[TUHS] looking for 4.1a BSD full kernel source

Clem Cole clemc at ccc.com
Fri Dec 2 07:28:32 AEST 2016


below...

On Thu, Dec 1, 2016 at 3:13 PM, Paul Ruizendaal <pnr at planet.nl> wrote:

> Clem,
>
> Many thanks for your informative post!

​Most welcome.​


>
>
> > Assuming I can read the tape, I know I do have a copy of 4.1a
> distribution on 9-track.
> That is great news. Let me know once you had a chance to look at it. I am
> in no hurry, so whenever you have time, even if that is months from now.

​My queue keeps getting bigger.  I need to retire so I have time for my
home projects ;-)​



>
>
> > Diomidis is correct, 4.1a use Joy/Cooper/Leffler reimplementation of of
> the BBN stack, not the original BBN stack.
> I suspected as much, but I am happy to hear it confirmed. I've been told
> on several occasions that 4.1a contained the original BBN stack.

​Truth is the basic stuff will be pretty similar.   The big thing they
share is mbuf code and how basic IP/TCP state machines.   The primary
changes will be that it's not using open/nami and started to get rid of the
ioctl hacks IIRC.​  It's been a long time since I looked at those stacks.

The original BBN release was >>tuned<< for 4.1BSD but was based on their
"portable" IP/TCP release.  So it's what was used for the HP3000 and few
other systems that DARPA wanted IP stacks.    Again this is more obvious
when you look at the mbuf stuff.  Rob needed a specific OS kernel
independent way of handling memory, so he wrote his own handler and then
spliced the few things it needed into the local kernel.

Joy kept that code for a long time, although in later and later versions of
the BSD stack (i.e. by NET2) the mbuf code has been rehashed by many and
deviated from Rob's stuff in many ways.




>
>
> > The BBN stack (Gurwitz et al) was for 4.1 (and other UNIX and non-Unix
> systems).   I do not believe I have a copy of that tape.  Rob or maybe Eric
> Cooper might.
> >
> > How Sam added the code into the UCB SCCS, I never knew (you can ask him
> directly, he is still findable these days).   Eric Cooper took the basic
> BBN distribution and put it on the UCB 4.1 systems around campus >>before<<
> Joy started the sockets work i.e. the installation was done by installing
> 4.1, then taking the BBN tape and installing it as is.
>
> I can confirm all this. I do have several copies of the BBN tapes from '81
> and '82, they survived in Kirk McKusick's archive.

​You mean on his CD's - if so can you send me a path.  I'll try to peak at
them.  I have the CD's at home, although not online.​




> They indeed contain material that is 'copied over' a clean 4 or 4.1 build
> tree. The first complete BBN beta is from May 1981 and Joy started on his
> network code in October 1981.

​FYI:  I don't remember it as a beta, but you might be right.  I thought it
was the official distribution.   BTW it supported the BBN 1822 interface
and the Xerox 3Mbit boards with the Xerox taps on the Vax.  I wrote a 3Com
driver for it and Sam and I hacked up an InterLan and DEC drivers when we
finally got 10M cable.     We had had a single (3Mbit) Xerox cable that
went through Cory and over to the 5th floor of Evan and then hit its limit.
  But at 3Mbit, it was a huge step of from the Berk-Net (9600 baud serial
over DZ-11 ports).​

3Com, DEC and InterLan all were working on Unibus ethernet board at
different stages of readiness.   3Com must have been shipping for about
18-24 mons by then, because I had used them on Tektronix (I was their first
customer) but UCB was using the Xerox gear when we started.   DEC had
donated some of their gear to the CAD team, so we had those board before
Joy/Sam did in Evans.   Since I had one "pure DEC" system of our 3 780s in
the CAD group (most of the Vaxen at UCB had non-DEC gear), I was often an
early "test system" for Sam.




>
>
> > BTW: about 3 years later, the BBN2 stack comes out from Rob and team and
> it is actually even more interesting; because it now uses the sockets
> interface (not the Chaosnet/UofI nami trick), and adds a number of both
> performance enhancements (Van Jacobson's work, etc.) as well as a more
> complete implementation of the IP stack.   I >>might<< have a copy of that
> tape squirreled away; but I'll have to look.
>
> I think this might be the material that appears in the BSD source tree in
> 1985, right?

​I don't remember/know if it ever got put directly into the BSD tree.   Rob
released it independent of the BSD kit and you needed and BBN license to
get it.   It was what we used at Stellar​, because it was easier to make
parallel and we were adding it into a more System V then BSD-ish kernel
(the Stellar box was a 4 "core" system in its CPU).

There was definitely some bad blood at the time.  BBN had the contract to
support IP/TCP not UCB.  So the stacks did diverge.   Most (commercial)
people used the BSD 4.2 ( and later NET2) implementation because they got
it from UCB and did not get the separate BBN license.  Arpanet contractors
got the BBN2 code automatically, but I don't think many of folks installed
it unless they needed some feature that BBN had that UCB did not.  The
National Labs are likely to have picked it up; but not all of the
Universities.




> Van Jacobson's work is ~1988, I think, but I could well be mistaken.

​Van was at LBL (up the hill as we used to say).    He started with the BSD
4.2 code, but he was talking with Rob pretty regularly.   I know by the
time we did Stellar,  IIRC:  both Van and Rob​ were part of the DARPA
advisors committee (predecessor to the IETF).  I do remember when we were
at Stellar, Van and Rob were talking because we were doing the
threading/parallel lock stuff and I'm pretty sure Rob traded some of our
work for something Van was doing at the time.



> I think the main performance improvement between the 1982 and the 1985
> version was a switch from a kernel thread model to a software interrupt
> model, but as yet I haven't grasped why this is better and my assumption
> may be incorrect.

​I should know, but frankly do not remember.   I suspect if I started to
stare at the code, some of those bits will refill the cache in my brain.
IIRC it was related to the amount of state that needed to be
saved/switched.  In the thread model, I believe you need the complete
context, but in the Interrupt model and are sharing whatever context the
kernel has at the time of the interrupt.​


​Clem​
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://minnie.tuhs.org/pipermail/tuhs/attachments/20161201/b543e35d/attachment.html>


More information about the TUHS mailing list