rfc1063.txt

来自「RFC 的详细文档！」· 文本 代码 · 共 619 行 · 第 1/2 页

Network Working Group                                      J.  Mogul
Request For Comments: 1063                                   C. Kent
                                                                 DEC
                                                        C. Partridge
                                                                 BBN
                                                       K. McCloghrie
                                                                 TWG
                                                           July 1988


                        IP MTU Discovery Options

STATUS OF THIS MEMO

   A pair of IP options that can be used to learn the minimum MTU of a
   path through an internet is described, along with its possible uses.
   This is a proposal for an Experimental protocol.  Distribution of
   this memo is unlimited.

INTRODUCTION

   Although the Internet Protocol allows gateways to fragment packets
   that are too large to forward, fragmentation is not always desirable.
   It can lead to poor performance or even total communication failure
   in circumstances that are surprisingly common.  (For a thorough
   discussion of this issue, see [1]).

   A datagram will be fragmented if it is larger than the Maximum
   Transmission Unit (MTU) of some network along the path it follows.
   In order to avoid fragmentation, a host sending an IP datagram must
   ensure that the datagram is no larger than the Minimum MTU (MINMTU)
   over the entire path.

   It has long been recognized that the methods for discovering the
   MINMTU of an IP internetwork path are inadequate.  The methods
   currently available fall into two categories: (1) choosing small MTUs
   to avoid fragmentation or (2) using additional probe packets to
   discover when fragmentation will occur.  Both methods have problems.

   Choosing MTUs requires a balance between network utilization (which
   requires the use of the largest possible datagram) and fragmentation
   avoidance (which in the absence of knowledge about the network path
   encourages the use of small, and thus too many, datagrams).  Any
   choice for the MTU size, without information from the network, is
   likely to either fail to properly utilize the network or fail to
   avoid fragmentation.

   Probe packets have the problem of burdening the network with



Mogul, Kent, Partridge, & McCloghrie                            [Page 1]

RFC 1063                IP MTU Discovery Options               July 1988


   unnecessary packets.  And because network paths often change during
   the lifetime of a TCP connection, probe packets will have to be sent
   on a regular basis to detect any changes in the effective MINMTU.

   Implementors sometimes mistake the TCP MSS option as a mechanism for
   learning the network MINMTU.  In fact, the MSS option is only a
   mechanism for learning about buffering capabilities at the two TCP
   peers.  Separate provisions must be made to learn the IP MINMTU.

   In this memo, we propose two new IP options that, when used in
   conjunction will permit two peers to determine the MINMTU of the
   paths between them.  In this scheme, one option is used to determine
   the lowest MTU in a path; the second option is used to convey this
   MTU back to the sender (possibly in the IP datagram containing the
   transport acknowledgement to the datagram which contained the MTU
   discovery option).

OPTION FORMATS

   Probe MTU Option (Number 11)

      Format

              +--------+--------+--------+--------+
              |00001011|00000100|   2 octet value |
              +--------+--------+--------+--------+

      Definition

      This option always contains the lowest MTU of all the networks
      that have been traversed so far by the datagram.

      A host that sends this option must initialize the value field to
      be the MTU of the directly-connected network.  If the host is
      multi-homed, this should be for the first-hop network.

      Each gateway that receives a datagram containing this option must
      compare the MTU field with the MTUs of the inbound and outbound
      links for the datagram.  If either MTU is lower than the value in
      the MTU field of the option, the option value should be set to the
      lower MTU.  (Note that gateways conforming to RFC-1009 may not
      know either the inbound interface or the outbound interface at the
      time that IP options are processed.  Accordingly, support for this
      option may require major gateway software changes).

      Any host receiving a datagram containing this option should
      confirm that value of the MTU field of the option is less than or
      equal to that of the inbound link, and if necessary, reduce the



Mogul, Kent, Partridge, & McCloghrie                            [Page 2]

RFC 1063                IP MTU Discovery Options               July 1988


      MTU field value, before processing the option.

      If the receiving host is not able to accept datagrams as large as
      specified by the value of the MTU field of the option, then it
      should reduce the MTU field to the size of the largest datagram it
      can accept.

   Reply MTU Option (Number 12)

      Format

              +--------+--------+--------+--------+
              |00001100|00000100|   2 octet value |
              +--------+--------+--------+--------+

      Definition

      This option is used to return the value learned from a Probe MTU
      option to the sender of the Probe MTU option.

RELATION TO TCP MSS

   Note that there are two superficially similar problems in choosing
   the size of a datagram.  First, there is the restriction [2] that a
   host not send a datagram larger than 576 octets unless it has
   assurance that the destination is prepared to accept a larger
   datagram.  Second, the sending host should not send a datagram larger
   than MINMTU, in order to avoid fragmentation.  The datagram size
   should normally be the minimum of these two lower bounds.

   In the past, the TCP MSS option [3] has been used to avoid sending
   packets larger than the destination can accept.  Unfortunately, this
   is not the most general mechanism; it is not available to other
   transport layers, and it cannot determine the MINMTU (because
   gateways do not parse TCP options).

   Because the MINMTU returned by a probe cannot be larger than the
   maximum datagram size that the destination can accept, this IP option
   could, in theory, supplant the use of the TCP MSS option, providing
   an economy of mechanism.  (Note however, that some researchers
   believe that the value of the TCP MSS is distinct from the path's
   MINMTU.  The MSS is the upper limit of the data size that the peer
   will accept, while the MINMTU represents a statement about the data
   size supported by the path).

