Network Working Group R. Braden
Request for Comments: 430 CCN/UCLA
NIC: 13299 7 February 1973 COMMENTS ON FILE TRANSFER PROTOCOL
On January 23, 1973, Jon Postel (NMC), Eric Harslem (RAND), Stephen
Wolfe (CCN), and Robert Braden (CCN), held and informal meeting at
UCLA on FTP. This RFC generally reports the consensus of that
meeting on the following issues: server-server transfers (ref. RFC
438 by Thomas and Clements); site-dependent information; and
miscellaneous questions/disagreements with RFC 354, 385, and 414.
There was also a discussion of the print file muddle, but that
subject is addressed in a separate RFC, No. 448.
Miscellaneous Comments on FTP
1. RFC 385, P. 1 (3)
The question of print files will be discussed at length in another
RFC. However, we did feel that the word "still" on the second
line from the bottom of Page 1 is gratuitous.
2. RFC 385, P. 2 (5.)
RFC 385, P. 3 (8.)
RFC 414, P. 4 (11.i)
To the extent that we understand these items, they seem to be
unnecessary and probably undesirable concessions to particular bad
implementations ("hacks"). In reference to the second item, No. 8
in RFC 385, one should note that in an asynchronous multi-process
system like the ARPA Network, the phrase "immediately after" has
little meaning. An implementation which depends upon "immediately
after" is erroneous and should be fixed. If the protocol as
defined has an intrinsic race condition, of course, the protocol
should be fixed, but we don't believe such a problem exists. It
would help if definitions of command-response sequences in the FTP
document were tightened up considerably. As for the last item, we
don't understand why Wayne Hathaway is so strongly opposed to
3. RFC 354, P. 13, Format Definitions for Block Mode
(a) The definition of the header length presumably is meant to be
the "smallest integral number of bytes whose length is greater
or equal to 24 bits".
(b) The same definitional problem occurs for restart markers.
(c) Why does the restart marker have to be greater than 8 bits?
(d) Note that changing the Descriptor coding to bit flags would
abolish the implied eor as well as the problem of RFC 385, P.
4. RFC 414, P. 5 (11.iii)
Note that text mode is not possible for any EBCDIC coded file.
Since EBCDIC is an 8-bit code, Telnet control characters
(128-255) cannot be used to distinguish either eor or eof.
Stream and block modes will work, however. We have found the
diagram on the last page to be useful for keeping track of the
three-dimensional space of FTP parameters.
5. RFC 354, P. 17, PASS Command
There is no mechanism within FTP for changing a password. A
user shouldn't have to use a different protocol (e.g., log
into a time sharing system) to merely change his password.
6. RFC 385, P. 3 (9.), TYPE Before BYTE
This admonition (to send TYPE before BYTE) should be clearly
labeled as a recommended procedure for user FTP, not a restriction
on a server FTP.
7. RFC 385, P. 2-3 (7) Order of 255 Reply
Some of the participants felt (strongly) that the timing problem
dealt with in this item is the result of bad NCP implementations
and ought not be dignified in the protocol. The issue here is the
old, familiar, and touchy one of queueing RFC's or not. (My own
view is that the protocol asymmetry forced by NCP's which can't
queue RFC's is at least unaesthetic, and makes some elegant
solutions impossible. For examples, see RFC 414 and the comments
below on server-server interaction, and RFC 438 on Reconnection
8. RFC 354, P. 15, Restart
Following a RESTart command, APPend and STORe presumably have
B. FTP Parameter Encoding
RFC 448, which discusses print files, points out that the print file
attribute is logically independent of the character code attribute
(ASCII vs. EBCDIC) in the type dimension; the set of allowable types
in FTP is the outer product of the individual attributes. Thus FTP
has (at least) four character types, summarized by the following two
x two matrix:
| ASCII | EBCDIC
Not Print File | |
Print File | |
I propose that the encoding in the TYPE command model this
interdependence of the types. Instead of using a distinct single
ASCII character for each type, we should use multiple ASCII
characters---qualifiers, if you wish. For example:
A represents ASCII code
E represents EBCDIC code
P represents print file
I represents image
L represents local byte
Then the legal types according to RFC 385 would be:
Note that the attributes under consideration here are type-like; they
are not (logically) concerned with the structure or the transmission
mode, only the internal encoding of the file.
