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>
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, in milli-watts (by 1 CPU or
17 by a registered device). It can be a total power: static and
19 * @cost: The cost coefficient associated with this level, used during
20 * energy calculation. Equal to: power * max_frequency / frequency
22 struct em_perf_state {
23 unsigned long frequency;
29 * em_perf_domain - Performance domain
30 * @table: List of performance states, in ascending order
31 * @nr_perf_states: Number of performance states
32 * @cpus: Cpumask covering the CPUs of the domain. It's here
33 * for performance reasons to avoid potential cache
34 * misses during energy calculations in the scheduler
35 * and simplifies allocating/freeing that memory region.
37 * In case of CPU device, a "performance domain" represents a group of CPUs
38 * whose performance is scaled together. All CPUs of a performance domain
39 * must have the same micro-architecture. Performance domains often have
40 * a 1-to-1 mapping with CPUFreq policies. In case of other devices the @cpus
43 struct em_perf_domain {
44 struct em_perf_state *table;
49 #define em_span_cpus(em) (to_cpumask((em)->cpus))
51 #ifdef CONFIG_ENERGY_MODEL
52 #define EM_MAX_POWER 0xFFFF
55 * Increase resolution of energy estimation calculations for 64-bit
56 * architectures. The extra resolution improves decision made by EAS for the
57 * task placement when two Performance Domains might provide similar energy
58 * estimation values (w/o better resolution the values could be equal).
60 * We increase resolution only if we have enough bits to allow this increased
61 * resolution (i.e. 64-bit). The costs for increasing resolution when 32-bit
62 * are pretty high and the returns do not justify the increased costs.
65 #define em_scale_power(p) ((p) * 1000)
67 #define em_scale_power(p) (p)
70 struct em_data_callback {
72 * active_power() - Provide power at the next performance state of
74 * @power : Active power at the performance state in mW
76 * @freq : Frequency at the performance state in kHz
78 * @dev : Device for which we do this operation (can be a CPU)
80 * active_power() must find the lowest performance state of 'dev' above
81 * 'freq' and update 'power' and 'freq' to the matching active power
84 * In case of CPUs, the power is the one of a single CPU in the domain,
85 * expressed in milli-watts. It is expected to fit in the
86 * [0, EM_MAX_POWER] range.
88 * Return 0 on success.
90 int (*active_power)(unsigned long *power, unsigned long *freq,
93 #define EM_DATA_CB(_active_power_cb) { .active_power = &_active_power_cb }
95 struct em_perf_domain *em_cpu_get(int cpu);
96 struct em_perf_domain *em_pd_get(struct device *dev);
97 int em_dev_register_perf_domain(struct device *dev, unsigned int nr_states,
98 struct em_data_callback *cb, cpumask_t *span);
99 void em_dev_unregister_perf_domain(struct device *dev);
102 * em_cpu_energy() - Estimates the energy consumed by the CPUs of a
104 * @pd : performance domain for which energy has to be estimated
105 * @max_util : highest utilization among CPUs of the domain
106 * @sum_util : sum of the utilization of all CPUs in the domain
108 * This function must be used only for CPU devices. There is no validation,
109 * i.e. if the EM is a CPU type and has cpumask allocated. It is called from
110 * the scheduler code quite frequently and that is why there is not checks.
112 * Return: the sum of the energy consumed by the CPUs of the domain assuming
113 * a capacity state satisfying the max utilization of the domain.
115 static inline unsigned long em_cpu_energy(struct em_perf_domain *pd,
116 unsigned long max_util, unsigned long sum_util)
118 unsigned long freq, scale_cpu;
119 struct em_perf_state *ps;
123 * In order to predict the performance state, map the utilization of
124 * the most utilized CPU of the performance domain to a requested
125 * frequency, like schedutil.
127 cpu = cpumask_first(to_cpumask(pd->cpus));
128 scale_cpu = arch_scale_cpu_capacity(cpu);
129 ps = &pd->table[pd->nr_perf_states - 1];
130 freq = map_util_freq(max_util, ps->frequency, scale_cpu);
133 * Find the lowest performance state of the Energy Model above the
134 * requested frequency.
136 for (i = 0; i < pd->nr_perf_states; i++) {
138 if (ps->frequency >= freq)
143 * The capacity of a CPU in the domain at the performance state (ps)
144 * can be computed as:
146 * ps->freq * scale_cpu
147 * ps->cap = -------------------- (1)
150 * So, ignoring the costs of idle states (which are not available in
151 * the EM), the energy consumed by this CPU at that performance state
154 * ps->power * cpu_util
155 * cpu_nrg = -------------------- (2)
158 * since 'cpu_util / ps->cap' represents its percentage of busy time.
160 * NOTE: Although the result of this computation actually is in
161 * units of power, it can be manipulated as an energy value
162 * over a scheduling period, since it is assumed to be
163 * constant during that interval.
165 * By injecting (1) in (2), 'cpu_nrg' can be re-expressed as a product
168 * ps->power * cpu_max_freq cpu_util
169 * cpu_nrg = ------------------------ * --------- (3)
172 * The first term is static, and is stored in the em_perf_state struct
175 * Since all CPUs of the domain have the same micro-architecture, they
176 * share the same 'ps->cost', and the same CPU capacity. Hence, the
177 * total energy of the domain (which is the simple sum of the energy of
178 * all of its CPUs) can be factorized as:
180 * ps->cost * \Sum cpu_util
181 * pd_nrg = ------------------------ (4)
184 return ps->cost * sum_util / scale_cpu;
188 * em_pd_nr_perf_states() - Get the number of performance states of a perf.
190 * @pd : performance domain for which this must be done
192 * Return: the number of performance states in the performance domain table
194 static inline int em_pd_nr_perf_states(struct em_perf_domain *pd)
196 return pd->nr_perf_states;
200 struct em_data_callback {};
201 #define EM_DATA_CB(_active_power_cb) { }
204 int em_dev_register_perf_domain(struct device *dev, unsigned int nr_states,
205 struct em_data_callback *cb, cpumask_t *span)
209 static inline void em_dev_unregister_perf_domain(struct device *dev)
212 static inline struct em_perf_domain *em_cpu_get(int cpu)
216 static inline struct em_perf_domain *em_pd_get(struct device *dev)
220 static inline unsigned long em_cpu_energy(struct em_perf_domain *pd,
221 unsigned long max_util, unsigned long sum_util)
225 static inline int em_pd_nr_perf_states(struct em_perf_domain *pd)