Rework signal protocol to minimize number of exchanged initial RP 512KB keys

client/internal/engine.go

@@ -1809,6 +1809,7 @@ func (e *Engine) createPeerConn(pubKey string, allowedIPs []netip.Prefix, agentV
 			PubKey:         e.getRosenpassPubKey(),
 			Addr:           e.getRosenpassAddr(),
 			PermissiveMode: e.config.RosenpassPermissive,
+			KeyResolver:    e.rosenpassKeyResolver(),
 		},
 		ICEConfig: e.createICEConfig(),
 	}
@@ -1879,6 +1880,8 @@ func (e *Engine) receiveSignalEvents() error {
 
 				log.Debugf("receiveMSG: took %s to get lock for peer %s with session id %s", gotLock, msg.Key, offerAnswer.SessionID)
 
+				e.applyRosenpassKeyExchange(msg, offerAnswer)
+
 				if msg.Body.Type == sProto.Body_OFFER {
 					conn.OnRemoteOffer(*offerAnswer)
 				} else {
@@ -2222,6 +2225,34 @@ func (e *Engine) getRosenpassAddr() string {
 	return ""
 }
 
+// rosenpassKeyResolver returns the Rosenpass manager as the offer/answer key
+// resolver, or a true nil interface when Rosenpass is disabled (returning the
+// typed-nil *Manager would make the interface non-nil and panic on use).
+func (e *Engine) rosenpassKeyResolver() peer.RosenpassKeyResolver {
+	if e.rpManager == nil {
+		return nil
+	}
+	return e.rpManager
+}
+
+// applyRosenpassKeyExchange reconciles the fingerprint/cache fields of an incoming
+// offer/answer against the Rosenpass manager's cache: it resolves the remote peer's
+// full public key (from the message or the cache) into the OfferAnswer, and records
+// whether the peer acknowledged holding our key. No-op when Rosenpass is disabled.
+func (e *Engine) applyRosenpassKeyExchange(msg *sProto.Message, oa *peer.OfferAnswer) {
+	if e.rpManager == nil {
+		return
+	}
+	cfg := msg.GetBody().GetRosenpassConfig()
+	if cfg == nil {
+		return
+	}
+
+	remoteWgKey := msg.GetKey()
+	oa.RosenpassPubKey = e.rpManager.ResolveRemotePubKey(remoteWgKey, cfg.GetRosenpassPubKey(), cfg.GetRosenpassPubKeyHash())
+	e.rpManager.SetRemoteAck(remoteWgKey, cfg.GetAcknowledgedRosenpassPubKeyHash())
+}
+
 // RunHealthProbes executes health checks for Signal, Management, Relay, and WireGuard services
 // and updates the status recorder with the latest states.
 //

client/internal/peer/conn.go

@@ -65,6 +65,20 @@ type WgConfig struct {
 	PreSharedKey *wgtypes.Key
 }
 
+// RosenpassKeyResolver lets the handshaker fill the fingerprint/cache fields of an
+// offer/answer without depending on the Rosenpass manager directly. Implemented by
+// rosenpass.Manager and wired in by the engine.
+type RosenpassKeyResolver interface {
+	// LocalPubKeyHash is the SHA256 of our own Rosenpass public key.
+	LocalPubKeyHash() []byte
+	// RemotePubKeyAck is the SHA256 of the remote peer's cached key (nil if we do
+	// not hold it), sent back as an acknowledgement.
+	RemotePubKeyAck(remoteWgKey string) []byte
+	// RemoteHasLocalKey reports whether the peer already holds our key, so the full
+	// key may be omitted.
+	RemoteHasLocalKey(remoteWgKey string) bool
+}
+
 type RosenpassConfig struct {
 	// RosenpassPubKey is this peer's Rosenpass public key
 	PubKey []byte
@@ -72,6 +86,10 @@ type RosenpassConfig struct {
 	Addr string
 
