Loading...
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
23struct sugov_tunables {
24 struct gov_attr_set attr_set;
25 unsigned int rate_limit_us;
26};
27
28struct 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
50struct 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
65static DEFINE_PER_CPU(struct sugov_cpu, sugov_cpu);
66
67/************************ Governor internals ***********************/
68
69static 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
90static 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 */
138static 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
154static 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
165static 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
179static 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
195static 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
220static 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
272static 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
299static 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
311static 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
335static struct sugov_tunables *global_tunables;
336static DEFINE_MUTEX(global_tunables_lock);
337
338static 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
343static 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
350static 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
368static struct governor_attr rate_limit_us = __ATTR_RW(rate_limit_us);
369
370static struct attribute *sugov_attributes[] = {
371 &rate_limit_us.attr,
372 NULL
373};
374
375static struct kobj_type sugov_tunables_ktype = {
376 .default_attrs = sugov_attributes,
377 .sysfs_ops = &governor_sysfs_ops,
378};
379
380/********************** cpufreq governor interface *********************/
381
382static struct cpufreq_governor schedutil_gov;
383
384static 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
397static void sugov_policy_free(struct sugov_policy *sg_policy)
398{
399 kfree(sg_policy);
400}
401
402static 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, ¶m);
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
440static 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
451static 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
464static 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
472static 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
529out:
530 mutex_unlock(&global_tunables_lock);
531 return 0;
532
533fail:
534 policy->governor_data = NULL;
535 sugov_tunables_free(tunables);
536
537stop_kthread:
538 sugov_kthread_stop(sg_policy);
539
540free_sg_policy:
541 mutex_unlock(&global_tunables_lock);
542
543 sugov_policy_free(sg_policy);
544
545disable_fast_switch:
546 cpufreq_disable_fast_switch(policy);
547
548 pr_err("initialization failed (error %d)\n", ret);
549 return ret;
550}
551
552static 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
572static 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
605static 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
621static 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
634static 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
645struct cpufreq_governor *cpufreq_default_governor(void)
646{
647 return &schedutil_gov;
648}
649#endif
650
651static int __init sugov_register(void)
652{
653 return cpufreq_register_governor(&schedutil_gov);
654}
655fs_initcall(sugov_register);
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * CPUFreq governor based on scheduler-provided CPU utilization data.
4 *
5 * Copyright (C) 2016, Intel Corporation
6 * Author: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
7 */
8
9#define IOWAIT_BOOST_MIN (SCHED_CAPACITY_SCALE / 8)
10
11struct sugov_tunables {
12 struct gov_attr_set attr_set;
13 unsigned int rate_limit_us;
14};
15
16struct sugov_policy {
17 struct cpufreq_policy *policy;
18
19 struct sugov_tunables *tunables;
20 struct list_head tunables_hook;
21
22 raw_spinlock_t update_lock;
23 u64 last_freq_update_time;
24 s64 freq_update_delay_ns;
25 unsigned int next_freq;
26 unsigned int cached_raw_freq;
27
28 /* The next fields are only needed if fast switch cannot be used: */
29 struct irq_work irq_work;
30 struct kthread_work work;
31 struct mutex work_lock;
32 struct kthread_worker worker;
33 struct task_struct *thread;
34 bool work_in_progress;
35
36 bool limits_changed;
37 bool need_freq_update;
38};
39
40struct sugov_cpu {
41 struct update_util_data update_util;
42 struct sugov_policy *sg_policy;
43 unsigned int cpu;
44
45 bool iowait_boost_pending;
46 unsigned int iowait_boost;
47 u64 last_update;
48
49 unsigned long util;
50 unsigned long bw_dl;
51 unsigned long max;
52
53 /* The field below is for single-CPU policies only: */
54#ifdef CONFIG_NO_HZ_COMMON
55 unsigned long saved_idle_calls;
56#endif
57};
58
59static DEFINE_PER_CPU(struct sugov_cpu, sugov_cpu);
60
61/************************ Governor internals ***********************/
62
63static bool sugov_should_update_freq(struct sugov_policy *sg_policy, u64 time)
64{
65 s64 delta_ns;
66
67 /*
68 * Since cpufreq_update_util() is called with rq->lock held for
69 * the @target_cpu, our per-CPU data is fully serialized.
