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 pr_fmt(fmt) KBUILD_MODNAME ": " fmt
10 
11 #include "sched.h"
12 
13 #include <linux/sched/cpufreq.h>
14 #include <trace/events/power.h>
15 
16 #define IOWAIT_BOOST_MIN	(SCHED_CAPACITY_SCALE / 8)
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			limits_changed;
44 	bool			need_freq_update;
45 };
46 
47 struct sugov_cpu {
48 	struct update_util_data	update_util;
49 	struct sugov_policy	*sg_policy;
50 	unsigned int		cpu;
51 
52 	bool			iowait_boost_pending;
53 	unsigned int		iowait_boost;
54 	u64			last_update;
55 
56 	unsigned long		bw_dl;
57 	unsigned long		max;
58 
59 	/* The field below is for single-CPU policies only: */
60 #ifdef CONFIG_NO_HZ_COMMON
61 	unsigned long		saved_idle_calls;
62 #endif
63 };
64 
65 static DEFINE_PER_CPU(struct sugov_cpu, sugov_cpu);
66 
67 /************************ Governor internals ***********************/
68 
69 static bool sugov_should_update_freq(struct sugov_policy *sg_policy, u64 time)
70 {
71 	s64 delta_ns;
72 
73 	/*
74 	 * Since cpufreq_update_util() is called with rq->lock held for
75 	 * the @target_cpu, our per-CPU data is fully serialized.
76 	 *
77 	 * However, drivers cannot in general deal with cross-CPU
78 	 * requests, so while get_next_freq() will work, our
79 	 * sugov_update_commit() call may not for the fast switching platforms.
80 	 *
81 	 * Hence stop here for remote requests if they aren't supported
82 	 * by the hardware, as calculating the frequency is pointless if
83 	 * we cannot in fact act on it.
84 	 *
85 	 * This is needed on the slow switching platforms too to prevent CPUs
86 	 * going offline from leaving stale IRQ work items behind.
87 	 */
88 	if (!cpufreq_this_cpu_can_update(sg_policy->policy))
89 		return false;
90 
91 	if (unlikely(sg_policy->limits_changed)) {
92 		sg_policy->limits_changed = false;
93 		sg_policy->need_freq_update = true;
94 		return true;
95 	}
96 
97 	delta_ns = time - sg_policy->last_freq_update_time;
98 
99 	return delta_ns >= sg_policy->freq_update_delay_ns;
100 }
101 
102 static bool sugov_update_next_freq(struct sugov_policy *sg_policy, u64 time,
103 				   unsigned int next_freq)
104 {
105 	if (sg_policy->next_freq == next_freq)
106 		return false;
107 
108 	sg_policy->next_freq = next_freq;
109 	sg_policy->last_freq_update_time = time;
110 
111 	return true;
112 }
113 
114 static void sugov_fast_switch(struct sugov_policy *sg_policy, u64 time,
115 			      unsigned int next_freq)
116 {
117 	struct cpufreq_policy *policy = sg_policy->policy;
118 	int cpu;
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 
129 	if (trace_cpu_frequency_enabled()) {
130 		for_each_cpu(cpu, policy->cpus)
131 			trace_cpu_frequency(next_freq, cpu);
132 	}
133 }
134 
135 static void sugov_deferred_update(struct sugov_policy *sg_policy, u64 time,
136 				  unsigned int next_freq)
137 {
138 	if (!sugov_update_next_freq(sg_policy, time, next_freq))
139 		return;
140 
141 	if (!sg_policy->work_in_progress) {
142 		sg_policy->work_in_progress = true;
143 		irq_work_queue(&sg_policy->irq_work);
144 	}
145 }
146 
147 /**
148  * get_next_freq - Compute a new frequency for a given cpufreq policy.
149  * @sg_policy: schedutil policy object to compute the new frequency for.
150  * @util: Current CPU utilization.
151  * @max: CPU capacity.
152  *
153  * If the utilization is frequency-invariant, choose the new frequency to be
154  * proportional to it, that is
155  *
156  * next_freq = C * max_freq * util / max
157  *
158  * Otherwise, approximate the would-be frequency-invariant utilization by
159  * util_raw * (curr_freq / max_freq) which leads to
160  *
161  * next_freq = C * curr_freq * util_raw / max
162  *
163  * Take C = 1.25 for the frequency tipping point at (util / max) = 0.8.
