On 11/9/2018 12:38 PM, Valentin Schneider wrote:
> Hi Steve,
>
> On 09/11/2018 12:50, Steve Sistare wrote:
> [...]
>> @@ -482,6 +484,10 @@ static void update_top_cache_domain(int cpu)
>> dirty_sched_domain_sysctl(cpu);
>> destroy_sched_domains(tmp);
>>
>> + sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
>> + cfs_overload_cpus = (sd ? sd->shared->cfs_overload_cpus : NULL);
>> + rcu_assign_pointer(rq->cfs_overload_cpus, cfs_overload_cpus);
>> +
>
> Why not do this in update_top_cache_domain() where we also look for the
> highest SD_SHARE_PKG_RESOURCES and setup shortcut pointers?
My snippet needs rq which is currently referenced in cpu_attach_domain() but
not in update_top_cache_domain(). I could just as easily do it in
update_top_cache_domain(). Either way is fine with me.
>> update_top_cache_domain(cpu);
>> }
>>
>> @@ -1619,9 +1625,19 @@ static void __sdt_free(const struct cpumask *cpu_map)
>> }
>> }
>>
>> +#define ZALLOC_MASK(maskp, nelems, node) \
>> + (!*(maskp) && !zalloc_sparsemask_node(maskp, nelems, \
>> + SPARSEMASK_DENSITY_DEFAULT, \
>> + GFP_KERNEL, node)) \
>> +
>> static int sd_llc_alloc(struct sched_domain *sd)
>> {
>> - /* Allocate sd->shared data here. Empty for now. */
>> + struct sched_domain_shared *sds = sd->shared;
>> + struct cpumask *span = sched_domain_span(sd);
>> + int nid = cpu_to_node(cpumask_first(span));
>> +
>> + if (ZALLOC_MASK(&sds->cfs_overload_cpus, nr_cpu_ids, nid))
>
> Mmm so this is called once on every CPU, but the !*(maskp) check in the
> macro makes it so there is only one allocation per sd_llc_shared.
>
> I wouldn't mind having that called out in a comment, or having the
> pointer check done explicitly outside of the macro.
OK, will add a comment. I like the macro because the code is cleaner if/when
multiple sets are created.
- Steve
