Re: [PATCH 0/5] Convert virtio-iommu to domain_alloc_paging()

[Robin Murphy <robin.murphy@arm.com>]

On 2025-02-12 6:50 pm, Jacob Pan wrote:
> Hi Jason,
> 
> On Wed, 12 Feb 2025 08:50:07 -0400
> Jason Gunthorpe <jgg@nvidia.com> wrote:
> 
>> On Tue, Feb 11, 2025 at 04:41:51PM -0800, Jacob Pan wrote:
>>
>>> I see the same on arm64 with v9.0, assigned an ixgbe nic via VFIO.
>>
>> Huh, I figured it worked on ARM..
>>
>> There is no requirement that an iommu driver implement identity, this
>> is all an optimization. If it doesn't work in alot of cases can we
>> just remove it? It would simplfy a lot..
>>
> Currently, identity domain type can be set via sysfs. To prevent that,
> maybe we need this?
> 
> +static int viommu_def_domain_type(struct device *dev)
> +{
> +  	return IOMMU_DOMAIN_DMA;

Ugh, please no more of that.


Yes, a user can *request* to set an identity domain via sysfs, and if 
the driver does not support identity domains then:

iommu_group_store_type()
     iommu_setup_default_domain()
         iommu_group_alloc_default_domain()
             __iommu_group_alloc_default_domain()
                 __iommu_alloc_identity_domain()

will return failure, and everything will remain fine (well, except on 
apple-dart for now...) Find an s390 system and try it yourself* ;)

Thanks,
Robin.



*quickly, before the patches land and that gains identity support too!

> +}
> +
>   static struct iommu_ops viommu_ops = {
>          .identity_domain        = &viommu_identity_domain.domain,
>          .capable                = viommu_capable,
> @@ -1096,6 +1106,7 @@ static struct iommu_ops viommu_ops = {
>          .device_group           = viommu_device_group,
>          .get_resv_regions       = viommu_get_resv_regions,
>          .of_xlate               = viommu_of_xlate,
> +       .def_domain_type = viommu_def_domain_type,
>          .owner                  = THIS_MODULE,
>