1/* SPDX-License-Identifier: GPL-2.0 */ 2#ifndef _LINUX_ENERGY_MODEL_H 3#define _LINUX_ENERGY_MODEL_H 4#include <linux/cpumask.h> 5#include <linux/device.h> 6#include <linux/jump_label.h> 7#include <linux/kobject.h> 8#include <linux/rcupdate.h> 9#include <linux/sched/cpufreq.h> 10#include <linux/sched/topology.h> 11#include <linux/types.h> 12 13/** 14 * em_perf_state - Performance state of a performance domain 15 * @frequency: The frequency in KHz, for consistency with CPUFreq 16 * @power: The power consumed at this level (by 1 CPU or by a registered 17 * device). It can be a total power: static and dynamic. 18 * @cost: The cost coefficient associated with this level, used during 19 * energy calculation. Equal to: power * max_frequency / frequency 20 */ 21struct em_perf_state { 22 unsigned long frequency; 23 unsigned long power; 24 unsigned long cost; 25}; 26 27/** 28 * em_perf_domain - Performance domain 29 * @table: List of performance states, in ascending order 30 * @nr_perf_states: Number of performance states 31 * @milliwatts: Flag indicating the power values are in milli-Watts 32 * or some other scale. 33 * @cpus: Cpumask covering the CPUs of the domain. It's here 34 * for performance reasons to avoid potential cache 35 * misses during energy calculations in the scheduler 36 * and simplifies allocating/freeing that memory region. 37 * 38 * In case of CPU device, a "performance domain" represents a group of CPUs 39 * whose performance is scaled together. All CPUs of a performance domain 40 * must have the same micro-architecture. Performance domains often have 41 * a 1-to-1 mapping with CPUFreq policies. In case of other devices the @cpus 42 * field is unused. 43 */ 44struct em_perf_domain { 45 struct em_perf_state *table; 46 int nr_perf_states; 47 int milliwatts; 48 unsigned long cpus[]; 49}; 50 51#define em_span_cpus(em) (to_cpumask((em)->cpus)) 52 53#ifdef CONFIG_ENERGY_MODEL 54#define EM_MAX_POWER 0xFFFF 55 56struct em_data_callback { 57 /** 58 * active_power() - Provide power at the next performance state of 59 * a device 60 * @power : Active power at the performance state 61 * (modified) 62 * @freq : Frequency at the performance state in kHz 63 * (modified) 64 * @dev : Device for which we do this operation (can be a CPU) 65 * 66 * active_power() must find the lowest performance state of 'dev' above 67 * 'freq' and update 'power' and 'freq' to the matching active power 68 * and frequency. 69 * 70 * In case of CPUs, the power is the one of a single CPU in the domain, 71 * expressed in milli-Watts or an abstract scale. It is expected to 72 * fit in the [0, EM_MAX_POWER] range. 73 * 74 * Return 0 on success. 75 */ 76 int (*active_power)(unsigned long *power, unsigned long *freq, 77 struct device *dev); 78}; 79#define EM_DATA_CB(_active_power_cb) { .active_power = &_active_power_cb } 80 81struct em_perf_domain *em_cpu_get(int cpu); 82struct em_perf_domain *em_pd_get(struct device *dev); 83int em_dev_register_perf_domain(struct device *dev, unsigned int nr_states, 84 struct em_data_callback *cb, cpumask_t *span, 85 bool milliwatts); 86void em_dev_unregister_perf_domain(struct device *dev); 87 88/** 89 * em_cpu_energy() - Estimates the energy consumed by the CPUs of a 90 performance domain 91 * @pd : performance domain for which energy has to be estimated 92 * @max_util : highest utilization among CPUs of the domain 93 * @sum_util : sum of the utilization of all CPUs in the domain 94 * 95 * This function must be used only for CPU devices. There is no validation, 96 * i.e. if the EM is a CPU type and has cpumask allocated. It is called from 97 * the scheduler code quite frequently and that is why there is not checks. 