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