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/kref.h>
9 #include <linux/rcupdate.h>
10 #include <linux/sched/cpufreq.h>
11 #include <linux/sched/topology.h>
12 #include <linux/types.h>
13
14 /**
15 * struct em_perf_state - Performance state of a performance domain
16 * @performance: CPU performance (capacity) at a given frequency
17 * @frequency: The frequency in KHz, for consistency with CPUFreq
18 * @power: The power consumed at this level (by 1 CPU or by a registered
19 * device). It can be a total power: static and dynamic.
20 * @cost: The cost coefficient associated with this level, used during
21 * energy calculation. Equal to: power * max_frequency / frequency
22 * @flags: see "em_perf_state flags" description below.
23 */
24 struct em_perf_state {
25 unsigned long performance;
26 unsigned long frequency;
27 unsigned long power;
28 unsigned long cost;
29 unsigned long flags;
30 };
31
32 /*
33 * em_perf_state flags:
34 *
35 * EM_PERF_STATE_INEFFICIENT: The performance state is inefficient. There is
36 * in this em_perf_domain, another performance state with a higher frequency
37 * but a lower or equal power cost. Such inefficient states are ignored when
38 * using em_pd_get_efficient_*() functions.
39 */
40 #define EM_PERF_STATE_INEFFICIENT BIT(0)
41
42 /**
43 * struct em_perf_table - Performance states table
44 * @rcu: RCU used for safe access and destruction
45 * @kref: Reference counter to track the users
46 * @state: List of performance states, in ascending order
47 */
48 struct em_perf_table {
49 struct rcu_head rcu;
50 struct kref kref;
51 struct em_perf_state state[];
52 };
53
54 /**
55 * struct em_perf_domain - Performance domain
56 * @em_table: Pointer to the runtime modifiable em_perf_table
57 * @nr_perf_states: Number of performance states
58 * @flags: See "em_perf_domain flags"
59 * @cpus: Cpumask covering the CPUs of the domain. It's here
60 * for performance reasons to avoid potential cache
61 * misses during energy calculations in the scheduler
62 * and simplifies allocating/freeing that memory region.
63 *
64 * In case of CPU device, a "performance domain" represents a group of CPUs
65 * whose performance is scaled together. All CPUs of a performance domain
66 * must have the same micro-architecture. Performance domains often have
67 * a 1-to-1 mapping with CPUFreq policies. In case of other devices the @cpus
68 * field is unused.
69 */
70 struct em_perf_domain {
71 struct em_perf_table __rcu *em_table;
72 int nr_perf_states;
73 unsigned long flags;
74 unsigned long cpus[];
75 };
76
77 /*
78 * em_perf_domain flags:
79 *
80 * EM_PERF_DOMAIN_MICROWATTS: The power values are in micro-Watts or some
81 * other scale.
82 *
83 * EM_PERF_DOMAIN_SKIP_INEFFICIENCIES: Skip inefficient states when estimating
84 * energy consumption.
85 *
86 * EM_PERF_DOMAIN_ARTIFICIAL: The power values are artificial and might be
87 * created by platform missing real power information
88 */
89 #define EM_PERF_DOMAIN_MICROWATTS BIT(0)
90 #define EM_PERF_DOMAIN_SKIP_INEFFICIENCIES BIT(1)
91 #define EM_PERF_DOMAIN_ARTIFICIAL BIT(2)
92
93 #define em_span_cpus(em) (to_cpumask((em)->cpus))
94 #define em_is_artificial(em) ((em)->flags & EM_PERF_DOMAIN_ARTIFICIAL)
95
96 #ifdef CONFIG_ENERGY_MODEL
97 /*
98 * The max power value in micro-Watts. The limit of 64 Watts is set as
99 * a safety net to not overflow multiplications on 32bit platforms. The
100 * 32bit value limit for total Perf Domain power implies a limit of
101 * maximum CPUs in such domain to 64.
102 */
103 #define EM_MAX_POWER (64000000) /* 64 Watts */
104
105 /*
106 * To avoid possible energy estimation overflow on 32bit machines add
107 * limits to number of CPUs in the Perf. Domain.
108 * We are safe on 64bit machine, thus some big number.
109 */
110 #ifdef CONFIG_64BIT
111 #define EM_MAX_NUM_CPUS 4096
112 #else
113 #define EM_MAX_NUM_CPUS 16
114 #endif
115
116 struct em_data_callback {
117 /**
118 * active_power() - Provide power at the next performance state of
119 * a device
120 * @dev : Device for which we do this operation (can be a CPU)
121 * @power : Active power at the performance state
122 * (modified)
123 * @freq : Frequency at the performance state in kHz
124 * (modified)
125 *
126 * active_power() must find the lowest performance state of 'dev' above
127 * 'freq' and update 'power' and 'freq' to the matching active power
128 * and frequency.
129 *
130 * In case of CPUs, the power is the one of a single CPU in the domain,
131 * expressed in micro-Watts or an abstract scale. It is expected to
132 * fit in the [0, EM_MAX_POWER] range.
