Network Working Group                                        A. McKenzie
Request for Comments: 714                                        BBN-NCC
NIC: 35144                                                    April 1976

            A Host/Host Protocol for an ARPANET-Type Network

   Recently we have been involved in the planning of a network, which,
   if implemented, would use ARPANET IMPs without modification, but
   would allow re-specification of Host/Host (and higher level)
   Protocol.  The remainder of this document is a slightly edited
   version of our recommendation for Host/Host protocol; we thought that
   it might be of interest to the ARPANET Community.

I. INTRODUCTION

   The Host/Host Protocol for the ARPANET was the first such protocol
   designed for use over a packet-switched network.  The current version
   has been in existence since early 1972 and has provided for the
   transportation of billions of bits over tens or hundreds of thousands
   of connections.  Clearly, the protocol is adequate for the job; this
   does not mean that it is ideal, however.  In particular, the ARPANET
   Host/Host protocol has been criticized on the following grounds
   (among others):

   (1) It is specified as a simplex protocol.  Each established
       connection is a simplex entity, thus two connections (one in each
       direction) must be established in order to carry out an exchange
       of messages.  This provides great generality but at a perhaps
       unacceptable cost in complexity.

   (2) It is not particularly robust, in that it cannot continue to
       operate correctly in the face of several types of message loss.
       While it is true that the ARPANET itself rarely loses messages,
       messages are occasionally lost, both by the network and by the
       Hosts.

   (3) Partly because of the simplex nature of connections, the flow
       control mechanisms defined in the ARPANET protocol do not make
       efficient use of the transactional nature of much of data
       processing.  Rather than carrying flow control information (in
       the form of permits, or requests for more information) in the
       reverse traffic, a separate channel is set up to convey this
       information.  Thus, for transactional systems, up to twice as
       many messages are exchanged (half for flow control information
       and half for data) as would be needed for data alone.

   (4) Prohibition against the multiple use of a connection termination
       point makes the establishment of communication with service
       facilities extremely cumbersome.

   The International Federation for Information Processing (IFIP)
   Working Group 6.1 (Packet-switched Network Internetworking) has
   recently approved a proposal for an internetwork end-to-end protocol.
   The IFIP Protocol is based on experience from the ARPANET, the
   (French) Cyclade Network, and the (British) NPL Network, as well as
   the plans of other networks.  Thus, one would expect that it would
   have all of the strengths and few (or none) of the weaknesses of the
   protocols which are in use on, or planned for, these networks.

   In fact, the IFIP Protocol avoids the deficiencies of the ARPANET
   protocol mentioned above.  Connections are treated as full-duplex
   entities, and this decision permits flow control information to be
   carried on the reverse channel in transaction-oriented systems where
   there is reverse channel traffic occurring naturally.  In addition,
   the IFIP Protocol is to some extent self synchronizing; in
   particular, there is no type of message loss from which the Protocol
   does not permit recovery in a graceful way.

   The IFIP Protocol makes a minimal number of assumptions about the
   network over which it will operate.  It is designed to permit
   fragmentation, as a message crosses from one network to another,
   without network reassembly.  It anticipates duplication, or non-
   delivery, of messages or message fragments and provides ways to
   recover from these conditions.  Finally, it permits delivery of
   messages at their destination Host in a completely different order
   from the order in which they were input by the source Host.
   Unfortunately, it achieves these advantages at a relatively high
   overhead cost in terms of transferred bits.  The complete source and
   destination process addresses are carried in every message, 24-bits
   of fragment identification are carried with each fragment and 16-bits
   of acknowledgement information are else carried in every message.

   When considering channel capacities of hundreds of kilobits (or
   more), message overhead of a few hundred bits is a modest price to
   pay in order to achieve great flexibility and generality.  However,
   for a stand-alone network of the type under consideration, and
   especially in view of the anticipated use of many circuits of 10kbs
   capacity, the IFIP Protocol offers far more generality than is
   needed, for which a fairly severe overhead price is paid.

   The virtual circuit protocols currently being debated within the
   International Telegraph and Telephone Consultative Committee (CCITT)
   are a step in the opposite direction.  Virtual circuit protocols
   attempt to make a packet switching network indistinguishable (from a

   customer's point of view) from a switched circuit network, except
   possibly in regard to error or delay characteristics.  Thus, virtual
   circuit protocols generally place responsibility for end-to-end
   communications control within the network rather than within the
   Hosts.  For example, when a receiving Host limits the rate at which
   it accepts data from the network, the network in turn limits the rate
   of input from the Host which is transmitting this data stream.  Host
   protocols which are designed for virtual circuit networks can be
   quite simple, if somewhat inflexible.  For example, the Host might
   give the network a "link number" or "index" and ask the network to
   set up a virtual circuit to some other Host to be associated with
   this number, and report back if and when the circuit is established.
   However, significant development would be required to add a virtual
   circuit capability to the ARPANET IMP software; the required changes
   would seem to be more expensive and carry greater uncertainty than
   they are worth.