 	PermissiveMode bool
+
+	// KeyResolver drives fingerprint-based key caching over signalling. Nil when
+	// Rosenpass is disabled, which makes the handshaker always send the full key.
+	KeyResolver RosenpassKeyResolver
 }
 
 // ConnConfig is a peer Connection configuration

client/internal/peer/handshaker.go

@@ -33,8 +33,13 @@ type OfferAnswer struct {
 	// Version of NetBird Agent
 	Version string
 	// RosenpassPubKey is the Rosenpass public key of the remote peer when receiving this message
-	// This value is the local Rosenpass server public key when sending the message
+	// This value is the local Rosenpass server public key when sending the message.
+	// May be empty on send when the remote peer has acknowledged already holding it (see RosenpassPubKeyAck).
 	RosenpassPubKey []byte
+	// RosenpassPubKeyHash is the SHA256 of the sender's own RosenpassPubKey. Always set when Rosenpass is enabled.
+	RosenpassPubKeyHash []byte
+	// RosenpassPubKeyAck is the SHA256 of the remote peer's key the sender holds cached; empty means "send it in full".
+	RosenpassPubKeyAck []byte
 	// RosenpassAddr is the Rosenpass server address (IP:port) of the remote peer when receiving this message
 	// This value is the local Rosenpass server address when sending the message
 	RosenpassAddr string
@@ -209,11 +214,11 @@ func (h *Handshaker) sendAnswer() error {
 
 func (h *Handshaker) buildOfferAnswer() OfferAnswer {
 	answer := OfferAnswer{
-		WgListenPort:    h.config.LocalWgPort,
-		Version:         version.NetbirdVersion(),
-		RosenpassPubKey: h.config.RosenpassConfig.PubKey,
-		RosenpassAddr:   h.config.RosenpassConfig.Addr,
+		WgListenPort:  h.config.LocalWgPort,
+		Version:       version.NetbirdVersion(),
+		RosenpassAddr: h.config.RosenpassConfig.Addr,
 	}
+	h.setRosenpassPubKey(&answer)
 
 	if h.ice != nil && h.RemoteICESupported() {
 		uFrag, pwd := h.ice.GetLocalUserCredentials()
@@ -230,6 +235,30 @@ func (h *Handshaker) buildOfferAnswer() OfferAnswer {
 	return answer
 }
 
+// setRosenpassPubKey fills the Rosenpass key fields of an outgoing offer/answer.
+// With a resolver wired it advertises our key hash and the ack for the remote key
+// we hold, and includes the full public key only when the peer has not yet
+// acknowledged holding it. Without a resolver (Rosenpass disabled, or an older
+// code path) it always sends the full key, preserving the previous behaviour.
+func (h *Handshaker) setRosenpassPubKey(answer *OfferAnswer) {
+	localKey := h.config.RosenpassConfig.PubKey
+	if len(localKey) == 0 {
+		return
+	}
+
+	resolver := h.config.RosenpassConfig.KeyResolver
+	if resolver == nil {
+		answer.RosenpassPubKey = localKey
+		return
+	}
+
+	answer.RosenpassPubKeyHash = resolver.LocalPubKeyHash()
+	answer.RosenpassPubKeyAck = resolver.RemotePubKeyAck(h.config.Key)
+	if !resolver.RemoteHasLocalKey(h.config.Key) {
+		answer.RosenpassPubKey = localKey
+	}
+}
+
 func (h *Handshaker) updateRemoteICEState(offer *OfferAnswer) {
 	hasICE := offer.hasICECredentials()
 	prev := h.remoteICESupported.Swap(hasICE)

