Re: [PATCH] [v2] linux: virtio: Standardize virtio's concept of "page size"

[Rusty Russell <rusty@rustcorp.com.au>]

On Thursday 13 November 2008 02:46:31 Hollis Blanchard wrote:
> On Wed, 2008-11-12 at 22:51 +1030, Rusty Russell wrote:
> > On Tuesday 11 November 2008 10:07:09 Hollis Blanchard wrote:
> > > Both sides of the virtio interface must agree about how big a pfn
> > > really is. This is particularly an issue on architectures where the
> > > page size is configurable (e.g. PowerPC, IA64) -- the interface must be
> > > independent of PAGE_SHIFT.
> > >
> > > Currently there are three distinct problems:
> > > * The shift count used when passing the physical address of the ring to
> > > a PCI-based back end.
> > > * The ring layout itself is padded to span at least two "pages".
> > > * The balloon driver operates in units of "pages".
> >
> > Hi Hollis,
> >
> >    The more I thought about this, the more I think we're not solving this
> > as neatly as we could.  The trigger was noting that we're breaking the
> > userspace API (vring_size and vring_init are exposed to userspace): I
> > know that qemu cut & pastes, but that's no excuse.
> >
> >    So instead, I've introduced separate constants for each use.  Yes,
> > all these constants are 12/4096.  But just to be contrary, at the end
> > is a patch to change lguest to 128.  And there's no reason this
> > couldn't change in future using some guest detection scheme.
>
> OK. I thought it was simpler to just say "4KB everywhere" in all aspects
> of the virtio interface, but I'm happy as long as we solve the problem
> somehow. :)

It is simpler, yes, but we can take this opportunity to deconflate them and 
make things clearer and better than the current code, not just "fix" it.

Note that I still don't have a balloon patch: want to send me one?

Thanks,
Rusty.