1 // SPDX-License-Identifier: GPL-2.0
3 * CPUFreq governor based on scheduler-provided CPU utilization data.
5 * Copyright (C) 2016, Intel Corporation
6 * Author: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
13 #include <linux/sched/cpufreq.h>
14 #include <trace/events/power.h>
16 #define IOWAIT_BOOST_MIN (SCHED_CAPACITY_SCALE / 8)
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;
58 /* The field below is for single-CPU policies only: */
59 #ifdef CONFIG_NO_HZ_COMMON
60 unsigned long saved_idle_calls;
64 static DEFINE_PER_CPU(struct sugov_cpu, sugov_cpu);
66 /************************ Governor internals ***********************/
68 static bool sugov_should_update_freq(struct sugov_policy *sg_policy, u64 time)
73 * Since cpufreq_update_util() is called with rq->lock held for
74 * the @target_cpu, our per-CPU data is fully serialized.
76 * However, drivers cannot in general deal with cross-CPU
77 * requests, so while get_next_freq() will work, our
78 * sugov_update_commit() call may not for the fast switching platforms.
80 * Hence stop here for remote requests if they aren't supported
81 * by the hardware, as calculating the frequency is pointless if
82 * we cannot in fact act on it.
84 * For the slow switching platforms, the kthread is always scheduled on
85 * the right set of CPUs and any CPU can find the next frequency and
86 * schedule the kthread.
88 if (sg_policy->policy->fast_switch_enabled &&
89 !cpufreq_this_cpu_can_update(sg_policy->policy))
92 if (unlikely(sg_policy->need_freq_update))
95 delta_ns = time - sg_policy->last_freq_update_time;
97 return delta_ns >= sg_policy->freq_update_delay_ns;
100 static bool sugov_update_next_freq(struct sugov_policy *sg_policy, u64 time,
101 unsigned int next_freq)
103 if (sg_policy->next_freq == next_freq)
106 sg_policy->next_freq = next_freq;
107 sg_policy->last_freq_update_time = time;
112 static void sugov_fast_switch(struct sugov_policy *sg_policy, u64 time,
113 unsigned int next_freq)
115 struct cpufreq_policy *policy = sg_policy->policy;
117 if (!sugov_update_next_freq(sg_policy, time, next_freq))
120 next_freq = cpufreq_driver_fast_switch(policy, next_freq);
124 policy->cur = next_freq;
125 trace_cpu_frequency(next_freq, smp_processor_id());
128 static void sugov_deferred_update(struct sugov_policy *sg_policy, u64 time,
129 unsigned int next_freq)
131 if (!sugov_update_next_freq(sg_policy, time, next_freq))
134 if (!sg_policy->work_in_progress) {
135 sg_policy->work_in_progress = true;
136 irq_work_queue(&sg_policy->irq_work);
141 * get_next_freq - Compute a new frequency for a given cpufreq policy.
142 * @sg_policy: schedutil policy object to compute the new frequency for.
143 * @util: Current CPU utilization.
144 * @max: CPU capacity.
146 * If the utilization is frequency-invariant, choose the new frequency to be
147 * proportional to it, that is
149 * next_freq = C * max_freq * util / max
151 * Otherwise, approximate the would-be frequency-invariant utilization by
152 * util_raw * (curr_freq / max_freq) which leads to
154 * next_freq = C * curr_freq * util_raw / max
156 * Take C = 1.25 for the frequency tipping point at (util / max) = 0.8.
158 * The lowest driver-supported frequency which is equal or greater than the raw
159 * next_freq (as calculated above) is returned, subject to policy min/max and
160 * cpufreq driver limitations.
162 static unsigned int get_next_freq(struct sugov_policy *sg_policy,
163 unsigned long util, unsigned long max)
165 struct cpufreq_policy *policy = sg_policy->policy;
166 unsigned int freq = arch_scale_freq_invariant() ?
167 policy->cpuinfo.max_freq : policy->cur;
169 freq = map_util_freq(util, freq, max);
171 if (freq == sg_policy->cached_raw_freq && !sg_policy->need_freq_update)
172 return sg_policy->next_freq;
174 sg_policy->need_freq_update = false;
175 sg_policy->cached_raw_freq = freq;
176 return cpufreq_driver_resolve_freq(policy, freq);
180 * This function computes an effective utilization for the given CPU, to be
181 * used for frequency selection given the linear relation: f = u * f_max.
