Merge remote-tracking branch 'origin/master' into multiport

We need to compare just the IPs here and not the IP+Port

This is a regression with the merge of v1.9.4 and the change to
netip.AddrPort

- https://github.com/slackhq/nebula/pull/768/commits/dabce8a1b416e1c0a70ba1254597a31b0d555c79

.github/workflows/smoke.yml

@@ -60,4 +60,12 @@ jobs:
       working-directory: ./.github/workflows/smoke
       run: NAME="smoke-p256" ./smoke.sh
 
+    - name: setup docker image for multiport
+      working-directory: ./.github/workflows/smoke
+      run: NAME="smoke-multiport" MULTIPORT_TX=true MULTIPORT_RX=true MULTIPORT_HANDSHAKE=true ./build.sh
+
+    - name: run smoke
+      working-directory: ./.github/workflows/smoke
+      run: NAME="smoke-multiport" ./smoke.sh
+
     timeout-minutes: 10

.github/workflows/smoke/genconfig.sh

@@ -48,6 +48,10 @@ listen:
 
 tun:
   dev: ${TUN_DEV:-tun0}
+  multiport:
+    tx_enabled: ${MULTIPORT_TX:-false}
+    rx_enabled: ${MULTIPORT_RX:-false}
+    tx_handshake: ${MULTIPORT_HANDSHAKE:-false}
 
 firewall:
   inbound_action: reject

CHANGELOG.md

@@ -7,6 +7,88 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [1.11.0] - 2026-07-23
+
+See the [v1.11.0](https://github.com/slackhq/nebula/milestone/25?closed=1) milestone for a complete list of changes.
+
+### Breaking
+
+- Logging has switched from logrus to Go's structured `slog`. Log output changes: levels are upper case
+  (`level=INFO`), trace prints as `level=DEBUG-4`, timestamps are always RFC3339Nano and `logging.timestamp_format`
+  is ignored, and some messages were reworded. Review any log parsing before upgrading. This is also an API break
+  for embedders, as constructors now take a `*slog.Logger`. (#1672, #1734, #1621)
+- `firewall.inbound_action` and `firewall.outbound_action` (used to set reject vs. drop policy) were each being
+  applied to the opposite direction, that is now corrected. This only affects how blocked packets are answered, not
+  which packets the firewall allows or denies. If you set either of these you are getting the behavior of the other
+  one today and likely want to swap them before upgrading. (#1798)
+- On Windows, Nebula now installs WFP PERMIT filters for the nebula adapter and the listener port by default. WFP
+  sits below Windows Defender Firewall, so any WDF inbound rules you rely on for either will no longer apply. Set
+  `tun.windows_bypass_wdf` and `listen.windows_bypass_wdf` to false to leave WDF in charge. (#1710)
+- On Windows, the nebula device is now set to the `private` network category instead of whatever Windows decided,
+  which is usually `Public`. This makes the host firewall less restrictive on the overlay. Set
+  `tun.network_category` to `unset` to keep the old behavior. (#1710)
+- Reject packets for non-TCP now use ICMP code 13, communication administratively prohibited, instead of code 3,
+  port unreachable. Anything keying off the old code needs updating. (#1766, #1768)
+- The SSH debug server's profiling commands are now confined to `sshd.sandbox_dir`, which defaults to
+  `$TMP/nebula-debug`. Relative paths resolve inside it and absolute paths outside it are rejected, so anything
+  scripting `start-cpu-profile`, `save-heap-profile`, or `save-mutex-profile` with a path elsewhere needs the
+  directory set. The directory is not created for you. (#1622)
+
+### Added
+
+- Sign the Windows release binaries. (#1718)
+- Generate IPv6 reject packets, matching the existing IPv4 behavior. (#1766, #1767, #1768)
+- Accept `-` in `nebula-cert` to read from stdin or write to stdout. (#1714)
+- Search for both `config.yml` and `config.yaml` in service and command line modes. (#1717)
+- Add version labels to the Docker/OCI images. (#1772)
+- Rebind the listener and re-query lighthouses on macOS when the underlay network changes, so devices moving
+  between wifi and wired or between networks recover without waiting for dead tunnel detection. Controlled by
+  `listen.rebind_on_network_change` (default `true`, not reloadable). (#1816)
+
+### Changed
+
+- Reload the firewall when the unsafe networks in the certificate change. (#1719)
+- Reconfigure, start, and stop the stats listener on a config reload instead of requiring a restart. (#1670)
+- Update a static host's addresses when they change on reload. (#1713)
+- Don't require a port on ICMP firewall rules. (#1609)
+- Connection track ICMP traffic. (#1602)
+- Return `NODATA` instead of `NXDOMAIN` from the DNS server for a name that exists but has no record of the
+  requested type, so clients that query `AAAA` first (busybox/Alpine) fall through to `A`. (#1668)
+- Record the local host's details in the DNS server. (#1716)
+- Install Windows unsafe routes as link routes. (#1709)
+- Reduce relay handshake log spam, and only log a handshake send error at error level when the remote list
+  changes. (#1733, #1765, #1810)
+- Start, stop, and reload subsystems (DNS, stats, conntrack, ssh, punchy) cleanly without leaking goroutines. (#1640, #1654, #1661, #1667, #1669, #1708, #1806, #1815)
+- `Control` is now safe to stop and wait on from any lifecycle state, and a new `Control.Wait` blocks until nebula
+  has fully stopped and returns the first fatal reader error. Failed starts release the udp sockets and tun fd
+  instead of leaking them. (#1794)
+- Trigger an immediate lighthouse update when reconnecting to or adding a lighthouse instead of waiting for the next update tick. (#1645)
+- Bring the Darwin and OpenBSD tun implementations in line with the other BSDs. (#1703)
+- Update to build against go v1.26. (#1818)
+- Various dependency updates. (#1586, #1587, #1604, #1617, #1618, #1627, #1628, #1629, #1652, #1664, #1665, #1697, #1721, #1732, #1742, #1743, #1750, #1763, #1771, #1782, #1800, #1807)
+
+### Fixed
+
+- Fix a data race on a host's remote address that could send packets to the wrong address during a roam. (#1773)
+- Fix tunnels that could permanently escape connection manager monitoring. (#1752)
+- Fix a crash when reloading the SSH server's trusted keys. (#1787)
+- Fix hostmap corruption when a host has multiple overlay addresses. Each address now gets its own list instead of
+  a single shared chain, which also fixes two latent bugs on the add and makePrimary paths. (#1788, #1790)
+- Apply `remote_allow_list` IPv4 rules to 4-in-6 mapped addresses. (#1786)
+- Don't panic in the DNS server on a short or empty query name. (#1635)
+- Advance the replay window on relayed packets so a relay drops replayed frames instead of re-forwarding them. (#1751)
+- Fix a race in relay state handling. (#1753)
+- Lock replay window updates so concurrent readers can't corrupt it. (#1802)
+- Reject malformed handshakes more reliably, including invalid ed25519 key lengths. (#1601, #1756)
+- Properly handle `closetunnel` packets. (#1638)
+- Fix an IPv6 extension-header length overflow that could make the firewall parse the wrong protocol and ports. (#1789)
+- Fix relay re-establishment when a handshake arrives over a relay entry that a one-sided teardown left
+  `Disestablished`, which silently dropped every send until dead tunnel detection forced a re-handshake. (#1805)
+- Don't build new relay state on a tunnel that was just discarded. (#1796)
+- Don't delete the wrong pending hostinfo in the handshake manager. (#1811)
+- Don't call the packet reader after a UDP error on Darwin. (#1755)
+- Open the FreeBSD tun device non blocking. (#1666)
+
 ## [1.10.3] - 2026-02-06
 
