
-		DefaultMTU:      c.GetInt("tun.mtu", DefaultMTU),
+		DefaultMTU:  c.GetInt("tun.mtu", DefaultMTU),
-		l:               l,
+		l:           l,
 	}
 
 	err = t.reload(c, true)
@@ -158,8 +155,8 @@ func newTunFromFd(_ *config.C, _ *slog.Logger, _ int, _ []netip.Prefix) (*tun, e
 }
 
 func (t *tun) Close() error {
-	if t.ReadWriteCloser != nil {
+	if t.f != nil {
-		return t.ReadWriteCloser.Close()
+		return t.f.Close()
 	}
 	return nil
 }
@@ -502,42 +499,103 @@ func delRoute(prefix netip.Prefix, gateway netroute.Addr) error {
 	return nil
 }
 
+// tunWritev and tunReadv are linkname'd to x/sys/unix's libc-routed writev/readv stubs so the
+// calls go through libSystem's pinned trampoline. A raw syscall.Syscall(SYS_WRITEV/SYS_READV, ...)
+// on darwin/arm64 emits an SVC #0x80 trap (see $GOROOT/src/syscall/asm_darwin_arm64.s), the path
+// Apple keeps warning they will eventually disallow. We pull the low-level stubs instead of calling
+// unix.Writev/unix.Readv because those take [][]byte and rebuild the []Iovec every call, which
+// heap-allocates the header; linkname'ing the stubs lets us hand them our own stack-allocated
+// iovecs. See golang/go#78049.
+
+//go:linkname tunWritev golang.org/x/sys/unix.writev
+//go:noescape
+func tunWritev(fd int, iovecs []unix.Iovec) (n int, err error)
+
+//go:linkname tunReadv golang.org/x/sys/unix.readv
+//go:noescape
+func tunReadv(fd int, iovecs []unix.Iovec) (n int, err error)
+
+// Read pulls one IP packet off the utun device, scattering the 4 byte protocol header away from
+// the packet so the payload lands directly in to.
 func (t *tun) Read(to []byte) (int, error) {
-	buf := make([]byte, len(to)+4)
+	var head [4]byte
 
-	n, err := t.ReadWriteCloser.Read(buf)
+	rc, err := t.f.SyscallConn()
+	if err != nil {
+		return 0, err
+	}
 
-	copy(to, buf[4:])
+	var n int
-	return n - 4, err
+	var callErr error
+	err = rc.Read(func(fd uintptr) bool {
+		iovecs := []unix.Iovec{
+			{Base: &head[0], Len: 4},
+			{Base: &to[0], Len: uint64(len(to))},
+		}
+		n, callErr = tunReadv(int(fd), iovecs)
+		if errno, ok := callErr.(syscall.Errno); ok && errno.Temporary() {
+			return false
+		}
+		return true
+	})
+	if err != nil {
+		return 0, err
+	}
+	if callErr != nil {
+		return 0, callErr
+	}
+	if n < 4 {
+		return 0, nil
+	}
+	return n - 4, nil
 }
 
-// Write is only valid for single threaded use
+// Write pushes one IP packet onto the utun device.
 func (t *tun) Write(from []byte) (int, error) {
-	buf := t.out
-	if cap(buf) < len(from)+4 {
-		buf = make([]byte, len(from)+4)
-		t.out = buf
-	}
-	buf = buf[:len(from)+4]
-
 	if len(from) == 0 {
 		return 0, syscall.EIO
 	}
 
-	// Determine the IP Family for the NULL L2 Header
 	ipVer := from[0] >> 4
-	if ipVer == 4 {
+	var head [4]byte
-		buf[3] = syscall.AF_INET
+	switch ipVer {
-	} else if ipVer == 6 {
+	case 4:
-		buf[3] = syscall.AF_INET6
+		head[3] = syscall.AF_INET
-	} else {
+	case 6:
+		head[3] = syscall.AF_INET6
+	default:
 		return 0, fmt.Errorf("unable to determine IP version from packet")
 	}
 
-	copy(buf[4:], from)
+	// Grab rc as a local so the compiler can devirtualize the call and keep the closure on the stack.
+	rc, err := t.f.SyscallConn()
+	if err != nil {
+		return 0, err
+	}
 
-	n, err := t.ReadWriteCloser.Write(buf)
+	var n int
-	return n - 4, err
+	var callErr error
+	err = rc.Write(func(fd uintptr) bool {
+		iovecs := []unix.Iovec{
+			{Base: &head[0], Len: 4},
+			{Base: &from[0], Len: uint64(len(from))},
+		}
+		n, callErr = tunWritev(int(fd), iovecs)
+		// Type-assert to syscall.Errno so the EAGAIN/EWOULDBLOCK/EINTR check doesn't box the errno
+		// constants into error interfaces on every call.
+		if errno, ok := callErr.(syscall.Errno); ok && errno.Temporary() {
+			return false
+		}
+		return true
+	})
+	if err != nil {
+		return 0, err
+	}
+	if callErr != nil {
+		return 0, callErr
+	}
+
+	return n - 4, nil
 }
 
 func (t *tun) Networks() []netip.Prefix {

overlay/tun_freebsd.go

@@ -659,7 +659,6 @@ func addRoute(prefix netip.Prefix, gateway netroute.Addr) error {
 				return fmt.Errorf("failed to create route.RouteMessage for change: %w", err)
 			}
 			_, err = unix.Write(sock, data[:])
-			fmt.Println("DOING CHANGE")
 			return err
 		}
 		return fmt.Errorf("failed to write route.RouteMessage to socket: %w", err)

overlay/tun_ios.go

@@ -18,6 +18,7 @@ import (
 	"github.com/slackhq/nebula/config"
 	"github.com/slackhq/nebula/routing"
 	"github.com/slackhq/nebula/util"
+	"golang.org/x/sys/unix"
 )
 
 type tun struct {
@@ -33,6 +34,12 @@ func newTun(_ *config.C, _ *slog.Logger, _ []netip.Prefix, _ bool) (*tun, error)
 }
 
 func newTunFromFd(c *config.C, l *slog.Logger, deviceFd int, vpnNetworks []netip.Prefix) (*tun, error) {
+	if err := unix.SetNonblock(deviceFd, true); err != nil {
+		// We own the fd from the moment it is handed to us, same as the reload error path below
+		_ = unix.Close(deviceFd)
+		return nil, fmt.Errorf("failed to set the tun fd to non-blocking mode: %w", err)
+	}
+
 	file := os.NewFile(uintptr(deviceFd), "/dev/tun")
 	t := &tun{
 		vpnNetworks:     vpnNetworks,
@@ -42,6 +49,7 @@ func newTunFromFd(c *config.C, l *slog.Logger, deviceFd int, vpnNetworks []netip
 
 	err := t.reload(c, true)
 	if err != nil {
+		_ = file.Close()
 		return nil, err
 	}
 

overlay/tun_openbsd.go

@@ -57,8 +57,6 @@ type tun struct {
 	l           *slog.Logger
 	f           *os.File
 	fd          int
-	// cache out buffer since we need to prepend 4 bytes for tun metadata
-	out []byte
 }
 
 var deviceNameRE = regexp.MustCompile(`^tun[0-9]+$`)
@@ -124,42 +122,103 @@ func (t *tun) Close() error {
 	return nil
 }
 
