1 /*
2  * CPUFreq governor based on scheduler-provided CPU utilization data.
3  *
4  * Copyright (C) 2016, Intel Corporation
5  * Author: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
6  *
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.
10  */
11 
12 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
13 
14 #include "sched.h"
15 
16 #include <trace/events/power.h>
17 
18 struct sugov_tunables {
19 	struct gov_attr_set	attr_set;
20 	unsigned int		rate_limit_us;
21 };
22 
23 struct sugov_policy {
24 	struct cpufreq_policy	*policy;
25 
26 	struct sugov_tunables	*tunables;
27 	struct list_head	tunables_hook;
28 
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;
34 
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;
42 
43 	bool			need_freq_update;
44 };
45 
46 struct sugov_cpu {
47 	struct update_util_data	update_util;
48 	struct sugov_policy	*sg_policy;
49 	unsigned int		cpu;
50 
51 	bool			iowait_boost_pending;
52 	unsigned int		iowait_boost;
53 	unsigned int		iowait_boost_max;
54 	u64			last_update;
55 
56 	/* The fields below are only needed when sharing a policy: */
57 	unsigned long		util_cfs;
58 	unsigned long		util_dl;
59 	unsigned long		max;
60 
61 	/* The field below is for single-CPU policies only: */
62 #ifdef CONFIG_NO_HZ_COMMON
63 	unsigned long		saved_idle_calls;
64 #endif
65 };
66 
67 static DEFINE_PER_CPU(struct sugov_cpu, sugov_cpu);
68 
69 /************************ Governor internals ***********************/
70 
71 static bool sugov_should_update_freq(struct sugov_policy *sg_policy, u64 time)
72 {
73 	s64 delta_ns;
74 
75 	/*
76 	 * Since cpufreq_update_util() is called with rq->lock held for
77 	 * the @target_cpu, our per-CPU data is fully serialized.
78 	 *
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.
82 	 *
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.
86 	 *
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.
90 	 */
91 	if (sg_policy->policy->fast_switch_enabled &&
92 	    !cpufreq_this_cpu_can_update(sg_policy->policy))
93 		return false;
94 
95 	if (unlikely(sg_policy->need_freq_update))
96 		return true;
97 
98 	delta_ns = time - sg_policy->last_freq_update_time;
99 
100 	return delta_ns >= sg_policy->freq_update_delay_ns;
101 }
102 
103 static bool sugov_update_next_freq(struct sugov_policy *sg_policy, u64 time,
104 				   unsigned int next_freq)
105 {
106 	if (sg_policy->next_freq == next_freq)
107 		return false;
108 
109 	sg_policy->next_freq = next_freq;
110 	sg_policy->last_freq_update_time = time;
111 
112 	return true;
113 }
114 
115 static void sugov_fast_switch(struct sugov_policy *sg_policy, u64 time,
116 			      unsigned int next_freq)
117 {
118 	struct cpufreq_policy *policy = sg_policy->policy;
119 
120 	if (!sugov_update_next_freq(sg_policy, time, next_freq))
121 		return;
122 
123 	next_freq = cpufreq_driver_fast_switch(policy, next_freq);
124 	if (!next_freq)
125 		return;
126 
127 	policy->cur = next_freq;
128 	trace_cpu_frequency(next_freq, smp_processor_id());
129 }
130 
131 static void sugov_deferred_update(struct sugov_policy *sg_policy, u64 time,
132 				  unsigned int next_freq)
133 {
134 	if (!sugov_update_next_freq(sg_policy, time, next_freq))
135 		return;
136 
137 	if (!sg_policy->work_in_progress) {
138 		sg_policy->work_in_progress = true;
139 		irq_work_queue(&sg_policy->irq_work);
140 	}
141 }
142 
143 /**
144  * get_next_freq - Compute a new frequency for a given cpufreq policy.
145  * @sg_policy: schedutil policy object to compute the new frequency for.
146  * @util: Current CPU utilization.
147  * @max: CPU capacity.
