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 	/* The next fields are only needed if fast switch cannot be used: */
29 	struct			irq_work irq_work;
30 	struct			kthread_work work;
31 	struct			mutex work_lock;
32 	struct			kthread_worker worker;
33 	struct task_struct	*thread;
34 	bool			work_in_progress;
35 
36 	bool			limits_changed;
37 	bool			need_freq_update;
38 };
39 
40 struct sugov_cpu {
41 	struct update_util_data	update_util;
42 	struct sugov_policy	*sg_policy;
43 	unsigned int		cpu;
44 
45 	bool			iowait_boost_pending;
46 	unsigned int		iowait_boost;
47 	u64			last_update;
48 
49 	unsigned long		util;
50 	unsigned long		bw_dl;
51 
52 	/* The field below is for single-CPU policies only: */
53 #ifdef CONFIG_NO_HZ_COMMON
54 	unsigned long		saved_idle_calls;
55 #endif
56 };
57 
58 static DEFINE_PER_CPU(struct sugov_cpu, sugov_cpu);
59 
60 /************************ Governor internals ***********************/
61 
62 static bool sugov_should_update_freq(struct sugov_policy *sg_policy, u64 time)
63 {
64 	s64 delta_ns;
65 
66 	/*
67 	 * Since cpufreq_update_util() is called with rq->lock held for
68 	 * the @target_cpu, our per-CPU data is fully serialized.
69 	 *
70 	 * However, drivers cannot in general deal with cross-CPU
71 	 * requests, so while get_next_freq() will work, our
72 	 * sugov_update_commit() call may not for the fast switching platforms.
73 	 *
74 	 * Hence stop here for remote requests if they aren't supported
75 	 * by the hardware, as calculating the frequency is pointless if
76 	 * we cannot in fact act on it.
77 	 *
78 	 * This is needed on the slow switching platforms too to prevent CPUs
79 	 * going offline from leaving stale IRQ work items behind.
80 	 */
81 	if (!cpufreq_this_cpu_can_update(sg_policy->policy))
82 		return false;
83 
84 	if (unlikely(sg_policy->limits_changed)) {
85 		sg_policy->limits_changed = false;
86 		sg_policy->need_freq_update = true;
87 		return true;
88 	}
89 
90 	delta_ns = time - sg_policy->last_freq_update_time;
91 
92 	return delta_ns >= sg_policy->freq_update_delay_ns;
93 }
94 
95 static bool sugov_update_next_freq(struct sugov_policy *sg_policy, u64 time,
96 				   unsigned int next_freq)
97 {
98 	if (sg_policy->need_freq_update)
99 		sg_policy->need_freq_update = cpufreq_driver_test_flags(CPUFREQ_NEED_UPDATE_LIMITS);
100 	else if (sg_policy->next_freq == next_freq)
101 		return false;
102 
103 	sg_policy->next_freq = next_freq;
104 	sg_policy->last_freq_update_time = time;
105 
106 	return true;
107 }
108 
109 static void sugov_deferred_update(struct sugov_policy *sg_policy)
110 {
111 	if (!sg_policy->work_in_progress) {
112 		sg_policy->work_in_progress = true;
113 		irq_work_queue(&sg_policy->irq_work);
114 	}
115 }
116 
117 /**
118  * get_next_freq - Compute a new frequency for a given cpufreq policy.
119  * @sg_policy: schedutil policy object to compute the new frequency for.
120  * @util: Current CPU utilization.
121  * @max: CPU capacity.
122  *
123  * If the utilization is frequency-invariant, choose the new frequency to be
124  * proportional to it, that is
125  *
126  * next_freq = C * max_freq * util / max
127  *
128  * Otherwise, approximate the would-be frequency-invariant utilization by
129  * util_raw * (curr_freq / max_freq) which leads to
130  *
131  * next_freq = C * curr_freq * util_raw / max
132  *
133  * Take C = 1.25 for the frequency tipping point at (util / max) = 0.8.
134  *
135  * The lowest driver-supported frequency which is equal or greater than the raw
136  * next_freq (as calculated above) is returned, subject to policy min/max and
137  * cpufreq driver limitations.
