Skip to contentSkip to conversation

sched_ext: Call wakeup_preempt() in local_dsq_post_enq()

PATCH Superseded Kuba Piecuch 2 participants 4 messages Apr 24 – Apr 27

Subsystems (2)

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 sched_ext: Call wakeup_preempt() in local_dsq_post_enq()

There are several edge cases (see linked thread) where an IMMED task can be left lingering on a local DSQ if an RT task swoops in at the wrong time. All of these edge cases are due to rq->next_class being idle even after dispatching a task to rq's local DSQ. We should bump rq->next_class to &ext_sched_class as soon as we've inserted a task into the local DSQ.

To optimize the common case of rq->next_class == &ext_sched_class, only call wakeup_preempt() if rq->next_class is below EXT. If next_class is EXT or above, wakeup_preempt() is a no-op anyway.

---
 kernel/sched/ext.c | 15 +++++++++++----
 1 file changed, 11 insertions(+), 4 deletions(-)
kernel/sched/ext.c +11 -4
index 1f670028bf19..034df77e3af1 100644
@@ -1393,7 +1393,6 @@ static void local_dsq_post_enq(struct scx_sched *sch, struct scx_dispatch_q *dsq
 			       struct task_struct *p, u64 enq_flags)
 {
 	struct rq *rq = container_of(dsq, struct rq, scx.local_dsq);
-	bool preempt = false;
 
 	call_task_dequeue(sch, rq, p, 0);
 
@@ -1408,11 +1407,19 @@ static void local_dsq_post_enq(struct scx_sched *sch, struct scx_dispatch_q *dsq
 	if ((enq_flags & SCX_ENQ_PREEMPT) && p != rq->curr &&
 	    rq->curr->sched_class == &ext_sched_class) {
 		rq->curr->scx.slice = 0;
-		preempt = true;
+		resched_curr(rq);
 	}
 
-	if (preempt || sched_class_above(&ext_sched_class, rq->curr->sched_class))
-		resched_curr(rq);
+	/*
+	 * If @rq->next_class is currently idle, we need to bump it
+	 * to &ext_sched_class using wakeup_preempt(). Otherwise, if we drop
+	 * the rq lock later in the pick and an RT task wakes up on @rq,
+	 * wakeup_preempt_idle() will be called during RT task wakeup and
+	 * SCX won't have an opportunity to re-enqueue IMMED tasks from @rq's
+	 * local DSQ.
+	 */
+	if (sched_class_above(&ext_sched_class, rq->next_class))
+		wakeup_preempt(rq, p, 0);
 }
 
 static void dispatch_enqueue(struct scx_sched *sch, struct rq *rq,
Signature · 1 line
2.54.0.rc2.544.gc7ae2d5bb8-goog
reviewer
#1 lore raw

Hello, Kuba.

kernel/sched/ext.c static void local_dsq_post_enq(struct scx_sched *sch, struct scx_dispatch_q *dsq · line 1415

Quoted 10 lines · Kuba Piecuch
kernel/sched/ext.c +1 -3
@@ -1408,11 +1407,19 @@ static void local_dsq_post_enq(struct scx_sched *sch, struct scx_dispatch_q *dsq
 	if ((enq_flags & SCX_ENQ_PREEMPT) && p != rq->curr &&
 	    rq->curr->sched_class == &ext_sched_class) {
 		rq->curr->scx.slice = 0;
-		preempt = true;
+		resched_curr(rq);
 	}
 
-	if (preempt || sched_class_above(&ext_sched_class, rq->curr->sched_class))
-		resched_curr(rq);

Hmm... I don't quite understand this part of the change. sched_class_above() got separated out into its own case but why is it dropping resched_curr() on SCX_ENQ_PREEMPT?

kernel/sched/ext.c static void local_dsq_post_enq(struct scx_sched *sch, struct scx_dispatch_q *dsq · line 1419

Quoted 7 lines
kernel/sched/ext.c +7 -0
+	/*
+	 * If @rq->next_class is currently idle, we need to bump it
+	 * to &ext_sched_class using wakeup_preempt(). Otherwise, if we drop
+	 * the rq lock later in the pick and an RT task wakes up on @rq,
+	 * wakeup_preempt_idle() will be called during RT task wakeup and
+	 * SCX won't have an opportunity to re-enqueue IMMED tasks from @rq's
+	 * local DSQ.

As this was really subtle, I think it warrants documenting all cases here.

Thanks.

Signature · 1 line
tejun
#2 lore raw

Hi Tejun,

Quoted 11 lines · Tejun Heo

Hello, Kuba.

Quoted 8 lines · Kuba Piecuch
kernel/sched/ext.c +1 -1
@@ -1408,11 +1407,19 @@ static void local_dsq_post_enq(struct scx_sched *sch, struct scx_dispatch_q *dsq
 	if ((enq_flags & SCX_ENQ_PREEMPT) && p != rq->curr &&
 	    rq->curr->sched_class == &ext_sched_class) {
 		rq->curr->scx.slice = 0;
-		preempt = true;
+		resched_curr(rq);
 	}
 
Quoted 6 lines

kernel/sched/ext.c static void local_dsq_post_enq(struct scx_sched *sch, struct scx_dispatch_q *dsq · line 1415

Quoted 2 lines
kernel/sched/ext.c +0 -2
-	if (preempt || sched_class_above(&ext_sched_class, rq->curr->sched_class))
-		resched_curr(rq);

Hmm... I don't quite understand this part of the change. sched_class_above() got separated out into its own case but why is it dropping resched_curr() on SCX_ENQ_PREEMPT?

