[PATCH 2/3] iommu/arm-smmu: Add initial driver support for ARM SMMUv3 devices

Will Deacon will.deacon at arm.com
Tue Jun 2 02:47:46 PDT 2015

On Tue, Jun 02, 2015 at 08:39:56AM +0100, Joerg Roedel wrote:
> On Mon, Jun 01, 2015 at 10:40:14AM +0100, Will Deacon wrote:
> > I like this proposal. The only remaining part is that the pgsize bitmap
> > inherited by the domain can only truly be finalised once the page table
> > is allocated, which in turn can only happen once we've identified the
> > IOMMU instance.
> We know the IOMMU instance for domains allocted with
> iommu_domain_alloc_for_group because every group is behind a single
> IOMMU instance. So the page-table can be allocated at domain creation
> time and mappings can be created before the group itself is attached.

Indeed -- I think this form of allocation makes a lot of sense.

> > In the case of iommu_domain_alloc_for_group, that's nice and
> > straightforward (i.e. as part of the call) but for the default
> > iommu_domain_alloc() case, we'd have to continue postponing things to
> > ->attach time before we could provide a reliable pgsize_bitmap. This is
> > to handle the case where an SMMU may be able to support only a subset of
> > the pgsize_bitmap on a per-domain basis.
> I don't think we need to postpone anything. Domains returned by
> iommu_domain_alloc() need to work for all groups. If there are multiple
> IOMMUs in the system with different capabilities/page-table formats the
> IOMMU core needs to build multiple page-tables for that domain.

I really don't think that's feasible. We support an awful lot of
combinations that affect the page table format on ARM: there are 5
different formats, each supporting different translation regimes (sets
of page size) and each of those needs configuring for input/output
address sizes to compute the number of levels. You could easily end up
with over 20 page tables and their corresponding sets of control
register values.

Given that we only install one page table in the end, I'm struggling to
see the issue with postponing its allocation until we've figured out
the IOMMU instance thanks to a device attach.

> VFIO already does that as a workaround of how things are done currently,
> but I really think this should be done in the IOMMU core code.

Postponing page table creation to ->attach works fine with VFIO today.
AFAICT, it never tries to map pages for an empty domain and therefore
the IOMMU instance is always known by the IOMMU driver at ->map time.

> > In which situation do you think the merged pgsize_bitmap would get used?
> > The only place I can think of is if we're trying to call iommu_map on a
> > domain with no devices attached. However, if that domain was created
> > using iommu_domain_alloc() then the pgsize_bitmap is the least of our
> > worries -- we don't even know the page table format!
> The first usecase for the merged pgsize_bitmap that comes to mind is to
> determine the minimum page-size that can be used for a domain. In the
> SMMUv3 case when there are IOMMUs with different minium page-sizes in one
> system, this would be the biggest minimum page-size of all IOMMUs.

Understood, but I'm trying to understand why you'd want to know that for
an empty (no devices attached) domain. I don't think we currently have
any use-cases for that and it's really not a practical thing to support.


More information about the linux-arm-kernel mailing list