164  *
165  * The lowest driver-supported frequency which is equal or greater than the raw
166  * next_freq (as calculated above) is returned, subject to policy min/max and
167  * cpufreq driver limitations.
168  */
169 static unsigned int get_next_freq(struct sugov_policy *sg_policy,
170 				  unsigned long util, unsigned long max)
171 {
172 	struct cpufreq_policy *policy = sg_policy->policy;
173 	unsigned int freq = arch_scale_freq_invariant() ?
174 				policy->cpuinfo.max_freq : policy->cur;
175 
176 	freq = map_util_freq(util, freq, max);
177 
178 	if (freq == sg_policy->cached_raw_freq && !sg_policy->need_freq_update)
179 		return sg_policy->next_freq;
180 
181 	sg_policy->need_freq_update = false;
182 	sg_policy->cached_raw_freq = freq;
183 	return cpufreq_driver_resolve_freq(policy, freq);
184 }
185 
186 /*
187  * This function computes an effective utilization for the given CPU, to be
188  * used for frequency selection given the linear relation: f = u * f_max.
189  *
190  * The scheduler tracks the following metrics:
191  *
192  *   cpu_util_{cfs,rt,dl,irq}()
193  *   cpu_bw_dl()
194  *
195  * Where the cfs,rt and dl util numbers are tracked with the same metric and
196  * synchronized windows and are thus directly comparable.
197  *
198  * The cfs,rt,dl utilization are the running times measured with rq->clock_task
199  * which excludes things like IRQ and steal-time. These latter are then accrued
200  * in the irq utilization.
201  *
202  * The DL bandwidth number otoh is not a measured metric but a value computed
203  * based on the task model parameters and gives the minimal utilization
204  * required to meet deadlines.
205  */
206 unsigned long schedutil_cpu_util(int cpu, unsigned long util_cfs,
207 				 unsigned long max, enum schedutil_type type,
208 				 struct task_struct *p)
209 {
210 	unsigned long dl_util, util, irq;
211 	struct rq *rq = cpu_rq(cpu);
212 
213 	if (!IS_BUILTIN(CONFIG_UCLAMP_TASK) &&
214 	    type == FREQUENCY_UTIL && rt_rq_is_runnable(&rq->rt)) {
215 		return max;
216 	}
217 
218 	/*
219 	 * Early check to see if IRQ/steal time saturates the CPU, can be
220 	 * because of inaccuracies in how we track these -- see
221 	 * update_irq_load_avg().
222 	 */
223 	irq = cpu_util_irq(rq);
224 	if (unlikely(irq >= max))
225 		return max;
226 
227 	/*
228 	 * Because the time spend on RT/DL tasks is visible as 'lost' time to
229 	 * CFS tasks and we use the same metric to track the effective
230 	 * utilization (PELT windows are synchronized) we can directly add them
231 	 * to obtain the CPU's actual utilization.
232 	 *
233 	 * CFS and RT utilization can be boosted or capped, depending on
234 	 * utilization clamp constraints requested by currently RUNNABLE
235 	 * tasks.
236 	 * When there are no CFS RUNNABLE tasks, clamps are released and
237 	 * frequency will be gracefully reduced with the utilization decay.
238 	 */
239 	util = util_cfs + cpu_util_rt(rq);
240 	if (type == FREQUENCY_UTIL)
241 		util = uclamp_rq_util_with(rq, util, p);
242 
243 	dl_util = cpu_util_dl(rq);
244 
245 	/*
246 	 * For frequency selection we do not make cpu_util_dl() a permanent part
247 	 * of this sum because we want to use cpu_bw_dl() later on, but we need
248 	 * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
249 	 * that we select f_max when there is no idle time.
250 	 *
251 	 * NOTE: numerical errors or stop class might cause us to not quite hit
252 	 * saturation when we should -- something for later.
253 	 */
254 	if (util + dl_util >= max)
255 		return max;
256 
257 	/*
258 	 * OTOH, for energy computation we need the estimated running time, so
259 	 * include util_dl and ignore dl_bw.
