Re: [RFC PATCH v3 05/10] sched/topology: Reference the Energy Model of CPUs when available
From: Quentin Perret
Date: Tue Jun 19 2018 - 13:13:39 EST
On Tuesday 19 Jun 2018 at 18:20:42 (+0200), Peter Zijlstra wrote:
> Right, so I would not do that many things at once. Also be more explicit
> about what data structure, and why.
OK, I can probably split that patch in two smaller patches. One that
introduces and enables the static_key (or something else to replace it,
see my reply below); and another one to create the list of frequency
domains.
> That said, I think the whole for_each_freq_domain() thing as done is
> broken. You've completely ignored the arguments to
> partition_sched_domains(). What happens if you create partitions right
> along the frequency domains?
>
> So you really want an argument to for_each_freq_domain() to indicate
> who's frequency domains you want to iterate. And then I think it's
> easiest if you hook into build_sched_domains() instead, because you
> really want a list per root_domain I suspect
Hmm right, the current implementation is broken with multiple root
domains... We rarely use that in mobile but the code should work
regardless I suppose.
I agree having one list per root_domain should be better. I'll change that
in v4. But I also think the idea of a global static_key is broken then.
We need a per root_domain thing as well. The SD_ASYM_CPUCAPACITY flag
might be set on one hierarchy and not the other for ex. Not sure if
attaching a static_key to each root_domain really makes sense though ...
Would replacing the static_key by a flag attached to the root_domain be
reasonable ? The CONFIG_ENERGY_MODEL config option could help us make
sure there is no performance impact when EAS can't be used with something
like this:
static bool sched_energy_enabled(struct root_domain *rd)
{
#ifdef CONFIG_ENERGY_MODEL
return rd->eas_enabled;
#else
return false;
#endif
}
Or maybe a sched_feat instead of that ifdefery ?
> (and an empty list if there is but one entry on).
Sorry but I didn't understand that ...
Thanks,
Quentin