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; /* For shared policies */
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 bw_dl;
57 unsigned long max;
58
59 /* The field below is for single-CPU policies only: */
60#ifdef CONFIG_NO_HZ_COMMON
61 unsigned long saved_idle_calls;
62#endif
63};
64
65static DEFINE_PER_CPU(struct sugov_cpu, sugov_cpu);
66
67/************************ Governor internals ***********************/
68
69static bool sugov_should_update_freq(struct sugov_policy *sg_policy, u64 time)
70{
71 s64 delta_ns;
72
73 /*
74 * Since cpufreq_update_util() is called with rq->lock held for
75 * the @target_cpu, our per-CPU data is fully serialized.
76 *
77 * However, drivers cannot in general deal with cross-CPU
78 * requests, so while get_next_freq() will work, our
79 * sugov_update_commit() call may not for the fast switching platforms.
80 *
81 * Hence stop here for remote requests if they aren't supported
82 * by the hardware, as calculating the frequency is pointless if
83 * we cannot in fact act on it.
84 *
85 * For the slow switching platforms, the kthread is always scheduled on
86 * the right set of CPUs and any CPU can find the next frequency and
87 * schedule the kthread.
88 */
89 if (sg_policy->policy->fast_switch_enabled &&
90 !cpufreq_this_cpu_can_update(sg_policy->policy))
91 return false;
92
93 if (unlikely(sg_policy->limits_changed)) {
94 sg_policy->limits_changed = false;
95 sg_policy->need_freq_update = true;
96 return true;
97 }
98
99 delta_ns = time - sg_policy->last_freq_update_time;
100
101 return delta_ns >= sg_policy->freq_update_delay_ns;
102}
103
104static bool sugov_update_next_freq(struct sugov_policy *sg_policy, u64 time,
105 unsigned int next_freq)
106{
107 if (sg_policy->next_freq == next_freq)
108 return false;
109
110 sg_policy->next_freq = next_freq;
111 sg_policy->last_freq_update_time = time;
112
113 return true;
114}
115
116static void sugov_fast_switch(struct sugov_policy *sg_policy, u64 time,
117 unsigned int next_freq)
118{
119 struct cpufreq_policy *policy = sg_policy->policy;
120 int cpu;
121
122 if (!sugov_update_next_freq(sg_policy, time, next_freq))
123 return;
124
125 next_freq = cpufreq_driver_fast_switch(policy, next_freq);
126 if (!next_freq)
127 return;
128
129 policy->cur = next_freq;
130
131 if (trace_cpu_frequency_enabled()) {
132 for_each_cpu(cpu, policy->cpus)
133 trace_cpu_frequency(next_freq, cpu);
134 }
135}
136
137static void sugov_deferred_update(struct sugov_policy *sg_policy, u64 time,
138 unsigned int next_freq)
139{
140 if (!sugov_update_next_freq(sg_policy, time, next_freq))
141 return;
142
143 if (!sg_policy->work_in_progress) {
144 sg_policy->work_in_progress = true;
145 irq_work_queue(&sg_policy->irq_work);
146 }
147}
148
149/**
150 * get_next_freq - Compute a new frequency for a given cpufreq policy.
151 * @sg_policy: schedutil policy object to compute the new frequency for.
152 * @util: Current CPU utilization.
153 * @max: CPU capacity.
154 *
155 * If the utilization is frequency-invariant, choose the new frequency to be
156 * proportional to it, that is
157 *
158 * next_freq = C * max_freq * util / max
159 *
160 * Otherwise, approximate the would-be frequency-invariant utilization by
161 * util_raw * (curr_freq / max_freq) which leads to
162 *
163 * next_freq = C * curr_freq * util_raw / max
164 *
165 * Take C = 1.25 for the frequency tipping point at (util / max) = 0.8.
166 *
167 * The lowest driver-supported frequency which is equal or greater than the raw
168 * next_freq (as calculated above) is returned, subject to policy min/max and
169 * cpufreq driver limitations.
