Re: [PATCH 0/10 v2] sched/fair: Fix statistics with delayed dequeue

[K Prateek Nayak <kprateek.nayak@amd.com>]

Hello Vincent, Mike,

On 12/2/2024 2:47 PM, Vincent Guittot wrote:
> On Sun, 1 Dec 2024 at 14:30, Mike Galbraith <efault@gmx.de> wrote:
>>
>> Greetings,
>>
>> On Fri, 2024-11-29 at 17:17 +0100, Vincent Guittot wrote:
>>> Delayed dequeued feature keeps a sleeping sched_entitiy enqueued until its
>>> lag has elapsed. As a result, it stays also visible in the statistics that
>>> are used to balance the system and in particular the field h_nr_running.
>>>
>>> This serie fixes those metrics by creating a new h_nr_queued that tracks
>>> all queued tasks. It renames h_nr_running into h_nr_runnable and restores
>>> the behavior of h_nr_running i.e. tracking the number of fair tasks that
>>>   want to run.
>>>
>>> h_nr_runnable is used in several places to make decision on load balance:
>>>    - PELT runnable_avg
>>>    - deciding if a group is overloaded or has spare capacity
>>>    - numa stats
>>>    - reduced capacity management
>>>    - load balance between groups
>>
>> I took the series for a spin in tip v6.12-10334-gb1b238fba309, but
>> runnable seems to have an off-by-one issue, causing it to wander ever
>> further south.
>>
>> patches 1-3 applied.
>>    .h_nr_runnable                 : -3046
>>    .runnable_avg                  : 450189777126
> 
> Yeah, I messed up something around finish_delayed_dequeue_entity().
> I'm' going to prepare a v3

I was looking into this and I have the below diff so far that seems to
solve the post boot negative values of h_nr_runnable on my setup; it is
only lightly tested so far:

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 87552870958c..423981e65aba 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -5464,6 +5464,10 @@ static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
  static void set_delayed(struct sched_entity *se)
  {
  	se->sched_delayed = 1;
+
+	if (!entity_is_task(se))
+		return;
+
  	for_each_sched_entity(se) {
  		struct cfs_rq *cfs_rq = cfs_rq_of(se);
  
@@ -5476,6 +5480,10 @@ static void set_delayed(struct sched_entity *se)
  static void clear_delayed(struct sched_entity *se)
  {
  	se->sched_delayed = 0;
+
+	if (!entity_is_task(se))
+		return;
+
  	for_each_sched_entity(se) {
  		struct cfs_rq *cfs_rq = cfs_rq_of(se);
  
@@ -6977,7 +6985,7 @@ enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
  	struct cfs_rq *cfs_rq;
  	struct sched_entity *se = &p->se;
  	int h_nr_idle = task_has_idle_policy(p);
-	int h_nr_runnable = 0;
+	int h_nr_runnable = 1;
  	int task_new = !(flags & ENQUEUE_WAKEUP);
  	int rq_h_nr_queued = rq->cfs.h_nr_queued;
  	u64 slice = 0;
@@ -7124,8 +7132,7 @@ static int dequeue_entities(struct rq *rq, struct sched_entity *se, int flags)
  		p = task_of(se);
  		h_nr_queued = 1;
  		h_nr_idle = task_has_idle_policy(p);
-		if (!task_sleep && !task_delayed)
-			h_nr_runnable = !se->sched_delayed;
+		h_nr_runnable = !se->sched_delayed;
  	} else {
  		cfs_rq = group_cfs_rq(se);
  		slice = cfs_rq_min_slice(cfs_rq);
diff --git a/kernel/sched/idle.c b/kernel/sched/idle.c
index ab911d1335ba..f4ef5aaa4674 100644
--- a/kernel/sched/idle.c
+++ b/kernel/sched/idle.c
@@ -457,6 +457,7 @@ static void put_prev_task_idle(struct rq *rq, struct task_struct *prev, struct t
  
  static void set_next_task_idle(struct rq *rq, struct task_struct *next, bool first)
  {
+	SCHED_WARN_ON(rq->cfs.h_nr_runnable);
  	update_idle_core(rq);
  	scx_update_idle(rq, true);
  	schedstat_inc(rq->sched_goidle);
--

I'm not sure if the change in dequeue_entities() is completely necessary
but I added it after seeing the (DEQUEUE_SLEEP | DEQUEUE_DELAYED) in
throttle_cfs_rq() but there the entity cannot possibly be a task so
perhaps that part is unnecessary ¯\_(ツ)_/¯

Still testing! Will keep an eye out for v3.

> 
>>
>> full set applied.
>>    .h_nr_runnable                 : -5707
>>    .runnable_avg                  : 4391793519526
>>
>>          -Mike

-- 
Thanks and Regards,
Prateek