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>
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.
24 struct em_perf_state {
25 unsigned long performance;
26 unsigned long frequency;
33 * em_perf_state flags:
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.
40 #define EM_PERF_STATE_INEFFICIENT BIT(0)
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
48 struct em_perf_table {
51 struct em_perf_state state[];
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.
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
70 struct em_perf_domain {
71 struct em_perf_table __rcu *em_table;
78 * em_perf_domain flags:
80 * EM_PERF_DOMAIN_MICROWATTS: The power values are in micro-Watts or some
83 * EM_PERF_DOMAIN_SKIP_INEFFICIENCIES: Skip inefficient states when estimating
86 * EM_PERF_DOMAIN_ARTIFICIAL: The power values are artificial and might be
87 * created by platform missing real power information
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)
93 #define em_span_cpus(em) (to_cpumask((em)->cpus))
94 #define em_is_artificial(em) ((em)->flags & EM_PERF_DOMAIN_ARTIFICIAL)
96 #ifdef CONFIG_ENERGY_MODEL
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.
103 #define EM_MAX_POWER (64000000) /* 64 Watts */
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.
111 #define EM_MAX_NUM_CPUS 4096
113 #define EM_MAX_NUM_CPUS 16
116 struct em_data_callback {
118 * active_power() - Provide power at the next performance state of
120 * @dev : Device for which we do this operation (can be a CPU)
121 * @power : Active power at the performance state
123 * @freq : Frequency at the performance state in kHz
126 * active_power() must find the lowest performance state of 'dev' above
127 * 'freq' and update 'power' and 'freq' to the matching active power
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.
134 * Return 0 on success.
136 int (*active_power)(struct device *dev, unsigned long *power,
137 unsigned long *freq);
140 * get_cost() - Provide the cost at the given performance state of
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
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.
151 * Return 0 on success, or appropriate error value in case of failure.
153 int (*get_cost)(struct device *dev, unsigned long freq,
154 unsigned long *cost);
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)
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,
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,
177 * em_pd_get_efficient_state() - Get an efficient performance state from the EM
178 * @table: List of performance states, in ascending order
179 * @nr_perf_states: Number of performance states
180 * @max_util: Max utilization to map with the EM
181 * @pd_flags: Performance Domain flags
183 * It is called from the scheduler code quite frequently and as a consequence
184 * doesn't implement any check.
186 * Return: An efficient performance state id, high enough to meet @max_util
190 em_pd_get_efficient_state(struct em_perf_state *table, int nr_perf_states,
191 unsigned long max_util, unsigned long pd_flags)
193 struct em_perf_state *ps;
196 for (i = 0; i < nr_perf_states; i++) {
198 if (ps->performance >= max_util) {
199 if (pd_flags & EM_PERF_DOMAIN_SKIP_INEFFICIENCIES &&
200 ps->flags & EM_PERF_STATE_INEFFICIENT)
206 return nr_perf_states - 1;
210 * em_cpu_energy() - Estimates the energy consumed by the CPUs of a
212 * @pd : performance domain for which energy has to be estimated
213 * @max_util : highest utilization among CPUs of the domain
214 * @sum_util : sum of the utilization of all CPUs in the domain
215 * @allowed_cpu_cap : maximum allowed CPU capacity for the @pd, which
216 * might reflect reduced frequency (due to thermal)
218 * This function must be used only for CPU devices. There is no validation,
219 * i.e. if the EM is a CPU type and has cpumask allocated. It is called from
220 * the scheduler code quite frequently and that is why there is not checks.
222 * Return: the sum of the energy consumed by the CPUs of the domain assuming
223 * a capacity state satisfying the max utilization of the domain.
225 static inline unsigned long em_cpu_energy(struct em_perf_domain *pd,
226 unsigned long max_util, unsigned long sum_util,
227 unsigned long allowed_cpu_cap)
229 struct em_perf_table *em_table;
230 struct em_perf_state *ps;
233 #ifdef CONFIG_SCHED_DEBUG
234 WARN_ONCE(!rcu_read_lock_held(), "EM: rcu read lock needed\n");
241 * In order to predict the performance state, map the utilization of
242 * the most utilized CPU of the performance domain to a requested
243 * performance, like schedutil. Take also into account that the real
244 * performance might be set lower (due to thermal capping). Thus, clamp
245 * max utilization to the allowed CPU capacity before calculating
246 * effective performance.