260 	 */
261 	if (type == ENERGY_UTIL)
262 		util += dl_util;
263 
264 	/*
265 	 * There is still idle time; further improve the number by using the
266 	 * irq metric. Because IRQ/steal time is hidden from the task clock we
267 	 * need to scale the task numbers:
268 	 *
269 	 *              max - irq
270 	 *   U' = irq + --------- * U
271 	 *                 max
272 	 */
273 	util = scale_irq_capacity(util, irq, max);
274 	util += irq;
275 
276 	/*
277 	 * Bandwidth required by DEADLINE must always be granted while, for
278 	 * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
279 	 * to gracefully reduce the frequency when no tasks show up for longer
280 	 * periods of time.
281 	 *
282 	 * Ideally we would like to set bw_dl as min/guaranteed freq and util +
283 	 * bw_dl as requested freq. However, cpufreq is not yet ready for such
284 	 * an interface. So, we only do the latter for now.
285 	 */
286 	if (type == FREQUENCY_UTIL)
287 		util += cpu_bw_dl(rq);
288 
289 	return min(max, util);
290 }
291 
292 static unsigned long sugov_get_util(struct sugov_cpu *sg_cpu)
293 {
294 	struct rq *rq = cpu_rq(sg_cpu->cpu);
295 	unsigned long util = cpu_util_cfs(rq);
296 	unsigned long max = arch_scale_cpu_capacity(sg_cpu->cpu);
297 
298 	sg_cpu->max = max;
299 	sg_cpu->bw_dl = cpu_bw_dl(rq);
300 
301 	return schedutil_cpu_util(sg_cpu->cpu, util, max, FREQUENCY_UTIL, NULL);
302 }
303 
304 /**
305  * sugov_iowait_reset() - Reset the IO boost status of a CPU.
306  * @sg_cpu: the sugov data for the CPU to boost
307  * @time: the update time from the caller
308  * @set_iowait_boost: true if an IO boost has been requested
309  *
310  * The IO wait boost of a task is disabled after a tick since the last update
311  * of a CPU. If a new IO wait boost is requested after more then a tick, then
312  * we enable the boost starting from IOWAIT_BOOST_MIN, which improves energy
313  * efficiency by ignoring sporadic wakeups from IO.
314  */
315 static bool sugov_iowait_reset(struct sugov_cpu *sg_cpu, u64 time,
316 			       bool set_iowait_boost)
317 {
318 	s64 delta_ns = time - sg_cpu->last_update;
319 
320 	/* Reset boost only if a tick has elapsed since last request */
321 	if (delta_ns <= TICK_NSEC)
322 		return false;
323 
324 	sg_cpu->iowait_boost = set_iowait_boost ? IOWAIT_BOOST_MIN : 0;
325 	sg_cpu->iowait_boost_pending = set_iowait_boost;
326 
327 	return true;
328 }
329 
330 /**
331  * sugov_iowait_boost() - Updates the IO boost status of a CPU.
332  * @sg_cpu: the sugov data for the CPU to boost
333  * @time: the update time from the caller
334  * @flags: SCHED_CPUFREQ_IOWAIT if the task is waking up after an IO wait
335  *
336  * Each time a task wakes up after an IO operation, the CPU utilization can be
337  * boosted to a certain utilization which doubles at each "frequent and
338  * successive" wakeup from IO, ranging from IOWAIT_BOOST_MIN to the utilization
339  * of the maximum OPP.
340  *
341  * To keep doubling, an IO boost has to be requested at least once per tick,
342  * otherwise we restart from the utilization of the minimum OPP.
343  */
344 static void sugov_iowait_boost(struct sugov_cpu *sg_cpu, u64 time,
345 			       unsigned int flags)
346 {
347 	bool set_iowait_boost = flags & SCHED_CPUFREQ_IOWAIT;
348 
349 	/* Reset boost if the CPU appears to have been idle enough */
350 	if (sg_cpu->iowait_boost &&
351 	    sugov_iowait_reset(sg_cpu, time, set_iowait_boost))
352 		return;
353 
354 	/* Boost only tasks waking up after IO */
355 	if (!set_iowait_boost)
356 		return;
357 
358 	/* Ensure boost doubles only one time at each request */
359 	if (sg_cpu->iowait_boost_pending)
360 		return;
361 	sg_cpu->iowait_boost_pending = true;
362 
363 	/* Double the boost at each request */
364 	if (sg_cpu->iowait_boost) {
365 		sg_cpu->iowait_boost =
366 			min_t(unsigned int, sg_cpu->iowait_boost << 1, SCHED_CAPACITY_SCALE);
367 		return;
368 	}
369 
370 	/* First wakeup after IO: start with minimum boost */
371 	sg_cpu->iowait_boost = IOWAIT_BOOST_MIN;
372 }
373 
374 /**
375  * sugov_iowait_apply() - Apply the IO boost to a CPU.
