Re: [PATCH resend 5/8] sched: cgroup cookie API for core scheduling

[Peter Zijlstra <peterz@infradead.org>]

On Wed, Mar 24, 2021 at 05:40:17PM -0400, Joel Fernandes (Google) wrote:
> From: Josh Don <joshdon@google.com>
> 
> This adds the API to set/get the cookie for a given cgroup. This
> interface lives at cgroup/cpu.core_tag.
> 
> The cgroup interface can be used to toggle a unique cookie value for all
> descendent tasks, preventing these tasks from sharing with any others.
> See Documentation/admin-guide/hw-vuln/core-scheduling.rst for a full
> rundown of both this and the per-task API.

I refuse to read RST. Life's too short for that.

> +u64 cpu_core_tag_read_u64(struct cgroup_subsys_state *css,
> +			  struct cftype *cft)
> +{
> +	return !!css_tg(css)->core_tagged;
> +}
> +
> +int cpu_core_tag_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
> +			   u64 val)
> +{
> +	static DEFINE_MUTEX(sched_core_group_mutex);
> +	struct task_group *tg = css_tg(css);
> +	struct cgroup_subsys_state *css_tmp;
> +	struct task_struct *p;
> +	unsigned long group_cookie;
> +	int ret = 0;
> +
> +	if (val > 1)
> +		return -ERANGE;
> +
> +	if (!static_branch_likely(&sched_smt_present))
> +		return -EINVAL;
> +
> +	mutex_lock(&sched_core_group_mutex);
> +
> +	if (!tg->core_tagged && val) {
> +		/* Tag is being set. Check ancestors and descendants. */
> +		if (cpu_core_get_group_cookie(tg) ||
> +		    cpu_core_check_descendants(tg, true /* tag */)) {
> +			ret = -EBUSY;
> +			goto out_unlock;
> +		}

So the desired semantics is to only allow a single tag on any upwards
path? Isn't that in conflict with the cgroup requirements?

TJ?

> +	} else if (tg->core_tagged && !val) {
> +		/* Tag is being reset. Check descendants. */
> +		if (cpu_core_check_descendants(tg, true /* tag */)) {

I'm struggling to understand this. If, per the above, you cannot set
when either a parent is already set or a child is set, then how can a
child be set to refuse clearing?

> +			ret = -EBUSY;
> +			goto out_unlock;
> +		}
> +	} else {
> +		goto out_unlock;
> +	}