MINIX - From the mouth of the horse
Andy Tanenbaum
porting MINIX to the Atari. For all I know a similar discussion is going on
in comp.sys.amiga and elsewhere. I am cross posting this to several groups
in case there are people there who are interested in porting MINIX to their
machine. If you missed the original note in mod.os.unix, I have just written
a UNIX clone that is available now with all the SOURCE CODE for $79.95 from
Prentice-Hall. There is also a book telling how it works inside.
I suggest that subsequent discussion go on in comp.sys.ibm.pc to avoid
scattering it all over the place. At the very least, crosspost comments to
comp.sys.ibm.pc under the subject MINIX. Later, we can set up comp.os.minix
if there is enough interest.
Read the comp.sys.atari group under the heading Forwarded for the last 2 weeks
to get the background for this note.
Although MINIX is copyright (not licensed), Prentice-Hall has agreed to permit
people to make a LIMITED number of copies for educational use, home study etc.
Posting the source code on the network (54K lines of C, kernel+utilities) is a
no no, but if each purchased copy doesn't generated more than say, 2 or 3
copies it is ok. While this is not public domain it is a lot better than Lotus,
Microsoft, Borland, and every other software company in the world's policy.
About the pricing. It is clearly not shareware, but by publishing the
source code on diskette, we are clearly not acting like AT&T either.
I had some discussion with people like Brian Kernighan (author of
the Software Tools package) and Doug Comer (author of Xinu, a little
embedded operating system for the LSI-11, although not really UNIX
like) and came to the following conclusion. I would like to see MINIX become
widespread, so the distribution mechanism is crucial. Having a major
publisher like Prentice-Hall advertise it, bring it to shows, send out
junk mail, etc. will get it a lot more attention than a note on the
net. Getting a commercial publisher like Prentice-Hall interested means
charging something. I think in the long run, this funny, copyrighted
but not real aggressive position we are taking will cause MINIX to become
widespread, new software to be made for it, etc. The GNU people are
upset because deep in their hearts they, too, know that people would rather
pay a reasonable price for good stuff than get empty promises for free.
Does anyone know how much GNU charges for its "free" software for the tape,
postage, handling etc? Berkeley generally charges something like $125
for its tapes, as I recall. If GNU also charges $125 for its "free" software
it seems to me that their moral indignation at Prentice-Hall's outrageous
$79.95 price is somewhat weakened. I mentioned MINIX in netnews last year
and I got quite a few reactions. I have also contacted lots of people for
various reasons. With two exceptions, everybody congratulated me and wished
me good luck. Some people, especially Martin Atkins, Charles Forsyth, and
Richard Gregg gave me a lot of help, for which I am grateful. Only two
people were really negative, almost bitter--both from the Free Software
Foundation (names withheld because I don't believe in character assassination
on a world-wide network). I am tempted to comment further, but I won't.
As to the port to the Atari/Amiga/etc as far as I see, there are no technical
problems with the MMU. The trouble is as follows. When you fork, the
child has to go somewhere else in memory than where the parent was.
Unfortunately, the child's stack contains absolute addresses, such as
the return from the fork routine. If the child runs somewhere other
than where the parent was, it will crash.
There are a couple of solutions, the simplest of which is this. When the
child is created, record in the process table where the parent was.
When it is time to run the child, just swap the parent and the child,
and actually run the child where it belongs. When the parent wants to
run, swap them again. Although this sounds horrendous, it is not at all
so bad. Swapping two 10K programs in memory might take 30 millisec.
MINIX programs are small. I am a strong believer in Small is Beautiful.
At the end of this note you will find the sizes of the MINIX utilities.
The only big ones are the compiler passes, cpp, cem, opt, cg, and asld.
Furthermore, 99.9% of the time, the child does an EXEC, at which point the
operating system can put the parent back where it belongs, and put the
new core image anywhere in memory. In practice, all this trick will
cost is about two copies of the forked core image, and it doesn't require
modifying the compiler. My experience with fragmentation is that it is not bad.
MINIX doesn't swap because one of the design goals was to have it run on CHEAP
hardware (meaning a 256K PC with 1 floppy disk) and floppies are not ideal as
swapping devices.
The MINIX memory management scheme is very simple, because the PC's hardware
is primitive. A core image consists of the text, the data, a gap, and then the
stack, growing downward. The stack and the data segments grow into the
gap. If they meet, BOOM! In practice, very few programs have wildly
growing stacks or data segments. I ran some statistics once, and for 90%
of the MINIX utilities, 2K stack is plenty. On the PC, the text is limited to
64K, and the data + stack is also limited to 64K. On a 68000, there would
be no need for such a limit. It comes from the 8088 architecture.
