Re: [PATCH v2 0/4] vfio-pci support pasid attach/detach

[Yi Liu <yi.l.liu@intel.com>]

On 2024/8/16 16:29, Vasant Hegde wrote:
> Yi,
> 
> 
> On 8/15/2024 7:42 AM, Yi Liu wrote:
>> On 2024/8/14 22:40, Jason Gunthorpe wrote:
>>> On Wed, Aug 14, 2024 at 04:19:13PM +0800, Yi Liu wrote:
>>>
>>>> /**
>>>>    * enum iommufd_hw_capabilities
>>>>    * @IOMMU_HW_CAP_DIRTY_TRACKING: IOMMU hardware support for dirty tracking
>>>>    *                               If available, it means the following APIs
>>>>    *                               are supported:
>>>>    *
>>>>    *                                   IOMMU_HWPT_GET_DIRTY_BITMAP
>>>>    *                                   IOMMU_HWPT_SET_DIRTY_TRACKING
>>>>    *
>>>>    */
>>>> enum iommufd_hw_capabilities {
>>>>      IOMMU_HW_CAP_DIRTY_TRACKING = 1 << 0,
>>>> };
>>>
>>> I think it would be appropriate to add the flag here
>>
>> ok.
>>
>>> Is it OK to rely on the PCI config space PASID enable? I see all the
>>> drivers right now are turning on PASID support during probe if the
>>> iommu supports it.
>>
>> Intel side is not ready yet as it enables pasid when the device is attached
>> to a non-blocking domain. I've chatted with Baolu, and he will kindly to
>> enable the pasid cap in the probe_device() op if both iommu and device has
>> this cap. After that, Intel side should be fine to rely on the PASID enable
>> bit in the PCI config space.
>>
>> How about SMMU and AMD iommu side? @Jason, @Suravee, @Vasant?
> 
> AMD driver currently discovers capability in probe_device() and enabled it in
> attach_dev() path.

I see. So AMD side also has a gap. Is it easy to make it suit Jason's
suggestion in the above?

-- 
Regards,
Yi Liu