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/kthread.h>
16 #include <uapi/linux/sched/types.h>
17 #include <linux/slab.h>
18 #include <trace/events/power.h>
22 #define SUGOV_KTHREAD_PRIORITY 50
24 struct sugov_tunables {
25 struct gov_attr_set attr_set;
26 unsigned int rate_limit_us;
30 struct cpufreq_policy *policy;
32 struct sugov_tunables *tunables;
33 struct list_head tunables_hook;
35 raw_spinlock_t update_lock; /* For shared policies */
36 u64 last_freq_update_time;
37 s64 freq_update_delay_ns;
38 unsigned int next_freq;
39 unsigned int cached_raw_freq;
41 /* The next fields are only needed if fast switch cannot be used. */
42 struct irq_work irq_work;
43 struct kthread_work work;
44 struct mutex work_lock;
45 struct kthread_worker worker;
46 struct task_struct *thread;
47 bool work_in_progress;
49 bool need_freq_update;
53 struct update_util_data update_util;
54 struct sugov_policy *sg_policy;
56 unsigned long iowait_boost;
57 unsigned long iowait_boost_max;
60 /* The fields below are only needed when sharing a policy. */
65 /* The field below is for single-CPU policies only. */
66 #ifdef CONFIG_NO_HZ_COMMON
67 unsigned long saved_idle_calls;
71 static DEFINE_PER_CPU(struct sugov_cpu, sugov_cpu);
73 /************************ Governor internals ***********************/
75 static bool sugov_should_update_freq(struct sugov_policy *sg_policy, u64 time)
79 if (sg_policy->work_in_progress)
82 if (unlikely(sg_policy->need_freq_update)) {
83 sg_policy->need_freq_update = false;
85 * This happens when limits change, so forget the previous
86 * next_freq value and force an update.
88 sg_policy->next_freq = UINT_MAX;
92 delta_ns = time - sg_policy->last_freq_update_time;
93 return delta_ns >= sg_policy->freq_update_delay_ns;
96 static void sugov_update_commit(struct sugov_policy *sg_policy, u64 time,
97 unsigned int next_freq)
99 struct cpufreq_policy *policy = sg_policy->policy;
101 if (sg_policy->next_freq == next_freq)
104 sg_policy->next_freq = next_freq;
105 sg_policy->last_freq_update_time = time;
107 if (policy->fast_switch_enabled) {
108 next_freq = cpufreq_driver_fast_switch(policy, next_freq);
109 if (next_freq == CPUFREQ_ENTRY_INVALID)
112 policy->cur = next_freq;
113 trace_cpu_frequency(next_freq, smp_processor_id());
115 sg_policy->work_in_progress = true;
116 irq_work_queue(&sg_policy->irq_work);
121 * get_next_freq - Compute a new frequency for a given cpufreq policy.
122 * @sg_policy: schedutil policy object to compute the new frequency for.
123 * @util: Current CPU utilization.
124 * @max: CPU capacity.
126 * If the utilization is frequency-invariant, choose the new frequency to be
127 * proportional to it, that is
129 * next_freq = C * max_freq * util / max
131 * Otherwise, approximate the would-be frequency-invariant utilization by
132 * util_raw * (curr_freq / max_freq) which leads to
134 * next_freq = C * curr_freq * util_raw / max
136 * Take C = 1.25 for the frequency tipping point at (util / max) = 0.8.
138 * The lowest driver-supported frequency which is equal or greater than the raw
139 * next_freq (as calculated above) is returned, subject to policy min/max and
140 * cpufreq driver limitations.
142 static unsigned int get_next_freq(struct sugov_policy *sg_policy,
143 unsigned long util, unsigned long max)
145 struct cpufreq_policy *policy = sg_policy->policy;
146 unsigned int freq = arch_scale_freq_invariant() ?
