Re: [PATCH 1/1] iommu/vt-d: Enable PASID during iommu device probe

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

On 2022/9/16 11:35, Ethan Zhao wrote:
> Baolu,
> 
> 在 2022/9/16 11:05, Baolu Lu 写道:
>> On 2022/9/16 10:40, Ethan Zhao wrote:
>>>>
>>>> I may not get you exactly. 😄 Some IOMMU features reply on PASID
>>>> capabilities on both IOMMU and device. The IOMMU drivers enumerate the
>>>> capabilities and enable them if they are supported.
>>> I might not express it straightforward,  I mean with this patch iommu 
>>> deals with
>>>
>>> the complexity of enabling PASID (globally?)  or not at probing stage 
>>> , instead
>>>
>>> of other device driver side decision to request IOMMU PASID enabling 
>>> during
>>>
>>> their setup state.  if so you move the decision to iommu probe stage. 
>>> hmmm...
>>
>> I am sorry that the commit message was a bit confusing. Actually we
>> always enable PASID at iommu probe path w/ or w/o this patch.
> Really ?  the commit message is quit clear to me ~~@
>>
>>>
>>> Pros,  iommu driver controls everything about PASID enabling.
>>>
>>> Cons,  iommu driver handles all possible complexity about capability 
>>> matching
>>
>> Do device drivers need to configure PCI PASID without IOMMU? I searched
>> the tree and found nothing.
> 
> Device knows if it has PCI PASID cap and its driver also could determine 
> to request
> 
> iommu to enable PASID or not by invoking
> 
> intel_iommu_enable_sva()->*intel_iommu_enable_pasid()*

PASID is a PCIe capability. Though SVA is built on it,it's not only
for SVA. Thus, the purpose of intel_iommu_enable_sva() is not for
enabling PASID.

> 
> *that is the old style without this patch.

No. Without this patch, PASID is also enabled in probe path. Calling
intel_iommu_enable_pasid() in enabling SVA path is actually duplicate.

The commit message for this patch is not correct. It's my bad. :-)

> *
> 
> *Iommu driver of course  also knows if devices in its group related the 
> iommu
> *
> 
> *have PASID cap support or not by enumerating them from the ACPI DMAR.
> *
> 
> This is my understanding, correct me if wrong.
> 
> While configuring device PCI PASID cap is another thing, from kernel or 
> userland,
> 
> compatible with the iommu cap, works, or not work.  Could you prevent anyone
> 
> from doing that ?

I do not object to adding a common interface to enable/disable PASID if
any device driver needs to manage PASID by itself. Before that, there
is no need to add complexity in IOMMU subsystem for a non-existent
requirement.

Best regards,
baolu