   In light of the above, our approach in defining this proposed
   protocol has been to start with the ARPANET Host/Host Protocol and
   modify it according to some of the concepts of the IFIP Protocol in
   order to remedy its major deficiencies.  The remainder of this
   document specifies the protocol, which we have designed for this
   purpose.

II.  COMMUNICATION CONCEPTS

   The IMP subnetwork imposes a number of physical restrictions on
   communications between Hosts.  These restrictions are presented in
   BBN Report No. 1822.  In particular, the concepts of leaders,
   messages, padding, message ID's and message types are of interest to
   the design of Host/Host Protocol.  The following discussion assumes
   that the reader is familiar with these concepts.

   The IMP subnetwork takes cognizance only of Hosts, but in general a
   Host connected to the network can support several users, several
   terminals, or several independent processes.  Since many or all of
   these users, terminals, or processes will need to use the network
   concurrently, a fundamental requirement of the Host/Host Protocol is
   to provide process-to-process communication over the network.  Thus,
   it is necessary for the Host/Host Protocol to provide a richer
   addressing structure than is required by the IMP subnetwork.

   Processes within a Host are envisioned as communicating with the rest
   of the network through a network control program (NCP) resident in
   that Host, which implements the Host/Host protocol.  The primary
   functions of an NCP are to establish connections, break connections,
   and control data flow over connections.  A connection couples two
   processes so that the output from one process is input to the other

   and vice versa.  The NCP may be implemented either as part of the
   Host's operating system or a separate user process, although it must
   have the capability of communicating with all of the processes or
   routines which are attempting to use the network.

   In order to accomplish its tasks, the NCP of one Host must
   communicate with the NCPs of other Hosts.  To this end, a particular
   communication path between each pair of Hosts has been designated as
   the control connection.  Messages transmitted over the control
   connection are called control messages, and must always be
   interpreted by an NCP as a sequence of one or more control commands.
   For example, one kind of control command is used to initiate a
   connection while another kind carries notification that a connection
   has been terminated.*

      * Note that in BBN Report No. 1822, messages of non-zero type are
        called control messages, and are used to control the flow of
        information between a Host and its IMP.  In this document the
        term "control message" is used for a message of type zero
        transmitted over the control connection.  The IMPs take no
        special notice of these messages.

   The maximum size of a message is limited by the IMP subnetwork to
   approximately 1000 8-bit bytes, and in fact may be further limited by
   the receiving Host for flow control reasons, as described later.

   Accordingly, the transmitting process, or its Network Control
   Program, must take responsibility for fragmenting long interprocess
   messages into messages of a size conforming to the Host/Host and
   Host/IMP protocols.  For this reason, it is impossible for a sending
   Host to guarantee that any significance should be attached to message
   boundaries by receiving processes.  Nevertheless, message boundaries
   will occur naturally, and should be used in a reasonable way wherever
   possible; that is, a sending process or its NCP should not act
   arbitrarily in deciding to fragment messages.  For example, this
   protocol specifies that each control message must contain an integral
   number of control commands and no single control command will be
   split into two pieces which are carried through the network in
   separate messages.

   A major concern of the Host/Host Protocol is the definition of the
   method for references to processes in other Hosts.  In order to
   facilitate this, a standard name space is used, with a separate
   portion of the name space allocated to each Host.  Each Host
   therefore must map internal process identifiers into its portion of
   this name space.  The elements of the name space are called sockets.
   A socket forms one end of a connection and a connection is fully
   specified by a pair of sockets, one in each Host.  A socket is

   identified by a Host number and a 16-bit socket number.  The same
   16-bit socket number in different Hosts represents difference
   sockets.  In order to avoid the transmission of a pair of 16-bit
   socket numbers in each message between these sockets, the process of
   connection establishment allows each Host to define a mapping, valid
   for the lifetime of the connection being established, from the 32
   bits which specify the socket pair to an 8-bit number.

   No constraints are placed on the assignment of socket numbers;
   however, since a pair of socket numbers defines a unique connection,
   it is clear that in assigning socket numbers, a Host must ensure that
   for each new connection at least one of the socket numbers is unique.
   For example, a Host which supports many terminals might choose to use
   a terminal's physical interface number as a portion of the socket
   number involved in any connection established on behalf of that
   terminal.  This would insure uniqueness at the terminal end.  Thus,
   no conflict would occur if several terminals attempted to access a
   common resource (identified by its own unique socket number).

   From the foregoing it should be clear that the Host/Host protocol
   allows a single socket to participate in several connections
   simultaneously.  This is quite similar to what happens in the
   telephone system, where a company, as well as an individual, can be
   identified with a phone number.  As seen from the outside, the phone
   number of a company is sharable, since several conversations can
   proceed at the same time and the caller does not have to worry about
   the already existing conversations.  Conversely, the phone number of
   an individual is not sharable, since he can process only one
   conversation at a time; the same is generally true of a connection to
   a terminal which might be using the network.

