Re: [PATCH 1/2] cpufreq: Add governor operation ongoing flag

[Xiaoguang Chen]

2013/8/14 Viresh Kumar <viresh.kumar@xxxxxxxxxx>:
> On 13 August 2013 12:39, Xiaoguang Chen <chenxg@xxxxxxxxxxx> wrote:
>> __cpufreq_governor operation needs to be executed one by one.
>> If one operation is ongoing, the other operation can't be executed.
>> If the order is not guaranteed, there may be unexpected behavior.
>
> What order??
I mean one stop operation is ongoing, one other process tries to call
a start operation.

>
>>  For example, governor is in enable state, and one process
>> tries to stop the goveror, but it is scheduled out before policy->
>> governor->governor() is executed, but the governor enable flag is
>> set to false already. Then one other process tries to start governor,
>> It finds enable flag is false, and it can process down to do governor
>> start operation, So the governor is started twice.
>
> That's not possible. A process will not and should not call START
> before calling STOP. And so the order of calling these routines must
> be forced.
>
> Hence, we may not need your patch.

Please see below code in __cpufreq_governor function

mutex_lock(&cpufreq_governor_lock);
if ((!policy->governor_enabled && (event == CPUFREQ_GOV_STOP)) ||
                            //////////// <1> Here one process A tries
to stop governor, it finds governor is enabled, so it will pass down.
   (policy->governor_enabled && (event == CPUFREQ_GOV_START))) {
                          /////////////<3> Process B tries to start
governor, it finds enable flag is false, so it can also pass down.
mutex_unlock(&cpufreq_governor_lock);
return -EBUSY;
}

if (event == CPUFREQ_GOV_STOP)
policy->governor_enabled = false;
                                                  //////////// < 2>
Here process A set flag to false and then process A is scheduled out
for some reasons(like interrupt or time slice end)
else if (event == CPUFREQ_GOV_START)
policy->governor_enabled = true;

mutex_unlock(&cpufreq_governor_lock);

ret = policy->governor->governor(policy, event);
                                           ///////////////<4>  Process
B executes the governor start operation, process A is not scheduled
back yet. as policy->governor->governor is not protected by


                the cpufreq_governor_lock, So this sequence can happen
really.









Thanks
Xiaoguang
--
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/