183 * The scheduler tracks the following metrics:
185 * cpu_util_{cfs,rt,dl,irq}()
188 * Where the cfs,rt and dl util numbers are tracked with the same metric and
189 * synchronized windows and are thus directly comparable.
191 * The cfs,rt,dl utilization are the running times measured with rq->clock_task
192 * which excludes things like IRQ and steal-time. These latter are then accrued
193 * in the irq utilization.
195 * The DL bandwidth number otoh is not a measured metric but a value computed
196 * based on the task model parameters and gives the minimal utilization
197 * required to meet deadlines.
199 unsigned long schedutil_freq_util(int cpu, unsigned long util_cfs,
200 unsigned long max, enum schedutil_type type)
202 unsigned long dl_util, util, irq;
203 struct rq *rq = cpu_rq(cpu);
205 if (type == FREQUENCY_UTIL && rt_rq_is_runnable(&rq->rt))
209 * Early check to see if IRQ/steal time saturates the CPU, can be
210 * because of inaccuracies in how we track these -- see
211 * update_irq_load_avg().
213 irq = cpu_util_irq(rq);
214 if (unlikely(irq >= max))
218 * Because the time spend on RT/DL tasks is visible as 'lost' time to
219 * CFS tasks and we use the same metric to track the effective
220 * utilization (PELT windows are synchronized) we can directly add them
221 * to obtain the CPU's actual utilization.
224 util += cpu_util_rt(rq);
226 dl_util = cpu_util_dl(rq);
229 * For frequency selection we do not make cpu_util_dl() a permanent part
230 * of this sum because we want to use cpu_bw_dl() later on, but we need
231 * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
232 * that we select f_max when there is no idle time.
234 * NOTE: numerical errors or stop class might cause us to not quite hit
235 * saturation when we should -- something for later.
237 if (util + dl_util >= max)
241 * OTOH, for energy computation we need the estimated running time, so
242 * include util_dl and ignore dl_bw.
244 if (type == ENERGY_UTIL)
248 * There is still idle time; further improve the number by using the
249 * irq metric. Because IRQ/steal time is hidden from the task clock we
250 * need to scale the task numbers:
253 * U' = irq + ------- * U
256 util = scale_irq_capacity(util, irq, max);
260 * Bandwidth required by DEADLINE must always be granted while, for
261 * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
262 * to gracefully reduce the frequency when no tasks show up for longer
265 * Ideally we would like to set bw_dl as min/guaranteed freq and util +
266 * bw_dl as requested freq. However, cpufreq is not yet ready for such
267 * an interface. So, we only do the latter for now.
269 if (type == FREQUENCY_UTIL)
270 util += cpu_bw_dl(rq);
272 return min(max, util);
275 static unsigned long sugov_get_util(struct sugov_cpu *sg_cpu)
277 struct rq *rq = cpu_rq(sg_cpu->cpu);
278 unsigned long util = cpu_util_cfs(rq);
279 unsigned long max = arch_scale_cpu_capacity(NULL, sg_cpu->cpu);
282 sg_cpu->bw_dl = cpu_bw_dl(rq);
284 return schedutil_freq_util(sg_cpu->cpu, util, max, FREQUENCY_UTIL);
288 * sugov_iowait_reset() - Reset the IO boost status of a CPU.
289 * @sg_cpu: the sugov data for the CPU to boost
290 * @time: the update time from the caller
291 * @set_iowait_boost: true if an IO boost has been requested
293 * The IO wait boost of a task is disabled after a tick since the last update
294 * of a CPU. If a new IO wait boost is requested after more then a tick, then
295 * we enable the boost starting from IOWAIT_BOOST_MIN, which improves energy
296 * efficiency by ignoring sporadic wakeups from IO.
298 static bool sugov_iowait_reset(struct sugov_cpu *sg_cpu, u64 time,
299 bool set_iowait_boost)
301 s64 delta_ns = time - sg_cpu->last_update;
303 /* Reset boost only if a tick has elapsed since last request */
304 if (delta_ns <= TICK_NSEC)
307 sg_cpu->iowait_boost = set_iowait_boost ? IOWAIT_BOOST_MIN : 0;
308 sg_cpu->iowait_boost_pending = set_iowait_boost;
314 * sugov_iowait_boost() - Updates the IO boost status of a CPU.
315 * @sg_cpu: the sugov data for the CPU to boost
316 * @time: the update time from the caller
317 * @flags: SCHED_CPUFREQ_IOWAIT if the task is waking up after an IO wait
319 * Each time a task wakes up after an IO operation, the CPU utilization can be
320 * boosted to a certain utilization which doubles at each "frequent and
321 * successive" wakeup from IO, ranging from IOWAIT_BOOST_MIN to the utilization
322 * of the maximum OPP.