248 max_util = min(max_util, allowed_cpu_cap);
251 * Find the lowest performance state of the Energy Model above the
252 * requested performance.
254 em_table = rcu_dereference(pd->em_table);
255 i = em_pd_get_efficient_state(em_table->state, pd->nr_perf_states,
256 max_util, pd->flags);
257 ps = &em_table->state[i];
260 * The performance (capacity) of a CPU in the domain at the performance
261 * state (ps) can be computed as:
263 * ps->freq * scale_cpu
264 * ps->performance = -------------------- (1)
267 * So, ignoring the costs of idle states (which are not available in
268 * the EM), the energy consumed by this CPU at that performance state
271 * ps->power * cpu_util
272 * cpu_nrg = -------------------- (2)
275 * since 'cpu_util / ps->performance' represents its percentage of busy
278 * NOTE: Although the result of this computation actually is in
279 * units of power, it can be manipulated as an energy value
280 * over a scheduling period, since it is assumed to be
281 * constant during that interval.
283 * By injecting (1) in (2), 'cpu_nrg' can be re-expressed as a product
286 * ps->power * cpu_max_freq
287 * cpu_nrg = ------------------------ * cpu_util (3)
288 * ps->freq * scale_cpu
290 * The first term is static, and is stored in the em_perf_state struct
293 * Since all CPUs of the domain have the same micro-architecture, they
294 * share the same 'ps->cost', and the same CPU capacity. Hence, the
295 * total energy of the domain (which is the simple sum of the energy of
296 * all of its CPUs) can be factorized as:
298 * pd_nrg = ps->cost * \Sum cpu_util (4)
300 return ps->cost * sum_util;
304 * em_pd_nr_perf_states() - Get the number of performance states of a perf.
306 * @pd : performance domain for which this must be done
308 * Return: the number of performance states in the performance domain table
310 static inline int em_pd_nr_perf_states(struct em_perf_domain *pd)
312 return pd->nr_perf_states;
316 * em_perf_state_from_pd() - Get the performance states table of perf.
318 * @pd : performance domain for which this must be done
320 * To use this function the rcu_read_lock() should be hold. After the usage
321 * of the performance states table is finished, the rcu_read_unlock() should
324 * Return: the pointer to performance states table of the performance domain
327 struct em_perf_state *em_perf_state_from_pd(struct em_perf_domain *pd)
329 return rcu_dereference(pd->em_table)->state;
333 struct em_data_callback {};
334 #define EM_ADV_DATA_CB(_active_power_cb, _cost_cb) { }
335 #define EM_DATA_CB(_active_power_cb) { }
336 #define EM_SET_ACTIVE_POWER_CB(em_cb, cb) do { } while (0)
339 int em_dev_register_perf_domain(struct device *dev, unsigned int nr_states,
340 struct em_data_callback *cb, cpumask_t *span,
345 static inline void em_dev_unregister_perf_domain(struct device *dev)
348 static inline struct em_perf_domain *em_cpu_get(int cpu)
352 static inline struct em_perf_domain *em_pd_get(struct device *dev)
356 static inline unsigned long em_cpu_energy(struct em_perf_domain *pd,
357 unsigned long max_util, unsigned long sum_util,
358 unsigned long allowed_cpu_cap)
362 static inline int em_pd_nr_perf_states(struct em_perf_domain *pd)
367 struct em_perf_table __rcu *em_table_alloc(struct em_perf_domain *pd)
371 static inline void em_table_free(struct em_perf_table __rcu *table) {}
373 int em_dev_update_perf_domain(struct device *dev,
374 struct em_perf_table __rcu *new_table)
379 struct em_perf_state *em_perf_state_from_pd(struct em_perf_domain *pd)
384 int em_dev_compute_costs(struct device *dev, struct em_perf_state *table,