Merge branches 'intel_pstate', 'pm-cpufreq' and 'pm-cpufreq-sched'
[sfrench/cifs-2.6.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         if (sg_policy->next_freq > next_freq)
105                 next_freq = (sg_policy->next_freq + next_freq) >> 1;
106
107         sg_policy->next_freq = next_freq;
108         sg_policy->last_freq_update_time = time;
109
110         if (policy->fast_switch_enabled) {
111                 next_freq = cpufreq_driver_fast_switch(policy, next_freq);
112                 if (next_freq == CPUFREQ_ENTRY_INVALID)
113                         return;
114
115                 policy->cur = next_freq;
116                 trace_cpu_frequency(next_freq, smp_processor_id());
117         } else {
118                 sg_policy->work_in_progress = true;
119                 irq_work_queue(&sg_policy->irq_work);
120         }
121 }
122
123 /**
124  * get_next_freq - Compute a new frequency for a given cpufreq policy.
125  * @sg_policy: schedutil policy object to compute the new frequency for.
126  * @util: Current CPU utilization.
127  * @max: CPU capacity.
128  *
129  * If the utilization is frequency-invariant, choose the new frequency to be
130  * proportional to it, that is
131  *
132  * next_freq = C * max_freq * util / max
133  *
134  * Otherwise, approximate the would-be frequency-invariant utilization by
135  * util_raw * (curr_freq / max_freq) which leads to
136  *
137  * next_freq = C * curr_freq * util_raw / max
138  *
139  * Take C = 1.25 for the frequency tipping point at (util / max) = 0.8.
140  *
141  * The lowest driver-supported frequency which is equal or greater than the raw
142  * next_freq (as calculated above) is returned, subject to policy min/max and
143  * cpufreq driver limitations.
144  */
145 static unsigned int get_next_freq(struct sugov_policy *sg_policy,
146                                   unsigned long util, unsigned long max)
147 {
148         struct cpufreq_policy *policy = sg_policy->policy;
149         unsigned int freq = arch_scale_freq_invariant() ?
150                                 policy->cpuinfo.max_freq : policy->cur;
151
152         freq = (freq + (freq >> 2)) * util / max;
153
154         if (freq == sg_policy->cached_raw_freq && sg_policy->next_freq != UINT_MAX)
155                 return sg_policy->next_freq;
156         sg_policy->cached_raw_freq = freq;
157         return cpufreq_driver_resolve_freq(policy, freq);
158 }
159
160 static void sugov_get_util(unsigned long *util, unsigned long *max)
161 {
162         struct rq *rq = this_rq();
163         unsigned long cfs_max;
164
165         cfs_max = arch_scale_cpu_capacity(NULL, smp_processor_id());
166
167         *util = min(rq->cfs.avg.util_avg, cfs_max);
168         *max = cfs_max;
169 }
170
171 static void sugov_set_iowait_boost(struct sugov_cpu *sg_cpu, u64 time,
172                                    unsigned int flags)
173 {
174         if (flags & SCHED_CPUFREQ_IOWAIT) {
175                 sg_cpu->iowait_boost = sg_cpu->iowait_boost_max;
176         } else if (sg_cpu->iowait_boost) {
177                 s64 delta_ns = time - sg_cpu->last_update;
178
179                 /* Clear iowait_boost if the CPU apprears to have been idle. */
180                 if (delta_ns > TICK_NSEC)
181                         sg_cpu->iowait_boost = 0;
182         }
183 }
184
185 static void sugov_iowait_boost(struct sugov_cpu *sg_cpu, unsigned long *util,
186                                unsigned long *max)
187 {
188         unsigned long boost_util = sg_cpu->iowait_boost;
189         unsigned long boost_max = sg_cpu->iowait_boost_max;
190
191         if (!boost_util)
192                 return;
193
194         if (*util * boost_max < *max * boost_util) {
195                 *util = boost_util;
196                 *max = boost_max;
197         }
198         sg_cpu->iowait_boost >>= 1;
199 }
200
201 #ifdef CONFIG_NO_HZ_COMMON
202 static bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu)
203 {
204         unsigned long idle_calls = tick_nohz_get_idle_calls();
205         bool ret = idle_calls == sg_cpu->saved_idle_calls;
206
207         sg_cpu->saved_idle_calls = idle_calls;
208         return ret;
209 }
210 #else
211 static inline bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu) { return false; }
212 #endif /* CONFIG_NO_HZ_COMMON */
213
214 static void sugov_update_single(struct update_util_data *hook, u64 time,
215                                 unsigned int flags)
216 {
217         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
218         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
219         struct cpufreq_policy *policy = sg_policy->policy;
220         unsigned long util, max;
