Re: [PATCH 01/11] dt-bindings: crypto: qcom,ice: Require power-domain and iface clk

[Krzysztof Kozlowski <krzk@kernel.org>]

On 03/02/2026 10:26, Harshal Dev wrote:
> Hi Krzysztof and Konrad,
> 
> On 1/26/2026 3:59 PM, Konrad Dybcio wrote:
>> On 1/23/26 12:04 PM, Harshal Dev wrote:
>>> Hi Krzysztof,
>>>
>>> On 1/23/2026 2:27 PM, Krzysztof Kozlowski wrote:
>>>> On 23/01/2026 08:11, Harshal Dev wrote:
>>>>> Update the inline-crypto engine DT binding to reflect that power-domain and
>>>>> clock-names are now mandatory. Also update the maximum number of clocks
>>>>> that can be specified to two. These new fields are mandatory because ICE
>>>>> needs to vote on the power domain before it attempts to vote on the core
>>>>> and iface clocks to avoid clock 'stuck' issues.
>>>>>
>>>>> Signed-off-by: Harshal Dev <harshal.dev@oss.qualcomm.com>
>>>>> ---
>>>>>  .../bindings/crypto/qcom,inline-crypto-engine.yaml         | 14 +++++++++++++-
>>>>>  1 file changed, 13 insertions(+), 1 deletion(-)
>>>>>
>>>>> diff --git a/Documentation/devicetree/bindings/crypto/qcom,inline-crypto-engine.yaml b/Documentation/devicetree/bindings/crypto/qcom,inline-crypto-engine.yaml
>>>>> index c3408dcf5d20..1c2416117d4c 100644
>>>>> --- a/Documentation/devicetree/bindings/crypto/qcom,inline-crypto-engine.yaml
>>>>> +++ b/Documentation/devicetree/bindings/crypto/qcom,inline-crypto-engine.yaml
>>>>> @@ -28,12 +28,20 @@ properties:
>>>>>      maxItems: 1
>>>>>  
>>>>>    clocks:
>>>>> +    maxItems: 2
>>>>
>>>> This is ABI break and your commit msg suggests things were not perfect,
>>>> but it is not explicit - was this working or not? How is it that ICE was
>>>> never tested?
>>>>
>>>
>>> I took some time to educate myself on the point of DT bindings stability being a
>>> strict requirement now, so I understand how these changes are breaking ABI, I'll
>>> send a better version of this again.
>>>
>>> As for your question of how it was working till now, it seems that
>>> things were tested with the 'clk_ignore_unused' flag, or with CONFIG_SCSI_UFS_QCOM
>>> flag being override set to 'y'. When this is done, QCOM-ICE (on which QCOM-UFS
>>> depends) initiates probe _before_ the unused clocks and power-domains are
>>> disabled by the kernel. And so, the un-clocked register access or clock 'stuck'
>>> isn't observed (since the clocks and power domains are already enabled).
>>> Perhaps I should write this scenario explicitly in the commit message?
>>>
>>> To maintain backward compatibility, let me introduce minItems and maxItems for clocks.
>>> When the Linux distro uses CONFIG_SCSI_UFS_QCOM=y, we can do with just 1 clock as
>>> before.
>>
>> You must not assume any particular kernel configuration
>>
>> clk_ignore_unused is a hack which leads to situations like this, since
>> the bootloader doesn't clean up clocks it turned on, which leads to
>> situations like this where someone who previously wrote this binding
>> didn't care enough to **actually** test whether this device can operate
>> with only the set of clocks it requires
>>
>> I believe in this case it absolutely makes sense to break things, but
>> you must put the backstory in writing, in the commit message
>>
> 
> I took some more time to think this through, and I agree with you now Konrad.
> 
> These DT bindings appear to be invalid from day-1. ICE being an independent
> and common IP for both UFS and SDCC, it cannot operate correctly without its
> power-domain and clocks being enabled first. Hence, it should be mandatory for
> them to be specified in the DT-node and the same should be reflected in the DT
> binding.
> 
> The only reason I can think of for omitting the 'power-domain' and 'iface' clock
> in the original DT-binding for ICE is because we failed to test the driver on
> a production kernel where the 'clk_ignore_unused' flag is not passed on the cmdline.

That's a reason to change ABI in the bindings, but not a reason to break
in-kernel or out of tree DTS.

> Or if we did test that way, we were just lucky to not run into a timing scenario
> where the probe for the driver is attempted _after_ the clocks are turned off by the
> kernel.
> 
> Sending a new patch, which makes these two resources optional (to preserve the DT
> binding) would either imply that we are make this bug fix optional as well or
> asking the reporter to resort to some workaround such as overriding
> CONFIG_SCSI_UFS_QCOM to 'y'.

Either I do not understand the point or you still insist on breaking a
working DTS on kernels with clk_ignore_unused, just because what
exactly? You claim it did not work, but in fact it did work. So you
claim it worked by luck, right? And what this patchset achieves? It
breaks this "work by luck" into "100% not working and broken". I do not
see how is this an improvement.

My NAK for driver change stays. This is wrong approach - you cannot
break working DTS.

Best regards,
Krzysztof