Re: [PATCH] cpufreq: conservative: fix requested_freq reduction issue

From: Stratos Karafotis
Date: Fri Nov 08 2013 - 12:43:34 EST


On Fri, Nov 8, 2013 at 6:55 AM, Viresh Kumar <viresh.kumar@xxxxxxxxxx> wrote:
> On 8 November 2013 00:36, Stratos Karafotis <skarafotis@xxxxxxxxx> wrote:
>> I think the existing code already checks if the requested_freq is greater
>> than policy->max in __cpufreq_driver_target.
>
> Yes it does. But the problem is:
> - cs_check_cpu() sets requested_freq above policy->max
> - We execute following code because (requested_freq != policy->max)
>
> dbs_info->requested_freq += get_freq_target(cs_tuners, policy);
> __cpufreq_driver_target(policy, dbs_info->requested_freq,
> CPUFREQ_RELATION_H);
> - In __cpufreq_driver_target(), we don't do anything and return early..
> - Above will keep on repeating all the time..
>
> If we change the code as I have suggested it to be:
> - After first loop where requested_freq went over policy->max, we will
> return early from cs_check_cpu(), but we have already set freq to max..
>
>> If we put this check earlier, cpufreq will never reach policy->max.
>
> Can you please explain why do you see that happening?

Please let me rephrase my previous post. In some circumstances (depending
on freq_step and freq_table values) CPU frequency will never reach to
policy->max.

For example suppose that (for simplicity values in MHz):
policy->max = 1000
policy->cur = 800
requested_freq = 800
freq_target = 300

In 'first' iteration, if we return early with this code (because
requested_freq will be
1100):
if (dbs_info->requested_freq >= policy->max)
return;

CPU freq will never go over 800MHz.

I think the current code works correctly.
- The requested freq will go to 1100 in first iteration.
- __cpufreq_driver_target will change CPU freq to 1000
- dbs_cpufreq_notifier will adjust the requested_freq to 1000

In next iteration the code:
if (dbs_info->requested_freq == policy->max)
return;

will keep the freq to max and break out early.

So, I think there is no need for an extra check because of
dbs_cpufreq_notifier code.


Thanks,
Stratos Karafotis
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/