221         unsigned int next_f;
222         bool busy;
223
224         sugov_set_iowait_boost(sg_cpu, time, flags);
225         sg_cpu->last_update = time;
226
227         if (!sugov_should_update_freq(sg_policy, time))
228                 return;
229
230         busy = sugov_cpu_is_busy(sg_cpu);
231
232         if (flags & SCHED_CPUFREQ_RT_DL) {
233                 next_f = policy->cpuinfo.max_freq;
234         } else {
235                 sugov_get_util(&util, &max);
236                 sugov_iowait_boost(sg_cpu, &util, &max);
237                 next_f = get_next_freq(sg_policy, util, max);
238                 /*
239                  * Do not reduce the frequency if the CPU has not been idle
240                  * recently, as the reduction is likely to be premature then.
241                  */
242                 if (busy && next_f < sg_policy->next_freq)
243                         next_f = sg_policy->next_freq;
244         }
245         sugov_update_commit(sg_policy, time, next_f);
246 }
247
248 static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu, u64 time)
249 {
250         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
251         struct cpufreq_policy *policy = sg_policy->policy;
252         unsigned long util = 0, max = 1;
253         unsigned int j;
254
255         for_each_cpu(j, policy->cpus) {
256                 struct sugov_cpu *j_sg_cpu = &per_cpu(sugov_cpu, j);
257                 unsigned long j_util, j_max;
258                 s64 delta_ns;
259
260                 /*
261                  * If the CPU utilization was last updated before the previous
262                  * frequency update and the time elapsed between the last update
263                  * of the CPU utilization and the last frequency update is long
264                  * enough, don't take the CPU into account as it probably is
265                  * idle now (and clear iowait_boost for it).
266                  */
267                 delta_ns = time - j_sg_cpu->last_update;
268                 if (delta_ns > TICK_NSEC) {
269                         j_sg_cpu->iowait_boost = 0;
270                         continue;
271                 }
272                 if (j_sg_cpu->flags & SCHED_CPUFREQ_RT_DL)
273                         return policy->cpuinfo.max_freq;
274
275                 j_util = j_sg_cpu->util;
276                 j_max = j_sg_cpu->max;
277                 if (j_util * max > j_max * util) {
278                         util = j_util;
279                         max = j_max;
280                 }
281
282                 sugov_iowait_boost(j_sg_cpu, &util, &max);
283         }
284
285         return get_next_freq(sg_policy, util, max);
286 }
287
288 static void sugov_update_shared(struct update_util_data *hook, u64 time,
289                                 unsigned int flags)
290 {
291         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
292         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
293         unsigned long util, max;
294         unsigned int next_f;
295
296         sugov_get_util(&util, &max);
297
298         raw_spin_lock(&sg_policy->update_lock);
299
300         sg_cpu->util = util;
301         sg_cpu->max = max;
302         sg_cpu->flags = flags;
303
304         sugov_set_iowait_boost(sg_cpu, time, flags);
305         sg_cpu->last_update = time;
306
307         if (sugov_should_update_freq(sg_policy, time)) {
308                 if (flags & SCHED_CPUFREQ_RT_DL)
309                         next_f = sg_policy->policy->cpuinfo.max_freq;
310                 else
311                         next_f = sugov_next_freq_shared(sg_cpu, time);
312
313                 sugov_update_commit(sg_policy, time, next_f);
314         }
315
316         raw_spin_unlock(&sg_policy->update_lock);
317 }
318
319 static void sugov_work(struct kthread_work *work)
320 {
321         struct sugov_policy *sg_policy = container_of(work, struct sugov_policy, work);
322
323         mutex_lock(&sg_policy->work_lock);
324         __cpufreq_driver_target(sg_policy->policy, sg_policy->next_freq,
325                                 CPUFREQ_RELATION_L);
326         mutex_unlock(&sg_policy->work_lock);
327
328         sg_policy->work_in_progress = false;
329 }
330
331 static void sugov_irq_work(struct irq_work *irq_work)
332 {
333         struct sugov_policy *sg_policy;
334
335         sg_policy = container_of(irq_work, struct sugov_policy, irq_work);
336
337         /*
338          * For RT and deadline tasks, the schedutil governor shoots the
339          * frequency to maximum. Special care must be taken to ensure that this
340          * kthread doesn't result in the same behavior.
