1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * CPUFreq governor based on scheduler-provided CPU utilization data.
4  *
5  * Copyright (C) 2016, Intel Corporation
6  * Author: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
7  */
8 
9 #define IOWAIT_BOOST_MIN	(SCHED_CAPACITY_SCALE / 8)
10 
11 struct sugov_tunables {
12 	struct gov_attr_set	attr_set;
13 	unsigned int		rate_limit_us;
14 };
15 
16 struct sugov_policy {
17 	struct cpufreq_policy	*policy;
18 
19 	struct sugov_tunables	*tunables;
20 	struct list_head	tunables_hook;
21 
22 	raw_spinlock_t		update_lock;
23 	u64			last_freq_update_time;
24 	s64			freq_update_delay_ns;
25 	unsigned int		next_freq;
26 	unsigned int		cached_raw_freq;
27 
28 	/* max CPU capacity, which is equal for all CPUs in freq. domain */
29 	unsigned long		max;
30 
31 	/* The next fields are only needed if fast switch cannot be used: */
32 	struct			irq_work irq_work;
33 	struct			kthread_work work;
34 	struct			mutex work_lock;
35 	struct			kthread_worker worker;
36 	struct task_struct	*thread;
37 	bool			work_in_progress;
38 
39 	bool			limits_changed;
40 	bool			need_freq_update;
41 };
42 
43 struct sugov_cpu {
44 	struct update_util_data	update_util;
45 	struct sugov_policy	*sg_policy;
46 	unsigned int		cpu;
47 
48 	bool			iowait_boost_pending;
49 	unsigned int		iowait_boost;
50 	u64			last_update;
51 
52 	unsigned long		util;
53 	unsigned long		bw_dl;
54 
55 	/* The field below is for single-CPU policies only: */
56 #ifdef CONFIG_NO_HZ_COMMON
57 	unsigned long		saved_idle_calls;
58 #endif
59 };
60 
61 static DEFINE_PER_CPU(struct sugov_cpu, sugov_cpu);
62 
63 /************************ Governor internals ***********************/
64 
65 static bool sugov_should_update_freq(struct sugov_policy *sg_policy, u64 time)
66 {
67 	s64 delta_ns;
68 
69 	/*
70 	 * Since cpufreq_update_util() is called with rq->lock held for
71 	 * the @target_cpu, our per-CPU data is fully serialized.
72 	 *
73 	 * However, drivers cannot in general deal with cross-CPU
74 	 * requests, so while get_next_freq() will work, our
75 	 * sugov_update_commit() call may not for the fast switching platforms.
76 	 *
77 	 * Hence stop here for remote requests if they aren't supported
78 	 * by the hardware, as calculating the frequency is pointless if
79 	 * we cannot in fact act on it.
80 	 *
81 	 * This is needed on the slow switching platforms too to prevent CPUs
82 	 * going offline from leaving stale IRQ work items behind.
83 	 */
84 	if (!cpufreq_this_cpu_can_update(sg_policy->policy))
85 		return false;
86 
87 	if (unlikely(sg_policy->limits_changed)) {
88 		sg_policy->limits_changed = false;
89 		sg_policy->need_freq_update = true;
90 		return true;
91 	}
92 
93 	delta_ns = time - sg_policy->last_freq_update_time;
94 
95 	return delta_ns >= sg_policy->freq_update_delay_ns;
96 }
97 
98 static bool sugov_update_next_freq(struct sugov_policy *sg_policy, u64 time,
99 				   unsigned int next_freq)
100 {
101 	if (sg_policy->need_freq_update)
102 		sg_policy->need_freq_update = cpufreq_driver_test_flags(CPUFREQ_NEED_UPDATE_LIMITS);
103 	else if (sg_policy->next_freq == next_freq)
104 		return false;
105 
106 	sg_policy->next_freq = next_freq;
107 	sg_policy->last_freq_update_time = time;
108 
109 	return true;
110 }
111 
112 static void sugov_deferred_update(struct sugov_policy *sg_policy)
113 {
114 	if (!sg_policy->work_in_progress) {
115 		sg_policy->work_in_progress = true;
116 		irq_work_queue(&sg_policy->irq_work);
117 	}
118 }
119 
120 /**
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.
