Re: [PATCH 2/5] x86: use common aperfmperf_khz_on_cpu() to calculate KHz using APERF/MPERF

From: Rafael J. Wysocki
Date: Mon Jun 19 2017 - 08:35:46 EST

On Friday, June 16, 2017 09:49:00 PM Len Brown wrote:
> On Fri, Jun 16, 2017 at 8:30 PM, Rafael J. Wysocki <rjw@xxxxxxxxxxxxx> wrote:
> > On Wednesday, June 07, 2017 07:39:13 PM Len Brown wrote:
> >> From: Len Brown <len.brown@xxxxxxxxx>


> >
> > I wonder if we could change intel_pstate_get() to simply return
> > aperfmperf_khz_on_cpu(cpu_num)?
> >
> > That would allow us to avoid the extra branch here and get rid of the
> > #ifdef x86 from the header.
> The reason I put the hook here is specifically so that the same
> code would always be called on the x86 architecture,
> no not matter what cpufreq driver is loaded.
> Yes, alternatively, all possible driver.get routines could be updated
> to call the same routine. That is acpi-cpufreq.c, intel_pstate.c,
> others?

Just acpi-cpufreq.c and intel_pstate.c.

Moreover, I wouldn't change the behavior on systems using acpi-cpufreq.c,
because why really?

And some users may complain that now what they see in cpuinfo_cur_freq is
different from what they saw there before.

That leaves us with just intel_pstate.

In addition to that, it would be good to avoid reading APERF and MPERF in the
cases when we already have the values. I have an idea on how to do that, but
I'll just need to prepare a patch.

> Do I think that saving a branch is meaningful in a system call to retrieve
> the average frequency from the kernel? No, I don't.
> I'm not religious about "if (CONFIG_X86)" appearing in cpufreq.h.
> If somebody really cares, then the purist approach would be
> to add an additional arch-specific-hook header -- though that seemed
> overkill to me...

I'm not a fan of arch stuff in general code, C or headers.

If it is necessary to add an arch header, let's add it, but at this point I'm
not sure about that even.