Re: [PATCH v3 04/10] iio: afe: rescale: reduce risk of integer overflow

From: Liam Beguin
Date: Mon Jul 05 2021 - 00:24:14 EST


On Sun Jul 4, 2021 at 12:36 PM EDT, Jonathan Cameron wrote:
> On Wed, 30 Jun 2021 21:00:28 -0400
> Liam Beguin <liambeguin@xxxxxxxxx> wrote:
>
> > From: Liam Beguin <lvb@xxxxxxxxxx>
> >
> > Reduce the risk of integer overflow by doing the scale calculation with
> > 64bit integers and looking for a Greatest Common Divider for both parts
> > of the fractional value.
> >
> > Signed-off-by: Liam Beguin <lvb@xxxxxxxxxx>
> > ---
> > drivers/iio/afe/iio-rescale.c | 10 +++++++---
> > 1 file changed, 7 insertions(+), 3 deletions(-)
> >
> > diff --git a/drivers/iio/afe/iio-rescale.c b/drivers/iio/afe/iio-rescale.c
> > index 774eb3044edd..98bcb5d418d6 100644
> > --- a/drivers/iio/afe/iio-rescale.c
> > +++ b/drivers/iio/afe/iio-rescale.c
> > @@ -39,7 +39,8 @@ static int rescale_read_raw(struct iio_dev *indio_dev,
> > int *val, int *val2, long mask)
> > {
> > struct rescale *rescale = iio_priv(indio_dev);
> > - unsigned long long tmp;
> > + s64 tmp, tmp2;
> > + u32 factor;
> > int ret;
> >
> > switch (mask) {
> > @@ -67,8 +68,11 @@ static int rescale_read_raw(struct iio_dev *indio_dev,
> > }
> > switch (ret) {
> > case IIO_VAL_FRACTIONAL:
> > - *val *= rescale->numerator;
> > - *val2 *= rescale->denominator;
> > + tmp = (s64)*val * rescale->numerator;
> > + tmp2 = (s64)*val2 * rescale->denominator;
> > + factor = gcd(tmp, tmp2);
>
> Hmm. I wonder if there are cases where this doesn't work and we end up
> truncating because the gcd is say 1. If all of val, val2,
> rescale->numerator,
> rescale->denominator are primes and the rescale values are moderately
> large
> then that might happen. We probably need a fallback position. Perhaps
> check tmp / factor and temp2/factor will fit in an int. If not, shift
> them until
> they do even if we have to dump some precision to do so.
>

I see what you mean. If we want to do that I guess it would also apply
to other areas of the driver.

> This stuff is getting fiddly enough we might want to figure out some
> self tests
> that exercise the various cases.
>

I never implemented kernel self tests before, I guess it should follow
the example of drivers/iio/test/iio-test-format.c?

Would you be okay to add this in a follow up series?

> > + *val = tmp / factor;
> > + *val2 = tmp2 / factor;
>
> This is doing 64 bit numbers divided by 32 bit ones. Doesn't that
> require
> use of do_div() etc on 32 bit platforms?
>

Apologies for that mistake, will fix.

> > return ret;
> > case IIO_VAL_INT:
> > *val *= rescale->numerator;