Re: [PATCH v2 06/22] sched/fair: Collect load-balancing stats for IPC classes

[Ricardo Neri]

On Wed, Dec 07, 2022 at 06:00:32PM +0100, Dietmar Eggemann wrote:
> On 28/11/2022 14:20, Ricardo Neri wrote:
> 
> [...]
> 
> > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > index 224107278471..3a1d6c50a19b 100644
> > --- a/kernel/sched/fair.c
> > +++ b/kernel/sched/fair.c
> > @@ -9100,6 +9100,57 @@ group_type group_classify(unsigned int imbalance_pct,
> >  	return group_has_spare;
> >  }
> >  
> > +struct sg_lb_ipcc_stats {
> > +	int min_score;	/* Min(score(rq->curr->ipcc)) */
> > +	int min_ipcc;	/* Min(rq->curr->ipcc) */
> > +	long sum_score; /* Sum(score(rq->curr->ipcc)) */
> > +};
> 
> Wouldn't it be cleaner to put `min_score`, `min_ipcc` and `sum_score`
> into `struct sg_lb_stats` next to `ipcc_score_{after, before}` under the
> same #ifdef CONFIG_IPC_CLASSES?

Yes, that is a good observation. I initially wanted to hide these
intermediate and only expose the end result ipcc_score_{after, before} to
struct sg_lb_stats. I agree, it would look cleaner as you suggest.
> 
> Looks like those IPCC stats would only be needed in the specific
> condition under which update_sg_lb_stats_scores() is called?

True.

> 
> > +#ifdef CONFIG_IPC_CLASSES
> > +static void init_rq_ipcc_stats(struct sg_lb_ipcc_stats *sgcs)
> > +{
> > +	*sgcs = (struct sg_lb_ipcc_stats) {
> > +		.min_score = INT_MAX,
> > +	};
> > +}
> > +
> > +/** Called only if cpu_of(@rq) is not idle and has tasks running. */
> > +static void update_sg_lb_ipcc_stats(struct sg_lb_ipcc_stats *sgcs,
> > +				    struct rq *rq)
> > +{
> > +	struct task_struct *curr;
> > +	unsigned short ipcc;
> > +	int score;
> > +
> > +	if (!sched_ipcc_enabled())
> > +		return;
> > +
> > +	curr = rcu_dereference(rq->curr);
> > +	if (!curr || (curr->flags & PF_EXITING) || is_idle_task(curr))
> 
> So the Idle task is excluded but RT, DL, (Stopper) tasks are not. Looks
> weird if non-CFS tasks could influence CFS load-balancing.
> AFAICS, RT and DL tasks could have p->ipcc != IPC_CLASS_UNCLASSIFIED?

Agreed. Perhaps I can also check for !(task_is_realtime()), which see
seems to cover all these cases.
> 
> [...]