2 * CPUFreq governor based on scheduler-provided CPU utilization data.
4 * Copyright (C) 2016, Intel Corporation
5 * Author: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
7 * This program is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License version 2 as
9 * published by the Free Software Foundation.
12 #include <linux/cpufreq.h>
13 #include <linux/module.h>
14 #include <linux/slab.h>
15 #include <trace/events/power.h>
19 struct sugov_tunables {
20 struct gov_attr_set attr_set;
21 unsigned int rate_limit_us;
25 struct cpufreq_policy *policy;
27 struct sugov_tunables *tunables;
28 struct list_head tunables_hook;
30 raw_spinlock_t update_lock; /* For shared policies */
31 u64 last_freq_update_time;
32 s64 freq_update_delay_ns;
33 unsigned int next_freq;
35 /* The next fields are only needed if fast switch cannot be used. */
36 struct irq_work irq_work;
37 struct work_struct work;
38 struct mutex work_lock;
39 bool work_in_progress;
41 bool need_freq_update;
45 struct update_util_data update_util;
46 struct sugov_policy *sg_policy;
48 /* The fields below are only needed when sharing a policy. */
54 static DEFINE_PER_CPU(struct sugov_cpu, sugov_cpu);
56 /************************ Governor internals ***********************/
58 static bool sugov_should_update_freq(struct sugov_policy *sg_policy, u64 time)
62 if (sg_policy->work_in_progress)
65 if (unlikely(sg_policy->need_freq_update)) {
66 sg_policy->need_freq_update = false;
68 * This happens when limits change, so forget the previous
69 * next_freq value and force an update.
71 sg_policy->next_freq = UINT_MAX;
75 delta_ns = time - sg_policy->last_freq_update_time;
76 return delta_ns >= sg_policy->freq_update_delay_ns;
79 static void sugov_update_commit(struct sugov_policy *sg_policy, u64 time,
80 unsigned int next_freq)
82 struct cpufreq_policy *policy = sg_policy->policy;
84 sg_policy->last_freq_update_time = time;
86 if (policy->fast_switch_enabled) {
87 if (sg_policy->next_freq == next_freq) {
88 trace_cpu_frequency(policy->cur, smp_processor_id());
91 sg_policy->next_freq = next_freq;
92 next_freq = cpufreq_driver_fast_switch(policy, next_freq);
93 if (next_freq == CPUFREQ_ENTRY_INVALID)
96 policy->cur = next_freq;
97 trace_cpu_frequency(next_freq, smp_processor_id());
98 } else if (sg_policy->next_freq != next_freq) {
99 sg_policy->next_freq = next_freq;
100 sg_policy->work_in_progress = true;
101 irq_work_queue(&sg_policy->irq_work);
106 * get_next_freq - Compute a new frequency for a given cpufreq policy.
107 * @policy: cpufreq policy object to compute the new frequency for.
108 * @util: Current CPU utilization.
109 * @max: CPU capacity.
111 * If the utilization is frequency-invariant, choose the new frequency to be
112 * proportional to it, that is
114 * next_freq = C * max_freq * util / max
116 * Otherwise, approximate the would-be frequency-invariant utilization by
117 * util_raw * (curr_freq / max_freq) which leads to
119 * next_freq = C * curr_freq * util_raw / max
121 * Take C = 1.25 for the frequency tipping point at (util / max) = 0.8.
123 static unsigned int get_next_freq(struct cpufreq_policy *policy,
124 unsigned long util, unsigned long max)
126 unsigned int freq = arch_scale_freq_invariant() ?
