31 Dec
2022
31 Dec
'22
12:24 a.m.
During the past year I’ve worked on and off on porting Sys III. My approach to studying
ancient Unix is to port it to a new system, as that forces one to consider many of the
design trade-offs in a code base. For 16 bit Unix I used the TI990 as a target, for the
1980-1985 period I was thinking about a 68030 based board. However, in the end using a
Riscv based target seemed more interesting and that is what I used (more specifically the
64-bit Allwinner D1 chip, for which several cheap boards are available, along with Fabrice
Ballard’s TinyEMU for simulation).
In general, my approach was to first select a toolchain and I settled on the Plan9
compiler suite, with the Riscv backend as done by Richard Miller, used together with a
library that enabled using this tool chain on modern Unix. This I then used to port the
educational XV6 system to the D1 chip, building on work done by Michael Engel. From there
I ported the SysIII kernel (reusing some bits of XV6 memory code), initially using the XV6
user land. From there I went on to port the original SysIII user land. Much to my
surprise, the original SysIII code can support the Plan 9 tool chain, just a dozen
standard libc routines needed to be added to give this SysIII port a native compiler.
There is a lot to report, and I will split it over several posts covering different
aspects. I hope that this will not be perceived as spamming this list. The topics are:
1. Riscv, Plan-9 and a toolchain
2. Porting the SysIII kernel: first steps & memory management
3. Porting the SysIII kernel: boot, config & device drivers
4. A few comments on porting the Bourne shell
5. Xv6, Linux and SysIII on a RV32 FPGA system
For those who just want to see the results, the tool chain is here:
https://gitlab.com/pnru/riscv-kencc
And the SysIII port is here:
https://gitlab.com/pnru/SysIII_rv64
Begin forwarded message:
From: Paul Ruizendaal <pnr(a)planet.nl>
Subject: Early Unix on (64-bit) Risc-V
Date: December 22, 2022 at 11:55:04 AM GMT+1
To: The Eunuchs Hysterical Society <tuhs(a)tuhs.org>
Cc: athornton(a)gmail.com, davida(a)pobox.com, luther(a)makerlisp.com
Adam Thorton wrote:
I mean all I really want for Christmas is a
64-bit v7 with TCP/IP support, a screen editor, and SMP support.
The third one is a solved problem. The second one would not be that hard to adapt, say,
uIP 0.9, to v7. That first one would require some work with C type sizes, but getting
larger is easier than the reverse. It's that last one.
Having said that...maybe what I really want is 64-bit 4.3 BSD?
I mean, just a Unix, without all the cruft of a modern Linux, but which can actually take
advantage of the resources of a modern machine. I don't care about a desktop, or
even a graphical environment, I don't care about all the strange syscalls that are
there to support particular databases, I don't care much about being a virtualization
host.
Luther Johnson wrote:
I'm in the process of building a system like
that for myself, but
perhaps a little smaller - mine will be based on an embedded
microprocessor I've developed (so much work still yet to do ! at least a
year out).
Earlier this year I ported VAX System III to Risc-V, to a simple Allwinner D1 based SBC.
This is RV64GC. Just ported to the console terminal.
It turned out that porting Sys III to 64 bit was surprisingly easy, most of the kernel
and user land appears to be 64 bit clean. It helps that I am using a LLP64 compiler,
though. Apart from networking Sys III also feels surprisingly modern (for an ancient Unix)
- its should get more attention than it does. The hardest work was in porting the VAX
memory code to Risc-V page tables (and to a lesser extent, updating libc for the different
FP formats).
The code is currently in an ugly state (with debug stuff in commented-out blocks, a mix
of ansi and K&R styles, an incoherent kludgy build system, etc.) and the shame stopped
me from putting it out on gitlab until I found enough time to clean this up. As there
seems to be some interest now, I’ll put it up anyway in the next week or so. There you go
Adam, half your wish comes true.
The kernel is about 60KB and most binaries are quite close in size to the VAX
equivalents.
My next goals for it are to re-implement the Reiser demand paging (I think I have a good
enough view of how that worked, but the proof of the pudding will be in the eating), and
to add TCP/IP networking, probably the BBN stack. Making it work on RV32 and exploring
early SMP work is also on my interest list.
===
David Arnold wrote:
I think xv6 does SMP? (param.h says NCPU = 8, at
least).
You’d need to add a network stack and a userland, but there are options for those …
For the above, making xv6 work on the D1 board was my first stepping stone, to proof the
tool chain and to get the basics right (hardware init, low-level I/O, etc.).
As an educational tool, I am sure that xv6 hits all the right spots, and it certainly
does SMP (the D1 is single hart, so I have not tried that myself). I like it a lot in that
context. However, as a simple Unix it is useless: from a user-land view it is less capable
than LSX. At minimum it needs fixes to make the file system less constrained.
In my view, for SMP Keith Kelleman’s work for Sys-V is probably a better place to start.