138  */
139 static unsigned int get_next_freq(struct sugov_policy *sg_policy,
140 				  unsigned long util, unsigned long max)
141 {
142 	struct cpufreq_policy *policy = sg_policy->policy;
143 	unsigned int freq = arch_scale_freq_invariant() ?
144 				policy->cpuinfo.max_freq : policy->cur;
145 
146 	util = map_util_perf(util);
147 	freq = map_util_freq(util, freq, max);
148 
149 	if (freq == sg_policy->cached_raw_freq && !sg_policy->need_freq_update)
150 		return sg_policy->next_freq;
151 
152 	sg_policy->cached_raw_freq = freq;
153 	return cpufreq_driver_resolve_freq(policy, freq);
154 }
155 
156 static void sugov_get_util(struct sugov_cpu *sg_cpu)
157 {
158 	struct rq *rq = cpu_rq(sg_cpu->cpu);
159 
160 	sg_cpu->bw_dl = cpu_bw_dl(rq);
161 	sg_cpu->util = effective_cpu_util(sg_cpu->cpu, cpu_util_cfs(sg_cpu->cpu),
162 					  FREQUENCY_UTIL, NULL);
163 }
164 
165 /**
166  * sugov_iowait_reset() - Reset the IO boost status of a CPU.
167  * @sg_cpu: the sugov data for the CPU to boost
168  * @time: the update time from the caller
169  * @set_iowait_boost: true if an IO boost has been requested
170  *
171  * The IO wait boost of a task is disabled after a tick since the last update
172  * of a CPU. If a new IO wait boost is requested after more then a tick, then
173  * we enable the boost starting from IOWAIT_BOOST_MIN, which improves energy
174  * efficiency by ignoring sporadic wakeups from IO.
175  */
176 static bool sugov_iowait_reset(struct sugov_cpu *sg_cpu, u64 time,
177 			       bool set_iowait_boost)
178 {
179 	s64 delta_ns = time - sg_cpu->last_update;
180 
181 	/* Reset boost only if a tick has elapsed since last request */
182 	if (delta_ns <= TICK_NSEC)
183 		return false;
184 
185 	sg_cpu->iowait_boost = set_iowait_boost ? IOWAIT_BOOST_MIN : 0;
186 	sg_cpu->iowait_boost_pending = set_iowait_boost;
187 
188 	return true;
189 }
190 
191 /**
192  * sugov_iowait_boost() - Updates the IO boost status of a CPU.
193  * @sg_cpu: the sugov data for the CPU to boost
194  * @time: the update time from the caller
195  * @flags: SCHED_CPUFREQ_IOWAIT if the task is waking up after an IO wait
196  *
197  * Each time a task wakes up after an IO operation, the CPU utilization can be
198  * boosted to a certain utilization which doubles at each "frequent and
199  * successive" wakeup from IO, ranging from IOWAIT_BOOST_MIN to the utilization
200  * of the maximum OPP.
201  *
202  * To keep doubling, an IO boost has to be requested at least once per tick,
203  * otherwise we restart from the utilization of the minimum OPP.
204  */
205 static void sugov_iowait_boost(struct sugov_cpu *sg_cpu, u64 time,
206 			       unsigned int flags)
207 {
208 	bool set_iowait_boost = flags & SCHED_CPUFREQ_IOWAIT;
209 
210 	/* Reset boost if the CPU appears to have been idle enough */
211 	if (sg_cpu->iowait_boost &&
212 	    sugov_iowait_reset(sg_cpu, time, set_iowait_boost))
213 		return;
214 
215 	/* Boost only tasks waking up after IO */
216 	if (!set_iowait_boost)
217 		return;
218 
219 	/* Ensure boost doubles only one time at each request */
220 	if (sg_cpu->iowait_boost_pending)
221 		return;
222 	sg_cpu->iowait_boost_pending = true;
223 
224 	/* Double the boost at each request */
225 	if (sg_cpu->iowait_boost) {
226 		sg_cpu->iowait_boost =
227 			min_t(unsigned int, sg_cpu->iowait_boost << 1, SCHED_CAPACITY_SCALE);
228 		return;
229 	}
230 
231 	/* First wakeup after IO: start with minimum boost */
232 	sg_cpu->iowait_boost = IOWAIT_BOOST_MIN;
233 }
234 
235 /**
236  * sugov_iowait_apply() - Apply the IO boost to a CPU.
