17 Jun
2022
17 Jun
'22
8:54 p.m.
As one of the few remaining people who has actually worked
with the original dmr stream I/O system, I'd love to dive
into the debate here, but I don't have time. I can't resist
commenting on a few things that have flown by, though. Maybe
I can find time to engage better on the weekend.
-- If you think there were no record delimiters in the stream
system, you don't know enough about it. A good resource to
repair that is Dennis's BSTJ paper:
https://www.bell-labs.com/usr/dmr/www/st.pdf
It's an early description and some of the details later
evolved (for example we realized that once pipes were
streams, we no longer needed pseudoterminals) but the
fundamentals remained constant.
See the section about Message blocks, and the distinction
between data and control blocks. Delimiters were one kind
of control; there were others, some of them potentially
specific to the network at hand. In particular, Datakit
(despite being a virtual-circuit network) had several sorts
of control words, including message delimiters. Delimiters
were necessary even just for terminals, though: how else
does the generic read(2) code know to return early, before
filling the buffer, when a complete line has arrived?
-- It's hard to compare sockets to streams and make much
sense, because they are very different kinds of thing.
When people talk about sockets, especially when complaining
that sockets are a mess, they usually mean the API: system
calls like connect and listen and getpeername and so on.
The stream system was mainly about the internal structure--
the composable modules and the general-purpose queue interface
between them, so that you could take a stream representing
an (already set up) network connection and push the tty
module on it and have a working remote terminal, no need
for user-mode programs and pseudo-terminals.
It's not inconceivable to build a system with socket-like
API and stream internals.
-- Connection setup was initially done with network-specific
magic messages and magic ioctls. Later we moved the knowledge
of that messy crap into network-specific daemons, so a user
program could make a network call just by calling
fd = ipcopen("string-destination-name")
without knowing or caring whether the network transport was
TCP or Datakit or involved forwarding over Datakit to a
gateway that then placed a TCP call to the Internet or whatnot.
That's what the connection server was all about:
https://www.bell-labs.com/usr/dmr/www/spe.pdf
Again, the API is not specific to the stream system. It
wouldn't be hard to write a connection server that provided
the same universal just-use-a-string interface (with the
privileged parts or network details left up to daemons)
on a system with only socket networking; the only subtle
bit is that it needs to be possible to pass an open file
descriptor from one process to another (on the same system),
which I don't think the socket world had early on but I
believe they added long ago.
-- There's nothing especially hard about UDP or broadcast.
It's not as if the socket abstraction has some sort of magic
datagram-specific file descriptor. Since every message sent
and every message received has to include the far end's
address info, you have to decide how to do that, whether
by specifying a format for the data (the first N bytes are
always the remote's address, for example) or provide an
out-of-band mechanism (some ioctl mess that lets you
supply it separately, a la sendto/recvfrom, and encodes it
as a control message).
There was an attempt to make UDP work in the 9th/10th edition
era. I don't think it ever worked very cleanly. When I
took an unofficial snapshot and started running the system
at home in the mid-1990s, I ended up just tossing UDP out,
because I didn't urgently need it (at the time TCP was good
enough for DNS, and I had to write my own DNS resolver anyway).
I figured I'd get around to fixing it later but never did.
But I think the only hard part is in deciding on an interface.
-- It's certainly true that the Research-system TCP/IP code
was never really production-quality (and I say that even
though I used it for my home firewall/gateway for 15 years).
TCP/IP wasn't taken as seriously as it ought to have been
by most of us in 1127 in the 1980s. But that wasn't because
of the stream structure--the IP implementation was in fact
a copy of that from 4.2 (I think) BSD, repackaged and
shoehorned into the stream world by Robert T Morris, and
later cleaned up quite a bit by Paul Glick. Maybe it would
have worked better had it been done from scratch by someone
who cared a lot about it, as the TCP/IP implementors in the
BSD world very much did. Certainly it's a non-trivial design
problem--the IP protocols and their sloppy observance of
layering (cf the `pseudo header' in the TCP and UDP standards)
make them more complicated to implement in a general-purpose
framework.
Or maybe it just can't be done, but I wish someone had
tried in the original simpler setup rather than the
cluttered SVr4 STREAMS.
Norman Wilson
Toronto ON