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 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_min;
51
52 /* The field below is for single-CPU policies only: */
53#ifdef CONFIG_NO_HZ_COMMON
54 unsigned long saved_idle_calls;
55#endif
56};
57
58static DEFINE_PER_CPU(struct sugov_cpu, sugov_cpu);
59
60/************************ Governor internals ***********************/
61
62static bool sugov_should_update_freq(struct sugov_policy *sg_policy, u64 time)
63{
64 s64 delta_ns;
65
66 /*
67 * Since cpufreq_update_util() is called with rq->lock held for
68 * the @target_cpu, our per-CPU data is fully serialized.
69 *
70 * However, drivers cannot in general deal with cross-CPU
71 * requests, so while get_next_freq() will work, our
72 * sugov_update_commit() call may not for the fast switching platforms.
73 *
74 * Hence stop here for remote requests if they aren't supported
75 * by the hardware, as calculating the frequency is pointless if
76 * we cannot in fact act on it.
77 *
78 * This is needed on the slow switching platforms too to prevent CPUs
79 * going offline from leaving stale IRQ work items behind.
80 */
81 if (!cpufreq_this_cpu_can_update(sg_policy->policy))
82 return false;
83
84 if (unlikely(sg_policy->limits_changed)) {
85 sg_policy->limits_changed = false;
86 sg_policy->need_freq_update = true;
87 return true;
88 }
89
90 delta_ns = time - sg_policy->last_freq_update_time;
91
92 return delta_ns >= sg_policy->freq_update_delay_ns;
93}
94
95static bool sugov_update_next_freq(struct sugov_policy *sg_policy, u64 time,
96 unsigned int next_freq)
97{
98 if (sg_policy->need_freq_update)
99 sg_policy->need_freq_update = cpufreq_driver_test_flags(CPUFREQ_NEED_UPDATE_LIMITS);
100 else if (sg_policy->next_freq == next_freq)
101 return false;
102
103 sg_policy->next_freq = next_freq;
104 sg_policy->last_freq_update_time = time;
105
106 return true;
107}
108
109static void sugov_deferred_update(struct sugov_policy *sg_policy)
110{
111 if (!sg_policy->work_in_progress) {
112 sg_policy->work_in_progress = true;
113 irq_work_queue(&sg_policy->irq_work);
114 }
115}
116
117/**
118 * get_capacity_ref_freq - get the reference frequency that has been used to
119 * correlate frequency and compute capacity for a given cpufreq policy. We use
120 * the CPU managing it for the arch_scale_freq_ref() call in the function.
121 * @policy: the cpufreq policy of the CPU in question.
122 *
123 * Return: the reference CPU frequency to compute a capacity.
124 */
125static __always_inline
126unsigned long get_capacity_ref_freq(struct cpufreq_policy *policy)
127{
128 unsigned int freq = arch_scale_freq_ref(policy->cpu);
129
130 if (freq)
131 return freq;
132
133 if (arch_scale_freq_invariant())
134 return policy->cpuinfo.max_freq;
135
136 /*
137 * Apply a 25% margin so that we select a higher frequency than
138 * the current one before the CPU is fully busy:
139 */
140 return policy->cur + (policy->cur >> 2);
141}
142
143/**
144 * get_next_freq - Compute a new frequency for a given cpufreq policy.
145 * @sg_policy: schedutil policy object to compute the new frequency for.
146 * @util: Current CPU utilization.
147 * @max: CPU capacity.
148 *
149 * If the utilization is frequency-invariant, choose the new frequency to be
150 * proportional to it, that is
151 *
152 * next_freq = C * max_freq * util / max
153 *
154 * Otherwise, approximate the would-be frequency-invariant utilization by
155 * util_raw * (curr_freq / max_freq) which leads to
156 *
157 * next_freq = C * curr_freq * util_raw / max
158 *
159 * Take C = 1.25 for the frequency tipping point at (util / max) = 0.8.
160 *
161 * The lowest driver-supported frequency which is equal or greater than the raw
162 * next_freq (as calculated above) is returned, subject to policy min/max and
163 * cpufreq driver limitations.