237  * @sg_cpu: the sugov data for the cpu to boost
238  * @time: the update time from the caller
239  * @max_cap: the max CPU capacity
240  *
241  * A CPU running a task which woken up after an IO operation can have its
242  * utilization boosted to speed up the completion of those IO operations.
243  * The IO boost value is increased each time a task wakes up from IO, in
244  * sugov_iowait_apply(), and it's instead decreased by this function,
245  * each time an increase has not been requested (!iowait_boost_pending).
246  *
247  * A CPU which also appears to have been idle for at least one tick has also
248  * its IO boost utilization reset.
249  *
250  * This mechanism is designed to boost high frequently IO waiting tasks, while
251  * being more conservative on tasks which does sporadic IO operations.
252  */
253 static void sugov_iowait_apply(struct sugov_cpu *sg_cpu, u64 time,
254 			       unsigned long max_cap)
255 {
256 	unsigned long boost;
257 
258 	/* No boost currently required */
259 	if (!sg_cpu->iowait_boost)
260 		return;
261 
262 	/* Reset boost if the CPU appears to have been idle enough */
263 	if (sugov_iowait_reset(sg_cpu, time, false))
264 		return;
265 
266 	if (!sg_cpu->iowait_boost_pending) {
267 		/*
268 		 * No boost pending; reduce the boost value.
269 		 */
270 		sg_cpu->iowait_boost >>= 1;
271 		if (sg_cpu->iowait_boost < IOWAIT_BOOST_MIN) {
272 			sg_cpu->iowait_boost = 0;
273 			return;
274 		}
275 	}
276 
277 	sg_cpu->iowait_boost_pending = false;
278 
279 	/*
280 	 * sg_cpu->util is already in capacity scale; convert iowait_boost
281 	 * into the same scale so we can compare.
282 	 */
283 	boost = (sg_cpu->iowait_boost * max_cap) >> SCHED_CAPACITY_SHIFT;
284 	boost = uclamp_rq_util_with(cpu_rq(sg_cpu->cpu), boost, NULL);
285 	if (sg_cpu->util < boost)
286 		sg_cpu->util = boost;
287 }
288 
289 #ifdef CONFIG_NO_HZ_COMMON
290 static bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu)
291 {
292 	unsigned long idle_calls = tick_nohz_get_idle_calls_cpu(sg_cpu->cpu);
293 	bool ret = idle_calls == sg_cpu->saved_idle_calls;
294 
295 	sg_cpu->saved_idle_calls = idle_calls;
296 	return ret;
297 }
298 #else
299 static inline bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu) { return false; }
300 #endif /* CONFIG_NO_HZ_COMMON */
301 
302 /*
303  * Make sugov_should_update_freq() ignore the rate limit when DL
304  * has increased the utilization.
305  */
306 static inline void ignore_dl_rate_limit(struct sugov_cpu *sg_cpu)
307 {
308 	if (cpu_bw_dl(cpu_rq(sg_cpu->cpu)) > sg_cpu->bw_dl)
309 		sg_cpu->sg_policy->limits_changed = true;
310 }
311 
312 static inline bool sugov_update_single_common(struct sugov_cpu *sg_cpu,
313 					      u64 time, unsigned long max_cap,
314 					      unsigned int flags)
315 {
316 	sugov_iowait_boost(sg_cpu, time, flags);
317 	sg_cpu->last_update = time;
318 
319 	ignore_dl_rate_limit(sg_cpu);
320 
321 	if (!sugov_should_update_freq(sg_cpu->sg_policy, time))
322 		return false;
323 
324 	sugov_get_util(sg_cpu);
325 	sugov_iowait_apply(sg_cpu, time, max_cap);
326 
327 	return true;
328 }
329 
330 static void sugov_update_single_freq(struct update_util_data *hook, u64 time,
331 				     unsigned int flags)
332 {
333 	struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
334 	struct sugov_policy *sg_policy = sg_cpu->sg_policy;
335 	unsigned int cached_freq = sg_policy->cached_raw_freq;
336 	unsigned long max_cap;
337 	unsigned int next_f;
338 
339 	max_cap = arch_scale_cpu_capacity(sg_cpu->cpu);
340 
341 	if (!sugov_update_single_common(sg_cpu, time, max_cap, flags))
342 		return;
343 
344 	next_f = get_next_freq(sg_policy, sg_cpu->util, max_cap);
345 	/*
346 	 * Do not reduce the frequency if the CPU has not been idle
347 	 * recently, as the reduction is likely to be premature then.
