Re: [PATCH 2/2] iio: Add driver for Murata IRS-D200

[Waqar Hameed]

On Sun, Jun 25, 2023 at 12:06 +0100 Jonathan Cameron <jic23@xxxxxxxxxx> wrote:

[...]

>> >> +static int irsd200_write_data_rate(struct irsd200_data *data, int val)
>> >> +{
>> >> +	size_t idx;
>> >> +	int ret;
>> >> +
>> >> +	for (idx = 0; idx < ARRAY_SIZE(irsd200_data_rates); ++idx) {
>> >> +		if (irsd200_data_rates[idx] == val)
>> >> +			break;
>> >> +	}
>> >> +
>> >> +	if (idx == ARRAY_SIZE(irsd200_data_rates))
>> >> +		return -ERANGE;
>> >> +
>> >> +	ret = regmap_write(data->regmap, IRS_REG_DATA_RATE, idx);
>> >> +	if (ret < 0) {
>> >> +		dev_err(data->dev, "Could not write data rate (%d)\n", ret);
>> >> +		return ret;
>> >> +	}
>> >> +
>> >> +	/* Data sheet says the device needs 3 seconds of settling time. */
>> >> +	ssleep(3);  
>> > You aren't preventing other userspace reads / writes during that time so
>> > this is a light protection at best.  
>> 
>> Yes, we aren't preventing anything. The hardware actually operates
>> without any "errors" during this period (e.g. buffer data and events
>> keep arriving). 
>> 
>> This is more of a guarantee that "within 3 s, the new data rate should
>> be in effect". When I think about it, we should maybe just remove this
>> sleep?
> Interesting corner case.  I'd keep as you have it but add a little
> more documentation as why.  Having this sleep will make it easy for a single
> thread in userspace to get what it expects.

Alright, let's update the comment then.

[...]

>> >> +		dev_err(data->dev, "Could not write hp filter frequency (%d)\n",
>> >> +			ret);
>> >> +		return ret;  
>> >
>> > drop this return ret out of the if block here.
>> >
>> > In general being able to ignore possibility of ret > 0 simplifies handling.  
>> 
>> I try to be consistent and it also "helps" the next person potentially
>> adding code after the `if`-statement and forgetting about adding
>> `return`. We can drop the `return here, but then we should do the same
>> in other places with a check just before the last `return` (like
>> `irsd200_write_timer()`, `irsd200_read_nr_count()`,
>> `irsd200_write_nr_count()` and many more), right?
>
> I don't feel particulartly strongly about this, but there are scripts
> that get used to scan for this pattern to simplify the code.
>
> Sure on the other cases. I don't tend to try and label all cases of things
> pointed out, just pick on one and rely on the patch author to generalise.

I don't have strong opinions on this either. Let's remove the `return`.

[...]

>> >> +static irqreturn_t irsd200_irq_thread(int irq, void *dev_id)
>> >> +{
>> >> +	struct iio_dev *indio_dev = dev_id;
>> >> +	struct irsd200_data *data = iio_priv(indio_dev);
>> >> +	enum iio_event_direction dir;
>> >> +	unsigned int lower_count;
>> >> +	unsigned int upper_count;
>> >> +	unsigned int status = 0;
>> >> +	unsigned int source = 0;
>> >> +	unsigned int clear = 0;
>> >> +	unsigned int count = 0;
>> >> +	int ret;
>> >> +
>> >> +	ret = regmap_read(data->regmap, IRS_REG_INTR, &source);
>> >> +	if (ret) {
>> >> +		dev_err(data->dev, "Could not read interrupt source (%d)\n",
>> >> +			ret);
>> >> +		return IRQ_NONE;
>> >> +	}
>> >> +
>> >> +	ret = regmap_read(data->regmap, IRS_REG_STATUS, &status);
>> >> +	if (ret) {
>> >> +		dev_err(data->dev, "Could not acknowledge interrupt (%d)\n",
>> >> +			ret);
>> >> +		return IRQ_NONE;
>> >> +	}
>> >> +
>> >> +	if (status & BIT(IRS_INTR_DATA) && iio_buffer_enabled(indio_dev)) {
>> >> +		iio_trigger_poll_nested(indio_dev->trig);
>> >> +		clear |= BIT(IRS_INTR_DATA);
>> >> +	}
>> >> +
>> >> +	if (status & BIT(IRS_INTR_TIMER) && source & BIT(IRS_INTR_TIMER)) {
>> >> +		iio_push_event(indio_dev,
>> >> +			       IIO_UNMOD_EVENT_CODE(IIO_PROXIMITY, 0,
>> >> +						    IIO_EV_TYPE_CHANGE,
>> >> +						    IIO_EV_DIR_NONE),  
>> >
>> > As below, I'd like more explanation of what this is.
>> > I can't find a datasheet to look it up in.  
>> 
>> This is a timer for the detection event window time, i.e. the signal
>> should pass the threshold values within this time in order to get an
>> interrupt (`IIO_EV_TYPE_THRESH`).
>> 
>> You setup the window time (`IIO_EV_INFO_TIMEOUT`), and when this timer
>> has expired, you get this interrupt (and thus `IIO_EV_TYPE_CHANGE`). I
>> couldn't find any other more fitting value in `enum iio_event_type`.
>
> I'm not totally following.   This is some sort of watchdog?  If threshold
> not exceeded for N seconds an interrupt occurs?  

