rfc963.txt

RFC 相关的技术文档

Network Working Group                                 Deepinder P. Sidhu
Request for Comments: 963                          Iowa State University
                                                           November 1985

              SOME PROBLEMS WITH THE SPECIFICATION OF THE
                  MILITARY STANDARD INTERNET PROTOCOL


STATUS OF THIS MEMO

   The purpose of this RFC is to provide helpful information on the
   Military Standard Internet Protocol (MIL-STD-1777) so that one can
   obtain a reliable implementation of this protocol standard.
   Distribution of this note is unlimited.

ABSTRACT

   This paper points out several significant problems in the
   specification of the Military Standard Internet Protocol
   (MIL-STD-1777, dated August 1983 [MILS83a]).  These results are based
   on an initial investigation of this protocol standard.  The problems
   are: (1) a failure to reassemble fragmented messages completely; (2)
   a missing state transition; (3) errors in testing for reassembly
   completion; (4) errors in computing fragment sizes; (5) minor errors
   in message reassembly; (6) incorrectly computed length for certain
   datagrams.  This note also proposes solutions to these problems.

1.  Introduction

   In recent years, much progress has been made in creating an
   integrated set of tools for developing reliable communication
   protocols.  These tools provide assistance in the specification,
   verification, implementation and testing of protocols.  Several
   protocols have been analyzed and developed using such tools.
   Examples of automated verification and implementation of several real
   world protocols are discussed in [BLUT82] [BLUT83] [SIDD83] [SIDD84].

   We are currently working on the automatic implementation of the
   Military Standard Internet Protocol (IP).  This analysis will be
   based on the published specification [MILS83a] of IP dated 12 August
   1983.

   While studying the MIL Standard IP specification, we have noticed
   numerous errors in the specification of this protocol.  One
   consequence of these errors is that the protocol will never deliver
   fragmented incoming datagrams; if this error is corrected, such
   datagrams will be missing some data and their lengths will be
   incorrectly reported.  In addition, outgoing datagrams that are
   divided into fragments will be missing some data.  The proof of these
   statements follows from the specification of IP [MILS83a] as
   discussed below.


Sidhu                                                           [Page 1]



RFC 963                                                    November 1985
Some Problems with MIL-STD IP


2.  Internet Protocol

   The Internet Protocol (IP) is a network layer protocol in the DoD
   protocol hierarchy which provides communication across interconnected
   packet-switched networks in an internetwork environment.  IP provides
   a pure datagram service with no mechanism for reliability, flow
   control, sequencing, etc.  Instead, these features are provided by a
   connection-oriented protocol, DoD Transmission Control Protocol (TCP)
   [MILS83b], which is implemented in the layer above IP.  TCP is
   designed to operate successfully over channels that are inherently
   unreliable, i.e., which can lose, damage, duplicate, and reorder
   packets.

   Over the years, DARPA has supported specifications of several
   versions of IP; the last one appeared in [POSJ81].  A few years ago,
   the Defense Communications Agency decided to standardize IP for use
   in DoD networks.  For this purpose, the DCA supported formal
   specification of this protocol, following the design discussed in
   [POSJ81] and the technique and organization defined in [SDC82].  A
   detailed specification of this protocol, given in [MILS83a], has been
   adopted as the DoD standard for the Internet Protocol.

   The specification of IP state transitions is organized into decision
   tables; the decision functions and action procedures are specified in
   a subset of Ada[1], and may employ a set of machine-specific data
   structures.  Decision tables are supplied for the pairs <state name,
   interface event> as follows: <inactive, send from upper layer>,
   <inactive, receive from lower layer>, and <reassembling, receive from
   lower layer>.  To provide an error indication in the case that some
   fragments of a datagram are received but some are missing, a decision
   table is also supplied for the pair <reassembling, reassembly time
   limit elapsed>.  (The event names are English descriptions and not
   the names employed by [MILS83a].)

3.  Problems with MIL Standard IP

   One of the major functions of IP is the fragmentation of datagrams
   that cannot be transmitted over a subnetwork in one piece, and their
   subsequent reassembly.  The specification has several problems in
   this area.  One of the most significant is the failure to insert the
   last fragment of an incoming datagram; this would cause datagrams to
   be delivered to the upper-level protocol (ULP) with some data
   missing. Another error in this area is that an incorrect value of the
   data length for reassembled datagrams is passed to the ULP, with
   unpredictable consequences.