341          *
342          * This is (mostly) guaranteed by the work_in_progress flag. The flag is
343          * updated only at the end of the sugov_work() function and before that
344          * the schedutil governor rejects all other frequency scaling requests.
345          *
346          * There is a very rare case though, where the RT thread yields right
347          * after the work_in_progress flag is cleared. The effects of that are
348          * neglected for now.
349          */
350         kthread_queue_work(&sg_policy->worker, &sg_policy->work);
351 }
352
353 /************************** sysfs interface ************************/
354
355 static struct sugov_tunables *global_tunables;
356 static DEFINE_MUTEX(global_tunables_lock);
357
358 static inline struct sugov_tunables *to_sugov_tunables(struct gov_attr_set *attr_set)
359 {
360         return container_of(attr_set, struct sugov_tunables, attr_set);
361 }
362
363 static ssize_t rate_limit_us_show(struct gov_attr_set *attr_set, char *buf)
364 {
365         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
366
367         return sprintf(buf, "%u\n", tunables->rate_limit_us);
368 }
369
370 static ssize_t rate_limit_us_store(struct gov_attr_set *attr_set, const char *buf,
371                                    size_t count)
372 {
373         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
374         struct sugov_policy *sg_policy;
375         unsigned int rate_limit_us;
376
377         if (kstrtouint(buf, 10, &rate_limit_us))
378                 return -EINVAL;
379
380         tunables->rate_limit_us = rate_limit_us;
381
382         list_for_each_entry(sg_policy, &attr_set->policy_list, tunables_hook)
383                 sg_policy->freq_update_delay_ns = rate_limit_us * NSEC_PER_USEC;
384
385         return count;
386 }
387
388 static struct governor_attr rate_limit_us = __ATTR_RW(rate_limit_us);
389
390 static struct attribute *sugov_attributes[] = {
391         &rate_limit_us.attr,
392         NULL
393 };
394
395 static struct kobj_type sugov_tunables_ktype = {
396         .default_attrs = sugov_attributes,
397         .sysfs_ops = &governor_sysfs_ops,
398 };
399
400 /********************** cpufreq governor interface *********************/
401
402 static struct cpufreq_governor schedutil_gov;
403
404 static struct sugov_policy *sugov_policy_alloc(struct cpufreq_policy *policy)
405 {
406         struct sugov_policy *sg_policy;
407
408         sg_policy = kzalloc(sizeof(*sg_policy), GFP_KERNEL);
409         if (!sg_policy)
410                 return NULL;
411
412         sg_policy->policy = policy;
413         raw_spin_lock_init(&sg_policy->update_lock);
414         return sg_policy;
415 }
416
417 static void sugov_policy_free(struct sugov_policy *sg_policy)
418 {
419         kfree(sg_policy);
420 }
421
422 static int sugov_kthread_create(struct sugov_policy *sg_policy)
423 {
424         struct task_struct *thread;
425         struct sched_param param = { .sched_priority = MAX_USER_RT_PRIO / 2 };
426         struct cpufreq_policy *policy = sg_policy->policy;
427         int ret;
428
429         /* kthread only required for slow path */
430         if (policy->fast_switch_enabled)
431                 return 0;
432
433         kthread_init_work(&sg_policy->work, sugov_work);
434         kthread_init_worker(&sg_policy->worker);
435         thread = kthread_create(kthread_worker_fn, &sg_policy->worker,
436                                 "sugov:%d",
437                                 cpumask_first(policy->related_cpus));
438         if (IS_ERR(thread)) {
439                 pr_err("failed to create sugov thread: %ld\n", PTR_ERR(thread));
440                 return PTR_ERR(thread);
441         }
442
443         ret = sched_setscheduler_nocheck(thread, SCHED_FIFO, &param);
444         if (ret) {