147 policy->cpuinfo.max_freq : policy->cur;
149 freq = (freq + (freq >> 2)) * util / max;
151 if (freq == sg_policy->cached_raw_freq && sg_policy->next_freq != UINT_MAX)
152 return sg_policy->next_freq;
153 sg_policy->cached_raw_freq = freq;
154 return cpufreq_driver_resolve_freq(policy, freq);
157 static void sugov_get_util(unsigned long *util, unsigned long *max)
159 struct rq *rq = this_rq();
160 unsigned long cfs_max;
162 cfs_max = arch_scale_cpu_capacity(NULL, smp_processor_id());
164 *util = min(rq->cfs.avg.util_avg, cfs_max);
168 static void sugov_set_iowait_boost(struct sugov_cpu *sg_cpu, u64 time,
171 if (flags & SCHED_CPUFREQ_IOWAIT) {
172 sg_cpu->iowait_boost = sg_cpu->iowait_boost_max;
173 } else if (sg_cpu->iowait_boost) {
174 s64 delta_ns = time - sg_cpu->last_update;
176 /* Clear iowait_boost if the CPU apprears to have been idle. */
177 if (delta_ns > TICK_NSEC)
178 sg_cpu->iowait_boost = 0;
182 static void sugov_iowait_boost(struct sugov_cpu *sg_cpu, unsigned long *util,
185 unsigned long boost_util = sg_cpu->iowait_boost;
186 unsigned long boost_max = sg_cpu->iowait_boost_max;
191 if (*util * boost_max < *max * boost_util) {
195 sg_cpu->iowait_boost >>= 1;
198 #ifdef CONFIG_NO_HZ_COMMON
199 static bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu)
201 unsigned long idle_calls = tick_nohz_get_idle_calls();
202 bool ret = idle_calls == sg_cpu->saved_idle_calls;
204 sg_cpu->saved_idle_calls = idle_calls;
208 static inline bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu) { return false; }
209 #endif /* CONFIG_NO_HZ_COMMON */
211 static void sugov_update_single(struct update_util_data *hook, u64 time,
214 struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
215 struct sugov_policy *sg_policy = sg_cpu->sg_policy;
216 struct cpufreq_policy *policy = sg_policy->policy;
217 unsigned long util, max;
221 sugov_set_iowait_boost(sg_cpu, time, flags);
222 sg_cpu->last_update = time;
224 if (!sugov_should_update_freq(sg_policy, time))
227 busy = sugov_cpu_is_busy(sg_cpu);
229 if (flags & SCHED_CPUFREQ_RT_DL) {
230 next_f = policy->cpuinfo.max_freq;
232 sugov_get_util(&util, &max);
233 sugov_iowait_boost(sg_cpu, &util, &max);
234 next_f = get_next_freq(sg_policy, util, max);
236 * Do not reduce the frequency if the CPU has not been idle
237 * recently, as the reduction is likely to be premature then.
239 if (busy && next_f < sg_policy->next_freq)
240 next_f = sg_policy->next_freq;
242 sugov_update_commit(sg_policy, time, next_f);
245 static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu, u64 time)
247 struct sugov_policy *sg_policy = sg_cpu->sg_policy;
248 struct cpufreq_policy *policy = sg_policy->policy;
249 unsigned long util = 0, max = 1;
252 for_each_cpu(j, policy->cpus) {
253 struct sugov_cpu *j_sg_cpu = &per_cpu(sugov_cpu, j);
254 unsigned long j_util, j_max;
258 * If the CPU utilization was last updated before the previous
259 * frequency update and the time elapsed between the last update
260 * of the CPU utilization and the last frequency update is long
261 * enough, don't take the CPU into account as it probably is
262 * idle now (and clear iowait_boost for it).
264 delta_ns = time - j_sg_cpu->last_update;
265 if (delta_ns > TICK_NSEC) {
266 j_sg_cpu->iowait_boost = 0;
269 if (j_sg_cpu->flags & SCHED_CPUFREQ_RT_DL)
270 return policy->cpuinfo.max_freq;
272 j_util = j_sg_cpu->util;
273 j_max = j_sg_cpu->max;
274 if (j_util * max > j_max * util) {
279 sugov_iowait_boost(j_sg_cpu, &util, &max);
282 return get_next_freq(sg_policy, util, max);
285 static void sugov_update_shared(struct update_util_data *hook, u64 time,
288 struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
289 struct sugov_policy *sg_policy = sg_cpu->sg_policy;
290 unsigned long util, max;
293 sugov_get_util(&util, &max);
295 raw_spin_lock(&sg_policy->update_lock);
299 sg_cpu->flags = flags;
301 sugov_set_iowait_boost(sg_cpu, time, flags);
302 sg_cpu->last_update = time;
304 if (sugov_should_update_freq(sg_policy, time)) {
305 if (flags & SCHED_CPUFREQ_RT_DL)
306 next_f = sg_policy->policy->cpuinfo.max_freq;
308 next_f = sugov_next_freq_shared(sg_cpu, time);
310 sugov_update_commit(sg_policy, time, next_f);
313 raw_spin_unlock(&sg_policy->update_lock);
316 static void sugov_work(struct kthread_work *work)
318 struct sugov_policy *sg_policy = container_of(work, struct sugov_policy, work);
320 mutex_lock(&sg_policy->work_lock);
321 __cpufreq_driver_target(sg_policy->policy, sg_policy->next_freq,
323 mutex_unlock(&sg_policy->work_lock);
325 sg_policy->work_in_progress = false;
328 static void sugov_irq_work(struct irq_work *irq_work)
330 struct sugov_policy *sg_policy;
332 sg_policy = container_of(irq_work, struct sugov_policy, irq_work);
335 * For RT and deadline tasks, the schedutil governor shoots the
336 * frequency to maximum. Special care must be taken to ensure that this
337 * kthread doesn't result in the same behavior.
