Re: [PATCH 1/1] sched/pelt: Refine the enqueue_load_avg calculate method

[Kuyo Chang]

On Mon, 2022-04-11 at 10:39 +0200, Vincent Guittot wrote:
> On Mon, 11 Apr 2022 at 08:17, Kuyo Chang <kuyo.chang@xxxxxxxxxxxx>
> wrote:
> > 
> > From: kuyo chang <kuyo.chang@xxxxxxxxxxxx>
> > 
> > I meet the warning message at cfs_rq_is_decayed at below code.
> > 
> > SCHED_WARN_ON(cfs_rq->avg.load_avg ||
> >                     cfs_rq->avg.util_avg ||
> >                     cfs_rq->avg.runnable_avg)
> > 
> > Following is the calltrace.
> > 
> > Call trace:
> > __update_blocked_fair
> > update_blocked_averages
> > newidle_balance
> > pick_next_task_fair
> > __schedule
> > schedule
> > pipe_read
> > vfs_read
> > ksys_read
> > 
> > After code analyzing and some debug messages, I found it exits a
> > corner
> > case at attach_entity_load_avg which will cause load_sum is zero
> > and
> > load_avg is not.
> > Consider se_weight is 88761 according by sched_prio_to_weight
> > table.
> > And assume the get_pelt_divider() is 47742, se->avg.load_avg is 1.
> > By the calculating for se->avg.load_sum as following will become
> > zero
> > as following.
> > se->avg.load_sum =
> >         div_u64(se->avg.load_avg * se->avg.load_sum,
> > se_weight(se));
> > se->avg.load_sum = 1*47742/88761 = 0.
> 
> The root problem is there, se->avg.load_sum must not be null if
> se->avg.load_avg is not null because the correct relation between
> _avg
> and _sum is:
> 
> load_avg = weight * load_sum / divider.
> 
> so the fix should be attach_entity_load_avg() and probably the below
> is enough
> 
> se->avg.load_sum = div_u64(se->avg.load_avg * se->avg.load_sum,
> se_weight(se)) + 1;

Thanks for your kindly suggestion.
+1 would make the calcuation for load_sum may be overestimate?
How about the below code make sense for fix the corner case?

--- 
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -3832,7 +3832,8 @@ static void attach_entity_load_avg(struct cfs_rq
*cfs_rq, struct sched_entity *s
 	se->avg.load_sum = divider;
 	if (se_weight(se)) {
 		se->avg.load_sum =
-			div_u64(se->avg.load_avg * se->avg.load_sum,
se_weight(se));
+			(se->avg.load_avg * se->avg.load_sum >
se_weight(se)) ?
+			div_u64(se->avg.load_avg * se->avg.load_sum,
se_weight(se)) : 1;
 	}
 
 	enqueue_load_avg(cfs_rq, se);
-- 
2.18.0


> > 
> > After enqueue_load_avg code as below.
> > cfs_rq->avg.load_avg += se->avg.load_avg;
> > cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
> > 
> > Then the load_sum for cfs_rq will be 1 while the load_sum for
> > cfs_rq is 0.
> > So it will hit the warning message.
> > 
> > After all, I refer the following commit patch to do the similar
> > thing at
> > enqueue_load_avg.
> > sched/pelt: Relax the sync of load_sum with load_avg
> > 
> > After long time testing, the kernel warning was gone and the system
> > runs
> > as well as before.
> > 
> > Signed-off-by: kuyo chang <kuyo.chang@xxxxxxxxxxxx>
> > ---
> >  kernel/sched/fair.c | 6 ++++--
> >  1 file changed, 4 insertions(+), 2 deletions(-)
> > 
> > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > index d4bd299d67ab..30d8b6dba249 100644
> > --- a/kernel/sched/fair.c
> > +++ b/kernel/sched/fair.c
> > @@ -3074,8 +3074,10 @@ account_entity_dequeue(struct cfs_rq
> > *cfs_rq, struct sched_entity *se)
> >  static inline void
> >  enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
> >  {
> > -       cfs_rq->avg.load_avg += se->avg.load_avg;
> > -       cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
> > +       add_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
> > +       add_positive(&cfs_rq->avg.load_sum, se_weight(se) * se-
> > >avg.load_sum);
> > +       cfs_rq->avg.load_sum = max_t(u32, cfs_rq->avg.load_sum,
> > +                                         cfs_rq->avg.load_avg *
> > PELT_MIN_DIVIDER);
> >  }
> > 
> >  static inline void
> > --
> > 2.18.0
> >