xref: /linux/kernel/sched/cpufreq_schedutil.c (revision 0d08df6c493898e679d9c517e77ea95c063d40ec)
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 #include <linux/cpufreq.h>
13 #include <linux/module.h>
14 #include <linux/slab.h>
15 #include <trace/events/power.h>
16 
17 #include "sched.h"
18 
19 struct sugov_tunables {
20 	struct gov_attr_set attr_set;
21 	unsigned int rate_limit_us;
22 };
23 
24 struct sugov_policy {
25 	struct cpufreq_policy *policy;
26 
27 	struct sugov_tunables *tunables;
28 	struct list_head tunables_hook;
29 
30 	raw_spinlock_t update_lock;  /* For shared policies */
31 	u64 last_freq_update_time;
32 	s64 freq_update_delay_ns;
33 	unsigned int next_freq;
34 
35 	/* The next fields are only needed if fast switch cannot be used. */
36 	struct irq_work irq_work;
37 	struct work_struct work;
38 	struct mutex work_lock;
39 	bool work_in_progress;
40 
41 	bool need_freq_update;
42 };
43 
44 struct sugov_cpu {
45 	struct update_util_data update_util;
46 	struct sugov_policy *sg_policy;
47 
48 	/* The fields below are only needed when sharing a policy. */
49 	unsigned long util;
50 	unsigned long max;
51 	u64 last_update;
52 };
53 
54 static DEFINE_PER_CPU(struct sugov_cpu, sugov_cpu);
55 
56 /************************ Governor internals ***********************/
57 
58 static bool sugov_should_update_freq(struct sugov_policy *sg_policy, u64 time)
59 {
60 	s64 delta_ns;
61 
62 	if (sg_policy->work_in_progress)
63 		return false;
64 
65 	if (unlikely(sg_policy->need_freq_update)) {
66 		sg_policy->need_freq_update = false;
67 		/*
68 		 * This happens when limits change, so forget the previous
69 		 * next_freq value and force an update.
70 		 */
71 		sg_policy->next_freq = UINT_MAX;
72 		return true;
73 	}
74 
75 	delta_ns = time - sg_policy->last_freq_update_time;
76 	return delta_ns >= sg_policy->freq_update_delay_ns;
77 }
78 
79 static void sugov_update_commit(struct sugov_policy *sg_policy, u64 time,
80 				unsigned int next_freq)
81 {
82 	struct cpufreq_policy *policy = sg_policy->policy;
83 
84 	sg_policy->last_freq_update_time = time;
85 
86 	if (policy->fast_switch_enabled) {
87 		if (sg_policy->next_freq == next_freq) {
88 			trace_cpu_frequency(policy->cur, smp_processor_id());
89 			return;
90 		}
91 		sg_policy->next_freq = next_freq;
92 		next_freq = cpufreq_driver_fast_switch(policy, next_freq);
93 		if (next_freq == CPUFREQ_ENTRY_INVALID)
94 			return;
95 
96 		policy->cur = next_freq;
97 		trace_cpu_frequency(next_freq, smp_processor_id());
98 	} else if (sg_policy->next_freq != next_freq) {
99 		sg_policy->next_freq = next_freq;
100 		sg_policy->work_in_progress = true;
101 		irq_work_queue(&sg_policy->irq_work);
102 	}
103 }
104 
105 /**
106  * get_next_freq - Compute a new frequency for a given cpufreq policy.
107  * @policy: cpufreq policy object to compute the new frequency for.
108  * @util: Current CPU utilization.
109  * @max: CPU capacity.
110  *
111  * If the utilization is frequency-invariant, choose the new frequency to be
112  * proportional to it, that is
113  *
114  * next_freq = C * max_freq * util / max
115  *
116  * Otherwise, approximate the would-be frequency-invariant utilization by
117  * util_raw * (curr_freq / max_freq) which leads to
118  *
119  * next_freq = C * curr_freq * util_raw / max
120  *
121  * Take C = 1.25 for the frequency tipping point at (util / max) = 0.8.
122  */
123 static unsigned int get_next_freq(struct cpufreq_policy *policy,
124 				  unsigned long util, unsigned long max)
125 {
126 	unsigned int freq = arch_scale_freq_invariant() ?