133 *
134 * Return 0 on success.
135 */
136 int (*active_power)(struct device *dev, unsigned long *power,
137 unsigned long *freq);
138
139 /**
140 * get_cost() - Provide the cost at the given performance state of
141 * a device
142 * @dev : Device for which we do this operation (can be a CPU)
143 * @freq : Frequency at the performance state in kHz
144 * @cost : The cost value for the performance state
145 * (modified)
146 *
147 * In case of CPUs, the cost is the one of a single CPU in the domain.
148 * It is expected to fit in the [0, EM_MAX_POWER] range due to internal
149 * usage in EAS calculation.
150 *
151 * Return 0 on success, or appropriate error value in case of failure.
152 */
153 int (*get_cost)(struct device *dev, unsigned long freq,
154 unsigned long *cost);
155 };
156 #define EM_SET_ACTIVE_POWER_CB(em_cb, cb) ((em_cb).active_power = cb)
157 #define EM_ADV_DATA_CB(_active_power_cb, _cost_cb) \
158 { .active_power = _active_power_cb, \
159 .get_cost = _cost_cb }
160 #define EM_DATA_CB(_active_power_cb) \
161 EM_ADV_DATA_CB(_active_power_cb, NULL)
162
163 struct em_perf_domain *em_cpu_get(int cpu);
164 struct em_perf_domain *em_pd_get(struct device *dev);
165 int em_dev_update_perf_domain(struct device *dev,
166 struct em_perf_table __rcu *new_table);
167 int em_dev_register_perf_domain(struct device *dev, unsigned int nr_states,
168 struct em_data_callback *cb, cpumask_t *span,
169 bool microwatts);
170 void em_dev_unregister_perf_domain(struct device *dev);
171 struct em_perf_table __rcu *em_table_alloc(struct em_perf_domain *pd);
172 void em_table_free(struct em_perf_table __rcu *table);
173 int em_dev_compute_costs(struct device *dev, struct em_perf_state *table,
174 int nr_states);
175 int em_dev_update_chip_binning(struct device *dev);
176
177 /**
178 * em_pd_get_efficient_state() - Get an efficient performance state from the EM
179 * @table: List of performance states, in ascending order
180 * @nr_perf_states: Number of performance states
181 * @max_util: Max utilization to map with the EM
182 * @pd_flags: Performance Domain flags
183 *
184 * It is called from the scheduler code quite frequently and as a consequence
185 * doesn't implement any check.
186 *
187 * Return: An efficient performance state id, high enough to meet @max_util
188 * requirement.
189 */
190 static inline int
191 em_pd_get_efficient_state(struct em_perf_state *table, int nr_perf_states,
192 unsigned long max_util, unsigned long pd_flags)
193 {
194 struct em_perf_state *ps;
195 int i;
196
197 for (i = 0; i < nr_perf_states; i++) {
198 ps = &table[i];
199 if (ps->performance >= max_util) {
200 if (pd_flags & EM_PERF_DOMAIN_SKIP_INEFFICIENCIES &&
201 ps->flags & EM_PERF_STATE_INEFFICIENT)
202 continue;
203 return i;
204 }
205 }
206
207 return nr_perf_states - 1;
208 }
209
210 /**
211 * em_cpu_energy() - Estimates the energy consumed by the CPUs of a
212 * performance domain
213 * @pd : performance domain for which energy has to be estimated
214 * @max_util : highest utilization among CPUs of the domain
215 * @sum_util : sum of the utilization of all CPUs in the domain
216 * @allowed_cpu_cap : maximum allowed CPU capacity for the @pd, which
217 * might reflect reduced frequency (due to thermal)
218 *
219 * This function must be used only for CPU devices. There is no validation,
220 * i.e. if the EM is a CPU type and has cpumask allocated. It is called from
221 * the scheduler code quite frequently and that is why there is not checks.
222 *
223 * Return: the sum of the energy consumed by the CPUs of the domain assuming
224 * a capacity state satisfying the max utilization of the domain.
225 */
226 static inline unsigned long em_cpu_energy(struct em_perf_domain *pd,
227 unsigned long max_util, unsigned long sum_util,
228 unsigned long allowed_cpu_cap)
229 {
230 struct em_perf_table *em_table;
231 struct em_perf_state *ps;
232 int i;
233
234 #ifdef CONFIG_SCHED_DEBUG
235 WARN_ONCE(!rcu_read_lock_held(), "EM: rcu read lock needed\n");
236 #endif
237
238 if (!sum_util)
239 return 0;
240
241 /*
242 * In order to predict the performance state, map the utilization of
243 * the most utilized CPU of the performance domain to a requested
244 * performance, like schedutil. Take also into account that the real
245 * performance might be set lower (due to thermal capping). Thus, clamp
246 * max utilization to the allowed CPU capacity before calculating
247 * effective performance.
