From: Su Zhidao <suzhidao@xiaomi.com>

scx_bypass_lb_timerfn() runs in softirq (BH) context, so
rcu_dereference_bh() is the correct RCU accessor. The previous
rcu_dereference_all() suppresses all sparse warnings and masks
potential RCU context issues.

Add a comment noting this is a transitional state: when
multi-scheduler support lands, the bypass LB timer will become
per-scheduler and the global scx_root reference will be removed.

Signed-off-by: Su Zhidao <suzhidao@xiaomi.com>
---
 kernel/sched/ext.c | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

diff --git a/kernel/sched/ext.c b/kernel/sched/ext.c
index 053d99c58802..c269e489902c 100644
--- a/kernel/sched/ext.c
+++ b/kernel/sched/ext.c
@@ -4170,7 +4170,13 @@ static void scx_bypass_lb_timerfn(struct timer_list *timer)
 	int node;
 	u32 intv_us;
 
-	sch = rcu_dereference_all(scx_root);
+	/*
+	 * scx_bypass_lb_timer is a global timer that fires in softirq
+	 * context while bypass mode is active. Use rcu_dereference_bh()
+	 * matching the BH context. When multi-scheduler support lands,
+	 * this timer will become per-scheduler instance.
+	 */
+	sch = rcu_dereference_bh(scx_root);
 	if (unlikely(!sch) || !READ_ONCE(scx_bypass_depth))
 		return;
 
-- 
2.43.0