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 <linux/cpufreq.h>
15 #include <linux/kthread.h>
16 #include <uapi/linux/sched/types.h>
17 #include <linux/slab.h>
18 #include <trace/events/power.h>
19 
20 #include "sched.h"
21 
22 #define SUGOV_KTHREAD_PRIORITY	50
23 
24 struct sugov_tunables {
25 	struct gov_attr_set attr_set;
26 	unsigned int rate_limit_us;
27 };
28 
29 struct sugov_policy {
30 	struct cpufreq_policy *policy;
31 
32 	struct sugov_tunables *tunables;
33 	struct list_head tunables_hook;
34 
35 	raw_spinlock_t update_lock;  /* For shared policies */
36 	u64 last_freq_update_time;
37 	s64 freq_update_delay_ns;
38 	unsigned int next_freq;
39 	unsigned int cached_raw_freq;
40 
41 	/* The next fields are only needed if fast switch cannot be used. */
42 	struct irq_work irq_work;
43 	struct kthread_work work;
44 	struct mutex work_lock;
45 	struct kthread_worker worker;
46 	struct task_struct *thread;
47 	bool work_in_progress;
48 
49 	bool need_freq_update;
50 };
51 
52 struct sugov_cpu {
53 	struct update_util_data update_util;
54 	struct sugov_policy *sg_policy;
55 
56 	unsigned long iowait_boost;
57 	unsigned long iowait_boost_max;
58 	u64 last_update;
59 
60 	/* The fields below are only needed when sharing a policy. */
61 	unsigned long util;
62 	unsigned long max;
63 	unsigned int flags;
64 
65 	/* The field below is for single-CPU policies only. */
66 #ifdef CONFIG_NO_HZ_COMMON
67 	unsigned long saved_idle_calls;
68 #endif
69 };
70 
71 static DEFINE_PER_CPU(struct sugov_cpu, sugov_cpu);
72 
73 /************************ Governor internals ***********************/
74 
75 static bool sugov_should_update_freq(struct sugov_policy *sg_policy, u64 time)
76 {
77 	s64 delta_ns;
78 
79 	if (sg_policy->work_in_progress)
80 		return false;
81 
82 	if (unlikely(sg_policy->need_freq_update)) {
83 		sg_policy->need_freq_update = false;
84 		/*
85 		 * This happens when limits change, so forget the previous
86 		 * next_freq value and force an update.
87 		 */
88 		sg_policy->next_freq = UINT_MAX;
89 		return true;
90 	}
91 
92 	delta_ns = time - sg_policy->last_freq_update_time;
93 	return delta_ns >= sg_policy->freq_update_delay_ns;
94 }
95 
96 static void sugov_update_commit(struct sugov_policy *sg_policy, u64 time,
97 				unsigned int next_freq)
98 {
99 	struct cpufreq_policy *policy = sg_policy->policy;
100 
101 	if (sg_policy->next_freq == next_freq)
102 		return;
103 
104 	sg_policy->next_freq = next_freq;
105 	sg_policy->last_freq_update_time = time;
106 
107 	if (policy->fast_switch_enabled) {
108 		next_freq = cpufreq_driver_fast_switch(policy, next_freq);
109 		if (next_freq == CPUFREQ_ENTRY_INVALID)
110 			return;
111 
112 		policy->cur = next_freq;
113 		trace_cpu_frequency(next_freq, smp_processor_id());
114 	} else {
115 		sg_policy->work_in_progress = true;
116 		irq_work_queue(&sg_policy->irq_work);
117 	}
118 }
119 
120 /**
121  * get_next_freq - Compute a new frequency for a given cpufreq policy.
122  * @sg_policy: schedutil policy object to compute the new frequency for.
123  * @util: Current CPU utilization.
124  * @max: CPU capacity.
125  *
126  * If the utilization is frequency-invariant, choose the new frequency to be
127  * proportional to it, that is
128  *
129  * next_freq = C * max_freq * util / max
130  *
131  * Otherwise, approximate the would-be frequency-invariant utilization by
132  * util_raw * (curr_freq / max_freq) which leads to
133  *
134  * next_freq = C * curr_freq * util_raw / max
135  *
136  * Take C = 1.25 for the frequency tipping point at (util / max) = 0.8.