70 *
71 * However, drivers cannot in general deal with cross-CPU
72 * requests, so while get_next_freq() will work, our
73 * sugov_update_commit() call may not for the fast switching platforms.
74 *
75 * Hence stop here for remote requests if they aren't supported
76 * by the hardware, as calculating the frequency is pointless if
77 * we cannot in fact act on it.
78 *
79 * This is needed on the slow switching platforms too to prevent CPUs
80 * going offline from leaving stale IRQ work items behind.
81 */
82 if (!cpufreq_this_cpu_can_update(sg_policy->policy))
83 return false;
84
85 if (unlikely(sg_policy->limits_changed)) {
86 sg_policy->limits_changed = false;
87 sg_policy->need_freq_update = true;
88 return true;
89 }
90
91 delta_ns = time - sg_policy->last_freq_update_time;
92
93 return delta_ns >= sg_policy->freq_update_delay_ns;
94}
95
96static bool sugov_update_next_freq(struct sugov_policy *sg_policy, u64 time,
97 unsigned int next_freq)
98{
99 if (sg_policy->need_freq_update)
100 sg_policy->need_freq_update = cpufreq_driver_test_flags(CPUFREQ_NEED_UPDATE_LIMITS);
101 else if (sg_policy->next_freq == next_freq)
102 return false;
103
104 sg_policy->next_freq = next_freq;
105 sg_policy->last_freq_update_time = time;
106
107 return true;
108}
109
110static void sugov_deferred_update(struct sugov_policy *sg_policy)
111{
112 if (!sg_policy->work_in_progress) {
113 sg_policy->work_in_progress = true;
114 irq_work_queue(&sg_policy->irq_work);
115 }
116}
117
118/**
119 * get_next_freq - Compute a new frequency for a given cpufreq policy.
120 * @sg_policy: schedutil policy object to compute the new frequency for.
121 * @util: Current CPU utilization.
122 * @max: CPU capacity.
123 *
124 * If the utilization is frequency-invariant, choose the new frequency to be
125 * proportional to it, that is
126 *
127 * next_freq = C * max_freq * util / max
128 *
129 * Otherwise, approximate the would-be frequency-invariant utilization by
130 * util_raw * (curr_freq / max_freq) which leads to
131 *
132 * next_freq = C * curr_freq * util_raw / max
133 *
134 * Take C = 1.25 for the frequency tipping point at (util / max) = 0.8.
135 *
136 * The lowest driver-supported frequency which is equal or greater than the raw
137 * next_freq (as calculated above) is returned, subject to policy min/max and
138 * cpufreq driver limitations.
139 */
140static unsigned int get_next_freq(struct sugov_policy *sg_policy,
141 unsigned long util, unsigned long max)
142{
143 struct cpufreq_policy *policy = sg_policy->policy;
144 unsigned int freq = arch_scale_freq_invariant() ?
145 policy->cpuinfo.max_freq : policy->cur;
146
147 util = map_util_perf(util);
148 freq = map_util_freq(util, freq, max);
149
150 if (freq == sg_policy->cached_raw_freq && !sg_policy->need_freq_update)
151 return sg_policy->next_freq;
152
153 sg_policy->cached_raw_freq = freq;
154 return cpufreq_driver_resolve_freq(policy, freq);
155}
156
157static void sugov_get_util(struct sugov_cpu *sg_cpu)
158{
159 struct rq *rq = cpu_rq(sg_cpu->cpu);
160
161 sg_cpu->max = arch_scale_cpu_capacity(sg_cpu->cpu);
162 sg_cpu->bw_dl = cpu_bw_dl(rq);
163 sg_cpu->util = effective_cpu_util(sg_cpu->cpu, cpu_util_cfs(sg_cpu->cpu),
164 FREQUENCY_UTIL, NULL);
165}
166
167/**
168 * sugov_iowait_reset() - Reset the IO boost status of a CPU.