+// tunWritev and tunReadv are linkname'd to x/sys/unix's libc-routed writev/readv stubs so the
+// calls go through libc's pinned trampoline. OpenBSD's pinsyscall protection rejects a raw
+// syscall.Syscall(SYS_WRITEV/SYS_READV, ...) because it doesn't originate from a libc-pinned
+// address, so we can't use the syscall.Syscall pattern that freebsd / netbsd use. We pull the
+// low-level stubs instead of calling unix.Writev/unix.Readv because those take [][]byte and rebuild
+// the []Iovec every call, which heap-allocates the header; linkname'ing the stubs lets us hand them
+// our own stack-allocated iovecs. See golang/go#78049.
+
+//go:linkname tunWritev golang.org/x/sys/unix.writev
+//go:noescape
+func tunWritev(fd int, iovecs []unix.Iovec) (n int, err error)
+
+//go:linkname tunReadv golang.org/x/sys/unix.readv
+//go:noescape
+func tunReadv(fd int, iovecs []unix.Iovec) (n int, err error)
+
+// Read pulls one IP packet off the tun device, scattering the 4 byte protocol header away from the
+// packet so the payload lands directly in to.
 func (t *tun) Read(to []byte) (int, error) {
-	buf := make([]byte, len(to)+4)
+	var head [4]byte
 
-	n, err := t.f.Read(buf)
+	rc, err := t.f.SyscallConn()
+	if err != nil {
+		return 0, err
+	}
 
-	copy(to, buf[4:])
+	var n int
-	return n - 4, err
+	var callErr error
+	err = rc.Read(func(fd uintptr) bool {
+		iovecs := []unix.Iovec{
+			{Base: &head[0], Len: 4},
+			{Base: &to[0], Len: uint64(len(to))},
+		}
+		n, callErr = tunReadv(int(fd), iovecs)
+		if errno, ok := callErr.(syscall.Errno); ok && errno.Temporary() {
+			return false
+		}
+		return true
+	})
+	if err != nil {
+		return 0, err
+	}
+	if callErr != nil {
+		return 0, callErr
+	}
+	if n < 4 {
+		return 0, nil
+	}
+	return n - 4, nil
 }
 
-// Write is only valid for single threaded use
+// Write pushes one IP packet onto the tun device.
 func (t *tun) Write(from []byte) (int, error) {
-	buf := t.out
-	if cap(buf) < len(from)+4 {
-		buf = make([]byte, len(from)+4)
-		t.out = buf
-	}
-	buf = buf[:len(from)+4]
-
 	if len(from) == 0 {
 		return 0, syscall.EIO
 	}
 
-	// Determine the IP Family for the NULL L2 Header
 	ipVer := from[0] >> 4
-	if ipVer == 4 {
+	var head [4]byte
-		buf[3] = syscall.AF_INET
+	switch ipVer {
-	} else if ipVer == 6 {
+	case 4:
-		buf[3] = syscall.AF_INET6
+		head[3] = syscall.AF_INET
-	} else {
+	case 6:
+		head[3] = syscall.AF_INET6
+	default:
 		return 0, fmt.Errorf("unable to determine IP version from packet")
 	}
 
-	copy(buf[4:], from)
+	// Grab rc as a local so the compiler can devirtualize the call and keep the closure on the stack.
+	rc, err := t.f.SyscallConn()
+	if err != nil {
+		return 0, err
+	}
 
-	n, err := t.f.Write(buf)
+	var n int
-	return n - 4, err
+	var callErr error
+	err = rc.Write(func(fd uintptr) bool {
+		iovecs := []unix.Iovec{
+			{Base: &head[0], Len: 4},
+			{Base: &from[0], Len: uint64(len(from))},
+		}
+		n, callErr = tunWritev(int(fd), iovecs)
+		// Type-assert to syscall.Errno so the EAGAIN/EWOULDBLOCK/EINTR check doesn't box the errno
+		// constants into error interfaces on every call.
+		if errno, ok := callErr.(syscall.Errno); ok && errno.Temporary() {
+			return false
+		}
+		return true
+	})
+	if err != nil {
+		return 0, err
+	}
+	if callErr != nil {
+		return 0, callErr
+	}
+
+	return n - 4, nil
 }
 
 func (t *tun) addIp(cidr netip.Prefix) error {

punchy.go

@@ -174,9 +174,9 @@ func (p *Punchy) SendPunch(hostinfo *HostInfo) {
 
 	if p.punchEverything.Load() {
 		p.sendPunchToAllRemotes(hostinfo)
-	} else if hostinfo.remote.IsValid() {
+	} else if hr := hostinfo.GetRemote(); hr.IsValid() {
 		p.metricPunchyTx.Inc(1)
-		p.punchConn.WriteTo([]byte{1}, hostinfo.remote)
+		p.punchConn.WriteTo([]byte{1}, hr)
 	}
 }
 

relay_manager.go

@@ -94,7 +94,7 @@ func (rm *relayManager) StartRelays(f *Interface, vpnIp netip.Addr, hh *Handshak
 		}
 
 		relayHostInfo := rm.hostmap.QueryVpnAddr(relay)
-		if relayHostInfo == nil || !relayHostInfo.remote.IsValid() {
+		if relayHostInfo == nil || !relayHostInfo.GetRemote().IsValid() {
 			hl.Log(context.Background(), level, "Establish tunnel to relay target", "relay", relay.String())
 			f.Handshake(relay)
 			continue
@@ -104,10 +104,13 @@ func (rm *relayManager) StartRelays(f *Interface, vpnIp netip.Addr, hh *Handshak
 		existingRelay, ok := relayHostInfo.relayState.QueryRelayForByIp(vpnIp)
 		if !ok {
 			// No relays exist or requested yet.
-			if relayHostInfo.remote.IsValid() {
+			if relayHostInfo.GetRemote().IsValid() {
 				idx, err := AddRelay(rm.l, relayHostInfo, rm.hostmap, vpnIp, nil, TerminalType, Requested)
 				if err != nil {
+					// No local relay state was installed, so a CreateRelayRequest would hand the
+					// peer an index we could never resolve. Skip it.
 					hl.Info("Failed to add relay to hostmap", "relay", relay.String(), "error", err)
+					continue
 				}
 
 				m := NebulaControl{
@@ -237,7 +240,12 @@ func AddRelay(l *slog.Logger, relayHostInfo *HostInfo, hm *HostMap, vpnIp netip.
 			// Avoid standing up a relay that can't be used since only the primary hostinfo
 			// will be pointed to by the relay logic
 			//TODO: if there was an existing primary and it had relay state, should we merge?
-			hm.unlockedMakePrimary(relayHostInfo)
+			if !hm.unlockedMakePrimary(relayHostInfo) {
+				// The tunnel was torn down after the caller grabbed relayHostInfo. A relay standing
+				// on an unlinked hostinfo would never carry traffic, and its Relays entry could
+				// never be reclaimed since the delete-time cleanup has already run.
+				return 0, errors.New("relay hostinfo is no longer in the hostmap")
+			}
 
 			hm.Relays[index] = relayHostInfo
 			newRelay := Relay{
@@ -309,6 +317,22 @@ func (rm *relayManager) HandleControlMsg(h *HostInfo, d []byte, f *Interface) {
 		v = cert.Version2
 	}
 