137  *
138  * The lowest driver-supported frequency which is equal or greater than the raw
139  * next_freq (as calculated above) is returned, subject to policy min/max and
140  * cpufreq driver limitations.
141  */
142 static unsigned int get_next_freq(struct sugov_policy *sg_policy,
143 				  unsigned long util, unsigned long max)
144 {
145 	struct cpufreq_policy *policy = sg_policy->policy;
146 	unsigned int freq = arch_scale_freq_invariant() ?
147 				policy->cpuinfo.max_freq : policy->cur;
148 
149 	freq = (freq + (freq >> 2)) * util / max;
150 
151 	if (freq == sg_policy->cached_raw_freq && sg_policy->next_freq != UINT_MAX)
152 		return sg_policy->next_freq;
153 	sg_policy->cached_raw_freq = freq;
154 	return cpufreq_driver_resolve_freq(policy, freq);
155 }
156 
157 static void sugov_get_util(unsigned long *util, unsigned long *max)
158 {
159 	struct rq *rq = this_rq();
160 	unsigned long cfs_max;
161 
162 	cfs_max = arch_scale_cpu_capacity(NULL, smp_processor_id());
163 
164 	*util = min(rq->cfs.avg.util_avg, cfs_max);
165 	*max = cfs_max;
166 }
167 
168 static void sugov_set_iowait_boost(struct sugov_cpu *sg_cpu, u64 time,
169 				   unsigned int flags)
170 {
171 	if (flags & SCHED_CPUFREQ_IOWAIT) {
172 		sg_cpu->iowait_boost = sg_cpu->iowait_boost_max;
173 	} else if (sg_cpu->iowait_boost) {
174 		s64 delta_ns = time - sg_cpu->last_update;
175 
176 		/* Clear iowait_boost if the CPU apprears to have been idle. */
177 		if (delta_ns > TICK_NSEC)
178 			sg_cpu->iowait_boost = 0;
179 	}
180 }
181 
182 static void sugov_iowait_boost(struct sugov_cpu *sg_cpu, unsigned long *util,
183 			       unsigned long *max)
184 {
185 	unsigned long boost_util = sg_cpu->iowait_boost;
186 	unsigned long boost_max = sg_cpu->iowait_boost_max;
187 
188 	if (!boost_util)
189 		return;
190 
191 	if (*util * boost_max < *max * boost_util) {
192 		*util = boost_util;
193 		*max = boost_max;
194 	}
195 	sg_cpu->iowait_boost >>= 1;
196 }
197 
198 #ifdef CONFIG_NO_HZ_COMMON
199 static bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu)
200 {
201 	unsigned long idle_calls = tick_nohz_get_idle_calls();
202 	bool ret = idle_calls == sg_cpu->saved_idle_calls;
203 
204 	sg_cpu->saved_idle_calls = idle_calls;
205 	return ret;
206 }
207 #else
208 static inline bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu) { return false; }
209 #endif /* CONFIG_NO_HZ_COMMON */
210 
211 static void sugov_update_single(struct update_util_data *hook, u64 time,
212 				unsigned int flags)
213 {
214 	struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
215 	struct sugov_policy *sg_policy = sg_cpu->sg_policy;
216 	struct cpufreq_policy *policy = sg_policy->policy;
217 	unsigned long util, max;
218 	unsigned int next_f;
219 	bool busy;
220 
221 	sugov_set_iowait_boost(sg_cpu, time, flags);
222 	sg_cpu->last_update = time;
223 
224 	if (!sugov_should_update_freq(sg_policy, time))
225 		return;
226 
227 	busy = sugov_cpu_is_busy(sg_cpu);
228 
229 	if (flags & SCHED_CPUFREQ_RT_DL) {
230 		next_f = policy->cpuinfo.max_freq;
231 	} else {
232 		sugov_get_util(&util, &max);
233 		sugov_iowait_boost(sg_cpu, &util, &max);
234 		next_f = get_next_freq(sg_policy, util, max);
235 		/*
236 		 * Do not reduce the frequency if the CPU has not been idle
237 		 * recently, as the reduction is likely to be premature then.
