Re: [RFC PATCH kernel] vfio: Skip checking for IOMMU_CAP_CACHE_COHERENCY on POWER and more

[Jason Gunthorpe <jgg@ziepe.ca>]

On Fri, Jul 01, 2022 at 07:10:45AM +0000, Tian, Kevin wrote:
> > From: Alexey Kardashevskiy <aik@ozlabs.ru>
> > Sent: Friday, July 1, 2022 2:18 PM
> > 
> > VFIO on POWER does not implement iommu_ops and therefore
> > iommu_capable()
> > always returns false and __iommu_group_alloc_blocking_domain() always
> > fails.
> > 
> > iommu_group_claim_dma_owner() in setting container fails for the same
> > reason - it cannot allocate a domain.
> > 
> > This skips the check for platforms supporting VFIO without implementing
> > iommu_ops which to my best knowledge is POWER only.
> > 
> > This also allows setting container in absence of iommu_ops.
> > 
> > Fixes: 70693f470848 ("vfio: Set DMA ownership for VFIO devices")
> > Fixes: e8ae0e140c05 ("vfio: Require that devices support DMA cache
> > coherence")
> > Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
> > ---
> > 
> > Not quite sure what the proper small fix is and implementing iommu_ops
> > on POWER is not going to happen any time soon or ever :-/
> 
> I'm not sure how others feel about checking bus->iommu_ops outside
> of iommu subsystem. This sounds a bit non-modular to me and it's not
> obvious from the caller side why lacking of iommu_ops implies the two
> relevant APIs are not usable.

The more I think about this, the more I think POWER should implement
partial iommu_ops to make this work. It would not support an UNMANAGED
domain, or default domains, but it would support blocking and the
coherency probe.

This makes everything work properly and keeps the mess out of the core
code.

It should not be hard to do if someone can share a bit about the ppc
code and test it..

Jason