Re: [PATCH v4 5/5] sched/core: Fix donor slice accounting under proxy execution

[Hui Su <sh_def@163.com>]

On Wed, Sep 09, 2026 at 06:29:01PM +0900, Hui Su wrote:
> Core scheduling uses __entity_slice_used() to decide whether the current
> scheduling context has consumed enough of its slice to let a force-idled
> SMT sibling run.
>
> The check is correctly made against rq->donor, since the slice belongs
> to the scheduling context. With proxy execution, however, task runtime
> is accounted to rq->curr. The donor's sum_exec_runtime therefore does
> not advance while another task executes on its behalf, causing
>
>         se->sum_exec_runtime - se->prev_sum_exec_runtime
>
> to remain near zero and preventing the force-idle reschedule from
> triggering.
>
> Using rq->curr is not correct either, as that would compare the execution
> task's runtime against its own slice rather than the donor's slice.
>
> Track the donor's task-clock timestamp when it is selected and measure
> the elapsed service using se->exec_start. update_se() advances the donor's
> exec_start from rq_clock_task() even under proxy execution, while the
> accumulated task runtime itself is charged to rq->curr.
>
> Keeping the comparison in the task-clock domain also avoids depending on
> the entity's weight. A vruntime delta accumulated across different
> weights cannot reliably be compared against a slice converted using only
> the current weight.
>
> Only snapshot task entities, as task_tick_core() performs the consumed
> slice check on the donor task.
>
> In non-proxy testing, the task-clock predicate matched the existing
> sum_exec_runtime predicate across HZ=100/250/1000 and nice -10/0/+10.
> Under proxy execution, the donor's sum_exec_runtime delta remained zero
> while the task-clock delta advanced and triggered the force-idle
> reschedule.

Following up on the task_tick_core() discussion with Tim and Chen Yu, I
reworked the implementation while keeping the slice check associated with
rq->donor, because the slice belongs to the scheduling context.

The new predicate measures consumed service in the task-clock domain rather
than using the donor's sum_exec_runtime:

  rtime = se->exec_start - rq->core_sched_start;

update_se() advances the donor's exec_start from rq_clock_task() during
proxy execution even though task-level sum_exec_runtime is charged to
rq->curr. This keeps both sides of the comparison in the task-clock domain.

The deadline/virtual-time prototype did not work reliably in my reproducer:
update_deadline() can advance the deadline before task_tick_core() evaluates
the predicate, making the reconstructed virtual service small again after
the donor has consumed service.

The baseline is stored in struct rq rather than every sched_entity. There is
one active donor per runqueue, so per-entity storage was unnecessary and
increased sched_entity from 224 to 256 bytes on i386 with
CONFIG_SCHED_CORE=y.

One proxy-specific detail is that an execution-owner handoff can retain the
same donor. The scheduler then uses a synthetic put_prev_task()/
set_next_task() pair for balance handling. That reselect must not reset the
donor's core_sched_start, or service before the handoff would be discarded.
The implementation refreshes the baseline only when selecting a different
donor or when reactivating a task after a scheduling-class transition.

In non-proxy testing, the new predicate matched the existing
sum_exec_runtime predicate in 89,900 samples across HZ=100/250/1000 and
nice -10/0/+10 with zero mismatches. The proxy-enabled
CONFIG_SCHED_CORE=y scheduler objects and full bzImage also build with this
change. A focused force-idle test which splits one donor's service across two
execution owners is still pending; I will not claim owner-handoff runtime
coverage until that case runs.

Could you take a look at whether this task-clock and per-rq baseline approach
is reasonable? If so, I will complete the focused owner-handoff test and carry
the result into v5 before posting the updated series.

For reference, the complete current core-slice patch follows.

