[PATCH v2] sched_ext: Update other_load_avgs on set_next_task_scx and idle entry
From: Peng Huang
Date: Sun Nov 22 2026 - 01:21:31 EST
When scx_switched_all() is enabled, scheduler_tick() skips
sched_balance_trigger(), so __update_blocked_others() no longer runs to
decay RT, DL, IRQ, and HW pressure load averages. Other-class load
averages then depend on scx_tick() and SCX_RQ_CAN_STOP_TICK transitions
in set_next_task_scx().
If SCX_RQ_CAN_STOP_TICK stays unchanged (e.g., remaining set across an
idle or RT interval on tickless CPUs, or remaining clear on finite-slice
switches before a tick fires), or when an SCX task goes idle after an
RT/DL interval without an intervening tick, stale RT/DL PELT values
can remain elevated when consumers such as effective_cpu_util() or
update_idle_rq_clock_pelt() read them.
Refresh update_other_load_avgs() once at the entry of set_next_task_scx()
(before SNT_REPICK early return and before ops.running() for SNT_PICK /
SNT_NORMAL), and in put_prev_task_scx() when transitioning to the idle
class.
Fixes: 22a920209ab6 ("sched_ext: Implement tickless support")
Cc: stable@xxxxxxxxxxxxxxx
Signed-off-by: Peng Huang <penghuang@xxxxxxxxxx>
---
Changes in v2:
- Call update_other_load_avgs() once before SNT_REPICK return and before
ops.running() in set_next_task_scx() so BPF callbacks such as
scx_bpf_cpuperf_set() see updated PELT values, and cover SNT_REPICK.
- Refresh other-class load_avgs in put_prev_task_scx() when transitioning to
idle_sched_class before ops.update_idle() and update_idle_rq_clock_pelt().
- Link to v1: https://lore.kernel.org/r/20261122033316.2819845-1-penghuang@xxxxxxxxxx
kernel/sched/ext/ext.c | 27 ++++++++++++++++++---------
1 file changed, 18 insertions(+), 9 deletions(-)
diff --git a/kernel/sched/ext/ext.c b/kernel/sched/ext/ext.c
index 10de6bd865e8..523f86be4d26 100644
--- a/kernel/sched/ext/ext.c
+++ b/kernel/sched/ext/ext.c
@@ -3579,6 +3579,15 @@ static void set_next_task_scx(struct rq *rq, struct task_struct *p, enum snt_e t
if (unlikely(rq->scx.kick_sync_pending) && (type == SNT_PICK || type == SNT_REPICK))
queue_balance_callback(rq, &rq->scx.kick_sync_bal_cb, kick_sync_wait_bal_cb);
+ /*
+ * Refresh other-class (RT/DL/IRQ/HW) load averages before ops.running()
+ * (or returning on SNT_REPICK) so BPF callbacks such as
+ * scx_bpf_cpuperf_set() observe fresh utilization while CFS PELT is
+ * dormant. In the future, we might want to add a mechanism to update
+ * load_avgs periodically on tick-stopped CPUs.
+ */
+ update_other_load_avgs(rq);
+
if (type == SNT_REPICK)
return;
@@ -3619,10 +3628,6 @@ static void set_next_task_scx(struct rq *rq, struct task_struct *p, enum snt_e t
* @p is getting newly scheduled or got kicked after someone updated its
* slice. Update SCX_RQ_CAN_STOP_TICK to reflect whether the tick can be
* stopped. See scx_can_stop_tick().
- *
- * Moreover, refresh the load_avgs just when transitioning in and out of
- * nohz. In the future, we might want to add a mechanism to update
- * load_avgs periodically on tick-stopped CPUs.
*/
if (can_stop_tick) {
if (!(rq->scx.flags & SCX_RQ_CAN_STOP_TICK)) {
@@ -3634,14 +3639,10 @@ static void set_next_task_scx(struct rq *rq, struct task_struct *p, enum snt_e t
*/
rq->scx.flags |= SCX_RQ_CAN_STOP_TICK;
sched_update_tick_dependency(rq);
-
- update_other_load_avgs(rq);
}
} else {
- if (rq->scx.flags & SCX_RQ_CAN_STOP_TICK) {
+ if (rq->scx.flags & SCX_RQ_CAN_STOP_TICK)
rq->scx.flags &= ~SCX_RQ_CAN_STOP_TICK;
- update_other_load_avgs(rq);
- }
/*
* @rq still references the outgoing scheduling context. A finite
@@ -3863,6 +3864,14 @@ static void put_prev_task_scx(struct rq *rq, struct task_struct *p,
if (next && next->sched_class != &ext_sched_class) {
scx_cid_sched_update(rq, NULL);
switch_class(rq, next);
+
+ /*
+ * Refresh other-class load averages before entering idle so
+ * ops.update_idle() and update_idle_rq_clock_pelt() observe
+ * decayed RT/DL/IRQ/HW metrics while CFS PELT is dormant.
+ */
+ if (sched_class_above(&ext_sched_class, next->sched_class))
+ update_other_load_avgs(rq);
}
}
--
2.56.1.750.gf70dc682cf-goog