client/internal/peer/handshaker_rosenpass_test.go

@@ -0,0 +1,65 @@
+package peer
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/require"
+)
+
+type fakeRPResolver struct {
+	localHash []byte
+	ack       []byte
+	hasLocal  bool
+}
+
+func (f fakeRPResolver) LocalPubKeyHash() []byte              { return f.localHash }
+func (f fakeRPResolver) RemotePubKeyAck(string) []byte        { return f.ack }
+func (f fakeRPResolver) RemoteHasLocalKey(remote string) bool { return f.hasLocal }
+
+func TestSetRosenpassPubKey_NoResolverAlwaysSendsFullKey(t *testing.T) {
+	localKey := []byte{1, 2, 3}
+	h := &Handshaker{config: ConnConfig{RosenpassConfig: RosenpassConfig{PubKey: localKey}}}
+
+	var a OfferAnswer
+	h.setRosenpassPubKey(&a)
+
+	require.Equal(t, localKey, a.RosenpassPubKey)
+	require.Nil(t, a.RosenpassPubKeyHash)
+	require.Nil(t, a.RosenpassPubKeyAck)
+}
+
+func TestSetRosenpassPubKey_ResolverIncludesFullKeyUntilAcked(t *testing.T) {
+	localKey := []byte{1, 2, 3}
+	res := fakeRPResolver{localHash: []byte{9}, ack: []byte{8}, hasLocal: false}
+	h := &Handshaker{config: ConnConfig{Key: "peerA", RosenpassConfig: RosenpassConfig{PubKey: localKey, KeyResolver: res}}}
+
+	var a OfferAnswer
+	h.setRosenpassPubKey(&a)
+
+	require.Equal(t, localKey, a.RosenpassPubKey, "full key must be sent until the peer acks it")
+	require.Equal(t, []byte{9}, a.RosenpassPubKeyHash)
+	require.Equal(t, []byte{8}, a.RosenpassPubKeyAck)
+}
+
+func TestSetRosenpassPubKey_ResolverOmitsFullKeyOnceAcked(t *testing.T) {
+	localKey := []byte{1, 2, 3}
+	res := fakeRPResolver{localHash: []byte{9}, ack: []byte{8}, hasLocal: true}
+	h := &Handshaker{config: ConnConfig{Key: "peerA", RosenpassConfig: RosenpassConfig{PubKey: localKey, KeyResolver: res}}}
+
+	var a OfferAnswer
+	h.setRosenpassPubKey(&a)
+
+	require.Nil(t, a.RosenpassPubKey, "full key must be omitted once the peer holds it")
+	require.Equal(t, []byte{9}, a.RosenpassPubKeyHash)
+	require.Equal(t, []byte{8}, a.RosenpassPubKeyAck)
+}
+
+func TestSetRosenpassPubKey_DisabledSetsNothing(t *testing.T) {
+	h := &Handshaker{config: ConnConfig{RosenpassConfig: RosenpassConfig{}}}
+
+	var a OfferAnswer
+	h.setRosenpassPubKey(&a)
+
+	require.Nil(t, a.RosenpassPubKey)
+	require.Nil(t, a.RosenpassPubKeyHash)
+}

client/internal/peer/signaler.go

@@ -61,11 +61,13 @@ func (s *Signaler) signalOfferAnswer(offerAnswer OfferAnswer, remoteKey string,
 			UFrag: offerAnswer.IceCredentials.UFrag,
 			Pwd:   offerAnswer.IceCredentials.Pwd,
 		},
-		RosenpassPubKey: offerAnswer.RosenpassPubKey,
-		RosenpassAddr:   offerAnswer.RosenpassAddr,
-		RelaySrvAddress: offerAnswer.RelaySrvAddress,
-		RelaySrvIP:      offerAnswer.RelaySrvIP,
-		SessionID:       sessionIDBytes,
+		RosenpassPubKey:     offerAnswer.RosenpassPubKey,
+		RosenpassPubKeyHash: offerAnswer.RosenpassPubKeyHash,
+		RosenpassPubKeyAck:  offerAnswer.RosenpassPubKeyAck,
+		RosenpassAddr:       offerAnswer.RosenpassAddr,
+		RelaySrvAddress:     offerAnswer.RelaySrvAddress,
+		RelaySrvIP:          offerAnswer.RelaySrvIP,
+		SessionID:           sessionIDBytes,
 	})
 	if err != nil {
 		return err

