Re: [RFC PATCH v1 4/8] sched/proxy: Tell the class about a proxy migration

From: John Stultz

Date: Mon Nov 16 2026 - 15:42:38 EST


On Mon, Nov 16, 2026 at 11:26 AM Mathieu Desnoyers
<mathieu.desnoyers@xxxxxxxxxxxx> wrote:
>
> find_proxy_task() moves a blocked donor to the runqueue of the mutex
> owner. proxy_set_task_cpu() does this with __set_task_cpu(), so that it
> can keep p->wake_cpu, the CPU the donor returns to.
>
> But __set_task_cpu() skips what set_task_cpu() does around it:
> - the migrate_task_rq() method of the class;
> - the migration count and the perf event;
> - the sched_migrate_task tracepoint.
>
> For a fair task, the dequeue on the old runqueue detaches its load.
> migrate_task_rq_fair() would then reset last_update_time, which makes
> the enqueue on the new runqueue attach the load. Without that call, the
> load is attached nowhere. When the donor is woken, the return migration
> goes through set_task_cpu(). It removes that load from the owner's
> runqueue, which never got it.
>
> So each round trip of a donor lowers the load and the utilization of
> the owner's runqueue. In a test, a CPU with two tasks that never sleep
> showed a mean util_avg of 195, where 1024 is expected. schedutil would
> then pick a low frequency for a CPU that is fully busy, and the load
> balancer would see that CPU as nearly idle.
>
> Fix this by doing in proxy_set_task_cpu() what set_task_cpu() does
> around __set_task_cpu(). Tracers and perf now see a proxy migration as
> a migration, and nr_migrations counts it.
>
> migrate_task_rq_dl() moves the bandwidth of a task that migrates to
> another root domain. "sched/deadline: Move a task's bandwidth to the
> root domain it runs in" added that. A proxy migration must not do it:
> the task does not run where it is queued. So make migrate_task_rq_dl()
> leave the bandwidth of such a task in its root domain. If its wake-up
> then picks a CPU of another root domain, the bandwidth moves at that
> time.
>
> Fixes: b049b81bdff6 ("sched: Handle blocked-waiter migration (and return migration)")
> Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@xxxxxxxxxxxx>
> ---
> kernel/sched/core.c | 13 +++++++++++++
> kernel/sched/deadline.c | 8 +++++++-
> 2 files changed, 20 insertions(+), 1 deletion(-)
>
> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> index 24105eeeb3c7..5d82c343662f 100644
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -6967,6 +6967,19 @@ static inline void proxy_set_task_cpu(struct task_struct *p, int cpu)
> * was previously runnable.
> */
> wake_cpu = p->wake_cpu;
> + /*
> + * Tell the class and the tracers about the migration, as
> + * set_task_cpu() does. The fair class then attaches the load of @p
> + * to the new runqueue. Without that, the return migration removes
> + * load that the runqueue never got.
> + */
> + trace_sched_migrate_task(p, cpu);
> + if (task_cpu(p) != cpu) {
> + if (p->sched_class->migrate_task_rq)
> + p->sched_class->migrate_task_rq(p, cpu);
> + p->se.nr_migrations++;
> + perf_event_task_migrate(p);
> + }
> __set_task_cpu(p, cpu);
> p->wake_cpu = wake_cpu;
> }

So I've got a pending fix from Yu-Ting in my tree (apologies for the
gerrit link, I've not cleaned up and pushed my tree out yet) that I
believe duplicates the bit above:
https://r.android.com/4410518


> diff --git a/kernel/sched/deadline.c b/kernel/sched/deadline.c
> index 3b62ee43c209..d14ce3615ba6 100644
> --- a/kernel/sched/deadline.c
> +++ b/kernel/sched/deadline.c
> @@ -3030,8 +3030,14 @@ static void migrate_task_rq_dl(struct task_struct *p, int new_cpu)
> * the bandwidth. A mask can hold CPUs of two root domains, and
> * select_fallback_rq() can widen it. Move the bandwidth there, as
> * dl_task_offline_migration() does for a throttled task.
> + *
> + * Not for a task that is blocked on a mutex and only queued for
> + * proxy execution. It does not run on the new CPU, so its bandwidth
> + * stays in its root domain. If its wake-up then picks a CPU of another
> + * root domain, the bandwidth moves at that time.
> */
> - if (unlikely(dl_task_bw_elsewhere(p, new_cpu)))
> + if (unlikely(dl_task_bw_elsewhere(p, new_cpu)) &&
> + !(p->is_blocked && task_on_rq_migrating(p)))
> dl_task_move_bw(p, new_cpu);
>
> if (READ_ONCE(p->__state) != TASK_WAKING)

I may need to pull this bit in separately.

thanks!
-john