Re: [PATCH v3 0/3] cpufreq: avoid redundant driver calls in schedutil

[Pingbo Wen]

On Thursday, July 14, 2016 04:25 AM, Steve Muckle wrote:
> Invoking the cpufreq driver to set a frequency can be expensive. On platforms
> with a cpufreq driver that does not support fast switching a thread must be
> woken to complete the operation. IPIs will also occur if and when support to
> process remote task wakeups is added in schedutil.
> 
> Currently schedutil calculates a raw frequency request from scheduler
> utilization data. This raw frequency request does not correlate to supported
> cpufreq driver frequencies aside from being capped by the CPU's maximum
> frequency. Consequently, there may be many consecutive requests for different
> raw frequency values which all translate to the same driver-supported
> frequency. For example on a platform with 3 supported frequencies 200MHz,
> 400MHz, and 600MHz, raw requests for 257MHz, 389MHz, and 307MHz all map to a
> driver-supported frequency of 400MHz in schedutil. Assuming these requests were
> consecutive and there were no changes in policy limits (min/max), there is no
> need to issue the second or third request.
> 
> In order to resolve a raw frequency request to a driver-supported one a new
> cpufreq API is added, cpufreq_driver_resolve_freq(). This API relies on a new
> cpufreq driver callback in the case of ->target() style drivers. Otherwise it
> uses the existing frequency table operations.
> 
> Lookups are cached both in cpufreq_driver_resolve_freq() (for the benefit of the
> driver) and in schedutil.
> 
> Changes since v2:
> - incorporated feedback from Viresh to use resolve_freq driver callbacks
>   only for ->target() style drivers, to use cpufreq's freq table operations,
>   and move freq mapping caching into cpufreq policy
> Changes since v1:
> - incorporated feedback from Rafael to avoid referencing freq_table from
>   schedutil by introducing a new cpufreq API
> 
> Steve Muckle (3):
>   cpufreq: add cpufreq_driver_resolve_freq()
>   cpufreq: schedutil: map raw required frequency to driver frequency

Tested the first two patches on db410c, only waking up irq_work 53
times, while previous was 257 times(79% decrease) in Android home idle
for 5 minutes.

>   cpufreq: acpi-cpufreq: use cached frequency mapping when possible
> 

Pingbo