127 				policy->cpuinfo.max_freq : policy->cur;
128 
129 	return (freq + (freq >> 2)) * util / max;
130 }
131 
132 static void sugov_update_single(struct update_util_data *hook, u64 time,
133 				unsigned long util, unsigned long max)
134 {
135 	struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
136 	struct sugov_policy *sg_policy = sg_cpu->sg_policy;
137 	struct cpufreq_policy *policy = sg_policy->policy;
138 	unsigned int next_f;
139 
140 	if (!sugov_should_update_freq(sg_policy, time))
141 		return;
142 
143 	next_f = util == ULONG_MAX ? policy->cpuinfo.max_freq :
144 			get_next_freq(policy, util, max);
145 	sugov_update_commit(sg_policy, time, next_f);
146 }
147 
148 static unsigned int sugov_next_freq_shared(struct sugov_policy *sg_policy,
149 					   unsigned long util, unsigned long max)
150 {
151 	struct cpufreq_policy *policy = sg_policy->policy;
152 	unsigned int max_f = policy->cpuinfo.max_freq;
153 	u64 last_freq_update_time = sg_policy->last_freq_update_time;
154 	unsigned int j;
155 
156 	if (util == ULONG_MAX)
157 		return max_f;
158 
159 	for_each_cpu(j, policy->cpus) {
160 		struct sugov_cpu *j_sg_cpu;
161 		unsigned long j_util, j_max;
162 		s64 delta_ns;
163 
164 		if (j == smp_processor_id())
165 			continue;
166 
167 		j_sg_cpu = &per_cpu(sugov_cpu, j);
168 		/*
169 		 * If the CPU utilization was last updated before the previous
170 		 * frequency update and the time elapsed between the last update
171 		 * of the CPU utilization and the last frequency update is long
172 		 * enough, don't take the CPU into account as it probably is
173 		 * idle now.
174 		 */
175 		delta_ns = last_freq_update_time - j_sg_cpu->last_update;
176 		if (delta_ns > TICK_NSEC)
177 			continue;
178 
179 		j_util = j_sg_cpu->util;
180 		if (j_util == ULONG_MAX)
181 			return max_f;
182 
183 		j_max = j_sg_cpu->max;
184 		if (j_util * max > j_max * util) {
185 			util = j_util;
186 			max = j_max;
187 		}
188 	}
189 
190 	return get_next_freq(policy, util, max);
191 }
192 
193 static void sugov_update_shared(struct update_util_data *hook, u64 time,
194 				unsigned long util, unsigned long max)
195 {
196 	struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
197 	struct sugov_policy *sg_policy = sg_cpu->sg_policy;
198 	unsigned int next_f;
199 
200 	raw_spin_lock(&sg_policy->update_lock);
201 
202 	sg_cpu->util = util;
203 	sg_cpu->max = max;
204 	sg_cpu->last_update = time;
205 
206 	if (sugov_should_update_freq(sg_policy, time)) {
207 		next_f = sugov_next_freq_shared(sg_policy, util, max);
208 		sugov_update_commit(sg_policy, time, next_f);
209 	}
210 
211 	raw_spin_unlock(&sg_policy->update_lock);
212 }
213 
214 static void sugov_work(struct work_struct *work)
215 {
216 	struct sugov_policy *sg_policy = container_of(work, struct sugov_policy, work);
217 
218 	mutex_lock(&sg_policy->work_lock);
219 	__cpufreq_driver_target(sg_policy->policy, sg_policy->next_freq,
220 				CPUFREQ_RELATION_L);
221 	mutex_unlock(&sg_policy->work_lock);
222 
223 	sg_policy->work_in_progress = false;
224 }
225 
226 static void sugov_irq_work(struct irq_work *irq_work)
227 {
228 	struct sugov_policy *sg_policy;
229 
230 	sg_policy = container_of(irq_work, struct sugov_policy, irq_work);
231 	schedule_work_on(smp_processor_id(), &sg_policy->work);
232 }
233 
234 /************************** sysfs interface ************************/
235 
236 static struct sugov_tunables *global_tunables;
237 static DEFINE_MUTEX(global_tunables_lock);
238 
239 static inline struct sugov_tunables *to_sugov_tunables(struct gov_attr_set *attr_set)
240 {
241 	return container_of(attr_set, struct sugov_tunables, attr_set);
242 }
243 
244 static ssize_t rate_limit_us_show(struct gov_attr_set *attr_set, char *buf)
245 {
246 	struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
247 
248 	return sprintf(buf, "%u\n", tunables->rate_limit_us);
249 }
250 