148  *
149  * If the utilization is frequency-invariant, choose the new frequency to be
150  * proportional to it, that is
151  *
152  * next_freq = C * max_freq * util / max
153  *
154  * Otherwise, approximate the would-be frequency-invariant utilization by
155  * util_raw * (curr_freq / max_freq) which leads to
156  *
157  * next_freq = C * curr_freq * util_raw / max
158  *
159  * Take C = 1.25 for the frequency tipping point at (util / max) = 0.8.
160  *
161  * The lowest driver-supported frequency which is equal or greater than the raw
162  * next_freq (as calculated above) is returned, subject to policy min/max and
163  * cpufreq driver limitations.
164  */
165 static unsigned int get_next_freq(struct sugov_policy *sg_policy,
166 				  unsigned long util, unsigned long max)
167 {
168 	struct cpufreq_policy *policy = sg_policy->policy;
169 	unsigned int freq = arch_scale_freq_invariant() ?
170 				policy->cpuinfo.max_freq : policy->cur;
171 
172 	freq = (freq + (freq >> 2)) * util / max;
173 
174 	if (freq == sg_policy->cached_raw_freq && !sg_policy->need_freq_update)
175 		return sg_policy->next_freq;
176 
177 	sg_policy->need_freq_update = false;
178 	sg_policy->cached_raw_freq = freq;
179 	return cpufreq_driver_resolve_freq(policy, freq);
180 }
181 
182 static void sugov_get_util(struct sugov_cpu *sg_cpu)
183 {
184 	struct rq *rq = cpu_rq(sg_cpu->cpu);
185 
186 	sg_cpu->max = arch_scale_cpu_capacity(NULL, sg_cpu->cpu);
187 	sg_cpu->util_cfs = cpu_util_cfs(rq);
188 	sg_cpu->util_dl  = cpu_util_dl(rq);
189 }
190 
191 static unsigned long sugov_aggregate_util(struct sugov_cpu *sg_cpu)
192 {
193 	struct rq *rq = cpu_rq(sg_cpu->cpu);
194 
195 	if (rq->rt.rt_nr_running)
196 		return sg_cpu->max;
197 
198 	/*
199 	 * Utilization required by DEADLINE must always be granted while, for
200 	 * FAIR, we use blocked utilization of IDLE CPUs as a mechanism to
201 	 * gracefully reduce the frequency when no tasks show up for longer
202 	 * periods of time.
203 	 *
204 	 * Ideally we would like to set util_dl as min/guaranteed freq and
205 	 * util_cfs + util_dl as requested freq. However, cpufreq is not yet
206 	 * ready for such an interface. So, we only do the latter for now.
207 	 */
208 	return min(sg_cpu->max, (sg_cpu->util_dl + sg_cpu->util_cfs));
209 }
210 
211 /**
212  * sugov_iowait_reset() - Reset the IO boost status of a CPU.
213  * @sg_cpu: the sugov data for the CPU to boost
214  * @time: the update time from the caller
215  * @set_iowait_boost: true if an IO boost has been requested
216  *
217  * The IO wait boost of a task is disabled after a tick since the last update
218  * of a CPU. If a new IO wait boost is requested after more then a tick, then
219  * we enable the boost starting from the minimum frequency, which improves
220  * energy efficiency by ignoring sporadic wakeups from IO.
221  */
222 static bool sugov_iowait_reset(struct sugov_cpu *sg_cpu, u64 time,
223 			       bool set_iowait_boost)
224 {
225 	s64 delta_ns = time - sg_cpu->last_update;
226 
227 	/* Reset boost only if a tick has elapsed since last request */
228 	if (delta_ns <= TICK_NSEC)
229 		return false;
230 
231 	sg_cpu->iowait_boost = set_iowait_boost
232 		? sg_cpu->sg_policy->policy->min : 0;
233 	sg_cpu->iowait_boost_pending = set_iowait_boost;
234 
235 	return true;
236 }
237 
238 /**
239  * sugov_iowait_boost() - Updates the IO boost status of a CPU.
