Re: [RFC PATCH v2 7/7] arm64: dts: qcom: Add the Inline Crypto Engine nodes

[Krzysztof Kozlowski <krzysztof.kozlowski@linaro.org>]

On 10/03/2023 10:21, Abel Vesa wrote:
>>>>>  			compatible = "qcom,sdm630-sdhci", "qcom,sdhci-msm-v5";
>>>>>  			reg = <0x0c0c4000 0x1000>,
>>>>> -			      <0x0c0c5000 0x1000>,
>>>>> -			      <0x0c0c8000 0x8000>;
>>>>> -			reg-names = "hc", "cqhci", "ice";
>>>>> +			      <0x0c0c5000 0x1000>;
>>>>> +			reg-names = "hc", "cqhci";
>>>>
>>>> I believe this will break the ICE on these platforms without valid
>>>> reason. The commit msg does not explain why you do it or why this is
>>>> necessary.
>>>>
>>>> We already we received comment that we keep breaking Qualcomm platforms
>>>> all the time and need to keep them in some shape.
>>>>
>>>> Also, patchset is non-applicable in current set (breaks users) and
>>>> neither commit nor cover letter mentions it.
>>>>
>>>
>>> FWIW, I tested this patchset on SDA845, and ICE continues to work fine.
>>
>> Really? I clearly see of_find_device_by_node -> "return NULL" and all
>> old code gone, so ABI is broken. Are you sure you applied patch 1-6 and
>> ICE was working?
> 
> of_qcom_ice_get will return the ICE instance if the consumer node has a
> qcom,ice property with a phandle for the ICE devicetree node.

When patches 1-6 are applied, there is no qcom,ice property in DTS. Thus
I don't consider the test as correct... Even if we skip entire ABI
discussion the patchset is non-bisectable thus the test was failing to
detect even that.

> It will
> return NULL otherwise. SDA845 has such ICE node added by this patch,
> therefore, it will work. All platforms that have such node will work
> functionally like before. But I'll take care of the legacy approach as
> well in v3 (see below).

At point of patch 6 none of nodes have it. That's the entire point of
bisectability.

What's more, if you reverse code and makes DTS patches before driver
hoping to fix bisectability - do you see ICE working on existing
platforms? I don't think it so...

> 
>>
>>>
>>> (Though if I understand the patchset correctly, the ICE clock is no longer
>>> turned off when the UFS host controller is suspended.  That isn't ideal as it
>>> wastes power.  I would like that to be fixed.)
>>>
>>> Anyway, when you say "break the ICE", do you really mean "make an incompatible
>>> change to the device-tree bindings"?
>>
>> It breaks existing users of DTS and kernel.
> 
> I assume you mean it breaks if someone is using old approach DTS with a
> kernel that would have ICE driver merged. Yes, that it does. And for
> that, in the v3, I'll make of_qcom_ice_get check if there is a reg entry
> with name "ice" and create an ICE instance but for the same dev as the
> consumer driver. OTOH, if there is no reg entry called "ice", it will
> look up a device based on phande of qcom,ice property. This will allow
> legacy style DTS to work fine, while using the unified driver as a
> library, in that case. For newer platforms, the recommended approach
> will be to add a new ICE node and use qcom,ice property.

For the driver this sounds good. I still think that existing (older) DTS
should not have regs removed, because this affects other users of kernel
DTS.


Best regards,
Krzysztof