348 	 *
349 	 * Except when the rq is capped by uclamp_max.
350 	 */
351 	if (!uclamp_rq_is_capped(cpu_rq(sg_cpu->cpu)) &&
352 	    sugov_cpu_is_busy(sg_cpu) && next_f < sg_policy->next_freq) {
353 		next_f = sg_policy->next_freq;
354 
355 		/* Restore cached freq as next_freq has changed */
356 		sg_policy->cached_raw_freq = cached_freq;
357 	}
358 
359 	if (!sugov_update_next_freq(sg_policy, time, next_f))
360 		return;
361 
362 	/*
363 	 * This code runs under rq->lock for the target CPU, so it won't run
364 	 * concurrently on two different CPUs for the same target and it is not
365 	 * necessary to acquire the lock in the fast switch case.
366 	 */
367 	if (sg_policy->policy->fast_switch_enabled) {
368 		cpufreq_driver_fast_switch(sg_policy->policy, next_f);
369 	} else {
370 		raw_spin_lock(&sg_policy->update_lock);
371 		sugov_deferred_update(sg_policy);
372 		raw_spin_unlock(&sg_policy->update_lock);
373 	}
374 }
375 
376 static void sugov_update_single_perf(struct update_util_data *hook, u64 time,
377 				     unsigned int flags)
378 {
379 	struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
380 	unsigned long prev_util = sg_cpu->util;
381 	unsigned long max_cap;
382 
383 	/*
384 	 * Fall back to the "frequency" path if frequency invariance is not
385 	 * supported, because the direct mapping between the utilization and
386 	 * the performance levels depends on the frequency invariance.
387 	 */
388 	if (!arch_scale_freq_invariant()) {
389 		sugov_update_single_freq(hook, time, flags);
390 		return;
391 	}
392 
393 	max_cap = arch_scale_cpu_capacity(sg_cpu->cpu);
394 
395 	if (!sugov_update_single_common(sg_cpu, time, max_cap, flags))
396 		return;
397 
398 	/*
399 	 * Do not reduce the target performance level if the CPU has not been
400 	 * idle recently, as the reduction is likely to be premature then.
401 	 *
402 	 * Except when the rq is capped by uclamp_max.
