Linux Audio

Check our new training course

Loading...
v6.2
  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
v4.17
 
  1/*
  2 * CPUFreq governor based on scheduler-provided CPU utilization data.
  3 *
  4 * Copyright (C) 2016, Intel Corporation
  5 * Author: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
  6 *
  7 * This program is free software; you can redistribute it and/or modify
  8 * it under the terms of the GNU General Public License version 2 as
  9 * published by the Free Software Foundation.
 10 */
 11
 12#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
 13
 14#include "sched.h"
 15
 16#include <trace/events/power.h>
 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			need_freq_update;
 44};
 45
 46struct sugov_cpu {
 47	struct update_util_data	update_util;
 48	struct sugov_policy	*sg_policy;
 49	unsigned int		cpu;
 50
 51	bool			iowait_boost_pending;
 52	unsigned int		iowait_boost;
 53	unsigned int		iowait_boost_max;
 54	u64 last_update;
 55
 56	/* The fields below are only needed when sharing a policy: */
 57	unsigned long		util_cfs;
 58	unsigned long		util_dl;
 59	unsigned long		max;
 60
 61	/* The field below is for single-CPU policies only: */
 62#ifdef CONFIG_NO_HZ_COMMON
 63	unsigned long		saved_idle_calls;
 64#endif
 65};
 66
 67static DEFINE_PER_CPU(struct sugov_cpu, sugov_cpu);
 68
 69/************************ Governor internals ***********************/
 70
 71static bool sugov_should_update_freq(struct sugov_policy *sg_policy, u64 time)
 72{
 73	s64 delta_ns;
 74
 75	/*
 76	 * Since cpufreq_update_util() is called with rq->lock held for
 77	 * the @target_cpu, our per-CPU data is fully serialized.
 78	 *
 79	 * However, drivers cannot in general deal with cross-CPU
 80	 * requests, so while get_next_freq() will work, our
 81	 * sugov_update_commit() call may not for the fast switching platforms.
 82	 *
 83	 * Hence stop here for remote requests if they aren't supported
 84	 * by the hardware, as calculating the frequency is pointless if
 85	 * we cannot in fact act on it.
 86	 *
 87	 * For the slow switching platforms, the kthread is always scheduled on
 88	 * the right set of CPUs and any CPU can find the next frequency and
 89	 * schedule the kthread.
 90	 */
 91	if (sg_policy->policy->fast_switch_enabled &&
 92	    !cpufreq_can_do_remote_dvfs(sg_policy->policy))
 93		return false;
 94
 95	if (sg_policy->work_in_progress)
 96		return false;
 97
 98	if (unlikely(sg_policy->need_freq_update)) {
 99		sg_policy->need_freq_update = false;
100		/*
101		 * This happens when limits change, so forget the previous
102		 * next_freq value and force an update.
103		 */
104		sg_policy->next_freq = UINT_MAX;
105		return true;
106	}
107
108	delta_ns = time - sg_policy->last_freq_update_time;
109
110	return delta_ns >= sg_policy->freq_update_delay_ns;
111}
112
113static void sugov_update_commit(struct sugov_policy *sg_policy, u64 time,
114				unsigned int next_freq)
115{
116	struct cpufreq_policy *policy = sg_policy->policy;
117
118	if (sg_policy->next_freq == next_freq)
119		return;
120
121	sg_policy->next_freq = next_freq;
122	sg_policy->last_freq_update_time = time;
123
124	if (policy->fast_switch_enabled) {
125		next_freq = cpufreq_driver_fast_switch(policy, next_freq);
126		if (!next_freq)
127			return;
128
129		policy->cur = next_freq;
130		trace_cpu_frequency(next_freq, smp_processor_id());
131	} else {
132		sg_policy->work_in_progress = true;
133		irq_work_queue(&sg_policy->irq_work);
134	}
135}
136
137/**
138 * get_next_freq - Compute a new frequency for a given cpufreq policy.
139 * @sg_policy: schedutil policy object to compute the new frequency for.
140 * @util: Current CPU utilization.
141 * @max: CPU capacity.
142 *
143 * If the utilization is frequency-invariant, choose the new frequency to be
144 * proportional to it, that is
145 *
146 * next_freq = C * max_freq * util / max
147 *
148 * Otherwise, approximate the would-be frequency-invariant utilization by
149 * util_raw * (curr_freq / max_freq) which leads to
150 *
151 * next_freq = C * curr_freq * util_raw / max
152 *
153 * Take C = 1.25 for the frequency tipping point at (util / max) = 0.8.
154 *
155 * The lowest driver-supported frequency which is equal or greater than the raw
156 * next_freq (as calculated above) is returned, subject to policy min/max and
157 * cpufreq driver limitations.
158 */
159static unsigned int get_next_freq(struct sugov_policy *sg_policy,
160				  unsigned long util, unsigned long max)
161{
162	struct cpufreq_policy *policy = sg_policy->policy;
163	unsigned int freq = arch_scale_freq_invariant() ?
164				policy->cpuinfo.max_freq : policy->cur;
165
166	freq = (freq + (freq >> 2)) * util / max;
 
167
168	if (freq == sg_policy->cached_raw_freq && sg_policy->next_freq != UINT_MAX)
169		return sg_policy->next_freq;
 