Yes, exactly.

> Change is definitely not indicating that, so not appropriate ABI to use.
> Timeout currently has a very specific defined meaning and it's not this
> (see the ABI docs - to do with adaptive algorithm jumps - we also have
> reset_timeout but that's different again).
>
> This probably needs some new ABI defining.  I'm not sure what will work
> best though as it's kind of a 'event did not happen' signal if I've understood
> it correctly?

Yeah, I'm not sure when this interrupt actually could be useful. Maybe
when you are testing and calibrating the device, it could help to know
that "these particular settings didn't cause the data to pass any
thresholds during the window time"?

The alternative would be to just ignore this interrupt and not signaling
any events for this. I don't think it would deteriorate the
functionality of the device (except the test/calibration situation
described above, which obviously _can_ be resolved in user space).

>> >> +			       iio_get_time_ns(indio_dev));
>> >> +		clear |= BIT(IRS_INTR_TIMER);
>> >> +	}
>> >> +
>> >> +	if (status & BIT(IRS_INTR_COUNT_THR_OR) &&
>> >> +	    source & BIT(IRS_INTR_COUNT_THR_OR)) {
>> >> +		/*
>> >> +		 * The register value resets to zero after reading. We therefore
>> >> +		 * need to read once and manually extract the lower and upper
>> >> +		 * count register fields.
>> >> +		 */
>> >> +		ret = regmap_read(data->regmap, IRS_REG_COUNT, &count);
>> >> +		if (ret)
>> >> +			dev_err(data->dev, "Could not read count (%d)\n", ret);
>> >> +
>> >> +		upper_count = IRS_UPPER_COUNT(count);
>> >> +		lower_count = IRS_LOWER_COUNT(count);
>> >> +
>> >> +		/*
>> >> +		 * We only check the OR mode to be able to push events for
>> >> +		 * rising and falling thresholds. AND mode is covered when both
>> >> +		 * upper and lower count is non-zero, and is signaled with
>> >> +		 * IIO_EV_DIR_EITHER.  
>> >
>> > Whey you say AND, you mean that both thresholds have been crossed but also that
>> > in that configuration either being crossed would also have gotten us to here?
>> > (as opposed to the interrupt only occurring if value is greater than rising threshold
>> >  and less than falling threshold?)
>> >
>> > If it's the first then just report two events.  Either means we don't know, rather
>> > than we know both occurred.  We don't document that well though - so something
>> > we should improved there. whilst a bit confusing: 
>> > https://elixir.bootlin.com/linux/v6.4-rc6/source/Documentation/ABI/testing/sysfs-bus-iio#L792
>> > talks about this.
>> >
>> > The bracketed case is more annoying to deal with so I hope you don't mean that.
>> > Whilst we've had sensors that support it in hardware for years, I don't think we
>> > ever came up with a usecase that really justified describing it.  
>> 
>> According to the data sheet (which will hopefully be soon publicly
>> available):
>> 
>> OR-interrupt:  (UPPER_COUNT + LOWER_COUNT >= NR_COUNT)
>> 
>> AND-interrupt: (UPPER_COUNT + LOWER_COUNT >= NR_COUNT) &&
>>                (UPPER_COUNT != 0) && (LOWER_COUNT != 0)
>>                
>> For example, consider the following situation:
>> 
>>                                ___
>>                               /   \
>> -----------------------------3------------------- Upper threshold
>>                ___          /       \
>> ______        /   \        /         \___________ Data signal
>>       \      /     \      /
>> -------1------------2---------------------------- Lower threshold
>>         \__/         \__/
>>         
>> When `NR_COUNT` is set to 3, we will get an OR-interrupt on point "3" in
>> the graph above. In this case `UPPER_COUNT = 1` and `LOWER_COUNT = 2`.
>> 
>> When `NR_COUNT` is set to 2, we will get an OR-interrupt on point "2"
>> instead. Here `UPPER_COUNT = 0` and `LOWER_COUNT = 2`.
>> 
>
> Thanks.  That is very odd definition of AND.  At least OR is close to normal
> though the way count is applied is unusual.  Most common thing similar to that
> is what we use period for in IIO - it's same count here, but it resets once
> the condition is no longer true.  Here we have a running total...
>
> Getting this into standard ABI or anything approaching it is going to be tricky.
>
> Firstly need a concept similar to period but with out the reset. That will at least
> allow us to comprehend the counts part.
>
> Either can then be used for the OR case.

Are you saying that the current implementation (with manually checking
the upper and lower counts only with the OR mode) wouldn't "fit" the
current ABI? It does cover the rising and falling directions correctly,
no? Could `IIO_EV_DIR_NONE` instead of `IIO_EV_DIR_EITHER` be used to
signal "both" then?

>
> The AND case is a mess so for now I'm stuck.  Will think some more on this.
> Out of curiosity does the datasheet include why that particular function makes
> any sense?  Feels like a rough attempt to approximate something they don't have
> hardware resources to actually estimate.

Unfortunately not. I guess there could be an application where you are
only interested if _both_ lower and upper threshold are exceeded. Maybe
in order to minimize small "false positives" movements in front of the
sensor? But as stated in the comments, one can cover this with only the
OR mode (and manually checking the upper and lower count as we do).