Re: [PATCH v7 1/4] cpufreq: Introduce an optional cpuinfo_avg_freq sysfs entry

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

On Tue, Oct 29, 2024 at 12:31:11PM +0100, Rafael J. Wysocki wrote:
> On Tue, Oct 29, 2024 at 8:04 AM Viresh Kumar <viresh.kumar@linaro.org> wrote:
> >
> > Apologies for the delay from my side. September was mostly holidays
> > for me and then I was stuck with other stuff plus email backlog and
> > this series was always a painful point to return to :(
> >
> > On 13-09-24, 14:29, Beata Michalska wrote:
> > > Currently the CPUFreq core exposes two sysfs attributes that can be used
> > > to query current frequency of a given CPU(s): namely cpuinfo_cur_freq
> > > and scaling_cur_freq. Both provide slightly different view on the
> > > subject and they do come with their own drawbacks.
> > >
> > > cpuinfo_cur_freq provides higher precision though at a cost of being
> > > rather expensive. Moreover, the information retrieved via this attribute
> > > is somewhat short lived as frequency can change at any point of time
> > > making it difficult to reason from.
> > >
> > > scaling_cur_freq, on the other hand, tends to be less accurate but then
> > > the actual level of precision (and source of information) varies between
> > > architectures making it a bit ambiguous.
> > >
> > > The new attribute, cpuinfo_avg_freq, is intended to provide more stable,
> > > distinct interface, exposing an average frequency of a given CPU(s), as
> > > reported by the hardware, over a time frame spanning no more than a few
> > > milliseconds. As it requires appropriate hardware support, this
> > > interface is optional.
> >
> > From what I recall, the plan is to:
> > - keep cpuinfo_cur_freq as it is, not expose for x86 and call ->get()
> >   for ARM.
> 
> Yes.
That one indeed remains unchanged.
> 
> > - introduce cpuinfo_avg_freq() and make it return frequency from hw
> >   counters for both ARM and Intel and others who provide the API.
> 
> Yes.
Will add changes for Intel as well.
> 
> > - update scaling_cur_freq() to only return the requested frequency or
> >   error in case of X86
> 
> Yes.
> 
> Preferably, -ENOTSUPP for "setpolicy" drivers without the .get() callback.
Right, my impression was that we want to leave that one as is.
Will add appropriate changes.
> 
> >   and update documentation to reflect the same.
> >   Right now or after some time ? How much time ?
> 
> After some time, I think at least two cycles, so people have the time
> to switch over, but much more may be necessary if someone is stuck
> with RHEL or similar user space.
> 
> Anyway, x86 will be the only one affected and there may be a Kconfig
> option even to allow it to be changed at the kernel build time.
> 
So just for my clarification we want a config switch to control what
scaling_cur_freq is to actually provide. It will keep the current behaviour as
default until we are ready to flip it and ultimately drop that temporary config
option ?
> The documentation for cpuinfo_avg_freq() needs to be added along with it.
That one is already provided unless you have smth else on mind ?
Like updating scaling_cur_freq to reference the new sysfs attribute ?

---
Best Regards
Beata
> 
> > > diff --git a/drivers/cpufreq/cpufreq.c b/drivers/cpufreq/cpufreq.c
> > > index 04fc786dd2c0..3493e5a9500d 100644
> > > --- a/drivers/cpufreq/cpufreq.c
> > > +++ b/drivers/cpufreq/cpufreq.c
> > > @@ -752,6 +752,16 @@ __weak unsigned int arch_freq_get_on_cpu(int cpu)
> > >       return 0;
> > >  }
> > >
> > > +__weak int arch_freq_avg_get_on_cpu(int cpu)
> > > +{
> > > +     return -EOPNOTSUPP;
> > > +}
> > > +
> > > +static inline bool cpufreq_avg_freq_supported(struct cpufreq_policy *policy)
> > > +{
> > > +     return arch_freq_avg_get_on_cpu(policy->cpu) >= 0;
> > > +}
> >
> > And why aren't we simply reusing arch_freq_get_on_cpu() here ?
> >
> > --
> > viresh