On Sat, Oct 03, 1998 at 01:49:18PM -0700, Andy Valencia wrote:
> [David Jeske <jeske@home.chat.net> writes:]
>
> >Andy: if a server setup a shared memory segment on 'physically mapped
> >pages', and then a client connects to it, is it currently possible for
> >the server to change the physical pages that shared memory segment
> >maps to after the fact?
>
> Hmmm, not really, since there's not (currently) any good way to name a
> shared object distinct from naming what it maps. I think it'd require a new
> page set (struct pset, psop_* stuff) object type. In fact, I think it would
> have to be a lot like a ZFOD set, with an operation to switch from the
> backing store memory set to a physical map set.
>
> >I believe this would be required for GGI to work, because the server
> >needs to be able to switch the client mapping from the 'real
> >framebuffer' to a 'backing store' without the client knowing about it.
>
> Delta some interesting race conditions. Switching the mapping of the page
> set isn't enough, because you also have to hammer the client's PTE's and TLB
> (if they're running when you do the switch). I think some of the current
> page stealing code would have to be generalized to support this.
>
> I'll do the VM work if somebody actually wants to try a GGI port!
I'll have to take another look at the current GGI stuff.. last I
looked, their stuff wasn't too ingrained into the 'Linux kernel view
of the world', but that was when they were doing proof of
concept. Things have certainly changed since then.
-- David Jeske (N9LCA) + http://www.chat.net/~jeske/ + jeske_at_chat.netReceived on Tue Oct 6 21:20:18 1998
This archive was generated by hypermail 2.1.8 : Thu Sep 22 2005 - 15:12:44 PDT