Re: [Qemu-devel] [PATCH RFC 4/7] vhost: set vring endianness for legacy virtio

[Cornelia Huck]

On Tue, 12 May 2015 18:40:35 +0200
"Michael S. Tsirkin" <address@hidden> wrote:

> On Tue, May 12, 2015 at 06:25:32PM +0200, Cornelia Huck wrote:
> > On Tue, 12 May 2015 17:15:53 +0200
> > "Michael S. Tsirkin" <address@hidden> wrote:

> > > One issue with virtio 1 patches as they are is with how features are
> > > handled ATM.  There are 3 types of features
> > > 
> > >   a. virtio 1 only features
> > >   b. virtio 0 only features
> > >   c. shared features
> > > 
> > > and 3 types of devices
> > >   a. legacy device: has b+c features
> > >   b. modern device: has a+c features
> > >   c. transitional device: has a+c features but exposes
> > >      only c through the legacy interface
> > 
> > Wouldn't a transitional device be able to expose b as well?
> 
> No because the virtio 1 spec says it shouldn't.

Can you point me to the part where it says that? Can't seem to find it.

> > > So I think a callback that gets features depending on guest
> > > version isn't a good way to model it because fundamentally device
> > > has one set of features.
> > > A better way to model this is really just a single
> > > host_features bitmask, and for transitional devices, a mask
> > > hiding a features - which are so far all bits > 31, so maybe
> > > for now we can just have a global mask.
> > 
> > How would this work for transitional presenting a modern device - would
> > you have a superset of bits and masks for legacy and modern?
> 
> Basically we expose through modern interface a superset
> of bits exposed through legacy.
> F_BAD for pci is probably the only exception.

ccw has F_BAD as well.

NOTIFY_ON_EMPTY is in the MAY be offered category.

I'm still wondering about the old legacy bits. Transitional devices not
offering them is easiest to implement from the host side, but I'm
wondering if all legacy devices will be able to function with missing
bits? Is breaking some legacy drivers OK?