125  *
126  * If the utilization is frequency-invariant, choose the new frequency to be
127  * proportional to it, that is
128  *
129  * next_freq = C * max_freq * util / max
130  *
131  * Otherwise, approximate the would-be frequency-invariant utilization by
132  * util_raw * (curr_freq / max_freq) which leads to
133  *
134  * next_freq = C * curr_freq * util_raw / max
135  *
136  * Take C = 1.25 for the frequency tipping point at (util / max) = 0.8.
137  *
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.
141  */
142 static unsigned int get_next_freq(struct sugov_policy *sg_policy,
143 				  unsigned long util, unsigned long max)
144 {
145 	struct cpufreq_policy *policy = sg_policy->policy;
146 	unsigned int freq = arch_scale_freq_invariant() ?
147 				policy->cpuinfo.max_freq : policy->cur;
148 
149 	util = map_util_perf(util);
150 	freq = map_util_freq(util, freq, max);
151 
152 	if (freq == sg_policy->cached_raw_freq && !sg_policy->need_freq_update)
153 		return sg_policy->next_freq;
154 
155 	sg_policy->cached_raw_freq = freq;
156 	return cpufreq_driver_resolve_freq(policy, freq);
157 }
158 
159 static void sugov_get_util(struct sugov_cpu *sg_cpu)
160 {
161 	struct rq *rq = cpu_rq(sg_cpu->cpu);
162 
163 	sg_cpu->bw_dl = cpu_bw_dl(rq);
164 	sg_cpu->util = effective_cpu_util(sg_cpu->cpu, cpu_util_cfs(sg_cpu->cpu),
165 					  FREQUENCY_UTIL, NULL);
166 }
167 
168 /**
169  * sugov_iowait_reset() - Reset the IO boost status of a CPU.
170  * @sg_cpu: the sugov data for the CPU to boost
171  * @time: the update time from the caller
172  * @set_iowait_boost: true if an IO boost has been requested
173  *
174  * The IO wait boost of a task is disabled after a tick since the last update
175  * of a CPU. If a new IO wait boost is requested after more then a tick, then
176  * we enable the boost starting from IOWAIT_BOOST_MIN, which improves energy
177  * efficiency by ignoring sporadic wakeups from IO.
178  */
179 static bool sugov_iowait_reset(struct sugov_cpu *sg_cpu, u64 time,
180 			       bool set_iowait_boost)
181 {
182 	s64 delta_ns = time - sg_cpu->last_update;
183 
184 	/* Reset boost only if a tick has elapsed since last request */
185 	if (delta_ns <= TICK_NSEC)
186 		return false;
187 
188 	sg_cpu->iowait_boost = set_iowait_boost ? IOWAIT_BOOST_MIN : 0;
189 	sg_cpu->iowait_boost_pending = set_iowait_boost;
190 
191 	return true;
192 }
193 
194 /**
195  * sugov_iowait_boost() - Updates the IO boost status of a CPU.
196  * @sg_cpu: the sugov data for the CPU to boost
197  * @time: the update time from the caller
198  * @flags: SCHED_CPUFREQ_IOWAIT if the task is waking up after an IO wait
199  *
200  * Each time a task wakes up after an IO operation, the CPU utilization can be
201  * boosted to a certain utilization which doubles at each "frequent and
202  * successive" wakeup from IO, ranging from IOWAIT_BOOST_MIN to the utilization
203  * of the maximum OPP.
204  *
205  * To keep doubling, an IO boost has to be requested at least once per tick,
206  * otherwise we restart from the utilization of the minimum OPP.