376  * @sg_cpu: the sugov data for the cpu to boost
377  * @time: the update time from the caller
378  * @util: the utilization to (eventually) boost
379  * @max: the maximum value the utilization can be boosted to
380  *
381  * A CPU running a task which woken up after an IO operation can have its
382  * utilization boosted to speed up the completion of those IO operations.
383  * The IO boost value is increased each time a task wakes up from IO, in
384  * sugov_iowait_apply(), and it's instead decreased by this function,
385  * each time an increase has not been requested (!iowait_boost_pending).
386  *
387  * A CPU which also appears to have been idle for at least one tick has also
388  * its IO boost utilization reset.
389  *
390  * This mechanism is designed to boost high frequently IO waiting tasks, while
391  * being more conservative on tasks which does sporadic IO operations.
392  */
393 static unsigned long sugov_iowait_apply(struct sugov_cpu *sg_cpu, u64 time,
394 					unsigned long util, unsigned long max)
395 {
396 	unsigned long boost;
397 
398 	/* No boost currently required */
399 	if (!sg_cpu->iowait_boost)
400 		return util;
401 
402 	/* Reset boost if the CPU appears to have been idle enough */
403 	if (sugov_iowait_reset(sg_cpu, time, false))
404 		return util;
405 
406 	if (!sg_cpu->iowait_boost_pending) {
407 		/*
408 		 * No boost pending; reduce the boost value.
409 		 */
410 		sg_cpu->iowait_boost >>= 1;
411 		if (sg_cpu->iowait_boost < IOWAIT_BOOST_MIN) {
412 			sg_cpu->iowait_boost = 0;
413 			return util;
414 		}
415 	}
416 
417 	sg_cpu->iowait_boost_pending = false;
418 
419 	/*
420 	 * @util is already in capacity scale; convert iowait_boost
421 	 * into the same scale so we can compare.
422 	 */
423 	boost = (sg_cpu->iowait_boost * max) >> SCHED_CAPACITY_SHIFT;
424 	return max(boost, util);
425 }
426 
427 #ifdef CONFIG_NO_HZ_COMMON
428 static bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu)
429 {
430 	unsigned long idle_calls = tick_nohz_get_idle_calls_cpu(sg_cpu->cpu);
431 	bool ret = idle_calls == sg_cpu->saved_idle_calls;
432 
433 	sg_cpu->saved_idle_calls = idle_calls;
434 	return ret;
435 }
436 #else
437 static inline bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu) { return false; }
438 #endif /* CONFIG_NO_HZ_COMMON */
439 
440 /*
441  * Make sugov_should_update_freq() ignore the rate limit when DL
442  * has increased the utilization.
443  */
444 static inline void ignore_dl_rate_limit(struct sugov_cpu *sg_cpu, struct sugov_policy *sg_policy)
445 {
446 	if (cpu_bw_dl(cpu_rq(sg_cpu->cpu)) > sg_cpu->bw_dl)
447 		sg_policy->limits_changed = true;
448 }
449 
450 static void sugov_update_single(struct update_util_data *hook, u64 time,
451 				unsigned int flags)
452 {
453 	struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
454 	struct sugov_policy *sg_policy = sg_cpu->sg_policy;
455 	unsigned long util, max;
456 	unsigned int next_f;
457 	bool busy;
458 
459 	sugov_iowait_boost(sg_cpu, time, flags);
460 	sg_cpu->last_update = time;
461 
462 	ignore_dl_rate_limit(sg_cpu, sg_policy);
463 
464 	if (!sugov_should_update_freq(sg_policy, time))
465 		return;
466 
467 	/* Limits may have changed, don't skip frequency update */
468 	busy = !sg_policy->need_freq_update && sugov_cpu_is_busy(sg_cpu);
469 
470 	util = sugov_get_util(sg_cpu);
471 	max = sg_cpu->max;
472 	util = sugov_iowait_apply(sg_cpu, time, util, max);
473 	next_f = get_next_freq(sg_policy, util, max);
474 	/*
475 	 * Do not reduce the frequency if the CPU has not been idle
476 	 * recently, as the reduction is likely to be premature then.
