Re: [PATCH v2 06/10] iio: adc: ad7476: Drop convstart chan_spec

[Matti Vaittinen <mazziesaccount@gmail.com>]

On 08/08/2025 12:00, Nuno Sá wrote:
> On Fri, Aug 08, 2025 at 08:37:07AM +0300, Matti Vaittinen wrote:
>> On 07/08/2025 16:10, Nuno Sá wrote:
>>> On Thu, Aug 07, 2025 at 01:41:31PM +0100, Nuno Sá wrote:
>>>> On Thu, Aug 07, 2025 at 12:34:52PM +0300, Matti Vaittinen wrote:
>>>>> The ad7476 driver defines separate chan_spec structures for operation
>>>>> with and without convstart GPIO. At quick glance this may seem as if the
>>>>> driver did provide more than 1 data-channel to users - one for the
>>>>> regular data, other for the data obtained with the convstart GPIO.
>>>>>
>>>>> The only difference between the 'convstart' and 'non convstart'
>>>>> -channels is presence / absence of the BIT(IIO_CHAN_INFO_RAW) in
>>>>> channel's flags.
>>>>>
>>>>> We can drop the convstart channel spec, and related convstart macro, by
>>>>> allocating a mutable per driver instance channel spec an adding the flag
>>>>> in probe if needed. This will simplify the driver with the cost of added
>>>>> memory consumption.
>>>>>
>>>>> Assuming there aren't systems with very many ADCs and very few
>>>>> resources, this tradeoff seems worth making.
>>>>>
>>>>> Simplify the driver by dropping the 'convstart' channel spec and
>>>>> allocating the chan spec for each driver instance.
>>>>
>>>> I do not agree with this one. Looking at the diff, code does not look
>>>> simpler to me...
>>>
>>> Ok, on a second thought I'm ok with this. It makes adding devices easier
>>> and (IIUC) for the one you're adding later we only have "convst_channel"
>>> channels.
>>
>> Yes, that's right. The BD79105 requires the convstart.
>>
>>> On comment though...
>>>
>>>>
>>>> - Nuno Sá
>>>>
>>>>>
>>>>> Signed-off-by: Matti Vaittinen <mazziesaccount@gmail.com>
>>>>>
>>>>> ---
>>>>> Revision history:
>>>>>    v1 => v2:
>>>>>    - New patch
>>>>>
>>>>> I considered squashing this change with the one limiting the chip_info
>>>>> scope. Having this as a separate change should help reverting if someone
>>>>> complains about the increased memory consumption though.
>>>>> ---
>>>>>    drivers/iio/adc/ad7476.c | 31 ++++++++++++++++++-------------
>>>>>    1 file changed, 18 insertions(+), 13 deletions(-)
>>>>>
>>>>> diff --git a/drivers/iio/adc/ad7476.c b/drivers/iio/adc/ad7476.c
>>>>> index e97742912b8e..a30eb016c11c 100644
>>>>> --- a/drivers/iio/adc/ad7476.c
>>>>> +++ b/drivers/iio/adc/ad7476.c
>>>>> @@ -29,8 +29,6 @@ struct ad7476_state;
>>>>>    struct ad7476_chip_info {
>>>>>    	unsigned int			int_vref_mv;
>>>>>    	struct iio_chan_spec		channel[2];
>>>>> -	/* channels used when convst gpio is defined */
>>>>> -	struct iio_chan_spec		convst_channel[2];
>>>>>    	void (*reset)(struct ad7476_state *);
>>>>>    	bool				has_vref;
>>>>>    	bool				has_vdrive;
>>>>> @@ -41,6 +39,7 @@ struct ad7476_state {
>>>>>    	struct gpio_desc		*convst_gpio;
>>>>>    	struct spi_transfer		xfer;
>>>>>    	struct spi_message		msg;
>>>>> +	struct iio_chan_spec		channel[2];
>>>>>    	int				scale_mv;
>>>>>    	/*
>>>>>    	 * DMA (thus cache coherency maintenance) may require the
>>>>> @@ -153,24 +152,18 @@ static int ad7476_read_raw(struct iio_dev *indio_dev,
>>>>>    #define AD7940_CHAN(bits) _AD7476_CHAN((bits), 15 - (bits), \
>>>>>    		BIT(IIO_CHAN_INFO_RAW))
>>>>>    #define AD7091R_CHAN(bits) _AD7476_CHAN((bits), 16 - (bits), 0)
>>>>> -#define AD7091R_CONVST_CHAN(bits) _AD7476_CHAN((bits), 16 - (bits), \
>>>>> -		BIT(IIO_CHAN_INFO_RAW))
>>>>>    #define ADS786X_CHAN(bits) _AD7476_CHAN((bits), 12 - (bits), \
