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
16 #include <trace/events/power.h>
18 struct sugov_tunables {
19 struct gov_attr_set attr_set;
20 unsigned int rate_limit_us;
24 struct cpufreq_policy *policy;
26 struct sugov_tunables *tunables;
27 struct list_head tunables_hook;
29 raw_spinlock_t update_lock; /* For shared policies */
30 u64 last_freq_update_time;
31 s64 freq_update_delay_ns;
32 unsigned int next_freq;
33 unsigned int cached_raw_freq;
35 /* The next fields are only needed if fast switch cannot be used: */
36 struct irq_work irq_work;
37 struct kthread_work work;
38 struct mutex work_lock;
39 struct kthread_worker worker;
40 struct task_struct *thread;
41 bool work_in_progress;
43 bool need_freq_update;
47 struct update_util_data update_util;
48 struct sugov_policy *sg_policy;
51 bool iowait_boost_pending;
52 unsigned int iowait_boost;
53 unsigned int iowait_boost_max;
56 /* The fields below are only needed when sharing a policy: */
57 unsigned long util_cfs;
58 unsigned long util_dl;
61 /* The field below is for single-CPU policies only: */
62 #ifdef CONFIG_NO_HZ_COMMON
63 unsigned long saved_idle_calls;
67 static DEFINE_PER_CPU(struct sugov_cpu, sugov_cpu);
69 /************************ Governor internals ***********************/
71 static bool sugov_should_update_freq(struct sugov_policy *sg_policy, u64 time)
76 * Since cpufreq_update_util() is called with rq->lock held for
77 * the @target_cpu, our per-CPU data is fully serialized.
79 * However, drivers cannot in general deal with cross-CPU
80 * requests, so while get_next_freq() will work, our
81 * sugov_update_commit() call may not for the fast switching platforms.
83 * Hence stop here for remote requests if they aren't supported
84 * by the hardware, as calculating the frequency is pointless if
85 * we cannot in fact act on it.
87 * For the slow switching platforms, the kthread is always scheduled on
88 * the right set of CPUs and any CPU can find the next frequency and
89 * schedule the kthread.
91 if (sg_policy->policy->fast_switch_enabled &&
92 !cpufreq_can_do_remote_dvfs(sg_policy->policy))
95 if (sg_policy->work_in_progress)
98 if (unlikely(sg_policy->need_freq_update)) {
99 sg_policy->need_freq_update = false;
101 * This happens when limits change, so forget the previous
102 * next_freq value and force an update.
104 sg_policy->next_freq = UINT_MAX;
108 delta_ns = time - sg_policy->last_freq_update_time;
110 return delta_ns >= sg_policy->freq_update_delay_ns;
113 static void sugov_update_commit(struct sugov_policy *sg_policy, u64 time,
114 unsigned int next_freq)
116 struct cpufreq_policy *policy = sg_policy->policy;
118 if (sg_policy->next_freq == next_freq)
121 sg_policy->next_freq = next_freq;
122 sg_policy->last_freq_update_time = time;
124 if (policy->fast_switch_enabled) {
125 next_freq = cpufreq_driver_fast_switch(policy, next_freq);
129 policy->cur = next_freq;
130 trace_cpu_frequency(next_freq, smp_processor_id());
132 sg_policy->work_in_progress = true;
133 irq_work_queue(&sg_policy->irq_work);
138 * get_next_freq - Compute a new frequency for a given cpufreq policy.
139 * @sg_policy: schedutil policy object to compute the new frequency for.
140 * @util: Current CPU utilization.
141 * @max: CPU capacity.
143 * If the utilization is frequency-invariant, choose the new frequency to be
144 * proportional to it, that is
146 * next_freq = C * max_freq * util / max
148 * Otherwise, approximate the would-be frequency-invariant utilization by
149 * util_raw * (curr_freq / max_freq) which leads to
151 * next_freq = C * curr_freq * util_raw / max
153 * Take C = 1.25 for the frequency tipping point at (util / max) = 0.8.
155 * The lowest driver-supported frequency which is equal or greater than the raw
156 * next_freq (as calculated above) is returned, subject to policy min/max and
157 * cpufreq driver limitations.
