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, in milli-watts (by 1 CPU or
17        by a registered device). It can be a total power: static and
18        dynamic.
19 * @cost:    The cost coefficient associated with this level, used during
20 *        energy calculation. Equal to: power * max_frequency / frequency
21 */
22struct em_perf_state {
23    unsigned long frequency;
24    unsigned long power;
25    unsigned long cost;
26};
27
28/**
29 * em_perf_domain - Performance domain
30 * @table:        List of performance states, in ascending order
31 * @nr_perf_states:    Number of performance states
32 * @milliwatts:        Flag indicating the power values are in milli-Watts
33 *            or some other scale.
34 * @cpus:        Cpumask covering the CPUs of the domain. It's here
35 *            for performance reasons to avoid potential cache
36 *            misses during energy calculations in the scheduler
37 *            and simplifies allocating/freeing that memory region.
38 *
39 * In case of CPU device, a "performance domain" represents a group of CPUs
40 * whose performance is scaled together. All CPUs of a performance domain
41 * must have the same micro-architecture. Performance domains often have
42 * a 1-to-1 mapping with CPUFreq policies. In case of other devices the @cpus
43 * field is unused.
44 */
45struct em_perf_domain {
46    struct em_perf_state *table;
47    int nr_perf_states;
48    int milliwatts;
49    unsigned long cpus[];
50};
51
52#define em_span_cpus(em) (to_cpumask((em)->cpus))
53
54#ifdef CONFIG_ENERGY_MODEL
55#define EM_MAX_POWER 0xFFFF
56
57/*
58 * Increase resolution of energy estimation calculations for 64-bit
59 * architectures. The extra resolution improves decision made by EAS for the
60 * task placement when two Performance Domains might provide similar energy
61 * estimation values (w/o better resolution the values could be equal).
62 *
63 * We increase resolution only if we have enough bits to allow this increased
64 * resolution (i.e. 64-bit). The costs for increasing resolution when 32-bit
65 * are pretty high and the returns do not justify the increased costs.
66 */
67#ifdef CONFIG_64BIT
68#define em_scale_power(p) ((p)*1000)
69#else
70#define em_scale_power(p) (p)
71#endif
72
73struct em_data_callback {
74    /**
75     * active_power() - Provide power at the next performance state of
76     *        a device
77     * @power    : Active power at the performance state in mW
78     *        (modified)
79     * @freq    : Frequency at the performance state in kHz
80     *        (modified)
81     * @dev        : Device for which we do this operation (can be a CPU)
82     *
83     * active_power() must find the lowest performance state of 'dev' above
84     * 'freq' and update 'power' and 'freq' to the matching active power
85     * and frequency.
86     *
87     * In case of CPUs, the power is the one of a single CPU in the domain,
88     * expressed in milli-watts. It is expected to fit in the
89     * [0, EM_MAX_POWER] range.
90     *
91     * Return 0 on success.
92     */
93    int (*active_power)(unsigned long *power, unsigned long *freq, struct device *dev);
94};
95#define EM_DATA_CB(_active_power_cb)                                                                                   \
96    {                                                                                                                  \
97        .active_power = &_active_power_cb                                                                              \
98    }
99
100struct em_perf_domain *em_cpu_get(int cpu);
101struct em_perf_domain *em_pd_get(struct device *dev);
102int em_dev_register_perf_domain(struct device *dev, unsigned int nr_states, struct em_data_callback *cb,
103                                cpumask_t *span, bool milliwatts);
104void em_dev_unregister_perf_domain(struct device *dev);
105
106/**
107 * em_cpu_energy() - Estimates the energy consumed by the CPUs of a
108        performance domain
109 * @pd        : performance domain for which energy has to be estimated
110 * @max_util    : highest utilization among CPUs of the domain
111 * @sum_util    : sum of the utilization of all CPUs in the domain
112 *
113 * This function must be used only for CPU devices. There is no validation,
114 * i.e. if the EM is a CPU type and has cpumask allocated. It is called from
115 * the scheduler code quite frequently and that is why there is not checks.
116 *
117 * Return: the sum of the energy consumed by the CPUs of the domain assuming
118 * a capacity state satisfying the max utilization of the domain.
