[PATCH] sched_ext:SCX is preemptible by the Fair scheduler

[PATCH] sched_ext:SCX is preemptible by the Fair scheduler

[Cui Jian]

sched_ext currently distinguishes preemption caused by STOP,
DL and RT scheduling classes, but preemption from the Fair
scheduler is reported as SCX_CPU_PREEMPT_UNKNOWN.

When analyzing SCX scheduling behavior, it is useful to know
whether a task was preempted by RT/DL classes or switched out
in favor of a Fair-class task. The lack of a dedicated reason
makes debugging and observability more difficult.

Add SCX_CPU_PREEMPT_FAIR and report it when the next scheduled
task belongs to fair_sched_class.

This change improves scheduling diagnostics and observability
without affecting scheduling behavior.

Signed-off-by: Cui Jian <cjian720@163.com>
---
 kernel/sched/ext.c          | 2 ++
 kernel/sched/ext_internal.h | 2 ++
 2 files changed, 4 insertions(+)

diff --git a/kernel/sched/ext.c b/kernel/sched/ext.c
index 5d2d19473a82..6e0281218197 100644
--- a/kernel/sched/ext.c
+++ b/kernel/sched/ext.c
@@ -3015,6 +3015,8 @@ preempt_reason_from_class(const struct sched_class *class)
 		return SCX_CPU_PREEMPT_DL;
 	if (class == &rt_sched_class)
 		return SCX_CPU_PREEMPT_RT;
+	if (class == &fair_sched_class)
+		return SCX_CPU_PREEMPT_FAIR;
 	return SCX_CPU_PREEMPT_UNKNOWN;
 }
 
diff --git a/kernel/sched/ext_internal.h b/kernel/sched/ext_internal.h
index a075732d4430..3ce07bec58a5 100644
--- a/kernel/sched/ext_internal.h
+++ b/kernel/sched/ext_internal.h
@@ -241,6 +241,8 @@ enum scx_cpu_preempt_reason {
 	SCX_CPU_PREEMPT_DL,
 	/* next task is being scheduled by &sched_class_stop */
 	SCX_CPU_PREEMPT_STOP,
+	/* next task is being scheduled by &sched_class_fair*/
+	SCX_CPU_PREEMPT_FAIR,
 	/* unknown reason for SCX being preempted */
 	SCX_CPU_PREEMPT_UNKNOWN,
 };
-- 
2.34.1

Re: [PATCH] sched_ext:SCX is preemptible by the Fair scheduler

[sashiko-bot]

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [Medium] The patch adds `SCX_CPU_PREEMPT_FAIR` to `enum scx_cpu_preempt_reason` in `kernel/sched/ext_internal.h`, but fails to update the checked-in autogenerated BPF tool headers in `tools/sched_ext/include/scx/enum_defs.autogen.h`.
--

commit 0e9cc05c20430d666c8c1a1d7ea0741d9c28b2d3
Author: Cui Jian <cjian720@163.com>

sched_ext:SCX is preemptible by the Fair scheduler

This commit adds a new preemption reason, SCX_CPU_PREEMPT_FAIR, to distinguish
when a sched_ext task is preempted by the Fair scheduler. It updates the
internal enum and mapping function to improve observability and diagnostics.

> diff --git a/kernel/sched/ext_internal.h b/kernel/sched/ext_internal.h
> index a075732d4430d..3ce07bec58a55 100644
> --- a/kernel/sched/ext_internal.h
> +++ b/kernel/sched/ext_internal.h
> @@ -241,6 +241,8 @@ enum scx_cpu_preempt_reason {
>  	SCX_CPU_PREEMPT_DL,
>  	/* next task is being scheduled by &sched_class_stop */
>  	SCX_CPU_PREEMPT_STOP,
> +	/* next task is being scheduled by &sched_class_fair*/
> +	SCX_CPU_PREEMPT_FAIR,

[Severity: Medium]
Should tools/sched_ext/include/scx/enum_defs.autogen.h be regenerated
and included in this patch?

When adding a new value to enum scx_cpu_preempt_reason, the autogenerated
BPF tool headers typically need updating so downstream BPF schedulers can
use feature detection macros like HAVE_SCX_CPU_PREEMPT_FAIR. 