159 static unsigned int get_next_freq(struct sugov_policy *sg_policy,
160 unsigned long util, unsigned long max)
162 struct cpufreq_policy *policy = sg_policy->policy;
163 unsigned int freq = arch_scale_freq_invariant() ?
164 policy->cpuinfo.max_freq : policy->cur;
166 freq = (freq + (freq >> 2)) * util / max;
168 if (freq == sg_policy->cached_raw_freq && sg_policy->next_freq != UINT_MAX)
169 return sg_policy->next_freq;
170 sg_policy->cached_raw_freq = freq;
171 return cpufreq_driver_resolve_freq(policy, freq);
174 static void sugov_get_util(struct sugov_cpu *sg_cpu)
176 struct rq *rq = cpu_rq(sg_cpu->cpu);
178 sg_cpu->max = arch_scale_cpu_capacity(NULL, sg_cpu->cpu);
179 sg_cpu->util_cfs = cpu_util_cfs(rq);
180 sg_cpu->util_dl = cpu_util_dl(rq);
183 static unsigned long sugov_aggregate_util(struct sugov_cpu *sg_cpu)
185 struct rq *rq = cpu_rq(sg_cpu->cpu);
188 if (rq->rt.rt_nr_running) {
191 util = sg_cpu->util_dl;
192 if (rq->cfs.h_nr_running)
193 util += sg_cpu->util_cfs;
197 * Ideally we would like to set util_dl as min/guaranteed freq and
198 * util_cfs + util_dl as requested freq. However, cpufreq is not yet
199 * ready for such an interface. So, we only do the latter for now.
201 return min(util, sg_cpu->max);
204 static void sugov_set_iowait_boost(struct sugov_cpu *sg_cpu, u64 time, unsigned int flags)
206 if (flags & SCHED_CPUFREQ_IOWAIT) {
207 if (sg_cpu->iowait_boost_pending)
210 sg_cpu->iowait_boost_pending = true;
212 if (sg_cpu->iowait_boost) {
213 sg_cpu->iowait_boost <<= 1;
214 if (sg_cpu->iowait_boost > sg_cpu->iowait_boost_max)
215 sg_cpu->iowait_boost = sg_cpu->iowait_boost_max;
217 sg_cpu->iowait_boost = sg_cpu->sg_policy->policy->min;
219 } else if (sg_cpu->iowait_boost) {
220 s64 delta_ns = time - sg_cpu->last_update;
222 /* Clear iowait_boost if the CPU apprears to have been idle. */
223 if (delta_ns > TICK_NSEC) {
224 sg_cpu->iowait_boost = 0;
225 sg_cpu->iowait_boost_pending = false;
230 static void sugov_iowait_boost(struct sugov_cpu *sg_cpu, unsigned long *util,
233 unsigned int boost_util, boost_max;
235 if (!sg_cpu->iowait_boost)
238 if (sg_cpu->iowait_boost_pending) {
239 sg_cpu->iowait_boost_pending = false;
241 sg_cpu->iowait_boost >>= 1;
242 if (sg_cpu->iowait_boost < sg_cpu->sg_policy->policy->min) {
243 sg_cpu->iowait_boost = 0;
248 boost_util = sg_cpu->iowait_boost;
249 boost_max = sg_cpu->iowait_boost_max;
251 if (*util * boost_max < *max * boost_util) {
257 #ifdef CONFIG_NO_HZ_COMMON
258 static bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu)
260 unsigned long idle_calls = tick_nohz_get_idle_calls_cpu(sg_cpu->cpu);
261 bool ret = idle_calls == sg_cpu->saved_idle_calls;
263 sg_cpu->saved_idle_calls = idle_calls;
267 static inline bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu) { return false; }
268 #endif /* CONFIG_NO_HZ_COMMON */
271 * Make sugov_should_update_freq() ignore the rate limit when DL
272 * has increased the utilization.
