On Tue, Aug 09, 2022 at 03:39:59PM -0700, Kees Cook wrote:
> On Tue, Aug 09, 2022 at 03:24:04PM -0700, Kees Cook wrote:
> > On Mon, Aug 08, 2022 at 03:15:04AM -0700, Guenter Roeck wrote:
> > > lm90_detect_nuvoton() is supposed to return NULL if it can not detect
> > > a chip, or a pointer to the chip name if it does. Under some circumstances
> > > it returns an error pointer instead. Some versions of gcc interpret an
> > > ERR_PTR as region of size 0 and generate an error message.
> > >
> > > In function ‘__fortify_strlen’,
> > > inlined from ‘strlcpy’ at ./include/linux/fortify-string.h:159:10,
> > > inlined from ‘lm90_detect’ at drivers/hwmon/lm90.c:2550:2:
> > > ./include/linux/fortify-string.h:50:33: error:
> > > ‘__builtin_strlen’ reading 1 or more bytes from a region of size 0
> > > 50 | #define __underlying_strlen __builtin_strlen
> > > | ^
> > > ./include/linux/fortify-string.h:141:24: note:
> > > in expansion of macro ‘__underlying_strlen’
> > > 141 | return __underlying_strlen(p);
> > > | ^~~~~~~~~~~~~~~~~~~
> > >
> > > Returning NULL instead of ERR_PTR() fixes the problem.
> > >
> > > Fixes: c7cebce984a2 ("hwmon: (lm90) Rework detect function")
> > > Reported-by: Ingo Molnar <mingo@xxxxxxxxxx>
> > > Cc: Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx>
> > > Cc: Kees Cook <keescook@xxxxxxxxxxxx>
>
> Reviewed-by: Kees Cook <keescook@xxxxxxxxxxxx>
>
> > > Signed-off-by: Guenter Roeck <linux@xxxxxxxxxxxx>
> > > ---
> > > It is interesting that some versions of gcc interpret an ERR_PTR this way.
> > > It did find a real bug, though the error message is quite confusing.
> > > Would it be possible to enhance the fortify functions to detect a constant
> > > ERR_PTR at compile time ? I think that might be quite useful.
> >
> > Yeah, that should be possible. I suspect something like this might work:
> >
> > BUILD_BUG_ON(__builtin_constant_p(src) && IS_ERR_VALUE(src));
> > BUILD_BUG_ON(__builtin_constant_p(dst) && IS_ERR_VALUE(dst));
> >
> > Though I'm not sure how it'd play with GCC value range checker.
>
> Yeah, looks like this doesn't work. These are all only able to check for
> a single value. The GCC diagnostics depend on its internal value range
> checking. This tripped because of the (sometimes buggy) "void * cast of
> a literal value is always a NULL pointer dereference, so its size must
> always be zero" which we've had to repeatedly work around. In this case,
> it was a real error, though. :P
>
Guess it would have been too easy. I wonder if it might be able to come up
with a coccinelle script to find such issues. Anyway, thanks for trying and
for the review.
Guenter