403 	 */
404 	if (!uclamp_rq_is_capped(cpu_rq(sg_cpu->cpu)) &&
405 	    sugov_cpu_is_busy(sg_cpu) && sg_cpu->util < prev_util)
406 		sg_cpu->util = prev_util;
407 
408 	cpufreq_driver_adjust_perf(sg_cpu->cpu, map_util_perf(sg_cpu->bw_dl),
409 				   map_util_perf(sg_cpu->util), max_cap);
410 
411 	sg_cpu->sg_policy->last_freq_update_time = time;
412 }
413 
414 static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu, u64 time)
415 {
416 	struct sugov_policy *sg_policy = sg_cpu->sg_policy;
417 	struct cpufreq_policy *policy = sg_policy->policy;
418 	unsigned long util = 0, max_cap;
419 	unsigned int j;
420 
421 	max_cap = arch_scale_cpu_capacity(sg_cpu->cpu);
422 
423 	for_each_cpu(j, policy->cpus) {
424 		struct sugov_cpu *j_sg_cpu = &per_cpu(sugov_cpu, j);
425 
426 		sugov_get_util(j_sg_cpu);
427 		sugov_iowait_apply(j_sg_cpu, time, max_cap);
428 
429 		util = max(j_sg_cpu->util, util);
430 	}
431 
432 	return get_next_freq(sg_policy, util, max_cap);
433 }
434 
435 static void
436 sugov_update_shared(struct update_util_data *hook, u64 time, unsigned int flags)
437 {
438 	struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
439 	struct sugov_policy *sg_policy = sg_cpu->sg_policy;
440 	unsigned int next_f;
441 
442 	raw_spin_lock(&sg_policy->update_lock);
443 
444 	sugov_iowait_boost(sg_cpu, time, flags);
445 	sg_cpu->last_update = time;
446 
447 	ignore_dl_rate_limit(sg_cpu);
448 
449 	if (sugov_should_update_freq(sg_policy, time)) {
450 		next_f = sugov_next_freq_shared(sg_cpu, time);
451 
452 		if (!sugov_update_next_freq(sg_policy, time, next_f))
453 			goto unlock;
454 
455 		if (sg_policy->policy->fast_switch_enabled)
456 			cpufreq_driver_fast_switch(sg_policy->policy, next_f);
457 		else
458 			sugov_deferred_update(sg_policy);
459 	}
460 unlock:
461 	raw_spin_unlock(&sg_policy->update_lock);
462 }
463 
464 static void sugov_work(struct kthread_work *work)
465 {
466 	struct sugov_policy *sg_policy = container_of(work, struct sugov_policy, work);
467 	unsigned int freq;
468 	unsigned long flags;
469 
470 	/*
471 	 * Hold sg_policy->update_lock shortly to handle the case where:
472 	 * in case sg_policy->next_freq is read here, and then updated by
473 	 * sugov_deferred_update() just before work_in_progress is set to false
474 	 * here, we may miss queueing the new update.
475 	 *
476 	 * Note: If a work was queued after the update_lock is released,
477 	 * sugov_work() will just be called again by kthread_work code; and the
478 	 * request will be proceed before the sugov thread sleeps.
479 	 */
480 	raw_spin_lock_irqsave(&sg_policy->update_lock, flags);
481 	freq = sg_policy->next_freq;
482 	sg_policy->work_in_progress = false;
483 	raw_spin_unlock_irqrestore(&sg_policy->update_lock, flags);
484 
485 	mutex_lock(&sg_policy->work_lock);
486 	__cpufreq_driver_target(sg_policy->policy, freq, CPUFREQ_RELATION_L);
487 	mutex_unlock(&sg_policy->work_lock);
488 }
489 
490 static void sugov_irq_work(struct irq_work *irq_work)
491 {
492 	struct sugov_policy *sg_policy;
493 
494 	sg_policy = container_of(irq_work, struct sugov_policy, irq_work);
495 
496 	kthread_queue_work(&sg_policy->worker, &sg_policy->work);
497 }
498 
499 /************************** sysfs interface ************************/
500 
501 static struct sugov_tunables *global_tunables;
502 static DEFINE_MUTEX(global_tunables_lock);
503 
504 static inline struct sugov_tunables *to_sugov_tunables(struct gov_attr_set *attr_set)