164 */
165static unsigned int get_next_freq(struct sugov_policy *sg_policy,
166 unsigned long util, unsigned long max)
167{
168 struct cpufreq_policy *policy = sg_policy->policy;
169 unsigned int freq;
170
171 freq = get_capacity_ref_freq(policy);
172 freq = map_util_freq(util, freq, max);
173
174 if (freq == sg_policy->cached_raw_freq && !sg_policy->need_freq_update)
175 return sg_policy->next_freq;
176
177 sg_policy->cached_raw_freq = freq;
178 return cpufreq_driver_resolve_freq(policy, freq);
179}
180
181unsigned long sugov_effective_cpu_perf(int cpu, unsigned long actual,
182 unsigned long min,
183 unsigned long max)
184{
185 /* Add dvfs headroom to actual utilization */
186 actual = map_util_perf(actual);
187 /* Actually we don't need to target the max performance */
188 if (actual < max)
189 max = actual;
190
191 /*
192 * Ensure at least minimum performance while providing more compute
193 * capacity when possible.
194 */
195 return max(min, max);
196}
197
198static void sugov_get_util(struct sugov_cpu *sg_cpu, unsigned long boost)
199{
200 unsigned long min, max, util = cpu_util_cfs_boost(sg_cpu->cpu);
201
202 util = effective_cpu_util(sg_cpu->cpu, util, &min, &max);
203 util = max(util, boost);
204 sg_cpu->bw_min = min;
205 sg_cpu->util = sugov_effective_cpu_perf(sg_cpu->cpu, util, min, max);
206}
207
208/**
209 * sugov_iowait_reset() - Reset the IO boost status of a CPU.
210 * @sg_cpu: the sugov data for the CPU to boost
211 * @time: the update time from the caller
212 * @set_iowait_boost: true if an IO boost has been requested
213 *
214 * The IO wait boost of a task is disabled after a tick since the last update
215 * of a CPU. If a new IO wait boost is requested after more then a tick, then
216 * we enable the boost starting from IOWAIT_BOOST_MIN, which improves energy
217 * efficiency by ignoring sporadic wakeups from IO.
218 */
219static bool sugov_iowait_reset(struct sugov_cpu *sg_cpu, u64 time,
220 bool set_iowait_boost)
221{
222 s64 delta_ns = time - sg_cpu->last_update;
223
224 /* Reset boost only if a tick has elapsed since last request */
225 if (delta_ns <= TICK_NSEC)
226 return false;
227
228 sg_cpu->iowait_boost = set_iowait_boost ? IOWAIT_BOOST_MIN : 0;
229 sg_cpu->iowait_boost_pending = set_iowait_boost;
230
231 return true;
232}
233
234/**
235 * sugov_iowait_boost() - Updates the IO boost status of a CPU.
236 * @sg_cpu: the sugov data for the CPU to boost
237 * @time: the update time from the caller
238 * @flags: SCHED_CPUFREQ_IOWAIT if the task is waking up after an IO wait
239 *
240 * Each time a task wakes up after an IO operation, the CPU utilization can be
241 * boosted to a certain utilization which doubles at each "frequent and
242 * successive" wakeup from IO, ranging from IOWAIT_BOOST_MIN to the utilization
243 * of the maximum OPP.
244 *
245 * To keep doubling, an IO boost has to be requested at least once per tick,
246 * otherwise we restart from the utilization of the minimum OPP.
247 */
248static void sugov_iowait_boost(struct sugov_cpu *sg_cpu, u64 time,
249 unsigned int flags)
250{
251 bool set_iowait_boost = flags & SCHED_CPUFREQ_IOWAIT;
252
253 /* Reset boost if the CPU appears to have been idle enough */
254 if (sg_cpu->iowait_boost &&
255 sugov_iowait_reset(sg_cpu, time, set_iowait_boost))
256 return;
257
258 /* Boost only tasks waking up after IO */
259 if (!set_iowait_boost)
260 return;
261
262 /* Ensure boost doubles only one time at each request */
263 if (sg_cpu->iowait_boost_pending)
264 return;
265 sg_cpu->iowait_boost_pending = true;
266
267 /* Double the boost at each request */
268 if (sg_cpu->iowait_boost) {
269 sg_cpu->iowait_boost =
270 min_t(unsigned int, sg_cpu->iowait_boost << 1, SCHED_CAPACITY_SCALE);
271 return;
272 }
273
274 /* First wakeup after IO: start with minimum boost */
275 sg_cpu->iowait_boost = IOWAIT_BOOST_MIN;
276}
277
278/**
279 * sugov_iowait_apply() - Apply the IO boost to a CPU.