98 * 99 * Return: the sum of the energy consumed by the CPUs of the domain assuming 100 * a capacity state satisfying the max utilization of the domain. 101 */ 102static inline unsigned long em_cpu_energy(struct em_perf_domain *pd, 103 unsigned long max_util, unsigned long sum_util) 104{ 105 unsigned long freq, scale_cpu; 106 struct em_perf_state *ps; 107 int i, cpu; 108 109 if (!sum_util) 110 return 0; 111 112 /* 113 * In order to predict the performance state, map the utilization of 114 * the most utilized CPU of the performance domain to a requested 115 * frequency, like schedutil. 116 */ 117 cpu = cpumask_first(to_cpumask(pd->cpus)); 118 scale_cpu = arch_scale_cpu_capacity(cpu); 119 ps = &pd->table[pd->nr_perf_states - 1]; 120 freq = map_util_freq(max_util, ps->frequency, scale_cpu); 121 122 /* 123 * Find the lowest performance state of the Energy Model above the 124 * requested frequency. 125 */ 126 for (i = 0; i < pd->nr_perf_states; i++) { 127 ps = &pd->table[i]; 128 if (ps->frequency >= freq) 129 break; 130 } 131 132 /* 133 * The capacity of a CPU in the domain at the performance state (ps) 134 * can be computed as: 135 * 136 * ps->freq * scale_cpu 137 * ps->cap = -------------------- (1) 138 * cpu_max_freq 139 * 140 * So, ignoring the costs of idle states (which are not available in 141 * the EM), the energy consumed by this CPU at that performance state 142 * is estimated as: 143 * 144 * ps->power * cpu_util 145 * cpu_nrg = -------------------- (2) 146 * ps->cap 147 * 148 * since 'cpu_util / ps->cap' represents its percentage of busy time. 149 * 150 * NOTE: Although the result of this computation actually is in 151 * units of power, it can be manipulated as an energy value 152 * over a scheduling period, since it is assumed to be 153 * constant during that interval. 154 * 155 * By injecting (1) in (2), 'cpu_nrg' can be re-expressed as a product 156 * of two terms: 157 * 158 * ps->power * cpu_max_freq cpu_util 159 * cpu_nrg = ------------------------ * --------- (3) 160 * ps->freq scale_cpu 161 * 162 * The first term is static, and is stored in the em_perf_state struct 163 * as 'ps->cost'. 164 * 165 * Since all CPUs of the domain have the same micro-architecture, they 166 * share the same 'ps->cost', and the same CPU capacity. Hence, the 167 * total energy of the domain (which is the simple sum of the energy of 168 * all of its CPUs) can be factorized as: 169 * 170 * ps->cost * \Sum cpu_util 171 * pd_nrg = ------------------------ (4) 172 * scale_cpu 173 */ 174 return ps->cost * sum_util / scale_cpu; 175} 176 177/** 178 * em_pd_nr_perf_states() - Get the number of performance states of a perf. 179 * domain 180 * @pd : performance domain for which this must be done 181 * 182 * Return: the number of performance states in the performance domain table 183 */ 184static inline int em_pd_nr_perf_states(struct em_perf_domain *pd) 185{ 186 return pd->nr_perf_states; 187} 188 189#else 190struct em_data_callback {}; 191#define EM_DATA_CB(_active_power_cb) { } 192 193static inline 194int em_dev_register_perf_domain(struct device *dev, unsigned int nr_states, 195 struct em_data_callback *cb, cpumask_t *span, 196 bool milliwatts) 197{ 198 return -EINVAL; 199} 200static inline void em_dev_unregister_perf_domain(struct device *dev) 201{ 202} 203static inline struct em_perf_domain *em_cpu_get(int cpu) 204{ 205 return NULL; 206} 207static inline struct em_perf_domain *em_pd_get(struct device *dev) 208{ 209 return NULL; 210} 211static inline unsigned long em_cpu_energy(struct em_perf_domain *pd, 212 unsigned long max_util, unsigned long sum_util) 213{ 214 return 0; 215} 216static inline int em_pd_nr_perf_states(struct em_perf_domain *pd) 217{ 218 return 0; 219} 220#endif 221 222#endif 223