1 // SPDX-License-Identifier: GPL-2.0
3 * Energy Model of devices
5 * Copyright (c) 2018-2021, Arm ltd.
6 * Written by: Quentin Perret, Arm ltd.
7 * Improvements provided by: Lukasz Luba, Arm ltd.
10 #define pr_fmt(fmt) "energy_model: " fmt
12 #include <linux/cpu.h>
13 #include <linux/cpufreq.h>
14 #include <linux/cpumask.h>
15 #include <linux/debugfs.h>
16 #include <linux/energy_model.h>
17 #include <linux/sched/topology.h>
18 #include <linux/slab.h>
21 * Mutex serializing the registrations of performance domains and letting
22 * callbacks defined by drivers sleep.
24 static DEFINE_MUTEX(em_pd_mutex);
26 static bool _is_cpu_device(struct device *dev)
28 return (dev->bus == &cpu_subsys);
31 #ifdef CONFIG_DEBUG_FS
32 static struct dentry *rootdir;
34 static void em_debug_create_ps(struct em_perf_state *ps, struct dentry *pd)
39 snprintf(name, sizeof(name), "ps:%lu", ps->frequency);
41 /* Create per-ps directory */
42 d = debugfs_create_dir(name, pd);
43 debugfs_create_ulong("frequency", 0444, d, &ps->frequency);
44 debugfs_create_ulong("power", 0444, d, &ps->power);
45 debugfs_create_ulong("cost", 0444, d, &ps->cost);
46 debugfs_create_ulong("inefficient", 0444, d, &ps->flags);
49 static int em_debug_cpus_show(struct seq_file *s, void *unused)
51 seq_printf(s, "%*pbl\n", cpumask_pr_args(to_cpumask(s->private)));
55 DEFINE_SHOW_ATTRIBUTE(em_debug_cpus);
57 static int em_debug_units_show(struct seq_file *s, void *unused)
59 struct em_perf_domain *pd = s->private;
60 char *units = (pd->flags & EM_PERF_DOMAIN_MILLIWATTS) ?
61 "milliWatts" : "bogoWatts";
63 seq_printf(s, "%s\n", units);
67 DEFINE_SHOW_ATTRIBUTE(em_debug_units);
69 static int em_debug_skip_inefficiencies_show(struct seq_file *s, void *unused)
71 struct em_perf_domain *pd = s->private;
72 int enabled = (pd->flags & EM_PERF_DOMAIN_SKIP_INEFFICIENCIES) ? 1 : 0;
74 seq_printf(s, "%d\n", enabled);
78 DEFINE_SHOW_ATTRIBUTE(em_debug_skip_inefficiencies);
80 static void em_debug_create_pd(struct device *dev)
85 /* Create the directory of the performance domain */
86 d = debugfs_create_dir(dev_name(dev), rootdir);
88 if (_is_cpu_device(dev))
89 debugfs_create_file("cpus", 0444, d, dev->em_pd->cpus,
92 debugfs_create_file("units", 0444, d, dev->em_pd, &em_debug_units_fops);
93 debugfs_create_file("skip-inefficiencies", 0444, d, dev->em_pd,
94 &em_debug_skip_inefficiencies_fops);
96 /* Create a sub-directory for each performance state */
97 for (i = 0; i < dev->em_pd->nr_perf_states; i++)
98 em_debug_create_ps(&dev->em_pd->table[i], d);
102 static void em_debug_remove_pd(struct device *dev)
104 struct dentry *debug_dir;
106 debug_dir = debugfs_lookup(dev_name(dev), rootdir);
107 debugfs_remove_recursive(debug_dir);
110 static int __init em_debug_init(void)
112 /* Create /sys/kernel/debug/energy_model directory */
113 rootdir = debugfs_create_dir("energy_model", NULL);
117 fs_initcall(em_debug_init);
118 #else /* CONFIG_DEBUG_FS */
119 static void em_debug_create_pd(struct device *dev) {}
120 static void em_debug_remove_pd(struct device *dev) {}
123 static int em_create_perf_table(struct device *dev, struct em_perf_domain *pd,
124 int nr_states, struct em_data_callback *cb)
126 unsigned long power, freq, prev_freq = 0, prev_cost = ULONG_MAX;
127 struct em_perf_state *table;
131 table = kcalloc(nr_states, sizeof(*table), GFP_KERNEL);
135 /* Build the list of performance states for this performance domain */
136 for (i = 0, freq = 0; i < nr_states; i++, freq++) {
138 * active_power() is a driver callback which ceils 'freq' to
139 * lowest performance state of 'dev' above 'freq' and updates
140 * 'power' and 'freq' accordingly.