170 */
171static unsigned int get_next_freq(struct sugov_policy *sg_policy,
172 unsigned long util, unsigned long max)
173{
174 struct cpufreq_policy *policy = sg_policy->policy;
175 unsigned int freq = arch_scale_freq_invariant() ?
176 policy->cpuinfo.max_freq : policy->cur;
177
178 freq = map_util_freq(util, freq, max);
179
180 if (freq == sg_policy->cached_raw_freq && !sg_policy->need_freq_update)
181 return sg_policy->next_freq;
182
183 sg_policy->need_freq_update = false;
184 sg_policy->cached_raw_freq = freq;
185 return cpufreq_driver_resolve_freq(policy, freq);
186}
187
188/*
189 * This function computes an effective utilization for the given CPU, to be
190 * used for frequency selection given the linear relation: f = u * f_max.
191 *
192 * The scheduler tracks the following metrics:
193 *
194 * cpu_util_{cfs,rt,dl,irq}()
195 * cpu_bw_dl()
196 *
197 * Where the cfs,rt and dl util numbers are tracked with the same metric and
198 * synchronized windows and are thus directly comparable.
199 *
200 * The cfs,rt,dl utilization are the running times measured with rq->clock_task
201 * which excludes things like IRQ and steal-time. These latter are then accrued
202 * in the irq utilization.
203 *
204 * The DL bandwidth number otoh is not a measured metric but a value computed
205 * based on the task model parameters and gives the minimal utilization
206 * required to meet deadlines.
207 */
208unsigned long schedutil_cpu_util(int cpu, unsigned long util_cfs,
209 unsigned long max, enum schedutil_type type,
210 struct task_struct *p)
211{
212 unsigned long dl_util, util, irq;
213 struct rq *rq = cpu_rq(cpu);
214
215 if (!IS_BUILTIN(CONFIG_UCLAMP_TASK) &&
216 type == FREQUENCY_UTIL && rt_rq_is_runnable(&rq->rt)) {
217 return max;
218 }
219
220 /*
221 * Early check to see if IRQ/steal time saturates the CPU, can be
222 * because of inaccuracies in how we track these -- see
223 * update_irq_load_avg().
224 */
225 irq = cpu_util_irq(rq);
226 if (unlikely(irq >= max))
227 return max;
228
229 /*
230 * Because the time spend on RT/DL tasks is visible as 'lost' time to
231 * CFS tasks and we use the same metric to track the effective
232 * utilization (PELT windows are synchronized) we can directly add them
233 * to obtain the CPU's actual utilization.
234 *
235 * CFS and RT utilization can be boosted or capped, depending on
236 * utilization clamp constraints requested by currently RUNNABLE
237 * tasks.
238 * When there are no CFS RUNNABLE tasks, clamps are released and
239 * frequency will be gracefully reduced with the utilization decay.
240 */
241 util = util_cfs + cpu_util_rt(rq);
242 if (type == FREQUENCY_UTIL)
243 util = uclamp_util_with(rq, util, p);
244
245 dl_util = cpu_util_dl(rq);
246
247 /*
248 * For frequency selection we do not make cpu_util_dl() a permanent part
249 * of this sum because we want to use cpu_bw_dl() later on, but we need
250 * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
251 * that we select f_max when there is no idle time.
252 *
253 * NOTE: numerical errors or stop class might cause us to not quite hit
254 * saturation when we should -- something for later.
255 */
256 if (util + dl_util >= max)
257 return max;
258
259 /*
260 * OTOH, for energy computation we need the estimated running time, so
261 * include util_dl and ignore dl_bw.
262 */
263 if (type == ENERGY_UTIL)
264 util += dl_util;
265
266 /*
267 * There is still idle time; further improve the number by using the
268 * irq metric. Because IRQ/steal time is hidden from the task clock we
269 * need to scale the task numbers:
270 *
271 * max - irq
272 * U' = irq + --------- * U
273 * max
274 */
275 util = scale_irq_capacity(util, irq, max);
276 util += irq;
277
278 /*
279 * Bandwidth required by DEADLINE must always be granted while, for
280 * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
281 * to gracefully reduce the frequency when no tasks show up for longer
282 * periods of time.