324 * To keep doubling, an IO boost has to be requested at least once per tick,
325 * otherwise we restart from the utilization of the minimum OPP.
327 static void sugov_iowait_boost(struct sugov_cpu *sg_cpu, u64 time,
330 bool set_iowait_boost = flags & SCHED_CPUFREQ_IOWAIT;
332 /* Reset boost if the CPU appears to have been idle enough */
333 if (sg_cpu->iowait_boost &&
334 sugov_iowait_reset(sg_cpu, time, set_iowait_boost))
337 /* Boost only tasks waking up after IO */
338 if (!set_iowait_boost)
341 /* Ensure boost doubles only one time at each request */
342 if (sg_cpu->iowait_boost_pending)
344 sg_cpu->iowait_boost_pending = true;
346 /* Double the boost at each request */
347 if (sg_cpu->iowait_boost) {
348 sg_cpu->iowait_boost =
349 min_t(unsigned int, sg_cpu->iowait_boost << 1, SCHED_CAPACITY_SCALE);
353 /* First wakeup after IO: start with minimum boost */
354 sg_cpu->iowait_boost = IOWAIT_BOOST_MIN;
358 * sugov_iowait_apply() - Apply the IO boost to a CPU.
359 * @sg_cpu: the sugov data for the cpu to boost
360 * @time: the update time from the caller
361 * @util: the utilization to (eventually) boost
362 * @max: the maximum value the utilization can be boosted to
364 * A CPU running a task which woken up after an IO operation can have its
365 * utilization boosted to speed up the completion of those IO operations.
366 * The IO boost value is increased each time a task wakes up from IO, in
367 * sugov_iowait_apply(), and it's instead decreased by this function,
368 * each time an increase has not been requested (!iowait_boost_pending).
370 * A CPU which also appears to have been idle for at least one tick has also
371 * its IO boost utilization reset.
373 * This mechanism is designed to boost high frequently IO waiting tasks, while
374 * being more conservative on tasks which does sporadic IO operations.
376 static unsigned long sugov_iowait_apply(struct sugov_cpu *sg_cpu, u64 time,
377 unsigned long util, unsigned long max)
381 /* No boost currently required */
382 if (!sg_cpu->iowait_boost)
385 /* Reset boost if the CPU appears to have been idle enough */
386 if (sugov_iowait_reset(sg_cpu, time, false))
389 if (!sg_cpu->iowait_boost_pending) {
391 * No boost pending; reduce the boost value.
393 sg_cpu->iowait_boost >>= 1;
394 if (sg_cpu->iowait_boost < IOWAIT_BOOST_MIN) {
395 sg_cpu->iowait_boost = 0;
400 sg_cpu->iowait_boost_pending = false;
403 * @util is already in capacity scale; convert iowait_boost
404 * into the same scale so we can compare.
406 boost = (sg_cpu->iowait_boost * max) >> SCHED_CAPACITY_SHIFT;
407 return max(boost, util);
410 #ifdef CONFIG_NO_HZ_COMMON
411 static bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu)
413 unsigned long idle_calls = tick_nohz_get_idle_calls_cpu(sg_cpu->cpu);
414 bool ret = idle_calls == sg_cpu->saved_idle_calls;
416 sg_cpu->saved_idle_calls = idle_calls;
420 static inline bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu) { return false; }
421 #endif /* CONFIG_NO_HZ_COMMON */
424 * Make sugov_should_update_freq() ignore the rate limit when DL
425 * has increased the utilization.
427 static inline void ignore_dl_rate_limit(struct sugov_cpu *sg_cpu, struct sugov_policy *sg_policy)
429 if (cpu_bw_dl(cpu_rq(sg_cpu->cpu)) > sg_cpu->bw_dl)
430 sg_policy->need_freq_update = true;
433 static void sugov_update_single(struct update_util_data *hook, u64 time,
436 struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
437 struct sugov_policy *sg_policy = sg_cpu->sg_policy;
438 unsigned long util, max;
442 sugov_iowait_boost(sg_cpu, time, flags);
443 sg_cpu->last_update = time;
445 ignore_dl_rate_limit(sg_cpu, sg_policy);
447 if (!sugov_should_update_freq(sg_policy, time))
450 busy = sugov_cpu_is_busy(sg_cpu);
452 util = sugov_get_util(sg_cpu);
454 util = sugov_iowait_apply(sg_cpu, time, util, max);
455 next_f = get_next_freq(sg_policy, util, max);
457 * Do not reduce the frequency if the CPU has not been idle
458 * recently, as the reduction is likely to be premature then.