274 static inline void ignore_dl_rate_limit(struct sugov_cpu *sg_cpu, struct sugov_policy *sg_policy)
276 if (cpu_util_dl(cpu_rq(sg_cpu->cpu)) > sg_cpu->util_dl)
277 sg_policy->need_freq_update = true;
280 static void sugov_update_single(struct update_util_data *hook, u64 time,
283 struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
284 struct sugov_policy *sg_policy = sg_cpu->sg_policy;
285 unsigned long util, max;
289 sugov_set_iowait_boost(sg_cpu, time, flags);
290 sg_cpu->last_update = time;
292 ignore_dl_rate_limit(sg_cpu, sg_policy);
294 if (!sugov_should_update_freq(sg_policy, time))
297 busy = sugov_cpu_is_busy(sg_cpu);
299 sugov_get_util(sg_cpu);
301 util = sugov_aggregate_util(sg_cpu);
302 sugov_iowait_boost(sg_cpu, &util, &max);
303 next_f = get_next_freq(sg_policy, util, max);
305 * Do not reduce the frequency if the CPU has not been idle
306 * recently, as the reduction is likely to be premature then.
308 if (busy && next_f < sg_policy->next_freq &&
309 sg_policy->next_freq != UINT_MAX) {
310 next_f = sg_policy->next_freq;
312 /* Reset cached freq as next_freq has changed */
313 sg_policy->cached_raw_freq = 0;
316 sugov_update_commit(sg_policy, time, next_f);
319 static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu, u64 time)
321 struct sugov_policy *sg_policy = sg_cpu->sg_policy;
322 struct cpufreq_policy *policy = sg_policy->policy;
323 unsigned long util = 0, max = 1;
326 for_each_cpu(j, policy->cpus) {
327 struct sugov_cpu *j_sg_cpu = &per_cpu(sugov_cpu, j);
328 unsigned long j_util, j_max;
331 sugov_get_util(j_sg_cpu);
334 * If the CFS CPU utilization was last updated before the
335 * previous frequency update and the time elapsed between the
336 * last update of the CPU utilization and the last frequency
337 * update is long enough, reset iowait_boost and util_cfs, as
338 * they are now probably stale. However, still consider the
339 * CPU contribution if it has some DEADLINE utilization
342 delta_ns = time - j_sg_cpu->last_update;
343 if (delta_ns > TICK_NSEC) {
344 j_sg_cpu->iowait_boost = 0;
345 j_sg_cpu->iowait_boost_pending = false;
348 j_max = j_sg_cpu->max;
349 j_util = sugov_aggregate_util(j_sg_cpu);
350 sugov_iowait_boost(j_sg_cpu, &j_util, &j_max);
351 if (j_util * max > j_max * util) {
357 return get_next_freq(sg_policy, util, max);
361 sugov_update_shared(struct update_util_data *hook, u64 time, unsigned int flags)
363 struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
364 struct sugov_policy *sg_policy = sg_cpu->sg_policy;
367 raw_spin_lock(&sg_policy->update_lock);
369 sugov_set_iowait_boost(sg_cpu, time, flags);
370 sg_cpu->last_update = time;
372 ignore_dl_rate_limit(sg_cpu, sg_policy);
374 if (sugov_should_update_freq(sg_policy, time)) {
375 next_f = sugov_next_freq_shared(sg_cpu, time);
376 sugov_update_commit(sg_policy, time, next_f);
379 raw_spin_unlock(&sg_policy->update_lock);
382 static void sugov_work(struct kthread_work *work)
384 struct sugov_policy *sg_policy = container_of(work, struct sugov_policy, work);
386 mutex_lock(&sg_policy->work_lock);
387 __cpufreq_driver_target(sg_policy->policy, sg_policy->next_freq,
389 mutex_unlock(&sg_policy->work_lock);
391 sg_policy->work_in_progress = false;
394 static void sugov_irq_work(struct irq_work *irq_work)
396 struct sugov_policy *sg_policy;
398 sg_policy = container_of(irq_work, struct sugov_policy, irq_work);
400 kthread_queue_work(&sg_policy->worker, &sg_policy->work);
403 /************************** sysfs interface ************************/
405 static struct sugov_tunables *global_tunables;
406 static DEFINE_MUTEX(global_tunables_lock);
