On Wed, Dec 15, 2021 at 12:41 AM David Laight [off-list ref] wrote:

From: Yury Norov

quoted

Sent: 14 December 2021 19:43

...

quoted

I think that for long bitmaps the most time consuming operation is moving
data to L1, and for short bitmaps the difference between approaches is
barely measurable.

But hweght_long on each iteration can't be more effective than the current
version. So, I'll drop this patch for v2 and keep things unchanged.

Actually do bitmap_full/empty() calls make any sense at all?
The result is stale since bitmaps are designed to do locked operations.
If you have a lock covering the bitmap then you should be using
something that uses non-locked accesses.
Rightly or wrongly that isn't the bitmap api.

Are you talking about __{set,clear}_bit()?
include/asm-generic/bitops/non-atomic.h