Re: [PATCH net-next v2 1/3] dt-bindings: sram: qcom,imem: Allow modem-tables

[Konrad Dybcio <konrad.dybcio@oss.qualcomm.com>]

On 7/15/25 8:37 AM, Krzysztof Kozlowski wrote:
> On 14/07/2025 19:53, Konrad Dybcio wrote:
>> On 5/27/25 1:42 PM, Krzysztof Kozlowski wrote:
>>> On 27/05/2025 13:36, Konrad Dybcio wrote:
>>>>>> diff --git a/Documentation/devicetree/bindings/sram/qcom,imem.yaml b/Documentation/devicetree/bindings/sram/qcom,imem.yaml
>>>>>> index 2711f90d9664b70fcd1e2f7e2dfd3386ed5c1952..7c882819222dc04190db357ac6f9a3a35137cc9e 100644
>>>>>> --- a/Documentation/devicetree/bindings/sram/qcom,imem.yaml
>>>>>> +++ b/Documentation/devicetree/bindings/sram/qcom,imem.yaml
>>>>>> @@ -51,6 +51,9 @@ properties:
>>>>>>      $ref: /schemas/power/reset/syscon-reboot-mode.yaml#
>>>>>>  
>>>>>>  patternProperties:
>>>>>> +  "^modem-tables@[0-9a-f]+$":
>>>>>> +    description: Region reserved for the IP Accelerator
>>>>>
>>>>> Missing additionalProperties: false, which would point you that this is
>>>>> incomplete (or useless because empty).
>>>>
>>>> How do I describe a 'stupid' node that is just a reg?
>>> With "reg" - similarly to many syscon bindings.
>>
>> Is this sort of inline style acceptable, or should I introduce
>> a separate file?
> 
> It's fine, assuming that it is desired in general. We do not describe
> individual memory regions of syscon nodes and this is a syscon.
> 
> If this is NVMEM (which it looks like), then could use NVMEM bindings to
> describe its cells - individual regions. But otherwise we just don't.

It's volatile on-chip memory

> There are many exceptions in other platforms, mostly old or even
> unreviewed by DT maintainers, so they are not a recommended example.
> 
> This would need serious justification WHY you need to describe the
> child. Why phandle to the main node is not enough for consumers.

It's simply a region of the SRAM, which needs to be IOMMU-mapped in a
specific manner (should IMEM move away from syscon+simple-mfd to
mmio-sram?). Describing slices is the DT way to pass them (like under
NVMEM providers).

> 
> If the reason is - to instantiate child device driver - then as well no.
> This has been NAKed on the lists many times - you need resources if the
> child should be a separate node. Address space is one resource but not
> enough, because it can easily be obtained from the parent/main node.

There is no additional driver for this

Konrad