Re: [PATCH v4 1/2] kernel.h: Introduce const_max() for VLA removal

[Miguel Ojeda]

On Fri, Mar 16, 2018 at 12:49 AM, Kees Cook <keescook@xxxxxxxxxxxx> wrote:
> On Thu, Mar 15, 2018 at 4:46 PM, Linus Torvalds
> <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
>> What I'm *not* so much ok with is "const_max(5,sizeof(x))" erroring
>> out, or silently causing insane behavior due to hidden subtle type
>> casts..
>
> Yup! I like it as an explicit argument. Thanks!
>

What about something like this?

#define INTMAXT_MAX LLONG_MAX
typedef int64_t intmax_t;

#define const_max(x, y)                                               \
        __builtin_choose_expr(                                        \
                !__builtin_constant_p(x) || !__builtin_constant_p(y), \
                __error_not_const_arg(),                              \
                __builtin_choose_expr(                                \
                        (x) > INTMAXT_MAX || (y) > INTMAXT_MAX,       \
                        __error_too_big(),                            \
                        __builtin_choose_expr(                        \
                                (intmax_t)(x) >= (intmax_t)(y),       \
                                (x),                                  \
                                (y)                                   \
                        )                                             \
                )                                                     \
        )

Works for different types, allows to mix negatives and positives and
returns the original type, e.g.:

  const_max(-1, sizeof(char));

is of type 'long unsigned int', but:

  const_max(2, sizeof(char));

is of type 'int'. While I am not a fan that the return type depends on
the arguments, it is useful if you are going to use the expression in
something that needs expects a precise (a printk() for instance?).

The check against the INTMAXT_MAX is there to avoid complexity (if we
do not handle those cases, it is safe to use intmax_t for the
comparison; otherwise you have to have another compile time branch for
the case positive-positive using uintmax_t) and also avoids odd
warnings for some cases above LLONG_MAX about comparisons with 0 for
unsigned expressions being always true. On the positive side, it
prevents using the macro for thing like "(size_t)-1".

Cheers,
Miguel