Re: [PATCH v3 1/4] hwmon: iio_hwmon: delay probing with late_initcall

[Quentin Schulz <quentin.schulz@free-electrons.com>]

On 15/08/2016 23:35, Jonathan Cameron wrote:
> 
> 
> On 15 August 2016 18:07:30 BST, Guenter Roeck <linux@roeck-us.net> wrote:
>> On Mon, Aug 15, 2016 at 04:40:21PM +0100, Jonathan Cameron wrote:
>>> On 26/07/16 17:04, Guenter Roeck wrote:
>>>> On Tue, Jul 26, 2016 at 12:00:33PM +0200, Alexander Stein wrote:
>>>>> On Tuesday 26 July 2016 11:33:59, Quentin Schulz wrote:
>>>>>> On 26/07/2016 11:05, Alexander Stein wrote:
>>>>>>> On Tuesday 26 July 2016 10:24:48, Quentin Schulz wrote:
>>>>>>>> On 26/07/2016 10:21, Alexander Stein wrote:
>>>>>>>>> On Tuesday 26 July 2016 09:43:44, Quentin Schulz wrote:
>>>>>>>>>> iio_channel_get_all returns -ENODEV when it cannot find
>> either phandles
>>>>>>>>>> and
>>>>>>>>>> properties in the Device Tree or channels whose
>> consumer_dev_name
>>>>>>>>>> matches
>>>>>>>>>> iio_hwmon in iio_map_list. The iio_map_list is filled in by
>> iio drivers
>>>>>>>>>> which might be probed after iio_hwmon.
>>>>>>>>>
>>>>>>>>> Would it work if iio_channel_get_all returning ENODEV is used
>> for
>>>>>>>>> returning
>>>>>>>>> EPROBE_DEFER in iio_channel_get_all? Using late initcalls for
>>>>>>>>> driver/device
>>>>>>>>> dependencies seems not right for me at this place.
>>>>>>>>
>>>>>>>> Then what if the iio_channel_get_all is called outside of the
>> probe of a
>>>>>>>> driver? We'll have to change the error code, things we are
>> apparently
>>>>>>>> trying to avoid (see v2 patches' discussions).
>>>>>>>
>>>>>>> Maybe I didn't express my idea enough. I don't want to change
>> the behavior
>>>>>>> of iio_channel_get_all at all. Just the result evaluation of
>>>>>>> iio_channel_get_all in iio_hwmon_probe. I have something link
>> the patch
>>>>>>> below in mind.
>>>>>>>
>>>>>>> Best regards,
>>>>>>> Alexander
>>>>>>> ---
>>>>>>> diff --git a/drivers/hwmon/iio_hwmon.c
>> b/drivers/hwmon/iio_hwmon.c
>>>>>>> index b550ba5..e32d150 100644
>>>>>>> --- a/drivers/hwmon/iio_hwmon.c
>>>>>>> +++ b/drivers/hwmon/iio_hwmon.c
>>>>>>> @@ -73,8 +73,12 @@ static int iio_hwmon_probe(struct
>> platform_device
>>>>>>> *pdev)
>>>>>>>
>>>>>>>                 name = dev->of_node->name;
>>>>>>>         
>>>>>>>         channels = iio_channel_get_all(dev);
>>>>>>>
>>>>>>> -       if (IS_ERR(channels))
>>>>>>> -               return PTR_ERR(channels);
>>>>>>> +       if (IS_ERR(channels)) {
>>>>>>> +               if (PTR_ERR(channels) == -ENODEV)
>>>>>>> +                       return -EPROBE_DEFER;
>>>>>>> +               else
>>>>>>> +                       return PTR_ERR(channels);
>>>>>>> +       }
>>>>>>>
>>>>>>>         st = devm_kzalloc(dev, sizeof(*st), GFP_KERNEL);
>>>>>>>         if (st == NULL) {
>>>>>>
>>>>>> Indeed, I misunderstood what you told me.
>>>>>>
>>>>>> Actually, the patch you proposed is part of my v1
>>>>>> (https://lkml.org/lkml/2016/6/28/203) and v2
>>>>>> (https://lkml.org/lkml/2016/7/15/215).
>>>>>> Jonathan and Guenter didn't really like the idea of changing the
>> -ENODEV
>>>>>> in -EPROBE_DEFER.
>>>>>
>>>>> Thanks for the links.
>>>>>
>>>>>> What I thought you were proposing was to change the -ENODEV
>> return code
>>>>>> inside iio_channel_get_all. This cannot be an option since the
>> function
>>>>>> might be called outside of a probe (it is not yet, but might be
>> in the
>>>>>> future?).
>>>>>
>>>>> AFAICS this is a helper function not knowing about device probing
>> itself. And 
>>>>> it should stay at that.
>>>>>
>>>>>> Of what I understood, two possibilities are then possible
>> (proposed
>>>>>> either by Guenter or Jonathan): either rework the iio framework
>> to
>>>>>> register iio map array earlier or to use late_initcall instead of
>> init
>>>>>> for the driver consuming the iio channels.
>>>>>
>>>>> Interestingly using this problem would not arise due to module
>> dependencies. 
>>>>> But using late_initcall would mean this needs to be done on any
>> driver using 
>>>>> iio channels? I would rather keep those consumers simple.
>>>>>
>>>> Me too, but that would imply a solution in iio. The change you
>> propose above
>>>> isn't exactly simple either, and would also be needed in each
>> consumer driver.
>>>>
>>>> Just for the record, I dislike the late_initcall solution as well,
>> but I prefer
>>>> it over blindly converting ENODEV to EPROBE_DEFER.
>>> I'm falling on the other side on this one right now. Though I'd be
>> tempted
>>> to renaming the function to something like
>> iio_channel_get_all_or_defer
>>> to make it explicit that it can result in deferred probing.
>>>
>> Would this new function return -EPROBE_DEFER instead of -ENODEV ?
> Yes. Though whether it really adds much over doing that in drivers isn't clear.
> 
> Hmm. Needs more thought...

Either we do the exact same "hack" as in the v2[1] in what you call
iio_channel_get_all_or_defer or we duplicate the code from
iio_channel_get_all in iio_channel_get_all_or_defer. Both do not seem
right to me but I really dislike the late_initcall method. With this
method we can only have one level of "channel dependency".

This means if we ever create a new driver which depends on channels from
the driver using late_initcall, we will also have to use late_initcall
and we can't be sure the new driver will always be probed after the
driver he depends on.

[1] https://lkml.org/lkml/2016/7/15/215

Quentin
>>
>> Thanks,
>> Guenter
>> --
>> To unsubscribe from this list: send the line "unsubscribe linux-iio" in
>> the body of a message to majordomo@vger.kernel.org
>> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>