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->need_freq_update) {
106 		if (sg_policy->next_freq == next_freq)
107 			return false;
108 	} else {
109 		sg_policy->need_freq_update = cpufreq_driver_test_flags(CPUFREQ_NEED_UPDATE_LIMITS);
110 	}
111 
112 	sg_policy->next_freq = next_freq;
113 	sg_policy->last_freq_update_time = time;
114 
115 	return true;
116 }
117 
118 static void sugov_fast_switch(struct sugov_policy *sg_policy, u64 time,
119 			      unsigned int next_freq)
120 {
121 	if (sugov_update_next_freq(sg_policy, time, next_freq))
122 		cpufreq_driver_fast_switch(sg_policy->policy, next_freq);
123 }
124 
125 static void sugov_deferred_update(struct sugov_policy *sg_policy, u64 time,
126 				  unsigned int next_freq)
127 {
128 	if (!sugov_update_next_freq(sg_policy, time, next_freq))
129 		return;
130 
131 	if (!sg_policy->work_in_progress) {
132 		sg_policy->work_in_progress = true;
133 		irq_work_queue(&sg_policy->irq_work);
134 	}
135 }
136 
137 /**
138  * get_next_freq - Compute a new frequency for a given cpufreq policy.
139  * @sg_policy: schedutil policy object to compute the new frequency for.
140  * @util: Current CPU utilization.
141  * @max: CPU capacity.
142  *
143  * If the utilization is frequency-invariant, choose the new frequency to be
144  * proportional to it, that is
145  *
146  * next_freq = C * max_freq * util / max
147  *
148  * Otherwise, approximate the would-be frequency-invariant utilization by
149  * util_raw * (curr_freq / max_freq) which leads to
150  *
151  * next_freq = C * curr_freq * util_raw / max
152  *
153  * Take C = 1.25 for the frequency tipping point at (util / max) = 0.8.
154  *
155  * The lowest driver-supported frequency which is equal or greater than the raw
156  * next_freq (as calculated above) is returned, subject to policy min/max and
157  * cpufreq driver limitations.
158  */
159 static unsigned int get_next_freq(struct sugov_policy *sg_policy,
160 				  unsigned long util, unsigned long max)
161 {
162 	struct cpufreq_policy *policy = sg_policy->policy;
163 	unsigned int freq = arch_scale_freq_invariant() ?
164 				policy->cpuinfo.max_freq : policy->cur;
165 
166 	freq = map_util_freq(util, freq, max);
167 
168 	if (freq == sg_policy->cached_raw_freq && !sg_policy->need_freq_update)
169 		return sg_policy->next_freq;
170 
171 	sg_policy->cached_raw_freq = freq;
172 	return cpufreq_driver_resolve_freq(policy, freq);
173 }
174 
175 /*
176  * This function computes an effective utilization for the given CPU, to be
177  * used for frequency selection given the linear relation: f = u * f_max.
178  *
179  * The scheduler tracks the following metrics:
180  *
181  *   cpu_util_{cfs,rt,dl,irq}()
182  *   cpu_bw_dl()
183  *
184  * Where the cfs,rt and dl util numbers are tracked with the same metric and
185  * synchronized windows and are thus directly comparable.
186  *
187  * The cfs,rt,dl utilization are the running times measured with rq->clock_task
188  * which excludes things like IRQ and steal-time. These latter are then accrued
189  * in the irq utilization.
190  *
191  * The DL bandwidth number otoh is not a measured metric but a value computed
192  * based on the task model parameters and gives the minimal utilization
193  * required to meet deadlines.
194  */
195 unsigned long schedutil_cpu_util(int cpu, unsigned long util_cfs,
196 				 unsigned long max, enum schedutil_type type,
197 				 struct task_struct *p)
198 {
199 	unsigned long dl_util, util, irq;
200 	struct rq *rq = cpu_rq(cpu);
201 
202 	if (!uclamp_is_used() &&
203 	    type == FREQUENCY_UTIL && rt_rq_is_runnable(&rq->rt)) {
204 		return max;
205 	}
206 
207 	/*
208 	 * Early check to see if IRQ/steal time saturates the CPU, can be
209 	 * because of inaccuracies in how we track these -- see
210 	 * update_irq_load_avg().
