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 * struct 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 * @flags: see "em_perf_state flags" description below. 21 */ 22 struct em_perf_state { 23 unsigned long frequency; 24 unsigned long power; 25 unsigned long cost; 26 unsigned long flags; 27 }; 28 29 /* 30 * em_perf_state flags: 31 * 32 * EM_PERF_STATE_INEFFICIENT: The performance state is inefficient. There is 33 * in this em_perf_domain, another performance state with a higher frequency 34 * but a lower or equal power cost. Such inefficient states are ignored when 35 * using em_pd_get_efficient_*() functions. 36 */ 37 #define EM_PERF_STATE_INEFFICIENT BIT(0) 38 39 /** 40 * struct em_perf_domain - Performance domain 41 * @table: List of performance states, in ascending order 42 * @nr_perf_states: Number of performance states 43 * @flags: See "em_perf_domain flags" 44 * @cpus: Cpumask covering the CPUs of the domain. It's here 45 * for performance reasons to avoid potential cache 46 * misses during energy calculations in the scheduler 47 * and simplifies allocating/freeing that memory region. 48 * 49 * In case of CPU device, a "performance domain" represents a group of CPUs 50 * whose performance is scaled together. All CPUs of a performance domain 51 * must have the same micro-architecture. Performance domains often have 52 * a 1-to-1 mapping with CPUFreq policies. In case of other devices the @cpus 53 * field is unused. 54 */ 55 struct em_perf_domain { 56 struct em_perf_state *table; 57 int nr_perf_states; 58 unsigned long flags; 59 unsigned long cpus[]; 60 }; 61 62 /* 63 * em_perf_domain flags: 64 * 65 * EM_PERF_DOMAIN_MILLIWATTS: The power values are in milli-Watts or some 66 * other scale. 67 * 68 * EM_PERF_DOMAIN_SKIP_INEFFICIENCIES: Skip inefficient states when estimating 69 * energy consumption. 70 * 71 * EM_PERF_DOMAIN_ARTIFICIAL: The power values are artificial and might be 72 * created by platform missing real power information 73 */ 74 #define EM_PERF_DOMAIN_MILLIWATTS BIT(0) 75 #define EM_PERF_DOMAIN_SKIP_INEFFICIENCIES BIT(1) 76 #define EM_PERF_DOMAIN_ARTIFICIAL BIT(2) 77 78 #define em_span_cpus(em) (to_cpumask((em)->cpus)) 79 #define em_is_artificial(em) ((em)->flags & EM_PERF_DOMAIN_ARTIFICIAL) 80 81 #ifdef CONFIG_ENERGY_MODEL 82 #define EM_MAX_POWER 0xFFFF 83 84 /* 85 * Increase resolution of energy estimation calculations for 64-bit 86 * architectures. The extra resolution improves decision made by EAS for the 87 * task placement when two Performance Domains might provide similar energy 88 * estimation values (w/o better resolution the values could be equal). 89 * 90 * We increase resolution only if we have enough bits to allow this increased 91 * resolution (i.e. 64-bit). The costs for increasing resolution when 32-bit 92 * are pretty high and the returns do not justify the increased costs. 93 */ 94 #ifdef CONFIG_64BIT 95 #define em_scale_power(p) ((p) * 1000) 96 #else 97 #define em_scale_power(p) (p) 98 #endif 99 100 struct em_data_callback { 101 /** 102 * active_power() - Provide power at the next performance state of 103 * a device 104 * @dev : Device for which we do this operation (can be a CPU) 105 * @power : Active power at the performance state 106 * (modified) 107 * @freq : Frequency at the performance state in kHz 108 * (modified) 109 * 110 * active_power() must find the lowest performance state of 'dev' above 111 * 'freq' and update 'power' and 'freq' to the matching active power 112 * and frequency. 113 * 114 * In case of CPUs, the power is the one of a single CPU in the domain, 115 * expressed in milli-Watts or an abstract scale. It is expected to 116 * fit in the [0, EM_MAX_POWER] range. 