207  */
208 static void sugov_iowait_boost(struct sugov_cpu *sg_cpu, u64 time,
209 			       unsigned int flags)
210 {
211 	bool set_iowait_boost = flags & SCHED_CPUFREQ_IOWAIT;
212 
213 	/* Reset boost if the CPU appears to have been idle enough */
214 	if (sg_cpu->iowait_boost &&
215 	    sugov_iowait_reset(sg_cpu, time, set_iowait_boost))
216 		return;
217 
218 	/* Boost only tasks waking up after IO */
219 	if (!set_iowait_boost)
220 		return;
221 
222 	/* Ensure boost doubles only one time at each request */
223 	if (sg_cpu->iowait_boost_pending)
224 		return;
225 	sg_cpu->iowait_boost_pending = true;
226 
227 	/* Double the boost at each request */
228 	if (sg_cpu->iowait_boost) {
229 		sg_cpu->iowait_boost =
230 			min_t(unsigned int, sg_cpu->iowait_boost << 1, SCHED_CAPACITY_SCALE);
231 		return;
232 	}
233 
234 	/* First wakeup after IO: start with minimum boost */
235 	sg_cpu->iowait_boost = IOWAIT_BOOST_MIN;
236 }
237 
238 /**
239  * sugov_iowait_apply() - Apply the IO boost to a CPU.
240  * @sg_cpu: the sugov data for the cpu to boost
241  * @time: the update time from the caller
242  *
243  * A CPU running a task which woken up after an IO operation can have its
244  * utilization boosted to speed up the completion of those IO operations.
245  * The IO boost value is increased each time a task wakes up from IO, in
246  * sugov_iowait_apply(), and it's instead decreased by this function,
247  * each time an increase has not been requested (!iowait_boost_pending).
248  *
249  * A CPU which also appears to have been idle for at least one tick has also
250  * its IO boost utilization reset.
251  *
252  * This mechanism is designed to boost high frequently IO waiting tasks, while
253  * being more conservative on tasks which does sporadic IO operations.
254  */
255 static void sugov_iowait_apply(struct sugov_cpu *sg_cpu, u64 time)
256 {
257 	struct sugov_policy *sg_policy = sg_cpu->sg_policy;
258 	unsigned long boost;
259 
260 	/* No boost currently required */
261 	if (!sg_cpu->iowait_boost)
262 		return;
263 
264 	/* Reset boost if the CPU appears to have been idle enough */
265 	if (sugov_iowait_reset(sg_cpu, time, false))
266 		return;
267 
268 	if (!sg_cpu->iowait_boost_pending) {
269 		/*
270 		 * No boost pending; reduce the boost value.
271 		 */
272 		sg_cpu->iowait_boost >>= 1;
273 		if (sg_cpu->iowait_boost < IOWAIT_BOOST_MIN) {
274 			sg_cpu->iowait_boost = 0;
275 			return;
276 		}
277 	}
278 
279 	sg_cpu->iowait_boost_pending = false;
280 
281 	/*
282 	 * sg_cpu->util is already in capacity scale; convert iowait_boost
283 	 * into the same scale so we can compare.
284 	 */
285 	boost = sg_cpu->iowait_boost * sg_policy->max;
286 	boost >>= SCHED_CAPACITY_SHIFT;
287 	boost = uclamp_rq_util_with(cpu_rq(sg_cpu->cpu), boost, NULL);
288 	if (sg_cpu->util < boost)
289 		sg_cpu->util = boost;
290 }
291 
292 #ifdef CONFIG_NO_HZ_COMMON
293 static bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu)
294 {
295 	unsigned long idle_calls = tick_nohz_get_idle_calls_cpu(sg_cpu->cpu);
296 	bool ret = idle_calls == sg_cpu->saved_idle_calls;
297 
298 	sg_cpu->saved_idle_calls = idle_calls;
299 	return ret;
300 }
301 #else
302 static inline bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu) { return false; }
303 #endif /* CONFIG_NO_HZ_COMMON */
304 
305 /*
306  * Make sugov_should_update_freq() ignore the rate limit when DL
307  * has increased the utilization.
