On Wednesday 10 Apr 2019 at 11:14:49 (+0530), Viresh Kumar wrote:
> On 28-03-19, 10:13, Quentin Perret wrote:
> > +static unsigned int get_state_freq(struct cpufreq_cooling_device
> > *cpufreq_cdev,
> > + unsigned long state)
> > +{
> > + struct cpufreq_policy *policy;
> > + unsigned long idx;
> > +
> > + /* Use the Energy Model table if available */
> > + if (cpufreq_cdev->em) {
> > + idx = cpufreq_cdev->max_level - state;
> > + return cpufreq_cdev->em->table[idx].frequency;
> > + }
> > +
> > + /* Otherwise, fallback on the CPUFreq table */
> > + policy = cpufreq_cdev->policy;
> > + if (policy->freq_table_sorted == CPUFREQ_TABLE_SORTED_ASCENDING)
>
> It is not guaranteed that the frequency table is sorted in any order, isn't
> it ?
Hmm, indeed... I thought cpufreq_table_validate_and_sort() was actively
sorting the table but it seems I was wrong.
But I _think_ in practice the freq table actually happens to be sorted
for the upstream cpufreq drivers with the CPUFREQ_IS_COOLING_DEV flag
set. Most of them use dev_pm_opp_init_cpufreq_table() which guarantees
the table is sorted and qoriq-cpufreq explicitly sorts the table. But
I'm not sure about qcom-cpufreq-hw ...
So, if the above is true, perhaps I could simply add a check to mandate
that policy->freq_table_sorted != CPUFREQ_TABLE_SORTED_UNSORTED for
cpu_cooling ? That shouldn't harm the existing users.
Do you happen to know a board where the table is unsorted ? Is it a
common use-case ?
If yes, then I'll probably need to drop the dependency on cpufreq's
freq_table and use something else to convert indexes into frequencies
(PM_OPP ?). Unless we can force-sort the table in the cpufreq core, but
that might require lots of changes to lots of drivers too.
>
> > + idx = cpufreq_cdev->max_level - state;
> > + else
> > + idx = state;
>
> --
> viresh
Thanks,
Quentin