238 		 */
239 		if (busy && next_f < sg_policy->next_freq)
240 			next_f = sg_policy->next_freq;
241 	}
242 	sugov_update_commit(sg_policy, time, next_f);
243 }
244 
245 static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu, u64 time)
246 {
247 	struct sugov_policy *sg_policy = sg_cpu->sg_policy;
248 	struct cpufreq_policy *policy = sg_policy->policy;
249 	unsigned long util = 0, max = 1;
250 	unsigned int j;
251 
252 	for_each_cpu(j, policy->cpus) {
253 		struct sugov_cpu *j_sg_cpu = &per_cpu(sugov_cpu, j);
254 		unsigned long j_util, j_max;
255 		s64 delta_ns;
256 
257 		/*
258 		 * If the CPU utilization was last updated before the previous
259 		 * frequency update and the time elapsed between the last update
260 		 * of the CPU utilization and the last frequency update is long
261 		 * enough, don't take the CPU into account as it probably is
262 		 * idle now (and clear iowait_boost for it).
263 		 */
264 		delta_ns = time - j_sg_cpu->last_update;
265 		if (delta_ns > TICK_NSEC) {
266 			j_sg_cpu->iowait_boost = 0;
267 			continue;
268 		}
269 		if (j_sg_cpu->flags & SCHED_CPUFREQ_RT_DL)
270 			return policy->cpuinfo.max_freq;
271 
272 		j_util = j_sg_cpu->util;
273 		j_max = j_sg_cpu->max;
274 		if (j_util * max > j_max * util) {
275 			util = j_util;
276 			max = j_max;
277 		}
278 
279 		sugov_iowait_boost(j_sg_cpu, &util, &max);
280 	}
281 
282 	return get_next_freq(sg_policy, util, max);
283 }
284 
285 static void sugov_update_shared(struct update_util_data *hook, u64 time,
286 				unsigned int flags)
287 {
288 	struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
289 	struct sugov_policy *sg_policy = sg_cpu->sg_policy;
290 	unsigned long util, max;
291 	unsigned int next_f;
292 
293 	sugov_get_util(&util, &max);
294 
295 	raw_spin_lock(&sg_policy->update_lock);
296 
297 	sg_cpu->util = util;
298 	sg_cpu->max = max;
299 	sg_cpu->flags = flags;
300 
301 	sugov_set_iowait_boost(sg_cpu, time, flags);
302 	sg_cpu->last_update = time;
303 
304 	if (sugov_should_update_freq(sg_policy, time)) {
305 		if (flags & SCHED_CPUFREQ_RT_DL)
306 			next_f = sg_policy->policy->cpuinfo.max_freq;
307 		else
308 			next_f = sugov_next_freq_shared(sg_cpu, time);
309 
310 		sugov_update_commit(sg_policy, time, next_f);
311 	}
312 
313 	raw_spin_unlock(&sg_policy->update_lock);
314 }
315 
316 static void sugov_work(struct kthread_work *work)
317 {
318 	struct sugov_policy *sg_policy = container_of(work, struct sugov_policy, work);
319 
320 	mutex_lock(&sg_policy->work_lock);
321 	__cpufreq_driver_target(sg_policy->policy, sg_policy->next_freq,
322 				CPUFREQ_RELATION_L);
323 	mutex_unlock(&sg_policy->work_lock);
324 
325 	sg_policy->work_in_progress = false;
326 }
327 
328 static void sugov_irq_work(struct irq_work *irq_work)
329 {
330 	struct sugov_policy *sg_policy;
331 
332 	sg_policy = container_of(irq_work, struct sugov_policy, irq_work);
333 
334 	/*
335 	 * For RT and deadline tasks, the schedutil governor shoots the
336 	 * frequency to maximum. Special care must be taken to ensure that this
337 	 * kthread doesn't result in the same behavior.
