| Message ID | 20200318114548.19916-1-lukasz.luba@arm.com |
|---|---|
| Headers | show |
| Series | Add support for devices in the Energy Model | expand |
Hi Lukasz, On 18/03/2020 12:45, Lukasz Luba wrote: > Add support of other devices into the Energy Model framework not only the > CPUs. Change the interface to be more unified which can handle other > devices as well. thanks for taking care of that. Overall I like the changes in this patch but it hard to review in details because the patch is too big :/ Could you split this patch into smaller ones? eg. (at your convenience) - One patch renaming s/cap/perf/ - One patch adding a new function: em_dev_register_perf_domain(struct device *dev, unsigned int nr_states, struct em_data_callback *cb); (+ EXPORT_SYMBOL_GPL) And em_register_perf_domain() using it. - One converting the em_register_perf_domain() user to em_dev_register_perf_domain - One adding the different new 'em' functions - And finally one removing em_register_perf_domain(). > Acked-by: Quentin Perret <qperret@google.com> > Signed-off-by: Lukasz Luba <lukasz.luba@arm.com> > --- [ ... ] > 2. Core APIs > @@ -70,14 +72,16 @@ CONFIG_ENERGY_MODEL must be enabled to use the EM framework. > Drivers are expected to register performance domains into the EM framework by > calling the following API:: > > - int em_register_perf_domain(cpumask_t *span, unsigned int nr_states, > - struct em_data_callback *cb); > + int em_register_perf_domain(struct device *dev, unsigned int nr_states, > + struct em_data_callback *cb, cpumask_t *cpus); Isn't possible to get rid of this cpumask by using cpufreq_cpu_get() which returns the cpufreq's policy and from their get the related cpus ? [ ... ]
On 18/03/2020 12:45, Lukasz Luba wrote: > From: Matthias Kaehlcke <mka@chromium.org> > > Now that devfreq supports limiting the frequency range of a device > through PM QoS make use of it instead of disabling OPPs that should > not be used. > > The switch from disabling OPPs to PM QoS introduces a subtle behavioral > change in case of conflicting requests (min > max): PM QoS gives > precedence to the MIN_FREQUENCY request, while higher OPPs disabled > with dev_pm_opp_disable() would override MIN_FREQUENCY. > > Signed-off-by: Matthias Kaehlcke <mka@chromium.org> > Reviewed-by: Lukasz Luba <lukasz.luba@arm.com> > Reviewed-by: Chanwoo Choi <cw00.choi@samsung.com> This patch is standalone, right? If yes, I will apply it.
On 03/04/2020 19:18, Matthias Kaehlcke wrote: > Hi Daniel, > > On Fri, Apr 03, 2020 at 06:43:20PM +0200, Daniel Lezcano wrote: >> On 18/03/2020 12:45, Lukasz Luba wrote: >>> From: Matthias Kaehlcke <mka@chromium.org> >>> >>> Now that devfreq supports limiting the frequency range of a device >>> through PM QoS make use of it instead of disabling OPPs that should >>> not be used. >>> >>> The switch from disabling OPPs to PM QoS introduces a subtle behavioral >>> change in case of conflicting requests (min > max): PM QoS gives >>> precedence to the MIN_FREQUENCY request, while higher OPPs disabled >>> with dev_pm_opp_disable() would override MIN_FREQUENCY. >>> >>> Signed-off-by: Matthias Kaehlcke <mka@chromium.org> >>> Reviewed-by: Lukasz Luba <lukasz.luba@arm.com> >>> Reviewed-by: Chanwoo Choi <cw00.choi@samsung.com> >> >> This patch is standalone, right? If yes, I will apply it. > > Yes, it is standalone, please apply Applied on 'testing', thanks
On 4/6/20 3:58 PM, Daniel Lezcano wrote: > > Hi Lukasz, > > > On 06/04/2020 15:29, Lukasz Luba wrote: >> Hi Daniel, >> >> Thank you for the review. >> >> On 4/3/20 5:05 PM, Daniel Lezcano wrote: >>> >>> Hi Lukasz, >>> >>> >>> On 18/03/2020 12:45, Lukasz Luba wrote: >>>> Add support of other devices into the Energy Model framework not only >>>> the >>>> CPUs. Change the interface to be more unified which can handle other >>>> devices as well. >>> >>> thanks for taking care of that. Overall I like the changes in this patch >>> but it hard to review in details because the patch is too big :/ >>> >>> Could you split this patch into smaller ones? >>> >>> eg. (at your convenience) >>> >>> - One patch renaming s/cap/perf/ >>> >>> - One patch adding a new function: >>> >>> em_dev_register_perf_domain(struct device *dev, >>> unsigned int nr_states, >>> struct em_data_callback *cb); >>> >>> (+ EXPORT_SYMBOL_GPL) >>> >>> And em_register_perf_domain() using it. >>> >>> - One converting the em_register_perf_domain() user to >>> em_dev_register_perf_domain >>> >>> - One adding the different new 'em' functions >>> >>> - And finally one removing em_register_perf_domain(). >> >> I agree and will do the split. I could also break the dependencies >> for future easier merge. >> >>> >>> >>>> Acked-by: Quentin Perret <qperret@google.com> >>>> Signed-off-by: Lukasz Luba <lukasz.luba@arm.com> >>>> --- >>> >>> [ ... ] >>> >>>> 2. Core APIs >>>> @@ -70,14 +72,16 @@ CONFIG_ENERGY_MODEL must be enabled to use the EM >>>> framework. >>>> Drivers are expected to register performance domains into the EM >>>> framework by >>>> calling the following API:: >>>> - int em_register_perf_domain(cpumask_t *span, unsigned int >>>> nr_states, >>>> - struct em_data_callback *cb); >>>> + int em_register_perf_domain(struct device *dev, unsigned int >>>> nr_states, >>>> + struct em_data_callback *cb, cpumask_t *cpus); >>> >>> Isn't possible to get rid of this cpumask by using >>> cpufreq_cpu_get() which returns the cpufreq's policy and from their get >>> the related cpus ? >> >> We had similar thoughts with Quentin and I've checked this. > > Yeah, I suspected you already think about that :) > >> Unfortunately, if the policy is a 'new policy' [1] it gets >> allocated and passed into cpufreq driver ->init(policy) [2]. >> Then that policy is set into per_cpu pointer for each related_cpu [3]: >> >> for_each_cpu(j, policy->related_cpus) >> per_cpu(cpufreq_cpu_data, j) = policy; >> >> >> Thus, any calls of functions (i.e. cpufreq_cpu_get()) which try to >> take this ptr before [3] won't work. >> >> We are trying to register EM from cpufreq_driver->init(policy) and the >> per_cpu policy is likely to be not populated at that phase. > > What is the problem of registering at the end of the cpufreq_online ? We want to enable driver developers to choose one of two options for the registration of Energy Model: 1. a simple one via dev_pm_opp_of_register_em(), which uses default callback function calculating power based on: voltage, freq and DT entry 'dynamic-power-coefficient' for each OPP 2. a more sophisticated, when driver provides callback function, which will be called from EM for each OPP to ask for related power; This interface could also be used by devices which relay not only on one source of 'voltage', i.e. manipulate body bias or have other controlling voltage for gates in the new 3D transistors. They might provide custom callback function in their cpufreq driver. This is used i.e. in cpufreq drivers which use firmware to get power, like scmi-cpufreq.c; To meet this requirement the registration of EM is moved into cpufreq drivers, not in the framework i.e cpufreq_online(). If we could limit the support for only option 1. then we could move the registration call into cpufreq framework and clean the cpufreq drivers. Regards, Lukasz