fips/docs/design/fips-session-layer.md

# FIPS Session Protocol (FSP)

The FIPS Session Protocol is the topmost layer of the FIPS protocol stack.
It sits above the FIPS Mesh Protocol (FMP) and below applications (native
FIPS API or IPv6 adapter). FSP provides end-to-end authenticated, encrypted
datagram delivery between any two FIPS nodes, regardless of how many
intermediate hops separate them.

## Role

FSP manages end-to-end communication sessions between FIPS nodes identified by
their public keys (npubs). Each session provides:

- **End-to-end encryption**: Payload confidentiality independent of how many
  intermediate nodes handle the traffic
- **Mutual authentication**: Both parties prove they control the private key
  for their claimed identity
- **Replay protection**: Counter-based nonces with sliding window, tolerant of
  UDP packet loss and reordering
- **Transport independence**: Sessions survive transport changes, route
  changes, and address changes — they are bound to npub identities, not to
  transport paths

FSP is a datagram session protocol. It provides encrypted datagrams, not
reliable streams. There is no FIPS equivalent of TCP; if applications need
reliability, ordering, or flow control, they provide it themselves (typically
by running TCP over the FIPS IPv6 adapter).

## Services Provided to Applications

Applications access the FIPS mesh through two interfaces, both served by FSP:

### Native FIPS API

Applications address destinations directly by npub or public key. The FIPS
stack resolves the destination's node_addr, establishes or reuses a session,
encrypts the payload, and routes through FMP. No DNS involvement.

### IPv6 Adapter

Unmodified IPv6 applications use a TUN device with `fd00::/8` routing. A local
DNS service maps npub → IPv6 address and primes the identity cache. Packets
arriving at the TUN are translated to FIPS datagrams and routed through FSP.

See [fips-ipv6-adapter.md](fips-ipv6-adapter.md) for the IPv6 adaptation
layer.

### Port-Based Service Dispatch

FSP DataPackets carry a 4-byte port header (source and destination port) inside
the AEAD envelope, enabling multiple services to share a single session. The
IPv6 adapter runs on port 256; the native FIPS API and future services
(gateways, application protocols) register on other ports. Port dispatch is
internal to the FSP layer — services see only their payload, not the port
header.

### What Applications Get

- **Authenticated datagram delivery**: Each datagram is encrypted and
  authenticated with session keys bound to both parties' npubs
- **Session transparency**: Sessions are established on demand and maintained
  automatically. Applications send packets; FSP handles session setup,
  encryption, and teardown.
- **Endpoint identity**: Applications address destinations by npub. The FIPS
  address is the public key.

### What Applications Do Not Get

- **Reliability**: Datagrams may be lost, duplicated, or delivered out of
  order. FSP provides no retransmission or ordering.
- **Path MTU discovery**: FSP does not signal MTU to applications. The IPv6
  adapter handles MTU enforcement via ICMP Packet Too Big and TCP MSS
  clamping.
- **Congestion control**: FSP does not throttle traffic. Applications running
  TCP over IPv6 get TCP's congestion control; native API applications must
  manage their own sending rate.

## Services Required from FMP

FSP treats FMP as a black box providing three services. FSP knows nothing about
transports, transport addresses, links, peers, spanning trees, coordinates,
bloom filters, hop counts, or network topology.

### SessionDatagram Forwarding

FMP accepts a SessionDatagram (source node_addr, destination node_addr, TTL,
path MTU, payload) and delivers it best-effort toward the destination. Delivery
may traverse multiple hops, each with independent link encryption.

### Error Signaling

FMP signals routing failures asynchronously:

- **CoordsRequired**: A transit node lacks the destination's tree coordinates.
  FSP responds by sending a standalone CoordsWarmup (0x14) message
  (rate-limited), re-initiating discovery, and resetting the coordinate warmup
  counter.
- **PathBroken**: Greedy routing reached a dead end. FSP responds by sending
  a standalone CoordsWarmup (rate-limited), re-discovering the destination's
  current coordinates, and resetting the warmup counter.
- **MtuExceeded**: A transit node cannot forward a SessionDatagram because
  the packet exceeds the next-hop link MTU; FSP adjusts its
  session-layer path MTU estimate from the reported bottleneck. See
  [fips-mtu.md](fips-mtu.md) for the full forward/reverse MTU model.