338 	 *
339 	 * This is (mostly) guaranteed by the work_in_progress flag. The flag is
340 	 * updated only at the end of the sugov_work() function and before that
341 	 * the schedutil governor rejects all other frequency scaling requests.
342 	 *
343 	 * There is a very rare case though, where the RT thread yields right
344 	 * after the work_in_progress flag is cleared. The effects of that are
345 	 * neglected for now.
346 	 */
347 	kthread_queue_work(&sg_policy->worker, &sg_policy->work);
348 }
349 
350 /************************** sysfs interface ************************/
351 
352 static struct sugov_tunables *global_tunables;
353 static DEFINE_MUTEX(global_tunables_lock);
354 
355 static inline struct sugov_tunables *to_sugov_tunables(struct gov_attr_set *attr_set)
356 {
357 	return container_of(attr_set, struct sugov_tunables, attr_set);
358 }
359 
360 static ssize_t rate_limit_us_show(struct gov_attr_set *attr_set, char *buf)
361 {
362 	struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
363 
364 	return sprintf(buf, "%u\n", tunables->rate_limit_us);
365 }
366 
367 static ssize_t rate_limit_us_store(struct gov_attr_set *attr_set, const char *buf,
368 				   size_t count)
369 {
370 	struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
371 	struct sugov_policy *sg_policy;
372 	unsigned int rate_limit_us;
373 
374 	if (kstrtouint(buf, 10, &rate_limit_us))
375 		return -EINVAL;
376 
377 	tunables->rate_limit_us = rate_limit_us;
378 
379 	list_for_each_entry(sg_policy, &attr_set->policy_list, tunables_hook)
380 		sg_policy->freq_update_delay_ns = rate_limit_us * NSEC_PER_USEC;
381 
382 	return count;
383 }
384 
385 static struct governor_attr rate_limit_us = __ATTR_RW(rate_limit_us);
386 
387 static struct attribute *sugov_attributes[] = {
388 	&rate_limit_us.attr,
389 	NULL
390 };
391 
392 static struct kobj_type sugov_tunables_ktype = {
393 	.default_attrs = sugov_attributes,
394 	.sysfs_ops = &governor_sysfs_ops,
395 };
396 
397 /********************** cpufreq governor interface *********************/
398 
399 static struct cpufreq_governor schedutil_gov;
400 
401 static struct sugov_policy *sugov_policy_alloc(struct cpufreq_policy *policy)
402 {
403 	struct sugov_policy *sg_policy;
404 
405 	sg_policy = kzalloc(sizeof(*sg_policy), GFP_KERNEL);
406 	if (!sg_policy)
407 		return NULL;
408 
409 	sg_policy->policy = policy;
410 	raw_spin_lock_init(&sg_policy->update_lock);
411 	return sg_policy;
412 }
413 
414 static void sugov_policy_free(struct sugov_policy *sg_policy)
415 {
416 	kfree(sg_policy);
417 }
418 
419 static int sugov_kthread_create(struct sugov_policy *sg_policy)
420 {
421 	struct task_struct *thread;
422 	struct sched_param param = { .sched_priority = MAX_USER_RT_PRIO / 2 };
423 	struct cpufreq_policy *policy = sg_policy->policy;
424 	int ret;
425 
426 	/* kthread only required for slow path */
427 	if (policy->fast_switch_enabled)
428 		return 0;
429 
430 	kthread_init_work(&sg_policy->work, sugov_work);
431 	kthread_init_worker(&sg_policy->worker);
432 	thread = kthread_create(kthread_worker_fn, &sg_policy->worker,
433 				"sugov:%d",
434 				cpumask_first(policy->related_cpus));
435 	if (IS_ERR(thread)) {
436 		pr_err("failed to create sugov thread: %ld\n", PTR_ERR(thread));
