6.4 TCP Layer Actions

6.4 TCP Layer Actions

The TCP layer must track the PMTU for the destination of a connection; it should not send datagrams that would be larger than this. A simple implementation could ask the IP layer for this value (using the GET_MAXSIZES interface described in [1]) each time it created a new segment, but this could be inefficient. Moreover, TCP implementations that follow the "slow-start" congestion-avoidance algorithm [4] typically calculate and cache several other values derived from the PMTU. It may be simpler to receive asynchronous notification when the PMTU changes, so that these variables may be updated.

A TCP implementation must also store the MSS value received from its peer (which defaults to 536), and not send any segment larger than this MSS, regardless of the PMTU. In 4.xBSD-derived implementations, this requires adding an additional field to the TCP state record.

Finally, when a Datagram Too Big message is received, it implies that a datagram was dropped by the router that sent the ICMP message. It is sufficient to treat this as any other dropped segment, and wait until the retransmission timer expires to cause retransmission of the segment. If the PMTU Discovery process requires several steps to estimate the right PMTU, this could delay the connection by many round-trip times.

Alternatively, the retransmission could be done in immediate response to a notification that the Path MTU has changed, but only for the specific connection specified by the Datagram Too Big message. The datagram size used in the retransmission should, of course, be no larger than the new PMTU.

Note: One MUST not retransmit in response to every Datagram Too Big message, since a burst of several oversized segments will give rise to several such messages and hence several retransmissions of the same data. If the new estimated PMTU is still wrong, the process repeats, and there is an exponential growth in the number of superfluous segments sent!

This means that the TCP layer must be able to recognize when a Datagram Too Big notification actually decreases the PMTU that it has already used to send a datagram on the given connection, and should ignore any other notifications.

Modern TCP implementations incorporate "congestion avoidance" and "slow-start" algorithms to improve performance [4]. Unlike a retransmission caused by a TCP retransmission timeout, a retransmission caused by a Datagram Too Big message should not change the congestion window. It should, however, trigger the slow-start mechanism (i.e., only one segment should be retransmitted until acknowledgements begin to arrive again).

TCP performance can be reduced if the sender's maximum window size is not an exact multiple of the segment size in use (this is not the congestion window size, which is always a multiple of the segment size). In many system (such as those derived from 4.2BSD), the segment size is often set to 1024 octets, and the maximum window size (the "send space") is usually a multiple of 1024 octets, so the proper relationship holds by default. If PMTU Discovery is used, however, the segment size may not be a submultiple of the send space, and it may change during a connection; this means that the TCP layer may need to change the transmission window size when PMTU Discovery changes the PMTU value. The maximum window size should be set to the greatest multiple of the segment size (PMTU - 40) that is less than or equal to the sender's buffer space size.

PMTU Discovery does not affect the value sent in the TCP MSS option, because that value is used by the other end of the connection, which may be using an unrelated PMTU value.