Re: [PATCH] sched/cpuset: distribute tasks within affinity masks

[Peter Zijlstra <peterz-wEGCiKHe2LqWVfeAwA7xHQ@public.gmane.org>]

On Thu, Feb 27, 2020 at 05:01:34PM -0800, Josh Don wrote:
> diff --git a/include/linux/sched.h b/include/linux/sched.h
> index 04278493bf15..a2aab6a8a794 100644
> --- a/include/linux/sched.h
> +++ b/include/linux/sched.h
> @@ -1587,6 +1587,8 @@ extern int task_can_attach(struct task_struct *p, const struct cpumask *cs_cpus_
>  #ifdef CONFIG_SMP
>  extern void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask);
>  extern int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask);
> +extern int set_cpus_allowed_ptr_distribute(struct task_struct *p,
> +				const struct cpumask *new_mask);

Why? Changelog doesn't seem to give a reason for adding another
interface.

> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> index 1a9983da4408..2336d6d66016 100644
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -1612,6 +1612,32 @@ void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
>  		set_next_task(rq, p);
>  }
>  
> +static DEFINE_PER_CPU(int, distribute_cpu_mask_prev);
> +
> +/*
> + * Returns an arbitrary cpu within *srcp1 & srcp2
> + *
> + * Iterated calls using the same srcp1 and srcp2, passing the previous cpu each
> + * time, will be distributed within their intersection.
> + */
> +static int distribute_to_new_cpumask(const struct cpumask *src1p,
> +				     const struct cpumask *src2p)
> +{
> +	int next, prev;
> +
> +	/* NOTE: our first selection will skip 0. */
> +	prev = __this_cpu_read(distribute_cpu_mask_prev);
> +
> +	next = cpumask_next_and(prev, src1p, src2p);
> +	if (next >= nr_cpu_ids)
> +		next = cpumask_first_and(src1p, src2p);
> +
> +	if (next < nr_cpu_ids)
> +		__this_cpu_write(distribute_cpu_mask_prev, next);
> +
> +	return next;
> +}

That's a valid implementation of cpumask_any_and(), it just has a really
weird name.

>  /*
>   * Change a given task's CPU affinity. Migrate the thread to a
>   * proper CPU and schedule it away if the CPU it's executing on
> @@ -1621,11 +1647,11 @@ void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
>   * task must not exit() & deallocate itself prematurely. The
>   * call is not atomic; no spinlocks may be held.
>   */
> -static int __set_cpus_allowed_ptr(struct task_struct *p,
> +static int __set_cpus_allowed_ptr(struct task_struct *p, bool distribute_cpus,
>  				  const struct cpumask *new_mask, bool check)
>  {
>  	const struct cpumask *cpu_valid_mask = cpu_active_mask;
> -	unsigned int dest_cpu;
> +	unsigned int dest_cpu, prev_cpu;
>  	struct rq_flags rf;
>  	struct rq *rq;
>  	int ret = 0;
> @@ -1652,8 +1678,33 @@ static int __set_cpus_allowed_ptr(struct task_struct *p,
>  	if (cpumask_equal(p->cpus_ptr, new_mask))
>  		goto out;
>  
> -	dest_cpu = cpumask_any_and(cpu_valid_mask, new_mask);
> -	if (dest_cpu >= nr_cpu_ids) {
> +	if (!cpumask_intersects(new_mask, cpu_valid_mask)) {
> +		ret = -EINVAL;
> +		goto out;
> +	}
> +
> +	prev_cpu = task_cpu(p);
> +	if (distribute_cpus) {
> +		dest_cpu = distribute_to_new_cpumask(new_mask,
> +						     cpu_valid_mask);
> +	} else {
> +		/*
> +		 * Can the task run on the task's current CPU? If so, we're
> +		 * done.
> +		 *
> +		 * We only enable this short-circuit in the case that we're
> +		 * not trying to distribute tasks.  As we may otherwise not
> +		 * distribute away from a loaded CPU, or make duplicate
> +		 * assignments to it.
> +		 */
> +		if (cpumask_test_cpu(prev_cpu, new_mask))
> +			dest_cpu = prev_cpu;
> +		else
> +			dest_cpu = cpumask_any_and(cpu_valid_mask, new_mask);
> +	}

That all seems overly complicated; what is wrong with just this:

	dest_cpu = cpumask_any_and_fancy(cpu_valid_mask, new_mask);

I don't really buy the argument why that shortcut is problematic; it's
all averages anyway, and keeping a task on a CPU where it's already
running seems like a win.

> +	/* May have raced with cpu_down */
> +	if (unlikely(dest_cpu >= nr_cpu_ids)) {
>  		ret = -EINVAL;
>  		goto out;
>  	}