11 Aug
2023
11 Aug
'23
3:05 p.m.
Date: Thu, 10 Aug 2023 03:17:25 +0000
From: segaloco
As far as I can tell the history of this code line goes back to 1977, when Jack Haverty at
BBN wrote a TCP/IP library (porting earlier work written in PDP-11 assembler) for a
slightly modified 6th Edition Unix. Fighting with 64KB core limits, throughput was
horrific and he concluded that a bigger PDP-11 was needed. Mike Wingfield then did a
re-implementation in C for a PDP-11/70. This worked in early 1979 and is arguably the
first Unix TCP/IP stack that can still interoperate with current IPv4. However, it was
still mostly a proof of concept user mode design (it was funded as a test vehicle for the
later abandoned Autodin-II fork of TCP).
BBN then got a contract to write a kernel mode TCP/IP stack for 4BSD (“VAX TCP” in the old
BBN doc’s). This work was performed by Rob Gurwitz under supervision of Jack Haverty. This
stack - although all new code - still showed its heritage: it was designed as a loosely
bound kernel process providing the NCP-Unix API. Some sources seem to imply that it was
developed first as a user mode process and once working in that context changed into a
kernel process / thread. Beta releases were available in 1981. It worked (and
interoperates with modern IPv4), but in my experiments a few years back it turned out that
it is difficult to get the scheduling for this kernel process right at higher system
loads.
Bill Joy of CSRG concluded that the BBN stack did not perform according to his
expectations. Note that CSRG was focused on usage over (thick) ethernet links, and BBN was
focused on usage over Arpanet and other wide-area networks (with much lower bandwidth, and
higher latency and error rates). He then in 1982 rewrote the stack to match the CSRG
environment, changing the design to use software interrupts instead of a kernel thread and
optimising the code (e.g. checksumming and fast code paths). It was a matter of debate how
new the code was, with the extremes being that it was written from scratch using the spec
versus it being mostly copied. Looking at it with a nearly 50 year distance, it seems in
between: small bits of surviving SCCS suggest CSRG starting with parts of BBN code
followed by rapid, massive modification; the end result is quite different but retained
the ‘mbuf’ core data structure and a BBN bug (off-by-one for OOB TCP segments).
The shift from the NCP-Unix API to sockets is separate from this and was planned. CSRG had
the contract to develop a new API for facilitating distributed systems with Unix and this
gelled into the sockets interface. The first prototypes for this were done in 1981.
Nearly all of the above source is available in the TUHS online Unix Tree (Wingfield,
VAX-TCP and two early versions from CSRG - one in 2.9BSD and one in 4.1cBSD).
>> TCP/IP, not datakit
All of the files that have timestamps at the top list 83/07/29, except ip_input.c which
has 83/08/16 instead. The V8 version has _device (device driver) and _ld (line
discipline) components that the 4.1cBSD code does not have. Many other files have analogs
between the two. The byte ordering subroutines have been copied into a file, goo.s, from
their home in 4.1cBSD in the C library (/usr/src/lib/libc/net/misc). When this work
originated someone else would need to answer, [...] 
12 Aug
12 Aug
11:08 a.m.
New subject: when did v8 or later get networking?
On Fri, Aug 11, 2023 at 3:05 AM Paul Ruizendaal <pnr(a)planet.nl> wrote:
Bill Joy of CSRG concluded that the BBN stack did not
perform according to
his expectations. Note that CSRG was focused on usage over (thick) ethernet
links, and BBN was focused on usage over Arpanet and other wide-area
networks (with much lower bandwidth, and higher latency and error rates).
He then in 1982 rewrote the stack to match the CSRG environment, changing
the design to use software interrupts instead of a kernel thread and
optimising the code (e.g. checksumming and fast code paths). It was a
matter of debate how new the code was, with the extremes being that it was
written from scratch using the spec versus it being mostly copied. Looking
at it with a nearly 50 year distance, it seems in between: small bits of
surviving SCCS suggest CSRG starting with parts of BBN code followed by
rapid, massive modification; the end result is quite different but retained
the ‘mbuf’ core data structure and a BBN bug (off-by-one for OOB TCP
segments).