283 *
284 * Ideally we would like to set bw_dl as min/guaranteed freq and util +
285 * bw_dl as requested freq. However, cpufreq is not yet ready for such
286 * an interface. So, we only do the latter for now.
287 */
288 if (type == FREQUENCY_UTIL)
289 util += cpu_bw_dl(rq);
290
291 return min(max, util);
292}
293
294static unsigned long sugov_get_util(struct sugov_cpu *sg_cpu)
295{
296 struct rq *rq = cpu_rq(sg_cpu->cpu);
297 unsigned long util = cpu_util_cfs(rq);
298 unsigned long max = arch_scale_cpu_capacity(sg_cpu->cpu);
299
300 sg_cpu->max = max;
301 sg_cpu->bw_dl = cpu_bw_dl(rq);
302
303 return schedutil_cpu_util(sg_cpu->cpu, util, max, FREQUENCY_UTIL, NULL);
304}
305
306/**
307 * sugov_iowait_reset() - Reset the IO boost status of a CPU.
308 * @sg_cpu: the sugov data for the CPU to boost
309 * @time: the update time from the caller
310 * @set_iowait_boost: true if an IO boost has been requested
311 *
312 * The IO wait boost of a task is disabled after a tick since the last update
313 * of a CPU. If a new IO wait boost is requested after more then a tick, then
314 * we enable the boost starting from IOWAIT_BOOST_MIN, which improves energy
315 * efficiency by ignoring sporadic wakeups from IO.
316 */
317static bool sugov_iowait_reset(struct sugov_cpu *sg_cpu, u64 time,
318 bool set_iowait_boost)
319{
320 s64 delta_ns = time - sg_cpu->last_update;
321
322 /* Reset boost only if a tick has elapsed since last request */
323 if (delta_ns <= TICK_NSEC)
324 return false;
325
326 sg_cpu->iowait_boost = set_iowait_boost ? IOWAIT_BOOST_MIN : 0;
327 sg_cpu->iowait_boost_pending = set_iowait_boost;
328
329 return true;
330}
331
332/**
333 * sugov_iowait_boost() - Updates the IO boost status of a CPU.
334 * @sg_cpu: the sugov data for the CPU to boost
335 * @time: the update time from the caller
336 * @flags: SCHED_CPUFREQ_IOWAIT if the task is waking up after an IO wait
337 *
338 * Each time a task wakes up after an IO operation, the CPU utilization can be
339 * boosted to a certain utilization which doubles at each "frequent and
340 * successive" wakeup from IO, ranging from IOWAIT_BOOST_MIN to the utilization
341 * of the maximum OPP.
342 *
343 * To keep doubling, an IO boost has to be requested at least once per tick,
344 * otherwise we restart from the utilization of the minimum OPP.
345 */
346static void sugov_iowait_boost(struct sugov_cpu *sg_cpu, u64 time,
347 unsigned int flags)
348{
349 bool set_iowait_boost = flags & SCHED_CPUFREQ_IOWAIT;
350
351 /* Reset boost if the CPU appears to have been idle enough */
352 if (sg_cpu->iowait_boost &&
353 sugov_iowait_reset(sg_cpu, time, set_iowait_boost))
354 return;
355
356 /* Boost only tasks waking up after IO */
357 if (!set_iowait_boost)
358 return;
359
360 /* Ensure boost doubles only one time at each request */
361 if (sg_cpu->iowait_boost_pending)
362 return;
363 sg_cpu->iowait_boost_pending = true;
364
365 /* Double the boost at each request */
366 if (sg_cpu->iowait_boost) {
367 sg_cpu->iowait_boost =
368 min_t(unsigned int, sg_cpu->iowait_boost << 1, SCHED_CAPACITY_SCALE);
369 return;
370 }
371
372 /* First wakeup after IO: start with minimum boost */
373 sg_cpu->iowait_boost = IOWAIT_BOOST_MIN;
374}
375
376/**
377 * sugov_iowait_apply() - Apply the IO boost to a CPU.
