Re: [PATCH v3 0/5] Introduce "non-pixel" sub node within iris video node

[Charan Teja Kalla]

On 7/4/2025 10:15 PM, Dmitry Baryshkov wrote:
> On Fri, Jul 04, 2025 at 09:23:06AM +0100, Bryan O'Donoghue wrote:
>> On 03/07/2025 22:23, Dmitry Baryshkov wrote:
>>>> I still give my RB for the series.
>>>>
>>>> To me the only question is should it be applied to sm8550 or to new SoCs
>>>> from now on, sa8775p, x1e and derivatives.
>>> I think we need to apply it to_all_ SoCs, maybe starting from MSM8x26
>>> and MSM8x16. Likewise, we need to think bout secure buffers usecase. And
>>> once we do that, we will realize that it also changes the ABI for all
>>> SoCs that support either Venus or Iris.
>> I think a dts change is a non-starter as its an ABI break.
>>
>> So if ABI break is out and reworking future dts to allow for a new device is
>> out, then some API change is needed to allow the driver to stop the kernel
>> automatically setting up the IOMMUs, create the new device as a platform
>> device not dependent on DT change and have the probe() of the relevant
>> drivers setup their own IOMMU extents based on - probably indexes we have in
>> the driver configuration parameters.
> 
> What about instead:
> 
> - keep IOMMU entries as is
> - Add iommu-maps, mapping the non-pixel SID

Otherways to avoid the ABI breakage:
a) Keep iommus max items as 2, which is unchanged.
b) Repeat the same sid for both entries, i.e.,
         iommus = <&apps_smmu 0x1940 0x0000>,
                - <&apps_smmu 0x1947 0x0000>;
	        + <&apps_smmu 0x1940 0x0000>;

and then create the new device as a platform device independent of dt.
RFC for this is posted[1].

However,  Dmitry(in offline forums) termed creating 2 same sid entries
as -- "ridiculous band-aid just to fulfill the "ABI" requirements and
really doesn't make sense". Looks Fair.

OTOH, To exactly implement the idea mentioned here, I have the below
questions, can you please help me with:
1) By keeping the entries in 'iommu=' as is, then add non-pixel sid in
iommu-maps actually creates the overlapping entries.

So, IIUC -- this requires changes to the iommu driver to ignore the
setting up the non-pixel sid(or whatever is mentioned in iommu-maps) and
then use newly created platform device to program the entries mentioned
in iommu-maps. Please CMIW.

If the above understanding correct -- Doesn't it look like we are trying
to fix in the iommu driver for the problem statement related to dt bindings?

2) However, I still think that problem statement of __non-eligibility
for video IP to create the sub devices in the dt__  is still valid and
can be taken separately and actually [1] is targetting this. I believe
this is atleast required for secure contexts. please CMIW.

[1]
https://lore.kernel.org/all/20250928171718.436440-1-charan.kalla@xxxxxxxxxxxxxxxx/

Thanks,
Charan> - In future expand iommu-maps, describing the secure contexts?