Re: [PATCH 2/6] iio: proximity: sx9324: fix proximity threshold writes

From: Andy Shevchenko

Date: Wed Nov 18 2026 - 02:01:25 EST


On Wed, Nov 18, 2026 at 12:40:29AM +0530, Arnav Kapoor wrote:
> On Tue, Nov 17, 2026 at 08:59:06PM +0200, Andy Shevchenko wrote:
> > > For odd r the read value is (r * r - 1) / 2,
> >
> > No, it's r²/2 but in integer arithmetics. That's why for odd number squares
> > it gives 1/2 that becomes 0.
>
> Same value, but your wording is clearer. I'll say in v2 that r * r / 2
> rounds down for odd r.
>
> > Yep, what we see is off-by-one for full squares. This is because division
> > by 2 in integer arithmetics makes it floor number. So, what you want
> > to have is a correction for all odd numbers.
>
> > And here I would rather use (val % 2) instead of 1. See above why.
>
> I don't think val % 2 works, because the correction depends on whether
> the register code r is odd, not whether val is.

Ah, yes, you are right. I was thinking about it, but got sideways and wrote
that.

> Odd codes read back as even values too, as in your table:
>
> reg read value int_sqrt(2 * val + val % 2) int_sqrt(2 * val + 1)
> 3 4 int_sqrt(8) = 2 int_sqrt(9) = 3
> 5 12 int_sqrt(24) = 4 int_sqrt(25) = 5
> 7 24 int_sqrt(48) = 6 int_sqrt(49) = 7
>
> With + 1: an odd r reads back as (r * r - 1) / 2,

Nope, this reads as usual, the / 2 makes it floor because of integer arithmetics.
While mathematically (in integer arithmetics) it gives the same result the
reasoning is wrong.

> so 2 * val + 1 is
> exactly r * r.

>From value point of view, yes for the full squares, but it's wrong for other
cases.

> An even r reads back as r * r / 2, so 2 * val + 1 is
> r * r + 1, and int_sqrt() still gives r. Checking all 256 codes,

For the sake of the correctness the whole set of 64k has to be checked. And in
most cases the + 1 there will be wrong. The only correction we need is on the
edges of the full squares, id est when val == r² - 1.

> + 1 round-trips every one of them, while val % 2 still fails on the same
> 127 codes as the current int_sqrt(2 * val). For values in between,
> + 1 also picks the largest code whose threshold does not exceed the
> requested value.

At bare minimum the commit message needs to be corrected and explained
these aspects that while the + 1 gives wrong result for most of the cases,
it's still okay to use as we work only in integer arithmetics. Also the formula
(r² - 1) / 2 has to be dropped as it's incorrect from mathematics point of
view. The floor() happens on / 2 and has nothing to do with (- 1) correction.

> > For the write part (the negative is problematic only on 32-bit case IIUC)
>
> I think it is on 64-bit too: val is an unsigned int, so 2 * val wraps
> at 32 bits before int_sqrt() sees it, whatever the size of long.

OK, thanks for checking.

> > > reg = SX9324_REG_PROX_CTRL6 + chan->channel / 2;
> >
> > This one is not used in the conditionals, so the new check may well be done
> > before it.
>
> > > + if (_val < 0)
> >
> > Checking _val after assigning val to it is confusing.
>
> > > if (val >= 1)
> >
> > First of all, I would drop the '=' here.
>
> Agreed on all three. In v2 the check on _val will come first, before
> val and reg are assigned, and the condition will be val > 1. I'll make
> the same changes in 5/6, which has the same code for sx9360.

--
With Best Regards,
Andy Shevchenko