Re: [PATCH v3 1/3] bitmap: generalize node_random()

[Yury Norov <yury.norov@gmail.com>]

On Tue, Jun 17, 2025 at 03:50:56PM -0700, Andrew Morton wrote:
> On Tue, 17 Jun 2025 16:08:51 -0400 Yury Norov <yury.norov@gmail.com> wrote:
> 
> > From: "Yury Norov [NVIDIA]" <yury.norov@gmail.com>
> > 
> > Generalize node_random() and make it available to general bitmaps and
> > cpumasks users.
> > 
> > Notice, find_first_bit() is generally faster than find_nth_bit(), and we
> > employ it when there's a single set bit in the bitmap.
> > 
> > See commit 3e061d924fe9c7b4 ("lib/nodemask: optimize node_random for
> > nodemask with single NUMA node").
> > 
> > ...
> >
> > --- a/include/linux/nodemask.h
> > +++ b/include/linux/nodemask.h
> > @@ -492,21 +492,7 @@ static __always_inline int num_node_state(enum node_states state)
> >  static __always_inline int node_random(const nodemask_t *maskp)
> >  {
> >  #if defined(CONFIG_NUMA) && (MAX_NUMNODES > 1)
> > -	int w, bit;
> > -
> > -	w = nodes_weight(*maskp);
> > -	switch (w) {
> > -	case 0:
> > -		bit = NUMA_NO_NODE;
> 
> If the mask has no bits set, return -1.

...

> If the mask has no bits set, return the mask's size.
> 
> > +	case 1:
> > +		/* Performance trick for single-bit bitmaps */
> > +		return find_first_bit(addr, size);
> > +	default:
> > +		return find_nth_bit(addr, size, get_random_u32_below(w));
> > +	}
> > +}
> 
> I'm not seeing how this is correct?

Ahh... Indeed you're right.

Thanks, Andrew, I'll send v4