* [PATCH v4] sched/stats: Fix run_delay over-count for migrated sched_delayed tasks
@ 2026-10-24 4:27 albin_yang
2026-10-25 13:03 ` albin_yang
0 siblings, 1 reply; 2+ messages in thread
From: albin_yang @ 2026-10-24 4:27 UTC (permalink / raw)
To: peterz, mingo, juri.lelli, vincent.guittot
Cc: dietmar.eggemann, rostedt, bsegall, mgorman, vschneid,
kprateek.nayak, chen.yu, kayracizmeci, mintaohuang, linux-kernel,
albin_yang, albinwyang, Chen Yu
From: Wei Yang <albinwyang@tencent.com>
With DELAY_DEQUEUE, a blocked task stays on the runqueue with
se.sched_delayed set and its sched_info.last_queued is cleared, so the sleep
is not counted into run_delay.
When such a delayed (sleeping) task is migrated across CPUs via the plain
migration paths (move_queued_task / move_queued_task_locked, the latter used
by __migrate_swap_task), activate_task(dst, 0) calls enqueue_task() without
ENQUEUE_RESTORE, re-arming last_queued to the migration timestamp while the
task is still sleeping. The later real wakeup (ENQUEUE_DELAYED) tries to
re-arm last_queued at wakeup time but is suppressed because last_queued is
already non-zero, so sched_info_arrive() folds the whole sleep duration
between migration and wakeup into run_delay.
Load balance is affected too: the sched_delayed check in can_migrate_task()
bails out only when env->migration_type != migrate_load, so it does not
block migration when the type is migrate_load - which active load balance
always uses (its lb_env leaves migration_type at 0 == migrate_load), and
which regular load balance can also use via calculate_imbalance(). Either
way the re-attach goes through attach_task() -> activate_task(rq, p,
ENQUEUE_NOCLOCK), without ENQUEUE_RESTORE.
Fix by not re-arming last_queued for a sched_delayed task in
sched_info_enqueue(). The wakeup path clears sched_delayed before reaching
sched_info_enqueue(), so it still re-arms at the real wakeup time. Plain
runnable tasks are unaffected.
Fixes: 152e11f6df29 ("sched/fair: Implement delayed dequeue")
Reported-by: MingTao Huang <mintaohuang@tencent.com>
Signed-off-by: Wei Yang <albinwyang@tencent.com>
Reviewed-by: Chen Yu <yu.c.chen@intel.com>
Reviewed-by: Kayra Cizmeci <kayracizmeci@gmail.com>
Reviewed-by: K Prateek Nayak <kprateek.nayak@amd.com>
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
---
Changes in v4:
- No code change; resending after the proxy exec discussion to resurface the
patch.
- Proxy exec has the same re-arm problem, but it is handled separately:
t->is_blocked cannot gate it (in ttwu_runnable() the real wakeup enqueue
sits inside the if (p->is_blocked) branch, while is_blocked is cleared
afterwards in ttwu_do_wakeup(), so !is_blocked would suppress the very
re-arm this patch relies on), and task_is_blocked() cannot replace
se.sched_delayed since a delayed sleeper has no blocked_on. Coupling them
would also prevent backporting to v6.12.y..v6.16.y, where the bug exists
(DELAY_DEQUEUE, v6.12) but proxy exec does not (v6.17).