339 * This is (mostly) guaranteed by the work_in_progress flag. The flag is
340 * updated only at the end of the sugov_work() function and before that
341 * the schedutil governor rejects all other frequency scaling requests.
343 * There is a very rare case though, where the RT thread yields right
344 * after the work_in_progress flag is cleared. The effects of that are
347 kthread_queue_work(&sg_policy->worker, &sg_policy->work);
350 /************************** sysfs interface ************************/
352 static struct sugov_tunables *global_tunables;
353 static DEFINE_MUTEX(global_tunables_lock);
355 static inline struct sugov_tunables *to_sugov_tunables(struct gov_attr_set *attr_set)
357 return container_of(attr_set, struct sugov_tunables, attr_set);
360 static ssize_t rate_limit_us_show(struct gov_attr_set *attr_set, char *buf)
362 struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
364 return sprintf(buf, "%u\n", tunables->rate_limit_us);
367 static ssize_t rate_limit_us_store(struct gov_attr_set *attr_set, const char *buf,
370 struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
371 struct sugov_policy *sg_policy;
372 unsigned int rate_limit_us;
374 if (kstrtouint(buf, 10, &rate_limit_us))
377 tunables->rate_limit_us = rate_limit_us;
379 list_for_each_entry(sg_policy, &attr_set->policy_list, tunables_hook)
380 sg_policy->freq_update_delay_ns = rate_limit_us * NSEC_PER_USEC;
385 static struct governor_attr rate_limit_us = __ATTR_RW(rate_limit_us);
387 static struct attribute *sugov_attributes[] = {
392 static struct kobj_type sugov_tunables_ktype = {
393 .default_attrs = sugov_attributes,
394 .sysfs_ops = &governor_sysfs_ops,
397 /********************** cpufreq governor interface *********************/
399 static struct cpufreq_governor schedutil_gov;
401 static struct sugov_policy *sugov_policy_alloc(struct cpufreq_policy *policy)
403 struct sugov_policy *sg_policy;
405 sg_policy = kzalloc(sizeof(*sg_policy), GFP_KERNEL);
409 sg_policy->policy = policy;
410 raw_spin_lock_init(&sg_policy->update_lock);
414 static void sugov_policy_free(struct sugov_policy *sg_policy)
419 static int sugov_kthread_create(struct sugov_policy *sg_policy)
421 struct task_struct *thread;
422 struct sched_param param = { .sched_priority = MAX_USER_RT_PRIO / 2 };
423 struct cpufreq_policy *policy = sg_policy->policy;
426 /* kthread only required for slow path */
427 if (policy->fast_switch_enabled)
430 kthread_init_work(&sg_policy->work, sugov_work);
431 kthread_init_worker(&sg_policy->worker);
432 thread = kthread_create(kthread_worker_fn, &sg_policy->worker,
434 cpumask_first(policy->related_cpus));
435 if (IS_ERR(thread)) {
436 pr_err("failed to create sugov thread: %ld\n", PTR_ERR(thread));
437 return PTR_ERR(thread);
440 ret = sched_setscheduler_nocheck(thread, SCHED_FIFO, ¶m);
442 kthread_stop(thread);
443 pr_warn("%s: failed to set SCHED_FIFO\n", __func__);
447 sg_policy->thread = thread;
448 kthread_bind_mask(thread, policy->related_cpus);
449 init_irq_work(&sg_policy->irq_work, sugov_irq_work);
450 mutex_init(&sg_policy->work_lock);
452 wake_up_process(thread);
457 static void sugov_kthread_stop(struct sugov_policy *sg_policy)
459 /* kthread only required for slow path */
460 if (sg_policy->policy->fast_switch_enabled)
463 kthread_flush_worker(&sg_policy->worker);
464 kthread_stop(sg_policy->thread);
465 mutex_destroy(&sg_policy->work_lock);
468 static struct sugov_tunables *sugov_tunables_alloc(struct sugov_policy *sg_policy)
470 struct sugov_tunables *tunables;
472 tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
474 gov_attr_set_init(&tunables->attr_set, &sg_policy->tunables_hook);
475 if (!have_governor_per_policy())
476 global_tunables = tunables;
481 static void sugov_tunables_free(struct sugov_tunables *tunables)
483 if (!have_governor_per_policy())