 ### Security

Makefile

@@ -268,6 +268,10 @@ smoke-relay-docker: bin-docker
 	cd .github/workflows/smoke/ && ./build-relay.sh
 	cd .github/workflows/smoke/ && ./smoke-relay.sh
 
+smoke-multiport-docker: bin-docker
+	cd .github/workflows/smoke/ && NAME="smoke-multiport" MULTIPORT_TX=true MULTIPORT_RX=true MULTIPORT_HANDSHAKE=true ./build.sh
+	cd .github/workflows/smoke/ && NAME="smoke-multiport" ./smoke.sh
+
 smoke-docker-ipv6: export SMOKE_OVERLAY_IPV6 = 1
 smoke-docker-ipv6: smoke-docker
 

config/multiport.go

@@ -0,0 +1,10 @@
+package config
+
+type MultiPortConfig struct {
+	Tx               bool
+	Rx               bool
+	TxBasePort       uint16
+	TxPorts          int
+	TxHandshake      bool
+	TxHandshakeDelay int64
+}

connection_state_test.go

@@ -8,6 +8,7 @@ import (
 	"github.com/flynn/noise"
 	"github.com/slackhq/nebula/cert"
 	ct "github.com/slackhq/nebula/cert_test"
+	"github.com/slackhq/nebula/config"
 	"github.com/slackhq/nebula/handshake"
 	"github.com/slackhq/nebula/header"
 	"github.com/stretchr/testify/assert"
@@ -55,6 +56,7 @@ func runTestHandshake(t *testing.T) (initR, respR *handshake.Result) {
 		cert.Version2, initCreds, verifier,
 		func() (uint32, error) { return 1000, nil },
 		true, header.HandshakeIXPSK0,
+		config.MultiPortConfig{},
 	)
 	require.NoError(t, err)
 
@@ -62,6 +64,7 @@ func runTestHandshake(t *testing.T) (initR, respR *handshake.Result) {
 		cert.Version2, respCreds, verifier,
 		func() (uint32, error) { return 2000, nil },
 		false, header.HandshakeIXPSK0,
+		config.MultiPortConfig{},
 	)
 	require.NoError(t, err)
 

examples/config.yml

@@ -328,6 +328,47 @@ tun:
   # SO_RCVBUFFORCE is used to avoid having to raise the system wide max
   #use_system_route_table_buffer_size: 0
 
+  # EXPERIMENTAL: This option may change or disappear in the future.
+  # Multiport spreads outgoing UDP packets across multiple UDP send ports,
+  # which allows nebula to work around any issues on the underlay network.
+  # Some example issues this could work around:
+  #  - UDP rate limits on a per flow basis.
+  #  - Partial underlay network failure in which some flows work and some don't
+  # Agreement is done during the handshake to decide if multiport mode will
+  # be used for a given tunnel (one side must have tx_enabled set, the other
+  # side must have rx_enabled set)
+  #
+  # NOTE: you cannot use multiport on a host if you are relying on UDP hole
+  # punching to get through a NAT or firewall.
+  #
+  # NOTE: Linux only (uses raw sockets to send). Also currently only works
+  # with IPv4 underlay network remotes.
+  #
+  # The default values are listed below:
+  #multiport:
+    # This host support sending via multiple UDP ports.
+    #tx_enabled: false
+    #
+    # This host supports receiving packets sent from multiple UDP ports.
+    #rx_enabled: false
+    #
+    # How many UDP ports to use when sending. The lowest source port will be
+    # listen.port and go up to (but not including) listen.port + tx_ports.
+    #tx_ports: 100
+    #
+    # NOTE: All of your hosts must be running a version of Nebula that supports
+    # multiport if you want to enable this feature. Older versions of Nebula
+    # will be confused by these multiport handshakes.
+    #
+    # If handshakes are not getting a response, attempt to transmit handshakes
+    # using random UDP source ports (to get around partial underlay network
+    # failures).
+    #tx_handshake: false
+    #
+    # How many unresponded handshakes we should send before we attempt to
+    # send multiport handshakes.
+    #tx_handshake_delay: 2
+
 # Configure logging level
 logging:
   # trace, debug, info, warn, or error. Default is info and is reloadable.

firewall/packet.go

@@ -3,6 +3,7 @@ package firewall
 import (
 	"encoding/json"
 	"fmt"
+	mathrand "math/rand"
 	"net/netip"
 )
 
@@ -65,3 +66,30 @@ func (fp Packet) MarshalJSON() ([]byte, error) {
 		"Fragment":   fp.Fragment,
 	})
 }
+
+// UDPSendPort calculates the UDP port to send from when using multiport mode.
+// The result will be from [0, numBuckets)
+func (fp Packet) UDPSendPort(numBuckets int) uint16 {
+	if numBuckets <= 1 {
+		return 0
+	}
+
+	// If there is no port (like an ICMP packet), pick a random UDP send port
+	if fp.LocalPort == 0 {
+		return uint16(mathrand.Intn(numBuckets))
+	}
+
+	// A decent enough 32bit hash function
+	// Prospecting for Hash Functions
+	// - https://nullprogram.com/blog/2018/07/31/
+	// - https://github.com/skeeto/hash-prospector
+	//   [16 21f0aaad 15 d35a2d97 15] = 0.10760229515479501
+	x := (uint32(fp.LocalPort) << 16) | uint32(fp.RemotePort)
+	x ^= x >> 16
+	x *= 0x21f0aaad
+	x ^= x >> 15
+	x *= 0xd35a2d97
+	x ^= x >> 15
+
+	return uint16(x) % uint16(numBuckets)
+}

handshake/handshake.proto

@@ -24,6 +24,14 @@ message NebulaHandshakeDetails {
   uint64 Cookie = 4 [deprecated = true];
   uint64 Time = 5;
   uint32 CertVersion = 8;
-  // reserved for WIP multiport
-  reserved 6, 7;
+
+  MultiPortDetails InitiatorMultiPort = 6;
+  MultiPortDetails ResponderMultiPort = 7;
+}
+
+message MultiPortDetails {
+  bool RxSupported = 1;
+  bool TxSupported = 2;
+  uint32 BasePort = 3;
+  uint32 TotalPorts = 4;
 }

handshake/helpers_test.go

@@ -8,6 +8,7 @@ import (
 	"github.com/flynn/noise"
 	"github.com/slackhq/nebula/cert"
 	ct "github.com/slackhq/nebula/cert_test"
+	"github.com/slackhq/nebula/config"
 	"github.com/slackhq/nebula/header"
 	"github.com/stretchr/testify/require"
 )
@@ -71,6 +72,7 @@ func newTestMachine(
 		cs.version, cs.getCredential,
 		verifier, func() (uint32, error) { return localIndex, nil },
 		initiator, header.HandshakeIXPSK0,
+		config.MultiPortConfig{},
 	)
 	require.NoError(t, err)
 	return m

handshake/machine.go

@@ -3,11 +3,13 @@ package handshake
 import (
 	"bytes"
 	"fmt"
+	"math"
 	"slices"
 	"time"
 