477 	 */
478 	if (busy && next_f < sg_policy->next_freq) {
479 		next_f = sg_policy->next_freq;
480 
481 		/* Reset cached freq as next_freq has changed */
482 		sg_policy->cached_raw_freq = 0;
483 	}
484 
485 	/*
486 	 * This code runs under rq->lock for the target CPU, so it won't run
487 	 * concurrently on two different CPUs for the same target and it is not
488 	 * necessary to acquire the lock in the fast switch case.
489 	 */
490 	if (sg_policy->policy->fast_switch_enabled) {
491 		sugov_fast_switch(sg_policy, time, next_f);
492 	} else {
493 		raw_spin_lock(&sg_policy->update_lock);
494 		sugov_deferred_update(sg_policy, time, next_f);
495 		raw_spin_unlock(&sg_policy->update_lock);
496 	}
497 }
498 
499 static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu, u64 time)
500 {
501 	struct sugov_policy *sg_policy = sg_cpu->sg_policy;
502 	struct cpufreq_policy *policy = sg_policy->policy;
503 	unsigned long util = 0, max = 1;
504 	unsigned int j;
505 
506 	for_each_cpu(j, policy->cpus) {
507 		struct sugov_cpu *j_sg_cpu = &per_cpu(sugov_cpu, j);
508 		unsigned long j_util, j_max;
509 
510 		j_util = sugov_get_util(j_sg_cpu);
511 		j_max = j_sg_cpu->max;
512 		j_util = sugov_iowait_apply(j_sg_cpu, time, j_util, j_max);
513 
514 		if (j_util * max > j_max * util) {
515 			util = j_util;
516 			max = j_max;
517 		}
518 	}
519 
520 	return get_next_freq(sg_policy, util, max);
521 }
522 
523 static void
524 sugov_update_shared(struct update_util_data *hook, u64 time, unsigned int flags)
525 {
526 	struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
527 	struct sugov_policy *sg_policy = sg_cpu->sg_policy;
528 	unsigned int next_f;
529 
530 	raw_spin_lock(&sg_policy->update_lock);
531 
532 	sugov_iowait_boost(sg_cpu, time, flags);
533 	sg_cpu->last_update = time;
534 
535 	ignore_dl_rate_limit(sg_cpu, sg_policy);
536 
537 	if (sugov_should_update_freq(sg_policy, time)) {
538 		next_f = sugov_next_freq_shared(sg_cpu, time);
539 
540 		if (sg_policy->policy->fast_switch_enabled)
541 			sugov_fast_switch(sg_policy, time, next_f);
542 		else
543 			sugov_deferred_update(sg_policy, time, next_f);
544 	}
545 
546 	raw_spin_unlock(&sg_policy->update_lock);
547 }
548 
549 static void sugov_work(struct kthread_work *work)
550 {
551 	struct sugov_policy *sg_policy = container_of(work, struct sugov_policy, work);
552 	unsigned int freq;
553 	unsigned long flags;
554 
555 	/*
556 	 * Hold sg_policy->update_lock shortly to handle the case where:
557 	 * incase sg_policy->next_freq is read here, and then updated by
558 	 * sugov_deferred_update() just before work_in_progress is set to false
559 	 * here, we may miss queueing the new update.
560 	 *
561 	 * Note: If a work was queued after the update_lock is released,
562 	 * sugov_work() will just be called again by kthread_work code; and the
563 	 * request will be proceed before the sugov thread sleeps.
