Date:   Mon, 7 Nov 2022 10:07:46 +0100
From:   Krzysztof Kozlowski <krzysztof.kozlowski@...aro.org>
To:     Bjorn Andersson <andersson@...nel.org>,
        Bhupesh Sharma <bhupesh.sharma@...aro.org>
Cc:     linux-crypto@...r.kernel.org, devicetree@...r.kernel.org,
        agross@...nel.org, herbert@...dor.apana.org.au,
        linux-kernel@...r.kernel.org, robh+dt@...nel.org,
        linux-arm-msm@...r.kernel.org, thara.gopinath@...il.com,
        robh@...nel.org, bhupesh.linux@...il.com, davem@...emloft.net,
        Jordan Crouse <jorcrous@...zon.com>
Subject: Re: [PATCH v7 6/9] crypto: qce: core: Add new compatibles for qce
 crypto driver

On 07/11/2022 04:43, Bjorn Andersson wrote:
> On Wed, Sep 21, 2022 at 12:22:03PM +0530, Bhupesh Sharma wrote:
>> On 9/21/22 11:57 AM, Krzysztof Kozlowski wrote:
>>> On 21/09/2022 08:16, Bhupesh Sharma wrote:
>>>>
>>>>
>>>> On 9/20/22 8:42 PM, Krzysztof Kozlowski wrote:
>>>>> On 20/09/2022 13:40, Bhupesh Sharma wrote:
>>>>>> Since we decided to use soc specific compatibles for describing
>>>>>> the qce crypto IP nodes in the device-trees, adapt the driver
>>>>>> now to handle the same.
>>>>>>
>>>>>> Keep the old deprecated compatible strings still in the driver,
>>>>>> to ensure backward compatibility.
>>>>>>
>>>>>> Cc: Bjorn Andersson <andersson@...nel.org>
>>>>>> Cc: Rob Herring <robh@...nel.org>
>>>>>> Cc: herbert@...dor.apana.org.au
>>>>>> Tested-by: Jordan Crouse <jorcrous@...zon.com>
>>>>>> Signed-off-by: Bhupesh Sharma <bhupesh.sharma@...aro.org>
>>>>>> ---
>>>>>>    drivers/crypto/qce/core.c | 9 +++++++++
>>>>>>    1 file changed, 9 insertions(+)
>>>>>>
>>>>>> diff --git a/drivers/crypto/qce/core.c b/drivers/crypto/qce/core.c
>>>>>> index 63be06df5519..99ed540611ab 100644
>>>>>> --- a/drivers/crypto/qce/core.c
>>>>>> +++ b/drivers/crypto/qce/core.c
>>>>>> @@ -291,8 +291,17 @@ static int qce_crypto_remove(struct platform_device *pdev)
>>>>>>    }
>>>>>>    static const struct of_device_id qce_crypto_of_match[] = {
>>>>>> +	/* Following two entries are deprecated (kept only for backward compatibility) */
>>>>>>    	{ .compatible = "qcom,crypto-v5.1", },
>>>>>>    	{ .compatible = "qcom,crypto-v5.4", },
>>>>>
>>>>> This is okay, so there is no ABI break.
>>>>
>>>> Great. Thanks for the confirmation.
>>>>
>>>>>> +	/* Add compatible strings as per updated dt-bindings, here: */
>>>>>> +	{ .compatible = "qcom,ipq4019-qce", },
>>>>>> +	{ .compatible = "qcom,ipq6018-qce", },
>>>>>> +	{ .compatible = "qcom,ipq8074-qce", },
>>>>>> +	{ .compatible = "qcom,msm8996-qce", },
>>>>>> +	{ .compatible = "qcom,sdm845-qce", },
>>>>>> +	{ .compatible = "qcom,sm8150-qce", },
>>>>>> +	{ .compatible = "qcom,sm8250-qce", },
>>>>>
>>>>> This is a bit odd... you have 7 devices which are simply compatible or
>>>>> even the same. This should be instead one compatible.
>>>>>
>>>>> I don't really get why do you want to deprecate "qcom,crypto-v5.1".
>>>>> Commit msg only says "we decided" but I do not know who is "we" and "why
>>>>> we decided like this". If you want to deprecate it, perfectly fine by
>>>>> me, but please say in commit msg why you are doing it.
>>>>
>>>> I understand. This patchset has been in flight for some time and hence I
>>>> might have missed sharing some detailed information about the review
>>>> comments and rework done along the way (in the cover letter for this
>>>> series).
>>>>
>>>> Coming back to your concern, here is the relevant background:
>>>> - Please see:
>>>> https://lore.kernel.org/linux-arm-msm/20210316222825.GA3792517@robh.at.kernel.org/
>>>>
>>>> - Rob shared some comments on the v1 series regarding the soc-specific
>>>> compatibles. He mentioned in the above thread that 'you should stick
>>>> with SoC specific compatibles as *everyone* else does (including most