280 * @sg_cpu: the sugov data for the cpu to boost
281 * @time: the update time from the caller
282 * @max_cap: the max CPU capacity
283 *
284 * A CPU running a task which woken up after an IO operation can have its
285 * utilization boosted to speed up the completion of those IO operations.
286 * The IO boost value is increased each time a task wakes up from IO, in
287 * sugov_iowait_apply(), and it's instead decreased by this function,
288 * each time an increase has not been requested (!iowait_boost_pending).
289 *
290 * A CPU which also appears to have been idle for at least one tick has also
291 * its IO boost utilization reset.
292 *
293 * This mechanism is designed to boost high frequently IO waiting tasks, while
294 * being more conservative on tasks which does sporadic IO operations.
295 */
296static unsigned long sugov_iowait_apply(struct sugov_cpu *sg_cpu, u64 time,
297 unsigned long max_cap)
298{
299 /* No boost currently required */
300 if (!sg_cpu->iowait_boost)
301 return 0;
302
303 /* Reset boost if the CPU appears to have been idle enough */
304 if (sugov_iowait_reset(sg_cpu, time, false))
305 return 0;
306
307 if (!sg_cpu->iowait_boost_pending) {
308 /*
309 * No boost pending; reduce the boost value.
310 */
311 sg_cpu->iowait_boost >>= 1;
312 if (sg_cpu->iowait_boost < IOWAIT_BOOST_MIN) {
313 sg_cpu->iowait_boost = 0;
314 return 0;
315 }
316 }
317
318 sg_cpu->iowait_boost_pending = false;
319
320 /*
321 * sg_cpu->util is already in capacity scale; convert iowait_boost
322 * into the same scale so we can compare.
323 */
324 return (sg_cpu->iowait_boost * max_cap) >> SCHED_CAPACITY_SHIFT;
325}
326
327#ifdef CONFIG_NO_HZ_COMMON
328static bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu)
329{
330 unsigned long idle_calls = tick_nohz_get_idle_calls_cpu(sg_cpu->cpu);
331 bool ret = idle_calls == sg_cpu->saved_idle_calls;
332
333 sg_cpu->saved_idle_calls = idle_calls;
334 return ret;
335}
336#else
337static inline bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu) { return false; }
338#endif /* CONFIG_NO_HZ_COMMON */
339
340/*
341 * Make sugov_should_update_freq() ignore the rate limit when DL
342 * has increased the utilization.
343 */
344static inline void ignore_dl_rate_limit(struct sugov_cpu *sg_cpu)
345{
346 if (cpu_bw_dl(cpu_rq(sg_cpu->cpu)) > sg_cpu->bw_min)
347 sg_cpu->sg_policy->limits_changed = true;
348}
349
350static inline bool sugov_update_single_common(struct sugov_cpu *sg_cpu,
351 u64 time, unsigned long max_cap,
352 unsigned int flags)
353{
354 unsigned long boost;
355
356 sugov_iowait_boost(sg_cpu, time, flags);
357 sg_cpu->last_update = time;
358
359 ignore_dl_rate_limit(sg_cpu);
360
361 if (!sugov_should_update_freq(sg_cpu->sg_policy, time))
362 return false;
363
364 boost = sugov_iowait_apply(sg_cpu, time, max_cap);
365 sugov_get_util(sg_cpu, boost);
366
367 return true;
368}
369
370static void sugov_update_single_freq(struct update_util_data *hook, u64 time,
371 unsigned int flags)
372{
373 struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
374 struct sugov_policy *sg_policy = sg_cpu->sg_policy;
375 unsigned int cached_freq = sg_policy->cached_raw_freq;
376 unsigned long max_cap;
377 unsigned int next_f;
378
379 max_cap = arch_scale_cpu_capacity(sg_cpu->cpu);
380
381 if (!sugov_update_single_common(sg_cpu, time, max_cap, flags))
382 return;
383
384 next_f = get_next_freq(sg_policy, sg_cpu->util, max_cap);
385 /*
386 * Do not reduce the frequency if the CPU has not been idle
387 * recently, as the reduction is likely to be premature then.
388 *
389 * Except when the rq is capped by uclamp_max.
