Re: [RFC PATCH v2 10/23] sched/topology: Add per-sd scratch for the load balance affinity score

[Peter Zijlstra]

On Thu, Aug 27, 2026 at 08:28:03PM +0800, Jianyong Wu wrote:
> The affinity score computed during load balancing needs two small arrays,
> one entry per LLC: the LLCs that sit closer to the balance destination than
> to the source, and the weight of each. Their size is only known once the
> sched domains have been built, which is why the load balancer currently has
> to allocate them on every call.
> 
> Signed-off-by: Jianyong Wu <wujianyong@xxxxxxxx>
> ---
>  include/linux/sched/topology.h |  6 ++++++
>  kernel/sched/topology.c        | 28 +++++++++++++++++++++++++++-
>  2 files changed, 33 insertions(+), 1 deletion(-)
> 
> diff --git a/include/linux/sched/topology.h b/include/linux/sched/topology.h
> index b31d2cf16592..569ff562dbdb 100644
> --- a/include/linux/sched/topology.h
> +++ b/include/linux/sched/topology.h
> @@ -122,6 +122,12 @@ struct sched_domain {
>  	unsigned int *llc_counts __counted_by_ptr(llc_max);
>  	unsigned long llc_bytes;
>  	unsigned int *numa_counts;
> +	/*
> +	 * Scratch for the load balance affinity score, llc_max entries each.
> +	 * Only the bottom sd owns them, like llc_counts above.
> +	 */
> +	int *affi_ids;
> +	int *affi_weights;

Per the below, these should also have: __counted_by_ptr(llc_max), right?

>  #endif

> +		p_ids = kcalloc_node(max_lid + 1, sizeof(int),
> +				 GFP_KERNEL, cpu_to_node(i));
> +		p_w = kcalloc_node(max_lid + 1, sizeof(int),
> +				 GFP_KERNEL, cpu_to_node(i));