[ih] NCP and TCP implementations

[Jack Haverty]

The first TCP implementation for Unix was done in PDP-11 assembly
language, running on a PDP-11/40 (with way too little memory or address
space).   It was built using code fragments excerpted from the LSI-11
TCP implementation provided by Jim Mathis, which ran under SRI's
home-built OS.  Jim's TCP was all written in PDP-11 assembler.  The code
was cross-compiled (assembled) on a PDP-10 Tenex system, and downloaded
over a TTY line to the PDP-11/40.  That was much easier and faster than
doing all the implementation work on the PDP-11.

The code architecture involved putting TCP itself at the user level,
communicating with its "customers" using Unix InterProcess
Communications facilities (Rand "Ports").   It would have been
preferable to implement TCP within the Unix kernel, but there was simply
not enough room due to the limited address space available on the 11/40
model.  Later implementations of TCP, on larger machines with twice the
address space, were done in the kernel.  In addition to the Berkeley BSD
work, I remember Gurwitz, Wingfield, Nemeth, and others working on TCP
implementation for the PDP-11/70 and Vax.

The initial Unix TCP implementation was for TCP version 2 (2.5 IIRC), as
was Jim's LSI-11 code.  This 2.5 implementation was one of the players
in the first "TCP Bakeoff" organized by Jon Postel and carried out on a
weekend at ISI before the quarterly Internet meeting.  The PDP-11/40 TCP
was modified extensively over the next year or so as TCP advanced
through 2.5, 2.5+, 3, and eventually stabilized at TCP4 (which it seems
we still have today, 40+ years later!)

The Unix TCP implementation required a small addition to the Unix kernel
code, to add the "await" and "capac" system calls.  Those calls were
necessary to enable the implementation of user-level code where the
traditional Unix "pipeline" model of programming
(input->process->process...->output) was inadequate for use in
multi-computer programming (such as FTP, Telnet, etc., - anywhere where
more than one computer was involved).

The code to add those new system calls was written in C, as was almost
all of the Unix OS itself.  The new system calls added the functionality
of "non-blocking I/O" which did not previously exist.  It involved very
few lines of code, since there wasn't room for very many more
instructions, and even so it required finding more space by shortening
many of the kernel error messages to save a few bytes here and there.

Randy Rettberg and I did that work, struggling to understand how Unix
kernel  internals worked, since neither of us had ever worked with Unix
before even as a user.   We did not try to "get it right" by making
significant changes to the basic Unix architecture.  That came later
with the Berkeley and Gurwitz efforts.  The PDP-11/40 was simply too
constrained to support such changes, and our mission was to get TCP
support on the machine, rather than develop the OS.

I think I speak authoritatively here, since I wrote and debugged that
first Unix TCP code.   I still have an old, yellowing listing of that
first Unix TCP.

FWIW, if there's interest in why certain languages were chosen, there's
a very simple explanation of why the Unix implementation was done in
assembler rather than C, the native language of Unix.  First, Jim
Mathis' code was in assembler, so it was easy to extract large chunks
and paste them into the Unix assembler implementation.  Second, and
probably most important, was that I was very accustomed to writing
assembler code and working at the processor instruction level.  But I
barely knew C existed, and was certainly not proficient in it, and we
needed the TCP working fast for use in other projects.  The choice was
very pragmatic, not based at all on technical issues of languages or
superiority of any architecture.

/Jack Haverty


On 3/9/20 11:14 PM, vinton cerf via Internet-history wrote:
> Steve Kirsch asks in what languages NCP and TCP were written.
>
> The Stanford first TCP implementation was done in BCPL by Richard Karp.
> Another version was written for PDP-11/23 by Jim Mathis but not clear in
> what language. Tenex was probably done in C at BBN. Was 360 done in PL/1??
> Dave Clark did one for IBM PC (assembly language/??)
>
> Other recollections much appreciated.
>
> vint