240  * @sg_cpu: the sugov data for the CPU to boost
241  * @time: the update time from the caller
242  * @flags: SCHED_CPUFREQ_IOWAIT if the task is waking up after an IO wait
243  *
244  * Each time a task wakes up after an IO operation, the CPU utilization can be
245  * boosted to a certain utilization which doubles at each "frequent and
246  * successive" wakeup from IO, ranging from the utilization of the minimum
247  * OPP to the utilization of the maximum OPP.
248  * To keep doubling, an IO boost has to be requested at least once per tick,
249  * otherwise we restart from the utilization of the minimum OPP.
250  */
251 static void sugov_iowait_boost(struct sugov_cpu *sg_cpu, u64 time,
252 			       unsigned int flags)
253 {
254 	bool set_iowait_boost = flags & SCHED_CPUFREQ_IOWAIT;
255 
256 	/* Reset boost if the CPU appears to have been idle enough */
257 	if (sg_cpu->iowait_boost &&
258 	    sugov_iowait_reset(sg_cpu, time, set_iowait_boost))
259 		return;
260 
261 	/* Boost only tasks waking up after IO */
262 	if (!set_iowait_boost)
263 		return;
264 
265 	/* Ensure boost doubles only one time at each request */
266 	if (sg_cpu->iowait_boost_pending)
267 		return;
268 	sg_cpu->iowait_boost_pending = true;
269 
270 	/* Double the boost at each request */
271 	if (sg_cpu->iowait_boost) {
272 		sg_cpu->iowait_boost <<= 1;
273 		if (sg_cpu->iowait_boost > sg_cpu->iowait_boost_max)
274 			sg_cpu->iowait_boost = sg_cpu->iowait_boost_max;
275 		return;
276 	}
277 
278 	/* First wakeup after IO: start with minimum boost */
279 	sg_cpu->iowait_boost = sg_cpu->sg_policy->policy->min;
280 }
281 
282 /**
283  * sugov_iowait_apply() - Apply the IO boost to a CPU.
284  * @sg_cpu: the sugov data for the cpu to boost
285  * @time: the update time from the caller
286  * @util: the utilization to (eventually) boost
287  * @max: the maximum value the utilization can be boosted to
288  *
289  * A CPU running a task which woken up after an IO operation can have its
290  * utilization boosted to speed up the completion of those IO operations.
291  * The IO boost value is increased each time a task wakes up from IO, in
292  * sugov_iowait_apply(), and it's instead decreased by this function,
293  * each time an increase has not been requested (!iowait_boost_pending).
294  *
295  * A CPU which also appears to have been idle for at least one tick has also
296  * its IO boost utilization reset.
297  *
298  * This mechanism is designed to boost high frequently IO waiting tasks, while
299  * being more conservative on tasks which does sporadic IO operations.
300  */
301 static void sugov_iowait_apply(struct sugov_cpu *sg_cpu, u64 time,
302 			       unsigned long *util, unsigned long *max)
303 {
304 	unsigned int boost_util, boost_max;
305 
306 	/* No boost currently required */
307 	if (!sg_cpu->iowait_boost)
308 		return;
309 
310 	/* Reset boost if the CPU appears to have been idle enough */
311 	if (sugov_iowait_reset(sg_cpu, time, false))
312 		return;
313 
314 	/*
315 	 * An IO waiting task has just woken up:
316 	 * allow to further double the boost value
317 	 */
318 	if (sg_cpu->iowait_boost_pending) {
319 		sg_cpu->iowait_boost_pending = false;
320 	} else {
321 		/*
322 		 * Otherwise: reduce the boost value and disable it when we
323 		 * reach the minimum.
324 		 */
325 		sg_cpu->iowait_boost >>= 1;
326 		if (sg_cpu->iowait_boost < sg_cpu->sg_policy->policy->min) {
327 			sg_cpu->iowait_boost = 0;
328 			return;
329 		}
330 	}
331 
332 	/*
333 	 * Apply the current boost value: a CPU is boosted only if its current
334 	 * utilization is smaller then the current IO boost level.
