Re: [PATCH] string: Introduce strtomem() and strtomem_pad()

[Kees Cook <keescook@chromium.org>]

On Thu, Sep 01, 2022 at 09:14:26PM +0200, Geert Uytterhoeven wrote:
> Hi Kees,
> 
> On Thu, Sep 1, 2022 at 8:35 PM Kees Cook <keescook@chromium.org> wrote:
> > On Thu, Sep 01, 2022 at 10:39:19AM +0200, Geert Uytterhoeven wrote:
> > > > --- a/include/linux/string.h
> > > > +++ b/include/linux/string.h
> > > > @@ -260,6 +260,49 @@ static inline const char *kbasename(const char *path)
> > > >  void memcpy_and_pad(void *dest, size_t dest_len, const void *src, size_t count,
> > > >                     int pad);
> > > >
> > > > +/**
> > > > + * strtomem_pad - Copy NUL-terminated string to non-NUL-terminated buffer
> > > > + *
> > > > + * @dest: Pointer of destination character array (marked as __nonstring)
> > > > + * @src: Pointer to NUL-terminated string
> > > > + * @pad: Padding character to fill any remaining bytes of @dest after copy
> > > > + *
> > > > + * This is a replacement for strncpy() uses where the destination is not
> > > > + * a NUL-terminated string, but with bounds checking on the source size, and
> > > > + * an explicit padding character. If padding is not required, use strtomem().
> > > > + *
> > > > + * Note that the size of @dest is not an argument, as the length of @dest
> > > > + * must be discoverable by the compiler.
> > > > + */
> > > > +#define strtomem_pad(dest, src, pad)   do {                            \
> > > > +       const size_t _dest_len = __builtin_object_size(dest, 1);        \
> > > > +                                                                       \
> > > > +       BUILD_BUG_ON(!__builtin_constant_p(_dest_len) ||                \
> > > > +                    _dest_len == (size_t)-1);                          \
> > >
> > > I think you want to include __must_be_array(dest) here.
> >
> > I didn't do that for the cases where we may be writing to non-array
> > destinations (e.g. see the cast from u64 in the strncpy use in
> > tools/perf/arch/x86/util/intel-pt.c). Since what we need to know is the
> > object size, it does not strictly need to be an array.
> 
> IC.  That does mean we cannot catch silly mistakes where the
> caller passes a pointer instead of the address of an array?

It's okay to pass a pointer as long as the compiler can reason about the
size of the object being pointed at (which is what __bos() does here).

> > > > +       for (int i = 2; i < sizeof(wrap.output); i++)
> > >
> > > unsigned int i (everywhere)
> >
> > I guess, but why? This could even be u8.
> 
> sizeof() is unsigned, so using int may cause signed/unsigned comparison
> warnings.

Do we have those warnings enabled anywhere? I thought solving that
warning was Sisyphean. But I guess, yes, better to avoid adding yet
more. :)

-- 
Kees Cook