248 */
249 max_util = min(max_util, allowed_cpu_cap);
250
251 /*
252 * Find the lowest performance state of the Energy Model above the
253 * requested performance.
254 */
255 em_table = rcu_dereference(pd->em_table);
256 i = em_pd_get_efficient_state(em_table->state, pd->nr_perf_states,
257 max_util, pd->flags);
258 ps = &em_table->state[i];
259
260 /*
261 * The performance (capacity) of a CPU in the domain at the performance
262 * state (ps) can be computed as:
263 *
264 * ps->freq * scale_cpu
265 * ps->performance = -------------------- (1)
266 * cpu_max_freq
267 *
268 * So, ignoring the costs of idle states (which are not available in
269 * the EM), the energy consumed by this CPU at that performance state
270 * is estimated as:
271 *
272 * ps->power * cpu_util
273 * cpu_nrg = -------------------- (2)
274 * ps->performance
275 *
276 * since 'cpu_util / ps->performance' represents its percentage of busy
277 * time.
278 *
279 * NOTE: Although the result of this computation actually is in
280 * units of power, it can be manipulated as an energy value
281 * over a scheduling period, since it is assumed to be
282 * constant during that interval.
283 *
284 * By injecting (1) in (2), 'cpu_nrg' can be re-expressed as a product
285 * of two terms:
286 *
287 * ps->power * cpu_max_freq
288 * cpu_nrg = ------------------------ * cpu_util (3)
289 * ps->freq * scale_cpu
290 *
291 * The first term is static, and is stored in the em_perf_state struct
292 * as 'ps->cost'.
293 *
294 * Since all CPUs of the domain have the same micro-architecture, they
295 * share the same 'ps->cost', and the same CPU capacity. Hence, the
296 * total energy of the domain (which is the simple sum of the energy of
297 * all of its CPUs) can be factorized as:
298 *
299 * pd_nrg = ps->cost * \Sum cpu_util (4)
300 */
301 return ps->cost * sum_util;
302 }
303
304 /**
305 * em_pd_nr_perf_states() - Get the number of performance states of a perf.
306 * domain
307 * @pd : performance domain for which this must be done
308 *
309 * Return: the number of performance states in the performance domain table
310 */
311 static inline int em_pd_nr_perf_states(struct em_perf_domain *pd)
312 {
313 return pd->nr_perf_states;
314 }
315
316 /**
317 * em_perf_state_from_pd() - Get the performance states table of perf.
318 * domain
319 * @pd : performance domain for which this must be done
320 *
321 * To use this function the rcu_read_lock() should be hold. After the usage
322 * of the performance states table is finished, the rcu_read_unlock() should
323 * be called.
324 *
325 * Return: the pointer to performance states table of the performance domain
326 */
327 static inline
328 struct em_perf_state *em_perf_state_from_pd(struct em_perf_domain *pd)
329 {
330 return rcu_dereference(pd->em_table)->state;
331 }
332
333 #else
334 struct em_data_callback {};
335 #define EM_ADV_DATA_CB(_active_power_cb, _cost_cb) { }
336 #define EM_DATA_CB(_active_power_cb) { }
337 #define EM_SET_ACTIVE_POWER_CB(em_cb, cb) do { } while (0)
338
339 static inline
340 int em_dev_register_perf_domain(struct device *dev, unsigned int nr_states,
341 struct em_data_callback *cb, cpumask_t *span,
342 bool microwatts)
343 {
344 return -EINVAL;
345 }
346 static inline void em_dev_unregister_perf_domain(struct device *dev)
347 {
348 }
349 static inline struct em_perf_domain *em_cpu_get(int cpu)
350 {
351 return NULL;
352 }
353 static inline struct em_perf_domain *em_pd_get(struct device *dev)
354 {
355 return NULL;
356 }
357 static inline unsigned long em_cpu_energy(struct em_perf_domain *pd,
358 unsigned long max_util, unsigned long sum_util,
359 unsigned long allowed_cpu_cap)
360 {
361 return 0;
362 }
363 static inline int em_pd_nr_perf_states(struct em_perf_domain *pd)
364 {
365 return 0;
366 }
367 static inline
368 struct em_perf_table __rcu *em_table_alloc(struct em_perf_domain *pd)
369 {
370 return NULL;
371 }
372 static inline void em_table_free(struct em_perf_table __rcu *table) {}
373 static inline
374 int em_dev_update_perf_domain(struct device *dev,
375 struct em_perf_table __rcu *new_table)
376 {
377 return -EINVAL;
378 }
379 static inline
380 struct em_perf_state *em_perf_state_from_pd(struct em_perf_domain *pd)
381 {
382 return NULL;
383 }
384 static inline
385 int em_dev_compute_costs(struct device *dev, struct em_perf_state *table,
386 int nr_states)
387 {
388 return -EINVAL;
389 }
390 static inline int em_dev_update_chip_binning(struct device *dev)
391 {
392 return -EINVAL;
393 }
394 #endif
395
396 #endif
397
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.