211 	 */
212 	irq = cpu_util_irq(rq);
213 	if (unlikely(irq >= max))
214 		return max;
215 
216 	/*
217 	 * Because the time spend on RT/DL tasks is visible as 'lost' time to
218 	 * CFS tasks and we use the same metric to track the effective
219 	 * utilization (PELT windows are synchronized) we can directly add them
220 	 * to obtain the CPU's actual utilization.
221 	 *
222 	 * CFS and RT utilization can be boosted or capped, depending on
223 	 * utilization clamp constraints requested by currently RUNNABLE
224 	 * tasks.
225 	 * When there are no CFS RUNNABLE tasks, clamps are released and
226 	 * frequency will be gracefully reduced with the utilization decay.
227 	 */
228 	util = util_cfs + cpu_util_rt(rq);
229 	if (type == FREQUENCY_UTIL)
230 		util = uclamp_rq_util_with(rq, util, p);
231 
232 	dl_util = cpu_util_dl(rq);
233 
234 	/*
235 	 * For frequency selection we do not make cpu_util_dl() a permanent part
236 	 * of this sum because we want to use cpu_bw_dl() later on, but we need
237 	 * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
238 	 * that we select f_max when there is no idle time.
239 	 *
240 	 * NOTE: numerical errors or stop class might cause us to not quite hit
241 	 * saturation when we should -- something for later.
242 	 */
243 	if (util + dl_util >= max)
244 		return max;
245 
246 	/*
247 	 * OTOH, for energy computation we need the estimated running time, so
248 	 * include util_dl and ignore dl_bw.
249 	 */
250 	if (type == ENERGY_UTIL)
251 		util += dl_util;
252 
253 	/*
254 	 * There is still idle time; further improve the number by using the
255 	 * irq metric. Because IRQ/steal time is hidden from the task clock we
256 	 * need to scale the task numbers:
257 	 *
258 	 *              max - irq
259 	 *   U' = irq + --------- * U
260 	 *                 max
261 	 */
262 	util = scale_irq_capacity(util, irq, max);
263 	util += irq;
264 
265 	/*
266 	 * Bandwidth required by DEADLINE must always be granted while, for
267 	 * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
268 	 * to gracefully reduce the frequency when no tasks show up for longer
269 	 * periods of time.
270 	 *
271 	 * Ideally we would like to set bw_dl as min/guaranteed freq and util +
272 	 * bw_dl as requested freq. However, cpufreq is not yet ready for such
273 	 * an interface. So, we only do the latter for now.
274 	 */
275 	if (type == FREQUENCY_UTIL)
276 		util += cpu_bw_dl(rq);
277 
278 	return min(max, util);
279 }
280 
281 static unsigned long sugov_get_util(struct sugov_cpu *sg_cpu)
282 {
283 	struct rq *rq = cpu_rq(sg_cpu->cpu);
284 	unsigned long util = cpu_util_cfs(rq);
285 	unsigned long max = arch_scale_cpu_capacity(sg_cpu->cpu);
286 
287 	sg_cpu->max = max;
288 	sg_cpu->bw_dl = cpu_bw_dl(rq);
289 
290 	return schedutil_cpu_util(sg_cpu->cpu, util, max, FREQUENCY_UTIL, NULL);
291 }
292 
293 /**
294  * sugov_iowait_reset() - Reset the IO boost status of a CPU.
295  * @sg_cpu: the sugov data for the CPU to boost
296  * @time: the update time from the caller
297  * @set_iowait_boost: true if an IO boost has been requested
298  *
299  * The IO wait boost of a task is disabled after a tick since the last update
300  * of a CPU. If a new IO wait boost is requested after more then a tick, then
301  * we enable the boost starting from IOWAIT_BOOST_MIN, which improves energy
302  * efficiency by ignoring sporadic wakeups from IO.