>>>> QCom bindings).'
>>>>
>>>> - So, while I had proposed "qcom,crypto-v5.1" (for ipq6018) and
>>>> "qcom,crypto-v5.4" (for sdm845, sm8150) etc. as the compatible(s) in the
>>>> v1 series, I shifted to using the soc-specific compatibles from the v2
>>>> series, onwards.
>>>
>>> Then the reason could be - Reviewers preferred SoC-based compatible
>>> instead of IP-block-version-based.
>>>
>>> What is confusing is the difference between that link and here. That
>>> link wanted to introduce 4 different compatibles... and here you have
>>> even 7 compatibles being the same.
>>
>> The link points to v1 version and we are on v7 currently. So there have been
>> other comments and reworks along the way :)
>>
>> All of these have been referred to in the cover letter logs.
>>
>> Again please refer to Vladimir's comments on v5 version here, where he
>> suggested adding soc compatibles for 'ipq8074' and 'msm8996' as well.
>>
>> -
>> https://lore.kernel.org/lkml/7328ae17-1dc7-eaa1-5993-411b986e5e02@linaro.org/
>> -
>> https://lore.kernel.org/lkml/f5b7c89c-3bdd-1e1e-772e-721aa5e95bbf@linaro.org/
>> -
>> https://lore.kernel.org/lkml/7328ae17-1dc7-eaa1-5993-411b986e5e02@linaro.org/
>>
>> Also the 7 SoC compatibles do not point to the same crypto IP version. We
>> have two IP versions currently supported upstream, "qcom,crypto-v5.1" and
>> "qcom,crypto-v5.4" (with patches for support for newer versions under work
>> and can be expected to land upstream in near future).
>>
>> However, if you suggest, we can add some comments in the dt-binding doc
>> to reflect which SoC supports which version.
>>
>>>> - Basically, since we are going to have newer qce IP versions available
>>>> in near future, e.g. "qcom,crypto-v5.5" etc, and we will have 2 or more
>>>> SoCs also sharing 1 version, these compatibles would grow and become
>>>> more confusing. IMO, having a soc-specific compatible in such cases is
>>>> probably a much cleaner approach.
>>>>
>>>> Hope this helps answer some of your concerns and provides some relevant
>>>> background information.
>>>
>>> Sure, but I still think you should have only one compatible in the
>>> driver in such case. You don't have differences between them from the
>>> driver point of view, so the devices seem to be compatible.
>>>
>>> If not, what are the differences?
>>
>> There can always be requirements for compatible specific handling done in
>> the driver. See Bjorn's comment here for example:
>> https://lore.kernel.org/lkml/YZKhqJuFlRVeQkCc@builder.lan/ , as an example
>> of 'clk_get' calls conditional based on the compatible instead.
>>
> 
> How about providing a generic compatible without the version number
> (i.e. qcom,crypto) and then in the DT binding require this and
> qcom,<platform>-crypto, and if we have such quirky integration behavior
> for a particular platform we can add the special handling in the driver
> for the platform compatible.
> 
> (And we obviously keep the two existing version-based compatibles in the
> driver, for backwards compatibility)

In general, compatibles should not be generic, just like wild-cards or
family-types. All are discouraged because they do not reflect hardware
or cause troubles later, when it turns out generic is not generic
enough... We had such cases where a wild-card (which is another way to
represent a "generic" compatible) turned out to be incompatible with a
specific version matched by it.

What are you going to do when this happens? Have a generic compatible
which covers only some of implementations? Using a generic compatible
for specific hardware is admitting "we do not know what is generic, so
we make something up and hope it stays generic".


Best regards,
Krzysztof

Powered by blists - more mailing lists