Re: [PATCH v2 2/2] i2c: add support for Socionext SynQuacer I2C controller

[Ard Biesheuvel]

On 26 February 2018 at 17:05, Andy Shevchenko <andy.shevchenko@xxxxxxxxx> wrote:
> On Mon, Feb 26, 2018 at 4:58 PM, Ard Biesheuvel
> <ard.biesheuvel@xxxxxxxxxx> wrote:
>> On 26 February 2018 at 11:35, Andy Shevchenko <andy.shevchenko@xxxxxxxxx> wrote:
>>> On Mon, Feb 26, 2018 at 11:59 AM, Ard Biesheuvel
>>> <ard.biesheuvel@xxxxxxxxxx> wrote:
>>>> On 23 February 2018 at 13:12, Andy Shevchenko <andy.shevchenko@xxxxxxxxx> wrote:
>>>>> On Fri, Feb 23, 2018 at 2:40 PM, Ard Biesheuvel
>>>>> <ard.biesheuvel@xxxxxxxxxx> wrote:
>
>>>>> Replace 'baudclk' with 'pclk' and p->uartclk with i2c->clkrate in
>>>>> above and you are almost done.
>
>>>> I don't think this is better.
>>>
>>> It's a pattern over ACPI vs. clk cases at least for now.
>>> But hold on. We have already an example of dealing with ACPI /
>>> non-ACPI cases for I2C controllers â i2c-designware-platdrv.c.
>>> Check how it's done there.
>>>
>>> I actually totally forgot about ACPI slaves described in the table. We
>>> need to take into account the ones with lowest bus speed.
>>>
>>
>> Wow, that code is absolutely terrible.
>
> To some degree I may say yes it is.
>
>> So even while _DSD device properties require vendor prefixes, which
>> are lacking in this case,
>
> What kind? clock-frequency? Does it require prefix?
>

What I remember from the _DSD discussions is that we should vendor
prefixes for per-device properties, and only use unprefixed names for
generic properties. However, looking more closely, I understand that
this undermines the idea of having parity between DT and ACPI, because
DT did not require vendor prefixes in the past (but it does now)

I guess 'clock-frequency' is one that would not require such a vendor prefix.

>> and the fact that the ACPI flavor of the
>> Designware I2C controller now provides two different ways to get the
>> timing parameters (using device properties or using SSCN/FMCN/etc ACPI
>> methods), you think this is a shining example of how this should be
>> implemented?
>
> No, those methods because of windows driver and existed ACPI tables at
> that time.
> You are not supposed to uglify your case.
>

OK, in that case, can you please spell out what you think is
mandatory? Because handwavy references to existing UART and I2C
drivers are not helping me here.

>> Also, I still think implementing a clock device using rate X just to
>> interrogate it for its rate (returning X) is absolutely pointless.
>
> OTOH the deviation in the driver is what I absolutely against of.
> Driver must not know the resource provider (ideally at all).
>

There is no 'resource provider'. There is only a single number, which
is the clock rate, and is only used to calculate some internal
dividers of the I2C IP block.

>> So what I can do is invent an ACPI method that returns the PCLK rate.
>> Would that work for you?
>
> Again, looking into existing examples (UART, I2C, etc) we better to
> create a generic helper in clock framework that would provide us a
> clock based on property value.
> But doing different paths for different resource providers is not what
> we are looking for.
>
> P.S. To move this somehow forward I may propose to submit an OF
> driver, and discuss ACPI part after.
>

Thanks, but that does not really work for me. What I can do is split
it into an initial DT only driver, and a followup ACPI patch.

Can you point me to an example of such a clock provider?