303  */
304 static bool sugov_iowait_reset(struct sugov_cpu *sg_cpu, u64 time,
305 			       bool set_iowait_boost)
306 {
307 	s64 delta_ns = time - sg_cpu->last_update;
308 
309 	/* Reset boost only if a tick has elapsed since last request */
310 	if (delta_ns <= TICK_NSEC)
311 		return false;
312 
313 	sg_cpu->iowait_boost = set_iowait_boost ? IOWAIT_BOOST_MIN : 0;
314 	sg_cpu->iowait_boost_pending = set_iowait_boost;
315 
316 	return true;
317 }
318 
319 /**
320  * sugov_iowait_boost() - Updates the IO boost status of a CPU.
321  * @sg_cpu: the sugov data for the CPU to boost
322  * @time: the update time from the caller
323  * @flags: SCHED_CPUFREQ_IOWAIT if the task is waking up after an IO wait
324  *
325  * Each time a task wakes up after an IO operation, the CPU utilization can be
326  * boosted to a certain utilization which doubles at each "frequent and
327  * successive" wakeup from IO, ranging from IOWAIT_BOOST_MIN to the utilization
328  * of the maximum OPP.
329  *
330  * To keep doubling, an IO boost has to be requested at least once per tick,
331  * otherwise we restart from the utilization of the minimum OPP.
332  */
333 static void sugov_iowait_boost(struct sugov_cpu *sg_cpu, u64 time,
334 			       unsigned int flags)
335 {
336 	bool set_iowait_boost = flags & SCHED_CPUFREQ_IOWAIT;
337 
338 	/* Reset boost if the CPU appears to have been idle enough */
339 	if (sg_cpu->iowait_boost &&
340 	    sugov_iowait_reset(sg_cpu, time, set_iowait_boost))
341 		return;
342 
343 	/* Boost only tasks waking up after IO */
344 	if (!set_iowait_boost)
345 		return;
346 
347 	/* Ensure boost doubles only one time at each request */
348 	if (sg_cpu->iowait_boost_pending)
349 		return;
350 	sg_cpu->iowait_boost_pending = true;
351 
352 	/* Double the boost at each request */
353 	if (sg_cpu->iowait_boost) {
354 		sg_cpu->iowait_boost =
355 			min_t(unsigned int, sg_cpu->iowait_boost << 1, SCHED_CAPACITY_SCALE);
356 		return;
357 	}
358 
359 	/* First wakeup after IO: start with minimum boost */
360 	sg_cpu->iowait_boost = IOWAIT_BOOST_MIN;
361 }
362 
363 /**
364  * sugov_iowait_apply() - Apply the IO boost to a CPU.
365  * @sg_cpu: the sugov data for the cpu to boost
366  * @time: the update time from the caller
367  * @util: the utilization to (eventually) boost
368  * @max: the maximum value the utilization can be boosted to
369  *
370  * A CPU running a task which woken up after an IO operation can have its
371  * utilization boosted to speed up the completion of those IO operations.
372  * The IO boost value is increased each time a task wakes up from IO, in
373  * sugov_iowait_apply(), and it's instead decreased by this function,
374  * each time an increase has not been requested (!iowait_boost_pending).
375  *
376  * A CPU which also appears to have been idle for at least one tick has also
377  * its IO boost utilization reset.
378  *
379  * This mechanism is designed to boost high frequently IO waiting tasks, while
380  * being more conservative on tasks which does sporadic IO operations.
381  */
382 static unsigned long sugov_iowait_apply(struct sugov_cpu *sg_cpu, u64 time,
383 					unsigned long util, unsigned long max)
384 {
385 	unsigned long boost;
386 
387 	/* No boost currently required */
388 	if (!sg_cpu->iowait_boost)
389 		return util;
390 
391 	/* Reset boost if the CPU appears to have been idle enough */
392 	if (sugov_iowait_reset(sg_cpu, time, false))
393 		return util;
394 
395 	if (!sg_cpu->iowait_boost_pending) {
396 		/*
397 		 * No boost pending; reduce the boost value.
