edler@cmcl2.UUCP (07/17/84)
The 68451 MMU is available in 64 pin DIP format, and we currently are using them at 10MHz with 10MHz 68010s. You can certainly do virtual memory with them, with fixed or variable sized segments. Jan Edler New York University cmcl2!edler edler@nyu (212) 460-7314
jcp@BRL-TGR.ARPA (07/20/84)
From: Joe Pistritto <jcp@BRL-TGR.ARPA> The 68451 certainly does exist and is available, (although not in massive quantities). The only format i have seen it in is the square LCC package though. In particular, more than one can be used at a time, and Omnibyte Corp. of Chicago makes a board called OB68K/MMU or something like that which has a 68010 with one to four 68451s on board. By the way, the types of memory management provided by the 68010 and the 68451 are DIFFERENT, the 68010 is for VIRTUAL memory implementations, (and by itself is used in the Sun processor, with custom memory management hardware), and the 68451 performs MEMORY RELOCATION only. The one thing you do need to do is always use a 68010 in ANY scheme that does dynamic memory management, (to wit, most UNIX implementations), as the 68000 has unrecoverable faults when a bus-error occurrs. (Not enough of the internal state of the CPU is accessable to re-run the bad instruction) The most ANNOYING thing is that the address pushed on the stack by the exception is NOT always the address of the bad instruction, (or the next instruction), it seems to be the pre-fetch pointer for the instruction pipeline, which can be in the middle of a DIFFERENT instruction from the one which caused the trap. BELIEVE ME, I tried for a couple of weeks to enable a 68000 based system to recover from faults, its NOT possible easily, anyway... -JCP-
johan@apple.UUCP (07/20/84)
>> The one thing you do need to do is always use a 68010 in ANY scheme >> that does dynamic memory management, (to wit, most UNIX implementations), as >> the 68000 has unrecoverable faults when a bus-error occurrs. > >Also not true - Masscomp uses two 68000s. The main 68K halts on a page fault, >and the other 68K services the page fault and restarts the main 68K. Doing it >with one 68000 isn't possible (to my knowledge - maybe somebody out there >*has* squeezed blood from a stone). The apple Lisa uses a plain 68000 and *does* squeeze blood out of the stone. Memory is segmented and code is demand loaded. The trick is that *some* instructions are restartable. There is much magic and kludges involved. ____________________________Computare necesse est!____________________________ Johan Strandberg Apple Computer Education Research Group [ERG] {mtxinu,dual,nsc,voder}!apple!johan
patcl@tekecs.UUCP (Pat Clancy) (07/21/84)
The Lisa does do demand loading of text segments, but not
data. This is possible because all the 68K instructions which could
fault on an instruction fetch are restartable (eg.: branch).
Data segments must be pre-loaded. I would be interested in any
information anyone might have indicating how useful text-only
virtual memory is.
Pat Clancy
{ucbvax,decvax,ihnp4,allegra,uw-beaver,hplabs}!tektronix!tekecs!patclguido@mcvax.UUCP (Guido van Rossum) (07/25/84)
>The apple Lisa uses a plain 68000 and *does* squeeze blood out of the stone. >Memory is segmented and code is demand loaded. The trick is that *some* >instructions are restartable. There is much magic and kludges involved. A very early UNIX implementation on a 68000 by ACE in the Netherlands did the same. Essentially, the compiler generated such a restartable instruction at the beginning of each C procedure which tried to reach to maximum needed stack address. The memory fault this could cause was handled by the kernel, and the program restarted. I guess the Bourne shell caused them the most problems. -- Guido van Rossum, "Stamp Out BASIC" Committee, CWI, Amsterdam guido @ mcvax
guy@rlgvax.UUCP (Guy Harris) (07/26/84)
> The 68451 certainly does exist and is available, (although not in > massive quantities). The only format i have seen it in is the square > LCC package though. We make (to, in my opinion, our sorrow) a 68K-based UNIX box which uses the 68451, in a DIP the same size as the 68000 chip itself. > By the way, the types of memory management provided by the 68010 > and the 68451 are DIFFERENT, the 68010 is for VIRTUAL memory > implementations, (and by itself is used in the Sun processor, with > custom memory management hardware), and the 68451 performs MEMORY > RELOCATION only. Not true - you can combine a 68010 and a 68451 and make virtual memory. It could be "demand segment loading" (which is virtual memory - ask any Burroughs mainframe programmer), or you can even do demand paging (by using the 68451 as a sort of translation lookaside buffer, with fixed-size small (~2KB, say) "segments". > The one thing you do need to do is always use a 68010 in ANY scheme > that does dynamic memory management, (to wit, most UNIX implementations), as > the 68000 has unrecoverable faults when a bus-error occurrs. Also not true - Masscomp uses two 68000s. The main 68K halts on a page fault, and the other 68K services the page fault and restarts the main 68K. Doing it with one 68000 isn't possible (to my knowledge - maybe somebody out there *has* squeezed blood from a stone). Guy Harris {seismo,ihnp4,allegra}!rlgvax!guy