Re: [PATCH v6 4/5] sched: Handle set_cpus_allowed_ptr() & sched_setaffinity() race

[Peter Zijlstra]

On Thu, Aug 25, 2022 at 09:01:18PM -0400, Waiman Long wrote:
> @@ -2722,6 +2734,7 @@ static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flag
>  			complete = true;
>  		}
>  
> +		swap_user_cpus_ptr(p, puser_mask);
>  		task_rq_unlock(rq, p, rf);
>  
>  		if (push_task) {
> @@ -2793,6 +2806,7 @@ static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flag
>  		if (flags & SCA_MIGRATE_ENABLE)
>  			p->migration_flags &= ~MDF_PUSH;
>  
> +		swap_user_cpus_ptr(p, puser_mask);
>  		task_rq_unlock(rq, p, rf);
>  
>  		if (!stop_pending) {
> @@ -2813,6 +2827,8 @@ static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flag
>  				complete = true;
>  			}
>  		}
> +
> +		swap_user_cpus_ptr(p, puser_mask);
>  		task_rq_unlock(rq, p, rf);
>  
>  		if (complete)

I'm not at all sure about those.

Would it not be much simpler to keep the update of cpus_mask and
cpus_user_mask together, always ensuring that cpus_user_mask is a strict
superset of cpus_mask ? That is, set_cpus_allowed_common() seems like
the right place to me.

I'm thinking this also means blowing away user_mask when we do a full
reset of the cpus_mask when we do an affnity break.