Key Change Strategies for TCP-MD5
RFC 4808

• Status
• IESG evaluation record
• IESG writeups
• Email expansions
• History

• Versions
• 04

Network Working Group                                        S. Bellovin
Request for Comments: 4808                           Columbia University
Category: Informational                                       March 2007

                   Key Change Strategies for TCP-MD5

Status of This Memo

   This memo provides information for the Internet community.  It does
   not specify an Internet standard of any kind.  Distribution of this
   memo is unlimited.

Copyright Notice

   Copyright (C) The IETF Trust (2007).

Abstract

   The TCP-MD5 option is most commonly used to secure BGP sessions
   between routers.  However, changing the long-term key is difficult,
   since the change needs to be synchronized between different
   organizations.  We describe single-ended strategies that will permit
   (mostly) unsynchronized key changes.

Bellovin                     Informational                      [Page 1]
RFC 4808                   TCP-MD5 Key Change                 March 2007

1.  Introduction

   The TCP-MD5 option [RFC2385] is most commonly used to secure BGP
   sessions between routers.  However, changing the long-term key is
   difficult, since the change needs to be synchronized between
   different organizations.  Worse yet, if the keys are out of sync, it
   may break the connection between the two routers, rendering repair
   attempts difficult.

   The proper solution involves some sort of key management protocol.
   Apart from the complexity of such things, RFC 2385 was not written
   with key changes in mind.  In particular, there is no KeyID field in
   the option, which means that even a key management protocol would run
   into the same problem.

   Fortunately, a heuristic permits key change despite this protocol
   deficiency.  The change can be installed unilaterally at one end of a
   connection; it is fully compatible with the existing protocol.

1.1.  Terminology

   The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
   "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this
   document are to be interpreted as described in [RFC2119].

2.  The Algorithm

   Separate algorithms are necessary for transmission and reception.
   Reception is easier; we explain it first.

2.1.  Reception

   A receiver has a list of valid keys.  Each key has a (conceptual)
   timestamp associated with it.  When a segment arrives, each key is
   tried in turn.  The segment is discarded if and only if it cannot be
   validated by any key in the list.

   In principle, there is no need to test keys in any particular order.
   For performance reasons, though, a simple most-recently-used (MRU)
   strategy -- try the last valid key first -- should work well.  More
   complex mechanisms, such as examining the TCP sequence number of an
   arriving segment to see whether it fits in a hole, are almost
   certainly unnecessary.  On the other hand, validating that a received
   segment is putatively legal, by checking its sequence number against
   the advertised window, can help avoid denial of service attacks.

   The newest key that has successfully validated a segment is marked as
   the "preferred" key; see below.

Bellovin                     Informational                      [Page 2]
RFC 4808                   TCP-MD5 Key Change                 March 2007

   Implicit in this scheme is the assumption that older keys will
   eventually be unneeded and can be removed.  Accordingly,
   implementations SHOULD provide an indication of when a key was last
   used successfully.

2.2.  Transmission

   Transmission is more complex, because the sender does not know which
   keys can be accepted at the far end.  Accordingly, the conservative
   strategy is to delay using any new keys for a considerable amount of
   time, probably measured in days.  This time interval is the amount of
   asynchronicity the parties wish to permit; it is agreed upon out of
   band and configured manually.

   Some automation is possible, however.  If a key has been used
   successfully to validate an incoming segment, clearly the other side
   knows it.  Accordingly, any key marked as "preferred" by the
   receiving part of a stack SHOULD be used for transmissions.

   A sophisticated implementation could try alternate keys if the TCP
   retransmission counter gets too high.  (This is analogous to dead
   gateway detection.)  In particular, if a key change has just been
   attempted but such segments are not acknowledged, it is reasonable to
   fall back to the previous key and issue an alert of some sort.
   Similarly, an implementation with a new but unused key could
   occasionally try to use it, much in the way that TCP implementations
   probe closed windows.  Doing this avoids the "silent host" problem
   discussed in Section 3.1.  This should be done at a moderately slow
   rate.

   Note that there is an ambiguity when an acknowledgment is received
   for a segment transmitted with two different keys.  The TCP Timestamp