390 */
391 if (!uclamp_rq_is_capped(cpu_rq(sg_cpu->cpu)) &&
392 sugov_cpu_is_busy(sg_cpu) && next_f < sg_policy->next_freq &&
393 !sg_policy->need_freq_update) {
394 next_f = sg_policy->next_freq;
395
396 /* Restore cached freq as next_freq has changed */
397 sg_policy->cached_raw_freq = cached_freq;
398 }
399
400 if (!sugov_update_next_freq(sg_policy, time, next_f))
401 return;
402
403 /*
404 * This code runs under rq->lock for the target CPU, so it won't run
405 * concurrently on two different CPUs for the same target and it is not
406 * necessary to acquire the lock in the fast switch case.
407 */
408 if (sg_policy->policy->fast_switch_enabled) {
409 cpufreq_driver_fast_switch(sg_policy->policy, next_f);
410 } else {
411 raw_spin_lock(&sg_policy->update_lock);
412 sugov_deferred_update(sg_policy);
413 raw_spin_unlock(&sg_policy->update_lock);
414 }
415}
416
417static void sugov_update_single_perf(struct update_util_data *hook, u64 time,
418 unsigned int flags)
419{
420 struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
421 unsigned long prev_util = sg_cpu->util;
422 unsigned long max_cap;
423
424 /*
425 * Fall back to the "frequency" path if frequency invariance is not
426 * supported, because the direct mapping between the utilization and
427 * the performance levels depends on the frequency invariance.
428 */
429 if (!arch_scale_freq_invariant()) {
430 sugov_update_single_freq(hook, time, flags);
431 return;
432 }
433
434 max_cap = arch_scale_cpu_capacity(sg_cpu->cpu);
435
436 if (!sugov_update_single_common(sg_cpu, time, max_cap, flags))
437 return;
438
439 /*
440 * Do not reduce the target performance level if the CPU has not been
441 * idle recently, as the reduction is likely to be premature then.
442 *
443 * Except when the rq is capped by uclamp_max.
444 */
445 if (!uclamp_rq_is_capped(cpu_rq(sg_cpu->cpu)) &&
446 sugov_cpu_is_busy(sg_cpu) && sg_cpu->util < prev_util)
447 sg_cpu->util = prev_util;
448
449 cpufreq_driver_adjust_perf(sg_cpu->cpu, sg_cpu->bw_min,
450 sg_cpu->util, max_cap);
451
452 sg_cpu->sg_policy->last_freq_update_time = time;
453}
454
455static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu, u64 time)
456{
457 struct sugov_policy *sg_policy = sg_cpu->sg_policy;
458 struct cpufreq_policy *policy = sg_policy->policy;
459 unsigned long util = 0, max_cap;
460 unsigned int j;
461
462 max_cap = arch_scale_cpu_capacity(sg_cpu->cpu);
463
464 for_each_cpu(j, policy->cpus) {
465 struct sugov_cpu *j_sg_cpu = &per_cpu(sugov_cpu, j);
466 unsigned long boost;
467
468 boost = sugov_iowait_apply(j_sg_cpu, time, max_cap);
469 sugov_get_util(j_sg_cpu, boost);
470
471 util = max(j_sg_cpu->util, util);
472 }
473
474 return get_next_freq(sg_policy, util, max_cap);
475}
476
477static void
478sugov_update_shared(struct update_util_data *hook, u64 time, unsigned int flags)
479{
480 struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
481 struct sugov_policy *sg_policy = sg_cpu->sg_policy;
482 unsigned int next_f;
483
484 raw_spin_lock(&sg_policy->update_lock);
485
486 sugov_iowait_boost(sg_cpu, time, flags);
487 sg_cpu->last_update = time;
488
489 ignore_dl_rate_limit(sg_cpu);
490
491 if (sugov_should_update_freq(sg_policy, time)) {
492 next_f = sugov_next_freq_shared(sg_cpu, time);
493
494 if (!sugov_update_next_freq(sg_policy, time, next_f))
495 goto unlock;
496
497 if (sg_policy->policy->fast_switch_enabled)
498 cpufreq_driver_fast_switch(sg_policy->policy, next_f);
499 else
500 sugov_deferred_update(sg_policy);
501 }
502unlock:
503 raw_spin_unlock(&sg_policy->update_lock);
504}
505
506static void sugov_work(struct kthread_work *work)
507{
508 struct sugov_policy *sg_policy = container_of(work, struct sugov_policy, work);
509 unsigned int freq;
510 unsigned long flags;
511
512 /*
513 * Hold sg_policy->update_lock shortly to handle the case where:
514 * in case sg_policy->next_freq is read here, and then updated by
515 * sugov_deferred_update() just before work_in_progress is set to false
516 * here, we may miss queueing the new update.