445                 kthread_stop(thread);
446                 pr_warn("%s: failed to set SCHED_FIFO\n", __func__);
447                 return ret;
448         }
449
450         sg_policy->thread = thread;
451         kthread_bind_mask(thread, policy->related_cpus);
452         init_irq_work(&sg_policy->irq_work, sugov_irq_work);
453         mutex_init(&sg_policy->work_lock);
454
455         wake_up_process(thread);
456
457         return 0;
458 }
459
460 static void sugov_kthread_stop(struct sugov_policy *sg_policy)
461 {
462         /* kthread only required for slow path */
463         if (sg_policy->policy->fast_switch_enabled)
464                 return;
465
466         kthread_flush_worker(&sg_policy->worker);
467         kthread_stop(sg_policy->thread);
468         mutex_destroy(&sg_policy->work_lock);
469 }
470
471 static struct sugov_tunables *sugov_tunables_alloc(struct sugov_policy *sg_policy)
472 {
473         struct sugov_tunables *tunables;
474
475         tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
476         if (tunables) {
477                 gov_attr_set_init(&tunables->attr_set, &sg_policy->tunables_hook);
478                 if (!have_governor_per_policy())
479                         global_tunables = tunables;
480         }
481         return tunables;
482 }
483
484 static void sugov_tunables_free(struct sugov_tunables *tunables)
485 {
486         if (!have_governor_per_policy())
487                 global_tunables = NULL;
488
489         kfree(tunables);
490 }
491
492 static int sugov_init(struct cpufreq_policy *policy)
493 {
494         struct sugov_policy *sg_policy;
495         struct sugov_tunables *tunables;
496         int ret = 0;
497
498         /* State should be equivalent to EXIT */
499         if (policy->governor_data)
500                 return -EBUSY;
501
502         cpufreq_enable_fast_switch(policy);
503
504         sg_policy = sugov_policy_alloc(policy);
505         if (!sg_policy) {
506                 ret = -ENOMEM;
507                 goto disable_fast_switch;
508         }
509
510         ret = sugov_kthread_create(sg_policy);
511         if (ret)
512                 goto free_sg_policy;
513
514         mutex_lock(&global_tunables_lock);
515
516         if (global_tunables) {
517                 if (WARN_ON(have_governor_per_policy())) {
518                         ret = -EINVAL;
519                         goto stop_kthread;
520                 }
521                 policy->governor_data = sg_policy;
522                 sg_policy->tunables = global_tunables;
523
524                 gov_attr_set_get(&global_tunables->attr_set, &sg_policy->tunables_hook);
525                 goto out;
526         }
527
528         tunables = sugov_tunables_alloc(sg_policy);
529         if (!tunables) {
530                 ret = -ENOMEM;
531                 goto stop_kthread;
532         }
533
534         if (policy->transition_delay_us) {
535                 tunables->rate_limit_us = policy->transition_delay_us;
536         } else {
537                 unsigned int lat;
538
539                 tunables->rate_limit_us = LATENCY_MULTIPLIER;
540                 lat = policy->cpuinfo.transition_latency / NSEC_PER_USEC;
541                 if (lat)
542                         tunables->rate_limit_us *= lat;
543         }
544
545         policy->governor_data = sg_policy;
546         sg_policy->tunables = tunables;
547
548         ret = kobject_init_and_add(&tunables->attr_set.kobj, &sugov_tunables_ktype,
549                                    get_governor_parent_kobj(policy), "%s",
550                                    schedutil_gov.name);
551         if (ret)
552                 goto fail;
553
554 out:
555         mutex_unlock(&global_tunables_lock);
556         return 0;
557
558 fail:
559         policy->governor_data = NULL;
560         sugov_tunables_free(tunables);
