Re: virtio-pci new configuration proposal

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

On Fri, Nov 04, 2011 at 03:53:24PM +0200, Sasha Levin wrote:
> On Fri, 2011-11-04 at 15:51 +0200, Michael S. Tsirkin wrote:
> > On Fri, Nov 04, 2011 at 02:32:19PM +0200, Sasha Levin wrote:
> > > On Fri, 2011-11-04 at 13:40 +0200, Michael S. Tsirkin wrote:
> > > > On Fri, Nov 04, 2011 at 08:14:43PM +1030, Rusty Russell wrote:
> > > > > > > 3) If we're changing the queue layout, it's a chance to fix a
> > > > > > >    longstanding bug: let the guest notify the host of preferred
> > > > > > >    queue size and alignment.
> > > > > > 
> > > > > > Yup, we can do that.
> > > > 
> > > > We don't need to change all of layout for that - just add another field
> > > > in the common config structure to supply the alignment.
> > > 
> > > How would you do it without changing the layout? Add another optional
> > > field at the end which will shift offsets based on whether the host and
> > > guest support this new feature or not?
> > > 
> > > This leads to 3 different things which now shift config offsets around.
> > 
> > No. Just put the field at offset 24 from the offset specified
> > by VIRTIO_PCI_CAP_COMMON_CFG.
> 
> Two questions here:
> 
> - What about backwards compatibility? How would the config space look
> when we're not using the new layout?

Exactly as it does now. You don't get to tweak alignment then.

> - How does it work with 64 bit features which are also located there?

Basically each field gets an offset.  E.g.
24 - features
28 - queue alignment

> > > As you said, the PCI cap list was introduced both to save space (which
> > > is not the motivation here), and because it's a very efficient
> > 
> > It's actually pretty inefficient - there's an overhead of 3 bytes for
> > each vendor specific option.
> 
> It's efficient because while you pay a small price for each optional
> option it also means that that option is optional and won't clutter the
> config space if it's not really in use.

I guess my assumption is that most options will be in use,
not discarded dead-ends.

> Think of how the PCI config space would look if all those caps wouldn't
> have been optional and would instead all of them would have just have
> been attached to the end of the config space.

It started out this way, but then they started running out
of space - it's only 256 bytes - so the capability mechanism
was invented.


> > 
> > > and easy way to manage optional features without requiring tricks
> > > which move offsets around like we do now.
> > 
> > Tricks with offsets only appeared because we had datapath, device
> > specific and common config in the same place.
> > feature list isn't needed to fix that.
> > 
> > > -- 
> > > 
> > > Sasha.
> 
> -- 
> 
> Sasha.