398 		 */
399 		sg_cpu->iowait_boost >>= 1;
400 		if (sg_cpu->iowait_boost < IOWAIT_BOOST_MIN) {
401 			sg_cpu->iowait_boost = 0;
402 			return util;
403 		}
404 	}
405 
406 	sg_cpu->iowait_boost_pending = false;
407 
408 	/*
409 	 * @util is already in capacity scale; convert iowait_boost
410 	 * into the same scale so we can compare.
411 	 */
412 	boost = (sg_cpu->iowait_boost * max) >> SCHED_CAPACITY_SHIFT;
413 	return max(boost, util);
414 }
415 
416 #ifdef CONFIG_NO_HZ_COMMON
417 static bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu)
418 {
419 	unsigned long idle_calls = tick_nohz_get_idle_calls_cpu(sg_cpu->cpu);
420 	bool ret = idle_calls == sg_cpu->saved_idle_calls;
421 
422 	sg_cpu->saved_idle_calls = idle_calls;
423 	return ret;
424 }
425 #else
426 static inline bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu) { return false; }
427 #endif /* CONFIG_NO_HZ_COMMON */
428 
429 /*
430  * Make sugov_should_update_freq() ignore the rate limit when DL
431  * has increased the utilization.
432  */
433 static inline void ignore_dl_rate_limit(struct sugov_cpu *sg_cpu, struct sugov_policy *sg_policy)
434 {
435 	if (cpu_bw_dl(cpu_rq(sg_cpu->cpu)) > sg_cpu->bw_dl)
436 		sg_policy->limits_changed = true;
437 }
438 
439 static void sugov_update_single(struct update_util_data *hook, u64 time,
440 				unsigned int flags)
441 {
442 	struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
443 	struct sugov_policy *sg_policy = sg_cpu->sg_policy;
444 	unsigned long util, max;
445 	unsigned int next_f;
446 	unsigned int cached_freq = sg_policy->cached_raw_freq;
447 
448 	sugov_iowait_boost(sg_cpu, time, flags);
449 	sg_cpu->last_update = time;
450 
451 	ignore_dl_rate_limit(sg_cpu, sg_policy);
452 
453 	if (!sugov_should_update_freq(sg_policy, time))
454 		return;
455 
456 	util = sugov_get_util(sg_cpu);
457 	max = sg_cpu->max;
458 	util = sugov_iowait_apply(sg_cpu, time, util, max);
459 	next_f = get_next_freq(sg_policy, util, max);
460 	/*
461 	 * Do not reduce the frequency if the CPU has not been idle
462 	 * recently, as the reduction is likely to be premature then.
463 	 */
464 	if (sugov_cpu_is_busy(sg_cpu) && next_f < sg_policy->next_freq) {
465 		next_f = sg_policy->next_freq;
466 
467 		/* Restore cached freq as next_freq has changed */
468 		sg_policy->cached_raw_freq = cached_freq;
469 	}
470 
471 	/*
472 	 * This code runs under rq->lock for the target CPU, so it won't run
473 	 * concurrently on two different CPUs for the same target and it is not
474 	 * necessary to acquire the lock in the fast switch case.
