On Fri, 2017-06-02 at 16:59 +0530, Viresh Kumar wrote:
The transition_latency_ns represents the maximum time it can take for the hardware to switch from/to any frequency for a CPU.
The transition_latency_ns is used currently for two purposes:
o To check if the hardware latency is over the maximum allowed for a governor (only for ondemand and conservative (why not schedutil?)) and to decide if the governor can be used or not.
o To calculate the sampling_rate or rate_limit for the governors by multiplying transition_latency_ns with a constant.
The platform drivers can also set this value to CPUFREQ_ETERNAL if they don't know this number and in that case we disallow use of ondemand and conservative governors as the latency would be higher than the maximum allowed for the governors.
In many cases this number is forged by the driver authors to get the default sampling rate to a desired value. Anyway, the actual latency values can differ from what is received from the hardware designers.
Over that, what is provided by the drivers is most likely the time it takes to change frequency of the hardware, which doesn't account the software overhead involved.
In order to have guarantees about this number, this patch tries to calculate the latency dynamically at cpufreq driver registration time by first switching to min frequency, then to the max and finally back to the initial frequency. And the maximum of all three is used as the target_latency. Specifically the time it takes to go from min to max frequency (when the software runs the slowest) should be good enough, and even if there is a delta involved then it shouldn't be a lot.
For now this patch limits this feature only for platforms which have set the transition latency to CPUFREQ_ETERNAL. Maybe we can convert everyone to use it in future, but lets see.
This is tested over ARM64 Hikey platform which currently sets "clock-latency" as 500 us from DT, while with this patch the actualy value increased to 800 us.
I remember checking if transition latency is correct for imx6q-cpufreq and it does not appear to be. Maybe because i2c latency of regulator adjustments is not counted in?
It seems to me it would be much nicer to have a special flag for this instead of overriding CPUFREQ_ETERNAL.
Also, wouldn't it be possible to update this dynamically? Just measure the duration every time it happens and do an update like latency = (latency * 7 + latest_latency) / 8.
-- Regards, Leonard