460 if (busy && next_f < sg_policy->next_freq) {
461 next_f = sg_policy->next_freq;
463 /* Reset cached freq as next_freq has changed */
464 sg_policy->cached_raw_freq = 0;
468 * This code runs under rq->lock for the target CPU, so it won't run
469 * concurrently on two different CPUs for the same target and it is not
470 * necessary to acquire the lock in the fast switch case.
472 if (sg_policy->policy->fast_switch_enabled) {
473 sugov_fast_switch(sg_policy, time, next_f);
475 raw_spin_lock(&sg_policy->update_lock);
476 sugov_deferred_update(sg_policy, time, next_f);
477 raw_spin_unlock(&sg_policy->update_lock);
481 static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu, u64 time)
483 struct sugov_policy *sg_policy = sg_cpu->sg_policy;
484 struct cpufreq_policy *policy = sg_policy->policy;
485 unsigned long util = 0, max = 1;
488 for_each_cpu(j, policy->cpus) {
489 struct sugov_cpu *j_sg_cpu = &per_cpu(sugov_cpu, j);
490 unsigned long j_util, j_max;
492 j_util = sugov_get_util(j_sg_cpu);
493 j_max = j_sg_cpu->max;
494 j_util = sugov_iowait_apply(j_sg_cpu, time, j_util, j_max);
496 if (j_util * max > j_max * util) {
502 return get_next_freq(sg_policy, util, max);
506 sugov_update_shared(struct update_util_data *hook, u64 time, unsigned int flags)
508 struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
509 struct sugov_policy *sg_policy = sg_cpu->sg_policy;
512 raw_spin_lock(&sg_policy->update_lock);
514 sugov_iowait_boost(sg_cpu, time, flags);
515 sg_cpu->last_update = time;
517 ignore_dl_rate_limit(sg_cpu, sg_policy);
519 if (sugov_should_update_freq(sg_policy, time)) {
520 next_f = sugov_next_freq_shared(sg_cpu, time);
522 if (sg_policy->policy->fast_switch_enabled)
523 sugov_fast_switch(sg_policy, time, next_f);
525 sugov_deferred_update(sg_policy, time, next_f);
528 raw_spin_unlock(&sg_policy->update_lock);
531 static void sugov_work(struct kthread_work *work)
533 struct sugov_policy *sg_policy = container_of(work, struct sugov_policy, work);
538 * Hold sg_policy->update_lock shortly to handle the case where:
539 * incase sg_policy->next_freq is read here, and then updated by
540 * sugov_deferred_update() just before work_in_progress is set to false
541 * here, we may miss queueing the new update.
543 * Note: If a work was queued after the update_lock is released,
544 * sugov_work() will just be called again by kthread_work code; and the
545 * request will be proceed before the sugov thread sleeps.
547 raw_spin_lock_irqsave(&sg_policy->update_lock, flags);
548 freq = sg_policy->next_freq;
549 sg_policy->work_in_progress = false;
550 raw_spin_unlock_irqrestore(&sg_policy->update_lock, flags);
552 mutex_lock(&sg_policy->work_lock);
553 __cpufreq_driver_target(sg_policy->policy, freq, CPUFREQ_RELATION_L);
554 mutex_unlock(&sg_policy->work_lock);
557 static void sugov_irq_work(struct irq_work *irq_work)
559 struct sugov_policy *sg_policy;
561 sg_policy = container_of(irq_work, struct sugov_policy, irq_work);
563 kthread_queue_work(&sg_policy->worker, &sg_policy->work);
566 /************************** sysfs interface ************************/
568 static struct sugov_tunables *global_tunables;
569 static DEFINE_MUTEX(global_tunables_lock);
571 static inline struct sugov_tunables *to_sugov_tunables(struct gov_attr_set *attr_set)
573 return container_of(attr_set, struct sugov_tunables, attr_set);
576 static ssize_t rate_limit_us_show(struct gov_attr_set *attr_set, char *buf)
578 struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
580 return sprintf(buf, "%u\n", tunables->rate_limit_us);
584 rate_limit_us_store(struct gov_attr_set *attr_set, const char *buf, size_t count)
586 struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
587 struct sugov_policy *sg_policy;
588 unsigned int rate_limit_us;
