[PATCH v2 14/18] iommu/arm-smmu-v3: Support domains with shared CDs

[Robin Murphy]

On 2023-06-14 10:17, Michael Shavit wrote:
> On Wed, Jun 7, 2023 at 1:09 AM Jason Gunthorpe <jgg at nvidia.com> wrote:
>>
>> On Tue, Jun 06, 2023 at 08:07:50PM +0800, Michael Shavit wrote:
>>> SVA may attach a CD to masters that have different upstream SMMU
>>> devices. The arm_smmu_domain structure can only be attached to a single
>>> upstream SMMU device however.
>>
>> Isn't that pretty much because we don't support replicating
>> invalidations to each of the different SMMU instances?
> 
> Looked into this some more, and supporting attach to multiple devices
> is still very hard:
> 1. When an arm_smmu_domain is first attached to a master, it
> initializes an io_pgtable_cfg object whose properties depend on the
> master's upstream SMMU device. This io_pgtable_cfg is then used to
> allocate an io_pgtable object, and the arm_smmu_ctx_desc's TTBR field
> points to that io_pgtable's TTBR (and ditto for the vttbr on stage 2
> domains). So then arm_smmu_domain needs to be split into two,
> arm_smmu_domain and arm_smmu_device_domain with the latter containing
> a per-SMMU device io_pgtable, arm_smmu_ctx_desc and arm_smmu_s2_cfg.
> Each iommu_domain_ops operation now needs to loop over each
> arm_smmu_device_domain.
> 2. Some of the iommu_domain fields also depend on the per-SMMU
> io_pgtable_cfg; specifically pgsize_bitmap and geometry.aperture_end.
> These need to be restricted as the domain is attached to more devices.
> 3. Attaching a domain to a new SMMU device must be prohibited after
> any call to map_pages or if iommu_domain.pgsize_bitmap and
> iommu-domain.geometry.aperture_end have been consumed by any system.
> The first is something the arm-smmu-v3.c driver could feasibly enforce
> on its own, the second requires changes to the iommu framework.

In practice it would be entirely reasonable to only support 
cross-instance attach between instances with matching capabilities such 
that they *can* share the pagetable directly.

> The arm-smmu-v3-sva.c implementation avoids all these problems because it
> doesn't need to allocate an io_pgtable;

SVA has all the same underlying problems - pagetable sharing is still 
pagetable sharing regardless of which code happens to allocate the 
physical pages - they're just avoided up-front by disallowing SVA at all 
if the relevant capabilities don't line up between SMMU and CPU.

Thanks,
Robin.