Re: [PATCH 01/14] sched: Employ sched_change guards

[Peter Zijlstra <peterz@infradead.org>]

On Thu, Sep 11, 2025 at 11:55:23AM +0200, Peter Zijlstra wrote:
> On Thu, Sep 11, 2025 at 02:36:21PM +0530, K Prateek Nayak wrote:
> > Hello Peter,
> > 
> > On 9/10/2025 9:14 PM, Peter Zijlstra wrote:
> > > @@ -9240,8 +9213,9 @@ static void sched_change_group(struct ta
> > >   */
> > >  void sched_move_task(struct task_struct *tsk, bool for_autogroup)
> > >  {
> > > -	int queued, running, queue_flags =
> > > +	unsigned int queue_flags =
> > >  		DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
> > 
> > nit.
> > 
> > Since we don't do a complete dequeue for delayed task in
> > sched_move_task(), can we get rid of that DEQUEUE_NOCLOCK and ...
> > 
> > > +	bool resched = false;
> > >  	struct rq *rq;
> > >  
> > >  	CLASS(task_rq_lock, rq_guard)(tsk);
> > > @@ -9249,28 +9223,12 @@ void sched_move_task(struct task_struct
> > >  
> > >  	update_rq_clock(rq);
> > 
> > ... this clock update and instead rely on sched_change_begin() to
> > handle it within the guard?
> 
> Yeah, I suppose we could. But let me try and do that in a later patch,
> on-top of all this.

Something like so?

---
 core.c     |   33 +++++++++++----------------------
 ext.c      |   12 ++++--------
 syscalls.c |    4 +---
 3 files changed, 16 insertions(+), 33 deletions(-)

--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -2359,10 +2359,8 @@ static void migrate_disable_switch(struc
 	if (p->cpus_ptr != &p->cpus_mask)
 		return;
 
-	scoped_guard (task_rq_lock, p) {
-		update_rq_clock(scope.rq);
+	scoped_guard (task_rq_lock, p)
 		do_set_cpus_allowed(p, &ac);
-	}
 }
 
 void migrate_disable(void)
@@ -2716,9 +2714,7 @@ void set_cpus_allowed_common(struct task
 static void
 do_set_cpus_allowed(struct task_struct *p, struct affinity_context *ctx)
 {
-	u32 flags = DEQUEUE_SAVE | DEQUEUE_NOCLOCK | DEQUEUE_LOCKED;
-
-	scoped_guard (sched_change, p, flags) {
+	scoped_guard (sched_change, p, DEQUEUE_SAVE | DEQUEUE_LOCKED) {
 		p->sched_class->set_cpus_allowed(p, ctx);
 		mm_set_cpus_allowed(p->mm, ctx->new_mask);
 	}
@@ -2740,10 +2736,8 @@ void set_cpus_allowed_force(struct task_
 		struct rcu_head rcu;
 	};
 
-	scoped_guard (__task_rq_lock, p) {
-		update_rq_clock(scope.rq);
+	scoped_guard (__task_rq_lock, p)
 		do_set_cpus_allowed(p, &ac);
-	}
 
 	/*
 	 * Because this is called with p->pi_lock held, it is not possible
@@ -9159,16 +9153,13 @@ static void sched_change_group(struct ta
  */
 void sched_move_task(struct task_struct *tsk, bool for_autogroup)
 {
-	unsigned int queue_flags =
-		DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK | DEQUEUE_LOCKED;
+	unsigned int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_LOCKED;
 	bool resched = false;
 	struct rq *rq;
 
 	CLASS(task_rq_lock, rq_guard)(tsk);
 	rq = rq_guard.rq;
 
-	update_rq_clock(rq);
-
 	scoped_guard (sched_change, tsk, queue_flags) {
 		sched_change_group(tsk);
 		if (!for_autogroup)
@@ -10852,19 +10843,17 @@ struct sched_change_ctx *sched_change_be
 	}
 #endif
 