All you have to do is change a couple of constants in the memory manager.
The book is already out. You should be able to order it at any book store.
The title is Operating Systems: Design and Implementation. The software
will be out in three weeks. It went into production about three weeks ago,
and it takes about six weeks. Don't ask me why. Probably the same reason
as why it takes Prentice-Hall 18 months to produce a book from the finished
manuscript (unless you give them camera ready copy, as I do). If you want to
get the software (either on PC diskettes or 9 track tape), first order the book
and then send back the business reply card (software order form) in the book.
I would like to see a MINIX version for the Atari/Amiga/etc. The 68000 is
clearly much better than the 8088, but the PC has a lot of software going for
it. Maybe a tolerable UNIX clone might help the Atari/Amiga/etc in its fight
against the monster from 8088-land. A colleague of mine at Philips has
already started to port MINIX to the Atari. He is an absolutely top rate
programmer, but he is VERY busy, so he doesn't have much time. I think he
has already rewritten the MINIX assembly code (low level interrupt
handlers, etc) for the 68000. I will check with him one of these days; he
seems to be away right now. What I would like to find is someone who:
(1) knows the Atari (Amiga, Macintosh, etc) hardware well
(2) knows UNIX well on the outside and moderately on the inside
(3) has a substantial amount of free time
(4) has access to an IBM PC for testing things etc. (not essential, but helps)
Perhaps such a person could do the port with a little assistance from me.
Unfortunately I don't have much time either, as Prentice-Hall is bugging me to
revise a book on networks I wrote a million years ago.
The main things to do, other than the 68000 assembly code, are the device
drivers, all of which are in C, but of course are totally different for
the PC and Atari etc. The PC version doesn't use the BIOS at all, because the
stupid thing doesn't use interrupts. When you start a background job up
and then start up the editor in the foreground, calling the BIOS to read
a character would put the whole computer in a tiny loop sitting and waiting
for the keyboard to produce a character. MINIX supports the full UNIX
multiprogramming, so I had to write all the drivers from scratch (in C). I
suspect that the Atari BIOS isn't any better, although maybe we could use the
screen output BIOS.
And here we come back to the $79 again. If the person doing the port does a
good job, Prentice-Hall could sell the other version on diskettes, source code
and all, for the same $79 as the PC version. I have enough clout with P-H
that I think I could arrange that. Needless to say, the person doing the
port would be remunerated for his efforts, probably in the form of a royalty
on each disk set sold. The royalty is typically only a couple of dollars,
but that small amount is why capitalism works and socialism doesn't.
If anyone is interested, let me know. I don't think it will be that difficult,
but you have to plow through much of a very tightly written 719 page book and
understand a 12,000 line program before you can even start, so it will no doubt
be months of work. Also, debugging operating system code on a bare machine
even a relatively nice one like the 68000, will be a fair amount of work.
One other point is the compiler. The compiler is based on ACK, which is
described in Communications of the ACM, Sept. 1983, pp. 654-660. ACK is a big
system for writing compilers. It is being distributed by UniPress in Edison
NJ and Transmediair in Utrecht, Holland. It uses the old UNCOL idea of having
front ends that generate a common intermediate code and then back ends that
compile from that code to the target machine. At present we have front ends
for C, Pascal, Modula 2, Basic, Occam, and even a subset of Ada. There are back
ends for virtually every micro around, from the 6502 to the 68020. The ACK
software is owned by the university I teach at. UniPress pays them a royalty on
each copy they sell (academic price is $995 for a source tape containing 6
megabytes, although Modula 2 and Occam aren't on the tape yet). Our department
doesn't have much money, and we use the royalties to allow grad students to go
to conferences and the like. For these reasons the compiler kit is not part
of MINIX. Furthermore even the 8088 C compiler source by itself fills 4
diskettes. If the compiler source were in the MINIX distribution, that would
have meant raising the basic price to over $100, very much against my idea of
keeping the price low. I personally wrote MINIX in my spare time, which is why
it doesn't have to follow the same rules as ACK.
Nevertheless, the source of the 8088 MINIX C compiler is available
as a separate package from UniPress. I suspect that the easiest way to get
a 68000 C compiler is for someone at a university to have their university buy
the ACK tape and use that to develop the 68000 compiler. When it is done, it
will be necessary to negotiate a deal with UniPress to allow it to be sold,
but I know Mark Krieger, the president of UniPress, and he is a reasonable guy,
so I am sure some deal can be worked out that won't raise the price too much.