When Kirk McKusick tells the story, UCB got a beta release (or early
access) of the BBN stack. UCB was supposed to add the socket interface to
whatever was there. But Bill Joy found it performed terribly (multiple
seconds to connect sometimes, single digit kB over 10Mb media, etc). He
optimized it to make it perform well. This was a combination of rewriting
chunks and tweaking other chunks, which matches your analysis of SCCS. When
BBN came back with their new, release ready stack Bill supposedly said
something like 'no thanks, we already got one that works way better.' This
is why much of the structure of the original BBN stack survived the
rewrite: if there wasn't a big issue with them, the design and mechanisms
wound up being conserved by this effort. It was too much work to move from
mbuf to something else, and too little gain.
I tried to find a good link, but they are in his BSD history retrospective
talks to differing degrees. Sorry I don't have an exact reference.
Warner
11:41 a.m.
New subject: when did v8 or later get networking?
Someone who worked at BBN told me they had no overhang budget for
improvements, they wrote code to fixed price contracts with DARPA and
'maybe we could do that better' was impossible without a second grant.
The butterfly we had at UCL had issues. No support worth writing home for.
Pre BGP routing was a bit of a disaster. Fixed size prefix tables and LRU
ejection. You could time out a telnet to the USA before the login: prompt
assuming you even had a route.
"Diamond" their SGML multimedia mailer was great, but a one-shot.
Smart people. Very focused on the bottom line.
(Apologies if this offends anyone ex BBN it's recollection of coffee room
gossip from 1985)
G
On Sat, 12 Aug 2023, 3:08 pm Warner Losh, <imp(a)bsdimp.com> wrote:
On Fri, Aug 11, 2023 at 3:05 AM Paul Ruizendaal <pnr(a)planet.nl> wrote:
Bill Joy of CSRG concluded that the BBN stack did
not perform according
to his expectations. Note that CSRG was focused on usage over (thick)
ethernet links, and BBN was focused on usage over Arpanet and other
wide-area networks (with much lower bandwidth, and higher latency and error
rates). He then in 1982 rewrote the stack to match the CSRG environment,
changing the design to use software interrupts instead of a kernel thread
and optimising the code (e.g. checksumming and fast code paths). It was a
matter of debate how new the code was, with the extremes being that it was
written from scratch using the spec versus it being mostly copied. Looking
at it with a nearly 50 year distance, it seems in between: small bits of
surviving SCCS suggest CSRG starting with parts of BBN code followed by
rapid, massive modification; the end result is quite different but retained
the ‘mbuf’ core data structure and a BBN bug (off-by-one for OOB TCP
segments).
When Kirk McKusick tells the story, UCB got a beta release (or early
access) of the BBN stack. UCB was supposed to add the socket interface to
whatever was there. But Bill Joy found it performed terribly (multiple
seconds to connect sometimes, single digit kB over 10Mb media, etc). He
optimized it to make it perform well. This was a combination of rewriting
chunks and tweaking other chunks, which matches your analysis of SCCS. When
BBN came back with their new, release ready stack Bill supposedly said
something like 'no thanks, we already got one that works way better.' This
is why much of the structure of the original BBN stack survived the
rewrite: if there wasn't a big issue with them, the design and mechanisms
wound up being conserved by this effort. It was too much work to move from
mbuf to something else, and too little gain.
I tried to find a good link, but they are in his BSD history retrospective
talks to differing degrees. Sorry I don't have an exact reference.
Warner
3:06 p.m.
New subject: when did v8 or later get networking?