408 static inline struct sugov_tunables *to_sugov_tunables(struct gov_attr_set *attr_set)
410 return container_of(attr_set, struct sugov_tunables, attr_set);
413 static ssize_t rate_limit_us_show(struct gov_attr_set *attr_set, char *buf)
415 struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
417 return sprintf(buf, "%u\n", tunables->rate_limit_us);
421 rate_limit_us_store(struct gov_attr_set *attr_set, const char *buf, size_t count)
423 struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
424 struct sugov_policy *sg_policy;
425 unsigned int rate_limit_us;
427 if (kstrtouint(buf, 10, &rate_limit_us))
430 tunables->rate_limit_us = rate_limit_us;
432 list_for_each_entry(sg_policy, &attr_set->policy_list, tunables_hook)
433 sg_policy->freq_update_delay_ns = rate_limit_us * NSEC_PER_USEC;
438 static struct governor_attr rate_limit_us = __ATTR_RW(rate_limit_us);
440 static struct attribute *sugov_attributes[] = {
445 static struct kobj_type sugov_tunables_ktype = {
446 .default_attrs = sugov_attributes,
447 .sysfs_ops = &governor_sysfs_ops,
450 /********************** cpufreq governor interface *********************/
452 static struct cpufreq_governor schedutil_gov;
454 static struct sugov_policy *sugov_policy_alloc(struct cpufreq_policy *policy)
456 struct sugov_policy *sg_policy;
458 sg_policy = kzalloc(sizeof(*sg_policy), GFP_KERNEL);
462 sg_policy->policy = policy;
463 raw_spin_lock_init(&sg_policy->update_lock);
467 static void sugov_policy_free(struct sugov_policy *sg_policy)
472 static int sugov_kthread_create(struct sugov_policy *sg_policy)
474 struct task_struct *thread;
475 struct sched_attr attr = {
476 .size = sizeof(struct sched_attr),
477 .sched_policy = SCHED_DEADLINE,
478 .sched_flags = SCHED_FLAG_SUGOV,
482 * Fake (unused) bandwidth; workaround to "fix"
483 * priority inheritance.
485 .sched_runtime = 1000000,
486 .sched_deadline = 10000000,
487 .sched_period = 10000000,
489 struct cpufreq_policy *policy = sg_policy->policy;
492 /* kthread only required for slow path */
493 if (policy->fast_switch_enabled)
496 kthread_init_work(&sg_policy->work, sugov_work);
497 kthread_init_worker(&sg_policy->worker);
498 thread = kthread_create(kthread_worker_fn, &sg_policy->worker,
500 cpumask_first(policy->related_cpus));
501 if (IS_ERR(thread)) {
502 pr_err("failed to create sugov thread: %ld\n", PTR_ERR(thread));
503 return PTR_ERR(thread);
506 ret = sched_setattr_nocheck(thread, &attr);
508 kthread_stop(thread);
509 pr_warn("%s: failed to set SCHED_DEADLINE\n", __func__);
513 sg_policy->thread = thread;
515 /* Kthread is bound to all CPUs by default */
516 if (!policy->dvfs_possible_from_any_cpu)
517 kthread_bind_mask(thread, policy->related_cpus);
519 init_irq_work(&sg_policy->irq_work, sugov_irq_work);
520 mutex_init(&sg_policy->work_lock);
522 wake_up_process(thread);
527 static void sugov_kthread_stop(struct sugov_policy *sg_policy)
529 /* kthread only required for slow path */
530 if (sg_policy->policy->fast_switch_enabled)
533 kthread_flush_worker(&sg_policy->worker);
534 kthread_stop(sg_policy->thread);
535 mutex_destroy(&sg_policy->work_lock);
538 static struct sugov_tunables *sugov_tunables_alloc(struct sugov_policy *sg_policy)
540 struct sugov_tunables *tunables;
542 tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
544 gov_attr_set_init(&tunables->attr_set, &sg_policy->tunables_hook);
545 if (!have_governor_per_policy())
546 global_tunables = tunables;
551 static void sugov_tunables_free(struct sugov_tunables *tunables)
553 if (!have_governor_per_policy())
554 global_tunables = NULL;
559 static int sugov_init(struct cpufreq_policy *policy)
561 struct sugov_policy *sg_policy;
