]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blob - kernel/sched/cpufreq_schedutil.c
cpufreq: conservative: Allow down_threshold to take values from 1 to 10
[mirror_ubuntu-bionic-kernel.git] / kernel / sched / cpufreq_schedutil.c
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)
246 {
247 struct sugov_policy *sg_policy = sg_cpu->sg_policy;
248 struct cpufreq_policy *policy = sg_policy->policy;
249 u64 last_freq_update_time = sg_policy->last_freq_update_time;
250 unsigned long util = 0, max = 1;
251 unsigned int j;
252
253 for_each_cpu(j, policy->cpus) {
254 struct sugov_cpu *j_sg_cpu = &per_cpu(sugov_cpu, j);
255 unsigned long j_util, j_max;
256 s64 delta_ns;
257
258 /*
259 * If the CPU utilization was last updated before the previous
260 * frequency update and the time elapsed between the last update
261 * of the CPU utilization and the last frequency update is long
262 * enough, don't take the CPU into account as it probably is
263 * idle now (and clear iowait_boost for it).
264 */
265 delta_ns = last_freq_update_time - j_sg_cpu->last_update;
266 if (delta_ns > TICK_NSEC) {
267 j_sg_cpu->iowait_boost = 0;
268 continue;
269 }
270 if (j_sg_cpu->flags & SCHED_CPUFREQ_RT_DL)
271 return policy->cpuinfo.max_freq;
272
273 j_util = j_sg_cpu->util;
274 j_max = j_sg_cpu->max;
275 if (j_util * max > j_max * util) {
276 util = j_util;
277 max = j_max;
278 }
279
280 sugov_iowait_boost(j_sg_cpu, &util, &max);
281 }
282
283 return get_next_freq(sg_policy, util, max);
284 }
285
286 static void sugov_update_shared(struct update_util_data *hook, u64 time,
287 unsigned int flags)
288 {
289 struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
290 struct sugov_policy *sg_policy = sg_cpu->sg_policy;
291 unsigned long util, max;
292 unsigned int next_f;
293
294 sugov_get_util(&util, &max);
295
296 raw_spin_lock(&sg_policy->update_lock);
297
298 sg_cpu->util = util;
299 sg_cpu->max = max;
300 sg_cpu->flags = flags;
301
302 sugov_set_iowait_boost(sg_cpu, time, flags);
303 sg_cpu->last_update = time;
304
305 if (sugov_should_update_freq(sg_policy, time)) {
306 if (flags & SCHED_CPUFREQ_RT_DL)
307 next_f = sg_policy->policy->cpuinfo.max_freq;
308 else
309 next_f = sugov_next_freq_shared(sg_cpu);
310
311 sugov_update_commit(sg_policy, time, next_f);
312 }
313
314 raw_spin_unlock(&sg_policy->update_lock);
315 }
316
317 static void sugov_work(struct kthread_work *work)
318 {
319 struct sugov_policy *sg_policy = container_of(work, struct sugov_policy, work);
320
321 mutex_lock(&sg_policy->work_lock);
322 __cpufreq_driver_target(sg_policy->policy, sg_policy->next_freq,
323 CPUFREQ_RELATION_L);
324 mutex_unlock(&sg_policy->work_lock);
325
326 sg_policy->work_in_progress = false;
327 }
328
329 static void sugov_irq_work(struct irq_work *irq_work)
330 {
331 struct sugov_policy *sg_policy;
332
333 sg_policy = container_of(irq_work, struct sugov_policy, irq_work);
334
335 /*
336 * For RT and deadline tasks, the schedutil governor shoots the
337 * frequency to maximum. Special care must be taken to ensure that this
338 * kthread doesn't result in the same behavior.
339 *
340 * This is (mostly) guaranteed by the work_in_progress flag. The flag is
341 * updated only at the end of the sugov_work() function and before that
342 * the schedutil governor rejects all other frequency scaling requests.
343 *
344 * There is a very rare case though, where the RT thread yields right
345 * after the work_in_progress flag is cleared. The effects of that are
346 * neglected for now.