484 global_tunables = NULL;
489 static int sugov_init(struct cpufreq_policy *policy)
491 struct sugov_policy *sg_policy;
492 struct sugov_tunables *tunables;
495 /* State should be equivalent to EXIT */
496 if (policy->governor_data)
499 cpufreq_enable_fast_switch(policy);
501 sg_policy = sugov_policy_alloc(policy);
504 goto disable_fast_switch;
507 ret = sugov_kthread_create(sg_policy);
511 mutex_lock(&global_tunables_lock);
513 if (global_tunables) {
514 if (WARN_ON(have_governor_per_policy())) {
518 policy->governor_data = sg_policy;
519 sg_policy->tunables = global_tunables;
521 gov_attr_set_get(&global_tunables->attr_set, &sg_policy->tunables_hook);
525 tunables = sugov_tunables_alloc(sg_policy);
531 if (policy->transition_delay_us) {
532 tunables->rate_limit_us = policy->transition_delay_us;
536 tunables->rate_limit_us = LATENCY_MULTIPLIER;
537 lat = policy->cpuinfo.transition_latency / NSEC_PER_USEC;
539 tunables->rate_limit_us *= lat;
542 policy->governor_data = sg_policy;
543 sg_policy->tunables = tunables;
545 ret = kobject_init_and_add(&tunables->attr_set.kobj, &sugov_tunables_ktype,
546 get_governor_parent_kobj(policy), "%s",
552 mutex_unlock(&global_tunables_lock);
556 policy->governor_data = NULL;
557 sugov_tunables_free(tunables);
560 sugov_kthread_stop(sg_policy);
563 mutex_unlock(&global_tunables_lock);
565 sugov_policy_free(sg_policy);
568 cpufreq_disable_fast_switch(policy);
570 pr_err("initialization failed (error %d)\n", ret);
574 static void sugov_exit(struct cpufreq_policy *policy)
576 struct sugov_policy *sg_policy = policy->governor_data;
577 struct sugov_tunables *tunables = sg_policy->tunables;
580 mutex_lock(&global_tunables_lock);
582 count = gov_attr_set_put(&tunables->attr_set, &sg_policy->tunables_hook);
583 policy->governor_data = NULL;
585 sugov_tunables_free(tunables);
587 mutex_unlock(&global_tunables_lock);
589 sugov_kthread_stop(sg_policy);
590 sugov_policy_free(sg_policy);
591 cpufreq_disable_fast_switch(policy);
594 static int sugov_start(struct cpufreq_policy *policy)
596 struct sugov_policy *sg_policy = policy->governor_data;
599 sg_policy->freq_update_delay_ns = sg_policy->tunables->rate_limit_us * NSEC_PER_USEC;
600 sg_policy->last_freq_update_time = 0;
601 sg_policy->next_freq = UINT_MAX;
602 sg_policy->work_in_progress = false;
603 sg_policy->need_freq_update = false;
604 sg_policy->cached_raw_freq = 0;
606 for_each_cpu(cpu, policy->cpus) {
607 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
609 memset(sg_cpu, 0, sizeof(*sg_cpu));
610 sg_cpu->sg_policy = sg_policy;
611 sg_cpu->flags = SCHED_CPUFREQ_RT;
612 sg_cpu->iowait_boost_max = policy->cpuinfo.max_freq;
613 cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util,
614 policy_is_shared(policy) ?
615 sugov_update_shared :
616 sugov_update_single);
621 static void sugov_stop(struct cpufreq_policy *policy)
623 struct sugov_policy *sg_policy = policy->governor_data;
626 for_each_cpu(cpu, policy->cpus)
627 cpufreq_remove_update_util_hook(cpu);
631 if (!policy->fast_switch_enabled) {
632 irq_work_sync(&sg_policy->irq_work);
633 kthread_cancel_work_sync(&sg_policy->work);
637 static void sugov_limits(struct cpufreq_policy *policy)
639 struct sugov_policy *sg_policy = policy->governor_data;
641 if (!policy->fast_switch_enabled) {
642 mutex_lock(&sg_policy->work_lock);
643 cpufreq_policy_apply_limits(policy);
644 mutex_unlock(&sg_policy->work_lock);
647 sg_policy->need_freq_update = true;
650 static struct cpufreq_governor schedutil_gov = {
652 .owner = THIS_MODULE,
655 .start = sugov_start,
657 .limits = sugov_limits,
660 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL
661 struct cpufreq_governor *cpufreq_default_governor(void)
663 return &schedutil_gov;
667 static int __init sugov_register(void)
669 return cpufreq_register_governor(&schedutil_gov);
671 fs_initcall(sugov_register);