378 * @sg_cpu: the sugov data for the cpu to boost
379 * @time: the update time from the caller
380 * @util: the utilization to (eventually) boost
381 * @max: the maximum value the utilization can be boosted to
382 *
383 * A CPU running a task which woken up after an IO operation can have its
384 * utilization boosted to speed up the completion of those IO operations.
385 * The IO boost value is increased each time a task wakes up from IO, in
386 * sugov_iowait_apply(), and it's instead decreased by this function,
387 * each time an increase has not been requested (!iowait_boost_pending).
388 *
389 * A CPU which also appears to have been idle for at least one tick has also
390 * its IO boost utilization reset.
391 *
392 * This mechanism is designed to boost high frequently IO waiting tasks, while
393 * being more conservative on tasks which does sporadic IO operations.
394 */
395static unsigned long sugov_iowait_apply(struct sugov_cpu *sg_cpu, u64 time,
396 unsigned long util, unsigned long max)
397{
398 unsigned long boost;
399
400 /* No boost currently required */
401 if (!sg_cpu->iowait_boost)
402 return util;
403
404 /* Reset boost if the CPU appears to have been idle enough */
405 if (sugov_iowait_reset(sg_cpu, time, false))
406 return util;
407
408 if (!sg_cpu->iowait_boost_pending) {
409 /*
410 * No boost pending; reduce the boost value.
411 */
412 sg_cpu->iowait_boost >>= 1;
413 if (sg_cpu->iowait_boost < IOWAIT_BOOST_MIN) {
414 sg_cpu->iowait_boost = 0;
415 return util;
416 }
417 }
418
419 sg_cpu->iowait_boost_pending = false;
420
421 /*
422 * @util is already in capacity scale; convert iowait_boost
423 * into the same scale so we can compare.
424 */
425 boost = (sg_cpu->iowait_boost * max) >> SCHED_CAPACITY_SHIFT;
426 return max(boost, util);
427}
428
429#ifdef CONFIG_NO_HZ_COMMON
430static bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu)
431{
432 unsigned long idle_calls = tick_nohz_get_idle_calls_cpu(sg_cpu->cpu);
433 bool ret = idle_calls == sg_cpu->saved_idle_calls;
434
435 sg_cpu->saved_idle_calls = idle_calls;
436 return ret;
437}
438#else
439static inline bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu) { return false; }
440#endif /* CONFIG_NO_HZ_COMMON */
441
442/*
443 * Make sugov_should_update_freq() ignore the rate limit when DL
444 * has increased the utilization.
445 */
446static inline void ignore_dl_rate_limit(struct sugov_cpu *sg_cpu, struct sugov_policy *sg_policy)
447{
448 if (cpu_bw_dl(cpu_rq(sg_cpu->cpu)) > sg_cpu->bw_dl)
449 sg_policy->limits_changed = true;
450}
451
452static void sugov_update_single(struct update_util_data *hook, u64 time,
453 unsigned int flags)
454{
455 struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
456 struct sugov_policy *sg_policy = sg_cpu->sg_policy;
457 unsigned long util, max;
458 unsigned int next_f;
459 bool busy;
460
461 sugov_iowait_boost(sg_cpu, time, flags);
462 sg_cpu->last_update = time;
463
464 ignore_dl_rate_limit(sg_cpu, sg_policy);
465
466 if (!sugov_should_update_freq(sg_policy, time))
467 return;
468
469 /* Limits may have changed, don't skip frequency update */
470 busy = !sg_policy->need_freq_update && sugov_cpu_is_busy(sg_cpu);
471
472 util = sugov_get_util(sg_cpu);
473 max = sg_cpu->max;
474 util = sugov_iowait_apply(sg_cpu, time, util, max);
475 next_f = get_next_freq(sg_policy, util, max);
476 /*
477 * Do not reduce the frequency if the CPU has not been idle
478 * recently, as the reduction is likely to be premature then.
