Re: [PATCH v2] port/xen: map memory directly in privcmd PRIVCMD_MMAPBATCH

[Manuel Bouyer <bouyer%antioche.eu.org@localhost>]

On Mon, Jun 25, 2012 at 06:35:15PM +0100, Roger Pau Monne wrote:
> Cherry G.Mathew wrote:
> >Hi Roger,
> >
> >
> >
> >     Roger>  Instead of mapping memory when a pagefault happens in the
> >     Roger>  memory region, map the memory directly on the pivcmd call (as
> >     Roger>  Linux does).  This fixes the problem with Qemu upstream, that
> >     Roger>  changes the location of the pfns after calling
> >     Roger>  xc_map_foreign_bulk.
> >
> >I'm not sure why exactly this change is needed, could you explain it
> >again ? Why is it not ok for the backing mfns to be lazy mapped in ?
> 
> I've tried explaining the problem here:
> 
> http://mail-index.netbsd.org/port-xen/2012/06/15/msg007293.html
> 
> Hopefully it's a good starting point.

In this mail, are 0x3f800 and 0xf0000 virtual, physical or machine addresses ?

> 
> >I see that you use pmap_enter() directly to map the MAs. This means that
> >uvm(9) doesn't know about the mapping, and if the qemu process were to
> >be swapped out of memory, the mapping would not be restored on a
> >subsequent swap in. Is this okay ?
> 
> Linux maps those pages as VM_RESERVED which prevents them from being
> swapped out, should we do the same (PMAP_WIRED)?

Probably, yes.

-- 
Manuel Bouyer <bouyer%antioche.eu.org@localhost>
     NetBSD: 26 ans d'experience feront toujours la difference
--