+	// validate:
+	switch msg.Type {
+	case NebulaControl_CreateRelayRequest, NebulaControl_CreateRelayResponse:
+		if msg.RelayFromAddr == nil {
+			if f.l.Enabled(context.Background(), slog.LevelDebug) {
+				h.logger(f.l).Debug("Control message received with nil RelayFromAddr", "type", msg.Type)
+			}
+			return
+		} else if msg.RelayToAddr == nil {
+			if f.l.Enabled(context.Background(), slog.LevelDebug) {
+				h.logger(f.l).Debug("Control message received with nil RelayToAddr", "type", msg.Type)
+			}
+			return
+		}
+	}
+
 	switch msg.Type {
 	case NebulaControl_CreateRelayRequest:
 		rm.handleCreateRelayRequest(v, h, f, msg)
@@ -318,6 +342,7 @@ func (rm *relayManager) HandleControlMsg(h *HostInfo, d []byte, f *Interface) {
 }
 
 func (rm *relayManager) handleCreateRelayResponse(v cert.Version, h *HostInfo, f *Interface, m *NebulaControl) {
+	//nil-checks for protoAddrToNetAddr handled by caller
 	relayFrom := protoAddrToNetAddr(m.RelayFromAddr)
 	relayTo := protoAddrToNetAddr(m.RelayToAddr)
 	rm.l.Info("handleCreateRelayResponse",
@@ -399,6 +424,7 @@ func (rm *relayManager) handleCreateRelayResponse(v cert.Version, h *HostInfo, f
 }
 
 func (rm *relayManager) handleCreateRelayRequest(v cert.Version, h *HostInfo, f *Interface, m *NebulaControl) {
+	//nil-checks for protoAddrToNetAddr handled by caller
 	from := protoAddrToNetAddr(m.RelayFromAddr)
 	target := protoAddrToNetAddr(m.RelayToAddr)
 
@@ -508,7 +534,7 @@ func (rm *relayManager) handleCreateRelayRequest(v cert.Version, h *HostInfo, f
 			f.Handshake(target)
 			return
 		}
-		if !peer.remote.IsValid() {
+		if !peer.GetRemote().IsValid() {
 			// Only create relays to peers for whom I have a direct connection
 			return
 		}

remote_list.go

@@ -344,6 +344,9 @@ func (r *RemoteList) CopyCache() *CacheMap {
 			}
 
 			for _, a := range mc.v4.reported {
+				if a == nil {
+					continue
+				}
 				c.Reported = append(c.Reported, protoV4AddrPortToNetAddrPort(a))
 			}
 		}
@@ -354,6 +357,9 @@ func (r *RemoteList) CopyCache() *CacheMap {
 			}
 
 			for _, a := range mc.v6.reported {
+				if a == nil {
+					continue
+				}
 				c.Reported = append(c.Reported, protoV6AddrPortToNetAddrPort(a))
 			}
 		}
@@ -582,6 +588,9 @@ func (r *RemoteList) unlockedCollect() {
 			}
 
 			for _, v := range c.v4.reported {
+				if v == nil {
+					continue
+				}
 				u := protoV4AddrPortToNetAddrPort(v)
 				if !r.unlockedIsBad(u) {
 					addrs = append(addrs, u)
@@ -598,6 +607,9 @@ func (r *RemoteList) unlockedCollect() {
 			}
 
 			for _, v := range c.v6.reported {
+				if v == nil {
+					continue
+				}
 				u := protoV6AddrPortToNetAddrPort(v)
 				if !r.unlockedIsBad(u) {
 					addrs = append(addrs, u)

service/service.go

@@ -43,12 +43,25 @@ type Service struct {
 	}
 }
 
-func New(control *nebula.Control) (*Service, error) {
+func New(control *nebula.Control) (_ *Service, reterr error) {
-	wait, err := control.Start()
+	// Check this before Start so a failure doesn't leave a running nebula
+	device, ok := control.Device().(*overlay.UserDevice)
+	if !ok {
+		return nil, errors.New("must be using user device")
+	}
+
+	err := control.Start()
 	if err != nil {
 		return nil, err
 	}
 
+	// Anything that fails after a successful Start must tear nebula back down
+	defer func() {
+		if reterr != nil {
+			control.Stop()
+		}
+	}()
+
 	ctx := control.Context()
 	eg, ctx := errgroup.WithContext(ctx)
 	s := Service{
@@ -57,11 +70,6 @@ func New(control *nebula.Control) (*Service, error) {
 	}
 	s.mu.listeners = map[uint16]*tcpListener{}
 
-	device, ok := control.Device().(*overlay.UserDevice)
-	if !ok {
-		return nil, errors.New("must be using user device")
-	}
-
 	s.ipstack = stack.New(stack.Options{
 		NetworkProtocols:   []stack.NetworkProtocolFactory{ipv4.NewProtocol, ipv6.NewProtocol},
 		TransportProtocols: []stack.TransportProtocolFactory{tcp.NewProtocol, udp.NewProtocol, icmp.NewProtocol4, icmp.NewProtocol6},
@@ -147,7 +155,7 @@ func New(control *nebula.Control) (*Service, error) {
 	// Add the nebula wait function to the group so a fatal reader error
 	// propagates out through errgroup.Wait().
 	eg.Go(func() error {
-		return wait()
+		return control.Wait()
 	})
 
 	return &s, nil

sshd/server.go

@@ -7,6 +7,7 @@ import (
 	"fmt"
 	"log/slog"
 	"net"
+	"sync"
 
 	"github.com/armon/go-radix"
 	"golang.org/x/crypto/ssh"
@@ -18,6 +19,8 @@ type SSHServer struct {
 
 	certChecker *ssh.CertChecker
 
+	// authLock guards trustedKeys and trustedCAs
+	authLock sync.RWMutex
 	// Map of user -> authorized keys
 	trustedKeys map[string]map[string]bool
 	trustedCAs  []ssh.PublicKey
@@ -45,6 +48,8 @@ func NewSSHServer(ctx context.Context, l *slog.Logger) (*SSHServer, error) {
 
 	cc := ssh.CertChecker{
 		IsUserAuthority: func(auth ssh.PublicKey) bool {
+			s.authLock.RLock()
+			defer s.authLock.RUnlock()
 			for _, ca := range s.trustedCAs {
 				if bytes.Equal(ca.Marshal(), auth.Marshal()) {
 					return true
@@ -57,6 +62,8 @@ func NewSSHServer(ctx context.Context, l *slog.Logger) (*SSHServer, error) {
 			pk := string(pubKey.Marshal())
 			fp := ssh.FingerprintSHA256(pubKey)
 
