[Intel-gfx] [PATCH v2 5/9] vfio/mdev: Consolidate all the device_api sysfs into the core code

[Alex Williamson]

On Fri, 10 Sep 2021 10:38:50 -0300
Jason Gunthorpe <jgg at nvidia.com> wrote:

> On Fri, Sep 10, 2021 at 01:10:46PM +0100, Christoph Hellwig wrote:
> > On Thu, Sep 09, 2021 at 04:38:45PM -0300, Jason Gunthorpe wrote:  
> > > Every driver just emits a static string, simply feed it through the ops
> > > and provide a standard sysfs show function.  
> > 
> > Looks sensible.  But can you make the attribute optional and add a
> > comment marking it deprecated?  Because it really is completely useless.
> > We don't version userspace APIs, userspae has to discover new features
> > individually by e.g. finding new sysfs files or just trying new ioctls.  
> 
> To be honest I have no idea what side effects that would have..
> 
> device code search tells me libvirt reads it and stuffs it into some
> XML
> 
> Something called mdevctl touches it, feeds it into some JSON and
> other stuff..
> 
> qemu has some VFIO_DEVICE_API_* constants but it is all dead code
> 
> I agree it shouldn't have been there in the first place
> 
> Cornelia? Alex? Any thoughts?

It's not a version, it's a means for userspace to determine the basic
API for an mdev device without needing to go through the process of
creating a container, adding the group, setting an IOMMU type, opening
the device before being able to call VFIO_DEVICE_GET_INFO to determine
the API.  For example, it wouldn't make sense for libvirt to attach a
vfio-ccw device to a PCIe root port in a VM.  It's a means to say this
mdev device is a vfio-pci or that mdev device is a vfio-ccw.  If it were
optional, then management tools would have no basic idea how to attach
the device to a VM without gaining access to the device themselves.
Thanks,

Alex