505 {
506 	return container_of(attr_set, struct sugov_tunables, attr_set);
507 }
508 
509 static ssize_t rate_limit_us_show(struct gov_attr_set *attr_set, char *buf)
510 {
511 	struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
512 
513 	return sprintf(buf, "%u\n", tunables->rate_limit_us);
514 }
515 
516 static ssize_t
517 rate_limit_us_store(struct gov_attr_set *attr_set, const char *buf, size_t count)
518 {
519 	struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
520 	struct sugov_policy *sg_policy;
521 	unsigned int rate_limit_us;
522 
523 	if (kstrtouint(buf, 10, &rate_limit_us))
524 		return -EINVAL;
525 
526 	tunables->rate_limit_us = rate_limit_us;
527 
528 	list_for_each_entry(sg_policy, &attr_set->policy_list, tunables_hook)
529 		sg_policy->freq_update_delay_ns = rate_limit_us * NSEC_PER_USEC;
530 
531 	return count;
532 }
533 
534 static struct governor_attr rate_limit_us = __ATTR_RW(rate_limit_us);
535 
536 static struct attribute *sugov_attrs[] = {
537 	&rate_limit_us.attr,
538 	NULL
539 };
540 ATTRIBUTE_GROUPS(sugov);
541 
542 static void sugov_tunables_free(struct kobject *kobj)
543 {
544 	struct gov_attr_set *attr_set = to_gov_attr_set(kobj);
545 
546 	kfree(to_sugov_tunables(attr_set));
547 }
548 
549 static const struct kobj_type sugov_tunables_ktype = {
550 	.default_groups = sugov_groups,
551 	.sysfs_ops = &governor_sysfs_ops,
552 	.release = &sugov_tunables_free,
553 };
554 
555 /********************** cpufreq governor interface *********************/
556 
557 struct cpufreq_governor schedutil_gov;
558 
559 static struct sugov_policy *sugov_policy_alloc(struct cpufreq_policy *policy)
560 {
561 	struct sugov_policy *sg_policy;
562 
563 	sg_policy = kzalloc(sizeof(*sg_policy), GFP_KERNEL);
564 	if (!sg_policy)
565 		return NULL;
566 
567 	sg_policy->policy = policy;
568 	raw_spin_lock_init(&sg_policy->update_lock);
569 	return sg_policy;
570 }
571 
572 static void sugov_policy_free(struct sugov_policy *sg_policy)
573 {
574 	kfree(sg_policy);
575 }
576 
577 static int sugov_kthread_create(struct sugov_policy *sg_policy)
578 {
579 	struct task_struct *thread;
580 	struct sched_attr attr = {
581 		.size		= sizeof(struct sched_attr),
582 		.sched_policy	= SCHED_DEADLINE,
583 		.sched_flags	= SCHED_FLAG_SUGOV,
584 		.sched_nice	= 0,
585 		.sched_priority	= 0,
586 		/*
587 		 * Fake (unused) bandwidth; workaround to "fix"
588 		 * priority inheritance.
589 		 */
590 		.sched_runtime	=  1000000,
591 		.sched_deadline = 10000000,
592 		.sched_period	= 10000000,
593 	};
594 	struct cpufreq_policy *policy = sg_policy->policy;
595 	int ret;
596 
597 	/* kthread only required for slow path */
598 	if (policy->fast_switch_enabled)
599 		return 0;
600 
601 	kthread_init_work(&sg_policy->work, sugov_work);
602 	kthread_init_worker(&sg_policy->worker);
603 	thread = kthread_create(kthread_worker_fn, &sg_policy->worker,
604 				"sugov:%d",
605 				cpumask_first(policy->related_cpus));
606 	if (IS_ERR(thread)) {
607 		pr_err("failed to create sugov thread: %ld\n", PTR_ERR(thread));
608 		return PTR_ERR(thread);
609 	}
610 
611 	ret = sched_setattr_nocheck(thread, &attr);
612 	if (ret) {
613 		kthread_stop(thread);
614 		pr_warn("%s: failed to set SCHED_DEADLINE\n", __func__);
615 		return ret;
616 	}
617 
618 	sg_policy->thread = thread;
619 	kthread_bind_mask(thread, policy->related_cpus);
620 	init_irq_work(&sg_policy->irq_work, sugov_irq_work);
621 	mutex_init(&sg_policy->work_lock);