590 if (kstrtouint(buf, 10, &rate_limit_us))
593 tunables->rate_limit_us = rate_limit_us;
595 list_for_each_entry(sg_policy, &attr_set->policy_list, tunables_hook)
596 sg_policy->freq_update_delay_ns = rate_limit_us * NSEC_PER_USEC;
601 static struct governor_attr rate_limit_us = __ATTR_RW(rate_limit_us);
603 static struct attribute *sugov_attributes[] = {
608 static struct kobj_type sugov_tunables_ktype = {
609 .default_attrs = sugov_attributes,
610 .sysfs_ops = &governor_sysfs_ops,
613 /********************** cpufreq governor interface *********************/
615 struct cpufreq_governor schedutil_gov;
617 static struct sugov_policy *sugov_policy_alloc(struct cpufreq_policy *policy)
619 struct sugov_policy *sg_policy;
621 sg_policy = kzalloc(sizeof(*sg_policy), GFP_KERNEL);
625 sg_policy->policy = policy;
626 raw_spin_lock_init(&sg_policy->update_lock);
630 static void sugov_policy_free(struct sugov_policy *sg_policy)
635 static int sugov_kthread_create(struct sugov_policy *sg_policy)
637 struct task_struct *thread;
638 struct sched_attr attr = {
639 .size = sizeof(struct sched_attr),
640 .sched_policy = SCHED_DEADLINE,
641 .sched_flags = SCHED_FLAG_SUGOV,
645 * Fake (unused) bandwidth; workaround to "fix"
646 * priority inheritance.
648 .sched_runtime = 1000000,
649 .sched_deadline = 10000000,
650 .sched_period = 10000000,
652 struct cpufreq_policy *policy = sg_policy->policy;
655 /* kthread only required for slow path */
656 if (policy->fast_switch_enabled)
659 kthread_init_work(&sg_policy->work, sugov_work);
660 kthread_init_worker(&sg_policy->worker);
661 thread = kthread_create(kthread_worker_fn, &sg_policy->worker,
663 cpumask_first(policy->related_cpus));
664 if (IS_ERR(thread)) {
665 pr_err("failed to create sugov thread: %ld\n", PTR_ERR(thread));
666 return PTR_ERR(thread);
669 ret = sched_setattr_nocheck(thread, &attr);
671 kthread_stop(thread);
672 pr_warn("%s: failed to set SCHED_DEADLINE\n", __func__);
676 sg_policy->thread = thread;
677 kthread_bind_mask(thread, policy->related_cpus);
678 init_irq_work(&sg_policy->irq_work, sugov_irq_work);
679 mutex_init(&sg_policy->work_lock);
681 wake_up_process(thread);
686 static void sugov_kthread_stop(struct sugov_policy *sg_policy)
688 /* kthread only required for slow path */
689 if (sg_policy->policy->fast_switch_enabled)
692 kthread_flush_worker(&sg_policy->worker);
693 kthread_stop(sg_policy->thread);
694 mutex_destroy(&sg_policy->work_lock);
697 static struct sugov_tunables *sugov_tunables_alloc(struct sugov_policy *sg_policy)
699 struct sugov_tunables *tunables;
701 tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
703 gov_attr_set_init(&tunables->attr_set, &sg_policy->tunables_hook);
704 if (!have_governor_per_policy())
705 global_tunables = tunables;
710 static void sugov_tunables_free(struct sugov_tunables *tunables)
712 if (!have_governor_per_policy())
713 global_tunables = NULL;
718 static int sugov_init(struct cpufreq_policy *policy)
720 struct sugov_policy *sg_policy;
721 struct sugov_tunables *tunables;
724 /* State should be equivalent to EXIT */
725 if (policy->governor_data)
728 cpufreq_enable_fast_switch(policy);
730 sg_policy = sugov_policy_alloc(policy);
733 goto disable_fast_switch;
736 ret = sugov_kthread_create(sg_policy);
740 mutex_lock(&global_tunables_lock);
742 if (global_tunables) {
743 if (WARN_ON(have_governor_per_policy())) {
747 policy->governor_data = sg_policy;
748 sg_policy->tunables = global_tunables;
750 gov_attr_set_get(&global_tunables->attr_set, &sg_policy->tunables_hook);
754 tunables = sugov_tunables_alloc(sg_policy);
760 tunables->rate_limit_us = cpufreq_policy_transition_delay_us(policy);
762 policy->governor_data = sg_policy;
763 sg_policy->tunables = tunables;
765 ret = kobject_init_and_add(&tunables->attr_set.kobj, &sugov_tunables_ktype,