169 * @sg_cpu: the sugov data for the CPU to boost
170 * @time: the update time from the caller
171 * @set_iowait_boost: true if an IO boost has been requested
172 *
173 * The IO wait boost of a task is disabled after a tick since the last update
174 * of a CPU. If a new IO wait boost is requested after more then a tick, then
175 * we enable the boost starting from IOWAIT_BOOST_MIN, which improves energy
176 * efficiency by ignoring sporadic wakeups from IO.
177 */
178static bool sugov_iowait_reset(struct sugov_cpu *sg_cpu, u64 time,
179 bool set_iowait_boost)
180{
181 s64 delta_ns = time - sg_cpu->last_update;
182
183 /* Reset boost only if a tick has elapsed since last request */
184 if (delta_ns <= TICK_NSEC)
185 return false;
186
187 sg_cpu->iowait_boost = set_iowait_boost ? IOWAIT_BOOST_MIN : 0;
188 sg_cpu->iowait_boost_pending = set_iowait_boost;
189
190 return true;
191}
192
193/**
194 * sugov_iowait_boost() - Updates the IO boost status of a CPU.
195 * @sg_cpu: the sugov data for the CPU to boost
196 * @time: the update time from the caller
197 * @flags: SCHED_CPUFREQ_IOWAIT if the task is waking up after an IO wait
198 *
199 * Each time a task wakes up after an IO operation, the CPU utilization can be
200 * boosted to a certain utilization which doubles at each "frequent and
201 * successive" wakeup from IO, ranging from IOWAIT_BOOST_MIN to the utilization
202 * of the maximum OPP.
203 *
204 * To keep doubling, an IO boost has to be requested at least once per tick,
205 * otherwise we restart from the utilization of the minimum OPP.
206 */
207static void sugov_iowait_boost(struct sugov_cpu *sg_cpu, u64 time,
208 unsigned int flags)
209{
210 bool set_iowait_boost = flags & SCHED_CPUFREQ_IOWAIT;
211
212 /* Reset boost if the CPU appears to have been idle enough */
213 if (sg_cpu->iowait_boost &&
214 sugov_iowait_reset(sg_cpu, time, set_iowait_boost))
215 return;
216
217 /* Boost only tasks waking up after IO */
218 if (!set_iowait_boost)
219 return;
220
221 /* Ensure boost doubles only one time at each request */
222 if (sg_cpu->iowait_boost_pending)
223 return;
224 sg_cpu->iowait_boost_pending = true;
225
226 /* Double the boost at each request */
227 if (sg_cpu->iowait_boost) {
228 sg_cpu->iowait_boost =
229 min_t(unsigned int, sg_cpu->iowait_boost << 1, SCHED_CAPACITY_SCALE);
230 return;
231 }
232
233 /* First wakeup after IO: start with minimum boost */
234 sg_cpu->iowait_boost = IOWAIT_BOOST_MIN;
235}
236
237/**
238 * sugov_iowait_apply() - Apply the IO boost to a CPU.
239 * @sg_cpu: the sugov data for the cpu to boost
240 * @time: the update time from the caller
241 *
242 * A CPU running a task which woken up after an IO operation can have its
243 * utilization boosted to speed up the completion of those IO operations.
244 * The IO boost value is increased each time a task wakes up from IO, in
245 * sugov_iowait_apply(), and it's instead decreased by this function,
246 * each time an increase has not been requested (!iowait_boost_pending).
247 *
248 * A CPU which also appears to have been idle for at least one tick has also
249 * its IO boost utilization reset.
250 *
251 * This mechanism is designed to boost high frequently IO waiting tasks, while
252 * being more conservative on tasks which does sporadic IO operations.
253 */
254static void sugov_iowait_apply(struct sugov_cpu *sg_cpu, u64 time)
255{
256 unsigned long boost;
257
258 /* No boost currently required */
259 if (!sg_cpu->iowait_boost)
260 return;
261
262 /* Reset boost if the CPU appears to have been idle enough */
263 if (sugov_iowait_reset(sg_cpu, time, false))
264 return;
265
266 if (!sg_cpu->iowait_boost_pending) {
267 /*
268 * No boost pending; reduce the boost value.