   A final major concept which should be explained is the "windowing"
   concept, which is used for flow control.  This concept is adapted
   from the IFIP protocol with some appropriate modifications for use in
   an ARPANET-type network.  When a connection is established, a
   sequence number is initialized to some specified starting point and
   the receiver allocates a certain number of credits to the sender.
   Each credit entitles the sender to transmit one message; that is, the
   receiver agrees to provide buffering for the number of messages
   specified by the number of credits granted.  If one thinks of
   sequence numbers advancing from left to right, the initial sequence
   number defines the left edge of a window into the entire sequence
   number space and the credit, when added to the initial sequence
   number, defines the right edge of the window.  The transmitting
   process is permitted to send as many messages and as would fill the
   window, but not more.

   When a receiver receives a message whose sequence number is at the
   left window edge (or several consecutive messages extending rightward
   from the left window edge) the receiver returns an acknowledgement
   for the rightmost such message, along with a new credit, and advances
   his own window; its new left edge immediately follows the last
   acknowledged message and it's new right edge is at the location
   defined by adding the new credit to the new left window edge.
   Similarly, when a sender receives an acknowledgement he advances his
   own left window edge to the location in the sequence number space
   specified by the acknowledgement and his own right window edge to the
   location specified by adding the new credit allocation to the left
   window edge.  Fields are reserved in each data message to carry an
   acknowledgement and a credit for traffic flowing in the reverse
   direction.  Thus, in the case of interactive or transactional
   exchanges, no control messages need to be sent.

   In the event that a sender does not receive acknowledgements for
   previously transmitted messages within some timeout period, the
   messages are transmitted again, using the same sequence number as was
   previously assigned.  This allows straightforward recovery from the
   situation of lost messages.  On the other hand, if it is the
   returning acknowledgement which is lost, the fact that the
   retransmitted message carries an identical sequence number allows the
   receiver to discard it.  However, the receiver should notice that at
   the time of retransmission the sender had not received an
   acknowledgement; therefore, the receiver should re-acknowledge this
   (and any subsequently received messages) by transmitting an
   acknowledgement bearing the current left window edge.  Thus, in both
   the case of lost data messages and the case of lost acknowledgements
   the protocol remains synchronized.

   The primary difference between this protocol and the IFIP Protocol is
   in the size of the sequence number field.  The IFIP Protocol is
   designed for interconnections of many networks with huge
   variabilities in delay and with a strong possibility that messages
   will not be delivered at the destination in the same order in which
   they were transmitted by the source.  Thus, the IFIP Protocol uses a
   16-bit sequence number field which, even at megabit per second rates
   cannot be completely cycled through in less than several hours.
   However, the proposed ARPANET-type network has the characteristic
   that delays are typically short, messages are rarely lost, and they
   are always delivered in the same order in which they were sent if
   they are delivered at all.  Therefore, this Host/Host Protocol uses
   only a 4-bit sequence number field which, of course, is cycled
   through every 16 messages.  This imposes the constraint that a window
   may never be larger than eight messages.  Since the sequence number
   is contained in a 4-bit field, it is also possible to use only four

   bits for each of the credit and acknowledgement fields; thus, this
   protocol uses only 12 bits in each message header rather than 40 bits
   used under the IFIP Protocol.

III.  NCP FUNCTIONS

   The functions of the NCP are to establish connections, terminate
   connections, control flow, transmit interrupts, and respond to test
   inquiries.  These functions are explained in this section, and
   control commands are introduced as needed.  In Section IV the formats
   of all control commands are presented together.

   Connection Establishment

      The command used to establish the connection is the RFC (request
      for connection).

          8*       16           16           8        16       8
      ----------------------------------------------------------------
      !  RFC  ! my-socket ! your-socket !  Index  !  size  ! credit  !
      ----------------------------------------------------------------

         * The number shown above each control command field is the
           length of that field in bits.

      The RFC command either requests the establishment of a connection
      between a pair of sockets or accepts a previously received request
      for connection.  Since the RFC command is used both for requesting
      and accepting the establishment of a connection, it is possible
      for either of two cooperating processes to initiate connection
      establishment.  Even if both processes were to simultaneously
      request the establishment of a connection, each would interpret
      receipt of the RFC sent by the other as an acceptance of its own
      RFC, and thus the connection would be established without
      difficulty.  The my-socket and your-socket fields in the RFC
      identify the sockets which terminate the ends of the connection at
      each Host.  The index field of the RFC specifies an index number
      which will be contained in each data transmission sent over this
      connection from the "my-socket" to the "your-socket" end of the
      connection.  The size field of the RFC specifies the maximum
      number of 8-bit bytes which are permitted to be sent from the
      "your-socket" to the "my-socket" end of the connection in any one
      message.  The credit field of the RFC specifies the initial size
      (in the range 0-7) of the window in the "your-socket" to the "my-
      socket" direction of the connection.  A pair of RFCs exchanged
      between two Hosts matches when the my-socket field of one equals
      your-socket field of the other, and vice versa.  The connection is
      established when a matching pair of RFCs has been exchanged.

