Re: rtVAX300 .. adding an CPU Type..

[Anders Magnusson <ragge%ludd.ltu.se@localhost>]

On 01/04/2013 09:29 PM, Jason Thorpe wrote:
> On Jan 4, 2013, at 2:39 AM, Mouse <mouse%Rodents-Montreal.ORG@localhost> wrote:
>
> > > > As some of you already know I'm fiddeling around with an ISA Board
> > > > in a PC that has an rtVAX 300 CPU on it.
> > > The primary difference between the rtVAX and normal VAX is that the rtVAX re$
> > It's not totally clear to me, yet, whether the rtVAX 300 works this way
> > or whether it has the normal VAX P0 and P1 setup.  The ROM code makes
> > me moderately sure that either it's a normal VAX in this respect or it
> > ignores the high bit of the P0BR - all ROM code that sets P0BR sets the
> > high bit except for one case which I suspect is an expected-to-fail
> > test.
> >
> > > This means all the process page tables have to be physically contiguous (as $
> > Perhaps proper KA620 (and other machines with that style of P0/P1 page
> > tables) support would mean building a pmap that migrates pages in an
> > attempt to maintain a large physically contiguous free block?
> ...or uses a single page table and zeroes / lazy-copies from process-specific 
> tables as needed.  (OMG that would be so slow... but easy :-)
Actually, after last pmap rewrite on vax (10-15 years ago? :-)  it works 
in a way similar to that.

The usrptmap is preallocated in kernel space and can at most
handle (currently) processes of size 1GB.
Space for processes are allocated out of this submap, and since
most processes are much much smaller than 1GB this is normally not a 
problem,
but if suddenly a process need lot of memory it will unmap some idle 
processes
and use their space in the usrpt.

This was very effective as it seems since:
- It's not often processes needs much memory.
- It do not need so much preallocated memory on the kernel page table.

1GB of user virtual memory takes 8MB kernel virtual memory
which in turn takes 64K physical memory.  If the indirection is missing
then the page tables would take 8MB physical memory instead,
or less, if the usrpt is shrunk in size.

-- Ragge