[PATCH v3] sched_ext: Trigger ops.update_idle() from pick_task_idle()

[Andrea Righi]

With the consolidation of put_prev_task/set_next_task(), see
commit 436f3eed5c69 ("sched: Combine the last put_prev_task() and the
first set_next_task()"), we are now skipping the transition between
these two functions when the previous and the next tasks are the same.

As a result, ops.update_idle() is now called only once when the CPU
transitions to the idle class. If the CPU stays active (e.g., through a
call to scx_bpf_kick_cpu()), ops.update_idle() will not be triggered
again since the task remains unchanged (rq->idle).

While this behavior seems generally correct, it can cause issues in
certain sched_ext scenarios.

For example, a BPF scheduler might use logic like the following to keep
the CPU active under specific conditions:

void BPF_STRUCT_OPS(sched_update_idle, s32 cpu, bool idle)
{
	if (!idle)
		return;
	if (condition)
		scx_bpf_kick_cpu(cpu, 0);
}

A call to scx_bpf_kick_cpu() wakes up the CPU, so in theory,
ops.update_idle() should be triggered again until the condition becomes
false. However, this doesn't happen, and scx_bpf_kick_cpu() doesn't
produce the expected effect.

In practice, this change badly impacts performance in user-space
schedulers that rely on ops.update_idle() to activate user-space
components.

For instance, in the case of scx_rustland, performance drops
significantly (e.g., gaming benchmarks fall from ~60fps to ~10fps).

To address this, trigger ops.update_idle() from pick_task_idle() rather
than set_next_task_idle(). This restores the correct behavior of
ops.update_idle() and it allows to fix the performance regression in
scx_rustland.

Fixes: 7c65ae81ea86 ("sched_ext: Don't call put_prev_task_scx() before picking the next task")
Signed-off-by: Andrea Righi <andrea.righi@xxxxxxxxx>
---
 kernel/sched/idle.c | 15 ++++++++++++++-
 1 file changed, 14 insertions(+), 1 deletion(-)

ChangeLog v2 -> v3:
  - add a comment to clarify why we need to update the scx idle state in
    pick_task()

ChangeLog v1 -> v2:
  - move the logic from put_prev_set_next_task() to scx_update_idle()

diff --git a/kernel/sched/idle.c b/kernel/sched/idle.c
index d2f096bb274c..d336a05a6006 100644
--- a/kernel/sched/idle.c
+++ b/kernel/sched/idle.c
@@ -459,13 +459,26 @@ static void put_prev_task_idle(struct rq *rq, struct task_struct *prev, struct t
 static void set_next_task_idle(struct rq *rq, struct task_struct *next, bool first)
 {
 	update_idle_core(rq);
-	scx_update_idle(rq, true);
 	schedstat_inc(rq->sched_goidle);
 	next->se.exec_start = rq_clock_task(rq);
 }
 
 struct task_struct *pick_task_idle(struct rq *rq)
 {
+	/*
+	 * When switching from a non-idle to the idle class, .set_next_task()
+	 * is called only once during the transition.
+	 *
+	 * However, the CPU may remain active for multiple rounds (e.g., by
+	 * calling scx_bpf_kick_cpu() from the ops.update_idle() callback).
+	 *
+	 * In such cases, we need to keep updating the scx idle state to
+	 * properly re-trigger the ops.update_idle() callback.
+	 *
+	 * Updating the state in .pick_task(), instead of .set_next_task(),
+	 * ensures correct handling of scx idle state transitions.
+	 */
+	scx_update_idle(rq, true);
 	return rq->idle;
 }
 
-- 
2.47.0