251 static ssize_t rate_limit_us_store(struct gov_attr_set *attr_set, const char *buf,
252 				   size_t count)
253 {
254 	struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
255 	struct sugov_policy *sg_policy;
256 	unsigned int rate_limit_us;
257 
258 	if (kstrtouint(buf, 10, &rate_limit_us))
259 		return -EINVAL;
260 
261 	tunables->rate_limit_us = rate_limit_us;
262 
263 	list_for_each_entry(sg_policy, &attr_set->policy_list, tunables_hook)
264 		sg_policy->freq_update_delay_ns = rate_limit_us * NSEC_PER_USEC;
265 
266 	return count;
267 }
268 
269 static struct governor_attr rate_limit_us = __ATTR_RW(rate_limit_us);
270 
271 static struct attribute *sugov_attributes[] = {
272 	&rate_limit_us.attr,
273 	NULL
274 };
275 
276 static struct kobj_type sugov_tunables_ktype = {
277 	.default_attrs = sugov_attributes,
278 	.sysfs_ops = &governor_sysfs_ops,
279 };
280 
281 /********************** cpufreq governor interface *********************/
282 
283 static struct cpufreq_governor schedutil_gov;
284 
285 static struct sugov_policy *sugov_policy_alloc(struct cpufreq_policy *policy)
286 {
287 	struct sugov_policy *sg_policy;
288 
289 	sg_policy = kzalloc(sizeof(*sg_policy), GFP_KERNEL);
290 	if (!sg_policy)
291 		return NULL;
292 
293 	sg_policy->policy = policy;
294 	init_irq_work(&sg_policy->irq_work, sugov_irq_work);
295 	INIT_WORK(&sg_policy->work, sugov_work);
296 	mutex_init(&sg_policy->work_lock);
297 	raw_spin_lock_init(&sg_policy->update_lock);
298 	return sg_policy;
299 }
300 
301 static void sugov_policy_free(struct sugov_policy *sg_policy)
302 {
303 	mutex_destroy(&sg_policy->work_lock);
304 	kfree(sg_policy);
305 }
306 
307 static struct sugov_tunables *sugov_tunables_alloc(struct sugov_policy *sg_policy)
308 {
309 	struct sugov_tunables *tunables;
310 
311 	tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
312 	if (tunables) {
313 		gov_attr_set_init(&tunables->attr_set, &sg_policy->tunables_hook);
314 		if (!have_governor_per_policy())
315 			global_tunables = tunables;
316 	}
317 	return tunables;
318 }
319 
320 static void sugov_tunables_free(struct sugov_tunables *tunables)
321 {
322 	if (!have_governor_per_policy())
323 		global_tunables = NULL;
324 
325 	kfree(tunables);
326 }
327 
328 static int sugov_init(struct cpufreq_policy *policy)
329 {
330 	struct sugov_policy *sg_policy;
331 	struct sugov_tunables *tunables;
332 	unsigned int lat;
333 	int ret = 0;
334 
335 	/* State should be equivalent to EXIT */
336 	if (policy->governor_data)
337 		return -EBUSY;
338 
339 	sg_policy = sugov_policy_alloc(policy);
340 	if (!sg_policy)
341 		return -ENOMEM;
342 
343 	mutex_lock(&global_tunables_lock);
344 
345 	if (global_tunables) {
346 		if (WARN_ON(have_governor_per_policy())) {
347 			ret = -EINVAL;
348 			goto free_sg_policy;
349 		}
350 		policy->governor_data = sg_policy;
351 		sg_policy->tunables = global_tunables;
352 
353 		gov_attr_set_get(&global_tunables->attr_set, &sg_policy->tunables_hook);
354 		goto out;
355 	}
356 
357 	tunables = sugov_tunables_alloc(sg_policy);
358 	if (!tunables) {
359 		ret = -ENOMEM;
360 		goto free_sg_policy;
361 	}
362 
363 	tunables->rate_limit_us = LATENCY_MULTIPLIER;
364 	lat = policy->cpuinfo.transition_latency / NSEC_PER_USEC;
365 	if (lat)
366 		tunables->rate_limit_us *= lat;
367 
368 	policy->governor_data = sg_policy;
369 	sg_policy->tunables = tunables;
370 
371 	ret = kobject_init_and_add(&tunables->attr_set.kobj, &sugov_tunables_ktype,
372 				   get_governor_parent_kobj(policy), "%s",
373 				   schedutil_gov.name);
374 	if (ret)
375 		goto fail;
376 
377  out:
378 	mutex_unlock(&global_tunables_lock);
379 
380 	cpufreq_enable_fast_switch(policy);
381 	return 0;
382 
383  fail:
384 	policy->governor_data = NULL;
385 	sugov_tunables_free(tunables);
386 
387  free_sg_policy:
388 	mutex_unlock(&global_tunables_lock);
389 
390 	sugov_policy_free(sg_policy);
391 	pr_err("cpufreq: schedutil governor initialization failed (error %d)\n", ret);