437 		return PTR_ERR(thread);
438 	}
439 
440 	ret = sched_setscheduler_nocheck(thread, SCHED_FIFO, &param);
441 	if (ret) {
442 		kthread_stop(thread);
443 		pr_warn("%s: failed to set SCHED_FIFO\n", __func__);
444 		return ret;
445 	}
446 
447 	sg_policy->thread = thread;
448 	kthread_bind_mask(thread, policy->related_cpus);
449 	init_irq_work(&sg_policy->irq_work, sugov_irq_work);
450 	mutex_init(&sg_policy->work_lock);
451 
452 	wake_up_process(thread);
453 
454 	return 0;
455 }
456 
457 static void sugov_kthread_stop(struct sugov_policy *sg_policy)
458 {
459 	/* kthread only required for slow path */
460 	if (sg_policy->policy->fast_switch_enabled)
461 		return;
462 
463 	kthread_flush_worker(&sg_policy->worker);
464 	kthread_stop(sg_policy->thread);
465 	mutex_destroy(&sg_policy->work_lock);
466 }
467 
468 static struct sugov_tunables *sugov_tunables_alloc(struct sugov_policy *sg_policy)
469 {
470 	struct sugov_tunables *tunables;
471 
472 	tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
473 	if (tunables) {
474 		gov_attr_set_init(&tunables->attr_set, &sg_policy->tunables_hook);
475 		if (!have_governor_per_policy())
476 			global_tunables = tunables;
477 	}
478 	return tunables;
479 }
480 
481 static void sugov_tunables_free(struct sugov_tunables *tunables)
482 {
483 	if (!have_governor_per_policy())
484 		global_tunables = NULL;
485 
486 	kfree(tunables);
487 }
488 
489 static int sugov_init(struct cpufreq_policy *policy)
490 {
491 	struct sugov_policy *sg_policy;
492 	struct sugov_tunables *tunables;
493 	int ret = 0;
494 
495 	/* State should be equivalent to EXIT */
496 	if (policy->governor_data)
497 		return -EBUSY;
498 
499 	cpufreq_enable_fast_switch(policy);
500 
501 	sg_policy = sugov_policy_alloc(policy);
502 	if (!sg_policy) {
503 		ret = -ENOMEM;
504 		goto disable_fast_switch;
505 	}
506 
507 	ret = sugov_kthread_create(sg_policy);
508 	if (ret)
509 		goto free_sg_policy;
510 
511 	mutex_lock(&global_tunables_lock);
512 
513 	if (global_tunables) {
514 		if (WARN_ON(have_governor_per_policy())) {
515 			ret = -EINVAL;
516 			goto stop_kthread;
517 		}
518 		policy->governor_data = sg_policy;
519 		sg_policy->tunables = global_tunables;
520 
521 		gov_attr_set_get(&global_tunables->attr_set, &sg_policy->tunables_hook);
522 		goto out;
523 	}
524 
525 	tunables = sugov_tunables_alloc(sg_policy);
526 	if (!tunables) {
527 		ret = -ENOMEM;
528 		goto stop_kthread;
529 	}
530 
531 	if (policy->transition_delay_us) {
532 		tunables->rate_limit_us = policy->transition_delay_us;
533 	} else {
534 		unsigned int lat;
535 
536 		tunables->rate_limit_us = LATENCY_MULTIPLIER;
537 		lat = policy->cpuinfo.transition_latency / NSEC_PER_USEC;
538 		if (lat)
539 			tunables->rate_limit_us *= lat;
540 	}
541 
542 	policy->governor_data = sg_policy;
543 	sg_policy->tunables = tunables;
544 
545 	ret = kobject_init_and_add(&tunables->attr_set.kobj, &sugov_tunables_ktype,
546 				   get_governor_parent_kobj(policy), "%s",
547 				   schedutil_gov.name);
548 	if (ret)
549 		goto fail;
550 
551 out:
552 	mutex_unlock(&global_tunables_lock);
553 	return 0;
554 
555 fail:
556 	policy->governor_data = NULL;
557 	sugov_tunables_free(tunables);
558 
559 stop_kthread:
560 	sugov_kthread_stop(sg_policy);
561 
562 free_sg_policy:
563 	mutex_unlock(&global_tunables_lock);
564 
565 	sugov_policy_free(sg_policy);
566 
567 disable_fast_switch:
568 	cpufreq_disable_fast_switch(policy);
569 
570 	pr_err("initialization failed (error %d)\n", ret);
571 	return ret;
572 }
573 
574 static void sugov_exit(struct cpufreq_policy *policy)
575 {
576 	struct sugov_policy *sg_policy = policy->governor_data;
577 	struct sugov_tunables *tunables = sg_policy->tunables;
578 	unsigned int count;
579 
580 	mutex_lock(&global_tunables_lock);
581 
582 	count = gov_attr_set_put(&tunables->attr_set, &sg_policy->tunables_hook);
583 	policy->governor_data = NULL;
584 	if (!count)
585 		sugov_tunables_free(tunables);
586 
587 	mutex_unlock(&global_tunables_lock);
588 
589 	sugov_kthread_stop(sg_policy);
590 	sugov_policy_free(sg_policy);
591 	cpufreq_disable_fast_switch(policy);
592 }
593 
594 static int sugov_start(struct cpufreq_policy *policy)
595 {
596 	struct sugov_policy *sg_policy = policy->governor_data;
597 	unsigned int cpu;
598 
599 	sg_policy->freq_update_delay_ns = sg_policy->tunables->rate_limit_us * NSEC_PER_USEC;
600 	sg_policy->last_freq_update_time = 0;
601 	sg_policy->next_freq = UINT_MAX;
602 	sg_policy->work_in_progress = false;
603 	sg_policy->need_freq_update = false;
604 	sg_policy->cached_raw_freq = 0;
605 
606 	for_each_cpu(cpu, policy->cpus) {
607 		struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
608 
609 		memset(sg_cpu, 0, sizeof(*sg_cpu));
610 		sg_cpu->sg_policy = sg_policy;
611 		sg_cpu->flags = SCHED_CPUFREQ_RT;
612 		sg_cpu->iowait_boost_max = policy->cpuinfo.max_freq;
613 	}
614 
615 	for_each_cpu(cpu, policy->cpus) {
616 		struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
617 
618 		cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util,
619 					     policy_is_shared(policy) ?
620 							sugov_update_shared :
621 							sugov_update_single);
622 	}
623 	return 0;
624 }
625 
626 static void sugov_stop(struct cpufreq_policy *policy)
627 {
628 	struct sugov_policy *sg_policy = policy->governor_data;
629 	unsigned int cpu;
630 
631 	for_each_cpu(cpu, policy->cpus)
632 		cpufreq_remove_update_util_hook(cpu);
633 
634 	synchronize_sched();
635 
636 	if (!policy->fast_switch_enabled) {
637 		irq_work_sync(&sg_policy->irq_work);
638 		kthread_cancel_work_sync(&sg_policy->work);
639 	}
640 }
641 
642 static void sugov_limits(struct cpufreq_policy *policy)
643 {
644 	struct sugov_policy *sg_policy = policy->governor_data;
645 
646 	if (!policy->fast_switch_enabled) {
647 		mutex_lock(&sg_policy->work_lock);
648 		cpufreq_policy_apply_limits(policy);
649 		mutex_unlock(&sg_policy->work_lock);
650 	}
651 
652 	sg_policy->need_freq_update = true;
653 }
654 
655 static struct cpufreq_governor schedutil_gov = {
656 	.name = "schedutil",
657 	.owner = THIS_MODULE,
658 	.init = sugov_init,
659 	.exit = sugov_exit,
660 	.start = sugov_start,
661 	.stop = sugov_stop,
662 	.limits = sugov_limits,
663 };
664 
665 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL
666 struct cpufreq_governor *cpufreq_default_governor(void)
667 {
668 	return &schedutil_gov;
669 }
670 #endif
671 
672 static int __init sugov_register(void)
673 {
674 	return cpufreq_register_governor(&schedutil_gov);
675 }
676 fs_initcall(sugov_register);
677