All three signals are generated by transit nodes (not the destination) and
travel back to the source inside a new SessionDatagram. They are plaintext
(not end-to-end encrypted) because transit nodes have no session with the
source.

### Local Delivery

When a SessionDatagram arrives with a destination node_addr matching the local
node, FMP delivers it to FSP for session-layer processing.

## Session Lifecycle

### Session Establishment

Sessions are established on demand when the first datagram needs to be sent to
a destination with no existing session.

FSP uses Noise XK for session key agreement (Noise Protocol Framework;
Perrin 2018). The initiator knows the destination's npub (required for
XK's pre-message `s` token); the responder learns the initiator's
identity from msg3 (not msg1, unlike IK at the link layer). This
provides stronger initiator identity hiding — the initiator's static
key is encrypted under the established shared secret rather than under
only the responder's static key.

The handshake is a three-message flow carried in SessionSetup, SessionAck,
and SessionMsg3:

1. **Initiator** sends SessionSetup containing Noise XK msg1 (ephemeral key
   only) and both parties' tree coordinates
2. **Responder** processes msg1, sends SessionAck containing Noise XK msg2
   (ephemeral key + encrypted epoch) and both parties' tree coordinates.
   The responder transitions to AwaitingMsg3 state.
3. **Initiator** processes msg2, sends SessionMsg3 containing the encrypted
   static key and encrypted epoch. Both parties derive identical symmetric
   session keys and the session is established.

Each side's epoch (an 8-byte random value generated at startup) is
exchanged encrypted in msg2 and msg3. On subsequent handshakes, an epoch
mismatch indicates the peer has restarted, triggering session
re-establishment — similar to IKEv2's INITIAL_CONTACT notification for
peer restart detection (RFC 7296 §2.4).

Packets that trigger session establishment are queued (with bounded buffer)
while a session or a coordinate lookup is in progress for the destination,
and transmitted after the session is established. When neither is in
progress, a TUN packet is answered at once with ICMPv6 Destination
Unreachable instead, and held packets are answered the same way once
nothing is working to deliver them. They are also answered once their
destination has been held longer than the lookup schedule plus the
handshake timeout (50 seconds by default). That age counts from when the
destination was first held or, if later, from the latest lookup or session
this node started for it, or from the first packet that rode a lookup
already pending for it; later packets riding the same lookup do not restart
it. A packet held before a lookup and then a session started for its
destination can therefore wait longer than 50 seconds, since each start is
new work that ends by its own timeout. A destination with no lookup
running (no peer's bloom filter reaches it, only a non-tree peer's does, or
its backoff window is open) is still held while the node has no peers and
for 30 seconds after it gains one, while routing information arrives.
Outside that window, a packet to a destination that is a direct peer or has
a next hop, whose session setup failed on the link, is dropped unanswered
rather than answered, so a retransmit retries the session.

### Self-Bootstrapping

SessionSetup is self-bootstrapping for routing. It carries the source's and
destination's tree coordinates in the clear (not inside the Noise payload).
As the message transits intermediate nodes, each node caches these coordinates,
warming the path for subsequent data packets that carry only addresses (no
coordinates).

SessionAck carries both the responder's and initiator's coordinates back
along the reverse path, warming caches in the other direction. This ensures
return-path transit nodes can route even when the reverse path diverges from
the forward path (e.g., after tree reconvergence).

### Simultaneous Initiation

When both nodes send a SessionSetup to each other at the same time
("crossing hellos"), the handshake initiated by the node with the smaller
address wins, as in IKEv2's simultaneous initiation (RFC 7296 §2.8). A
SessionSetup proves nothing about its sender until its msg3 arrives, so
the rule is applied in a way a forged setup cannot use to block the
address owner's session:

- The winner (smaller address) drops the peer's setup while its own setup
  is less than 5 s old, since the peer will answer it. If its own setup
  has gone unanswered for 5 s, it gives up its own leg and answers the
  peer's setup instead.
- The loser (larger address) answers the peer's setup and keeps its own
  initiator leg beside it, with its resends paused. The resends resume
  once the peer's setup has been held for 5 s with nothing completed, or
  at once if no half-open setup from the peer is left.

Each setup a node answers is held as a half-open responder setup with its
own handshake and SessionAck, up to four per address. If the loser's own
leg completes first, it keeps the winner's half-open setups beside its
session, and a msg3 completing one of them replaces its session, so both
sides end on the same handshake. The winner, once established, keeps none.