At present, this would be a trivial change. However, I foresee the
file transfer protocol expanding significantly over the next several
years as new types are added. Some servers will want to add server-
specific type variations, and the NWG will want to add some. How
about an APL character set? Or the multiple-overlay 256 character
ASCII which has been proposed? Multiple qualifiers (and later
perhaps more structure) in the type seems to be the cleanest escape
mechanism for future growth.
C. Server-Server Interaction
The FTP changes proposed by Thomas and Clements in RFC 438 are a
particular solution to a general problem inherent in the current
host-host protocol and higher-level protocols like FTP. There seems
to be a need for a secure and simple way for two (server) processes
in different hosts to exchange socket names (i.e., 40-bit numbers) so
they can subsequently exchange RFC's and establish a connection.
Current second-level (host-host) protocol provides exactly one
(secure) mechanism by which one host can learn a socket name of a
process at another host in order to establish a connection: ICP. The
ICP mechanism by itself is not adequate for server-server connection
in FTP. Therefore, Thomas and Clements have proposed an extension to
the FTP protocol, roughly as follows:
(1) A controller ("user") process at Host A uses ICP to invoke and
establish Telnet control connections to two automata
("server") processes at two other hosts. An automaton process
invoked in this manner then executes controller commands sent
in a standard command language over the Telnet control
(2) The controller process commands each automaton to reserve a
suitable data transfer socket and to return the socket name to
the controller over the control connection. An automaton
presumably negotiates with his own NCP in a host-dependent
manner to obtain the socket reservation.
(3) The controller now knows both data transfer socket names; he
will send them in subsequent commands to the automata so each
automaton will know the foreign socket name to which he is to
connect. Later commands cause the automata to issue RFC's and
open the data connection as needed.
This appears to be useful general model for process-process
interaction over the Network. Personally, I believe this symmetrical
model should be the basis of all FTP the controller and one of the
automata could be in the same host. Then the user/server problem
(for any pair of hosts to transfer files, one must have a server FTP
and the other a user FTP) would vanish. At least one host somewhere
in the Network would need a controller process; all other hosts would
need only an automaton process.
Perhaps at a future time the NWG should consider whether a socket-
reservation-and-passing mechanism ought to be incorporated into
second-level protocol rather than duplicated in a number of third-
level protocols. We should note that this model provides secure
sockets only if both user and server processes "release" the socket
reservations when the Telnet control connection breaks. The same
problem seems to occur with Thomas' Reconnection Protocol (426).
In any case, for the present we would endorse the general third-level
model of RFC 438. However, we would propose a slightly different,
and more symmetrical, approach.
1. The requirement in FTP that the FTP user listen on the data
socket before issuing a data transfer command should be
removed. The beauty of host-host protocol is that it doesn't
matter which host sends the first RFC, as long as they both
send matching RFC's "eventually". (Timeouts, of course, are
annoying, but I believe they are workable and ultimately
unavoidable); queueing RFC's is also necessary).
2. We propose, instead of LSTN, a new command GETSocket. The
controller (i.e., user FTP) process would send a GETSocket to
each automaton, probably after a successful login. Upon
receiving GETSocket, an automaton would assign a (send,
receive) pair of data transfer sockets and return the numbers
over the Telnet connection. (Alternatively, FTP could specify
that a (send, receive) pair of sockets always be assigned when
the server is first entered, and the numbers returned to the
user process via unsolicited 255 replies).
3. Then the user process would send the socket numbers to the
opposite hosts by sending SOCK commands to both.
