> > The problem is the parts of the kernel page table that must map the
> > user page tables, that must be in physical contiguous memory and
> > therefore allocated at boot.
> 
> I must be missing something.
> 
> Mapping 128M of user space requires 1M of user page tables in kernel
> virtual space; mapping 1M of kernel virtual space requires 8K of
> (physically contiguous) kernel page tables.
> 
Correct, but you must multiply this with the number of possible processes
(maxproc) that can be on your system. If you can have 1000 processes then 
it will be 8M instead.

> This doesn't sound too bad - especially since many of those pages of
> user PTEs will usually be identical no-access pages that could be
> collapsed (though I think I saw someone say that the VAX pmap doesn't).
> Is the pmap statically allocating MAXPROC times this 8K of space or
> something?
> 
The unused pages do not even have to be mapped. The only problem is the
system page table size. (that is statically allocated because it must
be located in physical contiguous memory).

> Also, do we support the KA620?  If so, I could see trouble finding
> physically contiguous space for the user page tables, but the VAX
> supported hardware list doesn't include the KA620 as far as I can see.
> 
No, it is not supported. With another handling of the user page tables
it could be supported (with some penalty) but because I lack any HW
to test it on it's unlikely to ever happen :-)

-- Ragge