347 */
348 kthread_queue_work(&sg_policy->worker, &sg_policy->work);
349 }
350
351 /************************** sysfs interface ************************/
352
353 static struct sugov_tunables *global_tunables;
354 static DEFINE_MUTEX(global_tunables_lock);
355
356 static inline struct sugov_tunables *to_sugov_tunables(struct gov_attr_set *attr_set)
357 {
358 return container_of(attr_set, struct sugov_tunables, attr_set);
359 }
360
361 static ssize_t rate_limit_us_show(struct gov_attr_set *attr_set, char *buf)
362 {
363 struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
364
365 return sprintf(buf, "%u\n", tunables->rate_limit_us);
366 }
367
368 static ssize_t rate_limit_us_store(struct gov_attr_set *attr_set, const char *buf,
369 size_t count)
370 {
371 struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
372 struct sugov_policy *sg_policy;
373 unsigned int rate_limit_us;
374
375 if (kstrtouint(buf, 10, &rate_limit_us))
376 return -EINVAL;
377
378 tunables->rate_limit_us = rate_limit_us;
379
380 list_for_each_entry(sg_policy, &attr_set->policy_list, tunables_hook)
381 sg_policy->freq_update_delay_ns = rate_limit_us * NSEC_PER_USEC;
382
383 return count;
384 }
385
386 static struct governor_attr rate_limit_us = __ATTR_RW(rate_limit_us);
387
388 static struct attribute *sugov_attributes[] = {
389 &rate_limit_us.attr,
390 NULL
391 };
392
393 static struct kobj_type sugov_tunables_ktype = {
394 .default_attrs = sugov_attributes,
395 .sysfs_ops = &governor_sysfs_ops,
396 };
397
398 /********************** cpufreq governor interface *********************/
399
400 static struct cpufreq_governor schedutil_gov;
401
402 static struct sugov_policy *sugov_policy_alloc(struct cpufreq_policy *policy)
403 {
404 struct sugov_policy *sg_policy;
405
406 sg_policy = kzalloc(sizeof(*sg_policy), GFP_KERNEL);
407 if (!sg_policy)
408 return NULL;
409
410 sg_policy->policy = policy;
411 raw_spin_lock_init(&sg_policy->update_lock);
412 return sg_policy;
413 }
414
415 static void sugov_policy_free(struct sugov_policy *sg_policy)
416 {
417 kfree(sg_policy);
418 }
419
420 static int sugov_kthread_create(struct sugov_policy *sg_policy)
421 {
422 struct task_struct *thread;
423 struct sched_param param = { .sched_priority = MAX_USER_RT_PRIO / 2 };
424 struct cpufreq_policy *policy = sg_policy->policy;
425 int ret;
426
427 /* kthread only required for slow path */
428 if (policy->fast_switch_enabled)
429 return 0;
430
431 kthread_init_work(&sg_policy->work, sugov_work);
432 kthread_init_worker(&sg_policy->worker);
433 thread = kthread_create(kthread_worker_fn, &sg_policy->worker,
434 "sugov:%d",
435 cpumask_first(policy->related_cpus));
436 if (IS_ERR(thread)) {
437 pr_err("failed to create sugov thread: %ld\n", PTR_ERR(thread));
438 return PTR_ERR(thread);
439 }
440
441 ret = sched_setscheduler_nocheck(thread, SCHED_FIFO, &param);
442 if (ret) {
443 kthread_stop(thread);
444 pr_warn("%s: failed to set SCHED_FIFO\n", __func__);
445 return ret;
446 }
447
448 sg_policy->thread = thread;
449 kthread_bind_mask(thread, policy->related_cpus);
450 init_irq_work(&sg_policy->irq_work, sugov_irq_work);
451 mutex_init(&sg_policy->work_lock);
452
453 wake_up_process(thread);
454
455 return 0;
456 }
457
458 static void sugov_kthread_stop(struct sugov_policy *sg_policy)
459 {
460 /* kthread only required for slow path */
461 if (sg_policy->policy->fast_switch_enabled)
462 return;
463
464 kthread_flush_worker(&sg_policy->worker);
465 kthread_stop(sg_policy->thread);
466 mutex_destroy(&sg_policy->work_lock);
467 }
468
469 static struct sugov_tunables *sugov_tunables_alloc(struct sugov_policy *sg_policy)
470 {
471 struct sugov_tunables *tunables;
472
473 tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
474 if (tunables) {
475 gov_attr_set_init(&tunables->attr_set, &sg_policy->tunables_hook);
476 if (!have_governor_per_policy())
477 global_tunables = tunables;
478 }
479 return tunables;
480 }
481
482 static void sugov_tunables_free(struct sugov_tunables *tunables)
483 {
484 if (!have_governor_per_policy())
485 global_tunables = NULL;
486
487 kfree(tunables);
488 }
489
490 static int sugov_init(struct cpufreq_policy *policy)
491 {
492 struct sugov_policy *sg_policy;
493 struct sugov_tunables *tunables;
494 int ret = 0;
495
496 /* State should be equivalent to EXIT */
497 if (policy->governor_data)
498 return -EBUSY;
499
500 cpufreq_enable_fast_switch(policy);
501
502 sg_policy = sugov_policy_alloc(policy);