308  */
309 static inline void ignore_dl_rate_limit(struct sugov_cpu *sg_cpu)
310 {
311 	if (cpu_bw_dl(cpu_rq(sg_cpu->cpu)) > sg_cpu->bw_dl)
312 		sg_cpu->sg_policy->limits_changed = true;
313 }
314 
315 static inline bool sugov_update_single_common(struct sugov_cpu *sg_cpu,
316 					      u64 time, unsigned int flags)
317 {
318 	sugov_iowait_boost(sg_cpu, time, flags);
319 	sg_cpu->last_update = time;
320 
321 	ignore_dl_rate_limit(sg_cpu);
322 
323 	if (!sugov_should_update_freq(sg_cpu->sg_policy, time))
324 		return false;
325 
326 	sugov_get_util(sg_cpu);
327 	sugov_iowait_apply(sg_cpu, time);
328 
329 	return true;
330 }
331 
332 static void sugov_update_single_freq(struct update_util_data *hook, u64 time,
333 				     unsigned int flags)
334 {
335 	struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
336 	struct sugov_policy *sg_policy = sg_cpu->sg_policy;
337 	unsigned int cached_freq = sg_policy->cached_raw_freq;
338 	unsigned int next_f;
339 
340 	if (!sugov_update_single_common(sg_cpu, time, flags))
341 		return;
342 
343 	next_f = get_next_freq(sg_policy, sg_cpu->util, sg_policy->max);
344 	/*
345 	 * Do not reduce the frequency if the CPU has not been idle
346 	 * recently, as the reduction is likely to be premature then.
347 	 *
348 	 * Except when the rq is capped by uclamp_max.
349 	 */
350 	if (!uclamp_rq_is_capped(cpu_rq(sg_cpu->cpu)) &&
351 	    sugov_cpu_is_busy(sg_cpu) && next_f < sg_policy->next_freq) {
352 		next_f = sg_policy->next_freq;
353 
354 		/* Restore cached freq as next_freq has changed */
355 		sg_policy->cached_raw_freq = cached_freq;
356 	}
357 
358 	if (!sugov_update_next_freq(sg_policy, time, next_f))
359 		return;
360 
361 	/*
362 	 * This code runs under rq->lock for the target CPU, so it won't run
363 	 * concurrently on two different CPUs for the same target and it is not
364 	 * necessary to acquire the lock in the fast switch case.
365 	 */
366 	if (sg_policy->policy->fast_switch_enabled) {
367 		cpufreq_driver_fast_switch(sg_policy->policy, next_f);
368 	} else {
369 		raw_spin_lock(&sg_policy->update_lock);
370 		sugov_deferred_update(sg_policy);
371 		raw_spin_unlock(&sg_policy->update_lock);
372 	}
373 }
374 
375 static void sugov_update_single_perf(struct update_util_data *hook, u64 time,
376 				     unsigned int flags)
377 {
378 	struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
379 	struct sugov_policy *sg_policy = sg_cpu->sg_policy;
380 	unsigned long prev_util = sg_cpu->util;
381 
382 	/*
383 	 * Fall back to the "frequency" path if frequency invariance is not
384 	 * supported, because the direct mapping between the utilization and
385 	 * the performance levels depends on the frequency invariance.
386 	 */
387 	if (!arch_scale_freq_invariant()) {
388 		sugov_update_single_freq(hook, time, flags);
389 		return;
390 	}
391 
392 	if (!sugov_update_single_common(sg_cpu, time, flags))
393 		return;
394 
395 	/*
396 	 * Do not reduce the target performance level if the CPU has not been
397 	 * idle recently, as the reduction is likely to be premature then.
398 	 *
399 	 * Except when the rq is capped by uclamp_max.