   As the specification [MILS83a] is now written, these errors are of


Sidhu                                                           [Page 2]



RFC 963                                                    November 1985
Some Problems with MIL-STD IP


   little consequence, since the test for reassembly completion will
   always fail, with the result that reassembled datagrams would never
   be delivered at all.

   In addition, a missing row in one of the decision tables creates the
   problem that network control (ICMP) messages that arrive in fragments
   will never be processed.  Among the other errors are the possibility
   that a few bytes will be discarded from each fragment transmitted and
   certain statements that will create run-time exceptions instead of
   performing their intended functions.

   A general problem with this specification is that the program
   language and action table portions of the specification were clearly
   not checked by any automatic syntax checking process.  Variable and
   procedure names are occasionally misspelled, and the syntax of the
   action statements is often incorrect.  We have enumerated some of
   these problems below as a set of cautionary notes to implementors,
   but we do not claim to have listed them all.  In particular, syntax
   errors are only discussed when they occur in conjunction with other
   problems.

   The following section discusses some of the serious errors that we
   have discovered with the MIL standard IP [MIL83a] during our initial
   study of this protocol.  We also propose corrections to each of these
   problems.

4.  Detailed Discussion of the Problems

   Problem 1: Failure to Insert Last Fragment

      This problem occurs in the decision table corresponding to the
      state reassembling and the input "receive from lower layer"
      [MILS83a, sec 9.4.6.1.3].  The problem occurs in the following row
      of this table:[2]

      ________________________________________________________
      check-    SNP      TTL    where    a     reass    ICMP
       sum     params   valid    to     frag   done    check-
      valid?   valid?     ?       ?      ?       ?      sum?
      __________________________________________________________________
      YES      YES      YES     ULP    YES     YES      d      reass_
                                                               delivery;
                                                               state :=
                                                                INACTIVE
      __________________________________________________________________

      The reass_done function, as will be seen below, returns YES if the


Sidhu                                                           [Page 3]



RFC 963                                                    November 1985
Some Problems with MIL-STD IP


      fragment just received is the last fragment needed to assemble a
      complete datagram and NO otherwise.  The action procedure
      reass_delivery simply delivers a completely reassembled datagram
      to the upper-level protocol.  It is the action procedure
      reassemble that inserts an incoming fragment into the datagram
      being assembled.  Since this row does not call reassemble, the
      result will be that every incoming fragmented datagram will be
      delivered to the upper layer with one fragment missing.  The
      solution is to rewrite this row of the table as follows:

      ________________________________________________________
      check-    SNP      TTL    where    a     reass    ICMP
       sum     params   valid    to     frag   done    check-
      valid?   valid?     ?       ?      ?       ?      sum?
      __________________________________________________________________
      YES      YES      YES     ULP    YES     YES      d    reassemble;
                                                               reass_
                                                               delivery;
                                                               state :=
                                                                INACTIVE
      __________________________________________________________________

      Incidentally, the mnemonic value of the name of the reass_done
      function is questionable, since at the moment this function is
      called datagram reassembly cannot possibly have been completed.  A
      better name for this function might be last_fragment.

   Problem 2: Missing State Transition

      This problem is the omission of a row of the same decision table
      [MILS83a, sec 9.4.6.1.3].  Incoming packets may be directed to an
      upper-level protocol (ULP), or they may be network control
      messages, which are marked ICMP (Internet Control Message
      Protocol).  When control messages have been completely assembled,
      they are processed by an IP procedure called analyze.  The
      decision table contains the row

      ________________________________________________________
      check-    SNP      TTL    where    a     reass    ICMP
       sum     params   valid    to     frag   done    check-
      valid?   valid?     ?       ?      ?       ?      sum?
      __________________________________________________________________
      YES      YES      YES    ICMP    YES     NO       d    reassemble;
      __________________________________________________________________





Sidhu                                                           [Page 4]



RFC 963                                                    November 1985
Some Problems with MIL-STD IP


      but makes no provision for the case in which where_to returns
      ICMP, a_frag returns YES, and reass_done returns YES.  An
      additional row should be inserted, which reads as follows:

      ________________________________________________________
      check-    SNP      TTL    where    a     reass    ICMP
       sum     params   valid    to     frag   done    check-
      valid?   valid?     ?       ?      ?       ?      sum?
      __________________________________________________________________
      YES      YES      YES    ICMP    YES     YES      d    reassemble;
                                                               analyze;
                                                               state :=
                                                                INACTIVE
      __________________________________________________________________

      Omitting this row means that incoming fragmented ICMP messages
      will never be analyzed, since the state machine does not have any
      action specified when the last fragment is received.

