Message ID | 20190128165522.31749-1-quentin.perret@arm.com |
---|---|
Headers | show |
Series | Register an Energy Model for Arm reference platforms | expand |
rHi Quentin, On Mon, Jan 28, 2019 at 04:55:16PM +0000, Quentin Perret wrote: > The Energy Model (EM) framework provides an API to let drivers register > the active power of CPUs. The drivers are expected to provide a callback > method which estimates the power consumed by a CPU at each available > performance levels. How exactly this should be implemented, however, > depends on the platform. > > On some systems, PM_OPP knows the voltage and frequency at which CPUs > can run. When coupled with the CPU 'capacitance' (as provided by the > 'dynamic-power-coefficient' devicetree binding), it is possible to > estimate the dynamic power consumption of a CPU as P = C * V^2 * f, with > C its capacitance and V and f respectively the voltage and frequency of > the OPP. The Intelligent Power Allocator (IPA) thermal governor already > implements that estimation method, in the thermal framework. > > However, this power estimation method can be applied to any platform > where all the parameters are known (C, V and f), and not only those > suffering thermal issues. As such, the code implementing this feature > can be re-used to also populate the EM framework now used by EAS. > > As a first step, introduce in PM_OPP a helper function which CPUFreq > drivers can use to register into the EM framework. This duplicates the > power estimation done in IPA until it can be migrated to using the EM > framework. This will be done later, once the EM framework has support > for at least all platforms currently supported by IPA. > > Signed-off-by: Quentin Perret <quentin.perret@arm.com> > --- > drivers/opp/of.c | 60 ++++++++++++++++++++++++++++++++++++++++++ > include/linux/pm_opp.h | 5 ++++ > 2 files changed, 65 insertions(+) > > diff --git a/drivers/opp/of.c b/drivers/opp/of.c > index 06f0f632ec47..7572a2eb2fd4 100644 > --- a/drivers/opp/of.c > +++ b/drivers/opp/of.c > @@ -1047,3 +1047,63 @@ struct device_node *dev_pm_opp_get_of_node(struct dev_pm_opp *opp) > return of_node_get(opp->np); > } > EXPORT_SYMBOL_GPL(dev_pm_opp_get_of_node); > + > +/** > + * of_dev_pm_opp_get_cpu_power() - Estimates the power of a CPU > + * @mW: pointer to the power estimate in milli-watts > + * @KHz: pointer to the OPP's frequency, in kilo-hertz nit: should be kHz > + * @cpu: CPU for which power needs to be estimated > + * > + * Computes the power estimated by @CPU at the first OPP above @KHz (ceil), > + * and updates @KHz and @mW accordingly. > + * > + * The power is estimated as P = C * V^2 * f, with C the CPU's capacitance > + * (read from the 'dynamic-power-coefficient' devicetree binding) and V and f > + * respectively the voltage and frequency of the OPP. > + * > + * Return: -ENODEV if the CPU device cannot be found, -EINVAL if the power > + * calculation failed because of missing parameters, 0 otherwise. > + */ > +int of_dev_pm_opp_get_cpu_power(unsigned long *mW, unsigned long *KHz, int cpu) I think it is more common to put the input parameters first, then the output ones, i.e. cpu, kHz, mW. > +{ > + unsigned long mV, Hz, MHz; > + struct device *cpu_dev; > + struct dev_pm_opp *opp; > + struct device_node *np; > + u32 cap; > + u64 tmp; > + int ret; > + > + cpu_dev = get_cpu_device(cpu); > + if (!cpu_dev) > + return -ENODEV; > + > + np = of_node_get(cpu_dev->of_node); > + if (!np) > + return -EINVAL; > + > + ret = of_property_read_u32(np, "dynamic-power-coefficient", &cap); > + of_node_put(np); > + if (ret) > + return -EINVAL; > + > + Hz = *KHz * 1000; > + opp = dev_pm_opp_find_freq_ceil(cpu_dev, &Hz); > + if (IS_ERR(opp)) > + return -EINVAL; > + > + mV = dev_pm_opp_get_voltage(opp) / 1000; > + dev_pm_opp_put(opp); > + if (!mV) > + return -EINVAL; > + > + MHz = Hz / 1000000; > + tmp = (u64)cap * mV * mV * MHz; > + do_div(tmp, 1000000000); > + > + *mW = (unsigned long)tmp; > + *KHz = Hz / 1000; > + > + return 