[PATCH v6 0/3] Fix uclamp race against non atomic static key enable

From: Qais Yousef

Date: Fri Nov 20 2026 - 23:15:21 EST


Latest version drops the partial revert (patch 3 in v5) as RT group
throttling is going to be marked as broken and anyway a proper fix will
require non trivial rewrite how dequeue happens for throttled tasks or
full revert of subtracting nr_running added 10 years ago. Both of which
is not appealing and with the proposal to mark this feature as broken,
better leave it and hopefully just delete the problem area.

So in summary RT throttling is not fixed and being ignored. Fair
throttling is fixed by pairing uclamp_rq_inc/dec with
add/sub_nr_running. For RT, we do that in inc/dec_tasks() which are
called when tasks are actually enqueued/dequeued - as stated will fail
in RT throttling case. Given the minor severity of the problem and
effort required to restructure part of RT group throttling that is being
marked broken as we speak, I think this should do.

Changes since v5:
- Drop patch 3 to partial revert subtracting rq->nr_running for
RT throttling.
- Drop fixes tag from patch 1 as the series now is partially
fixed and not suitable for LTS.

Changes since v4:
- Fix problems in fair/RT throttling logic that were exposed by
the original fix.

Changes since v3:
- Fix Sashiko report about force_idle paths where
rq->nr_running is not 0 highlighted by Christian
- Update patch 1 commit message to reflect jump_label call
sites on the same CPU might not be updated atomically too

Changes since v2:
- Use uclamp_rq_is_idle() instead of scanning buckets to detect
imbalance
- Do some cleanup on accessing the flag as an additional patch

Changes since v1:
- Fix kernel test robot report due to missing static inline in dummy
definition of init_uclamp_rq()

Tested with the patch below to force imbalance via sched_feat. I could see the
warning appearing and frequencies spiking during imbalance. Correct behavior is
restored when sched_feat is disabled where the idle entry reset fixes things
up.

I had a mix of uclamp_max mainly used in the system via schedqosd.

diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index a2035d66270b9..14f9f1e419e02 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -1889,6 +1889,11 @@ static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p, int flags
for_each_clamp_id(clamp_id)
uclamp_rq_inc_id(rq, p, clamp_id);

+ if (sched_feat(UCLAMP_IMBALANCED)) {
+ for_each_clamp_id(clamp_id)
+ uclamp_rq_inc_id(rq, p, clamp_id);
+ }
+
/* Reset clamp idle holding when there is one RUNNABLE task */
if (uclamp_rq_is_idle(rq))
uclamp_rq_clear_idle(rq);
diff --git a/kernel/sched/features.h b/kernel/sched/features.h
index 8f0dee8fc4755..b59e532500fd9 100644
--- a/kernel/sched/features.h
+++ b/kernel/sched/features.h
@@ -142,3 +142,5 @@ SCHED_FEAT(LATENCY_WARN, false)
*/
SCHED_FEAT(NI_RANDOM, true)
SCHED_FEAT(NI_RATE, true)
+
+SCHED_FEAT(UCLAMP_IMBALANCED, false)
diff --git a/kernel/sched/idle.c b/kernel/sched/idle.c
index 84301fc49e699..dbc3f5e1a7e01 100644
--- a/kernel/sched/idle.c
+++ b/kernel/sched/idle.c
@@ -519,7 +519,7 @@ static void set_next_task_idle(struct rq *rq, struct task_struct *next, enum snt
* are still queued.
*/
if (uclamp_is_used() &&
- !rq->nr_running && unlikely(!uclamp_rq_is_idle(rq)))
+ !rq->nr_running && unlikely(WARN_ON(!uclamp_rq_is_idle(rq))))
init_uclamp_rq(rq);
}

Qais Yousef (3):
sched/uclamp: Fix race condition leading to imbalanced uclamp_rq
sched/uclamp: Use wrappers instead of open coding UCLAMP_FLAG_IDLE
sched/uclamp: Move uclamp_rq_{inc, dec}() from core.c

kernel/sched/core.c | 49 +++++++++++++++---------------------------
kernel/sched/ext/ext.c | 6 ++----
kernel/sched/fair.c | 6 ++----
kernel/sched/idle.c | 19 ++++++++++++++++
kernel/sched/rt.c | 10 +++++----
kernel/sched/sched.h | 21 ++++++++++++++----
6 files changed, 63 insertions(+), 48 deletions(-)

--
2.53.0