On Tue, Sep 14, 2021 at 11:14 AM Linus Torvalds
[off-list ref] wrote:

All this pain could have been trivially avoided with just people
writing better code, knowing that multiplies and divides are
expensive, and that shift counts are small and cheap.

IOW, maybe the fix is just this attached trivial patch.

I didn't bother to change the order of the 'struct ndb_config'
structure. It would pack better if you put the (now 32-bit)
blksize_bits field next to the 'atomic_t' field, I think. But I wanted
to just see how a minimal patch looks.

I did make the debugfs interface reflect the change to blocksize_bits,
so this is visible in user space. But it's debugfs.

If people care, it could be made to use a DEFINE_SHOW_ATTRIBUTE()
function the way it already does for 'flags', so that's not a
fundamental issue, I just didn't bother.

Hmm?

Btw, I really think more of the block layer should perhaps think about
use bit shifts more, not expanded values.  Can things like the queue
'discard_alignment' really be non-powers-of-two?

          Linus