564 	 */
565 	raw_spin_lock_irqsave(&sg_policy->update_lock, flags);
566 	freq = sg_policy->next_freq;
567 	sg_policy->work_in_progress = false;
568 	raw_spin_unlock_irqrestore(&sg_policy->update_lock, flags);
569 
570 	mutex_lock(&sg_policy->work_lock);
571 	__cpufreq_driver_target(sg_policy->policy, freq, CPUFREQ_RELATION_L);
572 	mutex_unlock(&sg_policy->work_lock);
573 }
574 
575 static void sugov_irq_work(struct irq_work *irq_work)
576 {
577 	struct sugov_policy *sg_policy;
578 
579 	sg_policy = container_of(irq_work, struct sugov_policy, irq_work);
580 
581 	kthread_queue_work(&sg_policy->worker, &sg_policy->work);
582 }
583 
584 /************************** sysfs interface ************************/
585 
586 static struct sugov_tunables *global_tunables;
587 static DEFINE_MUTEX(global_tunables_lock);
588 
589 static inline struct sugov_tunables *to_sugov_tunables(struct gov_attr_set *attr_set)
590 {
591 	return container_of(attr_set, struct sugov_tunables, attr_set);
592 }
593 
594 static ssize_t rate_limit_us_show(struct gov_attr_set *attr_set, char *buf)
595 {
596 	struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
597 
598 	return sprintf(buf, "%u\n", tunables->rate_limit_us);
599 }
600 
601 static ssize_t
602 rate_limit_us_store(struct gov_attr_set *attr_set, const char *buf, size_t count)
603 {
604 	struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
605 	struct sugov_policy *sg_policy;
606 	unsigned int rate_limit_us;
607 
608 	if (kstrtouint(buf, 10, &rate_limit_us))
609 		return -EINVAL;
610 
611 	tunables->rate_limit_us = rate_limit_us;
612 
613 	list_for_each_entry(sg_policy, &attr_set->policy_list, tunables_hook)
614 		sg_policy->freq_update_delay_ns = rate_limit_us * NSEC_PER_USEC;
615 
616 	return count;
617 }
618 
619 static struct governor_attr rate_limit_us = __ATTR_RW(rate_limit_us);
620 
621 static struct attribute *sugov_attrs[] = {
622 	&rate_limit_us.attr,
623 	NULL
624 };
625 ATTRIBUTE_GROUPS(sugov);
626 
627 static struct kobj_type sugov_tunables_ktype = {
628 	.default_groups = sugov_groups,
629 	.sysfs_ops = &governor_sysfs_ops,
630 };
631 
632 /********************** cpufreq governor interface *********************/
633 
634 struct cpufreq_governor schedutil_gov;
635 
636 static struct sugov_policy *sugov_policy_alloc(struct cpufreq_policy *policy)
637 {
638 	struct sugov_policy *sg_policy;
639 
640 	sg_policy = kzalloc(sizeof(*sg_policy), GFP_KERNEL);
641 	if (!sg_policy)
642 		return NULL;
643 
644 	sg_policy->policy = policy;
645 	raw_spin_lock_init(&sg_policy->update_lock);
646 	return sg_policy;
647 }
648 
649 static void sugov_policy_free(struct sugov_policy *sg_policy)
650 {
651 	kfree(sg_policy);
652 }
653 
654 static int sugov_kthread_create(struct sugov_policy *sg_policy)
655 {
656 	struct task_struct *thread;
657 	struct sched_attr attr = {
658 		.size		= sizeof(struct sched_attr),
659 		.sched_policy	= SCHED_DEADLINE,
660 		.sched_flags	= SCHED_FLAG_SUGOV,
661 		.sched_nice	= 0,
662 		.sched_priority	= 0,
663 		/*
664 		 * Fake (unused) bandwidth; workaround to "fix"
665 		 * priority inheritance.