0; > +} > +EXPORT_SYMBOL_GPL(of_dev_pm_opp_get_cpu_power); > diff --git a/include/linux/pm_opp.h b/include/linux/pm_opp.h > index 0a2a88e5a383..fedde14f5187 100644 > --- a/include/linux/pm_opp.h > +++ b/include/linux/pm_opp.h > @@ -322,6 +322,7 @@ int dev_pm_opp_of_get_sharing_cpus(struct device *cpu_dev, struct cpumask *cpuma > struct device_node *dev_pm_opp_of_get_opp_desc_node(struct device *dev); > struct device_node *dev_pm_opp_get_of_node(struct dev_pm_opp *opp); > int of_get_required_opp_performance_state(struct device_node *np, int index); > +int of_dev_pm_opp_get_cpu_power(unsigned long *mW, unsigned long *KHz, int cpu); > #else > static inline int dev_pm_opp_of_add_table(struct device *dev) > { > @@ -364,6 +365,10 @@ static inline int of_get_required_opp_performance_state(struct device_node *np, > { > return -ENOTSUPP; > } > +static inline int of_dev_pm_opp_get_cpu_power(unsigned long *mW, unsigned long *KHz, int cpu) > +{ > + return -ENOTSUPP; > +} > #endif > > #endif /* __LINUX_OPP_H__ */ Besides the nits above: Reviewed-by: Matthias Kaehlcke <mka@chromium.org> Tested-by: Matthias Kaehlcke <mka@chromium.org>
On Mon, Jan 28, 2019 at 04:55:17PM +0000, Quentin Perret wrote: > Now that PM_OPP provides a helper function to estimate the power > consumed by CPUs, make sure to try and register an Energy Model (EM) > from cpufreq-dt, hence ensuring interested subsystems (the task > scheduler, for example) can make use of that information when available. > > Signed-off-by: Quentin Perret <quentin.perret@arm.com> > --- > drivers/cpufreq/cpufreq-dt.c | 7 ++++++- > 1 file changed, 6 insertions(+), 1 deletion(-) > > diff --git a/drivers/cpufreq/cpufreq-dt.c b/drivers/cpufreq/cpufreq-dt.c > index e58bfcb1169e..7556e07e7a9f 100644 > --- a/drivers/cpufreq/cpufreq-dt.c > +++ b/drivers/cpufreq/cpufreq-dt.c > @@ -16,6 +16,7 @@ > #include <linux/cpu_cooling.h> > #include <linux/cpufreq.h> > #include <linux/cpumask.h> > +#include <linux/energy_model.h> > #include <linux/err.h> > #include <linux/module.h> > #include <linux/of.h> > @@ -152,6 +153,7 @@ static int resources_available(void) > > static int cpufreq_init(struct cpufreq_policy *policy) > { > + struct em_data_callback em_cb = EM_DATA_CB(of_dev_pm_opp_get_cpu_power); > struct cpufreq_frequency_table *freq_table; > struct opp_table *opp_table = NULL; > struct private_data *priv; > @@ -160,7 +162,7 @@ static int cpufreq_init(struct cpufreq_policy *policy) > unsigned int transition_latency; > bool fallback = false; > const char *name; > - int ret; > + int ret, nr_opp; > > cpu_dev = get_cpu_device(policy->cpu); > if (!cpu_dev) { > @@ -237,6 +239,7 @@ static int cpufreq_init(struct cpufreq_policy *policy) > ret = -EPROBE_DEFER; > goto out_free_opp; > } > + nr_opp = ret; > > if (fallback) { > cpumask_setall(policy->cpus); > @@ -280,6 +283,8 @@ static int cpufreq_init(struct cpufreq_policy *policy) > policy->cpuinfo.transition_latency = transition_latency; > policy->dvfs_possible_from_any_cpu = true; > > + em_register_perf_domain(policy->cpus, nr_opp, &em_cb); The perf domain is registered here but not unregistered in exit() (there is currently no API do do this). When init() is called for the second (or third, ...) time em_register_perf_domain() is called again, though it skips the registration. To make things more explicit and not rely on internal behavior of em_register_perf_domain() you could place the registration inside an 'if (!em_cpu_get(cpu))' branch. You should probably check the return value of em_register_perf_domain() and at least print a warning in case of failure. This would require to call em_register_perf_domain() only when CONFIG_ENERGY_MODEL=y (the function returns -EINVAL otherwise). I think this patch will result in error messages at registration on platforms that use the cpufreq-dt driver and don't specify 'dynamic-power-coefficient' for the CPUs in the DT. Not sure if that's a problem as long as the cpufreq initialization succeeds regardless, it could be seen as a not-so-gentle nudge to add the values. Cheers Matthias