Each half-open setup expires on its own deadline, a handshake timeout
(30 s by default) after it was admitted, including those the loser keeps
beside an established session. With four held, a new setup replaces the
oldest one that is at least 5 s old, preferring one that arrived over the
same link peer, and is refused if none is that old. A forged setup
therefore cannot be evicted before it is 5 s old.

### Data Transfer

Once established, sessions carry encrypted data using the FSP pipeline. Each
encrypted message includes:

- A 12-byte cleartext header (used as AEAD AAD) containing the counter and
  flags (including the CP flag for coordinate cache warming)
- Optional cleartext coordinates when the CP flag is set
- An AEAD-encrypted payload containing a 6-byte inner header (session-relative
  timestamp, message type, inner flags) followed by DataPacket content: a 4-byte
  port header (src_port, dst_port) and the service payload. The receiver
  dispatches by destination port to the registered service handler.

### Session Idle Timeout

Sessions that see no traffic for a configurable duration (default 90s) are
torn down. When traffic resumes, a new session is established automatically.

Only DataPacket (message type 0x10) send/receive and session establishment
reset `last_activity`. MMP traffic — SenderReport, ReceiverReport, and
PathMtuNotification — does not reset the idle timer. This means a session
carrying only MMP reports and no application data will still tear down after
`node.session.idle_timeout_secs`. MMP reports continue flowing until the
session is torn down, providing final measurement data for the teardown log.

The idle timeout is deliberately shorter than the coordinate cache TTL (300s).
This ordering ensures that when traffic stops and the session tears down, the
transit node coordinate caches are still warm when a new session is established.
The fresh SessionSetup re-warms the caches, maintaining routing continuity.

### Session Independence from Transport

Sessions exist above the routing layer and are bound to npub identities, not
transport addresses or routing paths — following WireGuard's approach
(Donenfeld 2017) of binding cryptographic sessions to identity keys rather
than network addresses. A session survives:

- Transport failover (UDP → Ethernet → back to UDP)
- Route changes (different intermediate hops)
- Transport address changes (IP address or port changes)
- Topology changes (direct peer becomes multi-hop or vice versa)

## End-to-End Encryption

### Noise XK Pattern

FSP uses Noise XK for session encryption, distinct from the Noise IK
pattern used at the link layer. The full Noise descriptor is
`Noise_XK_secp256k1_ChaChaPoly_SHA256`, with one deviation recorded in
[the security reference](../reference/security.md): the handshake AEAD
uses an empty associated-data field where standard Noise
`EncryptAndHash` uses the handshake hash `h`.

The XK pattern (pre-message: `← s`):

- **msg1** (`→ e, es`): Initiator sends ephemeral key only. The initiator's
  static identity is not revealed in this message.
- **msg2** (`← e, ee`): Responder sends ephemeral key and encrypted epoch.
- **msg3** (`→ s, se`): Initiator sends encrypted static key and encrypted
  epoch. Both parties now share identical session keys.

After the handshake, Noise produces two directional symmetric keys
(`send_key`, `recv_key`) used with ChaCha20-Poly1305 for all subsequent data.

The XK pattern requires the initiator to know the responder's static key
in advance (the `← s` pre-message), which is satisfied by the discovery
or DNS lookup that precedes session establishment. In exchange, XK
provides stronger initiator identity protection than IK — the initiator's
static key is encrypted under the full shared secret (after three DH
operations) rather than under only the responder's static key.

### Cryptographic Primitives

FSP uses ChaCha20-Poly1305 with secp256k1 ECDH; see
[../reference/security.md](../reference/security.md) for the full
primitive table shared with the link layer.

### secp256k1 Parity Normalization

Nostr npubs encode x-only public keys (32 bytes, no y-coordinate parity). The
Noise XK pre-message mixes the responder's static key as a 33-byte compressed
key, and the default secp256k1 ECDH hash includes a parity-dependent version
byte.

Both operations are normalized to be parity-independent: the pre-message hash
uses even parity (`0x02` prefix), and ECDH hashes only the x-coordinate of the
result point. This ensures handshakes succeed regardless of the responder's
actual key parity.