142 ret = cb->active_power(&power, &freq, dev);
144 dev_err(dev, "EM: invalid perf. state: %d\n",
150 * We expect the driver callback to increase the frequency for
151 * higher performance states.
153 if (freq <= prev_freq) {
154 dev_err(dev, "EM: non-increasing freq: %lu\n",
160 * The power returned by active_state() is expected to be
161 * positive and to fit into 16 bits.
163 if (!power || power > EM_MAX_POWER) {
164 dev_err(dev, "EM: invalid power: %lu\n",
169 table[i].power = power;
170 table[i].frequency = prev_freq = freq;
173 /* Compute the cost of each performance state. */
174 fmax = (u64) table[nr_states - 1].frequency;
175 for (i = nr_states - 1; i >= 0; i--) {
176 unsigned long power_res = em_scale_power(table[i].power);
178 table[i].cost = div64_u64(fmax * power_res,
180 if (table[i].cost >= prev_cost) {
181 table[i].flags = EM_PERF_STATE_INEFFICIENT;
182 dev_dbg(dev, "EM: OPP:%lu is inefficient\n",
185 prev_cost = table[i].cost;
190 pd->nr_perf_states = nr_states;
199 static int em_create_pd(struct device *dev, int nr_states,
200 struct em_data_callback *cb, cpumask_t *cpus)
202 struct em_perf_domain *pd;
203 struct device *cpu_dev;
206 if (_is_cpu_device(dev)) {
207 pd = kzalloc(sizeof(*pd) + cpumask_size(), GFP_KERNEL);
211 cpumask_copy(em_span_cpus(pd), cpus);
213 pd = kzalloc(sizeof(*pd), GFP_KERNEL);
218 ret = em_create_perf_table(dev, pd, nr_states, cb);
224 if (_is_cpu_device(dev))
225 for_each_cpu(cpu, cpus) {
226 cpu_dev = get_cpu_device(cpu);
235 static void em_cpufreq_update_efficiencies(struct device *dev)
237 struct em_perf_domain *pd = dev->em_pd;
238 struct em_perf_state *table;
239 struct cpufreq_policy *policy;
243 if (!_is_cpu_device(dev) || !pd)
246 policy = cpufreq_cpu_get(cpumask_first(em_span_cpus(pd)));
248 dev_warn(dev, "EM: Access to CPUFreq policy failed");
254 for (i = 0; i < pd->nr_perf_states; i++) {
255 if (!(table[i].flags & EM_PERF_STATE_INEFFICIENT))
258 if (!cpufreq_table_set_inefficient(policy, table[i].frequency))
266 * Efficiencies have been installed in CPUFreq, inefficient frequencies
267 * will be skipped. The EM can do the same.