127 policy->cpuinfo.max_freq : policy->cur;
129 return (freq + (freq >> 2)) * util / max;
132 static void sugov_update_single(struct update_util_data *hook, u64 time,
133 unsigned long util, unsigned long max)
135 struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
136 struct sugov_policy *sg_policy = sg_cpu->sg_policy;
137 struct cpufreq_policy *policy = sg_policy->policy;
140 if (!sugov_should_update_freq(sg_policy, time))
143 next_f = util == ULONG_MAX ? policy->cpuinfo.max_freq :
144 get_next_freq(policy, util, max);
145 sugov_update_commit(sg_policy, time, next_f);
148 static unsigned int sugov_next_freq_shared(struct sugov_policy *sg_policy,
149 unsigned long util, unsigned long max)
151 struct cpufreq_policy *policy = sg_policy->policy;
152 unsigned int max_f = policy->cpuinfo.max_freq;
153 u64 last_freq_update_time = sg_policy->last_freq_update_time;
156 if (util == ULONG_MAX)
159 for_each_cpu(j, policy->cpus) {
160 struct sugov_cpu *j_sg_cpu;
161 unsigned long j_util, j_max;
164 if (j == smp_processor_id())
167 j_sg_cpu = &per_cpu(sugov_cpu, j);
169 * If the CPU utilization was last updated before the previous
170 * frequency update and the time elapsed between the last update
171 * of the CPU utilization and the last frequency update is long
172 * enough, don't take the CPU into account as it probably is
175 delta_ns = last_freq_update_time - j_sg_cpu->last_update;
176 if (delta_ns > TICK_NSEC)
179 j_util = j_sg_cpu->util;
180 if (j_util == ULONG_MAX)
183 j_max = j_sg_cpu->max;
184 if (j_util * max > j_max * util) {
190 return get_next_freq(policy, util, max);
193 static void sugov_update_shared(struct update_util_data *hook, u64 time,
194 unsigned long util, unsigned long max)
196 struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
197 struct sugov_policy *sg_policy = sg_cpu->sg_policy;
200 raw_spin_lock(&sg_policy->update_lock);
204 sg_cpu->last_update = time;
206 if (sugov_should_update_freq(sg_policy, time)) {
207 next_f = sugov_next_freq_shared(sg_policy, util, max);
208 sugov_update_commit(sg_policy, time, next_f);
211 raw_spin_unlock(&sg_policy->update_lock);
214 static void sugov_work(struct work_struct *work)
216 struct sugov_policy *sg_policy = container_of(work, struct sugov_policy, work);
218 mutex_lock(&sg_policy->work_lock);
219 __cpufreq_driver_target(sg_policy->policy, sg_policy->next_freq,
221 mutex_unlock(&sg_policy->work_lock);
223 sg_policy->work_in_progress = false;
226 static void sugov_irq_work(struct irq_work *irq_work)
228 struct sugov_policy *sg_policy;
230 sg_policy = container_of(irq_work, struct sugov_policy, irq_work);
231 schedule_work_on(smp_processor_id(), &sg_policy->work);
234 /************************** sysfs interface ************************/
236 static struct sugov_tunables *global_tunables;
237 static DEFINE_MUTEX(global_tunables_lock);
239 static inline struct sugov_tunables *to_sugov_tunables(struct gov_attr_set *attr_set)
241 return container_of(attr_set, struct sugov_tunables, attr_set);
244 static ssize_t rate_limit_us_show(struct gov_attr_set *attr_set, char *buf)
246 struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
248 return sprintf(buf, "%u\n", tunables->rate_limit_us);
251 static ssize_t rate_limit_us_store(struct gov_attr_set *attr_set, const char *buf,
254 struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
255 struct sugov_policy *sg_policy;
256 unsigned int rate_limit_us;
258 if (kstrtouint(buf, 10, &rate_limit_us))
261 tunables->rate_limit_us = rate_limit_us;
263 list_for_each_entry(sg_policy, &attr_set->policy_list, tunables_hook)
264 sg_policy->freq_update_delay_ns = rate_limit_us * NSEC_PER_USEC;
269 static struct governor_attr rate_limit_us = __ATTR_RW(rate_limit_us);
271 static struct attribute *sugov_attributes[] = {
276 static struct kobj_type sugov_tunables_ktype = {
277 .default_attrs = sugov_attributes,
278 .sysfs_ops = &governor_sysfs_ops,
281 /********************** cpufreq governor interface *********************/
283 static struct cpufreq_governor schedutil_gov;
285 static struct sugov_policy *sugov_policy_alloc(struct cpufreq_policy *policy)
287 struct sugov_policy *sg_policy;
289 sg_policy = kzalloc(sizeof(*sg_policy), GFP_KERNEL);
293 sg_policy->policy = policy;
294 init_irq_work(&sg_policy->irq_work, sugov_irq_work);
295 INIT_WORK(&sg_policy->work, sugov_work);
296 mutex_init(&sg_policy->work_lock);
297 raw_spin_lock_init(&sg_policy->update_lock);
301 static void sugov_policy_free(struct sugov_policy *sg_policy)
303 mutex_destroy(&sg_policy->work_lock);
307 static struct sugov_tunables *sugov_tunables_alloc(struct sugov_policy *sg_policy)
309 struct sugov_tunables *tunables;