335 	 */
336 	boost_util = sg_cpu->iowait_boost;
337 	boost_max = sg_cpu->iowait_boost_max;
338 	if (*util * boost_max < *max * boost_util) {
339 		*util = boost_util;
340 		*max = boost_max;
341 	}
342 }
343 
344 #ifdef CONFIG_NO_HZ_COMMON
345 static bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu)
346 {
347 	unsigned long idle_calls = tick_nohz_get_idle_calls_cpu(sg_cpu->cpu);
348 	bool ret = idle_calls == sg_cpu->saved_idle_calls;
349 
350 	sg_cpu->saved_idle_calls = idle_calls;
351 	return ret;
352 }
353 #else
354 static inline bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu) { return false; }
355 #endif /* CONFIG_NO_HZ_COMMON */
356 
357 /*
358  * Make sugov_should_update_freq() ignore the rate limit when DL
359  * has increased the utilization.
360  */
361 static inline void ignore_dl_rate_limit(struct sugov_cpu *sg_cpu, struct sugov_policy *sg_policy)
362 {
363 	if (cpu_util_dl(cpu_rq(sg_cpu->cpu)) > sg_cpu->util_dl)
364 		sg_policy->need_freq_update = true;
365 }
366 
367 static void sugov_update_single(struct update_util_data *hook, u64 time,
368 				unsigned int flags)
369 {
370 	struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
371 	struct sugov_policy *sg_policy = sg_cpu->sg_policy;
372 	unsigned long util, max;
373 	unsigned int next_f;
374 	bool busy;
375 
376 	sugov_iowait_boost(sg_cpu, time, flags);
377 	sg_cpu->last_update = time;
378 
379 	ignore_dl_rate_limit(sg_cpu, sg_policy);
380 
381 	if (!sugov_should_update_freq(sg_policy, time))
382 		return;
383 
384 	busy = sugov_cpu_is_busy(sg_cpu);
385 
386 	sugov_get_util(sg_cpu);
387 	max = sg_cpu->max;
388 	util = sugov_aggregate_util(sg_cpu);
389 	sugov_iowait_apply(sg_cpu, time, &util, &max);
390 	next_f = get_next_freq(sg_policy, util, max);
391 	/*
392 	 * Do not reduce the frequency if the CPU has not been idle
393 	 * recently, as the reduction is likely to be premature then.
394 	 */
395 	if (busy && next_f < sg_policy->next_freq) {
396 		next_f = sg_policy->next_freq;
397 
398 		/* Reset cached freq as next_freq has changed */
399 		sg_policy->cached_raw_freq = 0;
400 	}
401 
402 	/*
403 	 * This code runs under rq->lock for the target CPU, so it won't run
404 	 * concurrently on two different CPUs for the same target and it is not
405 	 * necessary to acquire the lock in the fast switch case.
406 	 */
407 	if (sg_policy->policy->fast_switch_enabled) {
408 		sugov_fast_switch(sg_policy, time, next_f);
409 	} else {
410 		raw_spin_lock(&sg_policy->update_lock);
411 		sugov_deferred_update(sg_policy, time, next_f);
412 		raw_spin_unlock(&sg_policy->update_lock);
413 	}
414 }
415 
416 static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu, u64 time)
417 {
418 	struct sugov_policy *sg_policy = sg_cpu->sg_policy;
419 	struct cpufreq_policy *policy = sg_policy->policy;
420 	unsigned long util = 0, max = 1;
421 	unsigned int j;
422 
423 	for_each_cpu(j, policy->cpus) {
424 		struct sugov_cpu *j_sg_cpu = &per_cpu(sugov_cpu, j);
425 		unsigned long j_util, j_max;
426 
427 		sugov_get_util(j_sg_cpu);
428 		j_max = j_sg_cpu->max;
429 		j_util = sugov_aggregate_util(j_sg_cpu);
430 		sugov_iowait_apply(j_sg_cpu, time, &j_util, &j_max);
431 
432 		if (j_util * max > j_max * util) {
433 			util = j_util;
434 			max = j_max;
435 		}
436 	}
437 
438 	return get_next_freq(sg_policy, util, max);
439 }
440 
441 static void
