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_flags_show(struct seq_file *s, void *unused)
59 struct em_perf_domain *pd = s->private;
61 seq_printf(s, "%#lx\n", pd->flags);
65 DEFINE_SHOW_ATTRIBUTE(em_debug_flags);
67 static void em_debug_create_pd(struct device *dev)
72 /* Create the directory of the performance domain */
73 d = debugfs_create_dir(dev_name(dev), rootdir);
75 if (_is_cpu_device(dev))
76 debugfs_create_file("cpus", 0444, d, dev->em_pd->cpus,
79 debugfs_create_file("flags", 0444, d, dev->em_pd,
80 &em_debug_flags_fops);
82 /* Create a sub-directory for each performance state */
83 for (i = 0; i < dev->em_pd->nr_perf_states; i++)
84 em_debug_create_ps(&dev->em_pd->table[i], d);
88 static void em_debug_remove_pd(struct device *dev)
90 debugfs_lookup_and_remove(dev_name(dev), rootdir);
93 static int __init em_debug_init(void)
95 /* Create /sys/kernel/debug/energy_model directory */
96 rootdir = debugfs_create_dir("energy_model", NULL);
100 fs_initcall(em_debug_init);
101 #else /* CONFIG_DEBUG_FS */
102 static void em_debug_create_pd(struct device *dev) {}
103 static void em_debug_remove_pd(struct device *dev) {}
106 static int em_create_perf_table(struct device *dev, struct em_perf_domain *pd,
107 int nr_states, struct em_data_callback *cb,
110 unsigned long power, freq, prev_freq = 0, prev_cost = ULONG_MAX;
111 struct em_perf_state *table;
115 table = kcalloc(nr_states, sizeof(*table), GFP_KERNEL);
119 /* Build the list of performance states for this performance domain */
120 for (i = 0, freq = 0; i < nr_states; i++, freq++) {
122 * active_power() is a driver callback which ceils 'freq' to
123 * lowest performance state of 'dev' above 'freq' and updates
124 * 'power' and 'freq' accordingly.
126 ret = cb->active_power(dev, &power, &freq);
128 dev_err(dev, "EM: invalid perf. state: %d\n",
134 * We expect the driver callback to increase the frequency for
135 * higher performance states.
137 if (freq <= prev_freq) {
138 dev_err(dev, "EM: non-increasing freq: %lu\n",
144 * The power returned by active_state() is expected to be
145 * positive and be in range.
147 if (!power || power > EM_MAX_POWER) {
148 dev_err(dev, "EM: invalid power: %lu\n",
153 table[i].power = power;
154 table[i].frequency = prev_freq = freq;
157 /* Compute the cost of each performance state. */
158 fmax = (u64) table[nr_states - 1].frequency;
159 for (i = nr_states - 1; i >= 0; i--) {
160 unsigned long power_res, cost;
162 if (flags & EM_PERF_DOMAIN_ARTIFICIAL) {
163 ret = cb->get_cost(dev, table[i].frequency, &cost);
164 if (ret || !cost || cost > EM_MAX_POWER) {
165 dev_err(dev, "EM: invalid cost %lu %d\n",
170 power_res = table[i].power;
171 cost = div64_u64(fmax * power_res, table[i].frequency);
174 table[i].cost = cost;
176 if (table[i].cost >= prev_cost) {
177 table[i].flags = EM_PERF_STATE_INEFFICIENT;
178 dev_dbg(dev, "EM: OPP:%lu is inefficient\n",
181 prev_cost = table[i].cost;
186 pd->nr_perf_states = nr_states;
195 static int em_create_pd(struct device *dev, int nr_states,
196 struct em_data_callback *cb, cpumask_t *cpus,
199 struct em_perf_domain *pd;
200 struct device *cpu_dev;
201 int cpu, ret, num_cpus;
203 if (_is_cpu_device(dev)) {
204 num_cpus = cpumask_weight(cpus);
206 /* Prevent max possible energy calculation to not overflow */
207 if (num_cpus > EM_MAX_NUM_CPUS) {
208 dev_err(dev, "EM: too many CPUs, overflow possible\n");
212 pd = kzalloc(sizeof(*pd) + cpumask_size(), GFP_KERNEL);
216 cpumask_copy(em_span_cpus(pd), cpus);
218 pd = kzalloc(sizeof(*pd), GFP_KERNEL);
223 ret = em_create_perf_table(dev, pd, nr_states, cb, flags);
229 if (_is_cpu_device(dev))
230 for_each_cpu(cpu, cpus) {
231 cpu_dev = get_cpu_device(cpu);
240 static void em_cpufreq_update_efficiencies(struct device *dev)
242 struct em_perf_domain *pd = dev->em_pd;
243 struct em_perf_state *table;
244 struct cpufreq_policy *policy;
248 if (!_is_cpu_device(dev) || !pd)
251 policy = cpufreq_cpu_get(cpumask_first(em_span_cpus(pd)));
253 dev_warn(dev, "EM: Access to CPUFreq policy failed");
259 for (i = 0; i < pd->nr_perf_states; i++) {
260 if (!(table[i].flags & EM_PERF_STATE_INEFFICIENT))
263 if (!cpufreq_table_set_inefficient(policy, table[i].frequency))
267 cpufreq_cpu_put(policy);
273 * Efficiencies have been installed in CPUFreq, inefficient frequencies
274 * will be skipped. The EM can do the same.
