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 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
14 #include <linux/cpufreq.h>
15 #include <linux/slab.h>
16 #include <trace/events/power.h>
20 struct sugov_tunables {
21 struct gov_attr_set attr_set;
22 unsigned int rate_limit_us;
26 struct cpufreq_policy *policy;
28 struct sugov_tunables *tunables;
29 struct list_head tunables_hook;
31 raw_spinlock_t update_lock; /* For shared policies */
32 u64 last_freq_update_time;
33 s64 freq_update_delay_ns;
34 unsigned int next_freq;
35 unsigned int cached_raw_freq;
37 /* The next fields are only needed if fast switch cannot be used. */
38 struct irq_work irq_work;
39 struct work_struct work;
40 struct mutex work_lock;
41 bool work_in_progress;
43 bool need_freq_update;
47 struct update_util_data update_util;
48 struct sugov_policy *sg_policy;
50 unsigned long iowait_boost;
51 unsigned long iowait_boost_max;
54 /* The fields below are only needed when sharing a policy. */
60 static DEFINE_PER_CPU(struct sugov_cpu, sugov_cpu);
62 /************************ Governor internals ***********************/
64 static bool sugov_should_update_freq(struct sugov_policy *sg_policy, u64 time)
68 if (sg_policy->work_in_progress)
71 if (unlikely(sg_policy->need_freq_update)) {
72 sg_policy->need_freq_update = false;
74 * This happens when limits change, so forget the previous
75 * next_freq value and force an update.
77 sg_policy->next_freq = UINT_MAX;
81 delta_ns = time - sg_policy->last_freq_update_time;
82 return delta_ns >= sg_policy->freq_update_delay_ns;
85 static void sugov_update_commit(struct sugov_policy *sg_policy, u64 time,
86 unsigned int next_freq)
88 struct cpufreq_policy *policy = sg_policy->policy;
90 sg_policy->last_freq_update_time = time;
92 if (policy->fast_switch_enabled) {
93 if (sg_policy->next_freq == next_freq) {
94 trace_cpu_frequency(policy->cur, smp_processor_id());
97 sg_policy->next_freq = next_freq;
98 next_freq = cpufreq_driver_fast_switch(policy, next_freq);
99 if (next_freq == CPUFREQ_ENTRY_INVALID)
102 policy->cur = next_freq;
103 trace_cpu_frequency(next_freq, smp_processor_id());
104 } else if (sg_policy->next_freq != next_freq) {
105 sg_policy->next_freq = next_freq;
106 sg_policy->work_in_progress = true;
107 irq_work_queue(&sg_policy->irq_work);
112 * get_next_freq - Compute a new frequency for a given cpufreq policy.
113 * @sg_cpu: schedutil cpu object to compute the new frequency for.
114 * @util: Current CPU utilization.
115 * @max: CPU capacity.
117 * If the utilization is frequency-invariant, choose the new frequency to be
118 * proportional to it, that is
120 * next_freq = C * max_freq * util / max
122 * Otherwise, approximate the would-be frequency-invariant utilization by
123 * util_raw * (curr_freq / max_freq) which leads to
125 * next_freq = C * curr_freq * util_raw / max
127 * Take C = 1.25 for the frequency tipping point at (util / max) = 0.8.
129 * The lowest driver-supported frequency which is equal or greater than the raw
130 * next_freq (as calculated above) is returned, subject to policy min/max and
131 * cpufreq driver limitations.
133 static unsigned int get_next_freq(struct sugov_cpu *sg_cpu, unsigned long util,
136 struct sugov_policy *sg_policy = sg_cpu->sg_policy;
137 struct cpufreq_policy *policy = sg_policy->policy;
138 unsigned int freq = arch_scale_freq_invariant() ?
