Re: [PATCH v5 04/27] iommu/io-pgtable-arm: Factor kernel specific code out

[Mostafa Saleh <smostafa@google.com>]

On Fri, Nov 28, 2025 at 12:45:41PM -0400, Jason Gunthorpe wrote:
> On Mon, Nov 17, 2025 at 06:47:51PM +0000, Mostafa Saleh wrote:
> > Some of the used APIs are only part of the kernel and are not
> > available in the hypervisor, factor those out:
> > - alloc/free memory
> 
> Why not provide the iommu-pages API for the hypervisor environment?
> 
> Same for virt_to_phys, that could be moved into an iommu-pages wrapper
> too..

I guess that's possible, but then we would have to stub dma_map/unmap,
which might a bit ugly. I will look more into it.

> 
> I want to change other parts of the driver to use iommu-pages in the
> long run as well, so putting the abstraction there is probably more
> valuable than this.
> 
> Also the genericpt stuff is merged, should you (eventually?) be making
> a pKVM hypervisor specific set of page table functions? Eg if all you
> are doing is mirroring the host stage 2 I think you can build
> something much more efficient...

Most of the code is re-used from the io-pgtable-arm, so when that is
converted to genericpt, it should get to the hypervisor also.
The only hypervisor specific logic is the memory allocation, virt/phys
conversion and CMOs, I can look more into painting those, we don’t have
to change much.

I have some plans to add support for sharing the CPU stage-2 page table,
which is going to be complicated as it requires changes to the core
hypervisor hyp/pgtable.c code, which will be in a subsequent series.
However, I reckon that we would keep the shadowing logic as not all
SMMUv3/platforms are capable of sharing page tables.

Thanks,
Mostafa

> 
> Jason
>