schedule_deferred() uses irq_work_queue() which always queues on the
calling CPU. The deferred work can run from any CPU correctly, and the
_locked() path already processes remote rqs from the calling CPU. However,
when falling through to the irq_work path, queuing on the target CPU is
preferable as the work can run sooner via IPI delivery rather than waiting
for the calling CPU to re-enable IRQs.

Currently, only reenqueue operations use this path - either BPF-initiated
reenqueue targeting a remote rq, or IMMED reenqueue when the target CPU is
busy running userspace (not in balance or wakeup, so the _locked() fast
paths aren't available). Use irq_work_queue_on() to target the owning CPU.

This improves IMMED reenqueue latency when tasks are dispatched to
remote local DSQs. Testing on a 24-CPU AMD Ryzen 3900X with scx_qmap
-I -F 50 (ALWAYS_ENQ_IMMED, every 50th enqueue forced to prev_cpu's
local DSQ) under heavy mixed load (2x CPU oversubscription, yield and
context-switch pressure, SCHED_FIFO bursts, periodic fork storms, mixed
nice levels, C-states disabled), measuring local DSQ residence time
(insert to remove) over 5 x 120s runs (~1.2M tasks per set):

  >128us outliers:  71 -> 39  (-45%)
  >256us outliers:  59 -> 36  (-39%)

Signed-off-by: Tejun Heo <tj@kernel.org>
---
 kernel/sched/ext.c |   14 +++++++++++---
 1 file changed, 11 insertions(+), 3 deletions(-)

--- a/kernel/sched/ext.c
+++ b/kernel/sched/ext.c
@@ -1164,10 +1164,18 @@ static void deferred_irq_workfn(struct i
 static void schedule_deferred(struct rq *rq)
 {
 	/*
-	 * Queue an irq work. They are executed on IRQ re-enable which may take
-	 * a bit longer than the scheduler hook in schedule_deferred_locked().
+	 * This is the fallback when schedule_deferred_locked() can't use
+	 * the cheaper balance callback or wakeup hook paths (the target
+	 * CPU is not in balance or wakeup). Currently, this is primarily
+	 * hit by reenqueue operations targeting a remote CPU.
+	 *
+	 * Queue on the target CPU. The deferred work can run from any CPU
+	 * correctly - the _locked() path already processes remote rqs from
+	 * the calling CPU - but targeting the owning CPU allows IPI delivery
+	 * without waiting for the calling CPU to re-enable IRQs and is
+	 * cheaper as the reenqueue runs locally.
 	 */
-	irq_work_queue(&rq->scx.deferred_irq_work);
+	irq_work_queue_on(&rq->scx.deferred_irq_work, cpu_of(rq));
 }

 /**
--
tejun

Applied to sched_ext/for-7.1.

Thanks.

--
tejun

On Sun Mar 22, 2026 at 4:33 PM EDT, Tejun Heo wrote:
> schedule_deferred() uses irq_work_queue() which always queues on the
> calling CPU. The deferred work can run from any CPU correctly, and the
> _locked() path already processes remote rqs from the calling CPU. However,
> when falling through to the irq_work path, queuing on the target CPU is
> preferable as the work can run sooner via IPI delivery rather than waiting
> for the calling CPU to re-enable IRQs.
>
> Currently, only reenqueue operations use this path - either BPF-initiated
> reenqueue targeting a remote rq, or IMMED reenqueue when the target CPU is
> busy running userspace (not in balance or wakeup, so the _locked() fast
> paths aren't available). Use irq_work_queue_on() to target the owning CPU.
>
> This improves IMMED reenqueue latency when tasks are dispatched to
> remote local DSQs. Testing on a 24-CPU AMD Ryzen 3900X with scx_qmap
> -I -F 50 (ALWAYS_ENQ_IMMED, every 50th enqueue forced to prev_cpu's
> local DSQ) under heavy mixed load (2x CPU oversubscription, yield and
> context-switch pressure, SCHED_FIFO bursts, periodic fork storms, mixed
> nice levels, C-states disabled), measuring local DSQ residence time
> (insert to remove) over 5 x 120s runs (~1.2M tasks per set):
>
>   >128us outliers:  71 -> 39  (-45%)
>   >256us outliers:  59 -> 36  (-39%)
>
> Signed-off-by: Tejun Heo <tj@kernel.org>

Reviewed-by: Emil Tsalapatis <emil@etsalapatis.com>

> ---
>  kernel/sched/ext.c |   14 +++++++++++---
>  1 file changed, 11 insertions(+), 3 deletions(-)
>
> --- a/kernel/sched/ext.c
> +++ b/kernel/sched/ext.c
> @@ -1164,10 +1164,18 @@ static void deferred_irq_workfn(struct i
>  static void schedule_deferred(struct rq *rq)
>  {
>  	/*
> -	 * Queue an irq work. They are executed on IRQ re-enable which may take
> -	 * a bit longer than the scheduler hook in schedule_deferred_locked().
> +	 * This is the fallback when schedule_deferred_locked() can't use
> +	 * the cheaper balance callback or wakeup hook paths (the target
> +	 * CPU is not in balance or wakeup). Currently, this is primarily
> +	 * hit by reenqueue operations targeting a remote CPU.
> +	 *
> +	 * Queue on the target CPU. The deferred work can run from any CPU
> +	 * correctly - the _locked() path already processes remote rqs from
> +	 * the calling CPU - but targeting the owning CPU allows IPI delivery
> +	 * without waiting for the calling CPU to re-enable IRQs and is
> +	 * cheaper as the reenqueue runs locally.
>  	 */
> -	irq_work_queue(&rq->scx.deferred_irq_work);
> +	irq_work_queue_on(&rq->scx.deferred_irq_work, cpu_of(rq));
>  }
>
>  /**
> --
> tejun