Re: [PATCH v2] pm: Add runtime PM statistics

[Kevin Hilman <khilman@deeprootsystems.com>]

Arjan van de Ven <arjan@linux.intel.com> writes:

> On 8/5/2010 4:20 PM, Kevin Hilman wrote:
>> Arjan van de Ven<arjan@linux.intel.com>  writes:
>>
>>    
>>> +
>>> +/**
>>> + * update_pm_runtime_accounting - Update the time accounting of power
>>> states
>>> + * @dev: Device to update the accounting for
>>> + *
>>> + * In order to be able to have time accounting of the various power states
>>> + * (as used by programs such as PowerTOP to show the effectiveness of
>>> runtime
>>> + * PM), we need to track the time spent in each state.
>>> + * update_pm_runtime_accounting must be called each time before the
>>> + * runtime_status field is updated, to account the time in the old state
>>> + * correctly.
>>> + */
>>> +void update_pm_runtime_accounting(struct device *dev)
>>> +{
>>> +    unsigned long now = jiffies;
>>> +    int delta;
>>> +
>>> +    delta = now - dev->power.accounting_timestamp;
>>> +
>>> +    if (delta<  0)
>>> +        delta = 0;
>>> +
>>> +    dev->power.accounting_timestamp = now;
>>> +
>>> +    if (dev->power.disable_depth>  0)
>>> +        return;
>>> +
>>> +    if (dev->power.runtime_status == RPM_SUSPENDED)
>>> +        dev->power.suspended_jiffies += delta;
>>> +    else
>>> +        dev->power.active_jiffies += delta;
>>> +}
>>>      
>> By using jiffies, I think we might miss events in drivers that are doing
>> runtime PM transitions in short bursts.  On embedded systems with slow
>> HZ, there could potentially be lots of transitions between ticks.
>>
>> It would be nicer to use clocksource-based time so transitions between
>> jiffies could still be factored into the accounting.
>>    
>
> you're absolutely right that the current mechanism is more "sampling
> accuracy" (similar to most /proc info that shows up with top and
> such).
>
> on the "slow HZ".. there is no more valid reason to not set HZ to
> 1000... 

Probably, especially with tickless idle, but not so sure there is total
agreement on this in the embedded world though...

> so we'll get 1 msec sampling rate basically.
>
> the problem with a more accurate clocksource is that it's
> expensive. And more... the path to such clocksource itself might be
> subject to power management ;-)

What about using read_persistent_clock() then?  Then the arch/platform
definition of this will determine the max sampling rate.

Kevin