      Connections are uniquely specified by the sockets which terminate
      the connection; thus, a pair of socket numbers cannot be used to
      identify two different connections simultaneously.  Similarly, the
      index is used to specify which connection a data message pertains
      to; thus, an index value cannot be reused while the connection to
      which it was first assigned is still active or in the process of
      being established.  For example, consider an RFC sent from Host A
      to Host B whose my-socket field contains the value X, your-socket
      field contains the value Y, and index contains the value Z.  Until
      the requested connection has been closed (even if it is never
      established) or reinitialized, Host A is prohibited from sending a
      different RFC to Host B whose my-socket field and your-socket
      fields are X and Y, or whose index field is Z.  Note that the
      prohibition against the reuse of the values X and Y treats them as
      a pair; that is, another RFC may be sent from Host A to Host B,
      whose my-socket field contains the value X so long as the your-
      socket field contains some value other than Y.

      In general there is no prescribed lifetime for an RFC.  A Host is
      permitted to queue incoming RFCs and withhold a response for an
      arbitrarily long time, or, alternatively, to reject requests
      immediately if it has not already sent a matching RFC.  Of course,
      the Host which originally sent the RFC may be unwilling to wait
      for an arbitrarily long time so it may abort the request.

      The decision to queue or not to queue incoming RFCs has important
      implications which must not be ignored.  Each RFC which is queued,
      of course, requires a small amount of memory in the Host doing the
      queuing.  If the incoming RFC is queued until a local process
      takes control of the local socket and accepts (or rejects) the
      RFC, but no local process ever takes control of the socket, the
      RFC must be queued "forever".  On the other hand, if no queuing is
      performed, the cooperating processes which may be attempting to
      establish communication may be able to establish this
      communication only by accident.

      The most reasonable solution to the problems posed above is for
      each NCP to give processes running in its own Host two options for
      attempting to initiate connections.  The first option would allow
      a process to cause an RFC to be sent to a specified remote socket,
      with the NCP notifying the process as to whether this RFC was
      accepted or rejected by the remote Host.  The second option would
      allow a process to tell its own NCP to "listen" for an RFC to a
      specified local socket from some remote socket (the process might
      also specify the particular remote socket and/or Host it wishes to
      communicate with) and to accept the RFC (i.e., return a matching

      RFC) if and when it arrives.  Note that this also involves queuing
      (of "listen" requests) but it is internal queuing, which is
      susceptible to reasonable management by the local Host.

   Connection Termination

      The command used to terminate a connection is CLS (close).

              8         16          16
          -----------------------------------
          !  CLS  ! my-socket ! your-socket !
          -----------------------------------

      The my-socket field and your-socket field of the CLS command
      identify the sockets which terminate the connection being closed.
      Each side must send and receive a CLS command before the
      connection termination is completed and prohibitions on the reuse
      of the socket pair and index value are ended.

      It is not necessary for connection to be established (i.e., for
      both RFCs to be exchanged) before connection termination begins.
      For example, if a Host wishes to refuse a request for connection
      it sends back a CLS instead of a matching RFC.  The refusing Host
      then waits for the initiating Host to acknowledge the refusal by
      returning a CLS.  Similarly, if a Host wishes to abort its
      outstanding request for connection it sends a CLS command.  The
      foreign Host is obliged to acknowledge the CLS with its own CLS.
      Note that even though the connection was never established, CLS
      commands must be exchanged before the prohibition on the reuse of
      the socket pair or the index is completely ended.  Under normal
      circumstances a Host should not send a CLS command for a
      connection on which that Host has unacknowledged data outstanding.
      Of course, the other Host may have just transmitted data so the
      sender of the CLS command may expect to receive additional data
      from the other Host.

      The Host should quickly acknowledge an incoming CLS so that the
      foreign Host can purge its tables.  In particular, in the absence
      of outstanding unacknowledged data a Host must acknowledge an
      incoming close within 60 seconds.  Following a 60 second period,
      the Host transmitting a CLS may regard the socket pair and the
      index as "unused" and it may delete the values from any tables
      describing active connections.  Of course, if the foreign Host
      malfunctions in such a way that the CLS is ignored for longer than
      60 seconds, subsequent attempts to establish connections or
      transmit data may lead to ambiguous results.  To deal with this
      possibility, a Host should in general "reinitialize" its use of
      connection parameters before attempting to establish a new

      connection to any Host which has failed to respond to CLS
      commands.  Methods for reinitializing connection parameter tables
      are described below.

Acknowledgement

      As described in the previous section, flow control is handled by a
      windowing scheme, based on sequence numbers.  Credits and
      acknowledgements can be piggybacked on data traveling over the
      reverse channel.  Thus, in general, acknowledgement of the receipt
      of messages will take place over the data connection rather than
      over the control connection.  However, there are some cases when
      it may be desirable to pass acknowledgements over the control
      connection (for example, when there is no data to be returned in
      the reverse direction).  In addition, for efficiency it may be
      desirable to negatively acknowledge data transmissions known not
      to have been delivered, rather than waiting for the timeout and
      retransmission mechanism to cause such messages to be
      retransmitted. [Note that such negative acknowledgement is not
      required, since timeout and retransmission is always sufficient to
      guarantee eventual delivery of all data, but may be used to
      increase the efficiency of communication.]  Since the frequency of
      use of the negative acknowledgement system over an ARPANET-type
      network will be extremely low, it is undesirable to leave space
      for negative acknowledgements in the header of every data message.
      Thus, negative acknowledgement can be most conveniently handled by
      control messages.

