Re: [PATCH] sched/core: Make fallback CPU selection NUMA-aware

[Yury Norov <ynorov@nvidia.com>]

Ping?

On Fri, Sep 04, 2026 at 11:36:56PM -0400, Yury Norov wrote:
> select_fallback_rq() checks the local node first, but then scans the task's
> affinity mask in a numerical order. On systems with more than two NUMA
> nodes, that can select a CPU farther away than necessary.
> 
> Walk the scheduler's NUMA hop masks and consider each newly reached CPU
> once, preserving locality across the full fallback search and after
> affinity relaxation.
> 
> The NUMA masks can be unavailable or incomplete while topology is rebuilt.
> Search any online CPUs not covered by the masks before relaxing affinity
> so fallback selection remains reliable during that window.
> 
> Signed-off-by: Yury Norov <ynorov@nvidia.com>
> ---
>  kernel/sched/core.c | 47 ++++++++++++++++++++++++---------------------
>  1 file changed, 25 insertions(+), 22 deletions(-)
> 
> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> index f78275192036..2155bad0e1ab 100644
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -3554,34 +3554,37 @@ EXPORT_SYMBOL_GPL(kick_process);
>   */
>  static int select_fallback_rq(int cpu, struct task_struct *p)
>  {
> -	int nid = cpu_to_node(cpu);
> -	const struct cpumask *nodemask = NULL;
> +	int nid = IS_ENABLED(CONFIG_NUMA) ? cpu_to_node(cpu) : NUMA_NO_NODE;
>  	enum { cpuset, possible, fail } state = cpuset;
>  	int dest_cpu;
>  
> -	/*
> -	 * If the node that the CPU is on has been offlined, cpu_to_node()
> -	 * will return -1. There is no CPU on the node, and we should
> -	 * select the CPU on the other node.
> -	 */
> -	if (nid != -1) {
> -		nodemask = cpumask_of_node(nid);
> -
> -		/* Look for allowed, online CPU in same node. */
> -		for_each_cpu(dest_cpu, nodemask) {
> -			if (is_cpu_allowed(p, dest_cpu))
> -				return dest_cpu;
> -		}
> -	}
> -
>  	for (;;) {
> -		/* Any allowed, online CPU? */
> -		for_each_cpu(dest_cpu, p->cpus_ptr) {
> -			if (!is_cpu_allowed(p, dest_cpu))
> -				continue;
> +		const struct cpumask *prev = cpu_none_mask, *cpus;
> +
> +		/* Look for the closest allowed, online CPU. */
> +		rcu_read_lock();
> +		for_each_numa_hop_mask(cpus, nid) {
> +			for_each_cpu_andnot(dest_cpu, cpus, prev) {
> +				if (is_cpu_allowed(p, dest_cpu)) {
> +					rcu_read_unlock();
> +					goto out;
> +				}
> +			}
> +			prev = cpus;
> +		}
>  
> -			goto out;
> +		/*
> +		 * NUMA masks may be unavailable or incomplete while the
> +		 * topology is being rebuilt. Search CPUs not covered by them
> +		 * before relaxing the task's affinity.
> +		 */
> +		for_each_cpu_andnot(dest_cpu, p->cpus_ptr, prev) {
> +			if (is_cpu_allowed(p, dest_cpu)) {
> +				rcu_read_unlock();
> +				goto out;
> +			}
>  		}
> +		rcu_read_unlock();
>  
>  		/* No more Mr. Nice Guy. */
>  		switch (state) {
> -- 
> 2.53.0