On 12 Aug 2023, at 07:08, Warner Losh
<imp(a)bsdimp.com> wrote:
On Fri, Aug 11, 2023 at 3:05 AM Paul Ruizendaal <pnr(a)planet.nl> wrote:
Bill Joy of CSRG concluded that the BBN stack did not perform according to his
expectations. Note that CSRG was focused on usage over (thick) ethernet links, and BBN was
focused on usage over Arpanet and other wide-area networks (with much lower bandwidth, and
higher latency and error rates). He then in 1982 rewrote the stack to match the CSRG
environment, changing the design to use software interrupts instead of a kernel thread and
optimising the code (e.g. checksumming and fast code paths). It was a matter of debate how
new the code was, with the extremes being that it was written from scratch using the spec
versus it being mostly copied. Looking at it with a nearly 50 year distance, it seems in
between: small bits of surviving SCCS suggest CSRG starting with parts of BBN code
followed by rapid, massive modification; the end result is quite different but retained
the ‘mbuf’ core data structure and a BBN bug (off-by-one for OOB TCP segments).
When Kirk McKusick tells the story, UCB got a beta release (or early access) of the BBN
stack. UCB was supposed to add the socket interface to whatever was there. But Bill Joy
found it performed terribly (multiple seconds to connect sometimes, single digit kB over
10Mb media, etc). He optimized it to make it perform well. This was a combination of
rewriting chunks and tweaking other chunks, which matches your analysis of SCCS. When BBN
came back with their new, release ready stack Bill supposedly said something like 'no
thanks, we already got one that works way better.' This is why much of the structure
of the original BBN stack survived the rewrite: if there wasn't a big issue with
them, the design and mechanisms wound up being conserved by this effort. It was too much
work to move from mbuf to something else, and too little gain.
I tried to find a good link, but they are in his BSD history retrospective talks to
differing degrees. Sorry I don't have an exact reference.
Warner
===
UCB got a beta release (or early access) of the BBN
stack.
This is certainly true. There are four surviving BBN tapes in the CSRG archive, from
memory the first was from August 1981 and the last from early 1982. Again from memory, the
oldest SCCS entries for the rewrite are from Oct ’81. All of this is on Kirk’s DVD.
Indeed the first tape is early code, written to just interface with Arpanet IMP’s. Things
like routing are rudimentary or hard-coded, etc. This is all filled out by the time of the
last tape, which also includes a token ring driver (written by Noel Chiappa, if I remember
well).
There are two editions of Mike Muuss’ TCP-DIGEST mailing list with posts on performance
from Joy and Gurwitz respectively (both from late 1981):
https://groups.google.com/g/fa.tcp-ip/c/WNE_j4mAbAE/m/3nCB79uvNcUJ
https://groups.google.com/g/fa.tcp-ip/c/zfYZh-kRlMg/m/pl-5oLQtYxIJ
Jonathan Gray sent me a good link of one of Kirk’s talks that addresses this bit of
history:
https://youtu.be/DEEr6dT-4uQ?t=706
===
Maybe Kirk was using some hyperbole in that talk. As far as I can tell the main issues
were the below:
- BBN coded the checksum routine in C and it compiled badly. Even Joy's
hand-optimised version still took some 25% of CPU when traffic maxed out.
- The time-out constants were 2s and 5s. This makes sense for the Arpanet of the time. For
local ethernet this was changed to 0.2s and 0.5s.
- The BBN code used at lot of bit fields. This too compiled badly, and was later changed
to and/or with #define’d constants
- The BBN code took a very layered approach, often abstracting small bits of functionality
into a separate routine. Without compiler inlining and a somewhat slow VAX subroutine
mechanism this had a cost. Some of these functions where later changed to #define’s
instead of functions.
- Although Kirk singles out replacing the state machine with a big switch statement in the
above talk, I’m not sure this was a major performance boost. It is certainly the most
visible/recognisable change in the source though. Somehow, this seems to have become core
to the debate at the time (ref. the BBN talk at the Summer '84 Usenix conference).
Maybe I underestimate the impact of this aspect.
- The TCP management process ran as a kernel thread with normal scheduling. On a loaded
machine this meant that it could take seconds for this process to be scheduled again.
Changing this to a software interrupt mechanism (somewhat similar to the runrun flag,
although Joy appears to have credited VMS for the idea) made things more responsive and
avoided context switches.
