Re: [PATCH v3 4/8] i2c: ov9650: use 64-bit arithmetic instead of 32-bit

[Gustavo A. R. Silva]

On 02/15/2018 07:52 AM, Hans Verkuil wrote:
> On 08/02/18 17:39, Gustavo A. R. Silva wrote:
> > Hi Sakari,
> >
> > On 02/07/2018 03:59 PM, Sakari Ailus wrote:
> > > Hi Gustavo,
> > >
> > > On Tue, Feb 06, 2018 at 10:47:50AM -0600, Gustavo A. R. Silva wrote:
> > > > Add suffix ULL to constants 10000 and 1000000 in order to give the
> > > > compiler complete information about the proper arithmetic to use.
> > > > Notice that these constants are used in contexts that expect
> > > > expressions of type u64 (64 bits, unsigned).
> > > >
> > > > The following expressions:
> > > >
> > > > (u64)(fi->interval.numerator * 10000)
> > > > (u64)(iv->interval.numerator * 10000)
> > > > fiv->interval.numerator * 1000000 / fiv->interval.denominator
> > > >
> > > > are currently being evaluated using 32-bit arithmetic.
> > > >
> > > > Notice that those casts to u64 for the first two expressions are only
> > > > effective after such expressions are evaluated using 32-bit arithmetic,
> > > > which leads to potential integer overflows. So based on those casts, it
> > > > seems that the original intention of the code is to actually use 64-bit
> > > > arithmetic instead of 32-bit.
> > > >
> > > > Also, notice that once the suffix ULL is added to the constants, the
> > > > outer casts to u64 are no longer needed.
> > > >
> > > > Addresses-Coverity-ID: 1324146 ("Unintentional integer overflow")
> > > > Fixes: 84a15ded76ec ("[media] V4L: Add driver for OV9650/52 image sensors")
> > > > Fixes: 79211c8ed19c ("remove abs64()")
> > > > Signed-off-by: Gustavo A. R. Silva <gustavo@xxxxxxxxxxxxxx>
> > > > ---
> > > > Changes in v2:
> > > >  Â - Update subject and changelog to better reflect the proposed code changes.
> > > >  Â - Add suffix ULL to constants instead of casting variables.
> > > >  Â - Remove unnecessary casts to u64 as part of the code change.
> > > >  Â - Extend the same code change to other similar expressions.
> > > >
> > > > Changes in v3:
> > > >  Â - None.
> > > >
> > > >  Â drivers/media/i2c/ov9650.c | 9 +++++----
> > > >  Â 1 file changed, 5 insertions(+), 4 deletions(-)
> > > >
> > > > diff --git a/drivers/media/i2c/ov9650.c b/drivers/media/i2c/ov9650.c
> > > > index e519f27..e716e98 100644
> > > > --- a/drivers/media/i2c/ov9650.c
> > > > +++ b/drivers/media/i2c/ov9650.c
> > > > @@ -1130,7 +1130,7 @@ static int __ov965x_set_frame_interval(struct ov965x *ov965x,
> > > >  ÂÂÂÂÂ if (fi->interval.denominator == 0)
> > > >  ÂÂÂÂÂÂÂÂÂ return -EINVAL;
> > > >  Â -ÂÂÂ req_int = (u64)(fi->interval.numerator * 10000) /
> > > > +ÂÂÂ req_int = fi->interval.numerator * 10000ULL /
> > > >  ÂÂÂÂÂÂÂÂÂ fi->interval.denominator;
> > >
> > > This has been addressed by your earlier patch "i2c: ov9650: fix potential integer overflow in
> > > __ov965x_set_frame_interval" I tweaked a little. It's not in media tree
> > > master yet.
> >
> > Yeah. Actually this patch is supposed to be an improved version of the one you mention. That is why this is version 3.
> >
> > Also, I wonder if the same issue you mention below regarding 32-bit ARM applies in this case too?
> >
> > > >  Â ÂÂÂÂÂ for (i = 0; i < ARRAY_SIZE(ov965x_intervals); i++) {
> > > > @@ -1139,7 +1139,7 @@ static int __ov965x_set_frame_interval(struct ov965x *ov965x,
> > > >  ÂÂÂÂÂÂÂÂÂ if (mbus_fmt->width != iv->size.width ||
> > > >  ÂÂÂÂÂÂÂÂÂÂÂÂÂ mbus_fmt->height != iv->size.height)
> > > >  ÂÂÂÂÂÂÂÂÂÂÂÂÂ continue;
> > > > -ÂÂÂÂÂÂÂ err = abs((u64)(iv->interval.numerator * 10000) /
> > > > +ÂÂÂÂÂÂÂ err = abs(iv->interval.numerator * 10000ULL /
> > >
> > > This and the chunk below won't work on e.g. 32-bit ARM. do_div(), please.
> >
> > Thanks for pointing this out.
> >
> > > >  ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ iv->interval.denominator - req_int);
> > > >  ÂÂÂÂÂÂÂÂÂ if (err < min_err) {
> > > >  ÂÂÂÂÂÂÂÂÂÂÂÂÂ fiv = iv;
> > > > @@ -1148,8 +1148,9 @@ static int __ov965x_set_frame_interval(struct ov965x *ov965x,
> > > >  ÂÂÂÂÂ }
> > > >  ÂÂÂÂÂ ov965x->fiv = fiv;
> > > >  Â -ÂÂÂ v4l2_dbg(1, debug, &ov965x->sd, "Changed frame interval to %u us\n",
> > > > -ÂÂÂÂÂÂÂÂ fiv->interval.numerator * 1000000 / fiv->interval.denominator);
> > > > +ÂÂÂ v4l2_dbg(1, debug, &ov965x->sd, "Changed frame interval to %llu us\n",
> > > > +ÂÂÂÂÂÂÂÂ fiv->interval.numerator * 1000000ULL /
> > > > +ÂÂÂÂÂÂÂÂ fiv->interval.denominator);
> >
> > I wonder if do_div should be used for the code above?
>
> Yes, do_div should be used.

I got it.

Thanks, Hans.
--
Gustavo