117 * 118 * Return 0 on success. 119 */ 120 int (*active_power)(struct device *dev, unsigned long *power, 121 unsigned long *freq); 122 123 /** 124 * get_cost() - Provide the cost at the given performance state of 125 * a device 126 * @dev : Device for which we do this operation (can be a CPU) 127 * @freq : Frequency at the performance state in kHz 128 * @cost : The cost value for the performance state 129 * (modified) 130 * 131 * In case of CPUs, the cost is the one of a single CPU in the domain. 132 * It is expected to fit in the [0, EM_MAX_POWER] range due to internal 133 * usage in EAS calculation. 134 * 135 * Return 0 on success, or appropriate error value in case of failure. 136 */ 137 int (*get_cost)(struct device *dev, unsigned long freq, 138 unsigned long *cost); 139 }; 140 #define EM_SET_ACTIVE_POWER_CB(em_cb, cb) ((em_cb).active_power = cb) 141 #define EM_ADV_DATA_CB(_active_power_cb, _cost_cb) \ 142 { .active_power = _active_power_cb, \ 143 .get_cost = _cost_cb } 144 #define EM_DATA_CB(_active_power_cb) \ 145 EM_ADV_DATA_CB(_active_power_cb, NULL) 146 147 struct em_perf_domain *em_cpu_get(int cpu); 148 struct em_perf_domain *em_pd_get(struct device *dev); 149 int em_dev_register_perf_domain(struct device *dev, unsigned int nr_states, 150 struct em_data_callback *cb, cpumask_t *span, 151 bool milliwatts); 152 void em_dev_unregister_perf_domain(struct device *dev); 153 154 /** 155 * em_pd_get_efficient_state() - Get an efficient performance state from the EM 156 * @pd : Performance domain for which we want an efficient frequency 157 * @freq : Frequency to map with the EM 158 * 159 * It is called from the scheduler code quite frequently and as a consequence 160 * doesn't implement any check. 161 * 162 * Return: An efficient performance state, high enough to meet @freq 163 * requirement. 164 */ 165 static inline 166 struct em_perf_state *em_pd_get_efficient_state(struct em_perf_domain *pd, 167 unsigned long freq) 168 { 169 struct em_perf_state *ps; 170 int i; 171 172 for (i = 0; i < pd->nr_perf_states; i++) { 173 ps = &pd->table[i]; 174 if (ps->frequency >= freq) { 175 if (pd->flags & EM_PERF_DOMAIN_SKIP_INEFFICIENCIES && 176 ps->flags & EM_PERF_STATE_INEFFICIENT) 177 continue; 178 break; 179 } 180 } 181 182 return ps; 183 } 184 185 /** 186 * em_cpu_energy() - Estimates the energy consumed by the CPUs of a 187 * performance domain 188 * @pd : performance domain for which energy has to be estimated 189 * @max_util : highest utilization among CPUs of the domain 190 * @sum_util : sum of the utilization of all CPUs in the domain 191 * @allowed_cpu_cap : maximum allowed CPU capacity for the @pd, which 192 * might reflect reduced frequency (due to thermal) 193 * 194 * This function must be used only for CPU devices. There is no validation, 195 * i.e. if the EM is a CPU type and has cpumask allocated. It is called from 196 * the scheduler code quite frequently and that is why there is not checks. 197 * 198 * Return: the sum of the energy consumed by the CPUs of the domain assuming 199 * a capacity state satisfying the max utilization of the domain. 200 */ 201 static inline unsigned long em_cpu_energy(struct em_perf_domain *pd, 202 unsigned long max_util, unsigned long sum_util, 203 unsigned long allowed_cpu_cap) 204 { 205 unsigned long freq, scale_cpu; 206 struct em_perf_state *ps; 207 int cpu; 208 209 if (!sum_util) 210 return 0; 211 212 /* 213 * In order to predict the performance state, map the utilization of 214 * the most utilized CPU of the performance domain to a requested 215 * frequency, like schedutil. Take also into account that the real 216 * frequency might be set lower (due to thermal capping). Thus, clamp 217 * max utilization to the allowed CPU capacity before calculating 218 * effective frequency. 