On 20/01/2021 12:57, Andy Shevchenko wrote:
> On Wed, Jan 20, 2021 at 06:21:41AM +0200, Laurent Pinchart wrote:
>> On Tue, Jan 19, 2021 at 07:51:14PM +0200, Andy Shevchenko wrote:
>>> On Tue, Jan 19, 2021 at 06:48:15PM +0200, Laurent Pinchart wrote:
>>>> On Tue, Jan 19, 2021 at 01:08:37PM +0200, Andy Shevchenko wrote:
>>>>> On Tue, Jan 19, 2021 at 10:40:42AM +0000, Daniel Scally wrote:
>>>>>> On 19/01/2021 09:24, Andy Shevchenko wrote:
>>>>>>>>>>> +static struct i2c_driver int3472_tps68470 = {
>>>>>>>>>>> + .driver = {
>>>>>>>>>>> + .name = "int3472-tps68470",
>>>>>>>>>>> + .acpi_match_table = int3472_device_id,
>>>>>>>>>>> + },
>>>>>>>>>>> + .probe_new = skl_int3472_tps68470_probe,
>>>>>>>>>>> +};
>>>>>>>>> I'm not sure we want to have like this. If I'm not mistaken the I²C driver can
>>>>>>>>> be separated without ACPI IDs (just having I²C IDs) and you may instantiate it
>>>>>>>>> via i2c_new_client_device() or i2c_acpi_new_device() whichever suits better...
>>>>>>>> Sorry, I'm a bit confused by this. The i2c device is already
>>>>>>>> present...we just want the driver to bind to them, so what role do those
>>>>>>>> functions have there?
>>>>>>> What I meant is something like
>>>>>>>
>>>>>>> *_i2c.c
>>>>>>> real I²C driver for the TPS chip, but solely with I²C ID table, no ACPI
>>>>>>> involved (and it sounds like it should be mfd/tps one, in which you
>>>>>>> just cut out ACPI IDs and convert to pure I²C one, that what I had
>>>>>>> suggested in the first place)
>>>>>>
>>>>>> Ahh; sorry - i misunderstood what you meant there. I understand now I
>>>>>> think, but there is one complication; the ACPI subsystem already creates
>>>>>> a client for that i2c adapter and address; i2c_new_client_device()
>>>>>> includes a check to see whether that adapter / address combination has
>>>>>> an i2c device already. So we would have to have the platform driver
>>>>>> with ACPI ID first find the existing i2c_client and unregister it before
>>>>>> registering the new one...the existing clients have a name matching the
>>>>>> ACPI device instance name (e.g i2c-INT3472:00) which we can't use as an
>>>>>> i2c_device_id of course.
>>>>>
>>>>> See how INT33FE is being handled. Hint: drivers/acpi/scan.c:~1600
>>>>>
>>>>> static const struct acpi_device_id i2c_multi_instantiate_ids[] = {
>>>>> {"BSG1160", },
>>>>> {"BSG2150", },
>>>>> {"INT33FE", },
>>>>> {"INT3515", },
>>>>> {}
>>>>> };
>>>>>
>>>>> So, we quirklist it here and instantiate manually from platform driver (new
>>>>> coming one).
>>>>
>>>> This is documented as used for devices that have multiple I2cSerialBus
>>>> resources. That's not the case for the INT3472 as far as I can tell. I
>>>> don't think we should abuse this mechanism.
>>>
>>> This is quite a similar case to that one. Let's avoid yak shaving, right?
>>
>> Exactly my point, that's why I think this patch is good overall, I don't
>> think it requires a complete rewrite.
>
> The approach in the series is to reinvent the MFD driver which I against of.
> I don;t think we need to kill it there and reborn in a new form and dragging
> code from there to here to there.
>
> On top of that the approach with a quirk driver in the middle seems to me
> cleaner than using different paths how the two drivers are being initialized.
> In the proposed approach there will be one making decision point and easy to
> understand what's going on.
>
> The bad example of two making decision points is acpi_lpss.c vs. individual
> drivers (however in that case it have different ID's, i.e. ACPI vs. PCI),
Right; so if I understand correctly, the proposal is:
1. Add INT3472 to the i2c_multi_instantiate_ids, which blocks it getting
created as an i2c device
2. instead of intel-skl-int3472 registering an i2c and a platform
driver, just register a platform driver that binds to the INT3472
acpi_device_id. We can check hardware type like in
intel_cht_int33fe_common.c and call either discrete probe that does what
the discrete driver is doing now, or else call tps68470 which is just a
stub driver registering an i2c device like intel_cht_int33fe_microb.c
3. Change the existing tps68470 mfd driver to match to that created i2c
device instead of ACPI match, and move the code from
intel_skl_int3472_tps68470.c to that driver instead
I think I finally got what you meant there, Andy, but correct me if I'm
wrong please.
I'm not sure that one's better than the other, to be honest. Either the
multi-function device functionality lives in the conventional place, or
else _all_ of the int3472 handling code lives together in one module.
>>>> Don't forget that the TPS68470 I2C driver needs to be ACPI-aware, as it
>>>> has to register an OpRegion for ACPI-based Chrome OS devices. On other
>>>> platforms (including DT platforms), it should only register regulators,
>>>> clocks and GPIOs. Given the differences between those platforms, I don't
>>>> think a TPS68470 driver that would fake being unaware of being probed
>>>> through ACPI would be a good idea. We can always refactor the code later
>>>> when we'll have a non-ACPI based platform using the TPS68470, without
>>>> such a platform there's no way we can test the I2C driver without ACPI
>>>> anyway.
>>>
>>> Are you agree that MFD approach should stay? How then we can manage to have an
>>> MFD driver cohabit with our new driver? I proposed a clean solution which will
>>> handle all possible cases via quirk driver. Having two drivers enumerated by
>>> different scenarios is a call for troubles (we have already with one of that
>>> sensors).
What kind of troubles do you anticipate here?
>> I think we should solve this problem when it will arise. Solving
>> problems with complex architectures without a platform to test the code
>> on is a pretty sure way to get the architecture design wrong. Let's get
>> this merged, it's an improvement compared to the current situation, and
>> then let's improve it further on top when we'll need to support more use
>> cases.
>
> But this is problem already here right now. The submitted code is to support
> a new platform that needs a quirk and treats INT3472 differently. The usual
> way is to refactor the existing solution to make them both to have a best
> compromise.
>
>>> And there is no "faking" anything, it's rather gating it depending on the
>>> platform.
>
