Re: [PATCH 2/5] iommu: Call helper function to get assigned pasid value

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

On 2023/8/9 22:48, Jason Gunthorpe wrote:
> On Wed, Aug 09, 2023 at 06:58:15PM +0800, Baolu Lu wrote:
>> On 2023/8/9 17:49, Tian, Kevin wrote:
>>>> From: Baolu Lu <baolu.lu@linux.intel.com>
>>>> Sent: Wednesday, August 9, 2023 8:22 AM
>>>>
>>>> On 2023/8/8 15:49, Tina Zhang wrote:
>>>>> Use the helper function mm_get_pasid() to get the mm assigned pasid
>>>>> value.
>>>>
>>>> For internal iommu drivers, perhaps we should use another helper.
>>>> Something like sva_domain_get_pasid()?
>>>>
>>>> Suppose that the iommu drivers should have no idea about the "mm".
>>>>
>>>
>>> Aren't all touched functions accept a struct mm_struct pointer?
>>
>> In the end we should remove all mm reference in the individual drivers.
>> The drivers only need to know what they need to know. All mm-aware code
>> should eventually be moved to the core.
>>
>> For now, at least we should avoid using mm in the set/remove_dev_pasid
>> code path. Later, once we complete consolidating mm notification in the
>> core, drivers will have no need to use "mm" anymore.
> 
> I'm not sure how this will play out...
> 
> We don't want extra function indirections here so ultimately the
> driver needs to hook the arch_invalidate_range() in the mm_notifier
> with its own function.

Agreed. Given the mm notification callback is a performance path, an
extra indirection call here is not good.

> 
> The core could put the mm_notifier in a common iommu_domain_sva struct
> and it could stick in the driver's invalidate ops, that would be a
> nice simplification (and discourage drivers from doing the crazy
> things they are currently doing)

Yes. So the iommu driver can retrieve the sva domain from struct
mmu_notifier. The callback implementation will still be domain centric.
Hence, there will be no need to use mm.

Best regards,
baolu