Without this update, BPF applications cannot conditionally check for this
new preemption feature across different kernel versions.

>  	/* unknown reason for SCX being preempted */
>  	SCX_CPU_PREEMPT_UNKNOWN,
>  };

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20260612161403.105775-1-cjian720@163.com?part=1

Re: [PATCH] sched_ext:SCX is preemptible by the Fair scheduler

[Andrea Righi]

Hi Cui,

On Sat, Jun 13, 2026 at 12:14:03AM +0800, Cui Jian wrote:
> sched_ext currently distinguishes preemption caused by STOP,
> DL and RT scheduling classes, but preemption from the Fair
> scheduler is reported as SCX_CPU_PREEMPT_UNKNOWN.
> 
> When analyzing SCX scheduling behavior, it is useful to know
> whether a task was preempted by RT/DL classes or switched out
> in favor of a Fair-class task. The lack of a dedicated reason
> makes debugging and observability more difficult.
> 
> Add SCX_CPU_PREEMPT_FAIR and report it when the next scheduled
> task belongs to fair_sched_class.
> 
> This change improves scheduling diagnostics and observability
> without affecting scheduling behavior.

The preempt reason is only used in ops.cpu_release(), which is a deprecated API
and will be removed in the next kernel versions (see the pr_warn() in
validate_ops(), kernel/sched/ext.c):

 pr_warn("ops->cpu_acquire/release() are deprecated, use sched_switch TP instead\n");

So, your patch makes sense, but I'm not sure it's worth updating the API at this
point.

Thanks,
-Andrea

> 
> Signed-off-by: Cui Jian <cjian720@163.com>
> ---
>  kernel/sched/ext.c          | 2 ++
>  kernel/sched/ext_internal.h | 2 ++
>  2 files changed, 4 insertions(+)
> 
> diff --git a/kernel/sched/ext.c b/kernel/sched/ext.c
> index 5d2d19473a82..6e0281218197 100644
> --- a/kernel/sched/ext.c
> +++ b/kernel/sched/ext.c
> @@ -3015,6 +3015,8 @@ preempt_reason_from_class(const struct sched_class *class)
>  		return SCX_CPU_PREEMPT_DL;
>  	if (class == &rt_sched_class)
>  		return SCX_CPU_PREEMPT_RT;
> +	if (class == &fair_sched_class)
> +		return SCX_CPU_PREEMPT_FAIR;
>  	return SCX_CPU_PREEMPT_UNKNOWN;
>  }
>  
> diff --git a/kernel/sched/ext_internal.h b/kernel/sched/ext_internal.h
> index a075732d4430..3ce07bec58a5 100644
> --- a/kernel/sched/ext_internal.h
> +++ b/kernel/sched/ext_internal.h
> @@ -241,6 +241,8 @@ enum scx_cpu_preempt_reason {
>  	SCX_CPU_PREEMPT_DL,
>  	/* next task is being scheduled by &sched_class_stop */
>  	SCX_CPU_PREEMPT_STOP,
> +	/* next task is being scheduled by &sched_class_fair*/
> +	SCX_CPU_PREEMPT_FAIR,
>  	/* unknown reason for SCX being preempted */
>  	SCX_CPU_PREEMPT_UNKNOWN,
>  };
> -- 
> 2.34.1
> 

Re: Re: [PATCH] sched_ext:SCX is preemptible by the Fair scheduler

[Andrea Righi]

On Sat, Jun 13, 2026 at 01:32:28AM +0800, cjian720 wrote:
> Hi Andrea, While this callback will be phased out eventually, SCX is still a relatively new scheduling framework. Incomplete preemption reason enums make it hard to analyze interactions between different schedulers, and filling this gap greatly boosts observability for existing SCX workloads. This patch does not touch core scheduling logic, carries zero performance overhead, and adds negligible long-term maintenance cost. 

You can get the same information directly from BPF using the sched_switch
tracepoint and looking at prev->sched_class and next->sched_class. As I
mentioned, ops.cpu_release() is a deprecated API, so I don't see much value
updating/improving it.

Thanks,
-Andrea