>>>>>    		BIT(IIO_CHAN_INFO_RAW))
>>>>>    static const struct ad7476_chip_info ad7091_chip_info = {
>>>>>    	.channel[0] = AD7091R_CHAN(12),
>>>>>    	.channel[1] = IIO_CHAN_SOFT_TIMESTAMP(1),
>>>>> -	.convst_channel[0] = AD7091R_CONVST_CHAN(12),
>>>>> -	.convst_channel[1] = IIO_CHAN_SOFT_TIMESTAMP(1),
>>>>>    	.reset = ad7091_reset,
>>>>>    };
>>>>>    static const struct ad7476_chip_info ad7091r_chip_info = {
>>>>>    	.channel[0] = AD7091R_CHAN(12),
>>>>>    	.channel[1] = IIO_CHAN_SOFT_TIMESTAMP(1),
>>>>> -	.convst_channel[0] = AD7091R_CONVST_CHAN(12),
>>>>> -	.convst_channel[1] = IIO_CHAN_SOFT_TIMESTAMP(1),
>>>>>    	.int_vref_mv = 2500,
>>>>>    	.has_vref = true,
>>>>>    	.reset = ad7091_reset,
>>>>> @@ -282,7 +275,7 @@ static int ad7476_probe(struct spi_device *spi)
>>>>>    	const struct ad7476_chip_info *chip_info;
>>>>>    	struct ad7476_state *st;
>>>>>    	struct iio_dev *indio_dev;
>>>>> -	int ret;
>>>>> +	int ret, i;
>>>>>    	indio_dev = devm_iio_device_alloc(&spi->dev, sizeof(*st));
>>>>>    	if (!indio_dev)
>>>>> @@ -332,16 +325,28 @@ static int ad7476_probe(struct spi_device *spi)
>>>>>    	if (IS_ERR(st->convst_gpio))
>>>>>    		return PTR_ERR(st->convst_gpio);
>>>>> +	/*
>>>>> +	 * This will never realize. Unless someone changes the channel specs
>>>>> +	 * in this driver. And if someone does, without changing the loop
>>>>> +	 * below, then we'd better immediately produce a big fat error, before
>>>>> +	 * the change proceeds from that developer's table.
>>>>> +	 */
>>>>> +	BUILD_BUG_ON(ARRAY_SIZE(st->channel) != ARRAY_SIZE(chip_info->channel));
>>>
>>> I guess it make sense but still looks too fancy for this :)
>>
>> Nothing else but a developer's carefulness keeps the number of channels "in
>> sync" for these two structs. I was originally doing WARN_ON() - but then I
>> thought that it's be even better to catch this at build time. Then I found
>> the BUILD_BUG_ON(). I see Andy suggested static_assert() instead - I've no
>> idea why one is preferred over other though. Let's see if I get educated by
>> Andy :)
>>
>>>
>>>>> +	for (i = 0; i < ARRAY_SIZE(st->channel); i++) {
>>>>> +		st->channel[i] = chip_info->channel[i];
>>>>> +		if (st->convst_gpio)
>>>
>>> I would flip this an do:
>>> 	if (!st->convst_gpio)
>>> 		break;
>>
>> To me this would just add an extra line of code, and more complex flow. I
>> would definitely agree if there were more operations to be done for the
>> 'convstart channels' - but since this is really just "if it's convstart,
>> then set a bit" - the
>>
>> if (foo)
>> 	bar;
>>
>> seems simpler than
>>
>> if (!foo)
>> 	break;
>> bar;
> 
> Yes but in this particular case, you likely would not need to do any
> line break afterward because of indentation. Logically it also makes
> sense because st->convst_gpio is a device property (not a channel one).
> So it makes no sense to check it for all channels (I know we only have two
> channels). So if you prefer, you could even do:
> 
> if (st->convst_gpio) {
> 	for (...)
> 		__set_bit(...);
> }
> 
> which also would make more sense to me.

I considered this option, but I need to populate all the channels in
st->channel with the template data from chip_info->channel anyways. 
Hence I want to loop through the channels also when the st->convst_gpio 
is not there :)

> 
> Up to you now :)

Well, I already sent the v3. (Sorry, I should've waited a bit more but 
wanted to get it out before the weekend). I kept the same logic as in 
v2. You can still suggest improvements there!

Yours,
	-- Matti