### Privacy Note

Noise XK provides stronger initiator identity protection than IK. In XK, the
initiator's static key is encrypted in msg3 under the full shared secret
(derived from three DH operations), so an attacker who compromises only the
responder's nsec cannot decrypt the initiator's identity from captured
handshake messages (they would also need the responder's ephemeral key).
This is the primary reason FSP uses XK rather than IK — session-layer
traffic traverses untrusted intermediate nodes, making initiator identity
protection more valuable than at the link layer.

### Data Packet Authentication

FSP uses AEAD authentication only — no per-packet signatures. The Noise
handshake binds session keys to both parties' static keys, so only holders of
the corresponding nsecs can derive the session keys. This provides implicit
authentication for every packet, matching WireGuard and Lightning's approach.

### Forward Secrecy

Ephemeral keys in the Noise handshake provide forward secrecy — a standard
property of Noise handshake patterns that include ephemeral key exchange.
Compromise of static keys (nsec) does not reveal past session keys, because
session keys are derived in part from ephemeral-ephemeral DH (`ee`), and
ephemeral keys are discarded after the handshake.

## Replay Protection

FSP uses explicit 8-byte counters on the wire for replay protection, adapted
from the DTLS anti-replay mechanism (Rescorla & Modadugu, RFC 6347). Each side
maintains a monotonically increasing send counter, included in the 12-byte
cleartext header of every encrypted message. The receiver maintains a sliding
window (2048-entry bitmap) tracking which counters have been seen.

This design is critical for operation over unreliable transports. Under UDP
packet loss or reordering, implicit nonce counters (where the receiver
increments on each decrypt attempt) would desynchronize permanently — a failed
`decrypt()` increments the nonce, and the desync grows with each lost packet.
Explicit counters with a sliding bitmap window — the standard DTLS approach —
allow the receiver to decrypt any packet independently, regardless of what
packets were lost or reordered.

The same `ReplayWindow` and `decrypt_with_replay_check()` implementation is
used at both the link and session layers.

## Hybrid Coordinate Warmup Strategy

Session establishment (SessionSetup/SessionAck) warms transit node coordinate
caches along the path. But coordinate caches have a finite TTL (default 300s),
and entries may be evicted under memory pressure. When a transit node's cache
entry expires, it cannot forward data packets (which carry only addresses, not
coordinates) and sends a CoordsRequired error.

FSP uses a hybrid warmup strategy combining proactive piggybacking with
reactive standalone messages to keep transit caches populated (Yggdrasil
uses a similar approach of embedding coordinates in session traffic to
warm transit node caches):

### Proactive Warmup Phase

After session establishment, the first N data packets (configurable, default 5)
per session attempt to piggyback source and destination coordinates via the CP
flag in the FSP common prefix. The coordinates appear in cleartext between the
12-byte header and the ciphertext, allowing transit nodes to cache them without
decryption.

If piggybacking coordinates would cause the total packet to exceed the
transport MTU, the source sends a standalone **CoordsWarmup** message (0x14)
first, followed by the data packet without the CP flag. This ensures transit
caches are warmed even when data packets are near the MTU limit.

### CoordsWarmup Message (0x14)

CoordsWarmup is an encrypted FSP message (phase 0x0) with the CP flag set and
inner msg_type 0x14. It carries cleartext source and destination coordinates
between the FSP header and the AEAD ciphertext, using the same format as
CP-flagged data packets. The inner body is empty — the message exists solely
to deliver coordinates to transit nodes.

Transit nodes extract coordinates via the existing `try_warm_coord_cache()`
code path with zero changes to the transit forwarding logic. CoordsWarmup is
indistinguishable from any other CP-flagged message at the transit layer.

On the wire, a CoordsWarmup message consists of the standard 12-byte FSP
header (used as AEAD AAD) with ver=0, phase=0, flags=CP, followed by
cleartext source and destination coordinates (same encoding as any
CP-flagged packet), followed by 22 bytes of AEAD ciphertext (6-byte inner
header and 16-byte Poly1305 tag). The total FSP payload is 12 bytes of
header, the coordinate data, and 22 bytes of ciphertext.

### Steady State

After the warmup count is reached, FSP clears the CP flag and sends minimal
data packets (12-byte header + ciphertext). Transit nodes serve from their
coordinate caches.