400 	 */
401 	if (!uclamp_rq_is_capped(cpu_rq(sg_cpu->cpu)) &&
402 	    sugov_cpu_is_busy(sg_cpu) && sg_cpu->util < prev_util)
403 		sg_cpu->util = prev_util;
404 
405 	cpufreq_driver_adjust_perf(sg_cpu->cpu, map_util_perf(sg_cpu->bw_dl),
406 				   map_util_perf(sg_cpu->util),
407 				   sg_policy->max);
408 
409 	sg_cpu->sg_policy->last_freq_update_time = time;
410 }
411 
412 static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu, u64 time)
413 {
414 	struct sugov_policy *sg_policy = sg_cpu->sg_policy;
415 	struct cpufreq_policy *policy = sg_policy->policy;
416 	unsigned long util = 0;
417 	unsigned int j;
418 
419 	for_each_cpu(j, policy->cpus) {
420 		struct sugov_cpu *j_sg_cpu = &per_cpu(sugov_cpu, j);
421 
422 		sugov_get_util(j_sg_cpu);
423 		sugov_iowait_apply(j_sg_cpu, time);
424 
425 		util = max(j_sg_cpu->util, util);
426 	}
427 
428 	return get_next_freq(sg_policy, util, sg_policy->max);
429 }
430 
431 static void
432 sugov_update_shared(struct update_util_data *hook, u64 time, unsigned int flags)
433 {
434 	struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
435 	struct sugov_policy *sg_policy = sg_cpu->sg_policy;
436 	unsigned int next_f;
437 
438 	raw_spin_lock(&sg_policy->update_lock);
439 
440 	sugov_iowait_boost(sg_cpu, time, flags);
441 	sg_cpu->last_update = time;
442 
443 	ignore_dl_rate_limit(sg_cpu);
444 
445 	if (sugov_should_update_freq(sg_policy, time)) {
446 		next_f = sugov_next_freq_shared(sg_cpu, time);
447 
448 		if (!sugov_update_next_freq(sg_policy, time, next_f))
449 			goto unlock;
450 
451 		if (sg_policy->policy->fast_switch_enabled)
452 			cpufreq_driver_fast_switch(sg_policy->policy, next_f);
453 		else
454 			sugov_deferred_update(sg_policy);
455 	}
456 unlock:
457 	raw_spin_unlock(&sg_policy->update_lock);
458 }
459 
460 static void sugov_work(struct kthread_work *work)
461 {
462 	struct sugov_policy *sg_policy = container_of(work, struct sugov_policy, work);
463 	unsigned int freq;
464 	unsigned long flags;
465 
466 	/*
467 	 * Hold sg_policy->update_lock shortly to handle the case where:
468 	 * in case sg_policy->next_freq is read here, and then updated by
469 	 * sugov_deferred_update() just before work_in_progress is set to false
470 	 * here, we may miss queueing the new update.
471 	 *
472 	 * Note: If a work was queued after the update_lock is released,
473 	 * sugov_work() will just be called again by kthread_work code; and the
474 	 * request will be proceed before the sugov thread sleeps.
475 	 */
476 	raw_spin_lock_irqsave(&sg_policy->update_lock, flags);
477 	freq = sg_policy->next_freq;
478 	sg_policy->work_in_progress = false;
479 	raw_spin_unlock_irqrestore(&sg_policy->update_lock, flags);
480 
481 	mutex_lock(&sg_policy->work_lock);
482 	__cpufreq_driver_target(sg_policy->policy, freq, CPUFREQ_RELATION_L);
483 	mutex_unlock(&sg_policy->work_lock);
484 }
485 
486 static void sugov_irq_work(struct irq_work *irq_work)
487 {
488 	struct sugov_policy *sg_policy;
489 
490 	sg_policy = container_of(irq_work, struct sugov_policy, irq_work);
491 
492 	kthread_queue_work(&sg_policy->worker, &sg_policy->work);
493 }
494 
495 /************************** sysfs interface ************************/
496 
497 static struct sugov_tunables *global_tunables;
498 static DEFINE_MUTEX(global_tunables_lock);
499 
500 static inline struct sugov_tunables *to_sugov_tunables(struct gov_attr_set *attr_set)
501 {
502 	return container_of(attr_set, struct sugov_tunables, attr_set);