+			s.authLock.RLock()
+			defer s.authLock.RUnlock()
 			tk, ok := s.trustedKeys[c.User()]
 			if !ok {
 				return nil, fmt.Errorf("unknown user %s", c.User())
@@ -105,11 +112,15 @@ func (s *SSHServer) SetHostKey(hostPrivateKey []byte) error {
 }
 
 func (s *SSHServer) ClearTrustedCAs() {
+	s.authLock.Lock()
 	s.trustedCAs = []ssh.PublicKey{}
+	s.authLock.Unlock()
 }
 
 func (s *SSHServer) ClearAuthorizedKeys() {
+	s.authLock.Lock()
 	s.trustedKeys = make(map[string]map[string]bool)
+	s.authLock.Unlock()
 }
 
 // AddTrustedCA adds a trusted CA for user certificates
@@ -119,7 +130,9 @@ func (s *SSHServer) AddTrustedCA(pubKey string) error {
 		return err
 	}
 
+	s.authLock.Lock()
 	s.trustedCAs = append(s.trustedCAs, pk)
+	s.authLock.Unlock()
 	s.l.Info("Trusted CA key", "sshKey", pubKey)
 	return nil
 }
@@ -131,6 +144,7 @@ func (s *SSHServer) AddAuthorizedKey(user, pubKey string) error {
 		return err
 	}
 
+	s.authLock.Lock()
 	tk, ok := s.trustedKeys[user]
 	if !ok {
 		tk = make(map[string]bool)
@@ -138,6 +152,7 @@ func (s *SSHServer) AddAuthorizedKey(user, pubKey string) error {
 	}
 
 	tk[string(pk.Marshal())] = true
+	s.authLock.Unlock()
 	s.l.Info("Authorized ssh key",
 		"sshKey", pubKey,
 		"sshUser", user,

udp/udp_linux.go

@@ -4,12 +4,13 @@
 package udp
 
 import (
-	"context"
 	"encoding/binary"
+	"errors"
 	"fmt"
 	"log/slog"
 	"net"
 	"net/netip"
+	"sync/atomic"
 	"syscall"
 	"unsafe"
 
@@ -19,58 +20,51 @@ import (
 )
 
 type StdConn struct {
-	udpConn *net.UDPConn
+	sysFd  int
-	rawConn syscall.RawConn
+	closed atomic.Bool
-	isV4    bool
+	isV4   bool
-	l       *slog.Logger
+	l      *slog.Logger
-	batch   int
+	batch  int
-}
-
-func setReusePort(network, address string, c syscall.RawConn) error {
-	var opErr error
-	err := c.Control(func(fd uintptr) {
-		opErr = unix.SetsockoptInt(int(fd), unix.SOL_SOCKET, unix.SO_REUSEPORT, 1)
-		//CloseOnExec already set by the runtime
-	})
-	if err != nil {
-		return err
-	}
-	return opErr
 }
 
 func NewListener(l *slog.Logger, ip netip.Addr, port int, multi bool, batch int) (Conn, error) {
-	listen := netip.AddrPortFrom(ip, uint16(port))
+	af := unix.AF_INET6
-	lc := net.ListenConfig{}
+	if ip.Is4() {
+		af = unix.AF_INET
+	}
+	syscall.ForkLock.RLock()
+	fd, err := unix.Socket(af, unix.SOCK_DGRAM, unix.IPPROTO_UDP)
+	if err == nil {
+		unix.CloseOnExec(fd)
+	}
+	syscall.ForkLock.RUnlock()
+	if err != nil {
+		return nil, fmt.Errorf("unable to open socket: %w", err)
+	}
+
 	if multi {
-		lc.Control = setReusePort
+		if err = unix.SetsockoptInt(fd, unix.SOL_SOCKET, unix.SO_REUSEPORT, 1); err != nil {
-	}
+			_ = unix.Close(fd)
-	//this context is only used during the bind operation, you can't cancel it to kill the socket
+			return nil, fmt.Errorf("unable to set SO_REUSEPORT: %w", err)
-	pc, err := lc.ListenPacket(context.Background(), "udp", listen.String())
+		}
-	if err != nil {
-		return nil, fmt.Errorf("unable to open socket: %s", err)
-	}
-	udpConn := pc.(*net.UDPConn)
-	rawConn, err := udpConn.SyscallConn()
-	if err != nil {
-		_ = udpConn.Close()
-		return nil, err
-	}
-	//gotta find out if we got an AF_INET6 socket or not:
-	out := &StdConn{
-		udpConn: udpConn,
-		rawConn: rawConn,
-		l:       l,
-		batch:   batch,
 	}
 
-	af, err := out.getSockOptInt(unix.SO_DOMAIN)
+	var sa unix.Sockaddr
-	if err != nil {
+	if ip.Is4() {
-		_ = out.Close()
+		sa4 := &unix.SockaddrInet4{Port: port}
-		return nil, err
+		sa4.Addr = ip.As4()
+		sa = sa4
+	} else {
+		sa6 := &unix.SockaddrInet6{Port: port}
+		sa6.Addr = ip.As16()
+		sa = sa6
+	}
+	if err = unix.Bind(fd, sa); err != nil {
+		_ = unix.Close(fd)
+		return nil, fmt.Errorf("unable to bind to socket: %w", err)
 	}
-	out.isV4 = af == unix.AF_INET
 
-	return out, nil
+	return &StdConn{sysFd: fd, isV4: ip.Is4(), l: l, batch: batch}, nil
 }
 
 func (u *StdConn) SupportsMultipleReaders() bool {
@@ -81,134 +75,111 @@ func (u *StdConn) Rebind() error {
 	return nil
 }
 
-func (u *StdConn) getSockOptInt(opt int) (int, error) {
-	if u.rawConn == nil {
-		return 0, fmt.Errorf("no UDP connection")
-	}
-	var out int
-	var opErr error
-	err := u.rawConn.Control(func(fd uintptr) {
-		out, opErr = unix.GetsockoptInt(int(fd), unix.SOL_SOCKET, opt)
-	})
-	if err != nil {
-		return 0, err
-	}
-	return out, opErr
-}
-
-func (u *StdConn) setSockOptInt(opt int, 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), unix.SOL_SOCKET, opt, n)
-	})
-	if err != nil {
-		return err
-	}
-	return opErr
-}
-
 func (u *StdConn) SetRecvBuffer(n int) error {
-	return u.setSockOptInt(unix.SO_RCVBUFFORCE, n)
+	return unix.SetsockoptInt(u.sysFd, unix.SOL_SOCKET, unix.SO_RCVBUFFORCE, n)
 }
 
 func (u *StdConn) SetSendBuffer(n int) error {
-	return u.setSockOptInt(unix.SO_SNDBUFFORCE, n)
+	return unix.SetsockoptInt(u.sysFd, unix.SOL_SOCKET, unix.SO_SNDBUFFORCE, n)
 }
 
