Re: [RFC PATCH v4 3/6] sched/cpufreq: Hook em_pd_get_higher_power() into get_next_freq()

[Douglas Raillard]

On 1/23/20 4:16 PM, Quentin Perret wrote:
> On Wednesday 22 Jan 2020 at 17:35:35 (+0000), Douglas RAILLARD wrote:
>> @@ -210,9 +211,16 @@ static unsigned int get_next_freq(struct sugov_policy *sg_policy,
>>  	struct cpufreq_policy *policy = sg_policy->policy;
>>  	unsigned int freq = arch_scale_freq_invariant() ?
>>  				policy->cpuinfo.max_freq : policy->cur;
>> +	struct em_perf_domain *pd = sugov_policy_get_pd(sg_policy);
>>  
>>  	freq = map_util_freq(util, freq, max);
>>  
>> +	/*
>> +	 * Try to get a higher frequency if one is available, given the extra
>> +	 * power we are ready to spend.
>> +	 */
>> +	freq = em_pd_get_higher_freq(pd, freq, 0);
> 
> I find it sad that the call just below to cpufreq_driver_resolve_freq()
> and cpufreq_frequency_table_target() iterates the OPPs all over again.
> It's especially a shame since most existing users of the EM stuff do
> have a cpufreq frequency table.
> 
> Have you looked at hooking this inside cpufreq_driver_resolve_freq()
> instead ? If we have a well-formed EM available, the call to
> cpufreq_frequency_table_target() feels redundant, so we might want to
> skip it.

We can't really move the call to em_pd_get_higher_freq() into
cpufreq_driver_resolve_freq() since that's a schedutil-specific feature,
and we would loose the !sg_policy->need_freq_update optimization.

Maybe we can add a flag to cpufreq_driver_resolve_freq() that promises
that the frequency is already a valid one. We have to be careful though,
since a number of things can make that untrue:
 - em_pd_get_higher_freq() will return the passed freq verbatim if it's
higher than the max freq, so em_pd_get_higher_freq() will have to set
the flag itself in case that logic changes.
 - policy limits can change the value
 - future things could tinker with the freq and forget to reset the flag.

If you think it's worth it I can make these changes.

> Thoughts ?
> 
> Quentin
> 

Cheers,
Douglas