Add a core event, SELECT_CPU_FALLBACK, which represents how many times
ops.select_cpu() returns a CPU that the task can't use.

__scx_add_event() is used since the caller holds a p->pi_lock,
so the preemption has already been disabled.

Signed-off-by: Changwoo Min <changwoo@igalia.com>
---
 kernel/sched/ext.c | 11 +++++++++++
 1 file changed, 11 insertions(+)

diff --git a/kernel/sched/ext.c b/kernel/sched/ext.c
index dbf659d593a3..727b1f8b623e 100644
--- a/kernel/sched/ext.c
+++ b/kernel/sched/ext.c
@@ -1462,6 +1462,11 @@ static struct task_struct *scx_task_iter_next_locked(struct scx_task_iter *iter)
  * Collection of event counters. Event types are placed in descending order.
  */
 struct scx_event_stats {
+	/*
+	 * If ops.select_cpu() returns a CPU which can't be used by the task,
+	 * the core scheduler code silently picks a fallback CPU.
+	 */
+	u64		SELECT_CPU_FALLBACK;
 };
 
 /*
@@ -3663,6 +3668,10 @@ static int select_task_rq_scx(struct task_struct *p, int prev_cpu, int wake_flag
 		cpu = SCX_CALL_OP_TASK_RET(SCX_KF_ENQUEUE | SCX_KF_SELECT_CPU,
 					   select_cpu, p, prev_cpu, wake_flags);
 		*ddsp_taskp = NULL;
+
+		if (unlikely(!is_cpu_allowed(p, cpu)))
+			__scx_add_event(SELECT_CPU_FALLBACK, 1);
+
 		if (ops_cpu_valid(cpu, "from ops.select_cpu()"))
 			return cpu;
 		else
@@ -5482,6 +5491,7 @@ static void scx_dump_state(struct scx_exit_info *ei, size_t dump_len)
 	dump_line(&s, "--------------");
 
 	scx_bpf_event_stats(&events, sizeof(events));
+	scx_dump_event(s, &events, SELECT_CPU_FALLBACK);
 
 	if (seq_buf_has_overflowed(&s) && dump_len >= sizeof(trunc_marker))
 		memcpy(ei->dump + dump_len - sizeof(trunc_marker),
@@ -7810,6 +7820,7 @@ __bpf_kfunc void scx_bpf_event_stats(struct scx_event_stats *events,
 	memset(&e_sys, 0, sizeof(e_sys));
 	for_each_possible_cpu(cpu) {
 		e_cpu = per_cpu_ptr(&event_stats_cpu, cpu);
+		scx_agg_event(&e_sys, e_cpu, SELECT_CPU_FALLBACK);
 	}
 
 	/*
-- 
2.48.1

Hello,

On Sun, Jan 26, 2025 at 07:16:06PM +0900, Changwoo Min wrote:
...
>  struct scx_event_stats {
> +	/*
> +	 * If ops.select_cpu() returns a CPU which can't be used by the task,
> +	 * the core scheduler code silently picks a fallback CPU.
> +	 */
> +	u64		SELECT_CPU_FALLBACK;

As C has one global namespace, we're gonna have to prefix these. Hopefully,
something not too long. SCX_EV_?

> @@ -3663,6 +3668,10 @@ static int select_task_rq_scx(struct task_struct *p, int prev_cpu, int wake_flag
>  		cpu = SCX_CALL_OP_TASK_RET(SCX_KF_ENQUEUE | SCX_KF_SELECT_CPU,
>  					   select_cpu, p, prev_cpu, wake_flags);
>  		*ddsp_taskp = NULL;
> +
> +		if (unlikely(!is_cpu_allowed(p, cpu)))
> +			__scx_add_event(SELECT_CPU_FALLBACK, 1);

This is trying to guess what select_task_rq() is going to do and then count
that as an event, which doesn't seem great. Can't we just record the picked
CPU here into a field in p->scx and then compare that against cpu_of(rq) in
enqueue_task_scx()? That way, we can reliably count events where CPU
selection was modified by the fallback logic regardless of why or how that
happens.

Thanks.

-- 
tejun

Hello,

On 25. 1. 28. 05:01, Tejun Heo wrote:
> Hello,
> 
> On Sun, Jan 26, 2025 at 07:16:06PM +0900, Changwoo Min wrote:
> ...
>>   struct scx_event_stats {
>> +	/*
>> +	 * If ops.select_cpu() returns a CPU which can't be used by the task,
>> +	 * the core scheduler code silently picks a fallback CPU.
>> +	 */
>> +	u64		SELECT_CPU_FALLBACK;
> 
> As C has one global namespace, we're gonna have to prefix these. Hopefully,
> something not too long. SCX_EV_?

Sure, I will add SCX_EV_ as a prefix for all events.

> 
>> @@ -3663,6 +3668,10 @@ static int select_task_rq_scx(struct task_struct *p, int prev_cpu, int wake_flag
>>   		cpu = SCX_CALL_OP_TASK_RET(SCX_KF_ENQUEUE | SCX_KF_SELECT_CPU,
>>   					   select_cpu, p, prev_cpu, wake_flags);
>>   		*ddsp_taskp = NULL;
>> +
>> +		if (unlikely(!is_cpu_allowed(p, cpu)))
>> +			__scx_add_event(SELECT_CPU_FALLBACK, 1);
> 
> This is trying to guess what select_task_rq() is going to do and then count
> that as an event, which doesn't seem great. Can't we just record the picked
> CPU here into a field in p->scx and then compare that against cpu_of(rq) in
> enqueue_task_scx()? That way, we can reliably count events where CPU
> selection was modified by the fallback logic regardless of why or how that
> happens.

That sounds good. I will add a field (say, selected_cpu?) to struct 
sched_ext_entity.

Regards,
Changwoo Min