6.2 Storing PMTU Information

6.2 Storing PMTU Information

In general, the IP layer should associate each PMTU value that it has learned with a specific path. A path is identified by a source address, a destination address and an IP type-of-service. (Some implementations do not record the source address of paths; this is acceptable for single-homed hosts, which have only one possible source address.)

Note: Some paths may be further distinguished by different security classifications. The details of such classifications are beyond the scope of this memo.

The obvious place to store this association is as a field in the routing table entries. A host will not have a route for every possible destination, but it should be able to cache a per-host route for every active destination. (This requirement is already imposed by the need to process ICMP Redirect messages.)

When the first packet is sent to a host for which no per-host route exists, a route is chosen either from the set of per-network routes, or from the set of default routes. The PMTU fields in these route entries should be initialized to be the MTU of the associated first-hop data link, and must never be changed by the PMTU Discovery process. (PMTU Discovery only creates or changes entries for per-host routes). Until a Datagram Too Big message is received, the PMTU associated with the initially-chosen route is presumed to be accurate.

When a Datagram Too Big message is received, the ICMP layer determines a new estimate for the Path MTU (either from a non-zero Next-Hop MTU value in the packet, or using the method described in section 5). If a per-host route for this path does not exist, then one is created (almost as if a per-host ICMP Redirect is being processed; the new route uses the same first-hop router as the current route). If the PMTU estimate associated with the per-host route is higher than the new estimate, then the value in the routing entry is changed.

The packetization layers must be notified about decreases in the PMTU. Any packetization layer instance (for example, a TCP connection) that is actively using the path must be notified if the PMTU estimate is decreased.

Note: even if the Datagram Too Big message contains an Original Datagram Header that refers to a UDP packet, the TCP layer must be notified if any of its connections use the given path.

Also, the instance that sent the datagram that elicited the Datagram Too Big message should be notified that its datagram has been dropped, even if the PMTU estimate has not changed, so that it may retransmit the dropped datagram.

Note: The notification mechanism can be analogous to the mechanism used to provide notification of an ICMP Source Quench message. In some implementations (such as 4.2BSD-derived systems), the existing notification mechanism is not able to identify the specific connection involved, and so an additional mechanism is necessary.

Alternatively, an implementation can avoid the use of an asynchronous notification mechanism for PMTU decreases by postponing notification until the next attempt to send a datagram larger than the PMTU estimate. In this approach, when an attempt is made to SEND a datagram with the DF bit set, and the datagram is larger than the PMTU estimate, the SEND function should fail and return a suitable error indication. This approach may be more suitable to a connectionless packetization layer (such as one using UDP), which (in some implementations) may be hard to "notify" from the ICMP layer. In this case, the normal timeout-based retransmission mechanisms would be used to recover from the dropped datagrams.

It is important to understand that the notification of the packetization layer instances using the path about the change in the PMTU is distinct from the notification of a specific instance that a packet has been dropped. The latter should be done as soon as practical (i.e., asynchronously from the point of view of the packetization layer instance), while the former may be delayed until a packetization layer instance wants to create a packet. Retransmission should be done for only for those packets that are known to be dropped, as indicated by a Datagram Too Big message.