[TUHS] Sockets and the true UNIX

[Paul Ruizendaal]

Gentlemen,

Below some additional thoughts on the various observations posted
about this. Note that I was not a contemporary of these developments,
and I may stand corrected on some views.

> I'm pretty sure the two main System V based TCP/IP stacks were STREAMS
> based: the Lachman one (which I ported to the ETA-10 and to SCO Unix)
> and the Mentat one that was done for Sun.  The socket API was sort of
> bolted on top of the STREAMS stuff, you could get to the STREAMS stuff
> directly (I think, it's been a long time).


Yes, that is my understanding too. I think it goes back to the two
roots of networking on Unix: the 1974 Spider network at Murray Hill and
the 1975 Arpanet implementation of the UoI.

It would seem that Spider chose to expose the network as a device, whereas
UoI chose to expose it as a kind of pipe. This seems to have continued in
derivative work (Datakit/streams/STREAMS and BBN/BSD sockets respectively).

When these systems were developed networking was mostly over serial lines,
and to use serial drivers was not illogical (i.e. streams->STREAMS). By 1980
fast local area networks were spreading, and the idea to see the network as
a serial device started to suck.

Much of the initial modification work that Joy did on the BBN code was to
make it perform on early ethernet -- it had been designed for 50 kbps arpanet
links. Some of his speed hacks (such as trailing headers) were later
discarded.

Interestingly, Spider was conceived as a fast network (1.5 Mbps); the local
network at Murray Hill operated at that speed, and things were designed to
work over long distance T1 connections as well. This integrated fast LAN/WAN
idea seems to have been abandoned in Datakit. I have a question out to Sandy
Fraser to ask about the origins of this, but have not yet received a reply.

> The sockets stuff was something Joy created to compete with the CMU Accent
> networking system.  [...]  CMU was developing Accent on the Triple Drip
> PascAlto (aka the Perq) and had a formal networking model that was very clean
> and sexy.  There were a lot of people interested in workstations, the Andrew
> project (MIT is about to start Athena etc). So Bill creates the sockets
> interface, and to show that UNIX could be just as modern as Accent.

I've always thought that the Joy/Leffler API was a gradual development of
the UoI/BBN API. The main conceptual change seem to have been support for
multiple network systems (selectable network stack, expansion
of the address space to 14 bytes).

I don't quite see the link to Accent and Wikipedia offers little help here
https://en.wikipedia.org/wiki/Accent_kernel
Could you elaborate on how Accent networking influenced Joy's sockets?

> * There's no reason for
>   a separate listen() call (it takes a "backlog" argument but
>   in practice everyone defaults it and the kernel does strange
>   manipulations on it.)

Perhaps there is. The UoI/BBN API did not have a listen() call;
instead the open() call - if it was for a listening connection - blocked until
a connection occurred. The server process would then fork of a worker process
and re-issue the listening open() call for the next connection. This left a
time gap where the server would not be 'listening'.

The listen() call would create up to 'backlog' connection blocks in the
network code, so that this many clients could connect simultaneously
without user space intervention. Each accept() would hand over a (now
connected) connection block and add a fresh unconnected one to the backlog
list. I think this idea came from Sam Leffler, but perhaps he was inspired
by something else (Accent?, Chaos?)

Of course, this can be done with fewer system calls. The UoI/BBN system
used the open() call, with a pointer to a parameter data block as the 2nd
argument. Perhaps Joy/Leffler were of the opinion that parameter data
blocks were not very Unix-y, and hence spread it out over
socket()/connect()/bind()/listen() instead.

The UoI choice to overload the open() call and not create a new call
(analogous to the pipe() call) was entirely pragmatic: they felt this
was easier for keeping up with the updates coming out of Murray Hill
all the time.

> In particular, I have often thought that it would have been a better
> and more consistent with the philosophy to have it implemented as
> open("/dev/tcp") and so on.

I think this is perhaps an orthogonal topic: how does one map network names
to network addresses. The most ambitious was perhaps the "portal()" system
call contemplated by Joy, but soon abandoned. It may have been implemented
in the early 90's in BSD, but I'm not sure this was fully the same idea.

That said, making the the name mapping a user concern rather than a kernel
concern is indeed a missed opportunity.

Last and least, when feeling argumentative I would claim that connection
strings like "/dev/tcp/host:port" are simply parameter data blocks encoded
in a string :^)

>   This also knocks out the need for
>   SO_REUSEADDR, because the kernel can tell at the time of
>   the call that you are asking to be a server.  Either someone
>   else already is (error) or you win (success).

Under TCP/IP I'm not sure you can. The protocol specifies that you must
wait for a certain period of time (120 sec, if memory serves my right)
before reusing an address/port combo, so that all in-flight packets have
disappeared from the network. Only if one is sure that this is not an
issue one can use SO_REUSEADDR.

> Also, the profusion of system calls (send, recv, sendmsg, recvmsg,
> recvfrom) is quite unnecessary: at most, one needs the equivalent
> of sendmsg/recvmsg.

Today that would indeed seem to make sense. Back in 1980 there seems
to have been a lot of confusion over message boundaries, even in
stream connections. My understanding is that originally send() and
recv() were intended to communicate a borderless stream, whereas
sendmsg() and recvmsg() were intended to communicate distinct
messages, even if transmitted over a stream protocol.

Paul