Re: [Qemu-devel] [PATCH V2 0/5] Set correct blk feature for virtio 1.0

[Cornelia Huck]

On Wed, 15 Jul 2015 14:49:43 +0300
"Michael S. Tsirkin" <address@hidden> wrote:

> On Wed, Jul 15, 2015 at 11:11:18AM +0200, Cornelia Huck wrote:
> > On Wed, 15 Jul 2015 13:29:56 +0800
> > Jason Wang <address@hidden> wrote:
> > 
> > > Hi all:
> > > 
> > > This series tries to set feature correctly for virtio-blk when virtio
> > > 1.0 is supported. Two isssues were addressed according to the spec:
> > > 
> > > - scsi passthrough was not support in 1.0. This is done by, 1) disable
> > >   scsi by defautl for 2.4 machine type and fail the initialization
> > >   when both scsi and 1.0 were set.
> > > - any layout must be set for transitional device. This is done by set
> > >   any layout when 1.0 is supported.
> > > 
> > > Please review
> > > 
> > > Changes from V1:
> > > - Split virtio-net changes out of the series
> > > - Enable VIRTIO_BLK_F_SCSI only when scsi is set
> > > - Disable scsi by default and compat it for legacy machine types
> > > - Let get_features() can fail and fail the initialization of
> > >   virito-blk when both 1.0 and scsi were supported.
> > 
> > Hm, this seems confusing to me mainly due to the different way
> > transitional devices are handled by pci and ccw.
> > 
> > For virtio-pci: (please correct me if I misunderstood)
> > - devices (except input) are transitional by default
> > - user can disable legacy or modern support
> > - drivers can use method they prefer, depending on VERSION_1
> > 
> > For virtio-ccw:
> > - transitional means "in limbo" regarding legacy or modern
> > - devices become legacy if features are read without negotiation of a
> >   revision, or if revision 0 is negotiated
> > - they become modern if revision 1 is negotiated
> > 
> > That implies that for ccw, a transitional device does not offer any
> > features: It either transitions to the legacy or modern state if a
> > revision was negotiated, or it becomes a legacy device by reading
> > features (which are the legacy features, then). While pci has "real"
> > transitional devices needing to offer a certain feature set.
> > 
> > I'm not sure what the solution is here: basically, ccw needs a dynamic
> > feature set, while pci does not.
> 
> I think the issue is theoretical since linux guests
> do not read feature bits > 31 unless revision 1
> is negotiated.

Again, VERSION_1 is not really the problem. (qemu clears it for
revisions < 1, and I plan to send a patch that effectively clears all
bits > 31.) I'm caring about the legacy bits (scsi, in that case) that
are not send to legacy drivers.