269 */
270 sg_cpu->iowait_boost >>= 1;
271 if (sg_cpu->iowait_boost < IOWAIT_BOOST_MIN) {
272 sg_cpu->iowait_boost = 0;
273 return;
274 }
275 }
276
277 sg_cpu->iowait_boost_pending = false;
278
279 /*
280 * sg_cpu->util is already in capacity scale; convert iowait_boost
281 * into the same scale so we can compare.
282 */
283 boost = (sg_cpu->iowait_boost * sg_cpu->max) >> SCHED_CAPACITY_SHIFT;
284 boost = uclamp_rq_util_with(cpu_rq(sg_cpu->cpu), boost, NULL);
285 if (sg_cpu->util < boost)
286 sg_cpu->util = boost;
287}
288
289#ifdef CONFIG_NO_HZ_COMMON
290static bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu)
291{
292 unsigned long idle_calls = tick_nohz_get_idle_calls_cpu(sg_cpu->cpu);
293 bool ret = idle_calls == sg_cpu->saved_idle_calls;
294
295 sg_cpu->saved_idle_calls = idle_calls;
296 return ret;
297}
298#else
299static inline bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu) { return false; }
300#endif /* CONFIG_NO_HZ_COMMON */
301
302/*
303 * Make sugov_should_update_freq() ignore the rate limit when DL
304 * has increased the utilization.
305 */
306static inline void ignore_dl_rate_limit(struct sugov_cpu *sg_cpu)
307{
308 if (cpu_bw_dl(cpu_rq(sg_cpu->cpu)) > sg_cpu->bw_dl)
309 sg_cpu->sg_policy->limits_changed = true;
310}
311
312static inline bool sugov_update_single_common(struct sugov_cpu *sg_cpu,
313 u64 time, unsigned int flags)
314{
315 sugov_iowait_boost(sg_cpu, time, flags);
316 sg_cpu->last_update = time;
317
318 ignore_dl_rate_limit(sg_cpu);
319
320 if (!sugov_should_update_freq(sg_cpu->sg_policy, time))
321 return false;
322
323 sugov_get_util(sg_cpu);
324 sugov_iowait_apply(sg_cpu, time);
325
326 return true;
327}
328
329static void sugov_update_single_freq(struct update_util_data *hook, u64 time,
330 unsigned int flags)
331{
332 struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
333 struct sugov_policy *sg_policy = sg_cpu->sg_policy;
334 unsigned int cached_freq = sg_policy->cached_raw_freq;
335 unsigned int next_f;
336
337 if (!sugov_update_single_common(sg_cpu, time, flags))
338 return;
339
340 next_f = get_next_freq(sg_policy, sg_cpu->util, sg_cpu->max);
341 /*
342 * Do not reduce the frequency if the CPU has not been idle
343 * recently, as the reduction is likely to be premature then.
344 *
345 * Except when the rq is capped by uclamp_max.
346 */
347 if (!uclamp_rq_is_capped(cpu_rq(sg_cpu->cpu)) &&
348 sugov_cpu_is_busy(sg_cpu) && next_f < sg_policy->next_freq) {
349 next_f = sg_policy->next_freq;
350
351 /* Restore cached freq as next_freq has changed */
352 sg_policy->cached_raw_freq = cached_freq;
353 }
354
355 if (!sugov_update_next_freq(sg_policy, time, next_f))
356 return;
357
358 /*
359 * This code runs under rq->lock for the target CPU, so it won't run
360 * concurrently on two different CPUs for the same target and it is not
361 * necessary to acquire the lock in the fast switch case.
362 */
363 if (sg_policy->policy->fast_switch_enabled) {
364 cpufreq_driver_fast_switch(sg_policy->policy, next_f);
365 } else {
366 raw_spin_lock(&sg_policy->update_lock);
367 sugov_deferred_update(sg_policy);
368 raw_spin_unlock(&sg_policy->update_lock);
369 }
370}
371
372static void sugov_update_single_perf(struct update_util_data *hook, u64 time,
373 unsigned int flags)
374{
375 struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
376 unsigned long prev_util = sg_cpu->util;
377
378 /*
379 * Fall back to the "frequency" path if frequency invariance is not
380 * supported, because the direct mapping between the utilization and
381 * the performance levels depends on the frequency invariance.