+	if (!(flags & DEQUEUE_NOCLOCK)) {
+		update_rq_clock(rq);
+		flags |= DEQUEUE_NOCLOCK;
+	}
+
 	if (flags & DEQUEUE_CLASS) {
 		if (WARN_ON_ONCE(flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)))
 			flags &= ~(DEQUEUE_SAVE | DEQUEUE_MOVE);
 
-		if (p->sched_class->switching_from) {
-			/*
-			 * switching_from_fair() assumes CLASS implies NOCLOCK;
-			 * fixing this assumption would mean switching_from()
-			 * would need to be able to change flags.
-			 */
-			WARN_ON(!(flags & DEQUEUE_NOCLOCK));
+		if (p->sched_class->switching_from)
 			p->sched_class->switching_from(rq, p);
-		}
 	}
 
 	*ctx = (struct sched_change_ctx){
@@ -10915,7 +10904,7 @@ void sched_change_end(struct sched_chang
 		p->sched_class->switching_to(rq, p);
 
 	if (ctx->queued)
-		enqueue_task(rq, p, ctx->flags | ENQUEUE_NOCLOCK);
+		enqueue_task(rq, p, ctx->flags);
 	if (ctx->running)
 		set_next_task(rq, p, ctx->flags);
 
--- a/kernel/sched/ext.c
+++ b/kernel/sched/ext.c
@@ -5018,14 +5018,12 @@ static void scx_disable_workfn(struct kt
 
 	scx_task_iter_start(&sti);
 	while ((p = scx_task_iter_next_locked(&sti))) {
-		unsigned int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE |
-					   DEQUEUE_NOCLOCK | DEQUEUE_LOCKED;
+		unsigned int queue_flags =
+			DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_LOCKED;
 		const struct sched_class *old_class = p->sched_class;
 		const struct sched_class *new_class =
 			__setscheduler_class(p->policy, p->prio);
 
-		update_rq_clock(task_rq(p));
-
 		if (old_class != new_class) {
 			queue_flags |= DEQUEUE_CLASS;
 			queue_flags &= ~(DEQUEUE_SAVE | DEQUEUE_MOVE);
@@ -5763,8 +5761,8 @@ static int scx_enable(struct sched_ext_o
 	percpu_down_write(&scx_fork_rwsem);
 	scx_task_iter_start(&sti);
 	while ((p = scx_task_iter_next_locked(&sti))) {
-		unsigned int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE |
-					   DEQUEUE_NOCLOCK | DEQUEUE_LOCKED;
+		unsigned int queue_flags =
+			DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_LOCKED;
 		const struct sched_class *old_class = p->sched_class;
 		const struct sched_class *new_class =
 			__setscheduler_class(p->policy, p->prio);
@@ -5772,8 +5770,6 @@ static int scx_enable(struct sched_ext_o
 		if (!tryget_task_struct(p))
 			continue;
 
-		update_rq_clock(task_rq(p));
-
 		if (old_class != new_class) {
 			queue_flags |= DEQUEUE_CLASS;
 			queue_flags &= ~(DEQUEUE_SAVE | DEQUEUE_MOVE);
--- a/kernel/sched/syscalls.c
+++ b/kernel/sched/syscalls.c
@@ -76,8 +76,6 @@ void set_user_nice(struct task_struct *p
 	CLASS(task_rq_lock, rq_guard)(p);
 	rq = rq_guard.rq;
 
-	update_rq_clock(rq);
-
 	/*
 	 * The RT priorities are set via sched_setscheduler(), but we still
 	 * allow the 'normal' nice value to be set - but as expected
@@ -89,7 +87,7 @@ void set_user_nice(struct task_struct *p
 		return;
 	}
 
-	scoped_guard (sched_change, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK | DEQUEUE_LOCKED) {
+	scoped_guard (sched_change, p, DEQUEUE_SAVE | DEQUEUE_LOCKED) {
 		p->static_prio = NICE_TO_PRIO(nice);
 		set_load_weight(p, true);
 		old_prio = p->prio;