Re: [PATCH 06/14] bitsperlong.h: introduce SMALL_CONST() macro

[Rasmus Villemoes]

On 18/02/2021 05.05, Yury Norov wrote:
> Many algorithms become simpler if they are passed with relatively small
> input values. One example is bitmap operations when the whole bitmap fits
> into one word. To implement such simplifications, linux/bitmap.h declares
> small_const_nbits() macro.
> 
> Other subsystems may also benefit from optimizations of this sort, like
> find_bit API in the following patches. So it looks helpful to generalize
> the macro and extend it's visibility.

Perhaps, but SMALL_CONST is too generic a name, it needs to keep "bits"
somewhere in there. So why not just keep it at small_const_nbits?

> Signed-off-by: Yury Norov <yury.norov@xxxxxxxxx>
> ---
>  include/asm-generic/bitsperlong.h |  2 ++
>  include/linux/bitmap.h            | 33 ++++++++++++++-----------------
>  2 files changed, 17 insertions(+), 18 deletions(-)
> 
> diff --git a/include/asm-generic/bitsperlong.h b/include/asm-generic/bitsperlong.h
> index 3905c1c93dc2..0eeb77544f1d 100644
> --- a/include/asm-generic/bitsperlong.h
> +++ b/include/asm-generic/bitsperlong.h
> @@ -23,4 +23,6 @@
>  #define BITS_PER_LONG_LONG 64
>  #endif
>  
> +#define SMALL_CONST(n) (__builtin_constant_p(n) && (unsigned long)(n) < BITS_PER_LONG)
> +
>  #endif /* __ASM_GENERIC_BITS_PER_LONG */
> diff --git a/include/linux/bitmap.h b/include/linux/bitmap.h
> index adf7bd9f0467..e89f1dace846 100644
> --- a/include/linux/bitmap.h
> +++ b/include/linux/bitmap.h
> @@ -224,9 +224,6 @@ extern int bitmap_print_to_pagebuf(bool list, char *buf,
>   * so make such users (should any ever turn up) call the out-of-line
>   * versions.
>   */
> -#define small_const_nbits(nbits) \
> -	(__builtin_constant_p(nbits) && (nbits) <= BITS_PER_LONG && (nbits) > 0)
> -
>  static inline void bitmap_zero(unsigned long *dst, unsigned int nbits)
>  {
>  	unsigned int len = BITS_TO_LONGS(nbits) * sizeof(unsigned long);
> @@ -278,7 +275,7 @@ extern void bitmap_to_arr32(u32 *buf, const unsigned long *bitmap,
>  static inline int bitmap_and(unsigned long *dst, const unsigned long *src1,
>  			const unsigned long *src2, unsigned int nbits)
>  {
> -	if (small_const_nbits(nbits))
> +	if (SMALL_CONST(nbits - 1))

Please don't force most users to be changed to something less readable.
What's wrong with just keeping small_const_nbits() the way it is,
avoiding all this churn and keeping the readability?

At a quick reading, one of the very few places where you end up not
passing nbits-1 but just nbits is this

 unsigned long find_next_zero_bit_le(const void *addr, unsigned
 		long size, unsigned long offset)
 {
+	if (SMALL_CONST(size)) {
+		unsigned long val = *(const unsigned long *)addr;
+
+		if (unlikely(offset >= size))
+			return size;

which is a regression, for much the same reason the nbits==0 case was
excluded from small_const_nbits in the first place. If size is 0, we
used to just return 0 early in _find_next_bit. But you've introduced a
dereference of addr before that check is now done, which is
theoretically an oops.

If find_next_zero_bit_le cannot handle nbits==BITS_PER_LONG efficiently
but requires one off-limits bit position, fine, so be it, add an extra
"small_const_nbits() && nbits < BITS_PER_LONG" (and a comment).

Rasmus