475 	 */
476 	if (sg_policy->policy->fast_switch_enabled) {
477 		sugov_fast_switch(sg_policy, time, next_f);
478 	} else {
479 		raw_spin_lock(&sg_policy->update_lock);
480 		sugov_deferred_update(sg_policy, time, next_f);
481 		raw_spin_unlock(&sg_policy->update_lock);
482 	}
483 }
484 
485 static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu, u64 time)
486 {
487 	struct sugov_policy *sg_policy = sg_cpu->sg_policy;
488 	struct cpufreq_policy *policy = sg_policy->policy;
489 	unsigned long util = 0, max = 1;
490 	unsigned int j;
491 
492 	for_each_cpu(j, policy->cpus) {
493 		struct sugov_cpu *j_sg_cpu = &per_cpu(sugov_cpu, j);
494 		unsigned long j_util, j_max;
495 
496 		j_util = sugov_get_util(j_sg_cpu);
497 		j_max = j_sg_cpu->max;
498 		j_util = sugov_iowait_apply(j_sg_cpu, time, j_util, j_max);
499 
500 		if (j_util * max > j_max * util) {
501 			util = j_util;
502 			max = j_max;
503 		}
504 	}
505 
506 	return get_next_freq(sg_policy, util, max);
507 }
508 
509 static void
510 sugov_update_shared(struct update_util_data *hook, u64 time, unsigned int flags)
511 {
512 	struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
513 	struct sugov_policy *sg_policy = sg_cpu->sg_policy;
514 	unsigned int next_f;
515 
516 	raw_spin_lock(&sg_policy->update_lock);
517 
518 	sugov_iowait_boost(sg_cpu, time, flags);
519 	sg_cpu->last_update = time;
520 
521 	ignore_dl_rate_limit(sg_cpu, sg_policy);
522 
523 	if (sugov_should_update_freq(sg_policy, time)) {
524 		next_f = sugov_next_freq_shared(sg_cpu, time);
525 
526 		if (sg_policy->policy->fast_switch_enabled)
527 			sugov_fast_switch(sg_policy, time, next_f);
528 		else
529 			sugov_deferred_update(sg_policy, time, next_f);
530 	}
531 
532 	raw_spin_unlock(&sg_policy->update_lock);
533 }
534 
535 static void sugov_work(struct kthread_work *work)
536 {
537 	struct sugov_policy *sg_policy = container_of(work, struct sugov_policy, work);
538 	unsigned int freq;
539 	unsigned long flags;
540 
541 	/*
542 	 * Hold sg_policy->update_lock shortly to handle the case where:
543 	 * incase sg_policy->next_freq is read here, and then updated by
544 	 * sugov_deferred_update() just before work_in_progress is set to false
545 	 * here, we may miss queueing the new update.
546 	 *
547 	 * Note: If a work was queued after the update_lock is released,
548 	 * sugov_work() will just be called again by kthread_work code; and the
549 	 * request will be proceed before the sugov thread sleeps.
550 	 */
551 	raw_spin_lock_irqsave(&sg_policy->update_lock, flags);
552 	freq = sg_policy->next_freq;
553 	sg_policy->work_in_progress = false;
554 	raw_spin_unlock_irqrestore(&sg_policy->update_lock, flags);
555 
556 	mutex_lock(&sg_policy->work_lock);
557 	__cpufreq_driver_target(sg_policy->policy, freq, CPUFREQ_RELATION_L);
558 	mutex_unlock(&sg_policy->work_lock);
559 }
560 
561 static void sugov_irq_work(struct irq_work *irq_work)
562 {
563 	struct sugov_policy *sg_policy;
564 
565 	sg_policy = container_of(irq_work, struct sugov_policy, irq_work);
566 
567 	kthread_queue_work(&sg_policy->worker, &sg_policy->work);
568 }
569 
570 /************************** sysfs interface ************************/
571 
572 static struct sugov_tunables *global_tunables;
573 static DEFINE_MUTEX(global_tunables_lock);
574 
575 static inline struct sugov_tunables *to_sugov_tunables(struct gov_attr_set *attr_set)
576 {
577 	return container_of(attr_set, struct sugov_tunables, attr_set);
578 }
579 
580 static ssize_t rate_limit_us_show(struct gov_attr_set *attr_set, char *buf)
581 {
582 	struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
583 
584 	return sprintf(buf, "%u\n", tunables->rate_limit_us);
585 }
586 
