Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Path: utzoo!mnetor!seismo!husc6!cmcl2!beta!unm-la!unmvax!hi!hc!ames!lll-lcc!lll-tis!ptsfa!hoptoad!gnu
From: gnu@hoptoad.uucp (John Gilmore)
Newsgroups: comp.sys.m68k
Subject: Re: How does one implement VM in a 68000?
Message-ID: <2351@hoptoad.uucp>
Date: Thu, 2-Jul-87 00:23:38 EDT
Article-I.D.: hoptoad.2351
Posted: Thu Jul  2 00:23:38 1987
Date-Received: Sat, 4-Jul-87 04:16:48 EDT
References: <1029@bloom-beacon.MIT.EDU>
Organization: Nebula Consultants in San Francisco
Lines: 36

wpd@athena.mit.edu (William P Doyle) wrote:
>                         ...because the 68000 could not restart
> instructions, VM wasn't possible.  His argument went something like,
> "You know that you will never page fault while in the supervisor mode,
> so you just insert probes in the user mode."

Early 68000 Unix systems worked this way; they did not do virtual memory,
though; they swapped, like V7 Unix and most System V's until recently.  
(The most serious constraint this gives is that you can't run programs
bigger than your physical memory.)

The "probe" method was used to dynamically allocate the stack.  Each
procedure prologue had a "tst sp@(-128)" instruction, or something similar.
If the stack was close to its end, this would get a bus error.  Enough
was known about the behaviour of the 68000 that restarting that simple
instruction was possible; e.g. we knew how far off the PC was from where
it should be.  If we didn't do the stack probe, some random instruction
that referenced the stack might get a bus error, and not enough information
would be there to restart it.

On the Macintosh, I understand that Apple did something similar for
procedure calls to non-resident libraries; they figured out what the
chip would do with an instruction like "jsr a0@" to a bad address, and
if it happens, they fix it up.

Apollo used to do virtual memory using a pair of 68000's; when the first
one took a bus error for a missing page, it would be frozen in mid-cycle,
and the second 68000 would take over, fetch the page, and resume the first.
This worked, but had the serious drawback that you can't run some other
process while waiting for the disk to spin.

68010's are now so cheap that there is no excuse for using a 68000 for a
system with an MMU.
-- 
{dasys1,ncoast,well,sun,ihnp4}!hoptoad!gnu	       gnu@ingres.berkeley.edu
Alt.all: the alternative radio of the Usenet. Contributions welcome - post 'em