4. When it receives a data transfer command, the automaton
(server) process would issue an RFC containing the two socket
numbers. When both servers are fired up, RFC's are exchanged
and data transfer starts.
D. Site-Dependent FTP Parameters
Some hosts will have a problem with the current FTP because their
file system needs additional host-specific parameters in certain
cases. As an example, the IBM operating systems tend to give the
programmer a number of options on the logical and physical mapping of
a file onto the disk.
This is true both of TSS/360 (see Wayne Hathaway's discussion of his
STOR command implementation, Page 5 of RFC 418), and OS/360. The
large set of options and parameters to the OS/360 file system is, in
fact, the (legitimate) origin of most complaints about OS Job Control
If the FTP user merely wants to store data without using it at one of
these sites, he has no problem; defaults can be chosen to handle any
reasonable FTP request. However, the FTP user who sends a file to an
IBM/360 for use there may need to specify local file system
parameters which are not derivable from any of the existing FTP
In designing an FTP server implementation for CCN, for example, we
first tried to handle the mapping problem by choosing a (possibly
different) default mapping for each combination of FTP parameters--
type, mode, and structure. We hoped that if a user chose
"reasonable" or "suitable" FTP parameters for a particular case
(e.g., "ASCII, stream, record" for source programs, and "image,
block, record" for load modules), then the right OS/360 file mapping
would result. We were forced to abandon this approach, however,
because of the following arguments:
1. Some user FTP's probably may not implement all FTP
type/mode/structure combinations (though they ought to!).
2. Some user FTP's may not give the user full or convenient
control over his type/mode/structure. Indeed, the mode should
be chosen on grounds of efficiency, not end use.
3. There weren't enough logically distinct combinations of FTP
4. The result would have been a set of hard-to-remember rules for
sending files to CCN for use here.
5. Some common cases require non-invertible transformations on the
data. For example, most IBM language processors (i.e.,
compilers) accept only fixed length records of (surprise!) 80
bytes each, i.e., literal card images. Such ugly (and
logically unnecessary) implementation stupidities in OS/360 are
a fact of life. Now if a FTP user innocently sent a data file
to CCN with the particular type/mode combination which
defaulted to card images, he would find his records truncated
to 80 bytes. That would be downright unfriendly.
Thus, the CCN server FTP would have to choose between being useful or
being friendly. We decided upon the following strategy:
1. The defaults will be friendly; we will accept any FTP
type/mode/structure and store it invertibly (except print
files). However, the user who uses only these defaults will
probably find he has to later run a utility under TSO to
reformat the data.
2. We will provide some mnmonic keywords associated with STOR
commands to choose the proper disk mapping. For example, if he
wants to STORe a Fortran source file for compilation at CCN,
the user will need only to specify "SOURCE" or "FORT" to get
reasonable and workable OS/360 file system parameters. In
addition, we will provide fairly complete "DD" parameters for
the sophisticated user. The syntax and semantics of these
keywords and parameters will be as close as possible to the
corresponding TSO commands. Full details will be published as
soon as the implementation is working.
All of this discussion leads to a general protocol question: how
should such host-dependent information appear within FTP? Hathaway
used the ALLO command (see RFC 418, P. 6). CCN, on the other hand,
feels that such information belongs in the only part of FTP syntax
which is already host-dependent: the pathname. So CCN plans to allow
a "generalized" pathname in a STOR command, a (full or partial) file
name optionally followed by one or keywords or keyword parameters
separated by commas.
A third possible solution might be for the user to precede his STORe
command by a server-dependent data set creation command, using
Hathaway's proposed SRVR command. The data set creation command
could then have all the parameters necessary for the server file
system. CCN might change to this approach if SRVR is adopted and if
people find the generalized pathname objectionable or unworkable.
For another interesting example of host-dependent problems, see
Hathaway's discussion of his DELE command in RFC 418 (pp.6-7).