562 struct sugov_tunables *tunables;
565 /* State should be equivalent to EXIT */
566 if (policy->governor_data)
569 cpufreq_enable_fast_switch(policy);
571 sg_policy = sugov_policy_alloc(policy);
574 goto disable_fast_switch;
577 ret = sugov_kthread_create(sg_policy);
581 mutex_lock(&global_tunables_lock);
583 if (global_tunables) {
584 if (WARN_ON(have_governor_per_policy())) {
588 policy->governor_data = sg_policy;
589 sg_policy->tunables = global_tunables;
591 gov_attr_set_get(&global_tunables->attr_set, &sg_policy->tunables_hook);
595 tunables = sugov_tunables_alloc(sg_policy);
601 tunables->rate_limit_us = cpufreq_policy_transition_delay_us(policy);
603 policy->governor_data = sg_policy;
604 sg_policy->tunables = tunables;
606 ret = kobject_init_and_add(&tunables->attr_set.kobj, &sugov_tunables_ktype,
607 get_governor_parent_kobj(policy), "%s",
613 mutex_unlock(&global_tunables_lock);
617 policy->governor_data = NULL;
618 sugov_tunables_free(tunables);
621 sugov_kthread_stop(sg_policy);
622 mutex_unlock(&global_tunables_lock);
625 sugov_policy_free(sg_policy);
628 cpufreq_disable_fast_switch(policy);
630 pr_err("initialization failed (error %d)\n", ret);
634 static void sugov_exit(struct cpufreq_policy *policy)
636 struct sugov_policy *sg_policy = policy->governor_data;
637 struct sugov_tunables *tunables = sg_policy->tunables;
640 mutex_lock(&global_tunables_lock);
642 count = gov_attr_set_put(&tunables->attr_set, &sg_policy->tunables_hook);
643 policy->governor_data = NULL;
645 sugov_tunables_free(tunables);
647 mutex_unlock(&global_tunables_lock);
649 sugov_kthread_stop(sg_policy);
650 sugov_policy_free(sg_policy);
651 cpufreq_disable_fast_switch(policy);
654 static int sugov_start(struct cpufreq_policy *policy)
656 struct sugov_policy *sg_policy = policy->governor_data;
659 sg_policy->freq_update_delay_ns = sg_policy->tunables->rate_limit_us * NSEC_PER_USEC;
660 sg_policy->last_freq_update_time = 0;
661 sg_policy->next_freq = UINT_MAX;
662 sg_policy->work_in_progress = false;
663 sg_policy->need_freq_update = false;
664 sg_policy->cached_raw_freq = 0;
666 for_each_cpu(cpu, policy->cpus) {
667 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
669 memset(sg_cpu, 0, sizeof(*sg_cpu));
671 sg_cpu->sg_policy = sg_policy;
672 sg_cpu->iowait_boost_max = policy->cpuinfo.max_freq;
675 for_each_cpu(cpu, policy->cpus) {
676 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
678 cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util,
679 policy_is_shared(policy) ?
680 sugov_update_shared :
681 sugov_update_single);
686 static void sugov_stop(struct cpufreq_policy *policy)
688 struct sugov_policy *sg_policy = policy->governor_data;
691 for_each_cpu(cpu, policy->cpus)
692 cpufreq_remove_update_util_hook(cpu);
696 if (!policy->fast_switch_enabled) {
697 irq_work_sync(&sg_policy->irq_work);
698 kthread_cancel_work_sync(&sg_policy->work);
702 static void sugov_limits(struct cpufreq_policy *policy)
704 struct sugov_policy *sg_policy = policy->governor_data;
706 if (!policy->fast_switch_enabled) {
707 mutex_lock(&sg_policy->work_lock);
708 cpufreq_policy_apply_limits(policy);
709 mutex_unlock(&sg_policy->work_lock);
712 sg_policy->need_freq_update = true;
715 static struct cpufreq_governor schedutil_gov = {
717 .owner = THIS_MODULE,
718 .dynamic_switching = true,
721 .start = sugov_start,
723 .limits = sugov_limits,
726 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL
727 struct cpufreq_governor *cpufreq_default_governor(void)
729 return &schedutil_gov;
733 static int __init sugov_register(void)
735 return cpufreq_register_governor(&schedutil_gov);
737 fs_initcall(sugov_register);