503 if (!sg_policy) {
504 ret = -ENOMEM;
505 goto disable_fast_switch;
506 }
507
508 ret = sugov_kthread_create(sg_policy);
509 if (ret)
510 goto free_sg_policy;
511
512 mutex_lock(&global_tunables_lock);
513
514 if (global_tunables) {
515 if (WARN_ON(have_governor_per_policy())) {
516 ret = -EINVAL;
517 goto stop_kthread;
518 }
519 policy->governor_data = sg_policy;
520 sg_policy->tunables = global_tunables;
521
522 gov_attr_set_get(&global_tunables->attr_set, &sg_policy->tunables_hook);
523 goto out;
524 }
525
526 tunables = sugov_tunables_alloc(sg_policy);
527 if (!tunables) {
528 ret = -ENOMEM;
529 goto stop_kthread;
530 }
531
532 if (policy->transition_delay_us) {
533 tunables->rate_limit_us = policy->transition_delay_us;
534 } else {
535 unsigned int lat;
536
537 tunables->rate_limit_us = LATENCY_MULTIPLIER;
538 lat = policy->cpuinfo.transition_latency / NSEC_PER_USEC;
539 if (lat)
540 tunables->rate_limit_us *= lat;
541 }
542
543 policy->governor_data = sg_policy;
544 sg_policy->tunables = tunables;
545
546 ret = kobject_init_and_add(&tunables->attr_set.kobj, &sugov_tunables_ktype,
547 get_governor_parent_kobj(policy), "%s",
548 schedutil_gov.name);
549 if (ret)
550 goto fail;
551
552 out:
553 mutex_unlock(&global_tunables_lock);
554 return 0;
555
556 fail:
557 policy->governor_data = NULL;
558 sugov_tunables_free(tunables);
559
560 stop_kthread:
561 sugov_kthread_stop(sg_policy);
562
563 free_sg_policy:
564 mutex_unlock(&global_tunables_lock);
565
566 sugov_policy_free(sg_policy);
567
568 disable_fast_switch:
569 cpufreq_disable_fast_switch(policy);
570
571 pr_err("initialization failed (error %d)\n", ret);
572 return ret;
573 }
574
575 static void sugov_exit(struct cpufreq_policy *policy)
576 {
577 struct sugov_policy *sg_policy = policy->governor_data;
578 struct sugov_tunables *tunables = sg_policy->tunables;
579 unsigned int count;
580
581 mutex_lock(&global_tunables_lock);
582
583 count = gov_attr_set_put(&tunables->attr_set, &sg_policy->tunables_hook);
584 policy->governor_data = NULL;
585 if (!count)
586 sugov_tunables_free(tunables);
587
588 mutex_unlock(&global_tunables_lock);
589
590 sugov_kthread_stop(sg_policy);
591 sugov_policy_free(sg_policy);
592 cpufreq_disable_fast_switch(policy);
593 }
594
595 static int sugov_start(struct cpufreq_policy *policy)
596 {
597 struct sugov_policy *sg_policy = policy->governor_data;
598 unsigned int cpu;
599
600 sg_policy->freq_update_delay_ns = sg_policy->tunables->rate_limit_us * NSEC_PER_USEC;
601 sg_policy->last_freq_update_time = 0;
602 sg_policy->next_freq = UINT_MAX;
603 sg_policy->work_in_progress = false;
604 sg_policy->need_freq_update = false;
605 sg_policy->cached_raw_freq = 0;
606
607 for_each_cpu(cpu, policy->cpus) {
608 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
609
610 memset(sg_cpu, 0, sizeof(*sg_cpu));
611 sg_cpu->sg_policy = sg_policy;
612 sg_cpu->flags = SCHED_CPUFREQ_RT;
613 sg_cpu->iowait_boost_max = policy->cpuinfo.max_freq;
614 cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util,
615 policy_is_shared(policy) ?
616 sugov_update_shared :
617 sugov_update_single);
618 }
619 return 0;
620 }
621
622 static void sugov_stop(struct cpufreq_policy *policy)
623 {
624 struct sugov_policy *sg_policy = policy->governor_data;
625 unsigned int cpu;
626
627 for_each_cpu(cpu, policy->cpus)
628 cpufreq_remove_update_util_hook(cpu);
629
630 synchronize_sched();
631
632 if (!policy->fast_switch_enabled) {
633 irq_work_sync(&sg_policy->irq_work);
634 kthread_cancel_work_sync(&sg_policy->work);
635 }
636 }
637
638 static void sugov_limits(struct cpufreq_policy *policy)
639 {
640 struct sugov_policy *sg_policy = policy->governor_data;
641
642 if (!policy->fast_switch_enabled) {
643 mutex_lock(&sg_policy->work_lock);
644 cpufreq_policy_apply_limits(policy);
645 mutex_unlock(&sg_policy->work_lock);
646 }
647
648 sg_policy->need_freq_update = true;
649 }
650
651 static struct cpufreq_governor schedutil_gov = {
652 .name = "schedutil",
653 .owner = THIS_MODULE,
654 .init = sugov_init,
655 .exit = sugov_exit,
656 .start = sugov_start,
657 .stop = sugov_stop,
658 .limits = sugov_limits,
659 };
660
661 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL
662 struct cpufreq_governor *cpufreq_default_governor(void)
663 {
664 return &schedutil_gov;
665 }
666 #endif
667
668 static int __init sugov_register(void)
669 {
670 return cpufreq_register_governor(&schedutil_gov);
671 }
672 fs_initcall(sugov_register);