* Re: [PATCH] sched/fair: Don't let task_tick_cache() IPI dyntick CPUs
2026-11-19 20:38 [PATCH] sched/fair: Don't let task_tick_cache() IPI dyntick CPUs Valentin Schneider
@ 2026-11-20 8:34 ` Peter Zijlstra
0 siblings, 0 replies; 2+ messages in thread
From: Peter Zijlstra @ 2026-11-20 8:34 UTC (permalink / raw)
To: Valentin Schneider
Cc: linux-kernel, Ingo Molnar, Juri Lelli, Vincent Guittot,
Dietmar Eggemann, Steven Rostedt, Ben Segall, Mel Gorman,
K Prateek Nayak
On Thu, Nov 19, 2026 at 09:38:25PM +0100, Valentin Schneider wrote:
> While testing some CPU isolation / NOHZ_FULL stuff I stumbled on something
> like so:
>
> kworker/u289:5-241489 [001] d..3. 2963.728734: ipi_send_cpu: cpu=6 callsite=task_work_add+0x84 callback=0x0
> kworker/u289:5-241489 [001] d..3. 2963.728735: kernel_stack: <stack trace >
> => trace_event_raw_event_ipi_send_cpu (ffffffffa6d7ba45)
> => kick_process (ffffffffa6d827a2)
> => task_work_add (ffffffffa6d645c4)
> => task_tick_fair (ffffffffa6d9cfbe)
> => sched_tick_remote (ffffffffa6d84c69)
> => process_one_work (ffffffffa6d5d57b)
> => worker_thread (ffffffffa6d5e1af)
> => kthread (ffffffffa6d66e85)
> => ret_from_fork (ffffffffa6cb3f2b)
> => ret_from_fork_asm (ffffffffa6c7741a)
>
> In this case CPU6 is isolated via isolcpus=, and is executing a single pure
> userspace task, so should be fully leveraging dynticks.
>
> However, because that task is SCHED_OTHER, housekeeping CPUs periodically
> handle a sched_tick_remote() on its behalf, and end up kicking that task
> out of userspace to run the task_cache_work().
>
> Let task_tick_cache() recognize when it is invoked via a remote tick, and
> bail out rather than disturb dyntick CPUs.
>
> Signed-off-by: Valentin Schneider <vschneid@redhat.com>
> ---
> kernel/sched/fair.c | 7 +++++++
> 1 file changed, 7 insertions(+)
>
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index d4cc0caee6265..42394745a124b 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -1925,6 +1925,13 @@ static void task_tick_cache(struct rq *rq, struct task_struct *p)
> !grp->pcpu_sched)
> return;
>
> + /*
> + * We got here from a sched_tick_remote(), meaning @p is making use of
> + * NOHZ_FULL. Don't disturb it with a task_work() IPI.
> + */
> + if (cpu_of(rq) != smp_processor_id())
> + return;
> +
> epoch = rq->cpu_epoch;
> /* avoid moving backwards */
> if (time_after_eq(grp->epoch, epoch))
Its early, and I definitely need more wake-up juice, but how is
task_tick_numa() different here?
Also, I'm not sure sched_tick_remote() is the right condition here. But
perhaps something like so? Does this make sense?
---
diff --git a/arch/x86/include/asm/stop_machine.h b/arch/x86/include/asm/stop_machine.h
deleted file mode 100644
index e69de29bb2d1..000000000000
diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 731d78155e65..ddcfd82a9e8e 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -4588,6 +4588,9 @@ static void task_tick_numa(struct rq *rq, struct task_struct *curr)
struct callback_head *work = &curr->numa_work;
u64 period, now;
+ if (!static_branch_unlikely(&sched_numa_balancing))
+ return;
+
/*
* We don't care about NUMA placement if we don't have memory.
*/
@@ -15401,15 +15404,20 @@ static void task_tick_fair(struct rq *rq, struct task_struct *curr, int hrtick)
if (hrtick)
return;
- if (static_branch_unlikely(&sched_numa_balancing))
- task_tick_numa(rq, curr);
+ task_tick_core(rq, curr);
+ /*
+ * If there is no root domain; there is no balancing, and none of this
+ * matters.
+ */
+ if (!rq->rd)
+ return;
+
+ task_tick_numa(rq, curr);
task_tick_cache(rq, curr);
update_misfit_status(curr, rq);
check_update_overutilized_status(task_rq(curr));
-
- task_tick_core(rq, curr);
}
/*
^ permalink raw reply [flat|nested] 2+ messages in thread