392 	return ret;
393 }
394 
395 static int sugov_exit(struct cpufreq_policy *policy)
396 {
397 	struct sugov_policy *sg_policy = policy->governor_data;
398 	struct sugov_tunables *tunables = sg_policy->tunables;
399 	unsigned int count;
400 
401 	cpufreq_disable_fast_switch(policy);
402 
403 	mutex_lock(&global_tunables_lock);
404 
405 	count = gov_attr_set_put(&tunables->attr_set, &sg_policy->tunables_hook);
406 	policy->governor_data = NULL;
407 	if (!count)
408 		sugov_tunables_free(tunables);
409 
410 	mutex_unlock(&global_tunables_lock);
411 
412 	sugov_policy_free(sg_policy);
413 	return 0;
414 }
415 
416 static int sugov_start(struct cpufreq_policy *policy)
417 {
418 	struct sugov_policy *sg_policy = policy->governor_data;
419 	unsigned int cpu;
420 
421 	sg_policy->freq_update_delay_ns = sg_policy->tunables->rate_limit_us * NSEC_PER_USEC;
422 	sg_policy->last_freq_update_time = 0;
423 	sg_policy->next_freq = UINT_MAX;
424 	sg_policy->work_in_progress = false;
425 	sg_policy->need_freq_update = false;
426 
427 	for_each_cpu(cpu, policy->cpus) {
428 		struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
429 
430 		sg_cpu->sg_policy = sg_policy;
431 		if (policy_is_shared(policy)) {
432 			sg_cpu->util = ULONG_MAX;
433 			sg_cpu->max = 0;
434 			sg_cpu->last_update = 0;
435 			cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util,
436 						     sugov_update_shared);
437 		} else {
438 			cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util,
439 						     sugov_update_single);
440 		}
441 	}
442 	return 0;
443 }
444 
445 static int sugov_stop(struct cpufreq_policy *policy)
446 {
447 	struct sugov_policy *sg_policy = policy->governor_data;
448 	unsigned int cpu;
449 
450 	for_each_cpu(cpu, policy->cpus)
451 		cpufreq_remove_update_util_hook(cpu);
452 
453 	synchronize_sched();
454 
455 	irq_work_sync(&sg_policy->irq_work);
456 	cancel_work_sync(&sg_policy->work);
457 	return 0;
458 }
459 
460 static int sugov_limits(struct cpufreq_policy *policy)
461 {
462 	struct sugov_policy *sg_policy = policy->governor_data;
463 
464 	if (!policy->fast_switch_enabled) {
465 		mutex_lock(&sg_policy->work_lock);
466 
467 		if (policy->max < policy->cur)
468 			__cpufreq_driver_target(policy, policy->max,
469 						CPUFREQ_RELATION_H);
470 		else if (policy->min > policy->cur)
471 			__cpufreq_driver_target(policy, policy->min,
472 						CPUFREQ_RELATION_L);
473 
474 		mutex_unlock(&sg_policy->work_lock);
475 	}
476 
477 	sg_policy->need_freq_update = true;
478 	return 0;
479 }
480 
481 int sugov_governor(struct cpufreq_policy *policy, unsigned int event)
482 {
483 	if (event == CPUFREQ_GOV_POLICY_INIT) {
484 		return sugov_init(policy);
485 	} else if (policy->governor_data) {
486 		switch (event) {
487 		case CPUFREQ_GOV_POLICY_EXIT:
488 			return sugov_exit(policy);
489 		case CPUFREQ_GOV_START:
490 			return sugov_start(policy);
491 		case CPUFREQ_GOV_STOP:
492 			return sugov_stop(policy);
493 		case CPUFREQ_GOV_LIMITS:
494 			return sugov_limits(policy);
495 		}
496 	}
497 	return -EINVAL;
498 }
499 
500 static struct cpufreq_governor schedutil_gov = {
501 	.name = "schedutil",
502 	.governor = sugov_governor,
503 	.owner = THIS_MODULE,
504 };
505 
506 static int __init sugov_module_init(void)
507 {
508 	return cpufreq_register_governor(&schedutil_gov);
509 }
510 
511 static void __exit sugov_module_exit(void)
512 {
513 	cpufreq_unregister_governor(&schedutil_gov);
514 }
515 
516 MODULE_AUTHOR("Rafael J. Wysocki <rafael.j.wysocki@intel.com>");
517 MODULE_DESCRIPTION("Utilization-based CPU frequency selection");
518 MODULE_LICENSE("GPL");
519 
520 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL
521 struct cpufreq_governor *cpufreq_default_governor(void)
522 {
523 	return &schedutil_gov;
524 }
525 
526 fs_initcall(sugov_module_init);
527 #else
528 module_init(sugov_module_init);
529 #endif
530 module_exit(sugov_module_exit);
531