 func (u *StdConn) SetSoMark(mark int) error {
-	return u.setSockOptInt(unix.SO_MARK, mark)
+	return unix.SetsockoptInt(u.sysFd, unix.SOL_SOCKET, unix.SO_MARK, mark)
 }
 
 func (u *StdConn) GetRecvBuffer() (int, error) {
-	return u.getSockOptInt(unix.SO_RCVBUF)
+	return unix.GetsockoptInt(u.sysFd, unix.SOL_SOCKET, unix.SO_RCVBUF)
 }
 
 func (u *StdConn) GetSendBuffer() (int, error) {
-	return u.getSockOptInt(unix.SO_SNDBUF)
+	return unix.GetsockoptInt(u.sysFd, unix.SOL_SOCKET, unix.SO_SNDBUF)
 }
 
 func (u *StdConn) GetSoMark() (int, error) {
-	return u.getSockOptInt(unix.SO_MARK)
+	return unix.GetsockoptInt(u.sysFd, unix.SOL_SOCKET, unix.SO_MARK)
 }
 
 func (u *StdConn) LocalAddr() (netip.AddrPort, error) {
-	a := u.udpConn.LocalAddr()
+	sa, err := unix.Getsockname(u.sysFd)
-
+	if err != nil {
-	switch v := a.(type) {
+		return netip.AddrPort{}, err
-	case *net.UDPAddr:
+	}
-		addr, ok := netip.AddrFromSlice(v.IP)
+	switch sa := sa.(type) {
-		if !ok {
+	case *unix.SockaddrInet4:
-			return netip.AddrPort{}, fmt.Errorf("LocalAddr returned invalid IP address: %s", v.IP)
+		return netip.AddrPortFrom(netip.AddrFrom4(sa.Addr), uint16(sa.Port)), nil
-		}
+	case *unix.SockaddrInet6:
-		return netip.AddrPortFrom(addr, uint16(v.Port)), nil
+		return netip.AddrPortFrom(netip.AddrFrom16(sa.Addr), uint16(sa.Port)), nil
-
 	default:
-		return netip.AddrPort{}, fmt.Errorf("LocalAddr returned: %#v", a)
+		return netip.AddrPort{}, fmt.Errorf("unsupported sock type: %T", sa)
 	}
 }
 
-func recvmmsg(fd uintptr, msgs []rawMessage) (int, bool, error) {
+// recvmmsg does one blocking recvmmsg (MSG_WAITFORONE), reading up to len(msgs) datagrams
-	var errno syscall.Errno
+func (u *StdConn) recvmmsg(msgs []rawMessage) (int, error) {
-	n, _, errno := unix.Syscall6(
+	r, _, errno := unix.Syscall6(
 		unix.SYS_RECVMMSG,
-		fd,
+		uintptr(u.sysFd),
 		uintptr(unsafe.Pointer(&msgs[0])),
 		uintptr(len(msgs)),
 		unix.MSG_WAITFORONE,
 		0,
 		0,
 	)
-	if errno == syscall.EAGAIN || errno == syscall.EWOULDBLOCK {
-		// No data available, block for I/O and try again.
-		return int(n), false, nil
-	}
 	if errno != 0 {
-		return int(n), true, &net.OpError{Op: "recvmmsg", Err: errno}
+		if u.closed.Load() {
-	}
+			return 0, net.ErrClosed
-	return int(n), true, nil
-}
-
-func (u *StdConn) listenOutSingle(r EncReader) error {
-	var err error
-	var n int
-	var from netip.AddrPort
-	buffer := make([]byte, MTU)
-
-	for {
-		n, from, err = u.udpConn.ReadFromUDPAddrPort(buffer)
-		if err != nil {
-			return err
 		}
-		from = netip.AddrPortFrom(from.Addr().Unmap(), from.Port())
+		return 0, &net.OpError{Op: "recvmmsg", Err: errno}
-		r(from, buffer[:n])
 	}
+	n := int(r)
+	if (n == 0 || msgs[0].Len == 0) && u.closed.Load() {
+		return 0, net.ErrClosed
+	}
+	return n, nil
 }
 
-func (u *StdConn) listenOutBatch(r EncReader) error {
+// recvmsg does one blocking recvmsg into msgs[0]
+func (u *StdConn) recvmsg(msgs []rawMessage) (int, error) {
+	r, _, errno := unix.Syscall6(
+		unix.SYS_RECVMSG,
+		uintptr(u.sysFd),
+		uintptr(unsafe.Pointer(&msgs[0].Hdr)),
+		0,
+		0,
+		0,
+		0,
+	)
+	if errno != 0 {
+		if u.closed.Load() {
+			return 0, net.ErrClosed
+		}
+		return 0, &net.OpError{Op: "recvmsg", Err: errno}
+	}
+	if r == 0 && u.closed.Load() {
+		return 0, net.ErrClosed
+	}
+	msgs[0].Len = uint32(r)
+	return 1, nil
+}
+
+func (u *StdConn) ListenOut(r EncReader) error {
 	var ip netip.Addr
-	var n int
-	var operr error
-
 	msgs, buffers, names := u.PrepareRawMessages(u.batch)
-
+	read := u.recvmmsg
-	//reader needs to capture variables from this function, since it's used as a lambda with rawConn.Read
+	if u.batch == 1 {
-	//defining it outside the loop so it gets re-used
+		read = u.recvmsg
-	reader := func(fd uintptr) (done bool) {
-		n, done, operr = recvmmsg(fd, msgs)
-		return done
 	}
 
 	for {
-		err := u.rawConn.Read(reader)
+		n, err := read(msgs)
 		if err != nil {
+			if errors.Is(err, unix.EINTR) {
+				continue // interrupted by a signal, retry the read
+			}
+			// net.ErrClosed after Close() is teardown, absorbed by the caller's
+			// closed flag like the other platforms; anything else is a real error.
 			return err
 		}
-		if operr != nil {
-			return operr
-		}
 
 		for i := 0; i < n; i++ {
 			// Its ok to skip the ok check here, the slicing is the only error that can occur and it will panic
@@ -222,26 +193,68 @@ func (u *StdConn) listenOutBatch(r EncReader) error {
 	}
 }
 
-func (u *StdConn) ListenOut(r EncReader) error {
+func (u *StdConn) WriteTo(b []byte, ip netip.AddrPort) error {
-	if u.batch == 1 {
+	if u.isV4 {
-		return u.listenOutSingle(r)
+		return u.writeTo4(b, ip)
-	} else {
+	}
-		return u.listenOutBatch(r)
+	return u.writeTo6(b, ip)
+}
+
+func (u *StdConn) writeTo6(b []byte, ip netip.AddrPort) error {
+	var rsa unix.RawSockaddrInet6
+	rsa.Family = unix.AF_INET6
+	rsa.Addr = ip.Addr().As16()
+	binary.BigEndian.PutUint16((*[2]byte)(unsafe.Pointer(&rsa.Port))[:], ip.Port())
+
+	for {
+		_, _, err := unix.Syscall6(
+			unix.SYS_SENDTO,
+			uintptr(u.sysFd),
+			uintptr(unsafe.Pointer(&b[0])),
+			uintptr(len(b)),
+			uintptr(0),
+			uintptr(unsafe.Pointer(&rsa)),
+			uintptr(unix.SizeofSockaddrInet6),
+		)
+		if err != 0 {
+			return &net.OpError{Op: "sendto", Err: err}
+		}
+		return nil
 	}
 }
 