276 pd->flags |= EM_PERF_DOMAIN_SKIP_INEFFICIENCIES;
280 * em_pd_get() - Return the performance domain for a device
281 * @dev : Device to find the performance domain for
283 * Returns the performance domain to which @dev belongs, or NULL if it doesn't
286 struct em_perf_domain *em_pd_get(struct device *dev)
288 if (IS_ERR_OR_NULL(dev))
293 EXPORT_SYMBOL_GPL(em_pd_get);
296 * em_cpu_get() - Return the performance domain for a CPU
297 * @cpu : CPU to find the performance domain for
299 * Returns the performance domain to which @cpu belongs, or NULL if it doesn't
302 struct em_perf_domain *em_cpu_get(int cpu)
304 struct device *cpu_dev;
306 cpu_dev = get_cpu_device(cpu);
310 return em_pd_get(cpu_dev);
312 EXPORT_SYMBOL_GPL(em_cpu_get);
315 * em_dev_register_perf_domain() - Register the Energy Model (EM) for a device
316 * @dev : Device for which the EM is to register
317 * @nr_states : Number of performance states to register
318 * @cb : Callback functions providing the data of the Energy Model
319 * @cpus : Pointer to cpumask_t, which in case of a CPU device is
320 * obligatory. It can be taken from i.e. 'policy->cpus'. For other
321 * type of devices this should be set to NULL.
322 * @microwatts : Flag indicating that the power values are in micro-Watts or
323 * in some other scale. It must be set properly.
325 * Create Energy Model tables for a performance domain using the callbacks
328 * The @microwatts is important to set with correct value. Some kernel
329 * sub-systems might rely on this flag and check if all devices in the EM are
330 * using the same scale.
332 * If multiple clients register the same performance domain, all but the first
333 * registration will be ignored.
335 * Return 0 on success
337 int em_dev_register_perf_domain(struct device *dev, unsigned int nr_states,
338 struct em_data_callback *cb, cpumask_t *cpus,
341 unsigned long cap, prev_cap = 0;
342 unsigned long flags = 0;
345 if (!dev || !nr_states || !cb)
349 * Use a mutex to serialize the registration of performance domains and
350 * let the driver-defined callback functions sleep.
352 mutex_lock(&em_pd_mutex);
359 if (_is_cpu_device(dev)) {
361 dev_err(dev, "EM: invalid CPU mask\n");
366 for_each_cpu(cpu, cpus) {
367 if (em_cpu_get(cpu)) {
368 dev_err(dev, "EM: exists for CPU%d\n", cpu);
373 * All CPUs of a domain must have the same
374 * micro-architecture since they all share the same
377 cap = arch_scale_cpu_capacity(cpu);
378 if (prev_cap && prev_cap != cap) {
379 dev_err(dev, "EM: CPUs of %*pbl must have the same capacity\n",
380 cpumask_pr_args(cpus));
390 flags |= EM_PERF_DOMAIN_MICROWATTS;
391 else if (cb->get_cost)
392 flags |= EM_PERF_DOMAIN_ARTIFICIAL;
394 ret = em_create_pd(dev, nr_states, cb, cpus, flags);
398 dev->em_pd->flags |= flags;
400 em_cpufreq_update_efficiencies(dev);
402 em_debug_create_pd(dev);
403 dev_info(dev, "EM: created perf domain\n");
406 mutex_unlock(&em_pd_mutex);
409 EXPORT_SYMBOL_GPL(em_dev_register_perf_domain);
412 * em_dev_unregister_perf_domain() - Unregister Energy Model (EM) for a device
413 * @dev : Device for which the EM is registered
415 * Unregister the EM for the specified @dev (but not a CPU device).
417 void em_dev_unregister_perf_domain(struct device *dev)
419 if (IS_ERR_OR_NULL(dev) || !dev->em_pd)
422 if (_is_cpu_device(dev))
426 * The mutex separates all register/unregister requests and protects
427 * from potential clean-up/setup issues in the debugfs directories.
428 * The debugfs directory name is the same as device's name.
430 mutex_lock(&em_pd_mutex);
431 em_debug_remove_pd(dev);
433 kfree(dev->em_pd->table);
436 mutex_unlock(&em_pd_mutex);
438 EXPORT_SYMBOL_GPL(em_dev_unregister_perf_domain);