479 */
480 if (busy && next_f < sg_policy->next_freq) {
481 next_f = sg_policy->next_freq;
482
483 /* Reset cached freq as next_freq has changed */
484 sg_policy->cached_raw_freq = 0;
485 }
486
487 /*
488 * This code runs under rq->lock for the target CPU, so it won't run
489 * concurrently on two different CPUs for the same target and it is not
490 * necessary to acquire the lock in the fast switch case.
491 */
492 if (sg_policy->policy->fast_switch_enabled) {
493 sugov_fast_switch(sg_policy, time, next_f);
494 } else {
495 raw_spin_lock(&sg_policy->update_lock);
496 sugov_deferred_update(sg_policy, time, next_f);
497 raw_spin_unlock(&sg_policy->update_lock);
498 }
499}
500
501static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu, u64 time)
502{
503 struct sugov_policy *sg_policy = sg_cpu->sg_policy;
504 struct cpufreq_policy *policy = sg_policy->policy;
505 unsigned long util = 0, max = 1;
506 unsigned int j;
507
508 for_each_cpu(j, policy->cpus) {
509 struct sugov_cpu *j_sg_cpu = &per_cpu(sugov_cpu, j);
510 unsigned long j_util, j_max;
511
512 j_util = sugov_get_util(j_sg_cpu);
513 j_max = j_sg_cpu->max;
514 j_util = sugov_iowait_apply(j_sg_cpu, time, j_util, j_max);
515
516 if (j_util * max > j_max * util) {
517 util = j_util;
518 max = j_max;
519 }
520 }
521
522 return get_next_freq(sg_policy, util, max);
523}
524
525static void
526sugov_update_shared(struct update_util_data *hook, u64 time, unsigned int flags)
527{
528 struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
529 struct sugov_policy *sg_policy = sg_cpu->sg_policy;
530 unsigned int next_f;
531
532 raw_spin_lock(&sg_policy->update_lock);
533
534 sugov_iowait_boost(sg_cpu, time, flags);
535 sg_cpu->last_update = time;
536
537 ignore_dl_rate_limit(sg_cpu, sg_policy);
538
539 if (sugov_should_update_freq(sg_policy, time)) {
540 next_f = sugov_next_freq_shared(sg_cpu, time);
541
542 if (sg_policy->policy->fast_switch_enabled)
543 sugov_fast_switch(sg_policy, time, next_f);
544 else
545 sugov_deferred_update(sg_policy, time, next_f);
546 }
547
548 raw_spin_unlock(&sg_policy->update_lock);
549}
550
551static void sugov_work(struct kthread_work *work)
552{
553 struct sugov_policy *sg_policy = container_of(work, struct sugov_policy, work);
554 unsigned int freq;
555 unsigned long flags;
556
557 /*
558 * Hold sg_policy->update_lock shortly to handle the case where:
559 * incase sg_policy->next_freq is read here, and then updated by
560 * sugov_deferred_update() just before work_in_progress is set to false
561 * here, we may miss queueing the new update.
562 *
563 * Note: If a work was queued after the update_lock is released,
564 * sugov_work() will just be called again by kthread_work code; and the
565 * request will be proceed before the sugov thread sleeps.