219 */ 220 cpu = cpumask_first(to_cpumask(pd->cpus)); 221 scale_cpu = arch_scale_cpu_capacity(cpu); 222 ps = &pd->table[pd->nr_perf_states - 1]; 223 224 max_util = map_util_perf(max_util); 225 max_util = min(max_util, allowed_cpu_cap); 226 freq = map_util_freq(max_util, ps->frequency, scale_cpu); 227 228 /* 229 * Find the lowest performance state of the Energy Model above the 230 * requested frequency. 231 */ 232 ps = em_pd_get_efficient_state(pd, freq); 233 234 /* 235 * The capacity of a CPU in the domain at the performance state (ps) 236 * can be computed as: 237 * 238 * ps->freq * scale_cpu 239 * ps->cap = -------------------- (1) 240 * cpu_max_freq 241 * 242 * So, ignoring the costs of idle states (which are not available in 243 * the EM), the energy consumed by this CPU at that performance state 244 * is estimated as: 245 * 246 * ps->power * cpu_util 247 * cpu_nrg = -------------------- (2) 248 * ps->cap 249 * 250 * since 'cpu_util / ps->cap' represents its percentage of busy time. 251 * 252 * NOTE: Although the result of this computation actually is in 253 * units of power, it can be manipulated as an energy value 254 * over a scheduling period, since it is assumed to be 255 * constant during that interval. 256 * 257 * By injecting (1) in (2), 'cpu_nrg' can be re-expressed as a product 258 * of two terms: 259 * 260 * ps->power * cpu_max_freq cpu_util 261 * cpu_nrg = ------------------------ * --------- (3) 262 * ps->freq scale_cpu 263 * 264 * The first term is static, and is stored in the em_perf_state struct 265 * as 'ps->cost'. 266 * 267 * Since all CPUs of the domain have the same micro-architecture, they 268 * share the same 'ps->cost', and the same CPU capacity. Hence, the 269 * total energy of the domain (which is the simple sum of the energy of 270 * all of its CPUs) can be factorized as: 271 * 272 * ps->cost * \Sum cpu_util 273 * pd_nrg = ------------------------ (4) 274 * scale_cpu 275 */ 276 return ps->cost * sum_util / scale_cpu; 277 } 278 279 /** 280 * em_pd_nr_perf_states() - Get the number of performance states of a perf. 281 * domain 282 * @pd : performance domain for which this must be done 283 * 284 * Return: the number of performance states in the performance domain table 285 */ 286 static inline int em_pd_nr_perf_states(struct em_perf_domain *pd) 287 { 288 return pd->nr_perf_states; 289 } 290 291 #else 292 struct em_data_callback {}; 293 #define EM_ADV_DATA_CB(_active_power_cb, _cost_cb) { } 294 #define EM_DATA_CB(_active_power_cb) { } 295 #define EM_SET_ACTIVE_POWER_CB(em_cb, cb) do { } while (0) 296 297 static inline 298 int em_dev_register_perf_domain(struct device *dev, unsigned int nr_states, 299 struct em_data_callback *cb, cpumask_t *span, 300 bool milliwatts) 301 { 302 return -EINVAL; 303 } 304 static inline void em_dev_unregister_perf_domain(struct device *dev) 305 { 306 } 307 static inline struct em_perf_domain *em_cpu_get(int cpu) 308 { 309 return NULL; 310 } 311 static inline struct em_perf_domain *em_pd_get(struct device *dev) 312 { 313 return NULL; 314 } 315 static inline unsigned long em_cpu_energy(struct em_perf_domain *pd, 316 unsigned long max_util, unsigned long sum_util, 317 unsigned long allowed_cpu_cap) 318 { 319 return 0; 320 } 321 static inline int em_pd_nr_perf_states(struct em_perf_domain *pd) 322 { 323 return 0; 324 } 325 #endif 326 327 #endif 328