 	"github.com/flynn/noise"
 	"github.com/slackhq/nebula/cert"
+	"github.com/slackhq/nebula/config"
 	"github.com/slackhq/nebula/header"
 )
 
@@ -39,6 +41,10 @@ type Result struct {
 	HandshakeTime uint64
 	MessageIndex  uint64 // number of messages exchanged during the handshake
 	Initiator     bool
+
+	MultiportRx       bool
+	MultiportTx       bool
+	MultiportBasePort uint16
 }
 
 // Machine drives a Noise handshake through N messages. It handles Noise
@@ -67,6 +73,8 @@ type Machine struct {
 	remoteCertSet  bool
 	payloadSet     bool
 	failed         bool
+
+	multiport config.MultiPortConfig
 }
 
 // NewMachine creates a handshake state machine. The subtype determines both
@@ -80,6 +88,7 @@ func NewMachine(
 	allocIndex IndexAllocator,
 	initiator bool,
 	subtype header.MessageSubType,
+	multiport config.MultiPortConfig,
 ) (*Machine, error) {
 	info, err := subtypeInfoFor(subtype)
 	if err != nil {
@@ -108,6 +117,8 @@ func NewMachine(
 			Initiator: initiator,
 			Cipher:    cred.cipherSuite,
 		},
+
+		multiport: multiport,
 	}, nil
 }
 
@@ -298,7 +309,7 @@ func (m *Machine) processPayload(msg []byte, flags msgFlags) error {
 	}
 
 	// Assert the payload contains exactly what we expect
-	hasPayloadData := payload.InitiatorIndex != 0 || payload.ResponderIndex != 0 || payload.Time != 0
+	hasPayloadData := payload.InitiatorIndex != 0 || payload.ResponderIndex != 0 || payload.Time != 0 || payload.InitiatorMultiPort != nil || payload.ResponderMultiPort != nil
 	if hasPayloadData != flags.expectsPayload {
 		m.failed = true
 		return ErrUnexpectedContent
@@ -315,8 +326,22 @@ func (m *Machine) processPayload(msg []byte, flags msgFlags) error {
 		var remoteIndex uint32
 		if m.result.Initiator {
 			remoteIndex = payload.ResponderIndex
+			if payload.ResponderMultiPort != nil {
+				m.result.MultiportRx = payload.ResponderMultiPort.RxSupported
+				m.result.MultiportTx = payload.ResponderMultiPort.TxSupported
+				if payload.ResponderMultiPort.BasePort <= math.MaxUint16 {
+					m.result.MultiportBasePort = uint16(payload.ResponderMultiPort.BasePort)
+				}
+			}
 		} else {
 			remoteIndex = payload.InitiatorIndex
+			if payload.InitiatorMultiPort != nil {
+				m.result.MultiportRx = payload.InitiatorMultiPort.RxSupported
+				m.result.MultiportTx = payload.InitiatorMultiPort.TxSupported
+				if payload.InitiatorMultiPort.BasePort <= math.MaxUint16 {
+					m.result.MultiportBasePort = uint16(payload.InitiatorMultiPort.BasePort)
+				}
+			}
 		}
 		// The payload presence check above can be satisfied by Time alone, so a payload
 		// could still carry a zero index here. We need to reject it.
@@ -397,11 +422,28 @@ func (m *Machine) marshalOutgoing(flags msgFlags) ([]byte, error) {
 
 		if m.result.Initiator {
 			p.InitiatorIndex = m.result.LocalIndex
+			if m.multiport.Rx || m.multiport.Tx {
+				p.InitiatorMultiPort = &PayloadMultiPortDetails{
+					RxSupported: m.multiport.Rx,
+					TxSupported: m.multiport.Tx,
+					BasePort:    uint32(m.multiport.TxBasePort),
+					TotalPorts:  uint32(m.multiport.TxPorts),
+				}
+			}
 		} else {
 			p.ResponderIndex = m.result.LocalIndex
 			p.InitiatorIndex = m.result.RemoteIndex
+			if m.multiport.Rx || m.multiport.Tx {
+				p.ResponderMultiPort = &PayloadMultiPortDetails{
+					RxSupported: m.multiport.Rx,
+					TxSupported: m.multiport.Tx,
+					BasePort:    uint32(m.multiport.TxBasePort),
+					TotalPorts:  uint32(m.multiport.TxPorts),
+				}
+			}
 		}
 		p.Time = uint64(time.Now().UnixNano())
+
 	}
 	if flags.expectsCert {
 		cred := m.getCred(m.myVersion)

handshake/machine_test.go

@@ -8,6 +8,7 @@ import (
 	"github.com/flynn/noise"
 	"github.com/slackhq/nebula/cert"
 	ct "github.com/slackhq/nebula/cert_test"
+	"github.com/slackhq/nebula/config"
 	"github.com/slackhq/nebula/header"
 	"github.com/slackhq/nebula/noiseutil"
 	"github.com/stretchr/testify/assert"
@@ -444,6 +445,7 @@ func TestMachineThreeMessagePattern(t *testing.T) {
 		initCS.getCredential, v,
 		func() (uint32, error) { return 1000, nil },
 		true, header.HandshakeXXPSK0,
+		config.MultiPortConfig{},
 	)
 	require.NoError(t, err)
 
@@ -452,6 +454,7 @@ func TestMachineThreeMessagePattern(t *testing.T) {
 		respCS.getCredential, v,
 		func() (uint32, error) { return 2000, nil },
 		false, header.HandshakeXXPSK0,
+		config.MultiPortConfig{},
 	)
 	require.NoError(t, err)
 

handshake/payload.go

@@ -20,6 +20,16 @@ type Payload struct {
 	ResponderIndex uint32
 	Time           uint64
 	CertVersion    uint32
+
+	InitiatorMultiPort *PayloadMultiPortDetails
+	ResponderMultiPort *PayloadMultiPortDetails
+}
+
+type PayloadMultiPortDetails struct {
+	RxSupported bool
+	TxSupported bool
+	BasePort    uint32
+	TotalPorts  uint32
 }
 
 // Proto field numbers for NebulaHandshakeDetails
@@ -29,6 +39,17 @@ const (
 	fieldResponderIndex = 3 // uint32
 	fieldTime           = 5 // uint64
 	fieldCertVersion    = 8 // uint32
+
+	fieldInitiatorMultiPort = 6 // MultiPortDetails
+	fieldResponderMultiPort = 7 // MultiPortDetails
+)
+
+// Proto field numbers for MultiPortDetails
+const (
+	fieldMultiportRxSupported = 1 // bool
+	fieldMultiportTxSupported = 2 // bool
+	fieldMultiportBasePort    = 3 // uint32
+	fieldMultiportTotalPorts  = 4 // uint32
 )
 
 // MarshalPayload encodes a handshake payload in protobuf wire format compatible
@@ -57,6 +78,16 @@ func MarshalPayload(out []byte, p Payload) []byte {
 		details = protowire.AppendTag(details, fieldCertVersion, protowire.VarintType)
 		details = protowire.AppendVarint(details, uint64(p.CertVersion))
 	}
+	if p.InitiatorMultiPort != nil {
+		details = protowire.AppendTag(details, fieldInitiatorMultiPort, protowire.BytesType)
+		details = protowire.AppendVarint(details, uint64(p.InitiatorMultiPort.size()))
+		details = p.InitiatorMultiPort.marshal(details)
+	}
+	if p.ResponderMultiPort != nil {
+		details = protowire.AppendTag(details, fieldResponderMultiPort, protowire.BytesType)
+		details = protowire.AppendVarint(details, uint64(p.ResponderMultiPort.size()))
+		details = p.ResponderMultiPort.marshal(details)
+	}
 
 	out = protowire.AppendTag(out, 1, protowire.BytesType)
 	out = protowire.AppendBytes(out, details)
@@ -64,6 +95,23 @@ func MarshalPayload(out []byte, p Payload) []byte {
 	return out
 }
 
