Re: virtio-pci new configuration proposal

["Michael S. Tsirkin" <mst@redhat.com>]

On Tue, Nov 08, 2011 at 08:51:38PM +1030, Rusty Russell wrote:
> On Tue, 8 Nov 2011 08:32:50 +0200, "Michael S. Tsirkin" <mst@redhat.com> wrote:
> > On Tue, Nov 08, 2011 at 10:23:33AM +1030, Rusty Russell wrote:
> > > Even so, should we just use the PCI cap list, and have each
> > > cap entry just contain a BIR & offset?
> > > 
> > > Thanks,
> > > Rusty.
> > 
> > And size :)
> > I say, Rusty, did you see my patch? That's what it's doing,
> > I also addressed the issue that with KVM on x86, PIO is faster
> > than MMIO so we need to use it for notifications/isr.
> 
> Faster?  Really?  Why?

I think, because it doesn't need to be emulated to get the address
and the value. There are so many ways on x86 to do memory
access, you need a lot of code to figure out what's going on.

> Were the ppc guys right that's it's obsolescent?

For PCI - surely not.  PCI Express spec tries to discourage its usage.
But support is there even in Express.

> Because PIO is going to be ugly AFAICT for every non-x86 platform.
> 
> Cheers,
> Rusty.
> 


So we should make the BAR MMIO on non-x86 host, PIO on x86.
Guest should just map and use what it's given.

-- 
MST