Messages in this thread | | | Subject | Re: [PATCH 1/4] sched/topology: Store root domain CPU capacity sum | From | Dietmar Eggemann <> | Date | Thu, 9 Apr 2020 15:50:10 +0200 |
| |
On 08.04.20 19:03, Vincent Guittot wrote: > On Wed, 8 Apr 2020 at 18:31, Dietmar Eggemann <dietmar.eggemann@arm.com> wrote: >> >> On 08.04.20 14:29, Vincent Guittot wrote: >>> On Wed, 8 Apr 2020 at 11:50, Dietmar Eggemann <dietmar.eggemann@arm.com> wrote: >> >> [...] >> >>>> /** >>>> diff --git a/kernel/sched/topology.c b/kernel/sched/topology.c >>>> index 8344757bba6e..74b0c0fa4b1b 100644 >>>> --- a/kernel/sched/topology.c >>>> +++ b/kernel/sched/topology.c >>>> @@ -2052,12 +2052,17 @@ build_sched_domains(const struct cpumask *cpu_map, struct sched_domain_attr *att >>>> /* Attach the domains */ >>>> rcu_read_lock(); >>>> for_each_cpu(i, cpu_map) { >>>> + unsigned long cap = arch_scale_cpu_capacity(i); >>> >>> Why do you replace the use of rq->cpu_capacity_orig by >>> arch_scale_cpu_capacity(i) ? >>> There is nothing about this change in the commit message >> >> True. And I can change this back. >> >> It seems though that the solution is not sufficient because of the >> 'rd->span &nsub cpu_active_mask' issue discussed under patch 2/4. >>ap >> But this remind me of another question I have. >> >> Currently we use arch_scale_cpu_capacity() more often (16 times) than >> capacity_orig_of()/rq->cpu_capacity_orig . >> >> What's hindering us to remove rq->cpu_capacity_orig and the code around >> it and solely rely on arch_scale_cpu_capacity()? I mean the arch >> implementation should be fast. > > Or we can do the opposite and only use capacity_orig_of()/rq->cpu_capacity_orig. > > Is there a case where the max cpu capacity changes over time ? So I > would prefer to use cpu_capacity_orig which is a field of scheduler > instead of always calling an external arch specific function
I see. So far it only changes during startup.
And it looks like that asym_cpu_capacity_level() [topology.c] would fail if we would use capacity_orig_of() instead of arch_scale_cpu_capacity().
post_init_entity_util_avg() [fair.c] and sugov_get_util() [cpufreq_schedutil.c] would be temporarily off until update_cpu_capacity() has updated cpu_rq(cpu)->cpu_capacity_orig.
compute_energy() [fair.c] is guarded by sched_energy_enabled() from being used at startup.
scale_rt_capacity() could be changed in case we call it after the cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu) in update_cpu_capacity().
The Energy Model (and CPUfreq cooling) code would need capacity_orig_of() exported. arch_scale_cpu_capacity() currently is exported via include/linux/sched/topology.h.
I guess Pelt and 'scale invariant Deadline bandwidth enforcement' should continue using arch_scale_cpu_capacity() in sync with arch_scale_freq_capacity().
IMHO it's hard to give clear advice when to use the one or the other.
We probably don't want to set cpu_rq(cpu)->cpu_capacity_orig in the arch cpu scale setter. We have arch_scale_cpu_capacity() to decouple that.
| |