On Wed, Dec 19, 2012 at 11:21:15PM +0100, Jean Delvare wrote:
> Hi Guenter,
>
> On Wed, 19 Dec 2012 06:40:15 -0800, Guenter Roeck wrote:
> > Commit 263a523 fixes a warning seen with W=1 due to change in
> > DIV_ROUND_CLOSEST. Unfortunately, the C compiler converts divide operations
> > with unsigned divisors to unsigned, even if the dividend is signed and
> > negative (for example, -10 / 5U = 858993457). The C standard says "If one
> > operand has unsigned int type, the other operand is converted to unsigned
> > int", so the compiler is not to blame.
>
> This is surprising to say the least. But if the C standard says so...
>
Agreed, but it is how it is.
> I wouldn't be surprised if there are bugs because of this in the kernel
> and in other projects.
>
Might easily be. This might make a good interview question - I suspect many
if not most engineers would fail it. At least I would have until yesterday :).
> > As a result, DIV_ROUND_CLOSEST(0, 2U) and similar operations now return
> > bad values, since the automatic conversion of expressions such as "0 - 2U/2"
> > to unsigned was not taken into account.
> >
> > Fix by checking for the divisor variable type when deciding which operation
> > to perform. This fixes DIV_ROUND_CLOSEST(0, 2U), but still returns bad values
> > for negative dividends divided by unsigned divisors. Mark the latter case as
> > unsupported.
>
> True but this last issue isn't specific to the DIV_ROUND_CLOSEST
> implementation, it would also happen with a simple division.
>
Correct, which is why I did not try to fix it. Still worth mentioning, though,
in my opinion.
> > Reported-by: Juergen Beisert <jbe@xxxxxxxxxxxxxx>
> > Tested-by: Juergen Beisert <jbe@xxxxxxxxxxxxxx>
> > Cc: Jean Delvare <khali@xxxxxxxxxxxx>
> > Signed-off-by: Guenter Roeck <linux@xxxxxxxxxxxx>
> > ---
> > v2: Description update (v1 wasn't supposed to make it to lkml)
> >
> > include/linux/kernel.h | 6 ++++--
> > 1 file changed, 4 insertions(+), 2 deletions(-)
> >
> > diff --git a/include/linux/kernel.h b/include/linux/kernel.h
> > index d97ed58..45726dc 100644
> > --- a/include/linux/kernel.h
> > +++ b/include/linux/kernel.h
> > @@ -77,13 +77,15 @@
> >
> > /*
> > * Divide positive or negative dividend by positive divisor and round
> > - * to closest integer. Result is undefined for negative divisors.
> > + * to closest integer. Result is undefined for negative divisors and
> > + * for negative dividends if the divisor variable type is unsigned.
>
> Thinking a bit more about this... Documenting the non-working cases is
> great, however I don't really expect all developers to pay attention. I
> can also imagine variable types changing from signed to unsigned later,
> and never thinking this can introduce a bug.
>
> So, is there nothing we can do to spot at least the second issue at
> build time? For regular division there's nothing we can do (although I
> don't understand why gcc doesn't warn...) but here we get the
> opportunity to report the issue, let's take it.
>
> And given that the divisor is almost always a constant,
> maybe we can check for negative divisors too, this would be safer and
> the code size increase would probably be very small in practice.
> Opinions?
>
Agreed, though we should fix the problem now and think about reporting
afterwards.
Guenter
> > */
> > #define DIV_ROUND_CLOSEST(x, divisor)( \
> > { \
> > typeof(x) __x = x; \
> > typeof(divisor) __d = divisor; \
> > - (((typeof(x))-1) > 0 || (__x) > 0) ? \
> > + (((typeof(x))-1) > 0 || \
> > + ((typeof(divisor))-1) > 0 || (__x) > 0) ? \
> > (((__x) + ((__d) / 2)) / (__d)) : \
> > (((__x) - ((__d) / 2)) / (__d)); \
> > } \
>
> Looks good.
>
> --
> Jean Delvare
>
_______________________________________________
lm-sensors mailing list
lm-sensors@xxxxxxxxxxxxxx
http://lists.lm-sensors.org/mailman/listinfo/lm-sensors
