[Qemu-devel] Re: [PATCH 0/2] fix virtio_blk serial pci config breakage

[Anthony Liguori]

Michael S. Tsirkin wrote:
> On Tue, Sep 29, 2009 at 09:14:56AM -0500, Anthony Liguori wrote:
>   
> > Michael S. Tsirkin wrote:
> >     
> > > I don't think it's such a good idea.  We want to keep exposing most of
> > > config space in i/o bar because of backwards compatibility.  And we want
> > > to keep datapath operations such as vq kicks, in i/o space.
> > >   
> > >       
> > We can do feature negotiation in virtio-pci.  That lets us continue  
> > exposing the first 246 bytes of the config space in PIO and guests can  
> > negotiate a second bar for access to larger config spaces.
> >
> > Regards,
> >
> > Anthony Liguori
> >     
>
> guests can't negotiate a bar.
>   

The host always exposes the config in two places.  config[0..235] in 
bytes 20-255 in BAR0 and the full config space in BAR2 (or BAR1 + offset 
if desired).  Old drivers will just ignore the new config location.

> What I think we want to do
> - Use i/o for small fields (2-4 bytes) as we did previously
>   we won't run out anytime soon
> - For large fields, put them in memory BAR1 and report the
>   offset in i/o space
>   

This is at best a band aid as a large number of small fields can result 
in the same problem.

> - For huge fields, provide a mailbox in i/o space where guest
>   writes an offset and reads out a value
>   

Regards,

Anthony Liguori