666 		 */
667 		.sched_runtime	=  1000000,
668 		.sched_deadline = 10000000,
669 		.sched_period	= 10000000,
670 	};
671 	struct cpufreq_policy *policy = sg_policy->policy;
672 	int ret;
673 
674 	/* kthread only required for slow path */
675 	if (policy->fast_switch_enabled)
676 		return 0;
677 
678 	kthread_init_work(&sg_policy->work, sugov_work);
679 	kthread_init_worker(&sg_policy->worker);
680 	thread = kthread_create(kthread_worker_fn, &sg_policy->worker,
681 				"sugov:%d",
682 				cpumask_first(policy->related_cpus));
683 	if (IS_ERR(thread)) {
684 		pr_err("failed to create sugov thread: %ld\n", PTR_ERR(thread));
685 		return PTR_ERR(thread);
686 	}
687 
688 	ret = sched_setattr_nocheck(thread, &attr);
689 	if (ret) {
690 		kthread_stop(thread);
691 		pr_warn("%s: failed to set SCHED_DEADLINE\n", __func__);
692 		return ret;
693 	}
694 
695 	sg_policy->thread = thread;
696 	kthread_bind_mask(thread, policy->related_cpus);
697 	init_irq_work(&sg_policy->irq_work, sugov_irq_work);
698 	mutex_init(&sg_policy->work_lock);
699 
700 	wake_up_process(thread);
701 
702 	return 0;
703 }
704 
705 static void sugov_kthread_stop(struct sugov_policy *sg_policy)
706 {
707 	/* kthread only required for slow path */
708 	if (sg_policy->policy->fast_switch_enabled)
709 		return;
710 
711 	kthread_flush_worker(&sg_policy->worker);
712 	kthread_stop(sg_policy->thread);
713 	mutex_destroy(&sg_policy->work_lock);
714 }
715 
716 static struct sugov_tunables *sugov_tunables_alloc(struct sugov_policy *sg_policy)
717 {
718 	struct sugov_tunables *tunables;
719 
720 	tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
721 	if (tunables) {
722 		gov_attr_set_init(&tunables->attr_set, &sg_policy->tunables_hook);
723 		if (!have_governor_per_policy())
724 			global_tunables = tunables;
725 	}
726 	return tunables;
727 }
728 
729 static void sugov_tunables_free(struct sugov_tunables *tunables)
730 {
731 	if (!have_governor_per_policy())
732 		global_tunables = NULL;
733 
734 	kfree(tunables);
735 }
736 
737 static int sugov_init(struct cpufreq_policy *policy)
738 {
739 	struct sugov_policy *sg_policy;
740 	struct sugov_tunables *tunables;
741 	int ret = 0;
742 
743 	/* State should be equivalent to EXIT */
744 	if (policy->governor_data)
745 		return -EBUSY;
746 
747 	cpufreq_enable_fast_switch(policy);
748 
749 	sg_policy = sugov_policy_alloc(policy);
750 	if (!sg_policy) {
751 		ret = -ENOMEM;
752 		goto disable_fast_switch;
753 	}
754 
755 	ret = sugov_kthread_create(sg_policy);
756 	if (ret)
757 		goto free_sg_policy;
758 
759 	mutex_lock(&global_tunables_lock);
760 
761 	if (global_tunables) {
762 		if (WARN_ON(have_governor_per_policy())) {
763 			ret = -EINVAL;
764 			goto stop_kthread;
765 		}
766 		policy->governor_data = sg_policy;
767 		sg_policy->tunables = global_tunables;
768 
769 		gov_attr_set_get(&global_tunables->attr_set, &sg_policy->tunables_hook);
770 		goto out;
771 	}
772 
773 	tunables = sugov_tunables_alloc(sg_policy);
774 	if (!tunables) {
775 		ret = -ENOMEM;
776 		goto stop_kthread;
777 	}
778 
779 	tunables->rate_limit_us = cpufreq_policy_transition_delay_us(policy);
780 
781 	policy->governor_data = sg_policy;
782 	sg_policy->tunables = tunables;
783 
784 	ret = kobject_init_and_add(&tunables->attr_set.kobj, &sugov_tunables_ktype,
785 				   get_governor_parent_kobj(policy), "%s",
786 				   schedutil_gov.name);
787 	if (ret)
788 		goto fail;
789 
790 out:
791 	mutex_unlock(&global_tunables_lock);
792 	return 0;
793 
794 fail:
795 	kobject_put(&tunables->attr_set.kobj);
796 	policy->governor_data = NULL;
797 	sugov_tunables_free(tunables);
798 
799 stop_kthread:
800 	sugov_kthread_stop(sg_policy);
801 	mutex_unlock(&global_tunables_lock);
802 
803 free_sg_policy:
804 	sugov_policy_free(sg_policy);
805 
806 disable_fast_switch:
807 	cpufreq_disable_fast_switch(policy);
808 
809 	pr_err("initialization failed (error %d)\n", ret);
810 	return ret;
811 }
812 
813 static void sugov_exit(struct cpufreq_policy *policy)
814 {
815 	struct sugov_policy *sg_policy = policy->governor_data;
816 	struct sugov_tunables *tunables = sg_policy->tunables;
817 	unsigned int count;
818 
819 	mutex_lock(&global_tunables_lock);
820 
821 	count = gov_attr_set_put(&tunables->attr_set, &sg_policy->tunables_hook);
822 	policy->governor_data = NULL;
823 	if (!count)
824 		sugov_tunables_free(tunables);
825 
826 	mutex_unlock(&global_tunables_lock);
827 
828 	sugov_kthread_stop(sg_policy);
829 	sugov_policy_free(sg_policy);
830 	cpufreq_disable_fast_switch(policy);
831 }
832 
833 static int sugov_start(struct cpufreq_policy *policy)
834 {
835 	struct sugov_policy *sg_policy = policy->governor_data;
836 	unsigned int cpu;
837 
838 	sg_policy->freq_update_delay_ns	= sg_policy->tunables->rate_limit_us * NSEC_PER_USEC;
839 	sg_policy->last_freq_update_time	= 0;
840 	sg_policy->next_freq			= 0;
841 	sg_policy->work_in_progress		= false;
842 	sg_policy->limits_changed		= false;
843 	sg_policy->need_freq_update		= false;
844 	sg_policy->cached_raw_freq		= 0;
845 
846 	for_each_cpu(cpu, policy->cpus) {
847 		struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
848 
849 		memset(sg_cpu, 0, sizeof(*sg_cpu));
850 		sg_cpu->cpu			= cpu;
851 		sg_cpu->sg_policy		= sg_policy;
852 	}
853 
854 	for_each_cpu(cpu, policy->cpus) {
855 		struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
856 
857 		cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util,
858 					     policy_is_shared(policy) ?