587 static ssize_t
588 rate_limit_us_store(struct gov_attr_set *attr_set, const char *buf, size_t count)
589 {
590 	struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
591 	struct sugov_policy *sg_policy;
592 	unsigned int rate_limit_us;
593 
594 	if (kstrtouint(buf, 10, &rate_limit_us))
595 		return -EINVAL;
596 
597 	tunables->rate_limit_us = rate_limit_us;
598 
599 	list_for_each_entry(sg_policy, &attr_set->policy_list, tunables_hook)
600 		sg_policy->freq_update_delay_ns = rate_limit_us * NSEC_PER_USEC;
601 
602 	return count;
603 }
604 
605 static struct governor_attr rate_limit_us = __ATTR_RW(rate_limit_us);
606 
607 static struct attribute *sugov_attrs[] = {
608 	&rate_limit_us.attr,
609 	NULL
610 };
611 ATTRIBUTE_GROUPS(sugov);
612 
613 static struct kobj_type sugov_tunables_ktype = {
614 	.default_groups = sugov_groups,
615 	.sysfs_ops = &governor_sysfs_ops,
616 };
617 
618 /********************** cpufreq governor interface *********************/
619 
620 struct cpufreq_governor schedutil_gov;
621 
622 static struct sugov_policy *sugov_policy_alloc(struct cpufreq_policy *policy)
623 {
624 	struct sugov_policy *sg_policy;
625 
626 	sg_policy = kzalloc(sizeof(*sg_policy), GFP_KERNEL);
627 	if (!sg_policy)
628 		return NULL;
629 
630 	sg_policy->policy = policy;
631 	raw_spin_lock_init(&sg_policy->update_lock);
632 	return sg_policy;
633 }
634 
635 static void sugov_policy_free(struct sugov_policy *sg_policy)
636 {
637 	kfree(sg_policy);
638 }
639 
640 static int sugov_kthread_create(struct sugov_policy *sg_policy)
641 {
642 	struct task_struct *thread;
643 	struct sched_attr attr = {
644 		.size		= sizeof(struct sched_attr),
645 		.sched_policy	= SCHED_DEADLINE,
646 		.sched_flags	= SCHED_FLAG_SUGOV,
647 		.sched_nice	= 0,
648 		.sched_priority	= 0,
649 		/*
650 		 * Fake (unused) bandwidth; workaround to "fix"
651 		 * priority inheritance.
652 		 */
653 		.sched_runtime	=  1000000,
654 		.sched_deadline = 10000000,
655 		.sched_period	= 10000000,
656 	};
657 	struct cpufreq_policy *policy = sg_policy->policy;
658 	int ret;
659 
660 	/* kthread only required for slow path */
661 	if (policy->fast_switch_enabled)
662 		return 0;
663 
664 	kthread_init_work(&sg_policy->work, sugov_work);
665 	kthread_init_worker(&sg_policy->worker);
666 	thread = kthread_create(kthread_worker_fn, &sg_policy->worker,
667 				"sugov:%d",
668 				cpumask_first(policy->related_cpus));
669 	if (IS_ERR(thread)) {
670 		pr_err("failed to create sugov thread: %ld\n", PTR_ERR(thread));
671 		return PTR_ERR(thread);
672 	}
673 
674 	ret = sched_setattr_nocheck(thread, &attr);
675 	if (ret) {
676 		kthread_stop(thread);
677 		pr_warn("%s: failed to set SCHED_DEADLINE\n", __func__);
678 		return ret;
679 	}
680 
681 	sg_policy->thread = thread;
682 	kthread_bind_mask(thread, policy->related_cpus);
683 	init_irq_work(&sg_policy->irq_work, sugov_irq_work);
684 	mutex_init(&sg_policy->work_lock);
685 
686 	wake_up_process(thread);
687 
688 	return 0;
689 }
690 
691 static void sugov_kthread_stop(struct sugov_policy *sg_policy)
692 {
693 	/* kthread only required for slow path */
694 	if (sg_policy->policy->fast_switch_enabled)
695 		return;
696 
697 	kthread_flush_worker(&sg_policy->worker);
698 	kthread_stop(sg_policy->thread);
699 	mutex_destroy(&sg_policy->work_lock);
700 }
701 
702 static struct sugov_tunables *sugov_tunables_alloc(struct sugov_policy *sg_policy)
703 {
704 	struct sugov_tunables *tunables;
705 
706 	tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
707 	if (tunables) {
708 		gov_attr_set_init(&tunables->attr_set, &sg_policy->tunables_hook);
709 		if (!have_governor_per_policy())
710 			global_tunables = tunables;
711 	}
712 	return tunables;