382 */
383 if (!arch_scale_freq_invariant()) {
384 sugov_update_single_freq(hook, time, flags);
385 return;
386 }
387
388 if (!sugov_update_single_common(sg_cpu, time, flags))
389 return;
390
391 /*
392 * Do not reduce the target performance level if the CPU has not been
393 * idle recently, as the reduction is likely to be premature then.
394 *
395 * Except when the rq is capped by uclamp_max.
396 */
397 if (!uclamp_rq_is_capped(cpu_rq(sg_cpu->cpu)) &&
398 sugov_cpu_is_busy(sg_cpu) && sg_cpu->util < prev_util)
399 sg_cpu->util = prev_util;
400
401 cpufreq_driver_adjust_perf(sg_cpu->cpu, map_util_perf(sg_cpu->bw_dl),
402 map_util_perf(sg_cpu->util), sg_cpu->max);
403
404 sg_cpu->sg_policy->last_freq_update_time = time;
405}
406
407static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu, u64 time)
408{
409 struct sugov_policy *sg_policy = sg_cpu->sg_policy;
410 struct cpufreq_policy *policy = sg_policy->policy;
411 unsigned long util = 0, max = 1;
412 unsigned int j;
413
414 for_each_cpu(j, policy->cpus) {
415 struct sugov_cpu *j_sg_cpu = &per_cpu(sugov_cpu, j);
416 unsigned long j_util, j_max;
417
418 sugov_get_util(j_sg_cpu);
419 sugov_iowait_apply(j_sg_cpu, time);
420 j_util = j_sg_cpu->util;
421 j_max = j_sg_cpu->max;
422
423 if (j_util * max > j_max * util) {
424 util = j_util;
425 max = j_max;
426 }
427 }
428
429 return get_next_freq(sg_policy, util, max);
430}
431
432static void
433sugov_update_shared(struct update_util_data *hook, u64 time, unsigned int flags)
434{
435 struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
436 struct sugov_policy *sg_policy = sg_cpu->sg_policy;
437 unsigned int next_f;
438
439 raw_spin_lock(&sg_policy->update_lock);
440
441 sugov_iowait_boost(sg_cpu, time, flags);
442 sg_cpu->last_update = time;
443
444 ignore_dl_rate_limit(sg_cpu);
445
446 if (sugov_should_update_freq(sg_policy, time)) {
447 next_f = sugov_next_freq_shared(sg_cpu, time);
448
449 if (!sugov_update_next_freq(sg_policy, time, next_f))
450 goto unlock;
451
452 if (sg_policy->policy->fast_switch_enabled)
453 cpufreq_driver_fast_switch(sg_policy->policy, next_f);
454 else
455 sugov_deferred_update(sg_policy);
456 }
457unlock:
458 raw_spin_unlock(&sg_policy->update_lock);
459}
460
461static void sugov_work(struct kthread_work *work)
462{
463 struct sugov_policy *sg_policy = container_of(work, struct sugov_policy, work);
464 unsigned int freq;
465 unsigned long flags;
466
467 /*
468 * Hold sg_policy->update_lock shortly to handle the case where:
469 * in case sg_policy->next_freq is read here, and then updated by
470 * sugov_deferred_update() just before work_in_progress is set to false
471 * here, we may miss queueing the new update.
472 *
473 * Note: If a work was queued after the update_lock is released,
474 * sugov_work() will just be called again by kthread_work code; and the
475 * request will be proceed before the sugov thread sleeps.
