Biografia e approfondimenti
Conversation with Steve Crocker
(Both my personal opinions and those that I believe represent
a
consensus of the Network Working Group at Project MAC are presented
here. The pronouns "I" and "we" are used to distinguish
between
these.)
On April 21 and 23 Thomas P. Skinner and I had telephone
conversations
with Steve Crocker at UCLA relating to the network protocol,
specifically regarding our proposal in NWG/RFC 46. The following items
were discussed. (I hope that Steve will pardon me if I happen to
misparaphrase him.)
1) Steve stated that he felt that a need for dynamic
reconnection would
later be recognized by the network participants. However, because of a
lack of consensus, it will not be included in the initial
implementation. (We at Project MAC favor this approach of not including
it initially.)
2) Steve supported the implementation of the INT network
command
described in NWG/RFC 46.
This command allows a process that has agreed to accept
interrupts over
a socket connection to be reliably interrupted by the process at the
other end. The interrupt causes a process to abey its current execution
and execute a procedure that it has specified as the INT handler. (The
NCP does not specify the INT handler. That is the function of higher
level protocols.)
The INT command is designed specifically for use by a
third level User
Control and Communication (UCC) protocol to implement a "quit"
signal.
Under such a protocol, both the requestor and the created process agree
that an INT related to a specific socket connection and transmitted over
the NCP control link to the created process is the standard "quit"
signal. The created process provides an INT handler that implements
this "quit" function. (This does not preclude a different
interpretation of INT by other third level protocols.)
Although many systems implement the "quit"
as a control character in the
Teletype input stream, systems such as CTSS, Multics, and others
implement it as a 200 ms spacing on the line. We at MAC think that the
first method is an undesirable implementation within the network (while
the second is impossible). I put forth several reasons why (and I think
Steve agreed).
(a) The link over which the quit character is to be transmitted
may be
blocked.
(b) While the interrupt is most effectively implemented
within the NCP,
it is undesirable for the NCP to place any particular structure on the
data being transmitted. (See discussion below.) This would be required
if the NCP were to scan a data stream for a control character.
(c) Scanning the input stream greatly reduces NCP efficiency
in a
subsystem where speed is critical to effective operation.
Steve pointed out that the implementation of INT as a
"quit" should not
necessarily preclude a HOST's interpretation of a control character in
the input stream from also acting as a "quit".
3) Steve is opposed both to including the instance tag
in the socket
identifier and reserving a null field in the identifier for future
definition. He cited several reasons:
(a) Multiple processes of a single user should be indistinguishable
to a
foreign process. (I agree with this in certain cases when processes are
co-ordinated in joint action. But what about the case where two
processes of the same user both want to independently use the network?)
(b) A process wishing to connect to one of a foreign
user's processes
does not know the instance tag of the particular process that he wants,
and he can't easily find out.
(c) If an instance tag should later prove desirable it
could be added
with some difficulty. (I claim that something as fundamental as the
length of a socket identifier will prove very resistant to change.)
Tom stated that perhaps the low order three bits of the
user code could
be reserved for later interpretation as an instance tag. He doesn't
think that a separate field is of great importance.
Steve's arguments seem to have merit. Perhaps Tom's suggestion
is the
way to go. I am currently undecided on this matter.
4) We all (Steve and MAC) seem to agree that at the NCP
level there
should be no special structure imposed on the data transmitted. To an
NCP all data to be transmitted are bit strings of arbitrary length. One
happy result is that the difficult question of character sets does not
have to be resolved at this protocol level. To include a character set
specification at the NCP level would delay agreement on the protocol and
make this character set more resistant to change. (If there is to be a
standard character set, we prefer ASCII. After all, it is the prefered
standard of our sponsoring organization.)
We also agree with Steve that there should be no optional
echoing of
messages at the NCP protocol level. (This is also the position of the
SDC people in RFC 44.)
5) Shoshani, Long, and Landsberg also state (RFC 33)
that they prefer to
align messages to end on a word boundary as opposed to double padding.
Steve agrees with us in not liking double padding.
6) In our proposal (RFC 46) we suggest that RFCs be queued
only for open
sockets, that RFCs to inactive or connected sockets are to be
automatically rejected via the CLS command. Steve proposes that RFCs to
these sockets be briefly queued. If the socket remains in an
unacceptable state for a specific interval after the RFC comes in, it
is
rejected. This scheme allows certain types of network command
interaction involving critical races to be implementable. Such a scheme
of limited queueing does not seem unreasonable to me.
7) Steve, Tom, and I discussed strategies for a User
Control and
Communication (UCC) Protocol. Steve said that he disliked our UCC
strategy (RFC 46) because it requires maintaining two full-duplex
connections to the requestor process and switching between them.