On 28-01-19, 16:55, Quentin Perret wrote: > diff --git a/drivers/opp/of.c b/drivers/opp/of.c > index 06f0f632ec47..7572a2eb2fd4 100644 > --- a/drivers/opp/of.c > +++ b/drivers/opp/of.c > @@ -1047,3 +1047,63 @@ struct device_node *dev_pm_opp_get_of_node(struct dev_pm_opp *opp) > return of_node_get(opp->np); > } > EXPORT_SYMBOL_GPL(dev_pm_opp_get_of_node); > + > +/** > + * of_dev_pm_opp_get_cpu_power() - Estimates the power of a CPU > + * @mW: pointer to the power estimate in milli-watts > + * @KHz: pointer to the OPP's frequency, in kilo-hertz > + * @cpu: CPU for which power needs to be estimated > + * > + * Computes the power estimated by @CPU at the first OPP above @KHz (ceil), > + * and updates @KHz and @mW accordingly. > + * > + * The power is estimated as P = C * V^2 * f, with C the CPU's capacitance > + * (read from the 'dynamic-power-coefficient' devicetree binding) and V and f > + * respectively the voltage and frequency of the OPP. > + * > + * Return: -ENODEV if the CPU device cannot be found, -EINVAL if the power > + * calculation failed because of missing parameters, 0 otherwise. > + */ > +int of_dev_pm_opp_get_cpu_power(unsigned long *mW, unsigned long *KHz, int cpu) I would suggest to change the return type to unsigned long and return 0 for errors and positive values for mW. > +{ > + unsigned long mV, Hz, MHz; MHz is used only once, I think you can get rid of it just open code the division by 1000000. Maybe remove Hz as well and directly use kHz. > + struct device *cpu_dev; > + struct dev_pm_opp *opp; > + struct device_node *np; > + u32 cap; > + u64 tmp; Name this mW with the above changes. > + int ret; > + > + cpu_dev = get_cpu_device(cpu); > + if (!cpu_dev) > + return -ENODEV; > + > + np = of_node_get(cpu_dev->of_node); > + if (!np) > + return -EINVAL; > + > + ret = of_property_read_u32(np, "dynamic-power-coefficient", &cap); > + of_node_put(np); > + if (ret) > + return -EINVAL; > + > + Hz = *KHz * 1000; > + opp = dev_pm_opp_find_freq_ceil(cpu_dev, &Hz); > + if (IS_ERR(opp)) > + return -EINVAL; > + > + mV = dev_pm_opp_get_voltage(opp) / 1000; > + dev_pm_opp_put(opp); > + if (!mV) > + return -EINVAL; > + > + MHz = Hz / 1000000; > + tmp = (u64)cap * mV * mV * MHz; > + do_div(tmp, 1000000000); > + > + *mW = (unsigned long)tmp; > + *KHz = Hz / 1000; > + > + return 0; > +} > +EXPORT_SYMBOL_GPL(of_dev_pm_opp_get_cpu_power); > diff --git a/include/linux/pm_opp.h b/include/linux/pm_opp.h > index 0a2a88e5a383..fedde14f5187 100644 > --- a/include/linux/pm_opp.h > +++ b/include/linux/pm_opp.h > @@ -322,6 +322,7 @@ int dev_pm_opp_of_get_sharing_cpus(struct device *cpu_dev, struct cpumask *cpuma > struct device_node *dev_pm_opp_of_get_opp_desc_node(struct device *dev); > struct device_node *dev_pm_opp_get_of_node(struct dev_pm_opp *opp); > int of_get_required_opp_performance_state(struct device_node *np, int index); > +int of_dev_pm_opp_get_cpu_power(unsigned long *mW, unsigned long *KHz, int cpu); > #else > static inline int dev_pm_opp_of_add_table(struct device *dev) > { > @@ -364,6 +365,10 @@ static inline int of_get_required_opp_performance_state(struct device_node *np, > { > return -ENOTSUPP; > } Add a blank line here. We missed adding the same before of_get_required_opp_performance_state() earlier, maybe add the new routine before of_get_required_opp_performance_state() and add blank line before and after it :) > +static inline int of_dev_pm_opp_get_cpu_power(unsigned long *mW, unsigned long *KHz, int cpu) > +{ > + return -ENOTSUPP; > +} > #endif > > #endif /* __LINUX_OPP_H__ */ > -- > 2.20.1
On 28-01-19, 11:36, Matthias Kaehlcke wrote: > I think this patch will result in error messages at registration on > platforms that use the cpufreq-dt driver and don't specify > 'dynamic-power-coefficient' for the CPUs in the DT. Not sure if that's > a problem as long as the cpufreq initialization succeeds regardless, > it could be seen as a not-so-gentle nudge to add the values. That wouldn't be acceptable.