476 */
477 raw_spin_lock_irqsave(&sg_policy->update_lock, flags);
478 freq = sg_policy->next_freq;
479 sg_policy->work_in_progress = false;
480 raw_spin_unlock_irqrestore(&sg_policy->update_lock, flags);
481
482 mutex_lock(&sg_policy->work_lock);
483 __cpufreq_driver_target(sg_policy->policy, freq, CPUFREQ_RELATION_L);
484 mutex_unlock(&sg_policy->work_lock);
485}
486
487static void sugov_irq_work(struct irq_work *irq_work)
488{
489 struct sugov_policy *sg_policy;
490
491 sg_policy = container_of(irq_work, struct sugov_policy, irq_work);
492
493 kthread_queue_work(&sg_policy->worker, &sg_policy->work);
494}
495
496/************************** sysfs interface ************************/
497
498static struct sugov_tunables *global_tunables;
499static DEFINE_MUTEX(global_tunables_lock);
500
501static inline struct sugov_tunables *to_sugov_tunables(struct gov_attr_set *attr_set)
502{
503 return container_of(attr_set, struct sugov_tunables, attr_set);
504}
505
506static ssize_t rate_limit_us_show(struct gov_attr_set *attr_set, char *buf)
507{
508 struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
509
510 return sprintf(buf, "%u\n", tunables->rate_limit_us);
511}
512
513static ssize_t
514rate_limit_us_store(struct gov_attr_set *attr_set, const char *buf, size_t count)
515{
516 struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
517 struct sugov_policy *sg_policy;
518 unsigned int rate_limit_us;
519
520 if (kstrtouint(buf, 10, &rate_limit_us))
521 return -EINVAL;
522
523 tunables->rate_limit_us = rate_limit_us;
524
525 list_for_each_entry(sg_policy, &attr_set->policy_list, tunables_hook)
526 sg_policy->freq_update_delay_ns = rate_limit_us * NSEC_PER_USEC;
527
528 return count;
529}
530
531static struct governor_attr rate_limit_us = __ATTR_RW(rate_limit_us);
532
533static struct attribute *sugov_attrs[] = {
534 &rate_limit_us.attr,
535 NULL
536};
537ATTRIBUTE_GROUPS(sugov);
538
539static void sugov_tunables_free(struct kobject *kobj)
540{
541 struct gov_attr_set *attr_set = to_gov_attr_set(kobj);
542
543 kfree(to_sugov_tunables(attr_set));
544}
545
546static struct kobj_type sugov_tunables_ktype = {
547 .default_groups = sugov_groups,
548 .sysfs_ops = &governor_sysfs_ops,
549 .release = &sugov_tunables_free,
550};
551
552/********************** cpufreq governor interface *********************/
553
554struct cpufreq_governor schedutil_gov;
555
556static struct sugov_policy *sugov_policy_alloc(struct cpufreq_policy *policy)
557{
558 struct sugov_policy *sg_policy;
559
560 sg_policy = kzalloc(sizeof(*sg_policy), GFP_KERNEL);
561 if (!sg_policy)
562 return NULL;
563
564 sg_policy->policy = policy;
565 raw_spin_lock_init(&sg_policy->update_lock);
566 return sg_policy;
567}
568
569static void sugov_policy_free(struct sugov_policy *sg_policy)
570{
571 kfree(sg_policy);
572}
573
574static int sugov_kthread_create(struct sugov_policy *sg_policy)
575{
576 struct task_struct *thread;
577 struct sched_attr attr = {
578 .size = sizeof(struct sched_attr),
579 .sched_policy = SCHED_DEADLINE,
580 .sched_flags = SCHED_FLAG_SUGOV,
581 .sched_nice = 0,
582 .sched_priority = 0,
583 /*
584 * Fake (unused) bandwidth; workaround to "fix"
585 * priority inheritance.
