Re: Two opens about iommufd pasid + IOPF

[Baolu Lu <baolu.lu@linux.intel.com>]

On 2024/8/16 20:47, Jason Gunthorpe wrote:
> On Fri, Aug 16, 2024 at 07:50:05PM +0800, Yi Liu wrote:
>> Hi Kevin, Baolu, Jason,
>>
>> I've got two opens when rebasing iommufd pasid series on top of 6.11.
>> Would like to hear about your ideas. 🙂
>>
>> 1) How to retrieve attached domain for pasid in the iommu core. This is
>>     needed when adding domain replacement for pasid. In the before, it can
>>     be done by searching the group->pasid_array[1]. However, the
>>     group->pasid_array stores iommu_attach_handle after the iopf patch [2].
>>     IIUC, iommu_attach_handle is passed only when the domain is
>>     iopf-capable. So, if the old domain does not support iopf, then it's not
>>     able to get this old domain from the group->pasid_array when trying to
>>     do domain replacement. 🙁
>>     Not sure if any side-effect if I always pass an iommu_attach_handle even
>>     if the hwpt/domain is not iopf-capable. @Baolu?
> You should always pass the handle
> 
> Alternatively you could use the xa based encoding I showed earlier to
> store either handles or naked domains in the xarray. But really it
> seems just fine to have iommufd allocate handle memory

That works.

An alternative approach is to separate replace_pasid from set_pasid. In
that case, both new and old domains could be passed from iommufd to
replace callback.

>> 2) Should I enable/disable IOPF when attaching/detaching pasid to/from
>>     domains just like patch [3] does?
>>
>> [1]https://lore.kernel.org/linux-iommu/20240628090557.50898-2-yi.l.liu@intel.com/
>> [2]https://lore.kernel.org/linux-iommu/20240702063444.105814-2-baolu.lu@linux.intel.com/
>> [3]https://lore.kernel.org/linux-iommu/20240702063444.105814-8-baolu.lu@linux.intel.com/
> I would really like to remove IOMMU_DEV_FEAT_IOPF...

That's reasonable. Then iopf is automatically enabled when the first
iopf-capable domain is attached and disabled when the last such domain
is detached.

Thanks,
baolu