Date:   Thu, 30 Aug 2018 11:00:20 +0100
From:   Patrick Bellasi <patrick.bellasi@....com>
To:     Quentin Perret <quentin.perret@....com>
Cc:     peterz@...radead.org, rjw@...ysocki.net,
        linux-kernel@...r.kernel.org, linux-pm@...r.kernel.org,
        gregkh@...uxfoundation.org, mingo@...hat.com,
        dietmar.eggemann@....com, morten.rasmussen@....com,
        chris.redpath@....com, valentin.schneider@....com,
        vincent.guittot@...aro.org, thara.gopinath@...aro.org,
        viresh.kumar@...aro.org, tkjos@...gle.com, joel@...lfernandes.org,
        smuckle@...gle.com, adharmap@...eaurora.org,
        skannan@...eaurora.org, pkondeti@...eaurora.org,
        juri.lelli@...hat.com, edubezval@...il.com,
        srinivas.pandruvada@...ux.intel.com, currojerez@...eup.net,
        javi.merino@...nel.org
Subject: Re: [PATCH v6 05/14] sched/topology: Reference the Energy Model of
 CPUs when available

On 29-Aug 17:56, Quentin Perret wrote:
> On Wednesday 29 Aug 2018 at 17:22:38 (+0100), Patrick Bellasi wrote:
> > > +static void build_perf_domains(const struct cpumask *cpu_map)
> > > +{
> > > +	struct perf_domain *pd = NULL, *tmp;
> > > +	int cpu = cpumask_first(cpu_map);
> > > +	struct root_domain *rd = cpu_rq(cpu)->rd;
> > > +	int i;
> > > +
> > > +	for_each_cpu(i, cpu_map) {
> > > +		/* Skip already covered CPUs. */
> > > +		if (find_pd(pd, i))
> > > +			continue;
> > > +
> > > +		/* Create the new pd and add it to the local list. */
> > > +		tmp = pd_init(i);
> > > +		if (!tmp)
> > > +			goto free;
> > > +		tmp->next = pd;
> > > +		pd = tmp;
> > > +	}
> > > +
> > > +	perf_domain_debug(cpu_map, pd);
> > > +
> > > +	/* Attach the new list of performance domains to the root domain. */
> > > +	tmp = rd->pd;
> > > +	rcu_assign_pointer(rd->pd, pd);
> > > +	if (tmp)
> > > +		call_rcu(&tmp->rcu, destroy_perf_domain_rcu);
> > 
> > We have:
> > 
> >   sched_cpu_activate/cpuset_cpu_inactive
> >     cpuset_cpu_active/sched_cpu_deactivate
> >       partition_sched_domains
> >         build_perf_domains
> > 
> > thus here we are building new SDs and, specifically, above we are
> > attaching the local list "pd" to a _new_ root domain... thus, there
> > cannot be already users of this new SDs and root domain at this stage,
> > isn't it ?
> 
> Hmm, actually you can end up here even if the rd isn't new. That would
> happen if you call rebuild_sched_domains() after the EM has been
> registered for example.
> At this point, you might skip
> detach_destroy_domains() and build_sched_domains() from
> partition_sched_domains(), but still call build_perf_domains(), which
> would then attach the pd list to the current rd.

Ok... then it's just me that need to go back and better study how and
when SD are rebuilds.

> That's one reason why rcu_assign_pointer() is probably a good idea. And
> it's also nice from a doc standpoint I suppose.

If we can call into build_perf_domains() and reach the assignement
above with an existing RD, then yes, it makes perfect sense.

> > Do we really need that rcu_assign_pointer ?
> > Is the rcu_assign_pointer there just to "match" the following call_rcu ?
> > 
> > What about this path:
> > 
> >   sched_init_domains
> >      partition_sched_domains
> > 
> > in which case we do not call build_perf_domains... is that intended ?
> 
> I assume you meant:
> 
>    sched_init_domains
>      build_sched_domains
> 
> Is that right ?

Mmm... yes... seems so.

> If yes, I didn't bother calling build_perf_domains() from there because
> I don't think there is a single platform out there which would have a
> registered Energy Model that early in the boot process. Or maybe there
> is one I don't know ?

Dunno... but, in any case, probably we don't care about using EAS until
the boot complete, isn't it?

Just to better understand, what is the most common activation path we expect ?

  1. system boot
  2. CPUs online
  3. CPUFreq initialization
  4. EM registered
  X. ???
  N. partition_sched_domains
        build_perf_domains

IOW, who do we expect to call build_perf_domains for the first time ?

> Anyway, that is probably easy to fix, if need be.
> 
> > > +
> > > +	return;
> > > +
> > > +free:
> > > +	free_pd(pd);
> > > +	tmp = rd->pd;
> > > +	rcu_assign_pointer(rd->pd, NULL);
> > > +	if (tmp)
> > > +		call_rcu(&tmp->rcu, destroy_perf_domain_rcu);
> > > +}
> > 
> > All the above functions use different naming conventions:
> > 
> >    "_pd" suffix, "pd_" prefix and "perf_domain_" prefix.
> > 
> > and you do it like that because it better matches the corresponding
> > call sites following down the file.
> 
> That's right. The functions are supposed to vaguely look like existing
> functions dealing with sched domains.
> 
> > However, since we are into a "CONFIG_ENERGY_MODEL" guarded section,
> > why not start using a common prefix for all PD related functions?
> > 
> > I very like "perf_domain_" (or "pd_") as a prefix and I would try to
> > use it for all the functions you defined above:
> > 
> >    perf_domain_free
> >    perf_domain_find
> >    perf_domain_debug
> >    perf_domain_destroy_rcu
> >    perf_domain_build
> 
> I kinda like the idea of keeping things consistent with the existing
> code TBH. Especially because I'm terrible at naming things ... But if
> there is a general agreement that I should rename everything I won't
> argue.

I've just got the impression that naming in this file is a bit
fuzzy and it could be worth a cleanup... but of course there is also
value in minimizing the changes.

Was just wondering if a better file organization in general could help
to make topology.c more easy to compile for humans... but yes... let's
keep this for another time ;)

Cheers Patrick

-- 
#include <best/regards.h>

Patrick Bellasi

Powered by blists - more mailing lists