On 28-01-19, 16:55, Quentin Perret wrote: > Now that PM_OPP provides a helper function to estimate the power > consumed by CPUs, make sure to try and register an Energy Model (EM) > from cpufreq-dt, hence ensuring interested subsystems (the task > scheduler, for example) can make use of that information when available. > > Signed-off-by: Quentin Perret <quentin.perret@arm.com> > --- > drivers/cpufreq/cpufreq-dt.c | 7 ++++++- > 1 file changed, 6 insertions(+), 1 deletion(-) > > diff --git a/drivers/cpufreq/cpufreq-dt.c b/drivers/cpufreq/cpufreq-dt.c > index e58bfcb1169e..7556e07e7a9f 100644 > --- a/drivers/cpufreq/cpufreq-dt.c > +++ b/drivers/cpufreq/cpufreq-dt.c > @@ -16,6 +16,7 @@ > #include <linux/cpu_cooling.h> > #include <linux/cpufreq.h> > #include <linux/cpumask.h> > +#include <linux/energy_model.h> > #include <linux/err.h> > #include <linux/module.h> > #include <linux/of.h> > @@ -152,6 +153,7 @@ static int resources_available(void) > > static int cpufreq_init(struct cpufreq_policy *policy) > { > + struct em_data_callback em_cb = EM_DATA_CB(of_dev_pm_opp_get_cpu_power); Hmm, so all the earlier comments regarding prototype of this routine are invalid because of this. > struct cpufreq_frequency_table *freq_table; > struct opp_table *opp_table = NULL; > struct private_data *priv; > @@ -160,7 +162,7 @@ static int cpufreq_init(struct cpufreq_policy *policy) > unsigned int transition_latency; > bool fallback = false; > const char *name; > - int ret; > + int ret, nr_opp; > > cpu_dev = get_cpu_device(policy->cpu); > if (!cpu_dev) { > @@ -237,6 +239,7 @@ static int cpufreq_init(struct cpufreq_policy *policy) > ret = -EPROBE_DEFER; > goto out_free_opp; > } > + nr_opp = ret; Instead use nr_opp instead of ret while calling dev_pm_opp_get_opp_count(). > > if (fallback) { > cpumask_setall(policy->cpus); > @@ -280,6 +283,8 @@ static int cpufreq_init(struct cpufreq_policy *policy) > policy->cpuinfo.transition_latency = transition_latency; > policy->dvfs_possible_from_any_cpu = true; > > + em_register_perf_domain(policy->cpus, nr_opp, &em_cb); > + > return 0; > > out_free_cpufreq_table: > -- > 2.20.1
On 28-01-19, 16:55, Quentin Perret wrote: > Now that PM_OPP provides a helper function to estimate the power > consumed by CPUs, make sure to try and register an Energy Model (EM) > from scpi-cpufreq, hence ensuring interested subsystems (the task > scheduler, for example) can make use of that information when available. > > Signed-off-by: Quentin Perret <quentin.perret@arm.com> > --- > drivers/cpufreq/scpi-cpufreq.c | 8 +++++++- > 1 file changed, 7 insertions(+), 1 deletion(-) > > diff --git a/drivers/cpufreq/scpi-cpufreq.c b/drivers/cpufreq/scpi-cpufreq.c > index 87a98ec77773..05fc7448f5cb 100644 > --- a/drivers/cpufreq/scpi-cpufreq.c > +++ b/drivers/cpufreq/scpi-cpufreq.c > @@ -23,6 +23,7 @@ > #include <linux/cpufreq.h> > #include <linux/cpumask.h> > #include <linux/cpu_cooling.h> > +#include <linux/energy_model.h> > #include <linux/export.h> > #include <linux/module.h> > #include <linux/of_platform.h> > @@ -98,11 +99,12 @@ scpi_get_sharing_cpus(struct device *cpu_dev, struct cpumask *cpumask) > > static int scpi_cpufreq_init(struct cpufreq_policy *policy) > { > - int ret; > + int ret, nr_opp; > unsigned int latency; > struct device *cpu_dev; > struct scpi_data *priv; > struct cpufreq_frequency_table *freq_table; > + struct em_data_callback em_cb = EM_DATA_CB(of_dev_pm_opp_get_cpu_power); > > cpu_dev = get_cpu_device(policy->cpu); > if (!cpu_dev) { > @@ -135,6 +137,7 @@ static int scpi_cpufreq_init(struct cpufreq_policy *policy) > ret = -EPROBE_DEFER; > goto out_free_opp; > } > + nr_opp = ret; Same comment here.
