Re: [PATCH treewide v2 1/3] bitfield: Add non-constant field_{prep,get}() helpers

[Yury Norov <yury.norov@xxxxxxxxx>]

On Tue, Feb 04, 2025 at 12:41:55AM +0900, Vincent Mailhol wrote:
> On 03/02/2025 at 22:59, Geert Uytterhoeven wrote:
> > Hi Vincent,
> > 
> > On Mon, 3 Feb 2025 at 14:37, Vincent Mailhol <mailhol.vincent@xxxxxxxxxx> wrote:
> >> On 03/02/2025 at 16:44, Johannes Berg wrote:
> >>> On Sun, 2025-02-02 at 12:53 -0500, Yury Norov wrote:
> >>>>> Instead of creating another variant for
> >>>>> non-constant bitfields, wouldn't it be better to make the existing macro
> >>>>> accept both?
> >>>>
> >>>> Yes, it would definitely be better IMO.
> >>>
> >>> On the flip side, there have been discussions in the past (though I
> >>> think not all, if any, on the list(s)) about the argument order. Since
> >>> the value is typically not a constant, requiring the mask to be a
> >>> constant has ensured that the argument order isn't as easily mixed up as
> >>> otherwise.
> >>
> >> If this is a concern, then it can be checked with:
> >>
> >>   BUILD_BUG_ON_MSG(!__builtin_constant_p(_mask) &&
> >>                    __builtin_constant_p(_val),
> >>                    _pfx "mask is not constant");
> >>
> >> It means that we forbid FIELD_PREP(non_const_mask, const_val) but allow
> >> any other combination.
> > 
> > Even that case looks valid to me. Actually there is already such a user
> > in drivers/iio/temperature/mlx90614.c:
> > 
> >     ret |= field_prep(chip_info->fir_config_mask, MLX90614_CONST_FIR);
> > 
> > So if you want enhanced safety, having both the safer/const upper-case
> > variants and the less-safe/non-const lower-case variants makes sense.

I agree with that. I just don't want the same shift-and operation to be
opencoded again and again.

What I actually meant is that I'm OK with whatever number of field_prep()
macro flavors, if we make sure that they don't duplicate each other. So
for me, something like this would be the best solution:

 #define field_prep(mask, val) \
       (((typeof(_mask))(_val) << __bf_shf(_mask)) & (_mask))

 #define FIELD_PREP(mask, val)                                         \
         (                                                             \
                 FIELD_PREP_INPUT_CHECK(_mask, _val,);                 \
                 field_prep(mask, val);                                \
         )
 
#define FIELD_PREP_CONST(_mask, _val)                                  \
        (                                                              \
                FIELD_PREP_CONST_INPUT_CHECK(mask, val);
                FIELD_PREP(mask, val); // or field_prep()
        )

We have a similar macro GENMASK() in linux/bits.h. It is implemented
like this:

 #define GENMASK_INPUT_CHECK(h, l) BUILD_BUG_ON_ZERO(const_true((l) > (h)))
 #define GENMASK(h, l) \
         (GENMASK_INPUT_CHECK(h, l) + __GENMASK(h, l))

And it works just well. Can we end up with a similar approach here?

> So, we are scared of people calling FIELD_PREP() with the arguments in
> the wrong order:
>
>   FIELD_PREP(val, mask)
> 
> thus adding the check that mask must be a compile time constant.

Don't be scared. Kernel coding implies that people get used to read
function declarations and comments on top of them before using
something.

Thansk,
Yury