503 }
504 
505 static ssize_t rate_limit_us_show(struct gov_attr_set *attr_set, char *buf)
506 {
507 	struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
508 
509 	return sprintf(buf, "%u\n", tunables->rate_limit_us);
510 }
511 
512 static ssize_t
513 rate_limit_us_store(struct gov_attr_set *attr_set, const char *buf, size_t count)
514 {
515 	struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
516 	struct sugov_policy *sg_policy;
517 	unsigned int rate_limit_us;
518 
519 	if (kstrtouint(buf, 10, &rate_limit_us))
520 		return -EINVAL;
521 
522 	tunables->rate_limit_us = rate_limit_us;
523 
524 	list_for_each_entry(sg_policy, &attr_set->policy_list, tunables_hook)
525 		sg_policy->freq_update_delay_ns = rate_limit_us * NSEC_PER_USEC;
526 
527 	return count;
528 }
529 
530 static struct governor_attr rate_limit_us = __ATTR_RW(rate_limit_us);
531 
532 static struct attribute *sugov_attrs[] = {
533 	&rate_limit_us.attr,
534 	NULL
535 };
536 ATTRIBUTE_GROUPS(sugov);
537 
538 static void sugov_tunables_free(struct kobject *kobj)
539 {
540 	struct gov_attr_set *attr_set = to_gov_attr_set(kobj);
541 
542 	kfree(to_sugov_tunables(attr_set));
543 }
544 
545 static struct kobj_type sugov_tunables_ktype = {
546 	.default_groups = sugov_groups,
547 	.sysfs_ops = &governor_sysfs_ops,
548 	.release = &sugov_tunables_free,
549 };
550 
551 /********************** cpufreq governor interface *********************/
552 
553 struct cpufreq_governor schedutil_gov;
554 
555 static struct sugov_policy *sugov_policy_alloc(struct cpufreq_policy *policy)
556 {
557 	struct sugov_policy *sg_policy;
558 
559 	sg_policy = kzalloc(sizeof(*sg_policy), GFP_KERNEL);
560 	if (!sg_policy)
561 		return NULL;
562 
563 	sg_policy->policy = policy;
564 	raw_spin_lock_init(&sg_policy->update_lock);
565 	return sg_policy;
566 }
567 
568 static void sugov_policy_free(struct sugov_policy *sg_policy)
569 {
570 	kfree(sg_policy);
571 }
572 
573 static int sugov_kthread_create(struct sugov_policy *sg_policy)
574 {
575 	struct task_struct *thread;
576 	struct sched_attr attr = {
577 		.size		= sizeof(struct sched_attr),
578 		.sched_policy	= SCHED_DEADLINE,
579 		.sched_flags	= SCHED_FLAG_SUGOV,
580 		.sched_nice	= 0,
581 		.sched_priority	= 0,
582 		/*
583 		 * Fake (unused) bandwidth; workaround to "fix"
584 		 * priority inheritance.
585 		 */
586 		.sched_runtime	=  1000000,
587 		.sched_deadline = 10000000,
588 		.sched_period	= 10000000,
589 	};
590 	struct cpufreq_policy *policy = sg_policy->policy;
591 	int ret;
592 
593 	/* kthread only required for slow path */
594 	if (policy->fast_switch_enabled)
595 		return 0;
596 
597 	kthread_init_work(&sg_policy->work, sugov_work);
598 	kthread_init_worker(&sg_policy->worker);
599 	thread = kthread_create(kthread_worker_fn, &sg_policy->worker,
600 				"sugov:%d",
601 				cpumask_first(policy->related_cpus));
602 	if (IS_ERR(thread)) {
603 		pr_err("failed to create sugov thread: %ld\n", PTR_ERR(thread));
604 		return PTR_ERR(thread);
605 	}
606 
607 	ret = sched_setattr_nocheck(thread, &attr);
608 	if (ret) {
609 		kthread_stop(thread);
610 		pr_warn("%s: failed to set SCHED_DEADLINE\n", __func__);
611 		return ret;
612 	}
613 
614 	sg_policy->thread = thread;
615 	kthread_bind_mask(thread, policy->related_cpus);
616 	init_irq_work(&sg_policy->irq_work, sugov_irq_work);
617 	mutex_init(&sg_policy->work_lock);
618 
619 	wake_up_process(thread);
620 
621 	return 0;
622 }