In the SCX_ENQ_PREEMPT case we call resched_curr() where we previously set preempt = true.

In the sched_class_above() case, wakeup_preempt() will call resched_curr() for us:

	void wakeup_preempt(struct rq *rq, struct task_struct *p, int flags)
	{
		[...]

		if (p->sched_class == rq->next_class) {
			rq->next_class->wakeup_preempt(rq, p, flags);

		} else if (sched_class_above(p->sched_class, rq->next_class)) {
			rq->next_class->wakeup_preempt(rq, p, flags);
		=====>	resched_curr(rq);  <=====
			rq->next_class = p->sched_class;
		}

		[...]
	}
Quoted 10 lines

kernel/sched/ext.c static void local_dsq_post_enq(struct scx_sched *sch, struct scx_dispatch_q *dsq · line 1419

Quoted 8 lines
kernel/sched/ext.c +7 -0
+	/*
+	 * If @rq->next_class is currently idle, we need to bump it
+	 * to &ext_sched_class using wakeup_preempt(). Otherwise, if we drop
+	 * the rq lock later in the pick and an RT task wakes up on @rq,
+	 * wakeup_preempt_idle() will be called during RT task wakeup and
+	 * SCX won't have an opportunity to re-enqueue IMMED tasks from @rq's
+	 * local DSQ.

As this was really subtle, I think it warrants documenting all cases here.

Yeah, I was trying to keep it concise. How about something like this:

/*
* Note that @rq's lock may be dropped between this enqueue and @p
* actually getting on CPU. This gives higher-class tasks (e.g. RT)
* an opportunity to wake up on @rq and prevent @p from running.
* Here are some concrete examples:
*
* Example 1:
*
* We dispatch two tasks from a single ops.dispatch():
* - First, a local task to this CPU's local DSQ;
* - Second, a local/remote task to a remote CPU's local DSQ.
* We must drop the local rq lock in order to finish the second
* dispatch. In that time, an RT task can wake up on the local rq. *
* Example 2:
*
* We dispatch a local/remote task to a remote CPU's local DSQ.
* We must drop the remote rq lock before the dispatched task can run,
* which gives an RT task an opportunity to wake up on the remote rq. *
* Both examples work the same if we replace dispatching with moving
* the tasks from a user-created DSQ.
*
* We must detect these wakeups so that we can re-enqueue IMMED tasks
* from @rq's local DSQ. scx_wakeup_preempt() serves exactly this
* purpose, but for it to be invoked, we must ensure that we bump
* @rq->next_class to &ext_sched_class if it's currently idle. *
* wakeup_preempt() does the bumping, and since we only invoke it if
* @rq->next_class is below &ext_sched_class, it will also
* resched_curr(rq).
*/

Thanks,
Kuba

reviewer
#3 lore raw
Quoted 9 lines · Kuba Piecuch
Quoted 3 lines

Hmm... I don't quite understand this part of the change. sched_class_above() got separated out into its own case but why is it dropping resched_curr() on SCX_ENQ_PREEMPT?

In the SCX_ENQ_PREEMPT case we call resched_curr() where we previously set preempt = true.

In the sched_class_above() case, wakeup_preempt() will call resched_curr() for us:

I see. It'd be helpful to note that in desc.

Quoted 36 lines
Quoted 9 lines

kernel/sched/ext.c static void local_dsq_post_enq(struct scx_sched *sch, struct scx_dispatch_q *dsq · line 1419

Quoted 7 lines
kernel/sched/ext.c +7 -0
+	/*
+	 * If @rq->next_class is currently idle, we need to bump it
+	 * to &ext_sched_class using wakeup_preempt(). Otherwise, if we drop
+	 * the rq lock later in the pick and an RT task wakes up on @rq,
+	 * wakeup_preempt_idle() will be called during RT task wakeup and
+	 * SCX won't have an opportunity to re-enqueue IMMED tasks from @rq's
+	 * local DSQ.

As this was really subtle, I think it warrants documenting all cases here.

Yeah, I was trying to keep it concise. How about something like this:

/*
* Note that @rq's lock may be dropped between this enqueue and @p
* actually getting on CPU. This gives higher-class tasks (e.g. RT)
* an opportunity to wake up on @rq and prevent @p from running.
* Here are some concrete examples:
*
* Example 1:
*
* We dispatch two tasks from a single ops.dispatch():
* - First, a local task to this CPU's local DSQ;
* - Second, a local/remote task to a remote CPU's local DSQ.
* We must drop the local rq lock in order to finish the second
* dispatch. In that time, an RT task can wake up on the local rq. *
* Example 2:
*
* We dispatch a local/remote task to a remote CPU's local DSQ.
* We must drop the remote rq lock before the dispatched task can run,
* which gives an RT task an opportunity to wake up on the remote rq. *
* Both examples work the same if we replace dispatching with moving
* the tasks from a user-created DSQ.
*
* We must detect these wakeups so that we can re-enqueue IMMED tasks

Quoted 8 lines

* from @rq's local DSQ. scx_wakeup_preempt() serves exactly this
* purpose, but for it to be invoked, we must ensure that we bump
* @rq->next_class to &ext_sched_class if it's currently idle. *
* wakeup_preempt() does the bumping, and since we only invoke it if
* @rq->next_class is below &ext_sched_class, it will also
* resched_curr(rq).
*/

Looks good to me.

Thanks.

Signature · 1 line
tejun