311 tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
313 gov_attr_set_init(&tunables->attr_set, &sg_policy->tunables_hook);
314 if (!have_governor_per_policy())
315 global_tunables = tunables;
320 static void sugov_tunables_free(struct sugov_tunables *tunables)
322 if (!have_governor_per_policy())
323 global_tunables = NULL;
328 static int sugov_init(struct cpufreq_policy *policy)
330 struct sugov_policy *sg_policy;
331 struct sugov_tunables *tunables;
335 /* State should be equivalent to EXIT */
336 if (policy->governor_data)
339 sg_policy = sugov_policy_alloc(policy);
343 mutex_lock(&global_tunables_lock);
345 if (global_tunables) {
346 if (WARN_ON(have_governor_per_policy())) {
350 policy->governor_data = sg_policy;
351 sg_policy->tunables = global_tunables;
353 gov_attr_set_get(&global_tunables->attr_set, &sg_policy->tunables_hook);
357 tunables = sugov_tunables_alloc(sg_policy);
363 tunables->rate_limit_us = LATENCY_MULTIPLIER;
364 lat = policy->cpuinfo.transition_latency / NSEC_PER_USEC;
366 tunables->rate_limit_us *= lat;
368 policy->governor_data = sg_policy;
369 sg_policy->tunables = tunables;
371 ret = kobject_init_and_add(&tunables->attr_set.kobj, &sugov_tunables_ktype,
372 get_governor_parent_kobj(policy), "%s",
378 mutex_unlock(&global_tunables_lock);
380 cpufreq_enable_fast_switch(policy);
384 policy->governor_data = NULL;
385 sugov_tunables_free(tunables);
388 mutex_unlock(&global_tunables_lock);
390 sugov_policy_free(sg_policy);
391 pr_err("cpufreq: schedutil governor initialization failed (error %d)\n", ret);
395 static int sugov_exit(struct cpufreq_policy *policy)
397 struct sugov_policy *sg_policy = policy->governor_data;
398 struct sugov_tunables *tunables = sg_policy->tunables;
401 cpufreq_disable_fast_switch(policy);
403 mutex_lock(&global_tunables_lock);
405 count = gov_attr_set_put(&tunables->attr_set, &sg_policy->tunables_hook);
406 policy->governor_data = NULL;
408 sugov_tunables_free(tunables);
410 mutex_unlock(&global_tunables_lock);
412 sugov_policy_free(sg_policy);
416 static int sugov_start(struct cpufreq_policy *policy)
418 struct sugov_policy *sg_policy = policy->governor_data;
421 sg_policy->freq_update_delay_ns = sg_policy->tunables->rate_limit_us * NSEC_PER_USEC;
422 sg_policy->last_freq_update_time = 0;
423 sg_policy->next_freq = UINT_MAX;
424 sg_policy->work_in_progress = false;
425 sg_policy->need_freq_update = false;
427 for_each_cpu(cpu, policy->cpus) {
428 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
430 sg_cpu->sg_policy = sg_policy;
431 if (policy_is_shared(policy)) {
432 sg_cpu->util = ULONG_MAX;
434 sg_cpu->last_update = 0;
435 cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util,
436 sugov_update_shared);
438 cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util,
439 sugov_update_single);
445 static int sugov_stop(struct cpufreq_policy *policy)
447 struct sugov_policy *sg_policy = policy->governor_data;
450 for_each_cpu(cpu, policy->cpus)
451 cpufreq_remove_update_util_hook(cpu);
455 irq_work_sync(&sg_policy->irq_work);
456 cancel_work_sync(&sg_policy->work);
460 static int sugov_limits(struct cpufreq_policy *policy)
462 struct sugov_policy *sg_policy = policy->governor_data;
464 if (!policy->fast_switch_enabled) {
465 mutex_lock(&sg_policy->work_lock);
467 if (policy->max < policy->cur)
468 __cpufreq_driver_target(policy, policy->max,
470 else if (policy->min > policy->cur)
471 __cpufreq_driver_target(policy, policy->min,
474 mutex_unlock(&sg_policy->work_lock);
477 sg_policy->need_freq_update = true;
481 int sugov_governor(struct cpufreq_policy *policy, unsigned int event)
483 if (event == CPUFREQ_GOV_POLICY_INIT) {
484 return sugov_init(policy);
485 } else if (policy->governor_data) {
487 case CPUFREQ_GOV_POLICY_EXIT:
488 return sugov_exit(policy);
489 case CPUFREQ_GOV_START:
490 return sugov_start(policy);
491 case CPUFREQ_GOV_STOP:
492 return sugov_stop(policy);
493 case CPUFREQ_GOV_LIMITS:
494 return sugov_limits(policy);
500 static struct cpufreq_governor schedutil_gov = {
502 .governor = sugov_governor,
503 .owner = THIS_MODULE,
506 static int __init sugov_module_init(void)
508 return cpufreq_register_governor(&schedutil_gov);
511 static void __exit sugov_module_exit(void)
513 cpufreq_unregister_governor(&schedutil_gov);
516 MODULE_AUTHOR("Rafael J. Wysocki <rafael.j.wysocki@intel.com>");
517 MODULE_DESCRIPTION("Utilization-based CPU frequency selection");
518 MODULE_LICENSE("GPL");
520 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL
521 struct cpufreq_governor *cpufreq_default_governor(void)
523 return &schedutil_gov;
526 fs_initcall(sugov_module_init);
528 module_init(sugov_module_init);
530 module_exit(sugov_module_exit);