622 
623 	wake_up_process(thread);
624 
625 	return 0;
626 }
627 
628 static void sugov_kthread_stop(struct sugov_policy *sg_policy)
629 {
630 	/* kthread only required for slow path */
631 	if (sg_policy->policy->fast_switch_enabled)
632 		return;
633 
634 	kthread_flush_worker(&sg_policy->worker);
635 	kthread_stop(sg_policy->thread);
636 	mutex_destroy(&sg_policy->work_lock);
637 }
638 
639 static struct sugov_tunables *sugov_tunables_alloc(struct sugov_policy *sg_policy)
640 {
641 	struct sugov_tunables *tunables;
642 
643 	tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
644 	if (tunables) {
645 		gov_attr_set_init(&tunables->attr_set, &sg_policy->tunables_hook);
646 		if (!have_governor_per_policy())
647 			global_tunables = tunables;
648 	}
649 	return tunables;
650 }
651 
652 static void sugov_clear_global_tunables(void)
653 {
654 	if (!have_governor_per_policy())
655 		global_tunables = NULL;
656 }
657 
658 static int sugov_init(struct cpufreq_policy *policy)
659 {
660 	struct sugov_policy *sg_policy;
661 	struct sugov_tunables *tunables;
662 	int ret = 0;
663 
664 	/* State should be equivalent to EXIT */
665 	if (policy->governor_data)
666 		return -EBUSY;
667 
668 	cpufreq_enable_fast_switch(policy);
669 
670 	sg_policy = sugov_policy_alloc(policy);
671 	if (!sg_policy) {
672 		ret = -ENOMEM;
673 		goto disable_fast_switch;
674 	}
675 
676 	ret = sugov_kthread_create(sg_policy);
677 	if (ret)
678 		goto free_sg_policy;
679 
680 	mutex_lock(&global_tunables_lock);
681 
682 	if (global_tunables) {
683 		if (WARN_ON(have_governor_per_policy())) {
684 			ret = -EINVAL;
685 			goto stop_kthread;
686 		}
687 		policy->governor_data = sg_policy;
688 		sg_policy->tunables = global_tunables;
689 
690 		gov_attr_set_get(&global_tunables->attr_set, &sg_policy->tunables_hook);
691 		goto out;
692 	}
693 
694 	tunables = sugov_tunables_alloc(sg_policy);
695 	if (!tunables) {
696 		ret = -ENOMEM;
697 		goto stop_kthread;
698 	}
699 
700 	tunables->rate_limit_us = cpufreq_policy_transition_delay_us(policy);
701 
702 	policy->governor_data = sg_policy;
703 	sg_policy->tunables = tunables;
704 
705 	ret = kobject_init_and_add(&tunables->attr_set.kobj, &sugov_tunables_ktype,
706 				   get_governor_parent_kobj(policy), "%s",
707 				   schedutil_gov.name);
708 	if (ret)
709 		goto fail;
710 
711 out:
712 	mutex_unlock(&global_tunables_lock);
713 	return 0;
714 
715 fail:
716 	kobject_put(&tunables->attr_set.kobj);
717 	policy->governor_data = NULL;
718 	sugov_clear_global_tunables();
719 
720 stop_kthread:
721 	sugov_kthread_stop(sg_policy);
722 	mutex_unlock(&global_tunables_lock);
723 
724 free_sg_policy:
725 	sugov_policy_free(sg_policy);
726 
727 disable_fast_switch:
728 	cpufreq_disable_fast_switch(policy);
729 
730 	pr_err("initialization failed (error %d)\n", ret);
731 	return ret;
732 }
733 
734 static void sugov_exit(struct cpufreq_policy *policy)
735 {
736 	struct sugov_policy *sg_policy = policy->governor_data;
737 	struct sugov_tunables *tunables = sg_policy->tunables;
738 	unsigned int count;
739 
740 	mutex_lock(&global_tunables_lock);
741 
742 	count = gov_attr_set_put(&tunables->attr_set, &sg_policy->tunables_hook);
743 	policy->governor_data = NULL;
744 	if (!count)
745 		sugov_clear_global_tunables();
746 
747 	mutex_unlock(&global_tunables_lock);
748 
749 	sugov_kthread_stop(sg_policy);
750 	sugov_policy_free(sg_policy);
751 	cpufreq_disable_fast_switch(policy);
