Re: [Qemu-devel] [PATCH 2/5] virtio-blk: disable scsi passthrough for 1.0 device

[Cornelia Huck]

On Wed, 15 Jul 2015 16:16:07 +0300
"Michael S. Tsirkin" <address@hidden> wrote:

> On Wed, Jul 15, 2015 at 02:43:51PM +0200, Cornelia Huck wrote:
> > On Wed, 15 Jul 2015 15:01:01 +0300
> > "Michael S. Tsirkin" <address@hidden> wrote:
> > 
> > > On Wed, Jul 15, 2015 at 01:46:38PM +0200, Cornelia Huck wrote:
> > > > On Wed, 15 Jul 2015 13:59:00 +0300
> > > > "Michael S. Tsirkin" <address@hidden> wrote:
> > > > 
> > > > > On Tue, Jul 14, 2015 at 07:43:44PM +0200, Cornelia Huck wrote:
> > > > > > > Yes, and that's because as written, transitional devices must set
> > > > > > > ANY_LAYOUT, but that's incompatible with scsi.
> > > > > > 
> > > > > > Hm, I had a patch before that dynamically allowed different feature
> > > > > > sets for legacy or modern, not only a subset. Probably won't apply
> > > > > > anymore, but I'd like to able to do the following:
> > > > > > 
> > > > > > - driver reads features without negotiating a revision: driver is
> > > > > >   legacy, offer legacy bits
> > > > > > - driver negotiates revision 0: dito
> > > > > > - driver negotiates revision >= 1: driver is modern, offer modern 
> > > > > > bits
> > > > > > 
> > > > > > That way we could offer SCSI and !ANY_LAYOUT (if scsi is enabled) 
> > > > > > in the
> > > > > > first two cases, and a new qemu could still offer scsi to old 
> > > > > > guests.
> > > > > > 
> > > > > > Would it be worth pursuing that idea?
> > > > > 
> > > > > Frankly, I don't think so: I don't see why it makes sense
> > > > > to expose more features on the legacy interface than
> > > > > on the modern one. Imagine updating drivers to fix a bug
> > > > > and losing some features. How does this make sense?
> > > > 
> > > > I don't think one should be a strict subset of the other. But I think
> > > > we don't want to withdraw features from legacy guests on qemu updates
> > > > either?
> > > 
> > > Absolutely. For now one has to enable the modern interface
> > > explicitly. Around 2.5 we might switch that around, we'll
> > > need to think hard about compatibility at that point.
> > > In any case, we must definitely keep the old capability for old machine
> > > types.
> > 
> > ccw only offers revision 0 (legacy) in 2.4. I plan to introduce
> > revision 1 in 2.5 and force revision to 0 for 2.4 compatibility (as 2.4
> > is the first versioned ccw machine).
> 
> I was talking about pci here actually.

Sure, and these are my plans for ccw ;)

> 
> > > 
> > > > > 
> > > > > I think the virtio TC's assumption was that the scsi passthrough was a
> > > > > bad idea, so in QEMU we only keep it around for legacy devices to 
> > > > > avoid
> > > > > regressions.
> > > > 
> > > > I'm not opposing this :)
> > > > 
> > > > > 
> > > > > If you disagree and think transitional devices need the SCSI feature,
> > > > > either try to convince pbonzini or rewrite the spec youself
> > > > > to support it in the virtio 1 mode.
> > > > 
> > > > This seems to boil down to the different meaning of "transitional" for
> > > > ccw and pci, see the other thread.
> > > 
> > > Before the revision is negotiated, ccw won't know whether
> > > it's a legacy driver - is that the difference?
> > 
> > I'd say it doesn't know whether the driver intends to use the modern
> > interface.
> 
> That's also the case for pci.

But does pci know the moment it first tries to get the device's
features? And does pci assume modern as default for transitional
devices?

> 
> > > Fine, but revision is negotiated way before features are
> > > probed so why does it make a practical difference?
> > 
> > Legacy drivers (that don't know about the set-revision command) will
> > read features without revision negotiation - we need to offer them the
> > legacy feature set.
> 
> Right. So simply do if (revision < 1) return features & 0xffffffff
> and that will do this, will it not?

Not for bits that we want to offer for legacy but not for modern.