566 */
567 raw_spin_lock_irqsave(&sg_policy->update_lock, flags);
568 freq = sg_policy->next_freq;
569 sg_policy->work_in_progress = false;
570 raw_spin_unlock_irqrestore(&sg_policy->update_lock, flags);
571
572 mutex_lock(&sg_policy->work_lock);
573 __cpufreq_driver_target(sg_policy->policy, freq, CPUFREQ_RELATION_L);
574 mutex_unlock(&sg_policy->work_lock);
575}
576
577static void sugov_irq_work(struct irq_work *irq_work)
578{
579 struct sugov_policy *sg_policy;
580
581 sg_policy = container_of(irq_work, struct sugov_policy, irq_work);
582
583 kthread_queue_work(&sg_policy->worker, &sg_policy->work);
584}
585
586/************************** sysfs interface ************************/
587
588static struct sugov_tunables *global_tunables;
589static DEFINE_MUTEX(global_tunables_lock);
590
591static inline struct sugov_tunables *to_sugov_tunables(struct gov_attr_set *attr_set)
592{
593 return container_of(attr_set, struct sugov_tunables, attr_set);
594}
595
596static ssize_t rate_limit_us_show(struct gov_attr_set *attr_set, char *buf)
597{
598 struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
599
600 return sprintf(buf, "%u\n", tunables->rate_limit_us);
601}
602
603static ssize_t
604rate_limit_us_store(struct gov_attr_set *attr_set, const char *buf, size_t count)
605{
606 struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
607 struct sugov_policy *sg_policy;
608 unsigned int rate_limit_us;
609
610 if (kstrtouint(buf, 10, &rate_limit_us))
611 return -EINVAL;
612
613 tunables->rate_limit_us = rate_limit_us;
614
615 list_for_each_entry(sg_policy, &attr_set->policy_list, tunables_hook)
616 sg_policy->freq_update_delay_ns = rate_limit_us * NSEC_PER_USEC;
617
618 return count;
619}
620
621static struct governor_attr rate_limit_us = __ATTR_RW(rate_limit_us);
622
623static struct attribute *sugov_attrs[] = {
624 &rate_limit_us.attr,
625 NULL
626};
627ATTRIBUTE_GROUPS(sugov);
628
629static struct kobj_type sugov_tunables_ktype = {
630 .default_groups = sugov_groups,
631 .sysfs_ops = &governor_sysfs_ops,
632};
633
634/********************** cpufreq governor interface *********************/
635
636struct cpufreq_governor schedutil_gov;
637
638static struct sugov_policy *sugov_policy_alloc(struct cpufreq_policy *policy)
639{
640 struct sugov_policy *sg_policy;
641
642 sg_policy = kzalloc(sizeof(*sg_policy), GFP_KERNEL);
643 if (!sg_policy)
644 return NULL;
645
646 sg_policy->policy = policy;
647 raw_spin_lock_init(&sg_policy->update_lock);
648 return sg_policy;
649}
650
651static void sugov_policy_free(struct sugov_policy *sg_policy)
652{
653 kfree(sg_policy);
654}
655
656static int sugov_kthread_create(struct sugov_policy *sg_policy)
657{
658 struct task_struct *thread;
659 struct sched_attr attr = {
660 .size = sizeof(struct sched_attr),
661 .sched_policy = SCHED_DEADLINE,
662 .sched_flags = SCHED_FLAG_SUGOV,
663 .sched_nice = 0,
664 .sched_priority = 0,
665 /*
666 * Fake (unused) bandwidth; workaround to "fix"
667 * priority inheritance.