269 pd->flags |= EM_PERF_DOMAIN_SKIP_INEFFICIENCIES;
273 * em_pd_get() - Return the performance domain for a device
274 * @dev : Device to find the performance domain for
276 * Returns the performance domain to which @dev belongs, or NULL if it doesn't
279 struct em_perf_domain *em_pd_get(struct device *dev)
281 if (IS_ERR_OR_NULL(dev))
286 EXPORT_SYMBOL_GPL(em_pd_get);
289 * em_cpu_get() - Return the performance domain for a CPU
290 * @cpu : CPU to find the performance domain for
292 * Returns the performance domain to which @cpu belongs, or NULL if it doesn't
295 struct em_perf_domain *em_cpu_get(int cpu)
297 struct device *cpu_dev;
299 cpu_dev = get_cpu_device(cpu);
303 return em_pd_get(cpu_dev);
305 EXPORT_SYMBOL_GPL(em_cpu_get);
308 * em_dev_register_perf_domain() - Register the Energy Model (EM) for a device
309 * @dev : Device for which the EM is to register
310 * @nr_states : Number of performance states to register
311 * @cb : Callback functions providing the data of the Energy Model
312 * @cpus : Pointer to cpumask_t, which in case of a CPU device is
313 * obligatory. It can be taken from i.e. 'policy->cpus'. For other
314 * type of devices this should be set to NULL.
315 * @milliwatts : Flag indicating that the power values are in milliWatts or
316 * in some other scale. It must be set properly.
318 * Create Energy Model tables for a performance domain using the callbacks
321 * The @milliwatts is important to set with correct value. Some kernel
322 * sub-systems might rely on this flag and check if all devices in the EM are
323 * using the same scale.
325 * If multiple clients register the same performance domain, all but the first
326 * registration will be ignored.
328 * Return 0 on success
330 int em_dev_register_perf_domain(struct device *dev, unsigned int nr_states,
331 struct em_data_callback *cb, cpumask_t *cpus,
334 unsigned long cap, prev_cap = 0;
337 if (!dev || !nr_states || !cb)
341 * Use a mutex to serialize the registration of performance domains and
342 * let the driver-defined callback functions sleep.
344 mutex_lock(&em_pd_mutex);
351 if (_is_cpu_device(dev)) {
353 dev_err(dev, "EM: invalid CPU mask\n");
358 for_each_cpu(cpu, cpus) {
359 if (em_cpu_get(cpu)) {
360 dev_err(dev, "EM: exists for CPU%d\n", cpu);
365 * All CPUs of a domain must have the same
366 * micro-architecture since they all share the same
369 cap = arch_scale_cpu_capacity(cpu);
370 if (prev_cap && prev_cap != cap) {
371 dev_err(dev, "EM: CPUs of %*pbl must have the same capacity\n",
372 cpumask_pr_args(cpus));
381 ret = em_create_pd(dev, nr_states, cb, cpus);
386 dev->em_pd->flags |= EM_PERF_DOMAIN_MILLIWATTS;
387 else if (cb->get_cost)
388 dev->em_pd->flags |= EM_PERF_DOMAIN_ARTIFICIAL;
390 em_cpufreq_update_efficiencies(dev);
392 em_debug_create_pd(dev);
393 dev_info(dev, "EM: created perf domain\n");
396 mutex_unlock(&em_pd_mutex);
399 EXPORT_SYMBOL_GPL(em_dev_register_perf_domain);
402 * em_dev_unregister_perf_domain() - Unregister Energy Model (EM) for a device
403 * @dev : Device for which the EM is registered
405 * Unregister the EM for the specified @dev (but not a CPU device).
407 void em_dev_unregister_perf_domain(struct device *dev)
409 if (IS_ERR_OR_NULL(dev) || !dev->em_pd)
412 if (_is_cpu_device(dev))
416 * The mutex separates all register/unregister requests and protects
417 * from potential clean-up/setup issues in the debugfs directories.
418 * The debugfs directory name is the same as device's name.
420 mutex_lock(&em_pd_mutex);
421 em_debug_remove_pd(dev);
423 kfree(dev->em_pd->table);
426 mutex_unlock(&em_pd_mutex);
428 EXPORT_SYMBOL_GPL(em_dev_unregister_perf_domain);