v3: https://lore.kernel.org/all/20260922082432.2987855-1-albin_yang@163.com/
v2: https://lore.kernel.org/all/20260920043116.1298017-1-albin_yang@163.com/
v1: https://lore.kernel.org/all/20260909133345.1572954-1-albin_yang@163.com/
kernel/sched/stats.h | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/kernel/sched/stats.h b/kernel/sched/stats.h
index ebe0a7765f98..dc626f99ffd9 100644
--- a/kernel/sched/stats.h
+++ b/kernel/sched/stats.h
@@ -290,7 +290,7 @@ static void sched_info_arrive(struct rq *rq, struct task_struct *t)
*/
static inline void sched_info_enqueue(struct rq *rq, struct task_struct *t)
{
- if (!t->sched_info.last_queued)
+ if (!t->sched_info.last_queued && !t->se.sched_delayed)
t->sched_info.last_queued = rq_clock(rq);
}
--
2.43.7
^ permalink raw reply [flat|nested] 2+ messages in thread* Re: [PATCH v4] sched/stats: Fix run_delay over-count for migrated sched_delayed tasks
2026-10-24 4:27 [PATCH v4] sched/stats: Fix run_delay over-count for migrated sched_delayed tasks albin_yang
@ 2026-10-25 13:03 ` albin_yang
0 siblings, 0 replies; 2+ messages in thread
From: albin_yang @ 2026-10-25 13:03 UTC (permalink / raw)
To: peterz
Cc: mingo, juri.lelli, vincent.guittot, dietmar.eggemann, rostedt,
bsegall, mgorman, vschneid, kprateek.nayak, chen.yu,
kayracizmeci, linux-kernel, albin_yang, albinwyang
Hi Peter,
Following up on the proxy exec discussion from the v3 thread
(https://lore.kernel.org/all/20260922082432.2987855-1-albin_yang@163.com/):
I have now reproduced the proxy case.
A userspace mutex is futex based and never sets blocked_on, so I used a small
test module exposing a kernel mutex, with the owner pinned to a known CPU.
While the donor is blocked, the module reports:
donor: blocked-in-module
pid=10364 cpu_before=0 cpu_now=1 on_rq=1 state=2
is_blocked=1 blocked_on=1
i.e. the donor stayed on the runqueue (on_rq=1) as a real proxy donor
(is_blocked=1, blocked_on=1) and was moved to the owner's CPU while blocked
(cpu_before=0, cpu_now=1).
Measured on a 2 CPU VM, 5 iterations each, hold 200 ms (run_delay is field 2
of /proc/<tid>/schedstat):
config block run_delay delta migrated
------------------------------ ------- --------------- --------
cross-cpu, unfixed 169.9 ms 167.1 ms 5/5
cross-cpu, v4 (!sched_delayed) 169.9 ms 167.1 ms 5/5
cross-cpu, + !task_is_blocked() 169.8 ms 0.01 ms 5/5
same-cpu (control) 167.3 ms 0.00 ms 0/5
So the same ~170 ms wait is charged ~167 ms or ~0 depending only on where the
owner happens to run, and v4 alone does not change it -- the two are
independent bugs.
The gate does not disable normal accounting either: with the donor runnable
and contending for the CPU, run_delay still grows by ~39 ms over an 80 ms
window (5/5 iterations, in both the cross-cpu and the same-cpu runs). So it
suppresses only the fake enqueue, not real waits.
On your original question: t->is_blocked still cannot be the gate. In
ttwu_runnable() the real wakeup enqueue sits inside the if (p->is_blocked)
branch, while is_blocked is cleared afterwards in ttwu_do_wakeup(), so
!is_blocked would suppress the very re-arm this patch relies on.
task_is_blocked() does work, but only as an additional condition: with
!sched_delayed && !task_is_blocked() the proxy case drops from 167.1 ms to
0.01 ms, while the delayed dequeue fix is unaffected, since a delayed sleeper
has no blocked_on.
Given the proxy migration only landed in v7.1 while the delayed dequeue bug
exists from v6.12, coupling them would either fail to build
(v6.12.y..v6.16.y, no proxy exec at all) or add a dependency on code that
cannot trigger (v6.17.y..v7.0.y, proxy exec present but no blocked-waiter
migration), so I'll send the proxy one as a separate patch on top of v4.
Thanks,
Wei Yang
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2026-10-25 13:05 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2026-10-24 4:27 [PATCH v4] sched/stats: Fix run_delay over-count for migrated sched_delayed tasks albin_yang
2026-10-25 13:03 ` albin_yang
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox
all inboxes | Powered by JetHome®