Re: [PATCH v4 3/5] sched/fair: Switch to task based throttle model

[Aaron Lu <ziqianlu@bytedance.com>]

On Thu, Sep 04, 2025 at 07:26:10PM +0800, Aaron Lu wrote:
> On Wed, Sep 03, 2025 at 01:55:36PM -0700, Benjamin Segall wrote:
> > Aaron Lu <ziqianlu@bytedance.com> writes:
> > 
> > > +static bool enqueue_throttled_task(struct task_struct *p)
> > > +{
> > > +	struct cfs_rq *cfs_rq = cfs_rq_of(&p->se);
> > > +
> > > +	/* @p should have gone through dequeue_throttled_task() first */
> > > +	WARN_ON_ONCE(!list_empty(&p->throttle_node));
> > > +
> > > +	/*
> > > +	 * If the throttled task @p is enqueued to a throttled cfs_rq,
> > > +	 * take the fast path by directly putting the task on the
> > > +	 * target cfs_rq's limbo list.
> > > +	 *
> > > +	 * Do not do that when @p is current because the following race can
> > > +	 * cause @p's group_node to be incorectly re-insterted in its rq's
> > > +	 * cfs_tasks list, despite being throttled:
> > > +	 *
> > > +	 *     cpuX                       cpuY
> > > +	 *   p ret2user
> > > +	 *  throttle_cfs_rq_work()  sched_move_task(p)
> > > +	 *  LOCK task_rq_lock
> > > +	 *  dequeue_task_fair(p)
> > > +	 *  UNLOCK task_rq_lock
> > > +	 *                          LOCK task_rq_lock
> > > +	 *                          task_current_donor(p) == true
> > > +	 *                          task_on_rq_queued(p) == true
> > > +	 *                          dequeue_task(p)
> > > +	 *                          put_prev_task(p)
> > > +	 *                          sched_change_group()
> > > +	 *                          enqueue_task(p) -> p's new cfs_rq
> > > +	 *                                             is throttled, go
> > > +	 *                                             fast path and skip
> > > +	 *                                             actual enqueue
> > > +	 *                          set_next_task(p)
> > > +	 *                    list_move(&se->group_node, &rq->cfs_tasks); // bug
> > > +	 *  schedule()
> > > +	 *
> > > +	 * In the above race case, @p current cfs_rq is in the same rq as
> > > +	 * its previous cfs_rq because sched_move_task() only moves a task
> > > +	 * to a different group from the same rq, so we can use its current
> > > +	 * cfs_rq to derive rq and test if the task is current.
> > > +	 */
> > > +	if (throttled_hierarchy(cfs_rq) &&
> > > +	    !task_current_donor(rq_of(cfs_rq), p)) {
> > > +		list_add(&p->throttle_node, &cfs_rq->throttled_limbo_list);
> > > +		return true;
> > > +	}
> > > +
> > > +	/* we can't take the fast path, do an actual enqueue*/
> > > +	p->throttled = false;
> > > +	return false;
> > > +}
> > > +
> > 
> > Is there a reason that __set_next_task_fair cannot check p->se.on_rq as
> > well as (or instead of) task_on_rq_queued()? All of the _entity parts of
> > set_next/put_prev check se.on_rq for this sort of thing, so that seems
> > fairly standard. And se.on_rq should exactly match if the task is on
> > cfs_tasks since that add/remove is done in account_entity_{en,de}queue.
> 
> Makes sense to me.
> 
> Only thing that feels a little strange is, a throttled/dequeued task is
> set as next now. Maybe not a big deal. I booted a VM and run some tests,
> didn't notice anything wrong but I could very well miss some cases.

Sorry, I should have added: the above test was done with following diff:

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index cb93e74a850e8..7a6782617c0e8 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -5836,38 +5836,8 @@ static bool enqueue_throttled_task(struct task_struct *p)
 	 * If the throttled task @p is enqueued to a throttled cfs_rq,
 	 * take the fast path by directly putting the task on the
 	 * target cfs_rq's limbo list.
-	 *
-	 * Do not do that when @p is current because the following race can
-	 * cause @p's group_node to be incorectly re-insterted in its rq's
-	 * cfs_tasks list, despite being throttled:
-	 *
-	 *     cpuX                       cpuY
-	 *   p ret2user
-	 *  throttle_cfs_rq_work()  sched_move_task(p)
-	 *  LOCK task_rq_lock
-	 *  dequeue_task_fair(p)
-	 *  UNLOCK task_rq_lock
-	 *                          LOCK task_rq_lock
-	 *                          task_current_donor(p) == true
-	 *                          task_on_rq_queued(p) == true
-	 *                          dequeue_task(p)
-	 *                          put_prev_task(p)
-	 *                          sched_change_group()
-	 *                          enqueue_task(p) -> p's new cfs_rq
-	 *                                             is throttled, go
-	 *                                             fast path and skip
-	 *                                             actual enqueue
-	 *                          set_next_task(p)
-	 *                    list_move(&se->group_node, &rq->cfs_tasks); // bug
-	 *  schedule()
-	 *
-	 * In the above race case, @p current cfs_rq is in the same rq as
-	 * its previous cfs_rq because sched_move_task() only moves a task
-	 * to a different group from the same rq, so we can use its current
-	 * cfs_rq to derive rq and test if the task is current.
 	 */
-	if (throttled_hierarchy(cfs_rq) &&
-	    !task_current_donor(rq_of(cfs_rq), p)) {
+	if (throttled_hierarchy(cfs_rq)) {
 		list_add(&p->throttle_node, &cfs_rq->throttled_limbo_list);
 		return true;
 	}
@@ -13256,7 +13226,7 @@ static void __set_next_task_fair(struct rq *rq, struct task_struct *p, bool firs
 {
 	struct sched_entity *se = &p->se;
 
-	if (task_on_rq_queued(p)) {
+	if (se->on_rq) {
 		/*
 		 * Move the next running task to the front of the list, so our
 		 * cfs_tasks list becomes MRU one.