Hi Matthias, On Monday 28 Jan 2019 at 11:02:51 (-0800), Matthias Kaehlcke wrote: > > +/** > > + * of_dev_pm_opp_get_cpu_power() - Estimates the power of a CPU > > + * @mW: pointer to the power estimate in milli-watts > > + * @KHz: pointer to the OPP's frequency, in kilo-hertz > > nit: should be kHz Right, I can change that :-) > > > + * @cpu: CPU for which power needs to be estimated > > + * > > + * Computes the power estimated by @CPU at the first OPP above @KHz (ceil), > > + * and updates @KHz and @mW accordingly. > > + * > > + * The power is estimated as P = C * V^2 * f, with C the CPU's capacitance > > + * (read from the 'dynamic-power-coefficient' devicetree binding) and V and f > > + * respectively the voltage and frequency of the OPP. > > + * > > + * Return: -ENODEV if the CPU device cannot be found, -EINVAL if the power > > + * calculation failed because of missing parameters, 0 otherwise. > > + */ > > +int of_dev_pm_opp_get_cpu_power(unsigned long *mW, unsigned long *KHz, int cpu) > > I think it is more common to put the input parameters first, then the > output ones, i.e. cpu, kHz, mW. Hmm, the issue is this must match the expectations of the EM framework: https://elixir.bootlin.com/linux/v5.0-rc4/source/include/linux/energy_model.h#L62 So, I don't really have a choice. Or we can change the core code if this _really_ is a problem -- we don't have users yet so now is the best time to do so I guess ... > > > +{ > > + unsigned long mV, Hz, MHz; > > + struct device *cpu_dev; > > + struct dev_pm_opp *opp; > > + struct device_node *np; > > + u32 cap; > > + u64 tmp; > > + int ret; > > + > > + cpu_dev = get_cpu_device(cpu); > > + if (!cpu_dev) > > + return -ENODEV; > > + > > + np = of_node_get(cpu_dev->of_node); > > + if (!np) > > + return -EINVAL; > > + > > + ret = of_property_read_u32(np, "dynamic-power-coefficient", &cap); > > + of_node_put(np); > > + if (ret) > > + return -EINVAL; > > + > > + Hz = *KHz * 1000; > > + opp = dev_pm_opp_find_freq_ceil(cpu_dev, &Hz); > > + if (IS_ERR(opp)) > > + return -EINVAL; > > + > > + mV = dev_pm_opp_get_voltage(opp) / 1000; > > + dev_pm_opp_put(opp); > > + if (!mV) > > + return -EINVAL; > > + > > + MHz = Hz / 1000000; > > + tmp = (u64)cap * mV * mV * MHz; > > + do_div(tmp, 1000000000); > > + > > + *mW = (unsigned long)tmp; > > + *KHz = Hz / 1000; > > + > > + return 0; > > +} > > +EXPORT_SYMBOL_GPL(of_dev_pm_opp_get_cpu_power); > > diff --git a/include/linux/pm_opp.h b/include/linux/pm_opp.h > > index 0a2a88e5a383..fedde14f5187 100644 > > --- a/include/linux/pm_opp.h > > +++ b/include/linux/pm_opp.h > > @@ -322,6 +322,7 @@ int dev_pm_opp_of_get_sharing_cpus(struct device *cpu_dev, struct cpumask *cpuma > > struct device_node *dev_pm_opp_of_get_opp_desc_node(struct device *dev); > > struct device_node *dev_pm_opp_get_of_node(struct dev_pm_opp *opp); > > int of_get_required_opp_performance_state(struct device_node *np, int index); > > +int of_dev_pm_opp_get_cpu_power(unsigned long *mW, unsigned long *KHz, int cpu); > > #else > > static inline int dev_pm_opp_of_add_table(struct device *dev) > > { > > @@ -364,6 +365,10 @@ static inline int of_get_required_opp_performance_state(struct device_node *np, > > { > > return -ENOTSUPP; > > } > > +static inline int of_dev_pm_opp_get_cpu_power(unsigned long *mW, unsigned long *KHz, int cpu) > > +{ > > + return -ENOTSUPP; > > +} > > #endif > > > > #endif /* __LINUX_OPP_H__ */ > > Besides the nits above: > > Reviewed-by: Matthias Kaehlcke <mka@chromium.org> > Tested-by: Matthias Kaehlcke <mka@chromium.org> Thank you ! Quentin
Hi Viresh, On Tuesday 29 Jan 2019 at 10:40:30 (+0530), Viresh Kumar wrote: > > +int of_dev_pm_opp_get_cpu_power(unsigned long *mW, unsigned long *KHz, int cpu) > > I would suggest to change the return type to unsigned long and return 0 for > errors and positive values for mW. Well, I can't really do that -- this must match the API of PM_EM: https://elixir.bootlin.com/linux/v5.0-rc4/source/include/linux/energy_model.h#L62 I'll make sure to make that explicit in the commit message for v2. > > > +{ > > + unsigned long mV, Hz, MHz; > > MHz is used only once, I think you can get rid of it just open code the division > by 1000000. Maybe remove Hz as well and directly use kHz. I would expect/hope the compiler will optimize things like that for me but yes, I can try to reduce a bit the temp variables if you prefer. > > > + struct device *cpu_dev; > > + struct