Re: [PATCH 2/2] iio: pressure: driver for Honeywell HSC/SSC series pressure sensors

[Andy Shevchenko <andriy.shevchenko@linux.intel.com>]

On Wed, Nov 22, 2023 at 08:08:27AM +0200, Petre Rodan wrote:
> On Mon, Nov 20, 2023 at 02:35:39PM +0200, Andy Shevchenko wrote:

...

> sorry, what is 'LKP' in this context and how do I reproduce?

It's an acronym for CI system run by Intel. You should have had an email in
your mailbox with complains. It also duplicates them to a mailing list which
address I don't know by heart.

...

> > Also there are missing at least these ones: array_size.h, types.h.
> 
> '#include <linux/array_size.h>' is a weird one.

Why?

> $ cd /usr/src/linux/drivers
> $ grep -r ARRAY_SIZE * | grep '\.c:' |  wc -l
>  32396
> $ grep -r 'include.*array_size\.h' * | grep -E '\.[ch]:' | wc -l
> 11
> $ grep -r 'include.*array_size\.h' * | grep -E '\.[ch]:' | grep -v '^pinctrl' | wc -l
> 0

Hint, use `git grep ...` which much, much faster against the Git indexed data.

> plus on a 6.1 version kernel, `make modules` actually reports that the header
> can't be found if I include it. can't comprehend that. so I'll be skipping
> that particular include.

No, the new code is always should be submitted against latest release cycle,
v6.7-rcX as of today. There is the header. Please, use it.

...

> > Can you utilize linear ranges data types and APIs? (linear_range.h)
> 
> not fit for this purpose, sorry.

NP.

...

> > > +	if (data->buffer[0] & 0xc0)
> > > +		return 0;
> > > +
> > > +	return 1;
> > 
> > You use bool and return integers.
> > 
> > Besides, it can be just a oneliner.
> 
> rewritten as a one-liner, without GENMASK.
> 
> > 	return !(buffer[0] & GENMASK(3, 2));
> > 
> > (Note, you will need bits.h for this.)
> > 
> > > +}

Why no GENMASK() ? What the meaning of the 0xc0?
Ideally it should be

#define ...meaningful name...  GENMASK()

...

> > > +		mutex_lock(&data->lock);
> > > +		ret = hsc_get_measurement(data);
> > > +		mutex_unlock(&data->lock);
> > 
> > Use guard() operator from cleanup.h.
> 
> I'm not familiar with that, for the time being I'll stick to
> mutex_lock/unlock if you don't mind.

I do mind. RAII is a method to make code more robust against forgotten
unlock/free calls.

...

> > > +		case IIO_PRESSURE:
> > > +			*val =
> > > +			    ((data->buffer[0] & 0x3f) << 8) + data->buffer[1];
> > > +			return IIO_VAL_INT;
> > > +		case IIO_TEMP:
> > > +			*val =
> > > +			    (data->buffer[2] << 3) +
> > > +			    ((data->buffer[3] & 0xe0) >> 5);
> > 
> > Is this some endianess / sign extension? Please convert using proper APIs.
> 
> the raw conversion data is spread over 4 bytes and interlaced with other info
> (see comment above the function).  I'm just cherry-picking the bits I'm
> interested in, in a way my brain can understand what is going on.

So, perhaps you need to use get_unaligned_.e32() and then FIELD_*() from
bitfield.h. This will be much better in terms of understanding the semantics
of these magic bit shifts and masks.

...

> > > +			ret = 0;
> > > +			if (!ret)
> > 
> > lol
> 
> I should leave that in for comic relief. missed it after a lot of code
> changes.

I understand, that's why no shame on you, just fun code to see :-)

...

> > Strange indentation of }:s...
> 
> I blame `indent -linux --line-length 80` for these and weirdly-spaced pointer
> declarations.  are you using something else?

Some maintainers suggest to use clang-format. I find it weird in some corner
cases. So, I would suggest to use it and reread the code and fix some
strangenesses.

...

> > > +	if (strcasecmp(hsc->range_str, "na") != 0) {
> > > +		// chip should be defined in the nomenclature
> > > +		for (index = 0; index < ARRAY_SIZE(hsc_range_config); index++) {
> > > +			if (strcasecmp
> > > +			    (hsc_range_config[index].name,
> > > +			     hsc->range_str) == 0) {
> > > +				hsc->pmin = hsc_range_config[index].pmin;
> > > +				hsc->pmax = hsc_range_config[index].pmax;
> > > +				found = 1;
> > > +				break;
> > > +			}
> > > +		}
> > 
> > Reinventing match_string() / sysfs_match_string() ?
> 
> match_string() is case-sensitive and operates on string arrays, so unfit for
> this purpose.

Let's put it this way: Why do you care of the relaxed case?
I.o.w. why can we be slightly stricter?

...

> > Can you use regmap I2C?
> 
> the communication is one-way as in the sensors do not expect anything except
> 4 bytes-worth of clock signals per 'packet' for both the i2c and spi
> versions.  regmap is suited to sensors with an actual memory map.

If not yet, worse to add in the comment area of the patch
(after the cutter '---' line).

...

> > No use of this function prototype, we have a new one.
> 
> oops, I was hoping my 6.1.38 kernel is using the same API as 6.7.0
> fixed.

Same way with a (new) header :-)

...

> > > +	ret = devm_regulator_get_enable_optional(dev, "vdd");
> > > +	if (ret == -EPROBE_DEFER)
> > > +		return -EPROBE_DEFER;
> > 
> > Oh, boy, this should check for ENODEV or so, yeah, regulator APIs a bit
> > interesting.
> 
> since I'm unable to test this I'd rather remove the block altogether.
> if I go the ENODEV route my module will never load since I can't see any
> vdd-supply support on my devboard.

No, what I meant is to have something like

	if (ret) {
		if (ret != -ENODEV)
			return ret;
		...regulator is not present...
	}

This is how it's being used in dozens of places in the kernel. Just utilize
`git grep ...` which should be a top-10 tool for the Linux kernel developer.

Q: ...
A: Try `git grep ...` to find your answer in the existing code.

...

> > > +	if (!dev_fwnode(dev))
> > > +		return -EOPNOTSUPP;
> > 
> > Why do you need this?
> > And why this error code?
> 
> it's intentional.
> this module has a hard requirement on the correct parameters (transfer
> function and pressure range) being provided in the devicetree.  without those
> I don't want to provide any measurements since there can't be a default
> transfer function and pressure range for a generic driver that supports an
> infinite combination of those.
> 
> echo hsc030pa 0x28 > /sys/bus/i2c/devices/i2c-0/new_device
> I want iio_info to detect 0 devices.

So, fine, but the very first mandatory property check will fail as it has
the very same check inside. So, why do you need a double check?

-- 
With Best Regards,
Andy Shevchenko