Re: [PATCH v9 09/15] sched: Introduce sched_energy_present static key

[Quentin Perret]

On Thursday 22 Nov 2018 at 09:17:16 (+0000), Quentin Perret wrote:
> On Wednesday 21 Nov 2018 at 15:14:45 (+0000), Quentin Perret wrote:
> > Hi Peter,
> > 
> > On Wednesday 21 Nov 2018 at 14:08:22 (+0100), Peter Zijlstra wrote:
> > > On Mon, Nov 19, 2018 at 02:18:51PM +0000, Quentin Perret wrote:
> > > 
> > > > +static void sched_energy_start(int ndoms_new, cpumask_var_t doms_new[])
> > > > +{
> > > > +	/*
> > > > +	 * The conditions for EAS to start are checked during the creation of
> > > > +	 * root domains. If one of them meets all conditions, it will have a
> > > > +	 * non-null list of performance domains.
> > > > +	 */
> > > > +	while (ndoms_new) {
> > > > +		if (cpu_rq(cpumask_first(doms_new[ndoms_new - 1]))->rd->pd)
> > > > +			goto enable;
> > > > +		ndoms_new--;
> > > > +	}
> > > 
> > > That seems quite ugly; can't you simply return a bool from
> > > build_perf_domains() ?
> > > 
> > > Something like the below, but less fugly ;-)
> > > 
> > > --- a/kernel/sched/topology.c
> > > +++ b/kernel/sched/topology.c
> > > @@ -299,7 +299,7 @@ static void destroy_perf_domain_rcu(stru
> > >  #define EM_MAX_COMPLEXITY 2048
> > >  
> > >  extern struct cpufreq_governor schedutil_gov;
> > > -static void build_perf_domains(const struct cpumask *cpu_map)
> > > +static bool build_perf_domains(const struct cpumask *cpu_map)
> > >  {
> > >  	int i, nr_pd = 0, nr_cs = 0, nr_cpus = cpumask_weight(cpu_map);
> > >  	struct perf_domain *pd = NULL, *tmp;
> > > @@ -365,7 +365,7 @@ static void build_perf_domains(const str
> > >  	if (tmp)
> > >  		call_rcu(&tmp->rcu, destroy_perf_domain_rcu);
> > >  
> > > -	return;
> > > +	return !!pd;
> > >  
> > >  free:
> > >  	free_pd(pd);
> > > @@ -373,6 +373,8 @@ static void build_perf_domains(const str
> > >  	rcu_assign_pointer(rd->pd, NULL);
> > >  	if (tmp)
> > >  		call_rcu(&tmp->rcu, destroy_perf_domain_rcu);
> > > +
> > > +	return false;
> > >  }
> > >  #else
> > >  static void free_pd(struct perf_domain *pd) { }
> > > @@ -2173,6 +2175,9 @@ void partition_sched_domains(int ndoms_n
> > >  	}
> > >  
> > >  #if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
> > > +	{
> > > +	bool has_eas = false;
> > 
> > That one wants to be declared '__maybe_unused' at the top of
> > partition_sched_domains() perhaps ? Just to avoid the { } section.
> > Hopefully the compiler will be smart enough to not use stack space for
> > it if not needed.
> > 
> > > +
> > >  	/* Build perf. domains: */
> > >  	for (i = 0; i < ndoms_new; i++) {
> > >  		for (j = 0; j < n && !sched_energy_update; j++) {
> > > @@ -2181,10 +2186,13 @@ void partition_sched_domains(int ndoms_n
> > >  				goto match3;
> > 
> > And we want to do 'has_eas = true' just before 'goto match3' here.
> > Otherwise we'll end up disabling EAS if we hit 'goto match3' for all
> > root domains.
> > 
> > >  		}
> > >  		/* No match - add perf. domains for a new rd */
> > > -		build_perf_domains(doms_new[i]);
> > > +		has_eas |= build_perf_domains(doms_new[i]);
> > >  match3:
> > >  		;
> > >  	}
> > > +
> > > +	sched_energy_set(has_eas);
> > > +	}
> > >  #endif
> > >  
> > >  	/* Remember the new sched domains: */
> > 
> > Other than that I guess that should work OK :-)
> 
> So I actually just came across your patch that does
> static_branch_{inc,dec} on sched_smt_present. Maybe I could do something
> similar here ?
> 
> How about something like the below (totally untested):
> 
> diff --git a/kernel/sched/topology.c b/kernel/sched/topology.c
> index 8f9dfea60344..10ac32a0dc2e 100644
> --- a/kernel/sched/topology.c
> +++ b/kernel/sched/topology.c
> @@ -298,6 +298,7 @@ static void destroy_perf_domain_rcu(struct rcu_head *rp)
>  
>         pd = container_of(rp, struct perf_domain, rcu);
>         free_pd(pd);
> +       static_branch_dec_cpuslocked(&sched_energy_present);
>  }
>  
>  static void sched_energy_start(int ndoms_new, cpumask_var_t doms_new[])
> @@ -421,6 +422,8 @@ static void build_perf_domains(const struct cpumask *cpu_map)
>         /* Attach the new list of performance domains to the root domain. */
>         tmp = rd->pd;
>         rcu_assign_pointer(rd->pd, pd);
> +       static_branch_inc_cpuslocked(&sched_energy_present);
> +
>         if (tmp)
>                 call_rcu(&tmp->rcu, destroy_perf_domain_rcu);
>  
> @@ -2246,7 +2249,6 @@ void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
>  match3:
>                 ;
>         }
> -       sched_energy_start(ndoms_new, doms_new);
>  #endif
>  
>         /* Remember the new sched domains: */
> 

Hmm, I went too fast, that's totally broken. But there's still something
we can do with static_branch_{inc,dec} I think. I'll come back later
with a better solution.

Sorry for the noise :/