      There are two commands dealing with acknowledgements.

              8       8       4       4
          ---------------------------------
          !  ACK  ! index !  seq  !  crd  !
          ---------------------------------

      The ACK (acknowledgement) command carries three data fields.  The
      index value is the index used by the sender of the acknowledgement
      to identify the connection.  The sequence ("seq") field contains
      the sequence number of the highest-numbered sequential data
      message correctly received over the connection.  [The very first
      data message to be transmitted over a newly established connection
      will have the sequence number one; until this data message is
      correctly received, any acknowledgement commands transmitted for
      this connection (for example, to change the credit value) will
      have the sequence field set to zero.  This applies whether the
      "acknowledgement" is carried by an ACK command or is contained in
      data messages being sent to the foreign Host over the connection.]
      The credit ("crd") field contains a number, in the range 0-7,

      which gives the size of the receive window.  This number, when
      added to the "seq", gives the sequence number of the highest
      numbered message which is permitted to be transmitted by the
      foreign Host.  Thus, a credit of zero says that the Host
      transmitting the ACK command is currently not prepared to accept
      any messages over the connection; and a credit of 7 says the Host
      is prepared to accept up to 7 messages over the connection.  Of
      course, since the sequence number is contained in a 4-bit field,
      the addition of the sequence number and the credit value must be
      performed modulo 16 (sequence number zero immediately follows
      sequence number 15).

      As noted above, the ACK command is intended for use with data
      connections where there is no data flow in one direction, for
      example, the transmission of a file to a line printer.  In fact it
      should be clear that, since transmission of control messages is
      not synchronized with transmission of data messages (either in the
      network or, more importantly, in the transmitting NCP), ACK
      commands should not be sent for any connection over which data is
      flowing in the same direction.  Thus, if an ACK command is
      generated, the NCP which transmits it must insure that the control
      message which contains it is transmitted prior to the transmission
      of new data messages for the same connection.

              8        8       8
          --------------------------
          !  NACK  ! index !  seq  !
          --------------------------

      The NACK (negative acknowledgement) command contains two data
      fields.  As with the positive acknowledgement command described
      above, the first field is the index number assigned to this
      connection by the sender of the NACK.  However, the second field
      contains only the 4-bit sequence number, right justified in an 8-
      bit field, of the data message for the connection in question
      which is being negatively acknowledged.  As previously noted, the
      NACK serves no vital function in the protocol but may occasionally
      allow more efficient communication.  The NACK is intended to be
      used when the window width is greater than one, the message at the
      left window edge has not been correctly received, and messages
      toward the right of the window have been correctly received.  A
      timeout will eventually cause the retransmission of the missing
      message, at which point the left window edge can be moved forward
      several messages.  Use of the NACK, however, could trigger the
      immediate retransmission of the missing message and thus reduce
      the delay.  Of course, if more than one message is missing it may

      be desirable to send several NACKs for one index in a single
      control message; the protocol permits this, although it is
      extremely unlikely to occur.

   Re-initialization

      Occasionally, due to lost control messages, system crashes, NCP
      errors, or other factors, communication between two NCPs will be
      disrupted.  One possible effect of any such disruption might be
      that neither of the involved NCPs could be sure that its stored
      information regarding connections with the other Host matched the
      information stored by the NCP of the other Host.  In this
      situation, an NCP may wish to reinitialize its tables and request
      that the other Host do likewise.  This re-initialization may be
      requested for a particular index and/or socket pair, or globally
      for all connections possibly established with the other Host.  For
      these purposes, the protocol provides three control commands as
      described below:

              8        16           16          8
          -------------------------------------------
          !  RCP  ! my-socket ! your-socket ! index !
          -------------------------------------------

      The RCP (reinitialize connection parameters) command contains
      three data fields.  The my-socket and your-socket fields contain a
      pair of socket numbers, which define a connection; the index field
      contains a value which would identify data messages over a
      connection.  When this command is received by an NCP it should
      purge its tables of any reference to a connection identified by
      the socket pair or any reference to a connection for which
      received data would be identified by the specified index value; of
      course, only connections using these values with the Host sending
      the RCP would be purged.  In effect, the Host sending the RCP
      command is saying: "I am about to send you an RFC using this
      socket pair and this index to identify a data connection, which I
      hope we can agree to establish.  I do not believe that any use of
      this socket pair or this index conflicts with any previous use,
      but if you believe it does, please record the fact (for later
      examination) as an error and then delete from your tables the
      conflicting information so that we may proceed to establish the
      connection."

      In case more global difficulties or loss of state information are
      suspected, the protocol provides the pair of control commands RST
      (reset) and RRP (reset reply).