713 }
714 
715 static void sugov_tunables_free(struct sugov_tunables *tunables)
716 {
717 	if (!have_governor_per_policy())
718 		global_tunables = NULL;
719 
720 	kfree(tunables);
721 }
722 
723 static int sugov_init(struct cpufreq_policy *policy)
724 {
725 	struct sugov_policy *sg_policy;
726 	struct sugov_tunables *tunables;
727 	int ret = 0;
728 
729 	/* State should be equivalent to EXIT */
730 	if (policy->governor_data)
731 		return -EBUSY;
732 
733 	cpufreq_enable_fast_switch(policy);
734 
735 	sg_policy = sugov_policy_alloc(policy);
736 	if (!sg_policy) {
737 		ret = -ENOMEM;
738 		goto disable_fast_switch;
739 	}
740 
741 	ret = sugov_kthread_create(sg_policy);
742 	if (ret)
743 		goto free_sg_policy;
744 
745 	mutex_lock(&global_tunables_lock);
746 
747 	if (global_tunables) {
748 		if (WARN_ON(have_governor_per_policy())) {
749 			ret = -EINVAL;
750 			goto stop_kthread;
751 		}
752 		policy->governor_data = sg_policy;
753 		sg_policy->tunables = global_tunables;
754 
755 		gov_attr_set_get(&global_tunables->attr_set, &sg_policy->tunables_hook);
756 		goto out;
757 	}
758 
759 	tunables = sugov_tunables_alloc(sg_policy);
760 	if (!tunables) {
761 		ret = -ENOMEM;
762 		goto stop_kthread;
763 	}
764 
765 	tunables->rate_limit_us = cpufreq_policy_transition_delay_us(policy);
766 
767 	policy->governor_data = sg_policy;
768 	sg_policy->tunables = tunables;
769 
770 	ret = kobject_init_and_add(&tunables->attr_set.kobj, &sugov_tunables_ktype,
771 				   get_governor_parent_kobj(policy), "%s",
772 				   schedutil_gov.name);
773 	if (ret)
774 		goto fail;
775 
776 out:
777 	mutex_unlock(&global_tunables_lock);
778 	return 0;
779 
780 fail:
781 	kobject_put(&tunables->attr_set.kobj);
782 	policy->governor_data = NULL;
783 	sugov_tunables_free(tunables);
784 
785 stop_kthread:
786 	sugov_kthread_stop(sg_policy);
787 	mutex_unlock(&global_tunables_lock);
788 
789 free_sg_policy:
790 	sugov_policy_free(sg_policy);
791 
792 disable_fast_switch:
793 	cpufreq_disable_fast_switch(policy);
794 
795 	pr_err("initialization failed (error %d)\n", ret);
796 	return ret;
797 }
798 
799 static void sugov_exit(struct cpufreq_policy *policy)
800 {
801 	struct sugov_policy *sg_policy = policy->governor_data;
802 	struct sugov_tunables *tunables = sg_policy->tunables;
803 	unsigned int count;
804 
805 	mutex_lock(&global_tunables_lock);
806 
807 	count = gov_attr_set_put(&tunables->attr_set, &sg_policy->tunables_hook);
808 	policy->governor_data = NULL;
809 	if (!count)
810 		sugov_tunables_free(tunables);
811 
812 	mutex_unlock(&global_tunables_lock);
813 
814 	sugov_kthread_stop(sg_policy);
815 	sugov_policy_free(sg_policy);
816 	cpufreq_disable_fast_switch(policy);
817 }
818 
819 static int sugov_start(struct cpufreq_policy *policy)
820 {
821 	struct sugov_policy *sg_policy = policy->governor_data;
822 	unsigned int cpu;
823 
824 	sg_policy->freq_update_delay_ns	= sg_policy->tunables->rate_limit_us * NSEC_PER_USEC;
825 	sg_policy->last_freq_update_time	= 0;
826 	sg_policy->next_freq			= 0;
827 	sg_policy->work_in_progress		= false;
828 	sg_policy->limits_changed		= false;
829 	sg_policy->cached_raw_freq		= 0;
830 
831 	sg_policy->need_freq_update = cpufreq_driver_test_flags(CPUFREQ_NEED_UPDATE_LIMITS);
832 
833 	for_each_cpu(cpu, policy->cpus) {
834 		struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
835 
836 		memset(sg_cpu, 0, sizeof(*sg_cpu));
837 		sg_cpu->cpu			= cpu;
838 		sg_cpu->sg_policy		= sg_policy;
839 	}
840 
841 	for_each_cpu(cpu, policy->cpus) {
842 		struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
843 
844 		cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util,
845 					     policy_is_shared(policy) ?