client/internal/rosenpass/cache_test.go

@@ -0,0 +1,62 @@
+package rosenpass
+
+import (
+	"bytes"
+	"testing"
+
+	"github.com/stretchr/testify/require"
+)
+
+func newCacheTestManager(spk []byte) *Manager {
+	return &Manager{
+		spk:               spk,
+		remotePubKeys:     make(map[string][]byte),
+		remoteHasLocalKey: make(map[string]bool),
+	}
+}
+
+func TestResolveRemotePubKey(t *testing.T) {
+	m := newCacheTestManager([]byte{0x01, 0x02})
+	full := bytes.Repeat([]byte{0xAB}, 64)
+
+	// a received full key is cached and returned
+	require.Equal(t, full, m.ResolveRemotePubKey("peerA", full, nil))
+
+	// a later hash-only message resolves from the cache
+	require.Equal(t, full, m.ResolveRemotePubKey("peerA", nil, rawRosenpassKeyHash(full)))
+
+	// hash mismatch is a cache miss
+	require.Nil(t, m.ResolveRemotePubKey("peerA", nil, bytes.Repeat([]byte{0x01}, 32)))
+
+	// no key and no hash (remote without Rosenpass) resolves to nil
+	require.Nil(t, m.ResolveRemotePubKey("peerB", nil, nil))
+}
+
+func TestRemotePubKeyAck(t *testing.T) {
+	m := newCacheTestManager([]byte{0x01})
+
+	// unknown peer -> no ack (signals "send me the full key")
+	require.Nil(t, m.RemotePubKeyAck("peerA"))
+
+	full := bytes.Repeat([]byte{0x09}, 48)
+	m.ResolveRemotePubKey("peerA", full, nil)
+	require.Equal(t, rawRosenpassKeyHash(full), m.RemotePubKeyAck("peerA"))
+}
+
+func TestSetRemoteAckAndRemoteHasLocalKey(t *testing.T) {
+	m := newCacheTestManager(bytes.Repeat([]byte{0x07}, 100))
+
+	require.False(t, m.RemoteHasLocalKey("peerA"))
+
+	// an ack matching our own key hash marks the peer as holding our key
+	m.SetRemoteAck("peerA", m.LocalPubKeyHash())
+	require.True(t, m.RemoteHasLocalKey("peerA"))
+
+	// empty ack clears it
+	m.SetRemoteAck("peerA", nil)
+	require.False(t, m.RemoteHasLocalKey("peerA"))
+
+	// a non-matching ack does not count
+	m.SetRemoteAck("peerA", bytes.Repeat([]byte{0x01}, 32))
+	require.False(t, m.RemoteHasLocalKey("peerA"))
+}

client/internal/rosenpass/manager.go

@@ -29,6 +29,11 @@ func hashRosenpassKey(key []byte) string {
 	return hex.EncodeToString(hasher.Sum(nil))
 }
 
+func rawRosenpassKeyHash(key []byte) []byte {
+	sum := sha256.Sum256(key)
+	return sum[:]
+}
+
 // rpServer is the subset of rp.Server used by Manager. Defined as an interface
 // so tests can substitute a mock without spinning up a real UDP server.
 type rpServer interface {
@@ -51,6 +56,15 @@ type Manager struct {
 	lock         sync.Mutex
 	port         int
 	wgIface      PresharedKeySetter
+
+	// remotePubKeys caches remote peers' full Rosenpass public keys keyed by their
+	// WireGuard public key, so a peer that already sent us its (large) key over
+	// signalling need only send its hash on subsequent offers/answers. RAM only —
+	// never persisted (1000 peers x ~512KB would be ~512MB on disk).
+	remotePubKeys map[string][]byte
+	// remoteHasLocalKey tracks, per remote WireGuard key, whether that peer has
+	// acknowledged holding our current Rosenpass public key, letting us omit it.
+	remoteHasLocalKey map[string]bool
 }
 