+func (p PayloadMultiPortDetails) marshal(details []byte) []byte {
+	details = protowire.AppendTag(details, fieldMultiportRxSupported, protowire.VarintType)
+	details = protowire.AppendVarint(details, protowire.EncodeBool(p.RxSupported))
+	details = protowire.AppendTag(details, fieldMultiportTxSupported, protowire.VarintType)
+	details = protowire.AppendVarint(details, protowire.EncodeBool(p.TxSupported))
+	details = protowire.AppendTag(details, fieldMultiportBasePort, protowire.VarintType)
+	details = protowire.AppendVarint(details, uint64(p.BasePort))
+	details = protowire.AppendTag(details, fieldMultiportTotalPorts, protowire.VarintType)
+	details = protowire.AppendVarint(details, uint64(p.TotalPorts))
+
+	return details
+}
+
+func (p PayloadMultiPortDetails) size() int {
+	return 4 + 2 + protowire.SizeVarint(uint64(p.BasePort)) + protowire.SizeVarint(uint64(p.TotalPorts))
+}
+
 // UnmarshalPayload decodes a protobuf-encoded NebulaHandshake message.
 func UnmarshalPayload(b []byte) (Payload, error) {
 	var p Payload
@@ -161,6 +209,97 @@ func unmarshalPayloadDetails(p *Payload, b []byte) error {
 			}
 			p.CertVersion = uint32(v)
 			b = b[n:]
+		case fieldInitiatorMultiPort:
+			if typ != protowire.BytesType {
+				return errInvalidHandshakeDetails
+			}
+			d, n := protowire.ConsumeBytes(b)
+			if n < 0 {
+				return errInvalidHandshakeMessage
+			}
+			b = b[n:]
+			p.InitiatorMultiPort = new(PayloadMultiPortDetails)
+			if err := unmarshalPayloadMultiPortDetails(p.InitiatorMultiPort, d); err != nil {
+				return err
+			}
+		case fieldResponderMultiPort:
+			if typ != protowire.BytesType {
+				return errInvalidHandshakeDetails
+			}
+			d, n := protowire.ConsumeBytes(b)
+			if n < 0 {
+				return errInvalidHandshakeMessage
+			}
+			b = b[n:]
+			p.ResponderMultiPort = new(PayloadMultiPortDetails)
+			if err := unmarshalPayloadMultiPortDetails(p.ResponderMultiPort, d); err != nil {
+				return err
+			}
+		default:
+			n := protowire.ConsumeFieldValue(num, typ, b)
+			if n < 0 {
+				return errInvalidHandshakeDetails
+			}
+			b = b[n:]
+		}
+	}
+	return nil
+}
+
+func unmarshalPayloadMultiPortDetails(p *PayloadMultiPortDetails, b []byte) error {
+	for len(b) > 0 {
+		num, typ, n := protowire.ConsumeTag(b)
+		if n < 0 {
+			return errInvalidHandshakeDetails
+		}
+		b = b[n:]
+
+		// For known field numbers, reject any non-matching wire type as a
+		// hard error rather than silently skipping. The caller will catch
+		// missing-field cases downstream, but a wire-type mismatch on a tag
+		// we know is a peer protocol violation worth flagging here.
+		// Repeated occurrences of a singular field follow proto3 last-wins.
+		switch num {
+		case fieldMultiportRxSupported:
+			if typ != protowire.VarintType {
+				return errInvalidHandshakeDetails
+			}
+			v, n := protowire.ConsumeVarint(b)
+			if n < 0 || v > math.MaxUint32 {
+				return errInvalidHandshakeDetails
+			}
+			p.RxSupported = protowire.DecodeBool(v)
+			b = b[n:]
+		case fieldMultiportTxSupported:
+			if typ != protowire.VarintType {
+				return errInvalidHandshakeDetails
+			}
+			v, n := protowire.ConsumeVarint(b)
+			if n < 0 || v > math.MaxUint32 {
+				return errInvalidHandshakeDetails
+			}
+			p.TxSupported = protowire.DecodeBool(v)
+			b = b[n:]
+		case fieldMultiportBasePort:
+			if typ != protowire.VarintType {
+				return errInvalidHandshakeDetails
+			}
+			v, n := protowire.ConsumeVarint(b)
+			if n < 0 || v > math.MaxUint32 {
+				return errInvalidHandshakeDetails
+			}
+			p.BasePort = uint32(v)
+			b = b[n:]
+		case fieldMultiportTotalPorts:
+			if typ != protowire.VarintType {
+				return errInvalidHandshakeDetails
+			}
+			v, n := protowire.ConsumeVarint(b)
+			if n < 0 || v > math.MaxUint32 {
+				return errInvalidHandshakeDetails
+			}
+			p.TotalPorts = uint32(v)
+			b = b[n:]
 		default:
 			n := protowire.ConsumeFieldValue(num, typ, b)
 			if n < 0 {

handshake/payload_test.go

@@ -117,24 +117,24 @@ func TestPayloadUnknownFields(t *testing.T) {
 		assert.Equal(t, uint32(88), got.ResponderIndex)
 	})
 
-	t.Run("reserved fields 6 and 7 are skipped", func(t *testing.T) {
-		// Fields 6 and 7 are reserved in the proto definition
-		var details []byte
-		details = protowire.AppendTag(details, fieldInitiatorIndex, protowire.VarintType)
-		details = protowire.AppendVarint(details, 100)
-		details = protowire.AppendTag(details, 6, protowire.VarintType)
-		details = protowire.AppendVarint(details, 1)
-		details = protowire.AppendTag(details, 7, protowire.VarintType)
-		details = protowire.AppendVarint(details, 2)
+	// t.Run("reserved fields 6 and 7 are skipped", func(t *testing.T) {
+	// 	// Fields 6 and 7 are reserved in the proto definition
+	// 	var details []byte
+	// 	details = protowire.AppendTag(details, fieldInitiatorIndex, protowire.VarintType)
+	// 	details = protowire.AppendVarint(details, 100)
+	// 	details = protowire.AppendTag(details, 6, protowire.VarintType)
+	// 	details = protowire.AppendVarint(details, 1)
+	// 	details = protowire.AppendTag(details, 7, protowire.VarintType)
+	// 	details = protowire.AppendVarint(details, 2)
 
-		var data []byte
-		data = protowire.AppendTag(data, 1, protowire.BytesType)
-		data = protowire.AppendBytes(data, details)
+	// 	var data []byte
+	// 	data = protowire.AppendTag(data, 1, protowire.BytesType)
+	// 	data = protowire.AppendBytes(data, details)
 