668 */
669 .sched_runtime = 1000000,
670 .sched_deadline = 10000000,
671 .sched_period = 10000000,
672 };
673 struct cpufreq_policy *policy = sg_policy->policy;
674 int ret;
675
676 /* kthread only required for slow path */
677 if (policy->fast_switch_enabled)
678 return 0;
679
680 kthread_init_work(&sg_policy->work, sugov_work);
681 kthread_init_worker(&sg_policy->worker);
682 thread = kthread_create(kthread_worker_fn, &sg_policy->worker,
683 "sugov:%d",
684 cpumask_first(policy->related_cpus));
685 if (IS_ERR(thread)) {
686 pr_err("failed to create sugov thread: %ld\n", PTR_ERR(thread));
687 return PTR_ERR(thread);
688 }
689
690 ret = sched_setattr_nocheck(thread, &attr);
691 if (ret) {
692 kthread_stop(thread);
693 pr_warn("%s: failed to set SCHED_DEADLINE\n", __func__);
694 return ret;
695 }
696
697 sg_policy->thread = thread;
698 kthread_bind_mask(thread, policy->related_cpus);
699 init_irq_work(&sg_policy->irq_work, sugov_irq_work);
700 mutex_init(&sg_policy->work_lock);
701
702 wake_up_process(thread);
703
704 return 0;
705}
706
707static void sugov_kthread_stop(struct sugov_policy *sg_policy)
708{
709 /* kthread only required for slow path */
710 if (sg_policy->policy->fast_switch_enabled)
711 return;
712
713 kthread_flush_worker(&sg_policy->worker);
714 kthread_stop(sg_policy->thread);
715 mutex_destroy(&sg_policy->work_lock);
716}
717
718static struct sugov_tunables *sugov_tunables_alloc(struct sugov_policy *sg_policy)
719{
720 struct sugov_tunables *tunables;
721
722 tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
723 if (tunables) {
724 gov_attr_set_init(&tunables->attr_set, &sg_policy->tunables_hook);
725 if (!have_governor_per_policy())
726 global_tunables = tunables;
727 }
728 return tunables;
729}
730
731static void sugov_tunables_free(struct sugov_tunables *tunables)
732{
733 if (!have_governor_per_policy())
734 global_tunables = NULL;
735
736 kfree(tunables);
737}
738
739static int sugov_init(struct cpufreq_policy *policy)
740{
741 struct sugov_policy *sg_policy;
742 struct sugov_tunables *tunables;
743 int ret = 0;
744
745 /* State should be equivalent to EXIT */
746 if (policy->governor_data)
747 return -EBUSY;
748
749 cpufreq_enable_fast_switch(policy);
750
751 sg_policy = sugov_policy_alloc(policy);
752 if (!sg_policy) {
753 ret = -ENOMEM;
754 goto disable_fast_switch;
755 }
756
757 ret = sugov_kthread_create(sg_policy);
758 if (ret)
759 goto free_sg_policy;
760
761 mutex_lock(&global_tunables_lock);
762
763 if (global_tunables) {
764 if (WARN_ON(have_governor_per_policy())) {
765 ret = -EINVAL;
766 goto stop_kthread;
767 }
768 policy->governor_data = sg_policy;
769 sg_policy->tunables = global_tunables;
770
771 gov_attr_set_get(&global_tunables->attr_set, &sg_policy->tunables_hook);
772 goto out;
773 }
774
775 tunables = sugov_tunables_alloc(sg_policy);
776 if (!tunables) {
777 ret = -ENOMEM;
778 goto stop_kthread;
779 }
780
781 tunables->rate_limit_us = cpufreq_policy_transition_delay_us(policy);
782
783 policy->governor_data = sg_policy;
784 sg_policy->tunables = tunables;
785
786 ret = kobject_init_and_add(&tunables->attr_set.kobj, &sugov_tunables_ktype,
787 get_governor_parent_kobj(policy), "%s",
788 schedutil_gov.name);
789 if (ret)
790 goto fail;
791
792out:
793 mutex_unlock(&global_tunables_lock);
794 return 0;
795
796fail:
797 kobject_put(&tunables->attr_set.kobj);
798 policy->governor_data = NULL;
799 sugov_tunables_free(tunables);
800
801stop_kthread:
802 sugov_kthread_stop(sg_policy);
803 mutex_unlock(&global_tunables_lock);
804
805free_sg_policy:
806 sugov_policy_free(sg_policy);
807
808disable_fast_switch:
809 cpufreq_disable_fast_switch(policy);
810
811 pr_err("initialization failed (error %d)\n", ret);
812 return ret;
813}
814
815static void sugov_exit(struct cpufreq_policy *policy)
816{
817 struct sugov_policy *sg_policy = policy->governor_data;
818 struct sugov_tunables *tunables = sg_policy->tunables;
819 unsigned int count;
820
821 mutex_lock(&global_tunables_lock);
822
823 count = gov_attr_set_put(&tunables->attr_set, &sg_policy->tunables_hook);
824 policy->governor_data = NULL;
825 if (!count)
826 sugov_tunables_free(tunables);
827
828 mutex_unlock(&global_tunables_lock);
829
830 sugov_kthread_stop(sg_policy);
831 sugov_policy_free(sg_policy);
832 cpufreq_disable_fast_switch(policy);
833}
834
835static int sugov_start(struct cpufreq_policy *policy)
836{
837 struct sugov_policy *sg_policy = policy->governor_data;
838 unsigned int cpu;
839
840 sg_policy->freq_update_delay_ns = sg_policy->tunables->rate_limit_us * NSEC_PER_USEC;
841 sg_policy->last_freq_update_time = 0;
842 sg_policy->next_freq = 0;
843 sg_policy->work_in_progress = false;
844 sg_policy->limits_changed = false;
845 sg_policy->need_freq_update = false;
846 sg_policy->cached_raw_freq = 0;
847
848 for_each_cpu(cpu, policy->cpus) {
849 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
850
851 memset(sg_cpu, 0, sizeof(*sg_cpu));
852 sg_cpu->cpu = cpu;
853 sg_cpu->sg_policy = sg_policy;
854 }
855
856 for_each_cpu(cpu, policy->cpus) {
857 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
858
859 cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util,
860 policy_is_shared(policy) ?