-func (u *StdConn) WriteTo(b []byte, ip netip.AddrPort) error {
+func (u *StdConn) writeTo4(b []byte, ip netip.AddrPort) error {
-	_, err := u.udpConn.WriteToUDPAddrPort(b, ip)
+	if !ip.Addr().Is4() {
-	return err
+		return ErrInvalidIPv6RemoteForSocket
+	}
+
+	var rsa unix.RawSockaddrInet4
+	rsa.Family = unix.AF_INET
+	rsa.Addr = ip.Addr().As4()
+	binary.BigEndian.PutUint16((*[2]byte)(unsafe.Pointer(&rsa.Port))[:], ip.Port())
+
+	for {
+		_, _, err := unix.Syscall6(
+			unix.SYS_SENDTO,
+			uintptr(u.sysFd),
+			uintptr(unsafe.Pointer(&b[0])),
+			uintptr(len(b)),
+			uintptr(0),
+			uintptr(unsafe.Pointer(&rsa)),
+			uintptr(unix.SizeofSockaddrInet4),
+		)
+		if err != 0 {
+			return &net.OpError{Op: "sendto", Err: err}
+		}
+		return nil
+	}
 }
 
 func (u *StdConn) ReloadConfig(c *config.C) {
 	b := c.GetInt("listen.read_buffer", 0)
 	if b > 0 {
-		err := u.SetRecvBuffer(b)
+		if err := u.SetRecvBuffer(b); err == nil {
-		if err == nil {
+			if s, err := u.GetRecvBuffer(); err == nil {
-			s, err := u.GetRecvBuffer()
-			if err == nil {
 				u.l.Info("listen.read_buffer was set", "size", s)
 			} else {
 				u.l.Warn("Failed to get listen.read_buffer", "error", err)
@@ -253,10 +266,8 @@ func (u *StdConn) ReloadConfig(c *config.C) {
 
 	b = c.GetInt("listen.write_buffer", 0)
 	if b > 0 {
-		err := u.SetSendBuffer(b)
+		if err := u.SetSendBuffer(b); err == nil {
-		if err == nil {
+			if s, err := u.GetSendBuffer(); err == nil {
-			s, err := u.GetSendBuffer()
-			if err == nil {
 				u.l.Info("listen.write_buffer was set", "size", s)
 			} else {
 				u.l.Warn("Failed to get listen.write_buffer", "error", err)
@@ -269,10 +280,8 @@ func (u *StdConn) ReloadConfig(c *config.C) {
 	b = c.GetInt("listen.so_mark", 0)
 	s, err := u.GetSoMark()
 	if b > 0 || (err == nil && s != 0) {
-		err := u.SetSoMark(b)
+		if err := u.SetSoMark(b); err == nil {
-		if err == nil {
+			if s, err := u.GetSoMark(); err == nil {
-			s, err := u.GetSoMark()
-			if err == nil {
 				u.l.Info("listen.so_mark was set", "mark", s)
 			} else {
 				u.l.Warn("Failed to get listen.so_mark", "error", err)
@@ -285,28 +294,20 @@ func (u *StdConn) ReloadConfig(c *config.C) {
 
 func (u *StdConn) getMemInfo(meminfo *[unix.SK_MEMINFO_VARS]uint32) error {
 	var vallen uint32 = 4 * unix.SK_MEMINFO_VARS
-
+	_, _, err := unix.Syscall6(unix.SYS_GETSOCKOPT, uintptr(u.sysFd), uintptr(unix.SOL_SOCKET), uintptr(unix.SO_MEMINFO), uintptr(unsafe.Pointer(meminfo)), uintptr(unsafe.Pointer(&vallen)), 0)
-	if u.rawConn == nil {
+	if err != 0 {
-		return fmt.Errorf("no UDP connection")
-	}
-	var opErr error
-	err := u.rawConn.Control(func(fd uintptr) {
-		_, _, syserr := unix.Syscall6(unix.SYS_GETSOCKOPT, fd, uintptr(unix.SOL_SOCKET), uintptr(unix.SO_MEMINFO), uintptr(unsafe.Pointer(meminfo)), uintptr(unsafe.Pointer(&vallen)), 0)
-		if syserr != 0 {
-			opErr = syserr
-		}
-	})
-	if err != nil {
 		return err
 	}
-	return opErr
+	return nil
 }
 
 func (u *StdConn) Close() error {
-	if u.udpConn != nil {
+	u.closed.Store(true)
-		return u.udpConn.Close()
+	// Wake the reader parked in recvmmsg/recvmsg. shutdown(2) on an unconnected socket
-	}
+	// returns ENOTCONN but still wakes it, so ignore the error.
-	return nil
+	// The reader then sees closed and stops touching the fd, making the Close below safe.
+	_ = unix.Shutdown(u.sysFd, unix.SHUT_RDWR)
+	return unix.Close(u.sysFd)
 }
 