dev_pm_opp *opp; > > + struct device_node *np; > > + u32 cap; > > + u64 tmp; > > Name this mW with the above changes. So that doesn't really work given what I mentioned above. > > + int ret; > > + > > + cpu_dev = get_cpu_device(cpu); > > + if (!cpu_dev) > > + return -ENODEV; > > + > > + np = of_node_get(cpu_dev->of_node); > > + if (!np) > > + return -EINVAL; > > + > > + ret = of_property_read_u32(np, "dynamic-power-coefficient", &cap); > > + of_node_put(np); > > + if (ret) > > + return -EINVAL; > > + > > + Hz = *KHz * 1000; > > + opp = dev_pm_opp_find_freq_ceil(cpu_dev, &Hz); > > + if (IS_ERR(opp)) > > + return -EINVAL; > > + > > + mV = dev_pm_opp_get_voltage(opp) / 1000; > > + dev_pm_opp_put(opp); > > + if (!mV) > > + return -EINVAL; > > + > > + MHz = Hz / 1000000; > > + tmp = (u64)cap * mV * mV * MHz; > > + do_div(tmp, 1000000000); > > + > > + *mW = (unsigned long)tmp; > > + *KHz = Hz / 1000; > > + > > + return 0; > > +} > > +EXPORT_SYMBOL_GPL(of_dev_pm_opp_get_cpu_power); > > diff --git a/include/linux/pm_opp.h b/include/linux/pm_opp.h > > index 0a2a88e5a383..fedde14f5187 100644 > > --- a/include/linux/pm_opp.h > > +++ b/include/linux/pm_opp.h > > @@ -322,6 +322,7 @@ int dev_pm_opp_of_get_sharing_cpus(struct device *cpu_dev, struct cpumask *cpuma > > struct device_node *dev_pm_opp_of_get_opp_desc_node(struct device *dev); > > struct device_node *dev_pm_opp_get_of_node(struct dev_pm_opp *opp); > > int of_get_required_opp_performance_state(struct device_node *np, int index); > > +int of_dev_pm_opp_get_cpu_power(unsigned long *mW, unsigned long *KHz, int cpu); > > #else > > static inline int dev_pm_opp_of_add_table(struct device *dev) > > { > > @@ -364,6 +365,10 @@ static inline int of_get_required_opp_performance_state(struct device_node *np, > > { > > return -ENOTSUPP; > > } > > Add a blank line here. We missed adding the same before > of_get_required_opp_performance_state() earlier, maybe add the new routine > before of_get_required_opp_performance_state() and add blank line before and > after it :) Sounds good ! > > > +static inline int of_dev_pm_opp_get_cpu_power(unsigned long *mW, unsigned long *KHz, int cpu) > > +{ > > + return -ENOTSUPP; > > +} > > #endif > > > > #endif /* __LINUX_OPP_H__ */ > > -- > > 2.20.1 > > -- > viresh Thanks, Quentin
On Tuesday 29 Jan 2019 at 11:00:41 (+0530), Viresh Kumar wrote: > On 28-01-19, 16:55, Quentin Perret wrote: > > Now that PM_OPP provides a helper function to estimate the power > > consumed by CPUs, make sure to try and register an Energy Model (EM) > > from cpufreq-dt, hence ensuring interested subsystems (the task > > scheduler, for example) can make use of that information when available. > > > > Signed-off-by: Quentin Perret <quentin.perret@arm.com> > > --- > > drivers/cpufreq/cpufreq-dt.c | 7 ++++++- > > 1 file changed, 6 insertions(+), 1 deletion(-) > > > > diff --git a/drivers/cpufreq/cpufreq-dt.c b/drivers/cpufreq/cpufreq-dt.c > > index e58bfcb1169e..7556e07e7a9f 100644 > > --- a/drivers/cpufreq/cpufreq-dt.c > > +++ b/drivers/cpufreq/cpufreq-dt.c > > @@ -16,6 +16,7 @@ > > #include <linux/cpu_cooling.h> > > #include <linux/cpufreq.h> > > #include <linux/cpumask.h> > > +#include <linux/energy_model.h> > > #include <linux/err.h> > > #include <linux/module.h> > > #include <linux/of.h> > > @@ -152,6 +153,7 @@ static int resources_available(void) > > > > static int cpufreq_init(struct cpufreq_policy *policy) > > { > > + struct em_data_callback em_cb = EM_DATA_CB(of_dev_pm_opp_get_cpu_power); > > Hmm, so all the earlier comments regarding prototype of this routine are invalid > because of this. Right. > > > struct cpufreq_frequency_table *freq_table; > > struct opp_table *opp_table = NULL; > > struct private_data *priv; > > @@ -160,7 +162,7 @@ static int cpufreq_init(struct cpufreq_policy *policy) > > unsigned int transition_latency; > > bool fallback = false; > > const char *name; > > - int ret; > > + int ret, nr_opp; > > > > cpu_dev = get_cpu_device(policy->cpu); > > if (!cpu_dev) { > > @@ -237,6 +239,7 @@ static int cpufreq_init(struct cpufreq_policy *policy) > > ret = -EPROBE_DEFER; > > goto out_free_opp; > > } > > + nr_opp = ret; > > Instead use nr_opp instead of ret while calling dev_pm_opp_get_opp_count(). Will do. > > > > > if (fallback) { > > cpumask_setall(policy->cpus); > > @@ -280,6 +283,8 @@ static int cpufreq_init(struct cpufreq_policy *policy) > > policy->cpuinfo.transition_latency = transition_latency; > > policy->dvfs_possible_from_any_cpu = true; > > > > + em_register_perf_domain(policy->cpus, nr_opp, &em_cb); > > + > > return 0; > > > > out_free_cpufreq_table: > > -- > > 2.20.1 > > -- > viresh Thanks, Quentin