861 sugov_update_shared :
862 sugov_update_single);
863 }
864 return 0;
865}
866
867static void sugov_stop(struct cpufreq_policy *policy)
868{
869 struct sugov_policy *sg_policy = policy->governor_data;
870 unsigned int cpu;
871
872 for_each_cpu(cpu, policy->cpus)
873 cpufreq_remove_update_util_hook(cpu);
874
875 synchronize_rcu();
876
877 if (!policy->fast_switch_enabled) {
878 irq_work_sync(&sg_policy->irq_work);
879 kthread_cancel_work_sync(&sg_policy->work);
880 }
881}
882
883static void sugov_limits(struct cpufreq_policy *policy)
884{
885 struct sugov_policy *sg_policy = policy->governor_data;
886
887 if (!policy->fast_switch_enabled) {
888 mutex_lock(&sg_policy->work_lock);
889 cpufreq_policy_apply_limits(policy);
890 mutex_unlock(&sg_policy->work_lock);
891 }
892
893 sg_policy->limits_changed = true;
894}
895
896struct cpufreq_governor schedutil_gov = {
897 .name = "schedutil",
898 .owner = THIS_MODULE,
899 .dynamic_switching = true,
900 .init = sugov_init,
901 .exit = sugov_exit,
902 .start = sugov_start,
903 .stop = sugov_stop,
904 .limits = sugov_limits,
905};
906
907#ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL
908struct cpufreq_governor *cpufreq_default_governor(void)
909{
910 return &schedutil_gov;
911}
912#endif
913
914static int __init sugov_register(void)
915{
916 return cpufreq_register_governor(&schedutil_gov);
917}
918fs_initcall(sugov_register);
919
920#ifdef CONFIG_ENERGY_MODEL
921extern bool sched_energy_update;
922extern struct mutex sched_energy_mutex;
923
924static void rebuild_sd_workfn(struct work_struct *work)
925{
926 mutex_lock(&sched_energy_mutex);
927 sched_energy_update = true;
928 rebuild_sched_domains();
929 sched_energy_update = false;
930 mutex_unlock(&sched_energy_mutex);
931}
932static DECLARE_WORK(rebuild_sd_work, rebuild_sd_workfn);
933
934/*
935 * EAS shouldn't be attempted without sugov, so rebuild the sched_domains
936 * on governor changes to make sure the scheduler knows about it.
937 */
938void sched_cpufreq_governor_change(struct cpufreq_policy *policy,
939 struct cpufreq_governor *old_gov)
940{
941 if (old_gov == &schedutil_gov || policy->governor == &schedutil_gov) {
942 /*
943 * When called from the cpufreq_register_driver() path, the
944 * cpu_hotplug_lock is already held, so use a work item to
945 * avoid nested locking in rebuild_sched_domains().
946 */
947 schedule_work(&rebuild_sd_work);
948 }
949
950}
951#endif