for_each_cpu() is widely used in the kernel, and it's beneficial to
create a NUMA-aware version of the macro.

Recently added for_each_numa_hop_mask() works, but switching existing
codebase to using it is not an easy process.

New for_each_numa_cpu() is designed to be similar to the for_each_cpu().
It allows to convert existing code to NUMA-aware as simple as adding a
hop iterator variable and passing it inside new macro. for_each_numa_cpu()
takes care of the rest.

At the moment, we have 2 users of NUMA-aware enumerators. One is
Melanox's in-tree driver, and another is Intel's in-review driver:

https://lore.kernel.org/lkml/20230216145455.661709-1-pawel.chmielewski@intel.com/

Both real-life examples follow the same pattern:

	for_each_numa_hop_mask(cpus, prev, node) {
 		for_each_cpu_andnot(cpu, cpus, prev) {
 			if (cnt++ == max_num)
 				goto out;
 			do_something(cpu);
 		}
		prev = cpus;
 	}

With the new macro, it would look like this:

	for_each_numa_cpu(cpu, hop, node, cpu_possible_mask) {
		if (cnt++ == max_num)
			break;
		do_something(cpu);
 	}

Straight conversion of existing for_each_cpu() codebase to NUMA-aware
version with for_each_numa_hop_mask() is difficult because it doesn't
take a user-provided cpu mask, and eventually ends up with open-coded
double loop. With for_each_numa_cpu() it shouldn't be a brainteaser.
Consider the NUMA-ignorant example:

	cpumask_t cpus = get_mask();
	int cnt = 0, cpu;

	for_each_cpu(cpu, cpus) {
		if (cnt++ == max_num)
			break;
		do_something(cpu);
 	}

Converting it to NUMA-aware version would be as simple as:

	cpumask_t cpus = get_mask();
	int node = get_node();
	int cnt = 0, hop, cpu;

	for_each_numa_cpu(cpu, hop, node, cpus) {
		if (cnt++ == max_num)
			break;
		do_something(cpu);
 	}

The latter looks more verbose and avoids from open-coding that annoying
double loop. Another advantage is that it works with a 'hop' parameter with
the clear meaning of NUMA distance, and doesn't make people not familiar
to enumerator internals bothering with current and previous masks machinery.

Signed-off-by: Yury Norov <yury.norov@gmail.com>
---
 include/linux/topology.h | 16 ++++++++++++++++
 1 file changed, 16 insertions(+)

diff --git a/include/linux/topology.h b/include/linux/topology.h
index 13209095d6e2..01fb3a55d7ce 100644
--- a/include/linux/topology.h
+++ b/include/linux/topology.h
@@ -286,4 +286,20 @@ sched_numa_hop_mask(unsigned int node, unsigned int hops)
 	     !IS_ERR_OR_NULL(mask);					       \
 	     __hops++)
 
+/**
+ * for_each_numa_cpu - iterate over cpus in increasing order taking into account
+ *		       NUMA distances from a given node.
+ * @cpu: the (optionally unsigned) integer iterator
+ * @hop: the iterator variable, must be initialized to a desired minimal hop.
+ * @node: the NUMA node to start the search from.
+ * @mask: the cpumask pointer
+ *
+ * Requires rcu_lock to be held.
+ */
+#define for_each_numa_cpu(cpu, hop, node, mask)					\
+	for ((cpu) = 0, (hop) = 0;						\
+		(cpu) = sched_numa_find_next_cpu((mask), (cpu), (node), &(hop)),\
+		(cpu) < nr_cpu_ids;						\
+		(cpu)++)
+
 #endif /* _LINUX_TOPOLOGY_H */
-- 
2.34.1