-		got, err := UnmarshalPayload(data)
-		require.NoError(t, err)
-		assert.Equal(t, uint32(100), got.InitiatorIndex)
-	})
+	// 	got, err := UnmarshalPayload(data)
+	// 	require.NoError(t, err)
+	// 	assert.Equal(t, uint32(100), got.InitiatorIndex)
+	// })
 }
 
 func TestPayloadBytesConsumed(t *testing.T) {

handshake_manager.go

@@ -14,6 +14,7 @@ import (
 
 	"github.com/rcrowley/go-metrics"
 	"github.com/slackhq/nebula/cert"
+	"github.com/slackhq/nebula/config"
 	"github.com/slackhq/nebula/handshake"
 	"github.com/slackhq/nebula/header"
 	"github.com/slackhq/nebula/udp"
@@ -71,6 +72,9 @@ type HandshakeManager struct {
 	f                      *Interface
 	l                      *slog.Logger
 
+	multiPort config.MultiPortConfig
+	udpRaw    *udp.RawConn
+
 	// can be used to trigger outbound handshake for the given vpnIp
 	trigger chan netip.Addr
 }
@@ -291,6 +295,7 @@ func (hm *HandshakeManager) handleOutbound(vpnIp netip.Addr, lighthouseTriggered
 
 	// Send the handshake to all known ips, stage 2 takes care of assigning the hostinfo.remote based on the first to reply
 	var sentTo []netip.AddrPort
+	var sentMultiport bool
 	hostinfo.remotes.ForEach(hm.mainHostMap.GetPreferredRanges(), func(addr netip.AddrPort, _ bool) {
 		hm.messageMetrics.Tx(header.Handshake, hh.machine.Subtype(), 1)
 		err := hm.outside.WriteTo(stage0, addr)
@@ -311,6 +316,29 @@ func (hm *HandshakeManager) handleOutbound(vpnIp netip.Addr, lighthouseTriggered
 		} else {
 			sentTo = append(sentTo, addr)
 		}
+
+		// Attempt a multiport handshake if we are past the TxHandshakeDelay attempts
+		if hm.multiPort.TxHandshake && hm.udpRaw != nil && hh.counter >= hm.multiPort.TxHandshakeDelay {
+			sentMultiport = true
+			// We need to re-allocate with 8 bytes at the start of SOCK_RAW
+			raw := hostinfo.HandshakePacket[0x80]
+			if raw == nil {
+				raw = make([]byte, len(hostinfo.HandshakePacket[0])+udp.RawOverhead)
+				copy(raw[udp.RawOverhead:], hostinfo.HandshakePacket[0])
+				hostinfo.HandshakePacket[0x80] = raw
+			}
+
+			hm.messageMetrics.Tx(header.Handshake, header.MessageSubType(hostinfo.HandshakePacket[0][1]), 1)
+			err = hm.udpRaw.WriteTo(raw, udp.RandomSendPort.UDPSendPort(hm.multiPort.TxPorts), addr)
+			if err != nil {
+				hostinfo.logger(hm.l).Error("Failed to send handshake message",
+					"error", err,
+					"udpAddr", addr,
+					"initiatorIndex", hostinfo.localIndexId,
+					"handshake", hsFields,
+				)
+			}
+		}
 	})
 
 	// Don't be too noisy or confusing if we fail to send a handshake - if we don't get through we'll eventually log a timeout,
@@ -320,6 +348,7 @@ func (hm *HandshakeManager) handleOutbound(vpnIp netip.Addr, lighthouseTriggered
 			"udpAddrs", sentTo,
 			"initiatorIndex", hostinfo.localIndexId,
 			"handshake", hsFields,
+			"multiportHandshake", sentMultiport,
 		)
 	} else if hm.l.Enabled(context.Background(), slog.LevelDebug) {
 		hostinfo.logger(hm.l).Debug("Handshake message sent",
@@ -672,6 +701,7 @@ func (hm *HandshakeManager) buildStage0Packet(hh *HandshakeHostInfo) bool {
 		v, cs.GetCredential,
 		hm.certVerifier(), func() (uint32, error) { return hm.allocateIndex(hh) },
 		true, header.HandshakeIXPSK0,
+		hm.multiPort,
 	)
 	if err != nil {
 		hm.f.l.Error("Failed to create handshake machine",
@@ -713,6 +743,7 @@ func (hm *HandshakeManager) beginHandshake(via ViaSender, packet []byte, h *head
 		v, cs.GetCredential,
 		hm.certVerifier(), func() (uint32, error) { return generateIndex(f.l) },
 		false, header.HandshakeIXPSK0,
+		hm.multiPort,
 	)
 	if err != nil {
 		f.l.Error("Failed to create handshake machine", "from", via, "error", err)
@@ -737,6 +768,12 @@ func (hm *HandshakeManager) beginHandshake(via ViaSender, packet []byte, h *head
 		return
 	}
 
+	if !via.IsRelayed && result.MultiportTx && result.MultiportBasePort != via.UdpAddr.Port() {
+		// The other side sent us a handshake from a different port, make sure
+		// we send responses back to the BasePort
+		via.UdpAddr = netip.AddrPortFrom(via.UdpAddr.Addr(), result.MultiportBasePort)
+	}
+
 	remoteCert := result.RemoteCert
 	if remoteCert == nil {
 		f.l.Error("Handshake did not produce a peer certificate", "from", via)
@@ -761,6 +798,8 @@ func (hm *HandshakeManager) beginHandshake(via ViaSender, packet []byte, h *head
 			relayForByAddr: map[netip.Addr]*Relay{},
 			relayForByIdx:  map[uint32]*Relay{},
 		},
+		multiportTx: hm.multiPort.Tx && result.MultiportRx,
+		multiportRx: hm.multiPort.Rx && result.MultiportTx,
 	}
 
 	msg := "Handshake message received"
@@ -777,6 +816,8 @@ func (hm *HandshakeManager) beginHandshake(via ViaSender, packet []byte, h *head
 		"initiatorIndex", result.RemoteIndex,
 		"responderIndex", result.LocalIndex,
 		"handshake", m{"stage": uint64(machine.MessageIndex()), "style": header.SubTypeName(header.Handshake, machine.Subtype())},
+		"multiportTx", hostinfo.multiportTx,
+		"multiportRx", hostinfo.multiportRx,
 	)
 
 	// packet aliases the listener's incoming buffer, so this copy must stay.
@@ -867,6 +908,14 @@ func (hm *HandshakeManager) continueHandshake(via ViaSender, hh *HandshakeHostIn
 		return
 	}
 
+	if !via.IsRelayed && result.MultiportTx && result.MultiportBasePort != via.UdpAddr.Port() {
+		// The other side sent us a handshake from a different port, make sure
+		// we send responses back to the BasePort
+		via.UdpAddr = netip.AddrPortFrom(via.UdpAddr.Addr(), result.MultiportBasePort)
+	}
+	hostinfo.multiportTx = hm.multiPort.Tx && result.MultiportRx
+	hostinfo.multiportRx = hm.multiPort.Rx && result.MultiportTx
+
 	// Handshake complete; build the ConnectionState now that we have keys and a verified peer cert.
 	hostinfo.ConnectionState = newConnectionStateFromResult(result)
 
