Re: [RFC 0/3] cpufreq: cppc: Add support for frequency invariance

From: Ionela Voinescu
Date: Tue Aug 25 2020 - 05:56:36 EST


Hi guys,

On Friday 24 Jul 2020 at 11:38:59 (+0200), Vincent Guittot wrote:
> On Fri, 10 Jul 2020 at 05:00, Viresh Kumar <viresh.kumar@xxxxxxxxxx> wrote:
> >
> > Thanks for the quick reply Ionela.
> >
> > On 09-07-20, 13:43, Ionela Voinescu wrote:
> > > I'll put all my comments here for now, as they refer more to the design
> > > of the solution.
> > >
> > > I hope it won't be too repetitive compared to what we previously discussed
> > > offline.
> >
> > > I understand you want to get additional points of view.
> >
> > Not necessarily, I knew you would be one of the major reviewers here
> > :)
> >
> > I posted so you don't need to review in private anymore and then the
> > code is somewhat updated since the previous time.
> >
> > > On Thursday 09 Jul 2020 at 15:43:32 (+0530), Viresh Kumar wrote:
> > > I believe the code is unnecessarily invasive for the functionality it
> > > tries to introduce and it does break existing functionality.
> > >
> > >
> > > - (1) From code readability and design point of view, this switching
> > > between an architectural method and a driver method complicates
> > > an already complicated situation. We already have code that
> > > chooses between a cpufreq-based method and a counter based method
> > > for frequency invariance. This would basically introduce a choice
> > > between a cpufreq-based method through arch_set_freq_scale(), an
> > > architectural counter-based method through arch_set_freq_tick(),
> > > and another cpufreq-based method that piggy-backs on the
> > > architectural arch_set_freq_tick().
> >
> > I agree.
> >
> > > As discussed offline, before I even try to begin accepting the
> > > possibility of this complicated mix, I would like to know why
> > > methods of obtaining the same thing by using the cpufreq
> > > arch_set_freq_scale()
> >
> > The problem is same as that was in case of x86, we don't know the real
> > frequency the CPU may be running at and we need something that fires
> > up periodically in a guaranteed way to capture the freq-scale.
>
> Yeah it's exactly the same behavior as x86 and re using the same
> mechanism seems the best solution
>
> The main problem is that AMU currently assumes that it will be the
> only to support such tick based mechanism whereas others like cppc can
> provides similar information
>

Yes, I agree that a similar method to the use of AMUs or APERF/MPERF would
result in a more accurate frequency scale factor.

> >
> > Though I am thinking now if we can trust the target_index() helper and
> > keep updating the freq-scale based on the delta between last call to
> > it and the latest call. I am not sure if it will be sufficient.
> >
> > > or even the more invasive wrapping of the
> > > counter read functions is not working.
> >
> > I am not sure I understood this one.
> >

I've been putting some more thought/code into this one and I believe
something as follows might look nicer as well as cover a few corner cases
(ignore implementation details for now, they can be changed):

- Have a per cpu value that marks the use of either AMUs, CPPC, or
cpufreq for freq invariance (can be done with per-cpu variable or with
cpumasks)

- arch_topology.c: initialization code as follows:

for_each_present_cpu(cpu) {
if (freq_inv_amus_valid(cpu) &&
!freq_inv_set_max_ratio(cpufreq_get_hw_max_freq(cpu) * 1000,
arch_timer_get_rate(), cpu)) {
per_cpu(inv_source, cpu) = INV_AMU_COUNTERS;
continue;
}
if (freq_inv_cppc_counters_valid(cpu) &&
!freq_inv_set_max_ratio(cppc_max_perf, cppc_ref_perf, cpu)) {
per_cpu(inv_source, cpu) = INV_CPPC_COUNTERS;
continue;
}
if (!cpufreq_supports_freq_invariance() ||
freq_inv_set_max_ratio(cpufreq_get_hw_max_freq(cpu),
1, cpu)) {
pr_info("FIE disabled: no valid source for CPU%d.", cpu);
return 0;
}
}
This would live in an equivalent of the init function we have now for
AMU counters only (init_amu_fie), made to handle more sources and moved
to arch_topology.c.
The freq_inv_set_max_ratio() would be a generic version of what is now
validate_cpu_freq_invariance_counters() (only the ratio computation and
setting).

- Finally,
void freq_inv_update_counter_refs(void)
{
this_cpu_write(arch_core_cycles_prev, read_corecnt());
this_cpu_write(arch_const_cycles_prev, read_constcnt());
}
This would be an equivalent of init_cpu_freq_invariance_counters().
There is the option of either read_{corecnt/constcnt}() to either do AMU
reads or CPPC counter reads depending on inv_source, or for either arm64
or cppc code to define the entire freq_inv_update_counter_refs().

- Given all of the above, topology_scale_freq_tick() can then be made generic

prev_const_cnt = this_cpu_read(arch_const_cycles_prev);
prev_core_cnt = this_cpu_read(arch_core_cycles_prev);
freq_inv_update_counter_refs();
const_cnt = this_cpu_read(arch_const_cycles_prev);
core_cnt = this_cpu_read(arch_core_cycles_prev);

I have some versions of code that do this generalisation for AMUs and cpufreq
with a common topology_set_freq_scale() that is to be used for both
arch_set_freq_scale() and arch_set_freq_tick(). But it's written with this
usecase in mind so it can be extended to use CPPC counters as source as well,
as detailed above.

So, this is basically what I had in mind when recommending "wrapping of the
counter read functions" :). This would basically reuse much of what is now
the AMU invariance code while allowing for CPPC counters as a possible source.

I'll stop here for now to see what you guys think about this.

Thanks,
Ionela.

> > > - (2) For 1/3, the presence of AMU counters does not guarantee their
> > > usability for frequency invariance. I know you wanted to avoid
> > > the complications of AMUs being marked as supporting invariance
> > > after the cpufreq driver init function, but this breaks the
> > > scenario in which the maximum frequency is invalid.
> >
> > Is that really a scenario ? i.e. Invalid maximum frequency ? Why would
> > that ever happen ?
> >
> > And I am not sure if this breaks anything which already exists,
> > because all we are doing in this case now is not registering cppc for
> > FI, which should be fine.
>
> IIUC, AMU must wait for cpufreq drivers to be registered in order to
> get the maximum freq and being able to enable its FIE support.
> Could we have a sequence like:
> cppc register its scale_freq_tick function
> AMU can then override the tick function for cpu which supports AMU
>
> >
> > > - (3) For 2/3, currently we support platforms that have partial support
> > > for AMUs, while this would not be supported here. The suggestions
> > > at (1) would give us this for free.
> >
> > As both were counter based mechanisms, I thought it would be better
> > and more consistent if only one of them is picked. Though partial
> > support of AMUs would still work without the CPPC driver.
> >
> > --
> > viresh