              8
          ---------
          !  RST  !
          ---------

              8
          ---------
          !  RRP  !
          ---------

      The RST command is to be interpreted by the Host receiving it as a
      signal to purge its tables of any entries which arose from
      communication with the Host which sent the RST.  The Host sending
      the RST should likewise purge its tables of any entries which
      arose from communication with the Host to which the RST was sent.
      The Host receiving the RST should acknowledge receipt by returning
      an RRP.  Once the first Host has sent an RST to the second Host,
      the first Host should not communicate with the second Host (except
      for responding to RST) until the second Host returns an RRP.  If
      both NCPs decide to send RSTs at approximately the same time, each
      Host will receive an RST and each must answer with an RRP even
      though its own RST has not been answered.

      A Host should not send an RRP when an RST has not been received.
      Further, a Host should send only one RST (and no other commands)
      in a single control message and should not send another RST to the
      same Host until either 60 seconds have elapsed or a command which
      is not an RST or RRP has been received from that Host.  Under
      these conditions, a single RRP constitutes an answer to all RSTs
      sent to that Host and any other RRPs arriving from that Host
      should be discarded.

   Interrupts

      It is sometimes necessary in a communication system to circumvent
      flow control mechanisms when serious errors or other important
      conditions are detected.  For example, the user of a time sharing
      terminal who creates and begins the execution of a program which
      contains an erroneous infinite loop may need to "attract the
      attention" of the operating system to ask it to cancel the
      execution of his program, even though the operating system may
      normally "listen" to the terminal only when the program in
      execution asks for input.  Similarly, in a computer communication
      network, where flow control may prevent the transmission of data
      from one process to another, under certain extraordinary
      conditions it may be necessary to pass a signal from one process
      to another.  Since the channel between the NCPs of two Hosts is
      not subject to the flow control mechanisms imposed on the data

      connections, it is possible to transmit such an "out-of-band"
      signal over the control connection, and for this purpose the INT
      (interrupt) command is provided.

              8       8       8
          -------------------------
          !  INT  ! index !  seq  !
          -------------------------

      The INT command contains two data fields.  The index field
      identifies the data connection to which the "interrupt" pertains;
      the sequence number ("seq"), which is four bits right-justified in
      an eight-bit field, gives the sequence number of the first data
      message which should "come after" the interrupt.  In other words,
      the INT command notifies the receiving NCP of an exception
      condition which must be synchronized with the data stream, and the
      sequence number provides the necessary synchronization.  Any data
      messages with sequence numbers to the left of the specified
      sequence number were generated before the exception condition
      arose.

      An NCP which receives an INT command should advance the right
      window edge of the specified data connection so that the window
      contains at least the sequence number specified in the interrupt
      command.  (It may be necessary to acknowledge data messages which
      were not correctly received or were not buffered in order to be
      able to advance the window to this point; justification is
      provided by the assumption that the INT was sent only because the
      flow control mechanisms were preventing the transmission of
      important information.)  Of course, the interrupt or exception
      signal itself is subject to the interpretation of the Host
      receiving the signal, but should have a meaning equivalent to:
      "notify the process in execution, or that process' superior, that
      something exceptional has happened and that the data now buffered
      is an important message."

   Test Inquiry

      It may sometimes be useful for one Host to determine if some other
      Host is carrying on network conversations.  The control command to
      be used for this purpose is ECO (echo).

              8       8
          ------------------
          !  ECO  !  data  !
          ------------------

      The data field of the ECO command may contain any bit
      configuration chosen by the Host sending the ECO.  Upon receiving
      an ECO command, an NCP should respond by returning the data to the
      sender in an ERP (echo reply) command.

              8       8
          ------------------
          !  ERP  !  data  !
          ------------------

      A Host should respond (with an ERP command) to an incoming ECO
      command within a reasonable time, here defined as sixty seconds or
      less.  A Host should not send an ERP when no ECO has been
      received.

   IV.  DECLARATIVE SPECIFICATIONS

   Message Format

      All Host-to-Host messages which conform to this protocol shall be
      constructed as follows:

      Bits 1-96: Leader - This field is as specified in BBN Report No.
      1822, with the following additional specifications.

      Bits 38-40: Maximum Message Size - This field should be zero for
      all control messages.  For messages sent over data connections,
      the value of this field should be calculated from the size
      received in the RFC which established the connection.

      Bits 65-76: Message-id - This field is subdivided into eight bits
      giving the index of the connection of which the message is a part,
      and four bits giving the sequence number of the message.  The
      index is contained in bits 65-72, and the sequence number in bits
      73-76.

      Bits 97-100: Acknowledgement - This field contains the four-bit
      sequence number of the highest-numbered data message to the left
      of the window for this connection; that is, the sequence number
      identifying the highest-numbered of the sequence of consecutively
      numbered (none missing) data messages which have been correctly
      received over this connection.  If no data messages have been
      received since the connection was established, this field must
      contain the value zero.  This field is not used (i.e., may have
      any value) in control messages.