859 							sugov_update_shared :
860 							sugov_update_single);
861 	}
862 	return 0;
863 }
864 
865 static void sugov_stop(struct cpufreq_policy *policy)
866 {
867 	struct sugov_policy *sg_policy = policy->governor_data;
868 	unsigned int cpu;
869 
870 	for_each_cpu(cpu, policy->cpus)
871 		cpufreq_remove_update_util_hook(cpu);
872 
873 	synchronize_rcu();
874 
875 	if (!policy->fast_switch_enabled) {
876 		irq_work_sync(&sg_policy->irq_work);
877 		kthread_cancel_work_sync(&sg_policy->work);
878 	}
879 }
880 
881 static void sugov_limits(struct cpufreq_policy *policy)
882 {
883 	struct sugov_policy *sg_policy = policy->governor_data;
884 
885 	if (!policy->fast_switch_enabled) {
886 		mutex_lock(&sg_policy->work_lock);
887 		cpufreq_policy_apply_limits(policy);
888 		mutex_unlock(&sg_policy->work_lock);
889 	}
890 
891 	sg_policy->limits_changed = true;
892 }
893 
894 struct cpufreq_governor schedutil_gov = {
895 	.name			= "schedutil",
896 	.owner			= THIS_MODULE,
897 	.dynamic_switching	= true,
898 	.init			= sugov_init,
899 	.exit			= sugov_exit,
900 	.start			= sugov_start,
901 	.stop			= sugov_stop,
902 	.limits			= sugov_limits,
903 };
904 
905 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL
906 struct cpufreq_governor *cpufreq_default_governor(void)
907 {
908 	return &schedutil_gov;
909 }
910 #endif
911 
912 static int __init sugov_register(void)
913 {
914 	return cpufreq_register_governor(&schedutil_gov);
915 }
916 core_initcall(sugov_register);
917 
918 #ifdef CONFIG_ENERGY_MODEL
919 extern bool sched_energy_update;
920 extern struct mutex sched_energy_mutex;
921 
922 static void rebuild_sd_workfn(struct work_struct *work)
923 {
924 	mutex_lock(&sched_energy_mutex);
925 	sched_energy_update = true;
926 	rebuild_sched_domains();
927 	sched_energy_update = false;
928 	mutex_unlock(&sched_energy_mutex);
929 }
930 static DECLARE_WORK(rebuild_sd_work, rebuild_sd_workfn);
931 
932 /*
933  * EAS shouldn't be attempted without sugov, so rebuild the sched_domains
934  * on governor changes to make sure the scheduler knows about it.
935  */
936 void sched_cpufreq_governor_change(struct cpufreq_policy *policy,
937 				  struct cpufreq_governor *old_gov)
938 {
939 	if (old_gov == &schedutil_gov || policy->governor == &schedutil_gov) {
940 		/*
941 		 * When called from the cpufreq_register_driver() path, the
942 		 * cpu_hotplug_lock is already held, so use a work item to
943 		 * avoid nested locking in rebuild_sched_domains().
944 		 */
945 		schedule_work(&rebuild_sd_work);
946 	}
947 
948 }
949 #endif
950