517 *
518 * Note: If a work was queued after the update_lock is released,
519 * sugov_work() will just be called again by kthread_work code; and the
520 * request will be proceed before the sugov thread sleeps.
521 */
522 raw_spin_lock_irqsave(&sg_policy->update_lock, flags);
523 freq = sg_policy->next_freq;
524 sg_policy->work_in_progress = false;
525 raw_spin_unlock_irqrestore(&sg_policy->update_lock, flags);
526
527 mutex_lock(&sg_policy->work_lock);
528 __cpufreq_driver_target(sg_policy->policy, freq, CPUFREQ_RELATION_L);
529 mutex_unlock(&sg_policy->work_lock);
530}
531
532static void sugov_irq_work(struct irq_work *irq_work)
533{
534 struct sugov_policy *sg_policy;
535
536 sg_policy = container_of(irq_work, struct sugov_policy, irq_work);
537
538 kthread_queue_work(&sg_policy->worker, &sg_policy->work);
539}
540
541/************************** sysfs interface ************************/
542
543static struct sugov_tunables *global_tunables;
544static DEFINE_MUTEX(global_tunables_lock);
545
546static inline struct sugov_tunables *to_sugov_tunables(struct gov_attr_set *attr_set)
547{
548 return container_of(attr_set, struct sugov_tunables, attr_set);
549}
550
551static ssize_t rate_limit_us_show(struct gov_attr_set *attr_set, char *buf)
552{
553 struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
554
555 return sprintf(buf, "%u\n", tunables->rate_limit_us);
556}
557
558static ssize_t
559rate_limit_us_store(struct gov_attr_set *attr_set, const char *buf, size_t count)
560{
561 struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
562 struct sugov_policy *sg_policy;
563 unsigned int rate_limit_us;
564
565 if (kstrtouint(buf, 10, &rate_limit_us))
566 return -EINVAL;
567
568 tunables->rate_limit_us = rate_limit_us;
569
570 list_for_each_entry(sg_policy, &attr_set->policy_list, tunables_hook)
571 sg_policy->freq_update_delay_ns = rate_limit_us * NSEC_PER_USEC;
572
573 return count;
574}
575
576static struct governor_attr rate_limit_us = __ATTR_RW(rate_limit_us);
577
578static struct attribute *sugov_attrs[] = {
579 &rate_limit_us.attr,
580 NULL
581};
582ATTRIBUTE_GROUPS(sugov);
583
584static void sugov_tunables_free(struct kobject *kobj)
585{
586 struct gov_attr_set *attr_set = to_gov_attr_set(kobj);
587
588 kfree(to_sugov_tunables(attr_set));
589}
590
591static const struct kobj_type sugov_tunables_ktype = {
592 .default_groups = sugov_groups,
593 .sysfs_ops = &governor_sysfs_ops,
594 .release = &sugov_tunables_free,
595};
596
597/********************** cpufreq governor interface *********************/
598
599#ifdef CONFIG_ENERGY_MODEL
600static void rebuild_sd_workfn(struct work_struct *work)
601{
602 rebuild_sched_domains_energy();
603}
604
605static DECLARE_WORK(rebuild_sd_work, rebuild_sd_workfn);
606
607/*
608 * EAS shouldn't be attempted without sugov, so rebuild the sched_domains
609 * on governor changes to make sure the scheduler knows about it.
610 */
611static void sugov_eas_rebuild_sd(void)
612{
613 /*
614 * When called from the cpufreq_register_driver() path, the
615 * cpu_hotplug_lock is already held, so use a work item to
616 * avoid nested locking in rebuild_sched_domains().
617 */
618 schedule_work(&rebuild_sd_work);
619}
620#else
621static inline void sugov_eas_rebuild_sd(void) { };
622#endif
623
624struct cpufreq_governor schedutil_gov;
625
626static struct sugov_policy *sugov_policy_alloc(struct cpufreq_policy *policy)
627{
628 struct sugov_policy *sg_policy;
629
630 sg_policy = kzalloc(sizeof(*sg_policy), GFP_KERNEL);
631 if (!sg_policy)
632 return NULL;
633
634 sg_policy->policy = policy;
635 raw_spin_lock_init(&sg_policy->update_lock);
636 return sg_policy;
637}
638
639static void sugov_policy_free(struct sugov_policy *sg_policy)
640{
641 kfree(sg_policy);
642}
643
644static int sugov_kthread_create(struct sugov_policy *sg_policy)
645{
646 struct task_struct *thread;
647 struct sched_attr attr = {
648 .size = sizeof(struct sched_attr),
649 .sched_policy = SCHED_DEADLINE,
650 .sched_flags = SCHED_FLAG_SUGOV,
651 .sched_nice = 0,
652 .sched_priority = 0,
653 /*
654 * Fake (unused) bandwidth; workaround to "fix"
655 * priority inheritance.