He is on the net (m...@unipress.uucp) if you have questions about all this.
The reason that I think this route is the easiest, is that ACK already has a
backend for the 68000, so there isn't much work to be done, but you really
need a VAX or a SUN or something like that to bring up the full ACK development
system. The compilers that are produced aren't so big, but the compiler-
compilers, and backend generators and the other meta-software doesn't really
fit easily on a PC. In addition to the 6 megabytes of source code on the
tape, you have to count on at least 20 megabytes of object files and working
space to compile everything. The 68000 compiler has been running for years
and it is pretty good. We recently rewrote the backend table for the
68000, 68010, and 68020 and the code quality seems very good (about 15% better
than the C compiler Motorola sells). I haven't even thought about using the
Pascal, Modula 2, Basic etc. front ends because I wanted the system to fit on,
and be able to recompile itself on a system without, a hard disk. This
succeeded. Technically there shouldn't be any big problem with the other
front ends. Note that UniPress has TWO packages: 8088 MINIX C compiler, and
full ACK. The former is 4 diskettes; the latter is 6 megabytes on a mag tape.
Andy Tanenbaum (mi...@cs.vu.nl, or in emergencies, mi...@vu44.uucp)
Sizes of the MINIX commands
Program Text Data Bss Total (all sizes in decimal bytes)
sync: 424 20 26 470
clr: 714 28 156 898
update: 836 42 58 936
sleep: 848 58 58 964
ln: 1070 86 74 1230
chmod: 984 98 156 1238
basename: 1060 66 156 1282
tee: 1228 82 94 1404
kill: 1240 78 156 1474
umount: 782 758 26 1566
touch: 1348 76 156 1580
sum: 1438 84 156 1678
mknod: 930 738 26 1694
mount: 890 782 26 1698
pwd: 1546 84 172 1802
mkdir: 1656 158 58 1872
split: 1656 132 130 1918
rev: 886 42 1052 1980
rmdir: 1802 196 188 2186
cat: 1212 722 540 2474
mv: 2444 244 58 2746
echo: 648 24 2076 2748
stty: 2336 260 170 2766
rm: 2468 276 30 2774
df: 2030 948 156 3134
time: 2584 666 266 3516
lpr: 1314 126 2114 3554
comm: 1822 104 2400 4326
tr: 1470 82 2852 4404
login: 3376 1436 28 4840
chmem: 3178 278 2074 5530
size: 3102 232 2208 5542
tar: 4010 456 1774 6240
head: 2762 168 3484 6414
wc: 4460 208 2104 6772
su: 3548 1498 2076 7122
dd: 4966 532 2148 7646
chown: 3522 2148 2074 7744
date: 5338 412 2104 7854
od: 5388 288 3654 9330
passwd: 4976 1764 2620 9360
pr: 5776 514 3110 9400
sort: 6404 694 2490 9588
grep: 6740 694 2208 9642
uniq: 4956 850 5150 10956
cc: 4404 826 5986 11216
cc.at: 4404 834 5986 11224
gres: 8496 768 2076 11340
tail: 4550 184 7200 11934
ar: 4900 488 11350 16738
mkfs: 8708 906 7376 16990
roff: 12742 820 4710 18272
cp: 1914 876 16412 19202
make: 15216 1976 3614 20806
shar: 980 82 20506 21568
cmp: 3372 224 18468 22064
dosread: 7440 3688 11992 23120
mined: 15680 2198 5308 23186
sh: 21536 1668 1310 24514
ls: 7164 584 17994 25742
cpp: 16896 3764 8580 29240 (Pass 1 of the C compiler)
asld: 14048 7882 7412 29342 (Pass 5 of the C compiler)
opt: 16400 9368 9494 35262 (Pass 3 of the C compiler)
cg: 24816 22968 10520 58304 (Pass 4 of the C compiler)
cem: 59856 10656 3164 73676 (Pass 2 of the C compiler)
to...@mit-eddie.uucp
> ... The GNU people are
>upset because deep in their hearts they, too, know that people would rather
>pay a reasonable price for good stuff than get empty promises for free.
First, I would like to commend ast for doing MINIX, and going a large part of
the way towards giving MINIX its freedom.
Second, GNU isn't an empty promise. GNU Emacs is out there. GDB (GNU's
Debugger) is out there. Bison, a YACC compatible Parser Generator, is out
there. The GNU C compiler (highly optimizing with VAX, 68000, and 68020 code
generators) will be released soon. Etc.