      Bits 101-104: Credit - This field contains a number in the range
      0-7.  Adding this number (modulo 16) to the sequence number in the
      acknowledgement field (bits 97-100) gives the highest sequence
      number which the foreign Host is permitted to send over this data
      connection.  Thus, a value of zero in this field indicates that no
      new data messages should be sent, and a value of seven indicates
      that the foreign Host may send up to seven messages beyond the
      message whose sequence number is specified by the acknowledgement
      bits.  Since flow control does not apply to messages sent over the
      control connection, this field may have any value in control
      messages.

      Bits 105 - ... : Text and padding - A sequence of 8-bit bytes of
      text, followed by padding, as specified in BBN Report No. 1822.

   Index Assignment

      Index values must be assigned (in bits 65-72) as follows:

       Number     Assignment

0 Identifies a control connection

1 Reserved for revisions to this protocol

        2-191     Identify data connections

      192-255     Reserved for expansion or for other protocols

   Sequence Number Assignment

      Every data message contains a sequence number in bits 73-76.  The
      sequence number is used by the receiver to detect the fact that a
      transmitted message has been lost, to identify the correct
      location in the data stream to insert a retransmitted (and
      therefore probably out of order) message which was previously lost
      (or to detect the retransmitted message as a duplicate) and to
      identify acknowledged messages (or sequences of messages) to the
      sender.  The sequence number is also used by the flow control
      mechanism.  Since the IMP subnetwork itself contains elaborate
      mechanisms to achieve these same goals, it is not anticipated that
      the error-recovery mechanisms based on the sequence numbers will
      be called into play frequently, and thus their efficiency is not
      of primary importance.

      Sequence numbers are assigned to the two directions of a
      connection independently.  For a given direction of a connection,
      the first data message transmitted after the connection is

      established must have sequence number one.  Subsequent messages
      are assigned sequentially increasing (modulo 16) sequence numbers;
      that is, sequence number zero is assigned to the message following
      message number 15.

      Sequence numbers are not assigned to control messages, since the
      protocol is designed to permit these messages to be delivered
      out-of-sequence without ill effect, and since flow control cannot
      be applied to the control link.

   Control Messages

      Messages sent over the control connection have the same format as
      other Host-to-Host messages, with the exceptions noted above.
      However, control messages may not contain more than 120 8-bit
      bytes of text.  Further, control messages must contain an integral
      number of control commands; a single control command must not be
      split into parts which are transmitted in different control
      messages.

   Message Transmission and Retransmission

      Control messages may be transmitted whenever they are required.
      Data messages, however, may be transmitted only when permitted by
      the flow control mechanism; that is, whenever the sequence number
      assigned to the message is within the "window" for the appropriate
      direction of the given connection.  The "left window edge" (LWE)
      is defined by the highest sequence number (modulo 16) which has
      been acknowledged (or zero, if no messages have been
      acknowledged).  The "right window edge" (RWE) is defined by adding
      (modulo 16) the most recently received credit to the left window
      edge. [Note that LWE=RWE if the most recently received credit is
      zero.]  A message with sequence number SEQ may be transmitted only
      if, prior to the (possible) reduction modulo 16 of the SEQ and/or
      RWE, it is true that

         LWE less-than SEQ less-than-or-equal RWE

      Messages should be retransmitted whenever any of the following
      conditions occur:

      - The IMP subnetwork has returned an "Incomplete transmission"
        (type 9) or "Error in Data" (type 8) response to the message
        (identified by having bits 41-76 of the response equal to those
        bits of the transmitted message).  Note that this condition
        applies to control messages as well as data messages.

      - The sequence number of this message is equal to (LWE + 1), and
        it has been more than 30 seconds since the message was last
        transmitted.

      - The sequence number of the message is specifically identified in
        a NACK command for this connection from the foreign Host.

      Since messages may occasionally have to be retransmitted, it is
      clear that they should not be discarded by the transmitting NCP
      until they have been acknowledged.  A message is considered to be
      acknowledged when its sequence number, or the sequence number of
      any message to the right of it in the same direction of the given
      connection, is returned in the acknowledgement field of a data
      message transmitted in the other direction over this connection,
      or is returned in an ACK command for this connection from the
      foreign Host.

   Control Commands

      Control commands are formatted in terms of 8-bit bytes.  Each
      command begins with a one byte opcode.  Opcodes are assigned the
      sequential values 0, 1, 2, ...  to permit table lookup upon
      receipt.  The conditions underlying the design and anticipated use
      of the control commands are described in Section III.

   NOP - No Operation

              8
          ---------
          !  NOP  !
          ---------

      The NOP command may be sent at any time and should be discarded by
      the receiver.  It may be useful for formatting control messages.

   RST - Reset

              8
          ---------
          !  RST  !
          ---------

      The RST command is used by one Host to inform another that all
      information regarding any previously existing connections between
      the two Hosts should be purged from the NCP tables of the Host
      receiving the RST.  Except for responding to RSTs, the Host which
      sent the RST should not communicate further with the other Host
      until an RRP is received in response.  When a Host is about to

      begin communicating (e.g., send an RFC command) to another Host
      with which it has no open connections, it is good practice to
      first send an RST command and wait for an RRP command.