Of all the above, it would seem to me that only the last point was fundamental to the
design. The other things appear relatively easy to fix.
===
Doug McIlroy wrote: "I recall expressions of surprise (dismay?) at the size of the
BSD internet code, but without it we'd have lost our place in the Unix
community."
An interesting question is how much of this size is unavoidable. Just counting file lines
(i.e. ‘wc -l’), 4.2BSD has some 2200 lines of TCP code (i.e. just the TCP part), the ‘82
BBN implementation 2400 lines, 8th edition had some 2400 lines, and the first version of
Plan9 some 2200 lines.
Some parts of TCP could perhaps be simplified (pseudo headers and ‘urgent' segments
come to mind), but not much. Maybe this is just the code size that it takes.
URP is smaller, but if I remember well it does not handle out-of-order packets or packet
duplicates (both of which do not occur in a (virtual) line switched context).
On the other hand, these sizes compare with 900 lines for the IL protocol in early Plan9.
4:29 p.m.
New subject: when did v8 or later get networking?
URP is smaller, but if I remember well it does not
handle out-of-order packets or packet duplicates (both of which do not occur in a
(virtual) line switched context).
Upon reflection, duplicates of course can occur when an ACK is lost.
9:20 p.m.
New subject: when did v8 or later get networking?
On Sat, Aug 12, 2023, 4:29 AM Paul Ruizendaal <pnr(a)planet.nl> wrote:
What's URP? Sounds a little like UDP, is it a typo or something else?
Warner
URP is
smaller, but if I remember well it does not handle out-of-order
packets or packet
duplicates (both of which do not occur in a (virtual)
line switched context).
Upon reflection, duplicates of course can occur when an ACK is lost.
9:24 p.m.
New subject: when did v8 or later get networking?
On Sat, Aug 12, 2023 at 11:20 AM Warner Losh <imp(a)bsdimp.com> wrote:
On Sat, Aug 12, 2023, 4:29 AM Paul Ruizendaal
<pnr(a)planet.nl> wrote:
What's URP? Sounds a little like UDP, is it a typo or something else?
I thought that might be a typo for UDP, but that doesn't have ACKs.
URP is
smaller, but if I remember well it does not handle out-of-order packets or packet
duplicates (both of which do not occur in a (virtual) line switched context).
Upon reflection, duplicates of course can occur when an ACK is lost.
10:12 p.m.
New subject: when did v8 or later get networking?
Universal Receiver Protocol, the protocol that Sandy Fraser invented for Datakit. The best
description of it (imho) is the patent:
https://patentimages.storage.googleapis.com/37/a9/70/46f8c44ebd4e3a/US48521…
At first the patent is very focused on hardware implementation, but towards the end it
gives the algorithms in pseudo-code.
Sandy Fraser wrote a 1993 retrospective paper “Early Experiments with Asynchronous Time
Division Networks”. This paper includes a short summary of the origins of the universal
receiver protocol, and how it evolved through 3 main steps. Unfortunately behind a
paywall.
The implementation for 8th edition (some 750 lines) is here:
https://www.tuhs.org/cgi-bin/utree.pl?file=V8/usr/sys/dev/dkp.c
I think early Plan9 had an implementation of URP as well.
On 12 Aug 2023, at 17:24, Dan Cross
<crossd(a)gmail.com> wrote:
On Sat, Aug 12, 2023 at 11:20 AM Warner Losh <imp(a)bsdimp.com> wrote:
On Sat, Aug 12, 2023, 4:29 AM Paul Ruizendaal
<pnr(a)planet.nl> wrote:
What's URP? Sounds a little like UDP, is it a typo or something else?
I thought that might be a typo for UDP, but that doesn't have ACKs. URP is
smaller, but if I remember well it does not handle out-of-order packets or packet
duplicates (both of which do not occur in a (virtual) line switched context).
Upon reflection, duplicates of course can occur when an ACK is lost.
498
days inactive
499
days old
7 comments
4 participants
participants (4)
-
Dan Cross -
George Michaelson -
Paul Ruizendaal -
Warner Losh