139 policy->cpuinfo.max_freq : policy->cur;
141 freq = (freq + (freq >> 2)) * util / max;
143 if (freq == sg_policy->cached_raw_freq && sg_policy->next_freq != UINT_MAX)
144 return sg_policy->next_freq;
145 sg_policy->cached_raw_freq = freq;
146 return cpufreq_driver_resolve_freq(policy, freq);
149 static void sugov_get_util(unsigned long *util, unsigned long *max)
151 struct rq *rq = this_rq();
152 unsigned long cfs_max;
154 cfs_max = arch_scale_cpu_capacity(NULL, smp_processor_id());
156 *util = min(rq->cfs.avg.util_avg, cfs_max);
160 static void sugov_set_iowait_boost(struct sugov_cpu *sg_cpu, u64 time,
163 if (flags & SCHED_CPUFREQ_IOWAIT) {
164 sg_cpu->iowait_boost = sg_cpu->iowait_boost_max;
165 } else if (sg_cpu->iowait_boost) {
166 s64 delta_ns = time - sg_cpu->last_update;
168 /* Clear iowait_boost if the CPU apprears to have been idle. */
169 if (delta_ns > TICK_NSEC)
170 sg_cpu->iowait_boost = 0;
174 static void sugov_iowait_boost(struct sugov_cpu *sg_cpu, unsigned long *util,
177 unsigned long boost_util = sg_cpu->iowait_boost;
178 unsigned long boost_max = sg_cpu->iowait_boost_max;
183 if (*util * boost_max < *max * boost_util) {
187 sg_cpu->iowait_boost >>= 1;
190 static void sugov_update_single(struct update_util_data *hook, u64 time,
193 struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
194 struct sugov_policy *sg_policy = sg_cpu->sg_policy;
195 struct cpufreq_policy *policy = sg_policy->policy;
196 unsigned long util, max;
199 sugov_set_iowait_boost(sg_cpu, time, flags);
200 sg_cpu->last_update = time;
202 if (!sugov_should_update_freq(sg_policy, time))
205 if (flags & SCHED_CPUFREQ_RT_DL) {
206 next_f = policy->cpuinfo.max_freq;
208 sugov_get_util(&util, &max);
209 sugov_iowait_boost(sg_cpu, &util, &max);
210 next_f = get_next_freq(sg_cpu, util, max);
212 sugov_update_commit(sg_policy, time, next_f);
215 static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu,
216 unsigned long util, unsigned long max,
219 struct sugov_policy *sg_policy = sg_cpu->sg_policy;
220 struct cpufreq_policy *policy = sg_policy->policy;
221 unsigned int max_f = policy->cpuinfo.max_freq;
222 u64 last_freq_update_time = sg_policy->last_freq_update_time;
225 if (flags & SCHED_CPUFREQ_RT_DL)
228 sugov_iowait_boost(sg_cpu, &util, &max);
230 for_each_cpu(j, policy->cpus) {
231 struct sugov_cpu *j_sg_cpu;
232 unsigned long j_util, j_max;
235 if (j == smp_processor_id())
238 j_sg_cpu = &per_cpu(sugov_cpu, j);
240 * If the CPU utilization was last updated before the previous
241 * frequency update and the time elapsed between the last update
242 * of the CPU utilization and the last frequency update is long
243 * enough, don't take the CPU into account as it probably is
244 * idle now (and clear iowait_boost for it).