 func NewUDPStatsEmitter(udpConns []Conn) func() {

udp/udp_linux_test.go

@@ -0,0 +1,179 @@
+//go:build linux && !android && !e2e_testing
+
+package udp
+
+import (
+	"errors"
+	"fmt"
+	"log/slog"
+	"net"
+	"net/netip"
+	"os"
+	"runtime"
+	"sync/atomic"
+	"testing"
+	"time"
+
+	"golang.org/x/sys/unix"
+)
+
+func testLogger() *slog.Logger {
+	return slog.New(slog.NewTextHandler(os.Stderr, &slog.HandlerOptions{Level: slog.LevelError}))
+}
+
+// TestShutdownWakesAfterRx_Mechanism exercises the kernel quirk our teardown
+// relies on: once a socket has received a packet, shutdown(2) wakes a blocked
+// recvmmsg with n>=1/Len==0 (not n==0). recvmmsg must turn that into net.ErrClosed
+// once Close set closed, so a parked reader exits instead of spinning.
+func TestShutdownWakesAfterRx_Mechanism(t *testing.T) {
+	c, err := NewListener(testLogger(), netip.MustParseAddr("127.0.0.1"), 0, true, 64)
+	if err != nil {
+		t.Fatalf("NewListener: %v", err)
+	}
+	sc := c.(*StdConn)
+	addr, err := sc.LocalAddr()
+	if err != nil {
+		t.Fatalf("LocalAddr: %v", err)
+	}
+	msgs, _, _ := sc.PrepareRawMessages(sc.batch)
+
+	// Receive a real packet so the socket has carried data.
+	send, err := net.Dial("udp", addr.String())
+	if err != nil {
+		t.Fatalf("dial: %v", err)
+	}
+	if _, err := send.Write([]byte("hello")); err != nil {
+		t.Fatalf("write: %v", err)
+	}
+	time.Sleep(50 * time.Millisecond)
+	n, err := sc.recvmmsg(msgs)
+	t.Logf("drain of real packet: n=%d err=%v msgs[0].Len=%d", n, err, msgs[0].Len)
+	_ = send.Close()
+
+	// Block a reader on the now-empty queue, then tear down as Close() does.
+	// recvmmsg must return net.ErrClosed (not hang, not spin) even post-rx.
+	done := make(chan error, 1)
+	go func() {
+		_, err := sc.recvmmsg(msgs)
+		done <- err
+	}()
+	time.Sleep(150 * time.Millisecond) // let it park in recvmmsg
+
+	sc.closed.Store(true)
+	if serr := unix.Shutdown(sc.sysFd, unix.SHUT_RDWR); serr != nil {
+		t.Logf("shutdown returned %v (expected ENOTCONN on unconnected UDP)", serr)
+	}
+
+	select {
+	case err := <-done:
+		if !errors.Is(err, net.ErrClosed) {
+			t.Errorf("recvmmsg after post-rx shutdown returned %v, want net.ErrClosed", err)
+		}
+	case <-time.After(2 * time.Second):
+		t.Fatalf("HANG: recvmmsg did not return after shutdown following a received packet")
+	}
+	_ = unix.Close(sc.sysFd)
+}
+
+// TestListenOutTeardown_TrafficPatterns reproduces the field report: a blocking
+// reader must tear down cleanly on Close() regardless of what the socket has
+// carried. The three cases the report called out:
+//
+//	no traffic ever      -> works (shutdown wakes recvmmsg with n==0)
+//	ping once, then idle -> historically HUNG: once the socket has received a
+//	                        packet, shutdown(2) wakes recvmmsg with n>=1/Len==0,
+//	                        which an n==0-only teardown check misses
+//	continuous traffic   -> works (a real packet is always arriving)
+//
+// All three must return within the deadline; a hang dumps goroutines so the
+// stuck reader is visible.
+func TestListenOutTeardown_TrafficPatterns(t *testing.T) {
+	cases := []struct {
+		name    string
+		traffic func(send net.Conn, stop <-chan struct{})
+	}{
+		{"no_traffic_ever", func(net.Conn, <-chan struct{}) {}},
+		{"ping_once_then_idle", func(send net.Conn, _ <-chan struct{}) {
+			_, _ = send.Write([]byte("hello"))
+		}},
+		{"continuous", func(send net.Conn, stop <-chan struct{}) {
+			for {
+				select {
+				case <-stop:
+					return
+				default:
+					_, _ = send.Write([]byte("hello"))
+					time.Sleep(2 * time.Millisecond)
+				}
+			}
+		}},
+	}
+
+	// batch 1 exercises the recvmsg path, batch 64 the recvmmsg path; both must
+	// tear down cleanly.
+	for _, batch := range []int{1, 64} {
+		for _, tc := range cases {
+			t.Run(fmt.Sprintf("batch%d/%s", batch, tc.name), func(t *testing.T) {
+				runTeardownCase(t, batch, tc.name, tc.traffic)
+			})
+		}
+	}
+}
+
+func runTeardownCase(t *testing.T, batch int, name string, traffic func(send net.Conn, stop <-chan struct{})) {
+	c, err := NewListener(testLogger(), netip.MustParseAddr("127.0.0.1"), 0, true, batch)
+	if err != nil {
+		t.Fatalf("NewListener: %v", err)
+	}
+	sc := c.(*StdConn)
+	addr, err := sc.LocalAddr()
+	if err != nil {
+		t.Fatalf("LocalAddr: %v", err)
+	}
+
+	var received atomic.Int64
+	loopDone := make(chan error, 1)
+	go func() {
+		loopDone <- sc.ListenOut(func(netip.AddrPort, []byte) {
+			received.Add(1)
+		})
+	}()
+
+	send, err := net.Dial("udp", addr.String())
+	if err != nil {
+		t.Fatalf("dial: %v", err)
+	}
+	defer send.Close()
+
+	stop := make(chan struct{})
+	trafficDone := make(chan struct{})
+	go func() {
+		traffic(send, stop)
+		close(trafficDone)
+	}()
+
+	// Let the pattern run and, for the idle case, the reader park again on an
+	// empty queue with the socket already having received a packet.
+	time.Sleep(500 * time.Millisecond)
+
+	start := time.Now()
+	if err := sc.Close(); err != nil {
+		t.Fatalf("Close: %v", err)
+	}
+	close(stop)
+
+	select {
+	case err := <-loopDone:
+		// Clean teardown surfaces as net.ErrClosed (propagated like the other
+		// platforms); the caller absorbs it via its closed flag.
+		if err != nil && !errors.Is(err, net.ErrClosed) {
+			t.Fatalf("%s: ListenOut returned unexpected error on teardown: %v", name, err)
+		}
+		t.Logf("%s: closed in %v (received %d packets)", name, time.Since(start), received.Load())
+	case <-time.After(3 * time.Second):
+		buf := make([]byte, 1<<20)
+		n := runtime.Stack(buf, true)
+		t.Fatalf("%s: HANG, ListenOut did not return within 3s of Close\n%s", name, buf[:n])
+	}
+	<-trafficDone
+}

wake_clock_darwin.go

@@ -0,0 +1,29 @@
+//go:build darwin
+
+package nebula
+
+import (
+	"time"
+
+	"golang.org/x/sys/unix"
+)
+
+// suspendClockDelta returns CLOCK_MONOTONIC - CLOCK_UPTIME_RAW. On macOS CLOCK_MONOTONIC keeps counting across
+// system sleep while CLOCK_UPTIME_RAW (mach_absolute_time) pauses, so the difference grows by time spent asleep.
+//
+// The pausing clock is read first so scheduling jitter between the two reads biases the delta positive; the
+// wakeDetector clamps out the noise.
+//
+// Caveat: on Apple Silicon the hardware timebase keeps ticking through sleep, which can make both clocks advance
+// and the spread stay flat, leaving this detector blind. That fails safe (no clears, behavior as before); IOKit
+// power notifications are the follow-up for full coverage on those machines.
+func suspendClockDelta() (time.Duration, bool) {
+	var uptime, mono unix.Timespec
+	if err := unix.ClockGettime(unix.CLOCK_UPTIME_RAW, &uptime); err != nil {
+		return 0, false
+	}
+	if err := unix.ClockGettime(unix.CLOCK_MONOTONIC, &mono); err != nil {
+		return 0, false
+	}
+	return time.Duration(mono.Nano() - uptime.Nano()), true
+}

wake_clock_generic.go

@@ -0,0 +1,11 @@
+//go:build !linux && !darwin && !windows
+
+package nebula
+
+import "time"
+
+// suspendClockDelta reports that this platform has no usable clock pair for detecting system sleep; the wake
+// detector stays dormant and dead tunnels are left to the normal traffic checks.
+func suspendClockDelta() (time.Duration, bool) {
+	return 0, false
+}