On 19/04/23 22:19, Yury Norov wrote:
> +/**
> + * for_each_numa_cpu - iterate over cpus in increasing order taking into account
> + *		       NUMA distances from a given node.
> + * @cpu: the (optionally unsigned) integer iterator
> + * @hop: the iterator variable, must be initialized to a desired minimal hop.
> + * @node: the NUMA node to start the search from.
> + * @mask: the cpumask pointer
> + *
> + * Requires rcu_lock to be held.
> + */
> +#define for_each_numa_cpu(cpu, hop, node, mask)					\
> +	for ((cpu) = 0, (hop) = 0;						\
> +		(cpu) = sched_numa_find_next_cpu((mask), (cpu), (node), &(hop)),\
> +		(cpu) < nr_cpu_ids;						\
> +		(cpu)++)
> +

I think we can keep sched_numa_find_next_cpu() as-is, but could we make
that macro use cpu_possible_mask by default? We can always add a variant
if/when we need to feed in a different mask.

>  #endif /* _LINUX_TOPOLOGY_H */
> -- 
> 2.34.1

On Tue, Apr 25, 2023 at 10:54:48AM +0100, Valentin Schneider wrote:
> On 19/04/23 22:19, Yury Norov wrote:
> > +/**
> > + * for_each_numa_cpu - iterate over cpus in increasing order taking into account
> > + *		       NUMA distances from a given node.
> > + * @cpu: the (optionally unsigned) integer iterator
> > + * @hop: the iterator variable, must be initialized to a desired minimal hop.
> > + * @node: the NUMA node to start the search from.
> > + * @mask: the cpumask pointer
> > + *
> > + * Requires rcu_lock to be held.
> > + */
> > +#define for_each_numa_cpu(cpu, hop, node, mask)					\
> > +	for ((cpu) = 0, (hop) = 0;						\
> > +		(cpu) = sched_numa_find_next_cpu((mask), (cpu), (node), &(hop)),\
> > +		(cpu) < nr_cpu_ids;						\
> > +		(cpu)++)
> > +
> 
> I think we can keep sched_numa_find_next_cpu() as-is, but could we make
> that macro use cpu_possible_mask by default? We can always add a variant
> if/when we need to feed in a different mask.

As mentioned in discussion to the driver's patch, all that numa things
imply only online CPUs, so cpu_possible_mask may mislead to some extent. 

Anyways, can you elaborate what you exactly want? Like this?

 #define for_each_numa_online_cpu(cpu, hop, node)       \
        for_each_numa_cpu(cpu, hop, node, cpu_online_mask)

On 25/04/23 22:32, Yury Norov wrote:
> On Tue, Apr 25, 2023 at 10:54:48AM +0100, Valentin Schneider wrote:
>> On 19/04/23 22:19, Yury Norov wrote:
>> > +/**
>> > + * for_each_numa_cpu - iterate over cpus in increasing order taking into account
>> > + *		       NUMA distances from a given node.
>> > + * @cpu: the (optionally unsigned) integer iterator
>> > + * @hop: the iterator variable, must be initialized to a desired minimal hop.
>> > + * @node: the NUMA node to start the search from.
>> > + * @mask: the cpumask pointer
>> > + *
>> > + * Requires rcu_lock to be held.
>> > + */
>> > +#define for_each_numa_cpu(cpu, hop, node, mask)					\
>> > +	for ((cpu) = 0, (hop) = 0;						\
>> > +		(cpu) = sched_numa_find_next_cpu((mask), (cpu), (node), &(hop)),\
>> > +		(cpu) < nr_cpu_ids;						\
>> > +		(cpu)++)
>> > +
>> 
>> I think we can keep sched_numa_find_next_cpu() as-is, but could we make
>> that macro use cpu_possible_mask by default? We can always add a variant
>> if/when we need to feed in a different mask.
>
> As mentioned in discussion to the driver's patch, all that numa things
> imply only online CPUs, so cpu_possible_mask may mislead to some extent. 
>
> Anyways, can you elaborate what you exactly want? Like this?
>
>  #define for_each_numa_online_cpu(cpu, hop, node)       \
>         for_each_numa_cpu(cpu, hop, node, cpu_online_mask)

Yeah, something like that. Like you said, the NUMA cpumasks built by the
scheduler reflect the online topology, so s/possible/online/ shouldn't
change much here.