The remaining large undone piece is the kernel. Work has started on that,
and its being leveraged off of existing code for a Unix style kernel, Trix,
written at MIT a while back.
GNU is a more ambitious project than MINIX, and rms hasn't had much more help
than ast. Most of rms' help has been volunteer. rms has also been working on
it for a shorter period of time.
Third, none of the GNU people I know of are upset. We are just sad that yet
more software has been chained up.
>Does anyone know how much GNU charges for its "free" software for the tape,
>postage, handling etc? Berkeley generally charges something like $125
>for its tapes, as I recall. If GNU also charges $125 for its "free" software
>it seems to me that their moral indignation at Prentice-Hall's outrageous
>$79.95 price is somewhat weakened.
First, "free" doesn't refer to cost, but to the freedom of the software.
Second, I would like to present some comparisons between GNU, MINIX, and
Unix. I know the facts are straight for GNU, correct me on the others.
GNU MINIX Unix
--- ----- ----
Is source code distributed? Yes Yes For many more $$
How many copies of the source can
you give away, legally? Unlimited 3-4 None
Can one legally restrict use by others? No YES YES
Can one legally post it on USENET? Yes NO NO
Can one legally ARPA ftp it, freely? Yes NO NO
Cost of non-ARPA distribution from
home organization: $ 150. $ 80. Many times more.
People are referred to:
- the GNU Public License
- the GNU Manifesto
- Minix's Licensing arrangements (I have yet to see these)
- AT&T and susbsidiary vendor Unix Licenses
for further details.
happy hacking, len tower
--
Len Tower, Project GNU of the Free Software Foundation
1000 Mass. Ave., Cambridge, MA 02138, USA +1 (617) 876-3296
HOME: 36 Porter Street, Somerville, MA 02143, USA +1 (617) 623-7739
UUCP: {}!mit-eddie!mit-prep!tower INTERNET: to...@prep.ai.mit.edu
to...@mit-eddie.uucp
of a...@cs.vu.nl (Andy Tanenbaum). I apologize for the delay (I was
keyboard-less at USENIX for the last week ;-} ).
When Andy Tanenbaum announced his plans for MINIX, I told him that he
could certainly use any of GNU in MINIX, as long as he followed the
terms, which say that everyone must be able to redistribute it in any
quantity to anyone. Also, I said that if he produced something that
fit the GNU system and was suitably available, I would use it.
I don't think this was an antagonistic response.
But I wasn't interested in more than passive cooperation, for two
reasons.
One was that the technical goals were very different and I doubted
that any code written for one system would really be suitable for the
other. He planned a small system to fit the machines now common. I
am aiming for a more powerful system that people will prefer to 4.2 or
system V, to run on the next generation of machine. Each of these
paths has its advantages and disadvantages which I'm sure the reader
can see.
The other is that I doubted that MINIX would ultimately be available
on terms that would allow GNU to use it. I wasn't interested in
investing any effort on it until this doubt was resolved. Now
it appears the resolution is that GNU can't use it.
Meanwhile, Tanenbaum hasn't used any GNU software, perhaps because
some is too big for today's IBM PC's or perhaps because GNU copylefts
would not permit their distribution on Prentice Hall's terms.
I do not understand why Tanenbaum calls the GNU project "empty
promises". Several pieces of GNU software are already in
distribution, complete with fanatical admirers and detractors.
I think we have demonstrated that we can deliver what we promise.
There is no charge whatever for using GNU software for any purpose.
The Free Software Foundation charges for mailing tapes, but this is
not the same as a charge for the software on the tape. That is free,
and you can make as many copies as you like for anyone at all. The
Free Software Foundation is a tax-exempt charitable organization and
the money that tape distribution brings in is spent on the creation of
more free software. (None of it goes to me personally.)
The GNU C compiler will be released for testing soon. It compiles
itself, GNU Emacs and Monardo's free TeX-in-C successfully, so it is
not far from ready. And it will be free, with sources. (TeX-in-C is
still being tested; the Free Software Foundation and probably others
will distribute it by and by. There will be announcements.)
Further questions on GNU, GNU mailing lists, and the availability
of GNU software can be directed to g...@prep.ai.mit.edu or
mit-eddie!prep!gnu or seismo!prep.ai.mit.edu!gnu.
rms (Richard M. Stallman) is directly reachable at
<r...@prep.ai.mit.edu>. Please realise that any time you spend
communicating with him will delay the delivery of GNU software, by the
time it takes him to read and reply.
happy hacking, -len tower