   RRP - Reset Reply

              8
          ---------
          !  RRP  !
          ---------

      The RRP command must be sent in reply to an RST command.

   RFC - Request for Connection

         8       16           16          8      16      8
      ---------------------------------------------------------
      ! RFC ! my-socket ! your-socket ! index ! size ! credit !
      ---------------------------------------------------------

      The RFC command is used to establish a connection.  The "my-
      socket" field specifies the socket local to the Host transmitting
      the RFC; the "your-socket" field specifies the socket local to the
      Host to which the RFC is transmitted.  The "index" field specifies
      the index value which will be given in bits 65-72 of each data
      message sent from "my-socket" to "your-socket".  The "size" field
      specifies the maximum number of 8-bit bytes which may be
      transmitted in any single message from "your-socket" to "my-
      socket".  The "credit" field specifies the size of the initial
      sequence number window (in the range 0-7) in the "your-socket" to
      "my-socket" direction.

   CLS - Close

              8        16           16
          -----------------------------------
          !  CLS  ! my-socket ! your-socket !
          -----------------------------------

      The CLS command is used to terminate a connection.  The connection
      need not be completely established before CLS is sent.

   RCP - Re-Initialize Connection Parameters

              8        16           16          8
          -------------------------------------------
          !  RCP  ! my-socket ! your-socket ! index !
          -------------------------------------------

      The RCP command is used by one Host to inform another that all
      information regarding a possibly previously-existing connection
      between "my-socket" and "your-socket" AND all information
      regarding a possibly previously-existing connection identified by
      "index" (between these Hosts) should be purged from the tables of
      the Host receiving the RCP.  The "my-socket", "your-socket", and
      "index" fields are defined as in the RFC command.

   ACK - Acknowledgement

              8       8       4       4
          ---------------------------------
          !  ACK  ! index !  seq  !  crd  !
          ---------------------------------

      The ACK command may be used to acknowledge received data, or to
      assign credit, without sending a data message.  The value in the
      index field identifies the data connection which uses the same
      index value (in the direction from the sender of the ACK to the
      receiver of the ACK).  The eight bits following the index field
      (the "seq" and "crd" field) have the same meaning as bits 97-104
      of the data message identified by the index value.

   NACK -- Negative Acknowledgement

              8        8       8
          --------------------------
          !  NACK  ! index !  seq  !
          --------------------------

      The NACK command informs the receiver of the NACK that it should
      immediately retransmit the data message identified by the
      remaining fields.  The index field is defined exactly as for the
      ACK command.  The "seq" field gives the 4-bit sequence number
      (right-justified) which should be immediately retransmitted.  Note
      that the data message to be retransmitted does not have an index
      value equal to "index", but instead is transmitted over the other
      direction of the data connection which the Host sending the NACK
      identifies by "index".  No Host is ever required to transmit or
      act upon a NACK command; however, use of the NACK may occasionally
      permit a decrease in retransmission delay.

   INT - Interrupt

              8       8       8
          -------------------------
          !  INT  ! index !  seq  !
          -------------------------

      The INT command is sent over the control link to provide an "out-
      of-band" (and hence not subject to flow control) signal for the
      data connection denoted by the index field.  The index value is
      the value which would appear in bits 65-72 of a data message sent
      from the sender of the INT command to the receiver of the INT
      command.  The means of synchronizing this signal with the data
      being transmitted over the data connection is the inclusion of a
      4-bit sequence number (right-justified) in the "seq" field.  The
      number specified by this field denotes the first data message
      which "follows" the out-of-band signal.

   ECO - Echo Request

              8       8
          ------------------
          !  ECO  !  data  !
          ------------------

      The ECO command is used only for test purposes.  The data field
      may be any bit configuration convenient to the Host sending the
      ECO command.

   ERP - Echo Reply

              8       8
          ------------------
          !  ERP  !  data  !
          ------------------

      The ERP command must be sent in reply to an ECO command.  The data
      field must be identical to the data field in the incoming ECO
      command.

   Opcode Assignment

      Opcodes are defined to be 8-bit unsigned binary numbers.  The
      values assigned to opcodes are:

          NOP  =  0

          INT  =  1

          RFC  =  2

          CLS  =  3

          ACK  =  4

          NACK =  5

          RCP  =  6

          RST  =  7

          RRP  =  8

          ECO  =  9

          ERP  = 10

         [ This RFC was put into machine readable form for entry ]
         [ into the online RFC archives by Alex McKenzie with    ]
         [ support from BBN Corp. and its successors.     7/2000 ]

I. INTRODUCTION

Acknowledgement

8 8 4 4 --------------------------------- ! ACK ! index ! seq ! crd ! ---------------------------------

0 Identifies a control connection

1 Reserved for revisions to this protocol

8 8 4 4
---------------------------------
! ACK ! index ! seq ! crd !
---------------------------------