Re: [qemu-s390x] [Qemu-devel] [PATCH 1/3] vfio-ccw: add capabilities chain

[Alex Williamson]

On Tue, 18 Dec 2018 18:24:00 +0100
Cornelia Huck <address@hidden> wrote:

> On Mon, 17 Dec 2018 16:53:34 -0500
> Eric Farman <address@hidden> wrote:
> 
> > On 11/22/2018 11:54 AM, Cornelia Huck wrote:
> > 
> > ...snip...
> >   
> > > diff --git a/include/uapi/linux/vfio.h b/include/uapi/linux/vfio.h
> > > index 813102810f53..565669f95534 100644
> > > --- a/include/uapi/linux/vfio.h
> > > +++ b/include/uapi/linux/vfio.h
> > > @@ -297,6 +297,7 @@ struct vfio_region_info_cap_type {
> > >   
> > >   #define VFIO_REGION_TYPE_PCI_VENDOR_TYPE        (1 << 31)
> > >   #define VFIO_REGION_TYPE_PCI_VENDOR_MASK        (0xffff)
> > > +#define VFIO_REGION_TYPE_CCW                     (1 << 30)    
> > 
> > Oof.  So the existing VFIO_REGION_TYPE_PCI_VENDOR_TYPE gets OR'd with 
> > another value (e.g., 8086).  But in 4.20, there was a 
> > VFIO_REGION_TYPE_GFX is added as simply "1" ... Which direction are 
> > these definitions being added from?  I guess asked another way, is 
> > _TYPE_CCW going to be OR'd with anything else that necessitates its 
> > presence as an identifier with some Other Thing, or should this follow 
> > the TYPE_GFX enumeration?  Perhaps the type field needs to be tidied up 
> > to help this sit more cleanly now?  (Sorry!)  
> 
> The semantics of that type stuff are really a bit unclear to me :(
> 
> I don't think we'll ever do any fancy mask handling for ccw. It is
> probably enough to have any kind of uniqueness within the different
> types, so maybe counting up would be indeed enough...

Just to confirm, this is the intended usage, simply reserve a new type
following the GFX region example.  We can define VFIO_REGION_TYPE_CCW
as 2 and then there's a whole address space of sub-types to fill in
within that.  I might have over-engineered PCI a bit with the address
space split, but it seemed like a good idea at the time to pre-define a
type address space for each vendor, such that they only need to define
the sub-types and we can avoid namespace collisions.  Unfortunately
this implicit definition for each PCI vendor also contributes to the
confusion here.  Sorry.  Thanks,

Alex