561
562 stop_kthread:
563         sugov_kthread_stop(sg_policy);
564
565 free_sg_policy:
566         mutex_unlock(&global_tunables_lock);
567
568         sugov_policy_free(sg_policy);
569
570 disable_fast_switch:
571         cpufreq_disable_fast_switch(policy);
572
573         pr_err("initialization failed (error %d)\n", ret);
574         return ret;
575 }
576
577 static void sugov_exit(struct cpufreq_policy *policy)
578 {
579         struct sugov_policy *sg_policy = policy->governor_data;
580         struct sugov_tunables *tunables = sg_policy->tunables;
581         unsigned int count;
582
583         mutex_lock(&global_tunables_lock);
584
585         count = gov_attr_set_put(&tunables->attr_set, &sg_policy->tunables_hook);
586         policy->governor_data = NULL;
587         if (!count)
588                 sugov_tunables_free(tunables);
589
590         mutex_unlock(&global_tunables_lock);
591
592         sugov_kthread_stop(sg_policy);
593         sugov_policy_free(sg_policy);
594         cpufreq_disable_fast_switch(policy);
595 }
596
597 static int sugov_start(struct cpufreq_policy *policy)
598 {
599         struct sugov_policy *sg_policy = policy->governor_data;
600         unsigned int cpu;
601
602         sg_policy->freq_update_delay_ns = sg_policy->tunables->rate_limit_us * NSEC_PER_USEC;
603         sg_policy->last_freq_update_time = 0;
604         sg_policy->next_freq = UINT_MAX;
605         sg_policy->work_in_progress = false;
606         sg_policy->need_freq_update = false;
607         sg_policy->cached_raw_freq = 0;
608
609         for_each_cpu(cpu, policy->cpus) {
610                 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
611
612                 memset(sg_cpu, 0, sizeof(*sg_cpu));
613                 sg_cpu->sg_policy = sg_policy;
614                 sg_cpu->flags = SCHED_CPUFREQ_RT;
615                 sg_cpu->iowait_boost_max = policy->cpuinfo.max_freq;
616                 cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util,
617                                              policy_is_shared(policy) ?
618                                                         sugov_update_shared :
619                                                         sugov_update_single);
620         }
621         return 0;
622 }
623
624 static void sugov_stop(struct cpufreq_policy *policy)
625 {
626         struct sugov_policy *sg_policy = policy->governor_data;
627         unsigned int cpu;
628
629         for_each_cpu(cpu, policy->cpus)
630                 cpufreq_remove_update_util_hook(cpu);
631
632         synchronize_sched();
633
634         if (!policy->fast_switch_enabled) {
635                 irq_work_sync(&sg_policy->irq_work);
636                 kthread_cancel_work_sync(&sg_policy->work);
637         }
638 }
639
640 static void sugov_limits(struct cpufreq_policy *policy)
641 {
642         struct sugov_policy *sg_policy = policy->governor_data;
643
644         if (!policy->fast_switch_enabled) {
645                 mutex_lock(&sg_policy->work_lock);
646                 cpufreq_policy_apply_limits(policy);
647                 mutex_unlock(&sg_policy->work_lock);
648         }
649
650         sg_policy->need_freq_update = true;
651 }
652
653 static struct cpufreq_governor schedutil_gov = {
654         .name = "schedutil",
655         .owner = THIS_MODULE,
656         .init = sugov_init,
657         .exit = sugov_exit,
658         .start = sugov_start,
659         .stop = sugov_stop,
660         .limits = sugov_limits,
661 };
662
663 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL
664 struct cpufreq_governor *cpufreq_default_governor(void)
665 {
666         return &schedutil_gov;
667 }
668 #endif
669
670 static int __init sugov_register(void)
671 {
672         return cpufreq_register_governor(&schedutil_gov);
673 }
674 fs_initcall(sugov_register);