766 get_governor_parent_kobj(policy), "%s",
772 mutex_unlock(&global_tunables_lock);
776 kobject_put(&tunables->attr_set.kobj);
777 policy->governor_data = NULL;
778 sugov_tunables_free(tunables);
781 sugov_kthread_stop(sg_policy);
782 mutex_unlock(&global_tunables_lock);
785 sugov_policy_free(sg_policy);
788 cpufreq_disable_fast_switch(policy);
790 pr_err("initialization failed (error %d)\n", ret);
794 static void sugov_exit(struct cpufreq_policy *policy)
796 struct sugov_policy *sg_policy = policy->governor_data;
797 struct sugov_tunables *tunables = sg_policy->tunables;
800 mutex_lock(&global_tunables_lock);
802 count = gov_attr_set_put(&tunables->attr_set, &sg_policy->tunables_hook);
803 policy->governor_data = NULL;
805 sugov_tunables_free(tunables);
807 mutex_unlock(&global_tunables_lock);
809 sugov_kthread_stop(sg_policy);
810 sugov_policy_free(sg_policy);
811 cpufreq_disable_fast_switch(policy);
814 static int sugov_start(struct cpufreq_policy *policy)
816 struct sugov_policy *sg_policy = policy->governor_data;
819 sg_policy->freq_update_delay_ns = sg_policy->tunables->rate_limit_us * NSEC_PER_USEC;
820 sg_policy->last_freq_update_time = 0;
821 sg_policy->next_freq = 0;
822 sg_policy->work_in_progress = false;
823 sg_policy->need_freq_update = false;
824 sg_policy->cached_raw_freq = 0;
826 for_each_cpu(cpu, policy->cpus) {
827 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
829 memset(sg_cpu, 0, sizeof(*sg_cpu));
831 sg_cpu->sg_policy = sg_policy;
834 for_each_cpu(cpu, policy->cpus) {
835 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
837 cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util,
838 policy_is_shared(policy) ?
839 sugov_update_shared :
840 sugov_update_single);
845 static void sugov_stop(struct cpufreq_policy *policy)
847 struct sugov_policy *sg_policy = policy->governor_data;
850 for_each_cpu(cpu, policy->cpus)
851 cpufreq_remove_update_util_hook(cpu);
855 if (!policy->fast_switch_enabled) {
856 irq_work_sync(&sg_policy->irq_work);
857 kthread_cancel_work_sync(&sg_policy->work);
861 static void sugov_limits(struct cpufreq_policy *policy)
863 struct sugov_policy *sg_policy = policy->governor_data;
865 if (!policy->fast_switch_enabled) {
866 mutex_lock(&sg_policy->work_lock);
867 cpufreq_policy_apply_limits(policy);
868 mutex_unlock(&sg_policy->work_lock);
871 sg_policy->need_freq_update = true;
874 struct cpufreq_governor schedutil_gov = {
876 .owner = THIS_MODULE,
877 .dynamic_switching = true,
880 .start = sugov_start,
882 .limits = sugov_limits,
885 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL
886 struct cpufreq_governor *cpufreq_default_governor(void)
888 return &schedutil_gov;
892 static int __init sugov_register(void)
894 return cpufreq_register_governor(&schedutil_gov);
896 fs_initcall(sugov_register);
898 #ifdef CONFIG_ENERGY_MODEL
899 extern bool sched_energy_update;
900 extern struct mutex sched_energy_mutex;
902 static void rebuild_sd_workfn(struct work_struct *work)
904 mutex_lock(&sched_energy_mutex);
905 sched_energy_update = true;
906 rebuild_sched_domains();
907 sched_energy_update = false;
908 mutex_unlock(&sched_energy_mutex);
910 static DECLARE_WORK(rebuild_sd_work, rebuild_sd_workfn);
913 * EAS shouldn't be attempted without sugov, so rebuild the sched_domains
914 * on governor changes to make sure the scheduler knows about it.
916 void sched_cpufreq_governor_change(struct cpufreq_policy *policy,
917 struct cpufreq_governor *old_gov)
919 if (old_gov == &schedutil_gov || policy->governor == &schedutil_gov) {
921 * When called from the cpufreq_register_driver() path, the
922 * cpu_hotplug_lock is already held, so use a work item to
923 * avoid nested locking in rebuild_sched_domains().
925 schedule_work(&rebuild_sd_work);