wake_clock_linux.go

@@ -0,0 +1,26 @@
+//go:build linux
+
+package nebula
+
+import (
+	"time"
+
+	"golang.org/x/sys/unix"
+)
+
+// suspendClockDelta returns CLOCK_BOOTTIME - CLOCK_MONOTONIC. CLOCK_MONOTONIC pauses while the system is suspended
+// and CLOCK_BOOTTIME does not, so the difference only ever grows, and only by time spent suspended. Both reads are
+// vDSO calls, cheap enough for a hot ticker.
+//
+// The pausing clock is read first so scheduling jitter between the two reads biases the delta positive; the
+// wakeDetector clamps out the noise.
+func suspendClockDelta() (time.Duration, bool) {
+	var mono, boot unix.Timespec
+	if err := unix.ClockGettime(unix.CLOCK_MONOTONIC, &mono); err != nil {
+		return 0, false
+	}
+	if err := unix.ClockGettime(unix.CLOCK_BOOTTIME, &boot); err != nil {
+		return 0, false
+	}
+	return time.Duration(boot.Nano() - mono.Nano()), true
+}

wake_clock_windows.go

@@ -0,0 +1,41 @@
+//go:build windows
+
+package nebula
+
+import (
+	"sync"
+	"time"
+	"unsafe"
+
+	"golang.org/x/sys/windows"
+)
+
+var (
+	procQueryInterruptTime         = windows.NewLazySystemDLL("kernelbase.dll").NewProc("QueryInterruptTime")
+	procQueryUnbiasedInterruptTime = windows.NewLazySystemDLL("kernel32.dll").NewProc("QueryUnbiasedInterruptTime")
+
+	// QueryInterruptTime needs Windows 10; probe once and stay dormant on anything older.
+	wakeClockAvailable = sync.OnceValue(func() bool {
+		return procQueryInterruptTime.Find() == nil && procQueryUnbiasedInterruptTime.Find() == nil
+	})
+)
+
+// suspendClockDelta returns interrupt time minus unbiased interrupt time, both in 100ns units. The unbiased count
+// excludes time the system spends suspended while the biased one includes it, so the difference grows by exactly
+// the time spent asleep.
+//
+// The pausing (unbiased) clock is read first so scheduling jitter between the two reads biases the delta positive;
+// the wakeDetector clamps out the noise.
+func suspendClockDelta() (time.Duration, bool) {
+	if !wakeClockAvailable() {
+		return 0, false
+	}
+
+	var unbiased, biased uint64
+	if r1, _, _ := procQueryUnbiasedInterruptTime.Call(uintptr(unsafe.Pointer(&unbiased))); r1 == 0 {
+		return 0, false
+	}
+	// Returns void, cannot fail once resolved.
+	_, _, _ = procQueryInterruptTime.Call(uintptr(unsafe.Pointer(&biased)))
+	return time.Duration(int64(biased-unbiased)) * 100, true
+}

wake_detector.go

@@ -0,0 +1,47 @@
+package nebula
+
+import "time"
+
+// wakeDetector notices when the machine has returned from system sleep and measures how long it was suspended.
+//
+// It samples the spread between two kernel clocks: one that pauses across a suspend and one that keeps counting
+// (suspendClockDelta, per platform). While the machine is awake the spread is constant no matter how starved,
+// stopped, or stepped this process is — SIGSTOP, debugger pauses, scheduler starvation, and NTP adjustments move
+// both clocks together or neither, so none of them can fake a wake. A true suspend is the only thing that grows
+// the spread, and it grows by exactly the time spent suspended.
+//
+// Sample is intended to piggyback on a ticker the caller already runs; it costs two clock reads. It is not safe
+// for concurrent use.
+type wakeDetector struct {
+	// read returns the current spread between the two clocks, false if this platform can't provide one.
+	read   func() (time.Duration, bool)
+	last   time.Duration
+	primed bool
+}
+
+func newWakeDetector() *wakeDetector {
+	return &wakeDetector{read: suspendClockDelta}
+}
+
+// Sample returns how long the machine was suspended since the previous call, 0 if it wasn't, and false if the
+// platform has no way to tell. The first call primes the baseline and always reports 0.
+func (w *wakeDetector) Sample() (time.Duration, bool) {
+	delta, ok := w.read()
+	if !ok {
+		return 0, false
+	}
+
+	if !w.primed {
+		w.primed = true
+		w.last = delta
+		return 0, true
+	}
+
+	slept := delta - w.last
+	w.last = delta
+	if slept < 0 {
+		// The clock pair is read non-atomically so tiny negative jitter is possible; it is never a wake.
+		slept = 0
+	}
+	return slept, true
+}

wake_detector_test.go

@@ -0,0 +1,67 @@
+package nebula
+
+import (
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/assert"
+)
+
+func TestWakeDetector(t *testing.T) {
+	delta := time.Duration(0)
+	ok := true
+	w := &wakeDetector{read: func() (time.Duration, bool) { return delta, ok }}
+
+	// The first sample primes the baseline and never reports a wake, even with a pre-existing spread
+	delta = 3 * time.Hour
+	slept, sok := w.Sample()
+	assert.True(t, sok)
+	assert.Equal(t, time.Duration(0), slept)
+
+	// A stable spread means the machine never slept
+	slept, sok = w.Sample()
+	assert.True(t, sok)
+	assert.Equal(t, time.Duration(0), slept)
+
+	// The spread grows by exactly the time spent suspended
+	delta += 42 * time.Second
+	slept, sok = w.Sample()
+	assert.True(t, sok)
+	assert.Equal(t, 42*time.Second, slept)
+
+	// A wake is reported once, then the baseline moves with it
+	slept, sok = w.Sample()
+	assert.True(t, sok)
+	assert.Equal(t, time.Duration(0), slept)
+
+	// Negative jitter from the non-atomic clock pair reads clamps to zero
+	delta -= time.Microsecond
+	slept, sok = w.Sample()
+	assert.True(t, sok)
+	assert.Equal(t, time.Duration(0), slept)
+
+	// Consecutive suspends both report; the clamped jitter moved the baseline so it is not double-counted
+	delta += time.Minute
+	slept, _ = w.Sample()
+	assert.Equal(t, time.Minute, slept)
+	delta += time.Hour
+	slept, _ = w.Sample()
+	assert.Equal(t, time.Hour, slept)
+
+	// An unsupported platform read reports not-ok
+	ok = false
+	_, sok = w.Sample()
+	assert.False(t, sok)
+}
+
+// TestWakeDetectorPlatformClock smoke tests the real clock pair: two samples close together must not report a
+// wake on a machine that isn't suspending mid-test.
+func TestWakeDetectorPlatformClock(t *testing.T) {
+	w := newWakeDetector()
+	if _, ok := w.Sample(); !ok {
+		t.Skip("platform has no suspend clock pair")
+	}
+	slept, ok := w.Sample()
+	assert.True(t, ok)
+	assert.Less(t, slept, time.Second)
+}