   Problem 3: Errors in reass_done

      The function reass_done, as can be seen from the above, determines
      whether the incoming subnetwork packet contains the last fragment
      needed to complete the reassembly of an IP datagram.  In order to
      understand the errors in this function, we must first understand
      how it employs its data structures.

      The reassembly of incoming fragments is accomplished by means of a
      bit map maintained separately for each state machine.  Since all
      fragments are not necessarily the same length, each bit in the map
      represents not a fragment, but a block, that is, a unit of eight
      octets.  Each fragment, with the possible exception of the "tail"
      fragment (we shall define this term below), is an integral number
      of consecutive blocks. Each fragment's offset from the beginning
      of the datagram is given, in units of blocks, by a field in the
      packet header of each incoming packet.  The total length of each
      fragment, including the fragment's header, is specified in the
      header field total_length; this length is given in octets.  The
      length of the header is specified in the field header_length; this
      length is given in words, that is, units of four octets.

      In analyzing this subroutine, we must distinguish between the
      "tail" fragment and the "last" fragment.  We define the last
      fragment as the one which is received last in time, that is, the
      fragment that permits reassembly to be completed.  The tail
      fragment is the fragment that is spatially last, that is, the
      fragment that is spatially located after any other fragment.  The


Sidhu                                                           [Page 5]



RFC 963                                                    November 1985
Some Problems with MIL-STD IP


      length and offset of the tail fragment make it possible to compute
      the length of the entire datagram.  This computation is actually
      done in the action procedure reassembly, and the result is saved
      in the state vector field total_data_length; if the tail fragment
      has not been received, this value is assumed to be zero.

      It is the task of the reass_done function [MILS83a, sec 9.4.6.2.6]
      to determine whether the incoming fragment is the last fragment.
      This determination is made as follows:

         1) If the tail fragment has not been received previously and
         the incoming fragment is not the tail fragment, then return NO.

         2) Otherwise, if the tail fragment has not been received, but
         the incoming fragment is the tail fragment, determine whether
         all fragments spatially preceding the tail fragment have also
         been received.

         3) Otherwise, if the tail fragment has been received earlier,
         determine whether the incoming fragment is the last one needed
         to complete reassembly.

      The evaluation of case (2) is accomplished by the following
      statment:

         if (state_vector.reassembly_map from 0 to
           (((from_SNP.dtgm.total_length -
               (from_SNP.dtgm.header_length * 4) + 7) / 8)
           + 7) / 8 is set)
         then return YES;

      The double occurrence of the subexpression " + 7 ) / 8" is
      apparently a misprint.  The function f(x) = (x + 7) / 8 will
      convert x from octets to blocks, rounding any remainder upward.
      There is no need for this function to be performed twice.  The
      second problem is that the fragment_offset field of the incoming
      packet is ignored.  The tail fragment specifies only its own
      length, not the length of the entire datagram; to determine the
      latter, the tail fragment's offset must be added to the tail
      fragment's own length.  The third problem hinges on the meaning of
      the English "... from ... to ..." phrase.  If this phrase has the
      same meaning as the ".." range indication in Ada [ADA83, sec 3.6],
      that is, includes both the upper and lower bounds, then it is
      necessary to subtract 1 from the final expression.

      The expression following the word to, above, should thus be
      changed to read


Sidhu                                                           [Page 6]



RFC 963                                                    November 1985
Some Problems with MIL-STD IP


         from_SNP.dtgm.fragment_offset +
             ((from_SNP.dtgm.total_length -
                 (from_SNP.dtgm.header_length * 4) + 7) / 8) - 1

      Another serious problem with this routine occurs when evaluating
      case (3).  In this case, the relevant statement is

         if (all reassembly map from 0 to
           (state_vector.total_data_length + 7)/8 is set
         then return YES

      If the tail fragment was received earlier, the code asks, in
      effect, whether all the bits in the reassembly map have been set.
      This, however, will not be the case even if the incoming fragment