246 delta_ns = last_freq_update_time - j_sg_cpu->last_update;
247 if (delta_ns > TICK_NSEC) {
248 j_sg_cpu->iowait_boost = 0;
251 if (j_sg_cpu->flags & SCHED_CPUFREQ_RT_DL)
254 j_util = j_sg_cpu->util;
255 j_max = j_sg_cpu->max;
256 if (j_util * max > j_max * util) {
261 sugov_iowait_boost(j_sg_cpu, &util, &max);
264 return get_next_freq(sg_cpu, util, max);
267 static void sugov_update_shared(struct update_util_data *hook, u64 time,
270 struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
271 struct sugov_policy *sg_policy = sg_cpu->sg_policy;
272 unsigned long util, max;
275 sugov_get_util(&util, &max);
277 raw_spin_lock(&sg_policy->update_lock);
281 sg_cpu->flags = flags;
283 sugov_set_iowait_boost(sg_cpu, time, flags);
284 sg_cpu->last_update = time;
286 if (sugov_should_update_freq(sg_policy, time)) {
287 next_f = sugov_next_freq_shared(sg_cpu, util, max, flags);
288 sugov_update_commit(sg_policy, time, next_f);
291 raw_spin_unlock(&sg_policy->update_lock);
294 static void sugov_work(struct work_struct *work)
296 struct sugov_policy *sg_policy = container_of(work, struct sugov_policy, work);
298 mutex_lock(&sg_policy->work_lock);
299 __cpufreq_driver_target(sg_policy->policy, sg_policy->next_freq,
301 mutex_unlock(&sg_policy->work_lock);
303 sg_policy->work_in_progress = false;
306 static void sugov_irq_work(struct irq_work *irq_work)
308 struct sugov_policy *sg_policy;
310 sg_policy = container_of(irq_work, struct sugov_policy, irq_work);
311 schedule_work_on(smp_processor_id(), &sg_policy->work);
314 /************************** sysfs interface ************************/
316 static struct sugov_tunables *global_tunables;
317 static DEFINE_MUTEX(global_tunables_lock);
319 static inline struct sugov_tunables *to_sugov_tunables(struct gov_attr_set *attr_set)
321 return container_of(attr_set, struct sugov_tunables, attr_set);
324 static ssize_t rate_limit_us_show(struct gov_attr_set *attr_set, char *buf)
326 struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
328 return sprintf(buf, "%u\n", tunables->rate_limit_us);
331 static ssize_t rate_limit_us_store(struct gov_attr_set *attr_set, const char *buf,
334 struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
335 struct sugov_policy *sg_policy;
336 unsigned int rate_limit_us;
338 if (kstrtouint(buf, 10, &rate_limit_us))
341 tunables->rate_limit_us = rate_limit_us;
343 list_for_each_entry(sg_policy, &attr_set->policy_list, tunables_hook)
344 sg_policy->freq_update_delay_ns = rate_limit_us * NSEC_PER_USEC;
349 static struct governor_attr rate_limit_us = __ATTR_RW(rate_limit_us);
351 static struct attribute *sugov_attributes[] = {
356 static void sugov_tunables_free(struct kobject *kobj)
358 struct gov_attr_set *attr_set = container_of(kobj, struct gov_attr_set, kobj);
360 kfree(to_sugov_tunables(attr_set));
363 static struct kobj_type sugov_tunables_ktype = {
364 .default_attrs = sugov_attributes,
365 .sysfs_ops = &governor_sysfs_ops,
366 .release = &sugov_tunables_free,
369 /********************** cpufreq governor interface *********************/
371 static struct cpufreq_governor schedutil_gov;
373 static struct sugov_policy *sugov_policy_alloc(struct cpufreq_policy *policy)
375 struct sugov_policy *sg_policy;
377 sg_policy = kzalloc(sizeof(*sg_policy), GFP_KERNEL);
381 sg_policy->policy = policy;
382 init_irq_work(&sg_policy->irq_work, sugov_irq_work);
383 INIT_WORK(&sg_policy->work, sugov_work);
384 mutex_init(&sg_policy->work_lock);
385 raw_spin_lock_init(&sg_policy->update_lock);
389 static void sugov_policy_free(struct sugov_policy *sg_policy)
391 mutex_destroy(&sg_policy->work_lock);
395 static struct sugov_tunables *sugov_tunables_alloc(struct sugov_policy *sg_policy)
397 struct sugov_tunables *tunables;
399 tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
