rfc636.txt

来自「RFC 的详细文档！」· 文本 代码 · 共 447 行 · 第 1/2 页


NWG/RFC# 636                 JDB BPC RST DCW3 MLK 23-OCT-75 22:27  30490
TIP/TENEX Reliability Improvements



RFC 636                                    J. Burchfiel  - BBN-TENEX
                                           B. Cosell     - BBN-NET
NIC 30490                                  R. Tomlinson  - BBN-TENEX
                                           D. Walden     - BBN-NET
                                                            10 June 1974
                                                                       
                   TIP/TENEX Reliability Improvements



                                                                       

During the past months we have felt strong pressure to improve the
reliability of TIP/TENEX network connection as improvement in the
reliability of users' connections between TENEXs and TIPs would have
major impact on the appearance of overall network reliability due to the
large number and high visibility of TENEXs and TIPs.  Despite the
emphasis on TIP/TENEX interaction, all work done applies equally well to
interactions between Hosts of any type.                                

The remainder of this RFC gives a sketch of our plan for improving the
reliability of connections bettween TIPs and TENEXs.  Major portions of
this plan have already been implemented (TIP version 322; TENEX version
1.32) and are now undergoing final test prior to release throughout the
network.  Completion of the implementation of the plan is expected in
the next quarter.                                                      

Our plan for improving the reliability of TIP/TENEX connections is
concerned with obtaining and maintaining TIP/TENEX connections,
gracefully recovering from lost connections, and providing clear
messages to the user whenever the state of his connection changes.     

When a TIP user attempts to open a connection to any Host, the Host may
be down.  In this case it would be helpful to provide the user with
information about the extent of the Host's unavailability. To facilitate
this, we modified the IMP program to accept and utilize information from
a Host about when the Host will be back up and for what reason it is
down.  TENEX is to be modified to supply such information before it goes
down, or through manual means, after it has gone down.  When the TIP
user then attempts to connect to the down TENEX, the IMP local to the
TENEX returns the information about why and for how long TENEX will be
down.  The TIP is to be modified to report this sort of information to
the user; e.g., "Host unavailable because of hardware maintenance --
expected available Tuesday at 16:30 GMT".                              

The TIP's logger is presently not reentrant.  Thus, no single TIP user
can be allowed to tie up the logger for too long at a time; and the TIP

NWG/RFC# 636                 JDB BPC RST DCW3 MLK 23-OCT-75 22:27  30490
TIP/TENEX Reliability Improvements



therefore enforces a timeout of arbitrary length (about 60 seconds) on
logger use.  However, a heavily loaded Host cannot be guaranteed always
to respond within 60 seconds to a TIP login request, and at present TIP
users sometimes cannot get connected to a heavily loaded TENEX.  To
correct this problem, the TIP logger will be made reentrant and the
timeout on logger use will be eliminated.                              

One notorious soft spot in the Host/Host protocol which degrades the
reliability of connections is the Host/Host protocol incremental
allocate mechanism.  Low frequency software bugs, intermittant hardware
bugs, etc., can lead to the incremental allocates associated with a
connection getting out of synchronization.  When this happens it usually
appears to the user as if the connection just "hung up".  A slight
addiition to the Host/Host protocol to allow connection allocates to be
resynchronized has been designed and implemented for both the TIP and
TENEX.                                                                 

TENEX has a number of internal consistency checks (called "bughalts")
which occasionally cause TENEX to halt.  Frequently, after diagnosis by
system personnel, TENEX can be made to proceed without loss from the
viewpoint of local users.  A mechanism is being provided which allows
TENEX to proceed in this case from the point of view of TIP users of
TENEX.                                                                 

The appropriate mechanism entails the following:  TENEX will not drop
its ready line during a bughalt (from which TENEX can usually proceed
successfully), nor will it clear its NCP tables and abort all
connections.  Instead, after a bughalt TENEX will:  discard the message
it is currently receiving, as the IMP has returned an Incomplete
Transmission to the source for this message; reinitialize the interface
to the IMP; and resynchronize, on all connections possible, Host/Host
protocol allocate inconsistencies due to lost messages, RFNMs etc.  The
latter is done with the same mechanism described above.  This procedure
is not guaranteed to save all data -- a tiny bit may be lost -- but this
is of secondary importance to maintaining the connection over the TENEX
bughalt.                                                              