### Reactive Re-Warm

When FSP receives a CoordsRequired or PathBroken signal:

1. A standalone CoordsWarmup message is sent immediately to re-warm transit
   caches, rate-limited at one per destination per configurable interval
   (default 2000ms, `node.session.coords_response_interval_ms`)
2. The warmup counter resets — subsequent data packets piggyback coordinates
   again when possible (or send additional CoordsWarmup messages when the
   data packet would exceed the MTU)
3. A new LookupRequest may be initiated to rediscover the destination's
   current coordinates (always for PathBroken; optionally for CoordsRequired)
4. When the LookupResponse arrives for an established session, the warmup
   counter resets again (handling the timing gap where warmup packets might
   fire before transit caches are repopulated by discovery)

The source-side rate limiting prevents amplification: at most one standalone
CoordsWarmup response per destination per `coords_response_interval_ms`
(default 2s). This is independent of the transit-node rate limiting on error
signal generation (100ms per destination).

### Warmup State Machine

![Coordinate warmup state machine](diagrams/session-warmup-fsm.svg)

## Identity Cache

The IPv6 adapter requires an identity cache to map `fd00::/8` addresses
back to `(NodeAddr, PublicKey)` for routing; see
[fips-ipv6-adapter.md](fips-ipv6-adapter.md#identity-cache) for the
cache rationale, eviction policy, and population mechanics.

## Coordinate Cache

The coordinate cache maps `NodeAddr → TreeCoordinate` and is the critical
data structure that enables efficient multi-hop routing. Without cached
coordinates for a destination, FMP cannot make forwarding decisions and must
either fall back to bloom-filter-only routing or signal CoordsRequired.

### Unified Cache

The coordinate cache is a single unified cache (merged from previously
separate coord_cache and route_cache). All coordinate sources — SessionSetup
transit, CP-flagged data packets, LookupResponse — write to the same cache.

### Eviction Policy

- **TTL-based expiration**: Entries expire after a configurable duration
  (default 300s)
- **Refresh on use**: Active routing through a cache entry resets its TTL,
  keeping hot entries alive
- **LRU eviction**: When the cache is full, the least recently used entry is
  evicted, skipping entries whose verification (from a lookup whose proof
  this node checked) is still live; a hint is refused only when every entry
  is live-verified. An expired entry is not sought out first: every hint
  carries the same TTL, so the least recently used entry is also the first
  to expire. The victim is found through an ordered index, so an insert at
  capacity costs O(log n) rather than a scan of the cache, and ties are
  broken by insertion order, never by address
- **Address bound**: Besides its entry count, the cache bounds the coordinate
  addresses it stores at 16 per configured entry. Each entry is charged its
  coordinate length or 16, whichever is larger, and a write that would exceed
  the bound evicts like a write to a full cache. Deep coordinates still cache
  and route; they reduce how many entries fit. A coordinate heavier than the
  whole bound is stored alone
- **Per-link-peer budget**: Each authenticated link peer pays for the hints it
  writes through this node, in units of the weight above, at 1,024 units a
  second up to 4,096; a newly seen peer starts with none, and a peer keeps
  what it accrues while it stays connected. A new destination the peer can pay
  for goes into the cache's main segment, or into probation if every main
  entry is live-verified. One it cannot pay for goes into a
  probation segment of `max(1, coord_size / 32)` entries on top, which evicts
  only its own least recently used entries, so the frame is still routed and
  nothing a paid write placed is displaced. A later write of the same
  destination that pays its whole weight, and at least the most weight
  probation has held for it, promotes it; using an entry does not. A write
  of an expired entry is placed as a new destination's would be. An update
  is charged only for depth beyond the most the entry has been charged for,
  and one that is not paid for keeps the old coordinates; making an entry
  shallower does not lower its charge, so a forged shallow rewrite
  does not make the honest correction cost anything. Writes this node's own
  handshakes make (the SessionAck's source coordinates) and verified lookup
  results are not charged. The `coord_hint_unfunded` counter counts unpaid
  writes
- **Flush on parent change**: When the local node's tree parent changes, the
  entire coordinate cache is flushed. Tree parent changes mean the node's own
  coordinates have changed, making cached coordinates for other nodes
  potentially stale for routing purposes.