Thanks,
Hui

---
diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index ed08b287016a..0fea899d61c7 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -6469,7 +6469,8 @@ dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
 }
 
 static void
-set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
+set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
+		bool reset_core_slice)
 {
 	/* 'current' is not kept within the tree. */
 	if (se->on_rq) {
@@ -6502,6 +6503,10 @@ set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
 	}
 
 	se->prev_sum_exec_runtime = se->sum_exec_runtime;
+#ifdef CONFIG_SCHED_CORE
+	if (reset_core_slice && entity_is_task(se))
+		rq_of(cfs_rq)->core_sched_start = se->exec_start;
+#endif
 }
 
 static bool __dequeue_task(struct rq *rq, struct task_struct *p, int flags);
@@ -14786,16 +14791,15 @@ static void rq_offline_fair(struct rq *rq)
 
 #ifdef CONFIG_SCHED_CORE
 static inline bool
-__entity_slice_used(struct sched_entity *se, int min_nr_tasks)
+__entity_slice_used(struct rq *rq, struct sched_entity *se, int min_nr_tasks)
 {
-	u64 rtime = se->sum_exec_runtime - se->prev_sum_exec_runtime;
-	u64 slice = se->slice;
+	u64 rtime = se->exec_start - rq->core_sched_start;
 
-	return (rtime * min_nr_tasks > slice);
+	return (rtime * min_nr_tasks > se->slice);
 }
 
 #define MIN_NR_TASKS_DURING_FORCEIDLE	2
-static inline void task_tick_core(struct rq *rq, struct task_struct *curr)
+static inline void task_tick_core(struct rq *rq, struct task_struct *donor)
 {
 	if (!sched_core_enabled(rq))
 		return;
@@ -14815,7 +14819,8 @@ static inline void task_tick_core(struct rq *rq, struct task_struct *curr)
 	 * if we need to give up the CPU.
 	 */
 	if (rq->core->core_forceidle_count && rq->cfs.h_nr_queued == 1 &&
-	    __entity_slice_used(&curr->se, MIN_NR_TASKS_DURING_FORCEIDLE))
+	    __entity_slice_used(rq, &donor->se,
+				MIN_NR_TASKS_DURING_FORCEIDLE))
 		resched_curr(rq);
 }
 
@@ -15049,7 +15054,7 @@ static int task_is_throttled_fair(struct task_struct *p, int cpu)
 	return throttled_hierarchy(cfs_rq);
 }
 #else /* !CONFIG_SCHED_CORE: */
-static inline void task_tick_core(struct rq *rq, struct task_struct *curr) {}
+static inline void task_tick_core(struct rq *rq, struct task_struct *donor) {}
 #endif /* !CONFIG_SCHED_CORE */
 
 /*
@@ -15257,6 +15262,8 @@ static void set_next_task_fair(struct rq *rq, struct task_struct *p, bool first)
 {
 	struct sched_entity *se = &p->se;
 	bool throttled = false;
+	/* Keep service accumulated across a same-donor proxy reselect. */
+	bool reset_core_slice = !first || rq->donor != p;
 	struct cfs_rq *cfs_rq = &rq->cfs;
 	unsigned long weight = NICE_0_LOAD;
 	bool on_rq = se->on_rq;
@@ -15271,7 +15278,7 @@ static void set_next_task_fair(struct rq *rq, struct task_struct *p, bool first)
 
 		if (!IS_ENABLED(CONFIG_FAIR_GROUP_SCHED) ||
 		    !first || !cfs_rq->h_curr)
-			set_next_entity(cfs_rq, se);
+			set_next_entity(cfs_rq, se, reset_core_slice);
 
 		/* ensure bandwidth has been allocated on our new cfs_rq */
 		throttled |= account_cfs_rq_runtime(cfs_rq, 0);
diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
index bb6f87552220..b430092fc9ad 100644
--- a/kernel/sched/sched.h
+++ b/kernel/sched/sched.h
@@ -1382,6 +1382,7 @@ struct rq {
 	unsigned int		core_forceidle_seq;
 	unsigned int		core_forceidle_occupation;
 	u64			core_forceidle_start;
+	u64			core_sched_start;
 	unsigned int		core_pick_in_flight;
 #endif /* CONFIG_SCHED_CORE */