Re: [PATCH v2 12/19] PM / devfreq: tegra: Avoid inconsistency of current frequency value

[Dmitry Osipenko]

16.04.2019 10:15, Chanwoo Choi ÐÐÑÐÑ:
> Hi,
> 
> On 19. 4. 15. ìí 11:54, Dmitry Osipenko wrote:
>> The frequency value potentially could change in-between. It doesn't
>> cause any real problem at all right now, but that could change in the
>> future. Hence let's avoid the inconsistency.
>>
>> Signed-off-by: Dmitry Osipenko <digetx@xxxxxxxxx>
>> ---
>>  drivers/devfreq/tegra-devfreq.c | 6 ++++--
>>  1 file changed, 4 insertions(+), 2 deletions(-)
>>
>> diff --git a/drivers/devfreq/tegra-devfreq.c b/drivers/devfreq/tegra-devfreq.c
>> index a668e4fbc874..f1a6f951813a 100644
>> --- a/drivers/devfreq/tegra-devfreq.c
>> +++ b/drivers/devfreq/tegra-devfreq.c
>> @@ -496,13 +496,15 @@ static int tegra_devfreq_get_dev_status(struct device *dev,
>>  {
>>  	struct tegra_devfreq *tegra = dev_get_drvdata(dev);
>>  	struct tegra_devfreq_device *actmon_dev;
>> +	unsigned long cur_freq;
>>  
>> -	stat->current_frequency = tegra->cur_freq * KHZ;
>> +	cur_freq = READ_ONCE(tegra->cur_freq);
>>  
>>  	/* To be used by the tegra governor */
>>  	stat->private_data = tegra;
>>  
>>  	/* The below are to be used by the other governors */
>> +	stat->current_frequency = cur_freq * KHZ;
>>  
>>  	actmon_dev = &tegra->devices[MCALL];
>>  
>> @@ -513,7 +515,7 @@ static int tegra_devfreq_get_dev_status(struct device *dev,
>>  	stat->busy_time *= 100 / BUS_SATURATION_RATIO;
>>  
>>  	/* Number of cycles in a sampling period */
>> -	stat->total_time = ACTMON_SAMPLING_PERIOD * tegra->cur_freq;
>> +	stat->total_time = ACTMON_SAMPLING_PERIOD * cur_freq;
>>  
>>  	stat->busy_time = min(stat->busy_time, stat->total_time);
>>  
>>
> 
> The read/write access of tegra->cur_freq is in the single routine
> of update_devfreq() as following. I think that there are no any
> potential problem about the inconsistency of tegra->cur_freq.

No, that's wrong assumption. The tegra->cur_freq is changed by the clock notifier that runs asynchronously with the devfreq driver when EMC clock rate is changed by something else in the kernel. 

> IMHO, if there are no any problem now, I'm not sure that we need
> to apply this patch.
> 
> update_devfreq()
> {
> 	devfreq->governor->get_target_freq()
> 		devfreq_update_stats(devfreq)
> 			tegra_devfreq_get_dev_status()
> 				stat->current_frequency = tegra->cur_freq * KHZ;
> 
> 	devfreq_set_target()
> 		tegra_devfreq_target()
> 			clk_set_min_rate(emc_rate, )
> 				tegra_actmon_rate_notify_cb()
> 					tegra->cur_freq = data->new_rate / KHZ;
> 		
> 			clk_set_rate(emc_rate, )
> 				tegra_actmon_rate_notify_cb()
> 					tegra->cur_freq = data->new_rate / KHZ;
> }
> 
> 

The cur_freq value is changed by the clock notifier that runs asynchronously with the rest of the devfreq driver. Hence potentially compiler may generate two separate fetches of the cur_freq value, then the clock rate could be changed by other CPU core simultaneously with tegra_devfreq_get_dev_status() or kernel may re-schedule preemptively, changing the clock rate in-between of the two fetches.