The TIP user must be kept fully informed as TENEX halts and then
continues.  Therefore, the TIP has been modified to report "Host not
responding -- connection suspended" when it senses that TENEX has halted
(it does this by properly interpreting messages returned by the
destination IMP).  When TENEX resumes service after proceeding from a
bughalt, the above procedure notifies the TIP that service is restored,
and the TIP has been modified to report "Service resumed" to all users
of that Host.                                                         

On the other hand, the service interruption may not be proceedable and





                                   1

NWG/RFC# 636                 JDB BPC RST DCW3 MLK 23-OCT-75 22:27  30490
TIP/TENEX Reliability Improvements



TENEX may have to do a total system reload and restart.  In this case
TENEX will clear its NCP connection tables and send a Host/Host protocol
reset command to all other Hosts.  On receiving this reset command, the
TIP will report "Host reset -- connection closed" to all users of that
Host with suspended connections.  The TIP user can then re-login to the
TENEX or to some other Host.                                          

Of couse, the user may not have the patience to wait for service to
resume after a TENEX bughalt.  Instead, he may unilaterally choose to
connect to some other Host, ignoring the previously suspended
connection.  If TENEX is then able to proceed, its NCP will still think
its connection to the TIP is good and suitable for use.  Thus, we have a
connection which the TIP thinks is closed and TENEX thinks is open, a
phenomenon known as the "half-closed connection".  An automatic
procedure for cleanly completing the closing of such a connection has
been specified and implemented for the TIP and TENEX.                 

Since TENEX will maintain connections across service interruptions, the
TIP user will be required to take the security procedure telling the TIP
to "forget" his suspended connection before abandoning his terminal.
The command @H 0 (for example) will guarantee that his connection will
not be reestablished on resumpption of service.  Otherwise, his job
would be left at the mercy of anyone who acquires that terminal.      

An appendix follows which describes the Host/Host protocol changes made.
These changes are backward compatible (with the exception that Hosts
which have not implemented these changes will sometimes receive
unrecognizable Host/Host protocol commands which they presumably discard
without suffering harm).  These protocol changes are ad hoc in nature
but in light of their backward compatibility and potential utility, ARPA
okayed their addition to the TIP and TENEX NCPs without (we believe) any
implication that other Hosts have to implement them (although we would
encourage their widespread implementation).                           



















                                   2

NWG/RFC# 636                 JDB BPC RST DCW3 MLK 23-OCT-75 22:27  30490
TIP/TENEX Reliability Improvements



             Appendix - Ad Hoc Change to Host-Host Protocol           

   A.1  Introduction                                                 

      The current Host-Host protocol (NIC #8246) contains no provisions
      for resynchronizing the status information kept at the two ends of
      each connection.  In particular, if either host suffers a service
      interruption, or if a control message is lost or corrupted in an
      interface or in the subnet, the status information at the two ends
      of the connection will be inconsistent.                       

      Since the current protocol provides no way to correct this
      condition, the NCPs at the two ends stay "confused" forever.  An
      occasional frustrating symptom of this effect is the "lost
      allocate" phenomenon, where the receiving NCP believes that it has
      bit and message allocations outstanding, while the sending NCP
      believes that it does not have any allocation.  As a result,
      information flow over that connection can never be restarted. 

      Use of the Host-Host RST (reset) command is inappropriate here, as
      it destroys all connections between the two hosts.  What is needed
      is a way to resynchronize only the affected connection without
      disturbing any others.                                        

      A second troublesome symptom of inconsistency in status
      information is the "half-closed" connection:  after a service
      interruption or network partitioning, one NCP may believe that a
      connection is still open, while the other believes that the
      connection is closed (does not exist).  When such an inconsistency
      is discovered, the "open" end of the connection should be closed.
                                                                    
   A.2  The RAR, RAS and RAP commands                               

      To achieve resynchronization of allocation, we add the following
      three commands to the host-host protocol.                     

              8 bits   8 bits
            -------------------
            !        !        !
         16 !  RAR   !  link  !
            !        !        !
            -------------------
         Reset Allocation by Receiver

              8 bits   8 bits
            -------------------
            !        !        !