[PATCH] sched/core: Preserve wake flags across queued wakeups

[Shubhang Kaushik (Ampere)]

The queued wakeup path only remembers whether the wakee migrated. When
the target CPU drains the wakelist, sched_ttwu_pending() turns that back
into either WF_MIGRATED or 0 and passes it to ttwu_do_activate().

That drops the rest of the wake flags seen by the direct wakeup path. In
particular, WF_SYNC, WF_TTWU and WF_RQ_SELECTED can still matter for
wakeup preemption, so queued wakeups can end up making a different
decision from a direct wakeup.

Store the relevant wake_flags bits instead of a single boolean and pass
them through when the wakelist is drained. Keep WF_CURRENT_CPU out of
the saved mask as it is a placement hint for CPU selection, not an
activation flag.

Signed-off-by: Shubhang Kaushik (Ampere) <sh@xxxxxxxxxx>
---
This aims to preserve activation-relevant wake flags across queued
remote wakeups, keeping the queued and direct wakeup paths consistent when
calling ttwu_do_activate().

Tested on an Ampere Altra 80 CPU system with perf bench sched messaging and
hackbench. Results were within run-to-run noise of the unpatched mainline
baseline.

Baseline: mainline origin/master at d96fcfe1b7f9.
---
 include/linux/sched.h | 6 +++---
 kernel/sched/core.c   | 9 +++++++--
 2 files changed, 10 insertions(+), 5 deletions(-)

diff --git a/include/linux/sched.h b/include/linux/sched.h
index 373bcc0598d10b4256a11f8c8373ece78fa5e0e5..5d55cdf696c13ed1c95869a1b6fb4dc1485727dc 100644
--- a/include/linux/sched.h
+++ b/include/linux/sched.h
@@ -1003,12 +1003,12 @@ struct task_struct {
 	 * schedule()			  if (p->on_rq && ..) // false
 	 *   smp_mb__after_spinlock();	  if (smp_load_acquire(&p->on_cpu) && //true
 	 *   deactivate_task()		      ttwu_queue_wakelist())
-	 *     p->on_rq = 0;			p->sched_remote_wakeup = Y;
+	 *     p->on_rq = 0;			p->sched_remote_wakeup_flags = Y;
 	 *
 	 * guarantees all stores of 'current' are visible before
-	 * ->sched_remote_wakeup gets used, so it can be in this word.
+	 * ->sched_remote_wakeup_flags gets used, so it can be in this word.
 	 */
-	unsigned			sched_remote_wakeup:1;
+	unsigned			sched_remote_wakeup_flags:8;
 #ifdef CONFIG_RT_MUTEXES
 	unsigned			sched_rt_mutex:1;
 #endif
diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index 96226707c2f6135341aa779b8262f113e103d8ad..0d2918b66be277ec09109022bf3ede4b9ec8220c 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -3902,13 +3902,17 @@ void sched_ttwu_pending(void *arg)
 	update_rq_clock(rq);
 
 	llist_for_each_entry_safe(p, t, llist, wake_entry.llist) {
+		int wake_flags;
+
 		if (WARN_ON_ONCE(p->on_cpu))
 			smp_cond_load_acquire(&p->on_cpu, !VAL);
 
 		if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq)))
 			set_task_cpu(p, cpu_of(rq));
 
-		ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
+		wake_flags = p->sched_remote_wakeup_flags;
+		p->sched_remote_wakeup_flags = 0;
+		ttwu_do_activate(rq, p, wake_flags, &rf);
 	}
 
 	/*
@@ -3951,7 +3955,8 @@ static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags
 {
 	struct rq *rq = cpu_rq(cpu);
 
-	p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
+	p->sched_remote_wakeup_flags = wake_flags &
+		(WF_TTWU | WF_SYNC | WF_MIGRATED | WF_RQ_SELECTED);
 
 	WRITE_ONCE(rq->ttwu_pending, 1);
 #ifdef CONFIG_SMP

---
base-commit: d96fcfe1b7f94ac742984ae7986b94a116abff1b
change-id: 20260710-b4-sched-ttwu-wake-flags-e87ada0e65e1

Best regards,
-- 
Shubhang Kaushik (Ampere) <sh@xxxxxxxxxx>