Re: [RFC PATCH v2 40/58] KVM: arm64: smmu-v3: Add map/unmap pages and iova_to_phys

[Mostafa Saleh <smostafa@google.com>]

Hi Jason,

On Thu, Dec 12, 2024 at 03:44:35PM -0400, Jason Gunthorpe wrote:
> On Thu, Dec 12, 2024 at 06:04:04PM +0000, Mostafa Saleh wrote:
> > Add map_pages and iova_to_phys HVC code, which
> > mainly calls the io-pgtable.
> > 
> > For unmap_pages, we rely on IO_PGTABLE_QUIRK_UNMAP_INVAL, where the
> > driver first calls unmap_pages which invalidate all the pages as
> > a typical unmap, issuing all the necessary TLB invalidations.
> > Then, we will start a page table with 2 callbacks:
> > - visit_leaf: for each unmapped leaf, it would decrement the refcount
> >   of the page using __pkvm_host_unuse_dma(), reversing the what IOMMU
> >   core does in map.
> > - visit_post_table: this would free any invalidated tables as they
> >   wouldn't be freed because of the quirk.
> 
> I don't know if the timelines will work out, but the pagetable stuff
> I'm working on will let you write a much more appropriate
> implementation for pkvm's usage than trying to hack it into the
> iopgtable code like this.

I didn’t check your new page table patches yet, but I would say it’s
more likely your patches would land first, because as mentioned in the
cover letter, there are still many dependencies for pKVM before IOMMU
support lands, so I don’t mind converging if possible.

Thanks,
Mostafa

> 
> Even the iommu focused routines I have got now would solve this
> problem because they allways spit out a linked list of all the memory
> to free after map/unmap and never internally free it..

> 
> Jason