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