586 */
587 .sched_runtime = 1000000,
588 .sched_deadline = 10000000,
589 .sched_period = 10000000,
590 };
591 struct cpufreq_policy *policy = sg_policy->policy;
592 int ret;
593
594 /* kthread only required for slow path */
595 if (policy->fast_switch_enabled)
596 return 0;
597
598 kthread_init_work(&sg_policy->work, sugov_work);
599 kthread_init_worker(&sg_policy->worker);
600 thread = kthread_create(kthread_worker_fn, &sg_policy->worker,
601 "sugov:%d",
602 cpumask_first(policy->related_cpus));
603 if (IS_ERR(thread)) {
604 pr_err("failed to create sugov thread: %ld\n", PTR_ERR(thread));
605 return PTR_ERR(thread);
606 }
607
608 ret = sched_setattr_nocheck(thread, &attr);
609 if (ret) {
610 kthread_stop(thread);
611 pr_warn("%s: failed to set SCHED_DEADLINE\n", __func__);
612 return ret;
613 }
614
615 sg_policy->thread = thread;
616 kthread_bind_mask(thread, policy->related_cpus);
617 init_irq_work(&sg_policy->irq_work, sugov_irq_work);
618 mutex_init(&sg_policy->work_lock);
619
620 wake_up_process(thread);
621
622 return 0;
623}
624
625static void sugov_kthread_stop(struct sugov_policy *sg_policy)
626{
627 /* kthread only required for slow path */
628 if (sg_policy->policy->fast_switch_enabled)
629 return;
630
631 kthread_flush_worker(&sg_policy->worker);
632 kthread_stop(sg_policy->thread);
633 mutex_destroy(&sg_policy->work_lock);
634}
635
636static struct sugov_tunables *sugov_tunables_alloc(struct sugov_policy *sg_policy)
637{
638 struct sugov_tunables *tunables;
639
640 tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
641 if (tunables) {
642 gov_attr_set_init(&tunables->attr_set, &sg_policy->tunables_hook);
643 if (!have_governor_per_policy())
644 global_tunables = tunables;
645 }
646 return tunables;
647}
648
649static void sugov_clear_global_tunables(void)
650{
651 if (!have_governor_per_policy())
652 global_tunables = NULL;
653}
654
655static int sugov_init(struct cpufreq_policy *policy)
656{
657 struct sugov_policy *sg_policy;
658 struct sugov_tunables *tunables;
659 int ret = 0;
660
661 /* State should be equivalent to EXIT */
662 if (policy->governor_data)
663 return -EBUSY;
664
665 cpufreq_enable_fast_switch(policy);
666
667 sg_policy = sugov_policy_alloc(policy);
668 if (!sg_policy) {
669 ret = -ENOMEM;
670 goto disable_fast_switch;
671 }
672
673 ret = sugov_kthread_create(sg_policy);
674 if (ret)
675 goto free_sg_policy;
676
677 mutex_lock(&global_tunables_lock);
678
679 if (global_tunables) {
680 if (WARN_ON(have_governor_per_policy())) {
681 ret = -EINVAL;
682 goto stop_kthread;
683 }
684 policy->governor_data = sg_policy;
685 sg_policy->tunables = global_tunables;
686
687 gov_attr_set_get(&global_tunables->attr_set, &sg_policy->tunables_hook);
688 goto out;
689 }
690
691 tunables = sugov_tunables_alloc(sg_policy);
692 if (!tunables) {
693 ret = -ENOMEM;
694 goto stop_kthread;
695 }
696
697 tunables->rate_limit_us = cpufreq_policy_transition_delay_us(policy);
698
699 policy->governor_data = sg_policy;
700 sg_policy->tunables = tunables;
701
702 ret = kobject_init_and_add(&tunables->attr_set.kobj, &sugov_tunables_ktype,
703 get_governor_parent_kobj(policy), "%s",
704 schedutil_gov.name);
705 if (ret)
706 goto fail;
707
708out:
709 mutex_unlock(&global_tunables_lock);
710 return 0;
711
712fail:
713 kobject_put(&tunables->attr_set.kobj);
714 policy->governor_data = NULL;
715 sugov_clear_global_tunables();
716
717stop_kthread:
718 sugov_kthread_stop(sg_policy);
719 mutex_unlock(&global_tunables_lock);
720
721free_sg_policy:
722 sugov_policy_free(sg_policy);
723
724disable_fast_switch:
725 cpufreq_disable_fast_switch(policy);
726
727 pr_err("initialization failed (error %d)\n", ret);
728 return ret;
729}
730
731static void sugov_exit(struct cpufreq_policy *policy)
732{