### Timer Ordering

Cache and session timers are ordered so that idle sessions tear down before
transit caches expire:

| Timer | Default | Purpose |
| ----- | ------- | ------- |
| Session idle timeout | 90s | Tear down unused sessions |
| Coordinate cache TTL | 300s | Expire stale coordinates |
| DNS TTL | 300s | Expire DNS resolutions |

When traffic stops: the session tears down at 90s. When traffic resumes: DNS
re-resolves the identity, a fresh SessionSetup carries coordinates, and transit
node caches (still within their 300s TTL) are re-warmed.

## Session-Layer MMP

FSP runs an MMP instance per established session for end-to-end metrics
independent of hop count. Reports are encrypted and forwarded through
every transit link, so bandwidth cost is proportional to path length;
the session-layer report intervals are correspondingly higher than the
link-layer intervals (clamped to `[500ms, 10s]` vs. `[1s, 5s]`).

The session-layer instance shares its algorithms (SRTT, jitter, loss,
ETX) and report wire format with link-layer MMP. The differences —
configuration namespace (`node.session_mmp.*`), routing scope,
send-failure backoff, idle-timeout interaction, and the
PathMtuNotification mechanism — are documented in the unified MMP
treatment at [fips-mmp.md](fips-mmp.md). For the end-to-end path-MTU
echo specifically, see [fips-mtu.md](fips-mtu.md). Reports and
PathMtuNotification do **not** reset the session idle timer, so a
session carrying only MMP traffic still tears down at the configured
idle threshold.

### MtuExceeded Handling

When FMP signals MtuExceeded (a transit node could not forward a
SessionDatagram because it exceeded the next-hop link MTU), FSP uses
the reported bottleneck MTU to adjust its session-layer path MTU
estimate immediately. See [fips-mtu.md](fips-mtu.md) for the full
reactive PMTUD mechanism.

## Implementation Status

| Feature | Status |
| ------- | ------ |
| Session establishment (Noise XK) | **Implemented** |
| Peer restart detection (epoch exchange) | **Implemented** |
| MtuExceeded handling | **Implemented** |
| End-to-end encryption (ChaCha20-Poly1305) | **Implemented** |
| Explicit counter replay protection | **Implemented** |
| Hybrid coordinate warmup (CP + CoordsWarmup) | **Implemented** |
| FSP wire format (prefix, AAD, inner header) | **Implemented** |
| Session-layer MMP (with send-failure backoff) | **Implemented** |
| Identity cache (LRU-only) | **Implemented** |
| Coordinate cache (unified, TTL + refresh) | **Implemented** |
| Session idle timeout | **Implemented** |
| CoordsRequired handling | **Implemented** |
| PathBroken handling | **Implemented** |
| Simultaneous initiation tie-breaker | **Implemented** |
| Flush coord cache on parent change | **Implemented** |
| Rekey | **Implemented** |
| Port-based service multiplexing | **Implemented** |
| IPv6 header compression (shim on port 256) | **Implemented** |
| Path MTU tracking (FMP SessionDatagram field) | **Implemented** |
| Path MTU notification (end-to-end echo) | **Implemented** |

## References

### FIPS Internal Documentation

- [fips-concepts.md](fips-concepts.md) — Protocol overview
- [fips-architecture.md](fips-architecture.md) — Layer architecture and
  identity model
- [fips-mesh-layer.md](fips-mesh-layer.md) — FMP specification (below FSP)
- [fips-ipv6-adapter.md](fips-ipv6-adapter.md) — IPv6 adaptation layer
  (above FSP)
- [fips-mesh-operation.md](fips-mesh-operation.md) — Routing, discovery,
  and error recovery
- [fips-mmp.md](fips-mmp.md) — Metrics Measurement Protocol (link + session)
- [fips-mtu.md](fips-mtu.md) — Path MTU model (PathMtuNotification,
  MtuExceeded, hysteresis)
- [fips-prior-work.md](fips-prior-work.md) — Noise XK, WireGuard,
  DTLS replay window, IKEv2 simultaneous initiation, hybrid coordinate
  warmup citations
- [../reference/wire-formats.md](../reference/wire-formats.md) — Wire
  format reference for all session message types
- [../reference/security.md](../reference/security.md) — Cryptographic
  primitives and rekey defaults

### External References

- Mogul, J., Deering, S. [RFC 1191](https://datatracker.ietf.org/doc/html/rfc1191):
  "Path MTU Discovery". 1990. *End-to-end path MTU discovery; FSP adapts this for
  overlay networks using transit-node min() propagation.*