> at least on the SPARC the kernel has to do
> callbacks to the PROM to set up VM mappings in other contexts.  (The
> kernel can't do this itself because one can't set up mappings in
> non-current contexts, but one can't run in a context that doesn't have
> the kernel mapped in.  Apparently the hardware has a mode where all
> memory references go to the PROM, which makes it easy - for the PROM.)

I was only vaguely aware of sparc "contexts" and certainly didn't
know about this characteristic.  That's pretty weird, but I guess
I can see how hardware designers might come up with such a solution
- especially if they thought that mappings in non-current contexts
would only happen at boot time.

> I'm sure someone who really knows those ports will correct me if I'm
> wrong here. :)

This discussion has been enlightening, for me at least, and I appreciate
all the input.

All these weird permutations just seem to validate the idea that
these really are machine dependent issues although there still may
be room for machine independent abstractions (like the work toward 
more common standalone code that seems to be underway).

Jason