Viresh,
On Tue, Nov 25, 2014 at 04:27:44PM +0530, Viresh Kumar wrote:
On 24 November 2014 at 23:40, Eduardo Valentin edubezval@gmail.com wrote:
Is it possible to have this registration only when we have a cpufreq driver up and running? The reasoning is that only after we have a way to control cpu frequencies, it makes sense to have the cpu_cooling device.
I am planing to have the following check in the cpu cooling code: if (!cpufreq_get_current_driver()) { dev_dbg(bgp->dev, "no cpufreq driver yet\n"); return -EPROBE_DEFER; }
that is the way I think of checking if the cpufreq layer is ready to have a cpu cooling on top of it. Currently, thermal drivers check this before calling cpu cooling registration. But instead of having this check in every driver, I would like to move it to cpu cooling.
However, for cpufreq-dt, the registration currently happens in the init phase, not in probe, so cpufreq driver is not registered, and thus the check won't work.
This is how the phases are present in cpufreq drivers: -> platform_init -> probe() ->cpufreq_driver->init()
You are right! I got confused because even with your patch, the sequencing is not working. Looking to that behavior I, somehow, thought the _init function in cpufreq-dt was about init() calls. But in fact, it is driver initialization callback.
And the cooling device is registered in cpufreq_driver->init() and by the time ->init() is called, cpufreq_driver is valid.
However, by the time of ->init() the cpufreq_driver is not really ready. Or at least, the cpufreq layer is not ready. A call to
cpufreq_frequency_get_table()
for instance, it is not working. The reasoning is because cpufreq_add_dev() is still not finalized by the time driver->init() is called. Meaning, the policy cannot be fetch because the cpu masks have not been set by that time. More important, in order to fetch the cpufreq table we would need to have the cpufreq_cpu_data properly initialized per cpu. In other words, when we call driver->init() that happens before the following code: cpumask_and(policy->cpus, policy->cpus, cpu_online_mask);
if (!recover_policy) { policy->user_policy.min = policy->min; policy->user_policy.max = policy->max; }
down_write(&policy->rwsem); write_lock_irqsave(&cpufreq_driver_lock, flags); for_each_cpu(j, policy->cpus) per_cpu(cpufreq_cpu_data, j) = policy; write_unlock_irqrestore(&cpufreq_driver_lock, flags);
The thing is, with current code: cpufreq_add_dev() -> cpufreq-dt->init() -> of_cpufreq_cooling_register() -> cpufreq_frequency_get_table() -> returns NULL;
Returns invalid data because it is not initialized yet.
The cpufreq-dt would need to add the of based cpufreq cooling only when cpufreq layer is ready. Any other better cpufreq driver callback to add the cpu cooling?
We could sort this out by polling in thermal layer for the cpufreq table until it gets ready, but I believe that would be a dirty hack.
Cheers,
Eduardo Valentin