@@ -961,6 +1010,8 @@ func (hm *HandshakeManager) continueHandshake(via ViaSender, hh *HandshakeHostIn
 		"handshake", m{"stage": uint64(machine.MessageIndex()), "style": header.SubTypeName(header.Handshake, machine.Subtype())},
 		"durationNs", duration,
 		"sentCachedPackets", len(hh.packetStore),
+		"multiportTx", hostinfo.multiportTx,
+		"multiportRx", hostinfo.multiportRx,
 	)
 
 	hostinfo.vpnAddrs = vpnAddrs
@@ -1102,6 +1153,10 @@ func (hm *HandshakeManager) handleCheckAndCompleteError(err error, existing, hos
 
 	switch err {
 	case ErrAlreadySeen:
+		if hostinfo.multiportRx {
+			// The other host is sending to us with multiport, so only grab the IP
+			via.UdpAddr = netip.AddrPortFrom(via.UdpAddr.Addr(), hostinfo.GetRemote().Port())
+		}
 		if existing.SetRemoteIfPreferred(f.hostMap, via) {
 			f.SendMessageToVpnAddr(header.Test, header.TestRequest, hostinfo.vpnAddrs[0], []byte(""), make([]byte, 12, 12), make([]byte, mtu))
 		}

hostmap.go

@@ -254,6 +254,12 @@ type HostInfo struct {
 	networks   *bart.Table[NetworkType]
 	relayState RelayState
 
+	// If true, we should send to this remote using multiport
+	multiportTx bool
+
+	// If true, we will receive from this remote using multiport
+	multiportRx bool
+
 	// HandshakePacket records the packets used to create this hostinfo
 	// We need these to avoid replayed handshake packets creating new hostinfos which causes churn
 	HandshakePacket map[uint8][]byte

inside.go

@@ -10,6 +10,7 @@ import (
 	"github.com/slackhq/nebula/iputil"
 	"github.com/slackhq/nebula/noiseutil"
 	"github.com/slackhq/nebula/routing"
+	"github.com/slackhq/nebula/udp"
 )
 
 func (f *Interface) consumeInsidePacket(packet []byte, fwPacket *firewall.Packet, nb, out []byte, q int, localCache firewall.ConntrackCache) {
@@ -73,7 +74,7 @@ func (f *Interface) consumeInsidePacket(packet []byte, fwPacket *firewall.Packet
 
 	dropReason := f.firewall.Drop(*fwPacket, false, hostinfo, f.pki.GetCAPool(), localCache)
 	if dropReason == nil {
-		f.sendNoMetrics(header.Message, 0, hostinfo.ConnectionState, hostinfo, netip.AddrPort{}, packet, nb, out, q)
+		f.sendNoMetrics(header.Message, 0, hostinfo.ConnectionState, hostinfo, netip.AddrPort{}, packet, nb, out, q, fwPacket)
 
 	} else {
 		f.rejectInside(packet, out, q)
@@ -122,7 +123,7 @@ func (f *Interface) rejectOutside(packet []byte, ci *ConnectionState, hostinfo *
 		return
 	}
 
-	f.sendNoMetrics(header.Message, 0, ci, hostinfo, netip.AddrPort{}, out, nb, packet, q)
+	f.sendNoMetrics(header.Message, 0, ci, hostinfo, netip.AddrPort{}, out, nb, packet, q, nil)
 }
 
 // Handshake will attempt to initiate a tunnel with the provided vpn address. This is a no-op if the tunnel is already established or being established
@@ -235,7 +236,7 @@ func (f *Interface) sendMessageNow(t header.MessageType, st header.MessageSubTyp
 		return
 	}
 
-	f.sendNoMetrics(header.Message, st, hostinfo.ConnectionState, hostinfo, netip.AddrPort{}, p, nb, out, 0)
+	f.sendNoMetrics(header.Message, st, hostinfo.ConnectionState, hostinfo, netip.AddrPort{}, p, nb, out, 0, nil)
 }
 
 // SendMessageToVpnAddr handles real addr:port lookup and sends to the current best known address for vpnAddr.
@@ -267,12 +268,12 @@ func (f *Interface) SendMessageToHostInfo(t header.MessageType, st header.Messag
 
 func (f *Interface) send(t header.MessageType, st header.MessageSubType, ci *ConnectionState, hostinfo *HostInfo, p, nb, out []byte) {
 	f.messageMetrics.Tx(t, st, 1)
-	f.sendNoMetrics(t, st, ci, hostinfo, netip.AddrPort{}, p, nb, out, 0)
+	f.sendNoMetrics(t, st, ci, hostinfo, netip.AddrPort{}, p, nb, out, 0, nil)
 }
 
 func (f *Interface) sendTo(t header.MessageType, st header.MessageSubType, ci *ConnectionState, hostinfo *HostInfo, remote netip.AddrPort, p, nb, out []byte) {
 	f.messageMetrics.Tx(t, st, 1)
-	f.sendNoMetrics(t, st, ci, hostinfo, remote, p, nb, out, 0)
+	f.sendNoMetrics(t, st, ci, hostinfo, remote, p, nb, out, 0, nil)
 }
 
 // SendVia sends a payload through a Relay tunnel. No authentication or encryption is done
@@ -340,10 +341,27 @@ func (f *Interface) SendVia(via *HostInfo,
 	f.connectionManager.RelayUsed(relay.LocalIndex)
 }
 
-func (f *Interface) sendNoMetrics(t header.MessageType, st header.MessageSubType, ci *ConnectionState, hostinfo *HostInfo, remote netip.AddrPort, p, nb, out []byte, q int) {
+func (f *Interface) sendNoMetrics(t header.MessageType, st header.MessageSubType, ci *ConnectionState, hostinfo *HostInfo, remote netip.AddrPort, p, nb, out []byte, q int, udpPortGetter udp.SendPortGetter) {
 	if ci.eKey == nil {
 		return
 	}
+
+	multiport := f.multiPort.Tx && hostinfo.multiportTx
+	rawOut := out
+	if multiport {
+		if len(out) < udp.RawOverhead {
+			// NOTE: This is because some spots in the code send us `out[:0]`, so
+			// we need to expand the slice back out to get our 8 bytes back.
+			out = out[:udp.RawOverhead]
+		}
+		// Preserve bytes needed for the raw socket
+		out = out[udp.RawOverhead:]
+
+		if udpPortGetter == nil {
+			udpPortGetter = udp.RandomSendPort
+		}
+	}
+
 	useRelay := !remote.IsValid() && !hostinfo.GetRemote().IsValid()
 	fullOut := out
 
@@ -396,7 +414,13 @@ func (f *Interface) sendNoMetrics(t header.MessageType, st header.MessageSubType
 	}
 
 	if remote.IsValid() {
-		err = f.writers[q].WriteTo(out, remote)
+		if multiport {
+			rawOut = rawOut[:len(out)+udp.RawOverhead]
+			port := udpPortGetter.UDPSendPort(f.multiPort.TxPorts)
+			err = f.udpRaw.WriteTo(rawOut, port, remote)
+		} else {
+			err = f.writers[q].WriteTo(out, remote)
+		}
 		if err != nil {
 			hostinfo.logger(f.l).Error("Failed to write outgoing packet",
 				"error", err,
@@ -404,7 +428,13 @@ func (f *Interface) sendNoMetrics(t header.MessageType, st header.MessageSubType
 			)
 		}
 	} else if hr := hostinfo.GetRemote(); hr.IsValid() {
-		err = f.writers[q].WriteTo(out, hr)
+		if multiport {
+			rawOut = rawOut[:len(out)+udp.RawOverhead]
+			port := udpPortGetter.UDPSendPort(f.multiPort.TxPorts)
+			err = f.udpRaw.WriteTo(rawOut, port, hr)
+		} else {
+			err = f.writers[q].WriteTo(out, hr)
+		}
 		if err != nil {
 			hostinfo.logger(f.l).Error("Failed to write outgoing packet",
 				"error", err,

interface.go

@@ -99,6 +99,9 @@ type Interface struct {
 	// triggerShutdown is a function that will be run exactly once, when onFatal swaps something non-nil into fatalErr
 	triggerShutdown func()
 