Steve put forth an alternate proposal: a process wishing
to create a
user process at a foreign HOST issues RFCs to sockets 0 and 1 belonging
to the user whose process he wishes to create. If these sockets are
inactive, the NCP automatically directs these requests to the foreign
HOST's logger process. The logger accepts connection and performs the
login ritual. If successful, the logger creates a user process and lets
go of the usurped sockets so that the created process may use them to
communicate with the requestor process. (I note that this does not use
reconnection at a network level, since the logger uses sockets belonging
to the ultimate user. However, it does involve internal reconnection.)
Tom and I objected to this because it introduces UCC
protocol into the
NCP level. (The NCP must direct all RFCs to inactive sockets 0 and 1 to
a logger process.) I made a quick suggestion that perhaps our two
proposals could be combined such that the requestor issues a
"signalling" RFC to a "signal" socket of the UCC process.
The UCC
rejects the RFC but remembers who is calling. It then tries to connect
two sockets of the process to be created to the requestor's sockets, and
conducts the login ritual through these. Steve liked this and suggested
that I write it up.
Following the conversation, I thought of several disadvantages
to this
UCC strategy:
(a) If the control sockets at a created process are limited
to 0 and 1,
there is the possibility that a rightful user may not be able to
communicate with a foreign UCC because the UCC already is using those
sockets to communicate with an imposter. The logger will discover this
and turn off the imposter, but this is an aggravating security breach.
A malicious process could issue simultaneous multiple requests to tie
up
the sockets and prevent access to a rightful user. A better solution is
to allow any socket pair of the potential user process to act as the
control path. This permits the UCC to conduct simultaneous
interrogations of competing requestors.
(b) A disadvantage of both Crocker's and the combined
UCC is that the
user to be logged in is specified by supplying a socket belonging to a
particular user. The logger must now make the additional check that the
user it is logging in actually belongs to the socket pair it is talking
over. This seems the reverse of the prefered process: to identify a
user and then determine the user code for his socket identifiers.
(c) The user may not know the socket user code of the
user he wishes to
log in at the foreign HOST. (After all, there is no basic reason why
the requestor and created processes should have the same user code so
long as the requestor satisfies the foreign logger.)
(d) In the combined strategy, there is no way for the
requestor to
specify which socket user code it wants. The only assumption that the
UCC can make is that the requestor process wishes to log in a process
having the same socket user code as itself. (This may not seem very
important, but I envision a scheme in which a local process exists to
allow consoles attached to the local HOST to login at a foreign HOST
without being logged in locally.)
(e) The idea of allowing a process to masquerade within
the network as
another process (even with the best of intentions) by using its socket
user code introduces a potentially dangerous security breach. I think
that it should be a basic protocol law that NO PROCESS WHATSOEVER may
request or accept connections or transmit or receive data over a socket
having a user code not its own. This does not apply to an NCP process
which has responsibility for such transmission, nor does it prevent a
priviliged process from closing or rejecting connections between a
foreign process and another local process.
I still think that the UCC proposal we advanced in RFC
46 is a good
workable scheme. It does not require socket reconnection (either
expressly throughout the network or implicitly within an NCP), nor do
any of the objections raised above apply. The only particular
disadvantage I see is that it requires the requestor process to maintain
and switch between two full-duplex connections. I don't see this as a
serious hindrance. I would like the comments of the network
participants on this point in particular.
Fortunately the UCC is a third level protocol. The second
level NCP can
be specified before we reach final agreement on a UCC, provided that the
NCP allows implementation of a workable UCC.
Steve expressed the thought that there need not be an
initial standard
UCC, that there might be several UCCs. We at MAC disagree. If we are
all to talk to each other, and not between limited subsets of HOSTs
within the network, there must be an initial standard UCC which
EVERYBODY implements. (Steve is of course correct that there can be
other experimental UCCs also implemented.)
It is theoretically possible for each HOST to provide
multiple sets of
software to allow a requestor process to communicate with the loggers
at
HOSTs implementing different UCCs. I don't think that it will work this
way in practice. Each HOST will implement the UCC protocol that is most
agreeable to it, and will provide one set of software so that a
requestor process can communicate only with those HOSTs which implement
similar UCCs.
I don't think that there is much enthusiasm at Project
MAC for
implementing a non-standard UCC just so we can talk to ourselves. We
want to implement a single UCC supported at all installations, so that
we can log in to all HOSTs using this protocol, and that users at all
foreign HOSTs can log in to us.
[ This RFC was put into machine readable form for entry
]
[ into the online RFC archives by Altair Petrofsky 7/97 ]