Re: [PATCH v6 06/25] iommu/io-pgtable-arm: Rework to use the iommu-pages API

[Jason Gunthorpe <jgg@ziepe.ca>]

On Fri, May 01, 2026 at 11:19:08AM +0000, Mostafa Saleh wrote:
> To prepare for supporting io-pgtable-arm in the pKVM hypervisor,
> we need to abstract away standard kernel allocations, frees, virt/phys
> conversions, and DMA API mapping.
> 
> This patch introduces a set of generic wrappers in iommu-pages.h:
> - iommu_alloc_data
> - iommu_free_data
> - iommu_virt_to_phys
> - iommu_phys_to_virt
> - iommu_pages_dma_map
> - iommu_pages_dma_mapping_error
> - iommu_pages_dma_unmap

Wah? This has nothing to do with iommu pages? This just leaking
everything iommu pages abstracted out :(

When I said to use iommu-pages, I meant to use the existing API, not a
completely different one.

From an iommu-pages perspective the issue is this code open codes
dma_map_single()/etc instead of using the API surface
iommu_pages_start_incoherent()

This is annoying to fix beacuse the external allocator messes it up,
but I think with some #ifdef you can probably fix it up.

So.. I suggest you update it to use the iommu_pages API, #ifdef out
the allocator so the pkvm pkvm doesn't need to deal with it. Then
compile a special iommu-pages for the pkvm side presenting the same
API.

You should have a pkvm shim header that provides
kmalloc/kfree/virt_to_phys in the normal way and just #include that in
io-pgtable when doing a pkvm build instead of hacking up all the code.

Jason