Re: [MiNT] MMU usage in MiNT, fork/mmap

["Evan K. Langlois" <Evan@CoolRunningConcepts.com>]

On Thu, 2005-04-07 at 23:30 +0200, Xavier Joubert wrote:
> Hi Evan,
> 
> 
> Selon "Evan K. Langlois" <Evan@CoolRunningConcepts.com>:
> > OK, I remember some older machines doing that - think my 3b2 works
> that
> > way.  That's gotta be a real performance killer though having to
> copy a
> > process at fork and again at every task switch.  Is anyone actively
> 
> You don't have to copy any data. A mere switch between two MMU trees
> is enough.

Thats my point.  Right now, MiNT does copies instead of sharing the text
pages and doing a copy-on-write for the data pages and just mapping the
memory with the MMU.  It doesn't have to if it would use the MMU, but I
remember talk about being compatible with the old 68000 and no MMU, and
no one wanted to add in the "has-an-MMU" case when the "no-MMU" case
already worked, no matter how inefficient.

> > working on a real virtual address space for MiNT?  I have a feeling
> this
> > will eventually be necessary.
> 
> I wrote a full virtual memory engine for TOS some years ago. ASM
> sources are
> available on Patrice Mandin's site (
> http://membres.lycos.fr/pmandin/download/src/sysinc.zip ). I was
> young. I
> didn't know mmap() at this time. I would write many things differently
> today.

Hmm .. was this about 10 years ago maybe?  I think I remember it, or
something similar.  Probably lots of ideas we can pull from looking at
the Linux and FreeBSD ports to the 68K as well.

> It's been a long time I would like to work on implementing this in
> MiNT. But I
> never found time to study current code. We need to implement a virtual
> address
> space first. To be clean, this should interact closely with disk cache
> management.

Agreed!

> > How different are the MMUs between the different processors?  Looks like
> > I need to get that stupid cross-compiler working soon.
> 
> 040 and 060 MMUs are a subset of 030 MMU. Don't use any 030 specific stuff.
> 
> Each processor family (030, 040 and 060) has different way to configure its MMU
> and to recover from access fault. We need a code path for each.
> 
> MMU trees are compatible.

Thats about what I figured.  Some minor differences in the setup and
other details, but the bulk would be compatible.  I suppose we should
also be thinking about Coldfire as well, as the 4e core has an MMU.
Believe it's almost identical to the 060's, but I'd have to check.

-- Evan