Re: [Qemu-devel] [PATCH 3/4] s390/kvm: Add a channel I/O based virtio transport driver.

[Cornelia Huck <cornelia.huck@de.ibm.com>]

On Tue, 14 Aug 2012 09:40:01 +0930
Rusty Russell <rusty@rustcorp.com.au> wrote:

> On Mon, 13 Aug 2012 10:56:38 +0200, Cornelia Huck <cornelia.huck@de.ibm.com> wrote:
> > On Wed, 08 Aug 2012 13:52:57 +0930
> > Rusty Russell <rusty@rustcorp.com.au> wrote:
> > 
> > > On Tue,  7 Aug 2012 16:52:47 +0200, Cornelia Huck <cornelia.huck@de.ibm.com> wrote:
> > > 1) Please don't limit yourself to 32 feature bits!  If you look at how
> > >    virtio_mmio does it, they use a selector to index into a
> > >    theoretically-infinite array of feature bits:
> > 
> > It should be easy to extend the data processed by the feature ccws to a
> > feature/index combination. Would it be practical to limit the index to
> > an 8 bit value?
> 
> 256 feature bits?  That seems like it could one day be limiting.  Or an
> 8 bit accessor into feature words?  8192 seems enough for anyone sane.

An 8 bit accessor. I hope everybody stays on the sane side :)

> 
> > > Note that we're also speculating a move to a new vring format, which
> > > will probably be little-endian.  But you probably want a completely new
> > > ccw code for that anyway.
> > 
> > Do you have a pointer to that discussion handy?
> > 
> > If the host may support different vring formats, I'll probably want to
> > add some kind of discovery mechanism for that as well (what discovery
> > mechanism depends on whether this would be per-device or per-machine).
> 
> It would be per-machine; per-device would be a bit crazy.  We'd
> deprecate the old ring format.
> 
> There's been no consistent thread on the ideas for a ring change,
> unfortunately, but you can find interesting parts here, off this thread:
> 
> Message-ID: <8762gj6q5r.fsf@rustcorp.com.au>
> Subject: Re: [RFC 7/11] virtio_pci: new, capability-aware driver.

I've read a bit through this and it looks like this is really virtio-2
or so. How about discoverability by the guest? Guests will likely have
to support both formats, and forcing them to look at the feature bits
for each device in order to figure out the queue format feels wrong if
it is going to be the same format for the whole machine anyway.