Re: Optimisation for smp_num_cpus loop in hotplug

[Rusty Russell <rusty@rustcorp.com.au>]

In message <200206211531.g5LFViZ07396@localhost.localdomain> you write:
> rusty@rustcorp.com.au said:
> > Yeah, it's simple, and none of the current ones are really critical.
> > But I think we're better off with:
> > 	for (i = first_cpu(); i < NR_CPUS; i = next_cpu(i)) {
> 
> > Which is simple enough not to need an iterator macro, and also has the
> > bonus of giving irq-balancing et al. an efficient, portable way of
> > looking for the "next" cpu. 
> 
> So you're thinking that next_cpu(i) is something like
> 
> __ffs((~(unsigned)((1<<i)-1) & cpu_online_map)
> 
> plus an extra exception piece to take next_cpu(i) above NR_CPUS if we have no
 
> remaining CPUs (because __ffs would be undefined)?  It's the exception piece 
> that I don't see how to do really efficiently.

find_next_bit already does this, but the generic one would look
something like:

	unsigned long mask = ~(unsigned long)((1<<(cpu+1))-1);
	if (mask & cpu_online_map)
		return _ffs(mask & cpu_online_map);
	return NR_CPUS;

Cheers!
Rusty.
--
  Anyone who quotes me in their sig is an idiot. -- Rusty Russell.