401 gov_attr_set_init(&tunables->attr_set, &sg_policy->tunables_hook);
402 if (!have_governor_per_policy())
403 global_tunables = tunables;
408 static void sugov_clear_global_tunables(void)
410 if (!have_governor_per_policy())
411 global_tunables = NULL;
414 static int sugov_init(struct cpufreq_policy *policy)
416 struct sugov_policy *sg_policy;
417 struct sugov_tunables *tunables;
421 /* State should be equivalent to EXIT */
422 if (policy->governor_data)
425 sg_policy = sugov_policy_alloc(policy);
429 mutex_lock(&global_tunables_lock);
431 if (global_tunables) {
432 if (WARN_ON(have_governor_per_policy())) {
436 policy->governor_data = sg_policy;
437 sg_policy->tunables = global_tunables;
439 gov_attr_set_get(&global_tunables->attr_set, &sg_policy->tunables_hook);
443 tunables = sugov_tunables_alloc(sg_policy);
449 tunables->rate_limit_us = LATENCY_MULTIPLIER;
450 lat = policy->cpuinfo.transition_latency / NSEC_PER_USEC;
452 tunables->rate_limit_us *= lat;
454 policy->governor_data = sg_policy;
455 sg_policy->tunables = tunables;
457 ret = kobject_init_and_add(&tunables->attr_set.kobj, &sugov_tunables_ktype,
458 get_governor_parent_kobj(policy), "%s",
464 mutex_unlock(&global_tunables_lock);
466 cpufreq_enable_fast_switch(policy);
470 policy->governor_data = NULL;
471 sugov_clear_global_tunables();
474 mutex_unlock(&global_tunables_lock);
476 sugov_policy_free(sg_policy);
477 pr_err("initialization failed (error %d)\n", ret);
481 static void sugov_exit(struct cpufreq_policy *policy)
483 struct sugov_policy *sg_policy = policy->governor_data;
484 struct sugov_tunables *tunables = sg_policy->tunables;
487 cpufreq_disable_fast_switch(policy);
489 mutex_lock(&global_tunables_lock);
491 count = gov_attr_set_put(&tunables->attr_set, &sg_policy->tunables_hook);
492 policy->governor_data = NULL;
494 sugov_clear_global_tunables();
496 mutex_unlock(&global_tunables_lock);
498 sugov_policy_free(sg_policy);
501 static int sugov_start(struct cpufreq_policy *policy)
503 struct sugov_policy *sg_policy = policy->governor_data;
506 sg_policy->freq_update_delay_ns = sg_policy->tunables->rate_limit_us * NSEC_PER_USEC;
507 sg_policy->last_freq_update_time = 0;
508 sg_policy->next_freq = UINT_MAX;
509 sg_policy->work_in_progress = false;
510 sg_policy->need_freq_update = false;
511 sg_policy->cached_raw_freq = 0;
513 for_each_cpu(cpu, policy->cpus) {
514 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
516 memset(sg_cpu, 0, sizeof(*sg_cpu));
517 sg_cpu->sg_policy = sg_policy;
518 sg_cpu->flags = SCHED_CPUFREQ_RT;
519 sg_cpu->iowait_boost_max = policy->cpuinfo.max_freq;
520 cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util,
521 policy_is_shared(policy) ?
522 sugov_update_shared :
523 sugov_update_single);
528 static void sugov_stop(struct cpufreq_policy *policy)
530 struct sugov_policy *sg_policy = policy->governor_data;
533 for_each_cpu(cpu, policy->cpus)
534 cpufreq_remove_update_util_hook(cpu);
538 irq_work_sync(&sg_policy->irq_work);
539 cancel_work_sync(&sg_policy->work);
542 static void sugov_limits(struct cpufreq_policy *policy)
544 struct sugov_policy *sg_policy = policy->governor_data;
546 if (!policy->fast_switch_enabled) {
547 mutex_lock(&sg_policy->work_lock);
548 cpufreq_policy_apply_limits(policy);
549 mutex_unlock(&sg_policy->work_lock);
552 sg_policy->need_freq_update = true;
555 static struct cpufreq_governor schedutil_gov = {
557 .owner = THIS_MODULE,
560 .start = sugov_start,
562 .limits = sugov_limits,
565 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL
566 struct cpufreq_governor *cpufreq_default_governor(void)
568 return &schedutil_gov;
572 static int __init sugov_register(void)
574 return cpufreq_register_governor(&schedutil_gov);
576 fs_initcall(sugov_register);