442 sugov_update_shared(struct update_util_data *hook, u64 time, unsigned int flags)
443 {
444 	struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
445 	struct sugov_policy *sg_policy = sg_cpu->sg_policy;
446 	unsigned int next_f;
447 
448 	raw_spin_lock(&sg_policy->update_lock);
449 
450 	sugov_iowait_boost(sg_cpu, time, flags);
451 	sg_cpu->last_update = time;
452 
453 	ignore_dl_rate_limit(sg_cpu, sg_policy);
454 
455 	if (sugov_should_update_freq(sg_policy, time)) {
456 		next_f = sugov_next_freq_shared(sg_cpu, time);
457 
458 		if (sg_policy->policy->fast_switch_enabled)
459 			sugov_fast_switch(sg_policy, time, next_f);
460 		else
461 			sugov_deferred_update(sg_policy, time, next_f);
462 	}
463 
464 	raw_spin_unlock(&sg_policy->update_lock);
465 }
466 
467 static void sugov_work(struct kthread_work *work)
468 {
469 	struct sugov_policy *sg_policy = container_of(work, struct sugov_policy, work);
470 	unsigned int freq;
471 	unsigned long flags;
472 
473 	/*
474 	 * Hold sg_policy->update_lock shortly to handle the case where:
475 	 * incase sg_policy->next_freq is read here, and then updated by
476 	 * sugov_deferred_update() just before work_in_progress is set to false
477 	 * here, we may miss queueing the new update.
478 	 *
479 	 * Note: If a work was queued after the update_lock is released,
480 	 * sugov_work() will just be called again by kthread_work code; and the
481 	 * request will be proceed before the sugov thread sleeps.
482 	 */
483 	raw_spin_lock_irqsave(&sg_policy->update_lock, flags);
484 	freq = sg_policy->next_freq;
485 	sg_policy->work_in_progress = false;
486 	raw_spin_unlock_irqrestore(&sg_policy->update_lock, flags);
487 
488 	mutex_lock(&sg_policy->work_lock);
489 	__cpufreq_driver_target(sg_policy->policy, freq, CPUFREQ_RELATION_L);
490 	mutex_unlock(&sg_policy->work_lock);
491 }
492 
493 static void sugov_irq_work(struct irq_work *irq_work)
494 {
495 	struct sugov_policy *sg_policy;
496 
497 	sg_policy = container_of(irq_work, struct sugov_policy, irq_work);
498 
499 	kthread_queue_work(&sg_policy->worker, &sg_policy->work);
500 }
501 
502 /************************** sysfs interface ************************/
503 
504 static struct sugov_tunables *global_tunables;
505 static DEFINE_MUTEX(global_tunables_lock);
506 
507 static inline struct sugov_tunables *to_sugov_tunables(struct gov_attr_set *attr_set)
508 {
509 	return container_of(attr_set, struct sugov_tunables, attr_set);
510 }
511 
512 static ssize_t rate_limit_us_show(struct gov_attr_set *attr_set, char *buf)
513 {
514 	struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
515 
516 	return sprintf(buf, "%u\n", tunables->rate_limit_us);
517 }
518 
519 static ssize_t
520 rate_limit_us_store(struct gov_attr_set *attr_set, const char *buf, size_t count)
521 {
522 	struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
523 	struct sugov_policy *sg_policy;
524 	unsigned int rate_limit_us;
525 
526 	if (kstrtouint(buf, 10, &rate_limit_us))
527 		return -EINVAL;
528 
529 	tunables->rate_limit_us = rate_limit_us;
530 
531 	list_for_each_entry(sg_policy, &attr_set->policy_list, tunables_hook)
532 		sg_policy->freq_update_delay_ns = rate_limit_us * NSEC_PER_USEC;
533 
534 	return count;
535 }
536 
537 static struct governor_attr rate_limit_us = __ATTR_RW(rate_limit_us);
538 
539 static struct attribute *sugov_attributes[] = {