623 
624 static void sugov_kthread_stop(struct sugov_policy *sg_policy)
625 {
626 	/* kthread only required for slow path */
627 	if (sg_policy->policy->fast_switch_enabled)
628 		return;
629 
630 	kthread_flush_worker(&sg_policy->worker);
631 	kthread_stop(sg_policy->thread);
632 	mutex_destroy(&sg_policy->work_lock);
633 }
634 
635 static struct sugov_tunables *sugov_tunables_alloc(struct sugov_policy *sg_policy)
636 {
637 	struct sugov_tunables *tunables;
638 
639 	tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
640 	if (tunables) {
641 		gov_attr_set_init(&tunables->attr_set, &sg_policy->tunables_hook);
642 		if (!have_governor_per_policy())
643 			global_tunables = tunables;
644 	}
645 	return tunables;
646 }
647 
648 static void sugov_clear_global_tunables(void)
649 {
650 	if (!have_governor_per_policy())
651 		global_tunables = NULL;
652 }
653 
654 static int sugov_init(struct cpufreq_policy *policy)
655 {
656 	struct sugov_policy *sg_policy;
657 	struct sugov_tunables *tunables;
658 	int ret = 0;
659 
660 	/* State should be equivalent to EXIT */
661 	if (policy->governor_data)
662 		return -EBUSY;
663 
664 	cpufreq_enable_fast_switch(policy);
665 
666 	sg_policy = sugov_policy_alloc(policy);
667 	if (!sg_policy) {
668 		ret = -ENOMEM;
669 		goto disable_fast_switch;
670 	}
671 
672 	ret = sugov_kthread_create(sg_policy);
673 	if (ret)
674 		goto free_sg_policy;
675 
676 	mutex_lock(&global_tunables_lock);
677 
678 	if (global_tunables) {
679 		if (WARN_ON(have_governor_per_policy())) {
680 			ret = -EINVAL;
681 			goto stop_kthread;
682 		}
683 		policy->governor_data = sg_policy;
684 		sg_policy->tunables = global_tunables;
685 
686 		gov_attr_set_get(&global_tunables->attr_set, &sg_policy->tunables_hook);
687 		goto out;
688 	}
689 
690 	tunables = sugov_tunables_alloc(sg_policy);
691 	if (!tunables) {
692 		ret = -ENOMEM;
693 		goto stop_kthread;
694 	}
695 
696 	tunables->rate_limit_us = cpufreq_policy_transition_delay_us(policy);
697 
698 	policy->governor_data = sg_policy;
699 	sg_policy->tunables = tunables;
700 
701 	ret = kobject_init_and_add(&tunables->attr_set.kobj, &sugov_tunables_ktype,
702 				   get_governor_parent_kobj(policy), "%s",
703 				   schedutil_gov.name);
704 	if (ret)
705 		goto fail;
706 
707 out:
708 	mutex_unlock(&global_tunables_lock);
709 	return 0;
710 
711 fail:
712 	kobject_put(&tunables->attr_set.kobj);
713 	policy->governor_data = NULL;
714 	sugov_clear_global_tunables();
715 
716 stop_kthread:
717 	sugov_kthread_stop(sg_policy);
718 	mutex_unlock(&global_tunables_lock);
719 
720 free_sg_policy:
721 	sugov_policy_free(sg_policy);
722 
723 disable_fast_switch:
724 	cpufreq_disable_fast_switch(policy);
725 
726 	pr_err("initialization failed (error %d)\n", ret);
727 	return ret;
728 }
729 
730 static void sugov_exit(struct cpufreq_policy *policy)
731 {
732 	struct sugov_policy *sg_policy = policy->governor_data;
733 	struct sugov_tunables *tunables = sg_policy->tunables;
734 	unsigned int count;
735 
736 	mutex_lock(&global_tunables_lock);
737 
738 	count = gov_attr_set_put(&tunables->attr_set, &sg_policy->tunables_hook);
739 	policy->governor_data = NULL;
740 	if (!count)
741 		sugov_clear_global_tunables();
742 
743 	mutex_unlock(&global_tunables_lock);
744 
745 	sugov_kthread_stop(sg_policy);
746 	sugov_policy_free(sg_policy);
747 	cpufreq_disable_fast_switch(policy);
748 }
749 
750 static int sugov_start(struct cpufreq_policy *policy)
751 {
