[PATCH] percpu: reduce the time complexity of mapping units and cpus

[PATCH] percpu: reduce the time complexity of mapping units and cpus

[Hyunmin Lee]

For mapping units and CPUs belonging to groups, it can be inefficient to
iterate through all CPUs by each group to find what CPUs belong to that
group.

Since group_map already has the information on which CPUs belong to which
group, CPUs can be directly mapped to a unit in the group to which the CPU
belongs.

Signed-off-by: Hyunmin Lee <hyunminrlee@gmail.com>
---
 mm/percpu.c | 10 +++++++---
 1 file changed, 7 insertions(+), 3 deletions(-)

diff --git a/mm/percpu.c b/mm/percpu.c
index b35494c8ede2..968aa0ace482 100644
--- a/mm/percpu.c
+++ b/mm/percpu.c
@@ -2906,6 +2906,13 @@ static struct pcpu_alloc_info * __init __flatten pcpu_build_alloc_info(
 	ai->atom_size = atom_size;
 	ai->alloc_size = alloc_size;
 
+	for_each_possible_cpu(cpu) {
+		group = group_map[cpu];
+		struct pcpu_group_info *gi = &ai->groups[group];
+
+		gi->cpu_map[gi->nr_units++] = cpu;
+	}
+
 	for (group = 0, unit = 0; group < nr_groups; group++) {
 		struct pcpu_group_info *gi = &ai->groups[group];
 
@@ -2916,9 +2923,6 @@ static struct pcpu_alloc_info * __init __flatten pcpu_build_alloc_info(
 		 */
 		gi->base_offset = unit * ai->unit_size;
 
-		for_each_possible_cpu(cpu)
-			if (group_map[cpu] == group)
-				gi->cpu_map[gi->nr_units++] = cpu;
 		gi->nr_units = roundup(gi->nr_units, upa);
 		unit += gi->nr_units;
 	}
-- 
2.43.0

Re: [PATCH] percpu: reduce the time complexity of mapping units and cpus

[Dennis Zhou]

Hello,

On Fri, Jun 13, 2025 at 01:21:38PM +0900, Hyunmin Lee wrote:
> For mapping units and CPUs belonging to groups, it can be inefficient to
> iterate through all CPUs by each group to find what CPUs belong to that
> group.
> 
> Since group_map already has the information on which CPUs belong to which
> group, CPUs can be directly mapped to a unit in the group to which the CPU
> belongs.
> 

Idk. For any single socket machine, it's even money. For large machines
with say 4+ numa nodes, you're not really going to notice the
microseconds the extra for loop is going to take.

I'm neutral but inclined to prefer what's already there.

> Signed-off-by: Hyunmin Lee <hyunminrlee@gmail.com>
> ---
>  mm/percpu.c | 10 +++++++---
>  1 file changed, 7 insertions(+), 3 deletions(-)
> 
> diff --git a/mm/percpu.c b/mm/percpu.c
> index b35494c8ede2..968aa0ace482 100644
> --- a/mm/percpu.c
> +++ b/mm/percpu.c
> @@ -2906,6 +2906,13 @@ static struct pcpu_alloc_info * __init __flatten pcpu_build_alloc_info(
>  	ai->atom_size = atom_size;
>  	ai->alloc_size = alloc_size;
>  
> +	for_each_possible_cpu(cpu) {
> +		group = group_map[cpu];
> +		struct pcpu_group_info *gi = &ai->groups[group];
> +
> +		gi->cpu_map[gi->nr_units++] = cpu;
> +	}
> +

Nit: the struct declaration goes at the top in the block.

for_each_possible_cpu(cpu) {
    struct pcpu_group_info *gi;

    group = group_map[cpu];
    gi = &ai->groups[group];

    gi->cpu_map[gi->nr_units++] = cpu;
}

Or condense it into a single line:

for_each_possible_cpu(cpu) {
    struct pcpu_group_info *gi = &ai->groups[group_map[cpu]];

    gi->cpu_map[gi->nr_units++] = cpu;
}

>  	for (group = 0, unit = 0; group < nr_groups; group++) {
>  		struct pcpu_group_info *gi = &ai->groups[group];
>  
> @@ -2916,9 +2923,6 @@ static struct pcpu_alloc_info * __init __flatten pcpu_build_alloc_info(
>  		 */
>  		gi->base_offset = unit * ai->unit_size;
>  
> -		for_each_possible_cpu(cpu)
> -			if (group_map[cpu] == group)
> -				gi->cpu_map[gi->nr_units++] = cpu;
>  		gi->nr_units = roundup(gi->nr_units, upa);
>  		unit += gi->nr_units;
>  	}
> -- 
> 2.43.0
> 

Thanks,
Dennis