[PATCH] arm64: topology: Cleanup init_amu_fie() a bit

Ionela Voinescu ionela.voinescu at arm.com
Thu Dec 10 08:26:39 EST 2020


On Thursday 10 Dec 2020 at 18:04:39 (+0530), Viresh Kumar wrote:
> On 10-12-20, 11:29, Ionela Voinescu wrote:
> > On Thursday 10 Dec 2020 at 16:25:06 (+0530), Viresh Kumar wrote:
> > > - But right after that we call static_branch_disable() if we aren't
> > >   invariant (call to topology_scale_freq_invariant()), and this will
> > >   happen if amu_fie_cpus doesn't have all the CPUs there. Isn't it? So
> > >   partial amu support is already disallowed, without cpufreq.
> > > 
> > 
> > This is the point that needs clarification:
> > 
> > topology_scale_freq_invariant()) = cpufreq_supports_freq_invariance() ||
> >                                    arch_freq_counters_available(cpu_online_mask);
> > 
> > This checks if the full system is invariant.
> > 
> > The possible scenarios are:
> > 
> >  - All online CPUs support AMUs - arch_freq_counters_available() will
> >    return true -> topology_scale_freq_invariant() will return true.
> > 
> >  - None of the CPUs support AMUs, or part of the CPUs support AMUs - the
> >    system is invariant only if cpufreq is invariant (dependent on
> >    whether the driver implements the proper callbacks that results in
> >    calling arch_set_freq_scale() in cpufreq core.
> > 
> >  - Either cpufreq does not support invariance or we don't have AMU
> >    support on all CPUs -> the system is not invariant so we disable
> >    the AMU static key that guards the calls to
> >    topology_scale_freq_tick() - we would not want to set a scale factor
> >    for only a part of the CPUs.
> > 
> > So whether were are or are not invariant does not depend only on the AMU
> > presence, but also on the cpufreq support for invariance. We have to
> > disable invariance altogether (including the AMU guarding static key)
> > if the system is not invariant (it no all CPUs have means to provide the
> > scale).
> 
> Okay, I think I mis-assumed that amu_fie_cpus will get set by
> enable_policy_freq_counters() even for CPUs where AMU support isn't
> there, it won't be though.
> 
> Having said that, this patch, along with the minor suggestion in the
> commit log, still stands fine, right ? The other patch which I sent is
> probably incorrect due to the above assumption I had.
> 

Yes, this one is good, although I would vote for the commit message
implementation. I'll wait for v2 for reviewed-by, in case you want to
push something for the second patch in the same series.

Ionela.

> -- 
> viresh



More information about the linux-arm-kernel mailing list