Re: [PATCH 1/3] powercap, intel_rapl, implement get_max_time_window

[Prarit Bhargava <prarit@redhat.com>]

On 12/17/2015 12:45 AM, Seiichi Ikarashi wrote:
> On 2015-12-15 22:02, Prarit Bhargava wrote:
>> The MSR_PKG_POWER_INFO register (Intel ASDM, section 14.9.3
>> "Package RAPL Domain") provides a maximum time window which the
>> system can support.  This window is read-only and is currently
>> not examined when setting the time windows for the package.
> 
> I have been having a question here long time.
> Maximum Time Window (bits 53:48) in MSR_PKG_POWER_INFO is only
> 6-bit length even though Time Window for Power Limit #1 (bits 23:17)
> and Time Window for Power Limit #2 (bits 55:49) in MSR_PKG_POWER_LIMIT 
> are both 7-bit length, not 6.

While looking at the MSR settings I had exactly the same question!  I too would
like to know the answer.

> 
> Do Intel guys have an answer for it?
> 
> 
> The patch itself looks good to me.
> Just minor comments below:
> 
>> diff --git a/drivers/powercap/intel_rapl.c b/drivers/powercap/intel_rapl.c
>> index cc97f08..f765b2c 100644
>> --- a/drivers/powercap/intel_rapl.c
>> +++ b/drivers/powercap/intel_rapl.c
>> @@ -493,13 +493,42 @@ static int get_current_power_limit(struct powercap_zone *power_zone, int id,
>>  	return ret;
>>  }
>>  
>> +static int get_max_time_window(struct powercap_zone *power_zone, int id,
> 
> The 2nd arg "id" is not necessary.

I'll drop this in v2.

> 
>> +			       u64 *data)
>> +{
>> +	struct rapl_domain *rd;
>> +	int ret = 0;
>> +	u64 val;
>> +
>> +	get_online_cpus();
>> +	rd = power_zone_to_rapl_domain(power_zone);
>> +
>> +	if (rapl_read_data_raw(rd, MAX_TIME_WINDOW, true, &val))
> 
> rapl_read_data_raw() can return -EINVAL and -ENODEV other than -EIO.
> 
>> +		ret = -EIO;
> 
> Is it OK to limit ret to -EIO here?

AFAICT it seems like it.  The only error that can occur here (at least by the
time this code is executed) is that there is a range error.  -EIO seems appropriate.

> 
>> +	else
>> +		*data = val;
>> +
>> +	put_online_cpus();
>> +	return ret;
>> +}
>> +
>>  static int set_time_window(struct powercap_zone *power_zone, int id,
>>  								u64 window)
>>  {
>>  	struct rapl_domain *rd;
>>  	int ret = 0;
>> +	u64 max_window;
>>  
>>  	get_online_cpus();
>> +	ret = get_max_time_window(power_zone, id, &max_window);
>> +	if (ret < 0)
>> +		goto out;
>> +
>> +	if (window > max_window) {
>> +		ret = -EINVAL;
>> +		goto out;
>> +	}
>> +
>>  	rd = power_zone_to_rapl_domain(power_zone);
>>  	switch (rd->rpl[id].prim_id) {
>>  	case PL1_ENABLE:
>> @@ -511,6 +540,7 @@ static int set_time_window(struct powercap_zone *power_zone, int id,
>>  	default:
>>  		ret = -EINVAL;
>>  	}
>> +out:
>>  	put_online_cpus();
>>  	return ret;
>>  }
>> @@ -590,6 +620,7 @@ static struct powercap_zone_constraint_ops constraint_ops = {
>>  	.set_time_window_us = set_time_window,
>>  	.get_time_window_us = get_time_window,
>>  	.get_max_power_uw = get_max_power,
>> +	.get_max_time_window_us = get_max_time_window,
>>  	.get_name = get_constraint_name,
>>  };
>>  
>> diff --git a/drivers/powercap/powercap_sys.c b/drivers/powercap/powercap_sys.c
>> index 84419af..7d77b83 100644
>> --- a/drivers/powercap/powercap_sys.c
>> +++ b/drivers/powercap/powercap_sys.c
>> @@ -101,7 +101,7 @@ static ssize_t store_constraint_##_attr(struct device *dev,\
>>  	int err; \
>>  	u64 value; \
>>  	struct powercap_zone *power_zone = to_powercap_zone(dev); \
>> -	int id; \
>> +	int id, ret; \
>>  	struct powercap_zone_constraint *pconst;\
>>  	\
>>  	if (!sscanf(dev_attr->attr.name, "constraint_%d_", &id)) \
>> @@ -113,8 +113,10 @@ static ssize_t store_constraint_##_attr(struct device *dev,\
>>  	if (err) \
>>  		return -EINVAL; \
>>  	if (pconst && pconst->ops && pconst->ops->set_##_attr) { \
>> -		if (!pconst->ops->set_##_attr(power_zone, id, value)) \
>> +		ret = pconst->ops->set_##_attr(power_zone, id, value); \
>> +		if (!ret) \
>>  			return count; \
>> +		return ret; \
> 
> An opposite question to above.
> Is it OK not to limit the return value to -EINVAL here?
> Do you want this function to return -EIO or something?

In this case, no, because the define is used by other values.  I think that
would limit all erros in the set_* functions to be -EIO.

P.