2013/6/5 Viresh Kumar viresh.kumar@linaro.org:

On 5 June 2013 08:54, Xiaoguang Chen chenxg.marvell@gmail.com wrote:

...

Hi, Guys

Hi Xiaoguang,

...

I met another question for cpufreq governor. When hotplug out one cpu, cpufreq will remove the cpu device and also call governor stop, if there are more than 1 cpu and cpu0 is the root for cpufreq, other cpus are linked to cpu0. then policy->cpu is 0 always.

Not really.. There are situations when this will not hold true.. If you hot-unplug cpu0, then next cpu in line, i.e. cpu 1, will become policy->cpu :)

Yes, I know this. but usually we dont' allow hotplug cpu0. So the root is cpu0 most of the time. this is just one example if policy->cpu is not changed.

...

in userspace governor, when governor stop is called, it will set per_cpu(cpu_is_managed, cpu) = 0

To make it clear, cpu_is_managed isn't telling which cpu is managing 'cpu' but if governor is enabled for cpu 'cpu' or not.

Yes, that should be the intent for this variable, but unluckily we get 'cpu' as 0 even when we hotplug cpu3.

...

but the cpu is policy->cpu which is 0 always.

   case CPUFREQ_GOV_STOP:
            mutex_lock(&userspace_mutex);
            cpus_using_userspace_governor--;
            if (cpus_using_userspace_governor == 0) {
                    cpufreq_unregister_notifier(
                                    &userspace_cpufreq_notifier_block,
                                    CPUFREQ_TRANSITION_NOTIFIER);
            }

            per_cpu(cpu_is_managed, cpu) = 0;
            per_cpu(cpu_min_freq, cpu) = 0;
            per_cpu(cpu_max_freq, cpu) = 0;
            per_cpu(cpu_set_freq, cpu) = 0;
            pr_debug("managing cpu %u stopped\n", cpu);
            mutex_unlock(&userspace_mutex);
            break;

also in userspace governor store_setspeed function cpufreq_set will retun error if per_cpu(cpu_is_managed, cpu) is 0 which means it will not change frequency .

static int cpufreq_set(struct cpufreq_policy *policy, unsigned int freq) { int ret = -EINVAL;

pr_debug("cpufreq_set for cpu %u, freq %u kHz\n", policy->cpu, freq);

mutex_lock(&userspace_mutex); if (!per_cpu(cpu_is_managed, policy->cpu)) goto err;

per_cpu(cpu_set_freq, policy->cpu) = freq; ...........

So consider below case:

1. cpu0 tries to hotplug cpu3, it calls userspace governor stop

function, it should stop cpu3's userspace governor, but it actually set cpu0's per_cpu(cpu_is_managed, cpu) to 0. 2) cpu0 tries to change cpu freuqency through userspace governor, but it will never succeed since cpufreq_set will return err if cpu0' percpu variable cpu_is_managed is 0. 3) unless there is another cpu hotplug in operation, in which it calls governor start and set cpu_is_managed to 1, then the frequency change will go on and doesn't report error.

The reason here is policy->cpu is always 0 if we use managed policy for all cpus. but we can't see which cpu's governor want's to start/stop in userspace goveror. we can only get the information from struct policy, but this policy doesn't include the exact cpu that wants to do governor start/stop.

Any suggestions for this issue?

Hmm.. the problem is bigger than what you observed.. I am starting to cleanup userspace governor now and will cc you for my patches. Please test them and give your Tested-by :)

Ok, thanks