+	udpRaw    *udp.RawConn
+	multiPort config.MultiPortConfig
+
 	metricHandshakes    metrics.Histogram
 	messageMetrics      *MessageMetrics
 	cachedPacketMetrics *cachedPacketMetrics
@@ -106,6 +109,15 @@ type Interface struct {
 	l *slog.Logger
 }
 
+type MultiPortConfig struct {
+	Tx               bool
+	Rx               bool
+	TxBasePort       uint16
+	TxPorts          int
+	TxHandshake      bool
+	TxHandshakeDelay int64
+}
+
 type EncWriter interface {
 	SendVia(via *HostInfo,
 		relay *Relay,
@@ -249,6 +261,8 @@ func (f *Interface) activate() error {
 
 	metrics.GetOrRegisterGauge("routines", nil).Update(int64(f.routines))
 
+	metrics.GetOrRegisterGauge("multiport.tx_ports", nil).Update(int64(f.multiPort.TxPorts))
+
 	// Prepare n tun queues
 	var reader io.ReadWriteCloser = f.inside
 	for i := 0; i < f.routines; i++ {
@@ -505,6 +519,8 @@ func (f *Interface) emitStats(ctx context.Context, i time.Duration) {
 
 	udpStats := udp.NewUDPStatsEmitter(f.writers)
 
+	var rawStats func()
+
 	certExpirationGauge := metrics.GetOrRegisterGauge("certificate.ttl_seconds", nil)
 	certInitiatingVersion := metrics.GetOrRegisterGauge("certificate.initiating_version", nil)
 	certMaxVersion := metrics.GetOrRegisterGauge("certificate.max_version", nil)
@@ -519,6 +535,13 @@ func (f *Interface) emitStats(ctx context.Context, i time.Duration) {
 		certExpirationGauge.Update(int64(defaultCrt.NotAfter().Sub(time.Now()) / time.Second))
 		certInitiatingVersion.Update(int64(defaultCrt.Version()))
 
+		if f.udpRaw != nil {
+			if rawStats == nil {
+				rawStats = udp.NewRawStatsEmitter(f.udpRaw)
+			}
+			rawStats()
+		}
+
 		// Report the max certificate version we are capable of using
 		if certState.v2Cert != nil {
 			certMaxVersion.Update(int64(certState.v2Cert.Version()))

main.go

@@ -244,6 +244,39 @@ func Main(c *config.C, configTest bool, buildVersion string, l *slog.Logger, dev
 		ifce.writers = udpConns
 		lightHouse.ifce = ifce
 
+		loadMultiPortConfig := func(c *config.C) {
+			ifce.multiPort.Rx = c.GetBool("tun.multiport.rx_enabled", false)
+
+			tx := c.GetBool("tun.multiport.tx_enabled", false)
+
+			if tx && ifce.udpRaw == nil {
+				ifce.udpRaw, err = udp.NewRawConn(l, c.GetString("listen.host", "0.0.0.0"), port, uint16(port))
+				if err != nil {
+					l.Error("Failed to get raw socket for tun.multiport.tx_enabled", "error", err)
+					ifce.udpRaw = nil
+					tx = false
+				}
+			}
+
+			if tx {
+				ifce.multiPort.TxBasePort = uint16(port)
+				ifce.multiPort.TxPorts = c.GetInt("tun.multiport.tx_ports", 100)
+				ifce.multiPort.TxHandshake = c.GetBool("tun.multiport.tx_handshake", false)
+				ifce.multiPort.TxHandshakeDelay = int64(c.GetInt("tun.multiport.tx_handshake_delay", 2))
+				ifce.udpRaw.ReloadConfig(c)
+			}
+			ifce.multiPort.Tx = tx
+
+			// TODO: if we upstream this, make this cleaner
+			handshakeManager.udpRaw = ifce.udpRaw
+			handshakeManager.multiPort = ifce.multiPort
+
+			l.Info("Multiport configured", "multiPort", ifce.multiPort)
+		}
+
+		loadMultiPortConfig(c)
+		c.RegisterReloadCallback(loadMultiPortConfig)
+
 		ifce.RegisterConfigChangeCallbacks(c)
 		ifce.reloadDisconnectInvalid(c)
 		ifce.reloadSendRecvError(c)

outside.go

@@ -264,6 +264,15 @@ func (f *Interface) sendCloseTunnel(h *HostInfo) {
 func (f *Interface) handleHostRoaming(hostinfo *HostInfo, via ViaSender) {
 	curRemote := hostinfo.GetRemote()
 	if !via.IsRelayed && curRemote != via.UdpAddr {
+		if hostinfo.multiportRx {
+			// If the remote is sending with multiport, we aren't roaming unless
+			// the IP has changed
+			if curRemote.Addr().Compare(via.UdpAddr.Addr()) == 0 {
+				return
+			}
+			// Keep the port from the original hostinfo, because the remote is transmitting from multiport ports
+			via.UdpAddr = netip.AddrPortFrom(via.UdpAddr.Addr(), curRemote.Port())
+		}
 		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)

udp/udp_raw.go

@@ -0,0 +1,16 @@
+package udp
+
+import mathrand "math/rand"
+
+type SendPortGetter interface {
+	// UDPSendPort returns the port to use
+	UDPSendPort(maxPort int) uint16
+}
+
+type randomSendPort struct{}
+
+func (randomSendPort) UDPSendPort(maxPort int) uint16 {
+	return uint16(mathrand.Intn(maxPort))
+}
+
+var RandomSendPort = randomSendPort{}

