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/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

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])
+	}
+}

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
+}

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/handshakes_test.go

@@ -725,6 +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

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

go.mod

@@ -24,12 +24,12 @@ 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 v1.0.1

go.sum

@@ -162,8 +162,8 @@ 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=
@@ -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=

handshake_manager.go

@@ -529,7 +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 {
@@ -1077,7 +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

@@ -287,7 +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
 }

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
 	}
@@ -151,7 +155,7 @@ func (f *Interface) readOutsidePackets(via ViaSender, out []byte, packet []byte,
 			// No-op, useful for the Roaming and connectionManager side-effects above
 		case header.TestRequest:
 			//recycle the input packet ciphertext as our output buffer
-			f.send(header.Test, header.TestReply, ci, hostinfo, out, nb, packet)
+			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
@@ -170,27 +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
@@ -214,7 +199,6 @@ func (f *Interface) handleOutsideRelayPacket(hostinfo *HostInfo, via ViaSender,
 		via = ViaSender{
 			UdpAddr:   via.UdpAddr,
 			relayHI:   hostinfo,
-			remoteIdx: relay.RemoteIndex,
 			relay:     relay,
 			IsRelayed: true,
 		}
@@ -235,9 +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
@@ -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 {

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)
+}