Re: [PATCH v5 2/3] sched/topology: Rework CPU capacity asymmetry detection

[Beata Michalska <beata.michalska@arm.com>]

On Tue, May 25, 2021 at 01:59:30PM +0200, Dietmar Eggemann wrote:
> On 25/05/2021 11:30, Beata Michalska wrote:
> > On Tue, May 25, 2021 at 10:25:36AM +0200, Dietmar Eggemann wrote:
> >> On 24/05/2021 12:16, Beata Michalska wrote:
> 
> [...]
> 
> >>> @@ -1266,6 +1266,112 @@ static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
> >>>  	update_group_capacity(sd, cpu);
> >>>  }
> >>>  
> >>> +/**
> >>> + * Asymmetric CPU capacity bits
> >>> + */
> >>> +struct asym_cap_data {
> >>> +	struct list_head link;
> >>> +	unsigned long    capacity;
> >>> +	struct cpumask   *cpu_mask;
> >>
> >> Not sure if this has been discussed already but shouldn't the flexible
> >> array members` approach known from struct sched_group, struct
> >> sched_domain or struct em_perf_domain be used here?
> >> IIRC the last time this has been discussed in this thread:
> >> https://lkml.kernel.org/r/20200910054203.525420-2-aubrey.li@intel.com
> >>
> > If I got right the discussion you have pointed to, it was about using
> > cpumask_var_t which is not the case here. I do not mind moving the code
> > to use the array but I am not sure if this changes much. Looking at the
> > code changes to support that (to_cpumask namely) it was introduced for
> > cases where cpumask_var_t was not appropriate, which again isn't the case
> > here.
> 
> Yeah, it was more about using `flexible array members` or allocating the
> cpumask separately.
> 
> Looks like you're using some kind of a mixed approach:
> 
> (1) struct asym_cap_data {
>         ...
>         struct cpumask *cpu_mask;
> 
> (2) entry = kzalloc(sizeof(*entry) + cpumask_size(), GFP_KERNEL);
> 
> (3) entry->cpu_mask = (struct cpumask *)((char *)entry +
>                       sizeof(*entry));
> 
> (4) cpumask_intersects(foo, entry->cpu_mask)
> 
> 
> E.g. struct em_perf_domain has
> 
> (1)  struct em_perf_domain {
>          ...
>          unsigned long cpus[];
> 
> (2) like yours
> 
> (3) is not needed.
> 
> (4) cpumask_copy(em_span_cpus(pd), foo)
> 
>     with #define em_span_cpus(em) (to_cpumask((em)->cpus))
> 
> IMHO, it's better to keep this approach aligned between the different
> data structures.
I would actually go the other way round as it seems more 'clean'
that way and it does not need the conversion but I don't mind playing along.

---
BR
B.