On Tuesday 29 Jan 2019 at 10:51:44 (+0530), Viresh Kumar wrote: > On 28-01-19, 11:36, Matthias Kaehlcke wrote: > > I think this patch will result in error messages at registration on > > platforms that use the cpufreq-dt driver and don't specify > > 'dynamic-power-coefficient' for the CPUs in the DT. Not sure if that's > > a problem as long as the cpufreq initialization succeeds regardless, > > it could be seen as a not-so-gentle nudge to add the values. > > That wouldn't be acceptable. Fair enough. What I can propose in this case is to have in PM_OPP a helper called 'dev_pm_opp_of_register_em()' or something like this. This function will check all prerequisites are present (we have the right values in DT, and so on) and then call (or not) em_register_perf_domain(). Then we can make the CPUFreq drivers use that instead of calling em_register_perf_domain() directly. That would also make it easy to implement Matthias' suggestion to not call em_register_perf_domain() if an EM is already present. Would that work ? Thanks, Quentin
Hi Quentin, On Tue, Jan 29, 2019 at 09:03:34AM +0000, Quentin Perret wrote: > Hi Matthias, > > On Monday 28 Jan 2019 at 11:02:51 (-0800), Matthias Kaehlcke wrote: > > > +/** > > > + * of_dev_pm_opp_get_cpu_power() - Estimates the power of a CPU > > > + * @mW: pointer to the power estimate in milli-watts > > > + * @KHz: pointer to the OPP's frequency, in kilo-hertz > > > > nit: should be kHz > > Right, I can change that :-) > > > > > > + * @cpu: CPU for which power needs to be estimated > > > + * > > > + * Computes the power estimated by @CPU at the first OPP above @KHz (ceil), > > > + * and updates @KHz and @mW accordingly. > > > + * > > > + * The power is estimated as P = C * V^2 * f, with C the CPU's capacitance > > > + * (read from the 'dynamic-power-coefficient' devicetree binding) and V and f > > > + * respectively the voltage and frequency of the OPP. > > > + * > > > + * Return: -ENODEV if the CPU device cannot be found, -EINVAL if the power > > > + * calculation failed because of missing parameters, 0 otherwise. > > > + */ > > > +int of_dev_pm_opp_get_cpu_power(unsigned long *mW, unsigned long *KHz, int cpu) > > > > I think it is more common to put the input parameters first, then the > > output ones, i.e. cpu, kHz, mW. > > Hmm, the issue is this must match the expectations of the EM framework: > > https://elixir.bootlin.com/linux/v5.0-rc4/source/include/linux/energy_model.h#L62 > > So, I don't really have a choice. Or we can change the core code if this > _really_ is a problem -- we don't have users yet so now is the best time > to do so I guess ... I see. I think it would be preferable to follow the common scheme, but it's also not a major issue, up to you whether you want to change the definition of the callback. It's certainly true that now would be the best time for a change. Cheers Matthias
On 29-01-19, 09:15, Quentin Perret wrote: > On Tuesday 29 Jan 2019 at 10:51:44 (+0530), Viresh Kumar wrote: > > On 28-01-19, 11:36, Matthias Kaehlcke wrote: > > > I think this patch will result in error messages at registration on > > > platforms that use the cpufreq-dt driver and don't specify > > > 'dynamic-power-coefficient' for the CPUs in the DT. Not sure if that's > > > a problem as long as the cpufreq initialization succeeds regardless, > > > it could be seen as a not-so-gentle nudge to add the values. > > > > That wouldn't be acceptable. > > Fair enough. What I can propose in this case is to have in PM_OPP a > helper called 'dev_pm_opp_of_register_em()' or something like this. This > function will check all prerequisites are present (we have the right > values in DT, and so on) and then call (or not) em_register_perf_domain(). > Then we can make the CPUFreq drivers use that instead of calling > em_register_perf_domain() directly. That should be fine. > That would also make it easy to implement Matthias' suggestion to not > call em_register_perf_domain() if an EM is already present. So you will track registration state within the OPP core for that ? Sorry but that doesn't sound right. What's wrong with having an unregister helper in energy-model to keep proper code flow everywhere ?