846 							sugov_update_shared :
847 							sugov_update_single);
848 	}
849 	return 0;
850 }
851 
852 static void sugov_stop(struct cpufreq_policy *policy)
853 {
854 	struct sugov_policy *sg_policy = policy->governor_data;
855 	unsigned int cpu;
856 
857 	for_each_cpu(cpu, policy->cpus)
858 		cpufreq_remove_update_util_hook(cpu);
859 
860 	synchronize_rcu();
861 
862 	if (!policy->fast_switch_enabled) {
863 		irq_work_sync(&sg_policy->irq_work);
864 		kthread_cancel_work_sync(&sg_policy->work);
865 	}
866 }
867 
868 static void sugov_limits(struct cpufreq_policy *policy)
869 {
870 	struct sugov_policy *sg_policy = policy->governor_data;
871 
872 	if (!policy->fast_switch_enabled) {
873 		mutex_lock(&sg_policy->work_lock);
874 		cpufreq_policy_apply_limits(policy);
875 		mutex_unlock(&sg_policy->work_lock);
876 	}
877 
878 	sg_policy->limits_changed = true;
879 }
880 
881 struct cpufreq_governor schedutil_gov = {
882 	.name			= "schedutil",
883 	.owner			= THIS_MODULE,
884 	.flags			= CPUFREQ_GOV_DYNAMIC_SWITCHING,
885 	.init			= sugov_init,
886 	.exit			= sugov_exit,
887 	.start			= sugov_start,
888 	.stop			= sugov_stop,
889 	.limits			= sugov_limits,
890 };
891 
892 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL
893 struct cpufreq_governor *cpufreq_default_governor(void)
894 {
895 	return &schedutil_gov;
896 }
897 #endif
898 
899 cpufreq_governor_init(schedutil_gov);
900 
901 #ifdef CONFIG_ENERGY_MODEL
902 extern bool sched_energy_update;
903 extern struct mutex sched_energy_mutex;
904 
905 static void rebuild_sd_workfn(struct work_struct *work)
906 {
907 	mutex_lock(&sched_energy_mutex);
908 	sched_energy_update = true;
909 	rebuild_sched_domains();
910 	sched_energy_update = false;
911 	mutex_unlock(&sched_energy_mutex);
912 }
913 static DECLARE_WORK(rebuild_sd_work, rebuild_sd_workfn);
914 
915 /*
916  * EAS shouldn't be attempted without sugov, so rebuild the sched_domains
917  * on governor changes to make sure the scheduler knows about it.
918  */
919 void sched_cpufreq_governor_change(struct cpufreq_policy *policy,
920 				  struct cpufreq_governor *old_gov)
921 {
922 	if (old_gov == &schedutil_gov || policy->governor == &schedutil_gov) {
923 		/*
924 		 * When called from the cpufreq_register_driver() path, the
925 		 * cpu_hotplug_lock is already held, so use a work item to
926 		 * avoid nested locking in rebuild_sched_domains().
927 		 */
928 		schedule_work(&rebuild_sd_work);
929 	}
930 
931 }
932 #endif
933