119 */
120static inline unsigned long em_cpu_energy(struct em_perf_domain *pd, unsigned long max_util, unsigned long sum_util)
121{
122    unsigned long freq, scale_cpu;
123    struct em_perf_state *ps;
124    int i, cpu;
125
126    if (!sum_util) {
127        return 0;
128    }
129
130    /*
131     * In order to predict the performance state, map the utilization of
132     * the most utilized CPU of the performance domain to a requested
133     * frequency, like schedutil.
134     */
135    cpu = cpumask_first(to_cpumask(pd->cpus));
136    scale_cpu = arch_scale_cpu_capacity(cpu);
137    ps = &pd->table[pd->nr_perf_states - 1];
138    freq = map_util_freq(max_util, ps->frequency, scale_cpu);
139
140    /*
141     * Find the lowest performance state of the Energy Model above the
142     * requested frequency.
143     */
144    for (i = 0; i < pd->nr_perf_states; i++) {
145        ps = &pd->table[i];
146        if (ps->frequency >= freq) {
147            break;
148        }
149    }
150
151    /*
152     * The capacity of a CPU in the domain at the performance state (ps)
153     * can be computed as:
154     *
155     *             ps->freq * scale_cpu
156     *   ps->cap = --------------------                          (1)
157     *                 cpu_max_freq
158     *
159     * So, ignoring the costs of idle states (which are not available in
160     * the EM), the energy consumed by this CPU at that performance state
161     * is estimated as:
162     *
163     *             ps->power * cpu_util
164     *   cpu_nrg = --------------------                          (2)
165     *                   ps->cap
166     *
167     * since 'cpu_util / ps->cap' represents its percentage of busy time.
168     *
169     *   NOTE: Although the result of this computation actually is in
170     *         units of power, it can be manipulated as an energy value
171     *         over a scheduling period, since it is assumed to be
172     *         constant during that interval.
173     *
174     * By injecting (1) in (2), 'cpu_nrg' can be re-expressed as a product
175     * of two terms:
176     *
177     *             ps->power * cpu_max_freq   cpu_util
178     *   cpu_nrg = ------------------------ * ---------          (3)
179     *                    ps->freq            scale_cpu
180     *
181     * The first term is static, and is stored in the em_perf_state struct
182     * as 'ps->cost'.
183     *
184     * Since all CPUs of the domain have the same micro-architecture, they
185     * share the same 'ps->cost', and the same CPU capacity. Hence, the
186     * total energy of the domain (which is the simple sum of the energy of
187     * all of its CPUs) can be factorized as:
188     *
189     *            ps->cost * \Sum cpu_util
190     *   pd_nrg = ------------------------                       (4)
191     *                  scale_cpu
192     */
193    return ps->cost * sum_util / scale_cpu;
194}
195
196/**
197 * em_pd_nr_perf_states() - Get the number of performance states of a perf.
198 *                domain
199 * @pd        : performance domain for which this must be done
200 *
201 * Return: the number of performance states in the performance domain table
202 */
203static inline int em_pd_nr_perf_states(struct em_perf_domain *pd)
204{
205    return pd->nr_perf_states;
206}
207
208#else
209struct em_data_callback {
210};
211#define EM_DATA_CB(_active_power_cb)                                                                                   \
212    {                                                                                                                  \
213    }
214
215static inline int em_dev_register_perf_domain(struct device *dev, unsigned int nr_states, struct em_data_callback *cb,
216                                              cpumask_t *span, bool milliwatts)
217{
218    return -EINVAL;
219}
220static inline void em_dev_unregister_perf_domain(struct device *dev)
221{
222}
223static inline struct em_perf_domain *em_cpu_get(int cpu)
224{
225    return NULL;
226}
227static inline struct em_perf_domain *em_pd_get(struct device *dev)
228{
229    return NULL;
230}
231static inline unsigned long em_cpu_energy(struct em_perf_domain *pd, unsigned long max_util, unsigned long sum_util)
232{
233    return 0;
234}
235static inline int em_pd_nr_perf_states(struct em_perf_domain *pd)
236{
237    return 0;
238}
239#endif
240
241#endif
242