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