656 */
657 .sched_runtime = 1000000,
658 .sched_deadline = 10000000,
659 .sched_period = 10000000,
660 };
661 struct cpufreq_policy *policy = sg_policy->policy;
662 int ret;
663
664 /* kthread only required for slow path */
665 if (policy->fast_switch_enabled)
666 return 0;
667
668 kthread_init_work(&sg_policy->work, sugov_work);
669 kthread_init_worker(&sg_policy->worker);
670 thread = kthread_create(kthread_worker_fn, &sg_policy->worker,
671 "sugov:%d",
672 cpumask_first(policy->related_cpus));
673 if (IS_ERR(thread)) {
674 pr_err("failed to create sugov thread: %ld\n", PTR_ERR(thread));
675 return PTR_ERR(thread);
676 }
677
678 ret = sched_setattr_nocheck(thread, &attr);
679 if (ret) {
680 kthread_stop(thread);
681 pr_warn("%s: failed to set SCHED_DEADLINE\n", __func__);
682 return ret;
683 }
684
685 sg_policy->thread = thread;
686 kthread_bind_mask(thread, policy->related_cpus);
687 init_irq_work(&sg_policy->irq_work, sugov_irq_work);
688 mutex_init(&sg_policy->work_lock);
689
690 wake_up_process(thread);
691
692 return 0;
693}
694
695static void sugov_kthread_stop(struct sugov_policy *sg_policy)
696{
697 /* kthread only required for slow path */
698 if (sg_policy->policy->fast_switch_enabled)
699 return;
700
701 kthread_flush_worker(&sg_policy->worker);
702 kthread_stop(sg_policy->thread);
703 mutex_destroy(&sg_policy->work_lock);
704}
705
706static struct sugov_tunables *sugov_tunables_alloc(struct sugov_policy *sg_policy)
707{
708 struct sugov_tunables *tunables;
709
710 tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
711 if (tunables) {
712 gov_attr_set_init(&tunables->attr_set, &sg_policy->tunables_hook);
713 if (!have_governor_per_policy())
714 global_tunables = tunables;
715 }
716 return tunables;
717}
718
719static void sugov_clear_global_tunables(void)
720{
721 if (!have_governor_per_policy())
722 global_tunables = NULL;
723}
724
725static int sugov_init(struct cpufreq_policy *policy)
726{
727 struct sugov_policy *sg_policy;
728 struct sugov_tunables *tunables;
729 int ret = 0;
730
731 /* State should be equivalent to EXIT */
732 if (policy->governor_data)
733 return -EBUSY;
734
735 cpufreq_enable_fast_switch(policy);
736
737 sg_policy = sugov_policy_alloc(policy);
738 if (!sg_policy) {
739 ret = -ENOMEM;
740 goto disable_fast_switch;
741 }
742
743 ret = sugov_kthread_create(sg_policy);
744 if (ret)
745 goto free_sg_policy;
746
747 mutex_lock(&global_tunables_lock);
748
749 if (global_tunables) {
750 if (WARN_ON(have_governor_per_policy())) {
751 ret = -EINVAL;
752 goto stop_kthread;
753 }
754 policy->governor_data = sg_policy;
755 sg_policy->tunables = global_tunables;
756
757 gov_attr_set_get(&global_tunables->attr_set, &sg_policy->tunables_hook);
758 goto out;
759 }
760
761 tunables = sugov_tunables_alloc(sg_policy);
762 if (!tunables) {
763 ret = -ENOMEM;
764 goto stop_kthread;
765 }
766
767 tunables->rate_limit_us = cpufreq_policy_transition_delay_us(policy);
768
769 policy->governor_data = sg_policy;
770 sg_policy->tunables = tunables;
771
772 ret = kobject_init_and_add(&tunables->attr_set.kobj, &sugov_tunables_ktype,
773 get_governor_parent_kobj(policy), "%s",
774 schedutil_gov.name);
775 if (ret)
776 goto fail;
777
778 sugov_eas_rebuild_sd();
779
780out:
781 mutex_unlock(&global_tunables_lock);
782 return 0;
783
784fail:
785 kobject_put(&tunables->attr_set.kobj);