 // NewManager creates a new Rosenpass manager. localWgKey is the local
@@ -85,8 +99,10 @@ func NewManager(preSharedKey *wgtypes.Key, wgIfaceName string, localWgKey wgtype
 		// nil receiver in addPeer -> m.rpWgHandler.AddPeer. generateConfig will
 		// replace it with a fresh handler on each Run() to clear stale peer
 		// state from previous engine sessions.
-		rpWgHandler: NewNetbirdHandler((*[32]byte)(preSharedKey), localWgKey),
-		lock:        sync.Mutex{},
+		rpWgHandler:       NewNetbirdHandler((*[32]byte)(preSharedKey), localWgKey),
+		lock:              sync.Mutex{},
+		remotePubKeys:     make(map[string][]byte),
+		remoteHasLocalKey: make(map[string]bool),
 	}, nil
 }
 
@@ -99,6 +115,68 @@ func (m *Manager) GetAddress() *net.UDPAddr {
 	return &net.UDPAddr{Port: m.port}
 }
 
+// LocalPubKeyHash returns the raw SHA256 of the local Rosenpass public key. It is
+// advertised on every offer/answer so the remote peer can tell (via its cache)
+// whether it already holds our full key.
+func (m *Manager) LocalPubKeyHash() []byte {
+	return rawRosenpassKeyHash(m.spk)
+}
+
+// RemotePubKeyAck returns the SHA256 of the remote peer's cached public key, used
+// as the acknowledgement we send back. Nil means we do not hold the peer's key,
+// which signals the peer to include its full key next time.
+func (m *Manager) RemotePubKeyAck(remoteWgKey string) []byte {
+	m.lock.Lock()
+	defer m.lock.Unlock()
+
+	key, ok := m.remotePubKeys[remoteWgKey]
+	if !ok {
+		return nil
+	}
+	return rawRosenpassKeyHash(key)
+}
+
+// RemoteHasLocalKey reports whether the remote peer acknowledged holding our
+// current public key, so we may omit the full key from the next offer/answer.
+func (m *Manager) RemoteHasLocalKey(remoteWgKey string) bool {
+	m.lock.Lock()
+	defer m.lock.Unlock()
+
+	return m.remoteHasLocalKey[remoteWgKey]
+}
+
+// ResolveRemotePubKey reconciles the Rosenpass key material from a received
+// offer/answer: it caches a received full key, or — when only a hash was sent —
+// returns the cached key matching that hash. It returns nil when the remote peer
+// does not use Rosenpass (no key, no hash) or on a cache miss (hash sent but not
+// held); a miss self-heals because our resulting empty ack makes the peer resend
+// its full key.
+func (m *Manager) ResolveRemotePubKey(remoteWgKey string, full, hash []byte) []byte {
+	m.lock.Lock()
+	defer m.lock.Unlock()
+
+	if len(full) > 0 {
+		m.remotePubKeys[remoteWgKey] = full
+		return full
+	}
+	if len(hash) == 0 {
+		return nil
+	}
+	if cached, ok := m.remotePubKeys[remoteWgKey]; ok && bytes.Equal(rawRosenpassKeyHash(cached), hash) {
+		return cached
+	}
+	return nil
+}
+
+// SetRemoteAck records whether the remote peer's acknowledgement matches our
+// current public key hash, i.e. whether it already holds our key.
+func (m *Manager) SetRemoteAck(remoteWgKey string, ack []byte) {
+	m.lock.Lock()
+	defer m.lock.Unlock()
+
+	m.remoteHasLocalKey[remoteWgKey] = len(ack) > 0 && bytes.Equal(ack, rawRosenpassKeyHash(m.spk))
+}
+
 // addPeer adds a new peer to the Rosenpass server
 func (m *Manager) addPeer(rosenpassPubKey []byte, rosenpassAddr string, wireGuardIP string, wireGuardPubKey string) error {
 	// Defense in depth against issue #4341 (Android crash): if Run() has not