Re: [PATCH v6 3/5] sched: Enforce user requested affinity

[Peter Zijlstra <peterz@infradead.org>]

On Thu, Aug 25, 2022 at 09:01:17PM -0400, Waiman Long wrote:

> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> index ac2b103d69dc..1c2f548e5369 100644
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -2928,11 +2928,40 @@ static int __set_cpus_allowed_ptr_locked(struct task_struct *p,
>  static int __set_cpus_allowed_ptr(struct task_struct *p,
>  				  const struct cpumask *new_mask, u32 flags)
>  {
> +	struct cpumask *alloc_mask = NULL;
>  	struct rq_flags rf;
>  	struct rq *rq;
> +	int ret;
>  
>  	rq = task_rq_lock(p, &rf);
> -	return __set_cpus_allowed_ptr_locked(p, new_mask, flags, rq, &rf);
> +	if (p->user_cpus_ptr) {
> +
> +		/*
> +		 * A scratch cpumask is allocated on the percpu runqueues
> +		 * to enable additional masking with user_cpus_ptr. This
> +		 * cpumask, once allocated, will not be freed.
> +		 */
> +		if (unlikely(!rq->scratch_mask)) {
> +			alloc_mask = kmalloc(cpumask_size(), GFP_ATOMIC);

This -- absolutely not. You can't have allocations under a
raw_spinlock_t.