Re: [PATCH v2 2/2] IIO: st_accel_i2c.c: Use probe_new() instead of probe()

[Nikolaus Voss <nikolaus.voss@loewensteinmedical.de>]

Hi Javier,

On Wed, 4 Jul 2018, Javier Martinez Canillas wrote:
> Hi Nikolaus,
>
> On Wed, Jul 4, 2018 at 1:15 PM, Nikolaus Voss
> <nikolaus.voss@loewensteinmedical.de> wrote:
>> On Wed, 4 Jul 2018, Javier Martinez Canillas wrote:
>
> [snip]
>
>>> I think Nikolaus is correct, assuming that the driver can be used on
>>> legacy
>>> platforms that register the I2C devices using board files / platform data.
>>> In that case you still need a I2C device ID table for (a) and (c) as he
>>> said.
>>>
>>> I don't buy on (b) though, that's a bug in my opinion. If you register an
>>> I2C
>>> device via DT then you must have a OF device ID entry that matches the
>>> device
>>> and the same for ACPI. You can't rely on the I2C device table to do the
>>> match.
>>
>>
>> Ok, in my opinion it is an elegant way of not bloating the driver when no
>> explicit handling (e.g. reading DT properties) is needed. Just adding an
>> of_device_id doesn't change any driver functionality then but only maps to
>> an already existing i2c_table_id.
>>
>
> I disagree, in the case of OF for example a compatible string is
> composed of both a vendor a device, the complete tuple is what should
> be matched.
>
> But with the fallback, only the device portion would be used so both
> <foo,bar> and <baz,bar> will match against the i2c device with id
> "bar". It may or may not be correct but the vendor portion is very
> important to disambiguate.

Disambiguation via DT implies there is already a name collision in i2c 
modalias name space, so adding an of_id would work around this collision 
for DT enumeration. I2c_board_info driver binding would still be broken. 
The right fix would be to remove the name collision.

>>>
>>> I would also remove the struct i2c_device_id .driver_data fields from the
>>> I2C
>>> device ID table, since are not used and just makes reading the code
>>> confusing
>>> (only the struct i2c_device_id .name is used as far as I can see).
>>
>>
>> Valid point, thanks. I will change that.
>>
>>>
>>>> Javier, just a summary of the above. Nikolaus switched one driver to
>>>> use ->probe_new() hook and left i2c ID table at the same time.
>>>> My understanding that this table is not anymore in use.
>>>>
>>>> But I have to admit I didn't see entire picture of this. Can you shed a
>>>> light?
>>>>
>>>
>>> So to shed some light, in the past even {OF,ACPI}-only drivers needed an
>>> I2C ID
>>> table because: 1) the .probe callback had a struct i2c_device_id *
>>> parameter
>>> and 2) the I2C core always reported a modalias of the form i2c:<foo> even
>>> for
>>> devices registered via OF.
>>
>>
>> It could have been a null pointer and device driver binding (and
>> auto-loading) done just via driver.name.
>>
>
> I'm not sure I understood this comment.

What I was trying to say is that if the i2c_device_id table information 
was not needed (i.d. only one single id), even the old probe() could be 
used without defining an i2c_device_id table, the i2c_device_id* argument 
to probe() being a nullptr.

Niko