On 10/07/2022 16:29, Jason Gunthorpe wrote:

On Sat, Jul 09, 2022 at 12:58:00PM +1000, Alexey Kardashevskiy wrote:
  

quoted

driver->ops->attach_group on POWER attaches a group so VFIO claims ownership
over a group, not devices. Underlying API (pnv_ioda2_take_ownership()) does
not need to keep track of the state, it is one group, one ownership
transfer, easy.

It should not change, I think you can just map the attach_dev to the group?

There are multiple devices in a group, cannot just map 1:1.

quoted

What is exactly the reason why iommu_group_claim_dma_owner() cannot stay
inside Type1 (sorry if it was explained, I could have missed)?

It has nothing to do with type1 - the ownership system is designed to
exclude other in-kernel drivers from using the group at the same time
vfio is using the group. power still needs this protection regardless
of if is using the formal iommu api or not.

POWER deals with it in vfio_iommu_driver_ops::attach_group.

quoted

Also, from another mail, you said iommu_alloc_default_domain() should fail
on power but at least IOMMU_DOMAIN_BLOCKED must be supported, or the whole
iommu_group_claim_dma_owner() thing falls apart.

Yes

quoted

And iommu_ops::domain_alloc() is not told if it is asked to create a default
domain, it only takes a type.

"default domain" refers to the default type pased to domain_alloc(),
it will never be blocking, so it will always fail on power.
"default domain" is better understood as the domain used by the DMA
API

The DMA API on POWER does not use iommu_ops, it is dma_iommu_ops from 
arch/powerpc/kernel/dma-iommu.c from before 2005. so the default domain 
is type == 0 where 0 == BLOCKED.


-- 
Alexey