   Note that a failure to observe the MINMTU restriction is not normally
   fatal; fragmentation will occur, but this is supposed to work.  A
   failure to observe the TCP MSS option, however, could be fatal



Mogul, Kent, Partridge, & McCloghrie                            [Page 3]

RFC 1063                IP MTU Discovery Options               July 1988


   because it might lead to datagrams that can never be accepted by the
   destination.  Therefore, unless and until the Probe MTU option is
   universally implemented, at least by hosts, the TCP MSS option must
   be used as well.

IMPLEMENTATION APPROACHES

   Who Sends the Option

      There are at least two ways to implement the MTU discovery scheme.
      One method makes the transport layer responsible for MTU
      discovery; the other method makes the IP layer responsible for MTU
      discovery.  A host system should support one of the two schemes.

   Transport Discovery

      In the transport case, the transport layer can include the Probe
      MTU option in an outbound datagram.  When a datagram containing
      the Probe MTU option is received, the option must be passed up to
      the receiving transport layer, which should then acknowledge the
      Probe with a Reply MTU option in the next return datagram.  Note
      that because the options are placed on unreliable datagrams, the
      original sender will have to resend Probes (possibly once per
      window of data) until it receives a Reply option.  Also note that
      the Reply MTU option may be returned on an IP datagram for a
      different transport protocol from which it was sent (e.g., TCP
      generated the probe but the Reply was received on a UDP datagram).

   IP Discovery

      A better scheme is to put MTU discovery into the IP layer, using
      control mechanisms in the routing cache.  Whenever an IP datagram
      is sent, the IP layer checks in the routing cache to see if a
      Probe or Reply MTU option needs to be inserted in the datagram.
      Whenever a datagram containing either option is received, the
      information in those options is placed in the routing cache.

      The basic working of the protocol is somewhat complex.  We trace
      it here through one round-trip.  Implementors should realize that
      there may be cases where both options are contained in one
      datagram.  For the purposes of this exposition, the sender of the
      probe is called the Probe-Sender and the receiver, Probe-Receiver.

      When the IP layer is asked to send a Probe MTU option (see the
      section below on when to probe), it makes some record in the
      routing cache that indicates the next IP datagram to Probe-
      Receiver should contain the Probe MTU option.




Mogul, Kent, Partridge, & McCloghrie                            [Page 4]

RFC 1063                IP MTU Discovery Options               July 1988


      When the next IP datagram to Probe-Receiver is sent, the Probe MTU
      option is inserted.  The IP layer in Probe-Sender should continue
      to send an occasional Probe MTU in subsequent datagrams until a
      Reply MTU option is received.  It is strongly recommended that the
      Probe MTU not be sent in all datagrams but only at such a rate
      that, on average, one Probe MTU will be sent per round-trip
      interval.  (Another way of saying this is that we would hope that
      only one datagram in a transport protocol window worth of data has
      the Probe MTU option set).  This mechanism might be implemented by
      sending every Nth packet, or, in those implementations where the
      round-trip time estimate to the destination is cached with the
      route, once every estimated RTT.

      When a Probe MTU option is received by Probe-Receiver, the
      receiving IP should place the value of this option in the next
      datagram it sends back to Probe-Sender.  The value is then
      discarded.  In other words, each Probe MTU option causes the Reply
      MTU option to be placed in one return datagram.

      When Probe-Sender receives the Reply MTU option, it should check
      the value of the option against the current MINMTU estimate in the
      routing cache.  If the option value is lower, it becomes the new
      MINMTU estimate.  If the option value is higher, Probe-Sender
      should be more conservative about changing the MINMTU estimate.
      If a route is flapping, the MINMTU may change frequently.  In such
      situations, keeping the smallest MINMTU of various routes in use
      is preferred.  As a result, a higher MINMTU estimate should only
      be accepted after a lower estimate has been permitted to "age" a
      bit.  In other words, if the probe value is higher than the
      estimated MINMTU, only update the estimate if the estimate is
      several seconds old or more.  Finally, whenever the Probe-Sender
      receives a Reply MTU option, it should stop retransmitting probes
      to Probe-Receiver.

      A few additional issues complicate this discussion.

      One problem is setting the default MINMTU when no Reply MTU
      options have been received.  We recommend the use of the minimum
      of the supported IP datagram size (576 octets) and the connected
      network MTU for destinations not on the local connected network,
      and the connected network MTU for hosts on the connected network.

      The MINMTU information, while kept by the Internet layer, is in
      fact, only of interest to the transport and higher layers.
      Accordingly, the Internet layer must keep the transport layer
      informed of the current value of the estimated MINMTU.
      Furthermore, minimal transport protocols, such as UDP, must be
      prepared to pass this information up to the transport protocol



Mogul, Kent, Partridge, & McCloghrie                            [Page 5]

RFC 1063                IP MTU Discovery Options               July 1988


      user.

      It is expected that there will be a transition period during which
      some hosts support this option and some do not.  As a result,
      hosts should stop sending Probe MTU options and refuse to send any
      further options if it does not receive either a Probe MTU option
      or Reply MTU option from the remote system after a certain number
      of Probe MTU options have been sent.  In short, if Probe-Sender
      has sent several probes but has gotten no indication that Probe-
      Receiver supports MTU probing, then Probe-Sender should assume
      that Probe-Receiver does not support probes.  (Obviously, if
      Probe-Sender later receives a probe option from Probe-Receiver, it
      should revise its opinion.)

      Implementations should not assume that routes to the same
      destination that have a different TOS have the same estimated
      MINMTU.  We recommend that the MTU be probed separately for each
      TOS.

   Respecting the TCP MSS

      One issue concerning TCP MSS is that it is usually negotiated
      assuming an IP header that contains no options.  If the transport
      layer is sending maximum size segments, it may not leave space for