Re: [PATCH v8 7/9] vfio/mdev: Add iommu related member in mdev_device

[Jason Gunthorpe]

On Tue, May 11, 2021 at 02:56:05PM +0800, Lu Baolu wrote:

> >     After my next series the mdev drivers will have direct access to
> >     the vfio_device. So an alternative to using the struct device, or
> >     adding 'if mdev' is to add an API to the vfio_device world to
> >     inject what iommu configuration is needed from that direction
> >     instead of trying to discover it from a struct device.
> 
> Just want to make sure that I understand you correctly.
> 
> We should use the existing IOMMU in-kernel APIs to connect mdev with the
> iommu subsystem, so that the upper lays don't need to use something
> like (if dev_is_mdev) to handle mdev differently. Do I get you
> correctly?

After going through all the /dev/ioasid stuff I'm pretty convinced
that none of the PASID use cases for mdev should need any iommu
connection from the mdev_device - this is an artifact of trying to
cram the vfio container and group model into the mdev world and is not
good design.

The PASID interfaces for /dev/ioasid should use the 'struct
pci_device' for everything and never pass in a mdev_device to the
iommu layer.

/dev/ioasid should be designed to support this operation and is why I
strongly want to see the actual vfio_device implementation handle the
connection to the iommu layer and not keep trying to hack through
building what is actually a vfio_device specific connection through
the type1 container code.

> > 3) The vfio_bus_is_mdev() and related symbol_get() nonsense in
> >     drivers/vfio/vfio_iommu_type1.c has to go, for the same reasons
> >     it was not acceptable to do this for the interrupt side either.
> 
> Yes. Agreed. I will look into it.

This will be harder, but the same logic applies - it serves to allow
controlling an ioasid without involving the vfio_device.

Jason