Hi Viresh, On Wednesday 30 Jan 2019 at 10:48:06 (+0530), Viresh Kumar wrote: > On 29-01-19, 09:15, Quentin Perret wrote: > > On Tuesday 29 Jan 2019 at 10:51:44 (+0530), Viresh Kumar wrote: > > > On 28-01-19, 11:36, Matthias Kaehlcke wrote: > > > > I think this patch will result in error messages at registration on > > > > platforms that use the cpufreq-dt driver and don't specify > > > > 'dynamic-power-coefficient' for the CPUs in the DT. Not sure if that's > > > > a problem as long as the cpufreq initialization succeeds regardless, > > > > it could be seen as a not-so-gentle nudge to add the values. > > > > > > That wouldn't be acceptable. > > > > Fair enough. What I can propose in this case is to have in PM_OPP a > > helper called 'dev_pm_opp_of_register_em()' or something like this. This > > function will check all prerequisites are present (we have the right > > values in DT, and so on) and then call (or not) em_register_perf_domain(). > > Then we can make the CPUFreq drivers use that instead of calling > > em_register_perf_domain() directly. > > That should be fine. > > > That would also make it easy to implement Matthias' suggestion to not > > call em_register_perf_domain() if an EM is already present. > > So you will track registration state within the OPP core for that ? What I had in mind is something as simple as: void of_dev_pm_opp_register_em(struct cpumask *cpus) { /* Bail out if an EM is there */ if (em_cpu_get(cpumask_first(cpus))) return; /* Check prerequisites: dpc coeff in DT, ... */ ... em_register_perf_domain(...); } IIUC, Matthias' point was that if the EM is already registered, there is no good reason to call em_register_perf_domain() again. Now, that should in fact be harmless because em_register_perf_domain() already does that check. It's just cleaner and easier to understand from a conceptual standpoint to not call that function several times for no reason I assume. > Sorry but that doesn't sound right. What's wrong with having an > unregister helper in energy-model to keep proper code flow everywhere > ? For the EM we basically allocate the tables in memory once and for all at boot time and never touch them again. That makes it easy for users (the scheduler as of now, IPA soon) to access them without messing around with RCU or so. So there isn't really a concept or unregistering a pd right now. Thanks, Quentin
On 30-01-19, 09:12, Quentin Perret wrote: > What I had in mind is something as simple as: > > void of_dev_pm_opp_register_em(struct cpumask *cpus) > { > /* Bail out if an EM is there */ > if (em_cpu_get(cpumask_first(cpus))) > return; > > /* Check prerequisites: dpc coeff in DT, ... */ > ... > > em_register_perf_domain(...); > } > > IIUC, Matthias' point was that if the EM is already registered, there is > no good reason to call em_register_perf_domain() again. Now, that should > in fact be harmless because em_register_perf_domain() already does that > check. It's just cleaner and easier to understand from a conceptual > standpoint to not call that function several times for no reason I > assume. If there is no good reason to call em_register_perf_domain() several times, then the same applies to of_dev_pm_opp_register_em() as well, isn't it ? This is init code anyway isn't going to run a lot, so I wouldn't suggest adding any such (duplicate) checks in of_dev_pm_opp_register_em().
On Wednesday 30 Jan 2019 at 15:47:15 (+0530), Viresh Kumar wrote: > On 30-01-19, 09:12, Quentin Perret wrote: > > What I had in mind is something as simple as: > > > > void of_dev_pm_opp_register_em(struct cpumask *cpus) > > { > > /* Bail out if an EM is there */ > > if (em_cpu_get(cpumask_first(cpus))) > > return; > > > > /* Check prerequisites: dpc coeff in DT, ... */ > > ... > > > > em_register_perf_domain(...); > > } > > > > IIUC, Matthias' point was that if the EM is already registered, there is > > no good reason to call em_register_perf_domain() again. Now, that should > > in fact be harmless because em_register_perf_domain() already does that > > check. It's just cleaner and easier to understand from a conceptual > > standpoint to not call that function several times for no reason I > > assume. > > If there is no good reason to call em_register_perf_domain() several > times, then the same applies to of_dev_pm_opp_register_em() as well, > isn't it ? Hrmpf, that is true ... :-) > This is init code anyway isn't going to run a lot, so I wouldn't > suggest adding any such (duplicate) checks in > of_dev_pm_opp_register_em(). Fair enough, I'll post a v2 w/o the check and we'll see if others have complaints. Thanks, Quentin