540 	&rate_limit_us.attr,
541 	NULL
542 };
543 
544 static struct kobj_type sugov_tunables_ktype = {
545 	.default_attrs = sugov_attributes,
546 	.sysfs_ops = &governor_sysfs_ops,
547 };
548 
549 /********************** cpufreq governor interface *********************/
550 
551 static struct cpufreq_governor schedutil_gov;
552 
553 static struct sugov_policy *sugov_policy_alloc(struct cpufreq_policy *policy)
554 {
555 	struct sugov_policy *sg_policy;
556 
557 	sg_policy = kzalloc(sizeof(*sg_policy), GFP_KERNEL);
558 	if (!sg_policy)
559 		return NULL;
560 
561 	sg_policy->policy = policy;
562 	raw_spin_lock_init(&sg_policy->update_lock);
563 	return sg_policy;
564 }
565 
566 static void sugov_policy_free(struct sugov_policy *sg_policy)
567 {
568 	kfree(sg_policy);
569 }
570 
571 static int sugov_kthread_create(struct sugov_policy *sg_policy)
572 {
573 	struct task_struct *thread;
574 	struct sched_attr attr = {
575 		.size		= sizeof(struct sched_attr),
576 		.sched_policy	= SCHED_DEADLINE,
577 		.sched_flags	= SCHED_FLAG_SUGOV,
578 		.sched_nice	= 0,
579 		.sched_priority	= 0,
580 		/*
581 		 * Fake (unused) bandwidth; workaround to "fix"
582 		 * priority inheritance.
583 		 */
584 		.sched_runtime	=  1000000,
585 		.sched_deadline = 10000000,
586 		.sched_period	= 10000000,
587 	};
588 	struct cpufreq_policy *policy = sg_policy->policy;
589 	int ret;
590 
591 	/* kthread only required for slow path */
592 	if (policy->fast_switch_enabled)
593 		return 0;
594 
595 	kthread_init_work(&sg_policy->work, sugov_work);
596 	kthread_init_worker(&sg_policy->worker);
597 	thread = kthread_create(kthread_worker_fn, &sg_policy->worker,
598 				"sugov:%d",
599 				cpumask_first(policy->related_cpus));
600 	if (IS_ERR(thread)) {
601 		pr_err("failed to create sugov thread: %ld\n", PTR_ERR(thread));
602 		return PTR_ERR(thread);
603 	}
604 
605 	ret = sched_setattr_nocheck(thread, &attr);
606 	if (ret) {
607 		kthread_stop(thread);
608 		pr_warn("%s: failed to set SCHED_DEADLINE\n", __func__);
609 		return ret;
610 	}
611 
612 	sg_policy->thread = thread;
613 	kthread_bind_mask(thread, policy->related_cpus);
614 	init_irq_work(&sg_policy->irq_work, sugov_irq_work);
615 	mutex_init(&sg_policy->work_lock);
616 
617 	wake_up_process(thread);
618 
619 	return 0;
620 }
621 
622 static void sugov_kthread_stop(struct sugov_policy *sg_policy)
623 {
624 	/* kthread only required for slow path */
625 	if (sg_policy->policy->fast_switch_enabled)
626 		return;
627 
628 	kthread_flush_worker(&sg_policy->worker);
629 	kthread_stop(sg_policy->thread);
630 	mutex_destroy(&sg_policy->work_lock);
631 }
632 
633 static struct sugov_tunables *sugov_tunables_alloc(struct sugov_policy *sg_policy)
634 {
635 	struct sugov_tunables *tunables;
636 
637 	tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
638 	if (tunables) {
639 		gov_attr_set_init(&tunables->attr_set, &sg_policy->tunables_hook);
640 		if (!have_governor_per_policy())
641 			global_tunables = tunables;
642 	}
643 	return tunables;
644 }
645 
646 static void sugov_tunables_free(struct sugov_tunables *tunables)
647 {
648 	if (!have_governor_per_policy())
649 		global_tunables = NULL;
650 
651 	kfree(tunables);
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 	policy->governor_data = NULL;
713 	sugov_tunables_free(tunables);
714 
715 stop_kthread:
716 	sugov_kthread_stop(sg_policy);
717 	mutex_unlock(&global_tunables_lock);
718 
719 free_sg_policy:
720 	sugov_policy_free(sg_policy);
721 
722 disable_fast_switch:
723 	cpufreq_disable_fast_switch(policy);
724 
725 	pr_err("initialization failed (error %d)\n", ret);
726 	return ret;
727 }
728 
729 static void sugov_exit(struct cpufreq_policy *policy)
730 {
731 	struct sugov_policy *sg_policy = policy->governor_data;
732 	struct sugov_tunables *tunables = sg_policy->tunables;
733 	unsigned int count;
734 
735 	mutex_lock(&global_tunables_lock);
736 
737 	count = gov_attr_set_put(&tunables->attr_set, &sg_policy->tunables_hook);
738 	policy->governor_data = NULL;
739 	if (!count)
740 		sugov_tunables_free(tunables);
741 
742 	mutex_unlock(&global_tunables_lock);
743 
744 	sugov_kthread_stop(sg_policy);
745 	sugov_policy_free(sg_policy);
746 	cpufreq_disable_fast_switch(policy);
747 }
748 
749 static int sugov_start(struct cpufreq_policy *policy)
750 {
751 	struct sugov_policy *sg_policy = policy->governor_data;
752 	unsigned int cpu;
753 
754 	sg_policy->freq_update_delay_ns	= sg_policy->tunables->rate_limit_us * NSEC_PER_USEC;
755 	sg_policy->last_freq_update_time	= 0;
756 	sg_policy->next_freq			= 0;
757 	sg_policy->work_in_progress		= false;
758 	sg_policy->need_freq_update		= false;
759 	sg_policy->cached_raw_freq		= 0;
760 
761 	for_each_cpu(cpu, policy->cpus) {
762 		struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
763 
764 		memset(sg_cpu, 0, sizeof(*sg_cpu));
765 		sg_cpu->cpu			= cpu;
766 		sg_cpu->sg_policy		= sg_policy;
767 		sg_cpu->iowait_boost_max	= policy->cpuinfo.max_freq;
768 	}
769 
770 	for_each_cpu(cpu, policy->cpus) {
771 		struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
772 
773 		cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util,
774 					     policy_is_shared(policy) ?
775 							sugov_update_shared :
776 							sugov_update_single);
777 	}
778 	return 0;
779 }
780 
781 static void sugov_stop(struct cpufreq_policy *policy)
782 {
783 	struct sugov_policy *sg_policy = policy->governor_data;
784 	unsigned int cpu;
785 
786 	for_each_cpu(cpu, policy->cpus)
787 		cpufreq_remove_update_util_hook(cpu);
788 
789 	synchronize_sched();
790 
791 	if (!policy->fast_switch_enabled) {
792 		irq_work_sync(&sg_policy->irq_work);
793 		kthread_cancel_work_sync(&sg_policy->work);
794 	}
795 }
796 
797 static void sugov_limits(struct cpufreq_policy *policy)
798 {
799 	struct sugov_policy *sg_policy = policy->governor_data;
800 
801 	if (!policy->fast_switch_enabled) {
802 		mutex_lock(&sg_policy->work_lock);
803 		cpufreq_policy_apply_limits(policy);
804 		mutex_unlock(&sg_policy->work_lock);
805 	}
806 
807 	sg_policy->need_freq_update = true;
808 }
809 
810 static struct cpufreq_governor schedutil_gov = {
811 	.name			= "schedutil",
812 	.owner			= THIS_MODULE,
813 	.dynamic_switching	= true,
814 	.init			= sugov_init,
815 	.exit			= sugov_exit,
816 	.start			= sugov_start,
817 	.stop			= sugov_stop,
818 	.limits			= sugov_limits,
819 };
820 
821 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL
822 struct cpufreq_governor *cpufreq_default_governor(void)
823 {
824 	return &schedutil_gov;
825 }
826 #endif
827 
828 static int __init sugov_register(void)
829 {
830 	return cpufreq_register_governor(&schedutil_gov);
831 }
832 fs_initcall(sugov_register);
833