Summary

  • RFC 3988 made an LSP’s advertised MTU the minimum of local and selected downstream constraints, not a property that one router could declare for the whole path.
  • A decrease had to travel quickly because stale optimism could discard oversized traffic; an increase could wait because stale caution mainly left capacity unused.

In February 2005, RFC 3988 proposed an Experimental extension to the Label Distribution Protocol. Its status record, errata record and Datatracker history preserve the exact problem. LDP could distribute labels, but it did not tell the ingress how large a packet an entire Label Switched Path could carry. Operators otherwise supplied that fact statically or through an offline system. A mistaken large value could let an edge router inject a packet that an interior label-switching router silently dropped.

The extension added one MTU Type-Length-Value field to a Label Mapping for a Forwarding Equivalence Class. The apparent simplicity concealed an important allocation of authority. A physical interface did not announce a universal answer. Each node calculated from what it knew locally, what selected downstream nodes claimed, and which neighbours the forwarding table actually used.

That last condition separated adjacency from consequence. A peer could exchange LDP messages yet not carry packets for the FEC. RFC 3988 defined downstream LSRs as the selected subset of peers that would forward the traffic. Only those paths belonged in the minimum. A control session was evidence of communication with a neighbour; it was not evidence that the neighbour sat on the active data path.

The arithmetic began with several differently scoped receipts. Link MTU included the IP header, payload and label stack, but excluded lower-layer headers. A “link” could itself be an interface, a GRE or IPsec tunnel, or another LSP. Hop MTU covered the usable limit between one upstream and one downstream LSR, taking the minimum when multiple forwarding links connected them. LSP MTU covered all valid forwarding paths from the calculating node toward the egresses.

RFC 3032, with its status, errata and Datatracker record, supplies the label-stack accounting. An ordinary MPLS label consumes four octets. RFC 3988’s worked path therefore turns a 1,500-octet link into a 1,496-octet hop for one additional label. A tunnel carrying another LSP can reduce the usable figure again. The number is not simply the physical port’s frame size; it is a claim about the packet as encapsulated at that point.

The composition algorithm flowed upstream. An egress began with 65,535 as a sentinel. For each selected downstream LSR, the current node took the smaller of the local Hop MTU and the downstream advertised LSP MTU. It then took the minimum across all selected downstreams. If a downstream Mapping lacked the extension, its advertised contribution was treated as 65,535, leaving known local constraints to do the bounding. That sentinel was not a measurement that the unknown path carried giant packets.

Policy could be conservative. A node was allowed to advertise less than it calculated, but never more. The rule reveals the direction of acceptable error. Understatement could waste capacity. Overstatement could admit packets that vanished beyond the ingress. The distributed minimum was a safety envelope, not a throughput promise.

RFC 3988 then made time part of the evidence. A changed downstream advertisement or a changed set of forwarding next hops forced recomputation. If the result fell, the node should re-advertise immediately. If it rose, the node could hold the update down. The old smaller value remained safe while the new topology settled; the old larger value did not. “Fresh enough” therefore depended on which direction reality had moved.

The unknown-TLV rule exposed another subtle boundary. Both propagation bits were set. A router that did not recognise the MTU extension had to ignore it for local processing and forward it upstream. The RFC openly conceded that this could produce an incorrect computation, while preserving the maximum available information for nodes that did understand it. Transparent carriage preserved provenance; it did not convert the intermediary into a verifier.

That behaviour came from the LDP machinery in RFC 3036, its status, errata and history. RFC 5036, with status, errata and Datatracker, later replaced the base LDP specification. Neither record proves that a particular network implemented RFC 3988. They define the envelope within which an opaque attribute could survive mixed capability.

At the ingress, the calculated LSP MTU became the next-hop network limit for the IPv4 logic in RFC 1191, its status, errata and Datatracker history. A permitted IPv4 packet could be fragmented; a DF packet that was too large had to be dropped with an ICMP fragmentation-needed report. That is an ingress reaction to a control-plane envelope. It still does not prove that a fitting packet crossed every hop, reached the endpoint or was accepted by an application.

RFC 3209, its status, errata and history supplied the RSVP-TE tunnel setting that RFC 3988 could treat as another link. Tunnels made the model recursive: a path limit could depend on another path limit plus another label. The operational ledger therefore needed the selected downstream set, each local aperture, encapsulation depth, received claim, claim age and computed result—not just one number labelled “MTU.”

RFC 3988’s historical lesson is an asymmetry of evidence. When reality becomes more restrictive, delay creates unsafe optimism. When it becomes less restrictive, delay creates conservative underuse. A control plane that understands that difference can prioritise the update that prevents loss without pretending that its final number is a delivery receipt.

Sources