Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
tjh.dev
kernel
os
linux
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 * @allowed_cpu_cap : maximum allowed CPU capacity for the @pd, which
95 might reflect reduced frequency (due to thermal)
96 *
97 * This function must be used only for CPU devices. There is no validation,
98 * i.e. if the EM is a CPU type and has cpumask allocated. It is called from
99 * the scheduler code quite frequently and that is why there is not checks.
100 *
101 * Return: the sum of the energy consumed by the CPUs of the domain assuming
102 * a capacity state satisfying the max utilization of the domain.
103 */
104static inline unsigned long em_cpu_energy(struct em_perf_domain *pd,
105 unsigned long max_util, unsigned long sum_util,
106 unsigned long allowed_cpu_cap)
107{
108 unsigned long freq, scale_cpu;
109 struct em_perf_state *ps;
110 int i, cpu;
111
112 if (!sum_util)
113 return 0;
114
115 /*
116 * In order to predict the performance state, map the utilization of
117 * the most utilized CPU of the performance domain to a requested
118 * frequency, like schedutil. Take also into account that the real
119 * frequency might be set lower (due to thermal capping). Thus, clamp
120 * max utilization to the allowed CPU capacity before calculating
121 * effective frequency.
122 */
123 cpu = cpumask_first(to_cpumask(pd->cpus));
124 scale_cpu = arch_scale_cpu_capacity(cpu);
125 ps = &pd->table[pd->nr_perf_states - 1];
126
127 max_util = map_util_perf(max_util);
128 max_util = min(max_util, allowed_cpu_cap);
129 freq = map_util_freq(max_util, ps->frequency, scale_cpu);
130
131 /*
132 * Find the lowest performance state of the Energy Model above the
133 * requested frequency.
134 */
135 for (i = 0; i < pd->nr_perf_states; i++) {
136 ps = &pd->table[i];
137 if (ps->frequency >= freq)
138 break;
139 }
140
141 /*
142 * The capacity of a CPU in the domain at the performance state (ps)
143 * can be computed as:
144 *
145 * ps->freq * scale_cpu
146 * ps->cap = -------------------- (1)
147 * cpu_max_freq
148 *
149 * So, ignoring the costs of idle states (which are not available in
150 * the EM), the energy consumed by this CPU at that performance state
151 * is estimated as:
152 *
153 * ps->power * cpu_util
154 * cpu_nrg = -------------------- (2)
155 * ps->cap
156 *
157 * since 'cpu_util / ps->cap' represents its percentage of busy time.
158 *
159 * NOTE: Although the result of this computation actually is in
160 * units of power, it can be manipulated as an energy value
161 * over a scheduling period, since it is assumed to be
162 * constant during that interval.
163 *
164 * By injecting (1) in (2), 'cpu_nrg' can be re-expressed as a product
165 * of two terms:
166 *
167 * ps->power * cpu_max_freq cpu_util
168 * cpu_nrg = ------------------------ * --------- (3)
169 * ps->freq scale_cpu
170 *
171 * The first term is static, and is stored in the em_perf_state struct
172 * as 'ps->cost'.
173 *
174 * Since all CPUs of the domain have the same micro-architecture, they
175 * share the same 'ps->cost', and the same CPU capacity. Hence, the
176 * total energy of the domain (which is the simple sum of the energy of
177 * all of its CPUs) can be factorized as:
178 *
179 * ps->cost * \Sum cpu_util
180 * pd_nrg = ------------------------ (4)
181 * scale_cpu
182 */
183 return ps->cost * sum_util / scale_cpu;
184}
185
186/**
187 * em_pd_nr_perf_states() - Get the number of performance states of a perf.
188 * domain
189 * @pd : performance domain for which this must be done
190 *
191 * Return: the number of performance states in the performance domain table
192 */
193static inline int em_pd_nr_perf_states(struct em_perf_domain *pd)
194{
195 return pd->nr_perf_states;
196}
197
198#else
199struct em_data_callback {};
200#define EM_DATA_CB(_active_power_cb) { }
201
202static inline
203int em_dev_register_perf_domain(struct device *dev, unsigned int nr_states,
204 struct em_data_callback *cb, cpumask_t *span,
205 bool milliwatts)
206{
207 return -EINVAL;
208}
209static inline void em_dev_unregister_perf_domain(struct device *dev)
210{
211}
212static inline struct em_perf_domain *em_cpu_get(int cpu)
213{
214 return NULL;
215}
216static inline struct em_perf_domain *em_pd_get(struct device *dev)
217{
218 return NULL;
219}
220static inline unsigned long em_cpu_energy(struct em_perf_domain *pd,
221 unsigned long max_util, unsigned long sum_util,
222 unsigned long allowed_cpu_cap)
223{
224 return 0;
225}
226static inline int em_pd_nr_perf_states(struct em_perf_domain *pd)
227{
228 return 0;
229}
230#endif
231
232#endif