752 }
753 
754 static int sugov_start(struct cpufreq_policy *policy)
755 {
756 	struct sugov_policy *sg_policy = policy->governor_data;
757 	void (*uu)(struct update_util_data *data, u64 time, unsigned int flags);
758 	unsigned int cpu;
759 
760 	sg_policy->freq_update_delay_ns	= sg_policy->tunables->rate_limit_us * NSEC_PER_USEC;
761 	sg_policy->last_freq_update_time	= 0;
762 	sg_policy->next_freq			= 0;
763 	sg_policy->work_in_progress		= false;
764 	sg_policy->limits_changed		= false;
765 	sg_policy->cached_raw_freq		= 0;
766 
767 	sg_policy->need_freq_update = cpufreq_driver_test_flags(CPUFREQ_NEED_UPDATE_LIMITS);
768 
769 	for_each_cpu(cpu, policy->cpus) {
770 		struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
771 
772 		memset(sg_cpu, 0, sizeof(*sg_cpu));
773 		sg_cpu->cpu			= cpu;
774 		sg_cpu->sg_policy		= sg_policy;
775 	}
776 
777 	if (policy_is_shared(policy))
778 		uu = sugov_update_shared;
779 	else if (policy->fast_switch_enabled && cpufreq_driver_has_adjust_perf())
780 		uu = sugov_update_single_perf;
781 	else
782 		uu = sugov_update_single_freq;
783 
784 	for_each_cpu(cpu, policy->cpus) {
785 		struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
786 
787 		cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util, uu);
788 	}
789 	return 0;
790 }
791 
792 static void sugov_stop(struct cpufreq_policy *policy)
793 {
794 	struct sugov_policy *sg_policy = policy->governor_data;
795 	unsigned int cpu;
796 
797 	for_each_cpu(cpu, policy->cpus)
798 		cpufreq_remove_update_util_hook(cpu);
799 
800 	synchronize_rcu();
801 
802 	if (!policy->fast_switch_enabled) {
803 		irq_work_sync(&sg_policy->irq_work);
804 		kthread_cancel_work_sync(&sg_policy->work);
805 	}
806 }
807 
808 static void sugov_limits(struct cpufreq_policy *policy)
809 {
810 	struct sugov_policy *sg_policy = policy->governor_data;
811 
812 	if (!policy->fast_switch_enabled) {
813 		mutex_lock(&sg_policy->work_lock);
814 		cpufreq_policy_apply_limits(policy);
815 		mutex_unlock(&sg_policy->work_lock);
816 	}
817 
818 	sg_policy->limits_changed = true;
819 }
820 
821 struct cpufreq_governor schedutil_gov = {
822 	.name			= "schedutil",
823 	.owner			= THIS_MODULE,
824 	.flags			= CPUFREQ_GOV_DYNAMIC_SWITCHING,
825 	.init			= sugov_init,
826 	.exit			= sugov_exit,
827 	.start			= sugov_start,
828 	.stop			= sugov_stop,
829 	.limits			= sugov_limits,
830 };
831 
832 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL
833 struct cpufreq_governor *cpufreq_default_governor(void)
834 {
835 	return &schedutil_gov;
836 }
837 #endif
838 
839 cpufreq_governor_init(schedutil_gov);
840 
841 #ifdef CONFIG_ENERGY_MODEL
842 static void rebuild_sd_workfn(struct work_struct *work)
843 {
844 	rebuild_sched_domains_energy();
845 }
846 static DECLARE_WORK(rebuild_sd_work, rebuild_sd_workfn);
847 
848 /*
849  * EAS shouldn't be attempted without sugov, so rebuild the sched_domains
850  * on governor changes to make sure the scheduler knows about it.
851  */
852 void sched_cpufreq_governor_change(struct cpufreq_policy *policy,
853 				  struct cpufreq_governor *old_gov)
854 {
855 	if (old_gov == &schedutil_gov || policy->governor == &schedutil_gov) {
856 		/*
857 		 * When called from the cpufreq_register_driver() path, the
858 		 * cpu_hotplug_lock is already held, so use a work item to
859 		 * avoid nested locking in rebuild_sched_domains().
860 		 */
861 		schedule_work(&rebuild_sd_work);
862 	}
863 
864 }
865 #endif
866