[PATCH] sched/pelt: Skip dormant cfs util_sum under scx_switched_all

From: Peng Huang

Date: Sat Nov 21 2026 - 20:57:36 EST


When scx_switched_all() is enabled, all fair-class tasks move to
sched_ext, scheduler_tick() skips sched_balance_trigger(), and fair_rq
scheduling is bypassed. As a result, rq->cfs.avg.util_sum stops updating
and retains whatever stale load was present prior to switching to SCX.

When a CPU later goes idle, update_idle_rq_clock_pelt() still adds this
dormant rq->cfs.avg.util_sum. If stale cfs.avg.util_sum plus active RT
and DL util_sum reaches divider, update_idle_rq_clock_pelt() treats the
rq as overutilized, continuously increments rq->lost_idle_time, and
stalls rq_clock_pelt(), which in turn freezes RT and DL PELT updates.

Exclude rq->cfs.avg.util_sum in update_idle_rq_clock_pelt() while
scx_switched_all() is active so dormant CFS load cannot stall idle
rq_clock_pelt().

Fixes: f0e1a0643a59 ("sched_ext: Implement BPF extensible scheduler class")
Cc: stable@xxxxxxxxxxxxxxx
Signed-off-by: Peng Huang <penghuang@xxxxxxxxxx>
---
kernel/sched/pelt.h | 7 ++++++-
1 file changed, 6 insertions(+), 1 deletion(-)

diff --git a/kernel/sched/pelt.h b/kernel/sched/pelt.h
index f921302dc40f..6d31e1fbd094 100644
--- a/kernel/sched/pelt.h
+++ b/kernel/sched/pelt.h
@@ -138,7 +138,12 @@ static inline void update_rq_clock_pelt(struct rq *rq, s64 delta)
static inline void update_idle_rq_clock_pelt(struct rq *rq)
{
u32 divider = ((LOAD_AVG_MAX - 1024) << SCHED_CAPACITY_SHIFT) - LOAD_AVG_MAX;
- u32 util_sum = rq->cfs.avg.util_sum;
+ /*
+ * Under scx_switched_all(), CFS PELT is dormant and may stay stale;
+ * exclude it so idle CPUs do not trip lost_idle_time and freeze RT
+ * PELT.
+ */
+ u32 util_sum = scx_switched_all() ? 0 : rq->cfs.avg.util_sum;
util_sum += rq->avg_rt.util_sum;
util_sum += rq->avg_dl.util_sum;

--
2.56.1.750.gf70dc682cf-goog