733 struct sugov_policy *sg_policy = policy->governor_data;
734 struct sugov_tunables *tunables = sg_policy->tunables;
735 unsigned int count;
736
737 mutex_lock(&global_tunables_lock);
738
739 count = gov_attr_set_put(&tunables->attr_set, &sg_policy->tunables_hook);
740 policy->governor_data = NULL;
741 if (!count)
742 sugov_clear_global_tunables();
743
744 mutex_unlock(&global_tunables_lock);
745
746 sugov_kthread_stop(sg_policy);
747 sugov_policy_free(sg_policy);
748 cpufreq_disable_fast_switch(policy);
749}
750
751static int sugov_start(struct cpufreq_policy *policy)
752{
753 struct sugov_policy *sg_policy = policy->governor_data;
754 void (*uu)(struct update_util_data *data, u64 time, unsigned int flags);
755 unsigned int cpu;
756
757 sg_policy->freq_update_delay_ns = sg_policy->tunables->rate_limit_us * NSEC_PER_USEC;
758 sg_policy->last_freq_update_time = 0;
759 sg_policy->next_freq = 0;
760 sg_policy->work_in_progress = false;
761 sg_policy->limits_changed = false;
762 sg_policy->cached_raw_freq = 0;
763
764 sg_policy->need_freq_update = cpufreq_driver_test_flags(CPUFREQ_NEED_UPDATE_LIMITS);
765
766 for_each_cpu(cpu, policy->cpus) {
767 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
768
769 memset(sg_cpu, 0, sizeof(*sg_cpu));
770 sg_cpu->cpu = cpu;
771 sg_cpu->sg_policy = sg_policy;
772 }
773
774 if (policy_is_shared(policy))
775 uu = sugov_update_shared;
776 else if (policy->fast_switch_enabled && cpufreq_driver_has_adjust_perf())
777 uu = sugov_update_single_perf;
778 else
779 uu = sugov_update_single_freq;
780
781 for_each_cpu(cpu, policy->cpus) {
782 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
783
784 cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util, uu);
785 }
786 return 0;
787}
788
789static void sugov_stop(struct cpufreq_policy *policy)
790{
791 struct sugov_policy *sg_policy = policy->governor_data;
792 unsigned int cpu;
793
794 for_each_cpu(cpu, policy->cpus)
795 cpufreq_remove_update_util_hook(cpu);
796
797 synchronize_rcu();
798
799 if (!policy->fast_switch_enabled) {
800 irq_work_sync(&sg_policy->irq_work);
801 kthread_cancel_work_sync(&sg_policy->work);
802 }
803}
804
805static void sugov_limits(struct cpufreq_policy *policy)
806{
807 struct sugov_policy *sg_policy = policy->governor_data;
808
809 if (!policy->fast_switch_enabled) {
810 mutex_lock(&sg_policy->work_lock);
811 cpufreq_policy_apply_limits(policy);
812 mutex_unlock(&sg_policy->work_lock);
813 }
814
815 sg_policy->limits_changed = true;
816}
817
818struct cpufreq_governor schedutil_gov = {
819 .name = "schedutil",
820 .owner = THIS_MODULE,
821 .flags = CPUFREQ_GOV_DYNAMIC_SWITCHING,
822 .init = sugov_init,
823 .exit = sugov_exit,
824 .start = sugov_start,
825 .stop = sugov_stop,
826 .limits = sugov_limits,
827};
828
829#ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL
830struct cpufreq_governor *cpufreq_default_governor(void)
831{
832 return &schedutil_gov;
833}
834#endif
835
836cpufreq_governor_init(schedutil_gov);
837
838#ifdef CONFIG_ENERGY_MODEL
839static void rebuild_sd_workfn(struct work_struct *work)
840{
841 rebuild_sched_domains_energy();
842}
843static DECLARE_WORK(rebuild_sd_work, rebuild_sd_workfn);
844
845/*
846 * EAS shouldn't be attempted without sugov, so rebuild the sched_domains
847 * on governor changes to make sure the scheduler knows about it.
848 */
849void sched_cpufreq_governor_change(struct cpufreq_policy *policy,
850 struct cpufreq_governor *old_gov)
851{
852 if (old_gov == &schedutil_gov || policy->governor == &schedutil_gov) {
853 /*
854 * When called from the cpufreq_register_driver() path, the
855 * cpu_hotplug_lock is already held, so use a work item to
856 * avoid nested locking in rebuild_sched_domains().
857 */
858 schedule_work(&rebuild_sd_work);
859 }
860
861}
862#endif