786 policy->governor_data = NULL;
787 sugov_clear_global_tunables();
788
789stop_kthread:
790 sugov_kthread_stop(sg_policy);
791 mutex_unlock(&global_tunables_lock);
792
793free_sg_policy:
794 sugov_policy_free(sg_policy);
795
796disable_fast_switch:
797 cpufreq_disable_fast_switch(policy);
798
799 pr_err("initialization failed (error %d)\n", ret);
800 return ret;
801}
802
803static void sugov_exit(struct cpufreq_policy *policy)
804{
805 struct sugov_policy *sg_policy = policy->governor_data;
806 struct sugov_tunables *tunables = sg_policy->tunables;
807 unsigned int count;
808
809 mutex_lock(&global_tunables_lock);
810
811 count = gov_attr_set_put(&tunables->attr_set, &sg_policy->tunables_hook);
812 policy->governor_data = NULL;
813 if (!count)
814 sugov_clear_global_tunables();
815
816 mutex_unlock(&global_tunables_lock);
817
818 sugov_kthread_stop(sg_policy);
819 sugov_policy_free(sg_policy);
820 cpufreq_disable_fast_switch(policy);
821
822 sugov_eas_rebuild_sd();
823}
824
825static int sugov_start(struct cpufreq_policy *policy)
826{
827 struct sugov_policy *sg_policy = policy->governor_data;
828 void (*uu)(struct update_util_data *data, u64 time, unsigned int flags);
829 unsigned int cpu;
830
831 sg_policy->freq_update_delay_ns = sg_policy->tunables->rate_limit_us * NSEC_PER_USEC;
832 sg_policy->last_freq_update_time = 0;
833 sg_policy->next_freq = 0;
834 sg_policy->work_in_progress = false;
835 sg_policy->limits_changed = false;
836 sg_policy->cached_raw_freq = 0;
837
838 sg_policy->need_freq_update = cpufreq_driver_test_flags(CPUFREQ_NEED_UPDATE_LIMITS);
839
840 if (policy_is_shared(policy))
841 uu = sugov_update_shared;
842 else if (policy->fast_switch_enabled && cpufreq_driver_has_adjust_perf())
843 uu = sugov_update_single_perf;
844 else
845 uu = sugov_update_single_freq;
846
847 for_each_cpu(cpu, policy->cpus) {
848 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
849
850 memset(sg_cpu, 0, sizeof(*sg_cpu));
851 sg_cpu->cpu = cpu;
852 sg_cpu->sg_policy = sg_policy;
853 cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util, uu);
854 }
855 return 0;
856}
857
858static void sugov_stop(struct cpufreq_policy *policy)
859{
860 struct sugov_policy *sg_policy = policy->governor_data;
861 unsigned int cpu;
862
863 for_each_cpu(cpu, policy->cpus)
864 cpufreq_remove_update_util_hook(cpu);
865
866 synchronize_rcu();
867
868 if (!policy->fast_switch_enabled) {
869 irq_work_sync(&sg_policy->irq_work);
870 kthread_cancel_work_sync(&sg_policy->work);
871 }
872}
873
874static void sugov_limits(struct cpufreq_policy *policy)
875{
876 struct sugov_policy *sg_policy = policy->governor_data;
877
878 if (!policy->fast_switch_enabled) {
879 mutex_lock(&sg_policy->work_lock);
880 cpufreq_policy_apply_limits(policy);
881 mutex_unlock(&sg_policy->work_lock);
882 }
883
884 sg_policy->limits_changed = true;
885}
886
887struct cpufreq_governor schedutil_gov = {
888 .name = "schedutil",
889 .owner = THIS_MODULE,
890 .flags = CPUFREQ_GOV_DYNAMIC_SWITCHING,
891 .init = sugov_init,
892 .exit = sugov_exit,
893 .start = sugov_start,
894 .stop = sugov_stop,
895 .limits = sugov_limits,
896};
897
898#ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL
899struct cpufreq_governor *cpufreq_default_governor(void)
900{
901 return &schedutil_gov;
902}
903#endif
904
905cpufreq_governor_init(schedutil_gov);