Re: [RFC PATCH 07/11] sched: Add proxy execution

[Dietmar Eggemann]

On 03/10/2022 23:44, Connor O'Brien wrote:
> From: Peter Zijlstra <peterz@xxxxxxxxxxxxx>

[...]

> + * Returns the task that is going to be used as execution context (the one
> + * that is actually going to be put to run on cpu_of(rq)).
> + */
> +static struct task_struct *
> +proxy(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
> +{

[...]

> +migrate_task:

[...]

> +	/*
> +	 * Since we're going to drop @rq, we have to put(@next) first,
> +	 * otherwise we have a reference that no longer belongs to us.  Use
> +	 * @fake_task to fill the void and make the next pick_next_task()
           ^^^^^^^^^^

There was a `static struct task_struct fake_task` in
https://lkml.kernel.org/r/20181009092434.26221-6-juri.lelli@xxxxxxxxxx
but now IMHO we use `rq->idle` <-- (1)

> +	 * invocation happy.
> +	 *
> +	 * XXX double, triple think about this.
> +	 * XXX put doesn't work with ON_RQ_MIGRATE
> +	 *
> +	 * CPU0				CPU1
> +	 *
> +	 *				B mutex_lock(X)
> +	 *
> +	 * A mutex_lock(X) <- B
> +	 * A __schedule()
> +	 * A pick->A
> +	 * A proxy->B
> +	 * A migrate A to CPU1
> +	 *				B mutex_unlock(X) -> A
> +	 *				B __schedule()
> +	 *				B pick->A
> +	 *				B switch_to (A)
> +	 *				A ... does stuff
> +	 * A ... is still running here
> +	 *
> +	 *		* BOOM *
> +	 */
> +	put_prev_task(rq, next);
> +	if (curr_in_chain) {
> +		rq->proxy = rq->idle;
> +		set_tsk_need_resched(rq->idle);
> +		/*
> +		 * XXX [juril] don't we still need to migrate @next to
> +		 * @owner's CPU?
> +		 */
> +		return rq->idle;
> +	}

--> (1)

> +	rq->proxy = rq->idle;
> +
> +	for (; p; p = p->blocked_proxy) {
> +		int wake_cpu = p->wake_cpu;
> +
> +		WARN_ON(p == rq->curr);
> +
> +		deactivate_task(rq, p, 0);
> +		set_task_cpu(p, that_cpu);
> +		/*
> +		 * We can abuse blocked_entry to migrate the thing, because @p is
> +		 * still on the rq.
> +		 */
> +		list_add(&p->blocked_entry, &migrate_list);
> +
> +		/*
> +		 * Preserve p->wake_cpu, such that we can tell where it
> +		 * used to run later.
> +		 */
> +		p->wake_cpu = wake_cpu;
> +	}
> +
> +	rq_unpin_lock(rq, rf);
> +	raw_spin_rq_unlock(rq);

Don't we run into rq_pin_lock()'s:

SCHED_WARN_ON(rq->balance_callback && rq->balance_callback !=
&balance_push_callback)

by releasing rq lock between queue_balance_callback(, push_rt/dl_tasks)
and __balance_callbacks()?

[...]