udp/udp_raw_linux.go

@@ -0,0 +1,191 @@
+//go:build !android && !e2e_testing
+// +build !android,!e2e_testing
+
+package udp
+
+import (
+	"encoding/binary"
+	"fmt"
+	"log/slog"
+	"net"
+	"net/netip"
+	"syscall"
+	"unsafe"
+
+	"github.com/rcrowley/go-metrics"
+	"github.com/slackhq/nebula/config"
+	"golang.org/x/net/ipv4"
+	"golang.org/x/sys/unix"
+)
+
+// RawOverhead is the number of bytes that need to be reserved at the start of
+// the raw bytes passed to (*RawConn).WriteTo. This is used by WriteTo to prefix
+// the IP and UDP headers.
+const RawOverhead = 28
+
+type RawConn struct {
+	sysFd    int
+	basePort uint16
+	l        *slog.Logger
+}
+
+func NewRawConn(l *slog.Logger, ip string, port int, basePort uint16) (*RawConn, error) {
+	syscall.ForkLock.RLock()
+	// With IPPROTO_UDP, the linux kernel tries to deliver every UDP packet
+	// received in the system to our socket. This constantly overflows our
+	// buffer and marks our socket as having dropped packets. This makes the
+	// stats on the socket useless.
+	//
+	// In contrast, IPPROTO_RAW is not delivered any packets and thus our read
+	// buffer will not fill up and mark as having dropped packets. The only
+	// difference is that we have to assemble the IP header as well, but this
+	// is fairly easy since Linux does the checksum for us.
+	//
+	// TODO: How to get this working with Inet6 correctly? I was having issues
+	// with the source address when testing before, probably need to `bind(2)`?
+	fd, err := unix.Socket(unix.AF_INET, unix.SOCK_RAW, unix.IPPROTO_RAW)
+	if err == nil {
+		unix.CloseOnExec(fd)
+	}
+	syscall.ForkLock.RUnlock()
+	if err != nil {
+		return nil, err
+	}
+
+	// We only want to send, not recv. This will hopefully help the kernel avoid
+	// wasting time on us
+	if err = unix.SetsockoptInt(fd, unix.SOL_SOCKET, unix.SO_RCVBUF, 0); err != nil {
+		return nil, fmt.Errorf("unable to set SO_RCVBUF: %s", err)
+	}
+
+	var lip [16]byte
+	copy(lip[:], net.ParseIP(ip))
+
+	// TODO do we need to `bind(2)` so that we send from the correct address/interface?
+	if err = unix.Bind(fd, &unix.SockaddrInet6{Addr: lip, Port: port}); err != nil {
+		return nil, fmt.Errorf("unable to bind to socket: %s", err)
+	}
+
+	return &RawConn{
+		sysFd:    fd,
+		basePort: basePort,
+		l:        l,
+	}, nil
+}
+
+// WriteTo must be called with raw leaving the first `udp.RawOverhead` bytes empty,
+// for the IP/UDP headers.
+func (u *RawConn) WriteTo(raw []byte, fromPort uint16, ip netip.AddrPort) error {
+	var rsa unix.RawSockaddrInet4
+	rsa.Family = unix.AF_INET
+	rsa.Addr = ip.Addr().As4()
+
+	totalLen := len(raw)
+	udpLen := totalLen - ipv4.HeaderLen
+
+	// IP header
+	raw[0] = byte(ipv4.Version<<4 | (ipv4.HeaderLen >> 2 & 0x0f))
+	raw[1] = 0 // tos
+	binary.BigEndian.PutUint16(raw[2:4], uint16(totalLen))
+	binary.BigEndian.PutUint16(raw[4:6], 0)   // id (linux does it for us)
+	binary.BigEndian.PutUint16(raw[6:8], 0)   // frag options
+	raw[8] = byte(64)                         // ttl
+	raw[9] = byte(17)                         // protocol
+	binary.BigEndian.PutUint16(raw[10:12], 0) // checksum (linux does it for us)
+	binary.BigEndian.PutUint32(raw[12:16], 0) // src (linux does it for us)
+	copy(raw[16:20], rsa.Addr[:])             // dst
+
+	// UDP header
+	fromPort = u.basePort + fromPort
+	binary.BigEndian.PutUint16(raw[20:22], uint16(fromPort))  // src port
+	binary.BigEndian.PutUint16(raw[22:24], uint16(ip.Port())) // dst port
+	binary.BigEndian.PutUint16(raw[24:26], uint16(udpLen))    // UDP length
+	binary.BigEndian.PutUint16(raw[26:28], 0)                 // checksum (optional)
+
+	for {
+		_, _, err := unix.Syscall6(
+			unix.SYS_SENDTO,
+			uintptr(u.sysFd),
+			uintptr(unsafe.Pointer(&raw[0])),
+			uintptr(len(raw)),
+			uintptr(0),
+			uintptr(unsafe.Pointer(&rsa)),
+			uintptr(unix.SizeofSockaddrInet4),
+		)
+
+		if err != 0 {
+			return &net.OpError{Op: "sendto", Err: err}
+		}
+
+		//TODO: handle incomplete writes
+
+		return nil
+	}
+}
+
+func (u *RawConn) ReloadConfig(c *config.C) {
+	b := c.GetInt("listen.write_buffer", 0)
+	if b <= 0 {
+		return
+	}
+
+	if err := u.SetSendBuffer(b); err != nil {
+		u.l.Error("Failed to set listen.write_buffer", "error", err)
+		return
+	}
+
+	s, err := u.GetSendBuffer()
+	if err != nil {
+		u.l.Warn("Failed to get listen.write_buffer", "error", err)
+		return
+	}
+
+	u.l.Info("listen.write_buffer was set", "size", s)
+}
+
+func (u *RawConn) SetSendBuffer(n int) error {
+	return unix.SetsockoptInt(u.sysFd, unix.SOL_SOCKET, unix.SO_SNDBUFFORCE, n)
+}
+
+func (u *RawConn) GetSendBuffer() (int, error) {
+	return unix.GetsockoptInt(u.sysFd, unix.SOL_SOCKET, unix.SO_SNDBUF)
+}
+
+func (u *RawConn) 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 err != 0 {
+		return err
+	}
+	return nil
+}
+
+func NewRawStatsEmitter(rawConn *RawConn) func() {
+	// Check if our kernel supports SO_MEMINFO before registering the gauges
+	var gauges [unix.SK_MEMINFO_VARS]metrics.Gauge
+	var meminfo [unix.SK_MEMINFO_VARS]uint32
+	if err := rawConn.getMemInfo(&meminfo); err == nil {
+		gauges = [unix.SK_MEMINFO_VARS]metrics.Gauge{
+			metrics.GetOrRegisterGauge("raw.rmem_alloc", nil),
+			metrics.GetOrRegisterGauge("raw.rcvbuf", nil),
+			metrics.GetOrRegisterGauge("raw.wmem_alloc", nil),
+			metrics.GetOrRegisterGauge("raw.sndbuf", nil),
+			metrics.GetOrRegisterGauge("raw.fwd_alloc", nil),
+			metrics.GetOrRegisterGauge("raw.wmem_queued", nil),
+			metrics.GetOrRegisterGauge("raw.optmem", nil),
+			metrics.GetOrRegisterGauge("raw.backlog", nil),
+			metrics.GetOrRegisterGauge("raw.drops", nil),
+		}
+	} else {
+		// return no-op because we don't support SO_MEMINFO
+		return func() {}
+	}
+
+	return func() {
+		if err := rawConn.getMemInfo(&meminfo); err == nil {
+			for j := 0; j < unix.SK_MEMINFO_VARS; j++ {
+				gauges[j].Update(int64(meminfo[j]))
+			}
+		}
+	}
+}

udp/udp_raw_unsupported.go

@@ -0,0 +1,29 @@
+//go:build !linux || android || e2e_testing
+// +build !linux android e2e_testing
+
+package udp
+
+import (
+	"fmt"
+	"log/slog"
+	"net/netip"
+	"runtime"
+
+	"github.com/slackhq/nebula/config"
+)
+
+const RawOverhead = 0
+
+type RawConn struct{}
+
+func NewRawConn(l *slog.Logger, ip string, port int, basePort uint16) (*RawConn, error) {
+	return nil, fmt.Errorf("multiport tx is not supported on %s", runtime.GOOS)
+}
+
+func (u *RawConn) WriteTo(raw []byte, fromPort uint16, addr netip.AddrPort) error {
+	return fmt.Errorf("multiport tx is not supported on %s", runtime.GOOS)
+}
+
+func (u *RawConn) ReloadConfig(c *config.C) {}
+
+func NewRawStatsEmitter(rawConn *RawConn) func() { return func() {} }