Geoff Collyer

Plan 9 and (l)unix system programmer and administrator.

I've been working for Google since September 2014.

From 1994 through 2001 and again from 2006 to 2014, I was a Member of Technical Staff at Bell Labs (originally a part of AT&T, now a part of Alcatel-Lucent) in Murray Hill, NJ, where I worked on operating systems and messaging systems and protocols, among other things.

pgp public key(s)

My current RSA public key was generated by PGP 2.6.2 (i.e. it's in the second of three or four incompatible formats) and its fingerprint is

Type bits/keyID    Date       User ID
pub  1024/1F08CB75 1996/06/24 Geoff Collyer <geoff@collyer.net>
                              Geoff Collyer <geoff@world.std.com>
                              Geoff Collyer <geoff@bell-labs.com>
          Key fingerprint =  67 08 1E F6 6B F3 1B 1B  CF 27 87 7B 3D 59 30 A1

ranting and raving

the sad state of crytographic infrastructure, 17 May 2012

I've been looking at ssh and IPsec again and it's made me realise again that we (in the computing field) have made a mess of the artifacts of cryptography and it's entirely our own fault. Starting from goals such as privacy and authentication of communication, we have taken aim at our feet and fired with

oids (object ids)
the X.509 binary encoding used for key certificates
dozens of certificate `authorities'
protocols with so many optional parts and choices that systems can barely interoperate
an array of binary key formats
and more

PGP is a sad example: by changing key formats and algorithms repeatedly (sometimes for compelling legal reasons), interoperability is so impaired that you pretty much need to know in advance which version of PGP each recipient uses.

The triumph of ssh v2 (a classic second system) over v1 is another example: whatever the theoretical weaknesses of v1 were, it was surely better to encrypt one's communication than not. We need a v3, a stripped-down and streamlined version closer to v1 than v2.

We seem to be better at building great steaming piles of crypto (e.g., OpenSSL) than getting to the core of what needs to be done and doing just that.

i hate git, 17 December 2015

I have used plenty of revision control systems and don't really like any of them (though I can use them), but I truly detest git, which I find unusable. It's slow, wasteful, bureaucratic, and over-complicated, which does make it a good match for Linux, for which it was written.

computing experience

ported Plan 9 to Xilinx Virtex 4 and 5 Power-PC-based evaluation boards, ARM-based systems including the Sheevaplug, Gumstix Overo and Trimslice, and MIPS-based systems including the Routerboard RB450G.
wrote an NVMe driver for Plan 9.
maintained the original Plan 9 file server kernel (now upgraded to 64-bit sizes) and contribute to Plan 9 generally.
- merged the big-endian and little-endian MIPS compilers, assemblers and loaders.
- wrote the current (new) Plan 9 tar, which is cleaner and faster than the old one, and is POSIX-compliant (i.e., can cope with longer path names).
- modified the file server kernel to use 64-bit file sizes, offsets and block numbers internally, ported the CPU kernel's IDE code, thus enabling the use of DMA, RWM, etc., and ported the CPU kernel's Intel gigabit Ethernet driver.
- added greylisting and other anti-spam measures to smtpd.
- added multicasting (for IPv6) and shutdown entry points to drivers for all modern ethernet cards. updated the connection server to recognise icmpv6. fixed the kernel to recognise v6 addresses starting with ``20'' as global.
- ported dbm and mdbm libraries and made them use a consistent byte order on disk, rendering the databases portable across CPU architectures.
- arranged completely diskless booting of Plan 9 PCs using compact flash cards
- added device mirroring (think RAID 1) to the Plan 9 file server kernel. Combined with support for IDE disks, this enables use of inexpensive hardware to form a large Plan 9 file server.
- set up and ran the largest Plan 9 file server (with optical disc jukebox) inside Bell Labs other than the main Plan 9 development file server. Extended support for jukeboxes with optical disks of varying sizes. The smaller jukeboxes can now be had second-hand quite cheaply.
on the inferno distributed operating system, built a scalable message store that served a broken-out view of MIME messages, and designed and implemented rsmtp, an SMTP replacement. Got Inferno running on Mac OS X.
configuring IPv6 tunnelling networks
developed the operating system for the now-defunct AT&T HomeCenter™ set-top box, notably DSP communication, and the C library.
Unix device drivers (e.g. for Tektronix storage (display) tubes (for V6 Unix), Nuclear Medicine scintillation counters (for V6 Unix), terminal multiplexors, disks), bootstraps, and other kernel work (e.g. putting /tmp in memory).
reverse-engineering (``decompiling'') manual pages from nroff output to troff -man input: nam
invented the term const poisoning during development of the ANSI C standard
Unix system programming and running Unix systems (including 8th and 9th edition) for almost a decade at the University of Toronto, notably for the computing center, undergraduate computer science instruction, and the statistics department.
set up the original bell-labs.com servers during the ATT/Lucent split, notably DNS and mail servers on Plan 9 and internal and external web servers.

publications

assigning interrupt priorities in software via interrupt queueing, in Computing Systems, Spring 1996
evolution of operating systems: Research Unix in html and postscript (updated 23 Dec 1998 to more closely match the printed article in Microsystems in 1984). Also, a new, improved draft that fixes a number of errors we missed the first time around and adds real references, in html and postscript.
software reimplementation: the shell, including stak.c from the appendix, and a newer, tidier stak.c that eliminates struct alignment problems that showed up on the Alpha. I've retrofitted the above memory-management, directory-reading and a few other changes into a V7 sh that should run on POSIXy systems. It's been run on the BSDs and Linux on x86 and amd64 PCs, and ARM64 systems without trouble. Among its other features, it lacks the GNU readline library, so what you type is what you get. Update! Mon Mar 7 04:03:13 PST 2005: To get it running on 64-bit Suns, I had to change two declarations; v7sh.tar and the newer stak.c have been updated.
performance tuning and more software reimplementation: C news netnews software
programming style: no ifdefs, can't happen (postscript)
automated game referees: othello

computing interests

Plan 9, the successor to Unix; a distributed system drawing on over 20 years of experience with Unix
- my own contributions of Plan 9 software
- an early outline of a proposed-but-rejected book about why Plan 9 matters. Pardon the formatting; it was converted to HTML mechanically.
the inferno distributed system, derived from Plan 9, and the limbo language, which features particularly helpful support for processes sharing address space. If POSIX pthreads give you the willies, limbo might be for you. inferno runs on bare 386 or Sparc hardware and as a `guest OS' on the usual (l)Unixes and MS Windows.
archiving and optical media, notably CD-R. Plan 9 has a convenient backup system using optical disks, discussed in its overview, file server and port papers. The venti block store is a worthy replacement when combined with fossil, the file server built on top of venti.
indexing. This is the distribution of my nov news overview database code. Nov is now incorporated into essentially all serious newsreaders.
software reimplementation: More Taste, Less Greed
strong cryptography as an aid to privacy and civil rights

UTF and Unicode examples

Geoff Collyer
geoff at collyer.net