752 	struct sugov_policy *sg_policy = policy->governor_data;
753 	void (*uu)(struct update_util_data *data, u64 time, unsigned int flags);
754 	unsigned int cpu = cpumask_first(policy->cpus);
755 
756 	sg_policy->freq_update_delay_ns	= sg_policy->tunables->rate_limit_us * NSEC_PER_USEC;
757 	sg_policy->last_freq_update_time	= 0;
758 	sg_policy->next_freq			= 0;
759 	sg_policy->work_in_progress		= false;
760 	sg_policy->limits_changed		= false;
761 	sg_policy->cached_raw_freq		= 0;
762 	sg_policy->max				= arch_scale_cpu_capacity(cpu);
763 
764 	sg_policy->need_freq_update = cpufreq_driver_test_flags(CPUFREQ_NEED_UPDATE_LIMITS);
765 
766 	for_each_cpu(cpu, policy->cpus) {
767 		struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
768 
769 		memset(sg_cpu, 0, sizeof(*sg_cpu));
770 		sg_cpu->cpu			= cpu;
771 		sg_cpu->sg_policy		= sg_policy;
772 	}
773 
774 	if (policy_is_shared(policy))
775 		uu = sugov_update_shared;
776 	else if (policy->fast_switch_enabled && cpufreq_driver_has_adjust_perf())
777 		uu = sugov_update_single_perf;
778 	else
779 		uu = sugov_update_single_freq;
780 
781 	for_each_cpu(cpu, policy->cpus) {
782 		struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
783 
784 		cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util, uu);
785 	}
786 	return 0;
787 }
788 
789 static void sugov_stop(struct cpufreq_policy *policy)
790 {
791 	struct sugov_policy *sg_policy = policy->governor_data;
792 	unsigned int cpu;
793 
794 	for_each_cpu(cpu, policy->cpus)
795 		cpufreq_remove_update_util_hook(cpu);
796 
797 	synchronize_rcu();
798 
799 	if (!policy->fast_switch_enabled) {
800 		irq_work_sync(&sg_policy->irq_work);
801 		kthread_cancel_work_sync(&sg_policy->work);
802 	}
803 }
804 
805 static void sugov_limits(struct cpufreq_policy *policy)
806 {
807 	struct sugov_policy *sg_policy = policy->governor_data;
808 
809 	if (!policy->fast_switch_enabled) {
810 		mutex_lock(&sg_policy->work_lock);
811 		cpufreq_policy_apply_limits(policy);
812 		mutex_unlock(&sg_policy->work_lock);
813 	}
814 
815 	sg_policy->limits_changed = true;
816 }
817 
818 struct cpufreq_governor schedutil_gov = {
819 	.name			= "schedutil",
820 	.owner			= THIS_MODULE,
821 	.flags			= CPUFREQ_GOV_DYNAMIC_SWITCHING,
822 	.init			= sugov_init,
823 	.exit			= sugov_exit,
824 	.start			= sugov_start,
825 	.stop			= sugov_stop,
826 	.limits			= sugov_limits,
827 };
828 
829 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL
830 struct cpufreq_governor *cpufreq_default_governor(void)
831 {
832 	return &schedutil_gov;
833 }
834 #endif
835 
836 cpufreq_governor_init(schedutil_gov);
837 
838 #ifdef CONFIG_ENERGY_MODEL
839 static void rebuild_sd_workfn(struct work_struct *work)
840 {
841 	rebuild_sched_domains_energy();
842 }
843 static DECLARE_WORK(rebuild_sd_work, rebuild_sd_workfn);
844 
845 /*
846  * EAS shouldn't be attempted without sugov, so rebuild the sched_domains
847  * on governor changes to make sure the scheduler knows about it.
848  */
849 void sched_cpufreq_governor_change(struct cpufreq_policy *policy,
850 				  struct cpufreq_governor *old_gov)
851 {
852 	if (old_gov == &schedutil_gov || policy->governor == &schedutil_gov) {
853 		/*
854 		 * When called from the cpufreq_register_driver() path, the
855 		 * cpu_hotplug_lock is already held, so use a work item to
856 		 * avoid nested locking in rebuild_sched_domains().
857 		 */
858 		schedule_work(&rebuild_sd_work);
859 	}
860 
861 }
862 #endif
863