170	sg_policy->cached_raw_freq = freq;
171	return cpufreq_driver_resolve_freq(policy, freq);
172}
173
174static void sugov_get_util(struct sugov_cpu *sg_cpu)
175{
176	struct rq *rq = cpu_rq(sg_cpu->cpu);
177
178	sg_cpu->max = arch_scale_cpu_capacity(NULL, sg_cpu->cpu);
179	sg_cpu->util_cfs = cpu_util_cfs(rq);
180	sg_cpu->util_dl  = cpu_util_dl(rq);
 
181}
182
183static unsigned long sugov_aggregate_util(struct sugov_cpu *sg_cpu)
 
 
 
 
 
 
 
 
 
 
 
 
184{
185	struct rq *rq = cpu_rq(sg_cpu->cpu);
186	unsigned long util;
 
 
 
187
188	if (rq->rt.rt_nr_running) {
189		util = sg_cpu->max;
190	} else {
191		util = sg_cpu->util_dl;
192		if (rq->cfs.h_nr_running)
193			util += sg_cpu->util_cfs;
194	}
195
196	/*
197	 * Ideally we would like to set util_dl as min/guaranteed freq and
198	 * util_cfs + util_dl as requested freq. However, cpufreq is not yet
199	 * ready for such an interface. So, we only do the latter for now.
200	 */
201	return min(util, sg_cpu->max);
202}
203
204static void sugov_set_iowait_boost(struct sugov_cpu *sg_cpu, u64 time, unsigned int flags)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
205{
206	if (flags & SCHED_CPUFREQ_IOWAIT) {
207		if (sg_cpu->iowait_boost_pending)
208			return;
 
 
 
209
210		sg_cpu->iowait_boost_pending = true;
 
 
211
212		if (sg_cpu->iowait_boost) {
213			sg_cpu->iowait_boost <<= 1;
214			if (sg_cpu->iowait_boost > sg_cpu->iowait_boost_max)
215				sg_cpu->iowait_boost = sg_cpu->iowait_boost_max;
216		} else {
217			sg_cpu->iowait_boost = sg_cpu->sg_policy->policy->min;
218		}
219	} else if (sg_cpu->iowait_boost) {
220		s64 delta_ns = time - sg_cpu->last_update;
221
222		/* Clear iowait_boost if the CPU apprears to have been idle. */
223		if (delta_ns > TICK_NSEC) {
224			sg_cpu->iowait_boost = 0;
225			sg_cpu->iowait_boost_pending = false;
226		}
227	}
 
 
 
228}
229
230static void sugov_iowait_boost(struct sugov_cpu *sg_cpu, unsigned long *util,
231			       unsigned long *max)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
232{
233	unsigned int boost_util, boost_max;
234
 
235	if (!sg_cpu->iowait_boost)
236		return;
237
238	if (sg_cpu->iowait_boost_pending) {
239		sg_cpu->iowait_boost_pending = false;
240	} else {
 
 
 
 
 
241		sg_cpu->iowait_boost >>= 1;
242		if (sg_cpu->iowait_boost < sg_cpu->sg_policy->policy->min) {
243			sg_cpu->iowait_boost = 0;
244			return;
245		}
246	}
247
248	boost_util = sg_cpu->iowait_boost;
249	boost_max = sg_cpu->iowait_boost_max;
250
251	if (*util * boost_max < *max * boost_util) {
252		*util = boost_util;
253		*max = boost_max;
254	}
 
 
 
 
255}
256
257#ifdef CONFIG_NO_HZ_COMMON
258static bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu)
259{
260	unsigned long idle_calls = tick_nohz_get_idle_calls_cpu(sg_cpu->cpu);
261	bool ret = idle_calls == sg_cpu->saved_idle_calls;
262
263	sg_cpu->saved_idle_calls = idle_calls;
264	return ret;
265}
266#else
267static inline bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu) { return false; }
268#endif /* CONFIG_NO_HZ_COMMON */
269
270/*
271 * Make sugov_should_update_freq() ignore the rate limit when DL
272 * has increased the utilization.
273 */
274static inline void ignore_dl_rate_limit(struct sugov_cpu *sg_cpu, struct sugov_policy *sg_policy)
 
 
 
 
 
 
 
275{
276	if (cpu_util_dl(cpu_rq(sg_cpu->cpu)) > sg_cpu->util_dl)
277		sg_policy->need_freq_update = true;
 
 
 
 
 
 
 
 
 
 
278}
279
280static void sugov_update_single(struct update_util_data *hook, u64 time,
281				unsigned int flags)
282{
283	struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
284	struct sugov_policy *sg_policy = sg_cpu->sg_policy;
285	unsigned long util, max;
286	unsigned int next_f;
287	bool busy;
288
289	sugov_set_iowait_boost(sg_cpu, time, flags);
290	sg_cpu->last_update = time;
291
292	ignore_dl_rate_limit(sg_cpu, sg_policy);
293
294	if (!sugov_should_update_freq(sg_policy, time))
295		return;
296
297	busy = sugov_cpu_is_busy(sg_cpu);
298
299	sugov_get_util(sg_cpu);
300	max = sg_cpu->max;
301	util = sugov_aggregate_util(sg_cpu);
302	sugov_iowait_boost(sg_cpu, &util, &max);
303	next_f = get_next_freq(sg_policy, util, max);
304	/*
305	 * Do not reduce the frequency if the CPU has not been idle
306	 * recently, as the reduction is likely to be premature then.
 
 
307	 */
308	if (busy && next_f < sg_policy->next_freq &&
309	    sg_policy->next_freq != UINT_MAX) {
310		next_f = sg_policy->next_freq;
311
312		/* Reset cached freq as next_freq has changed */
313		sg_policy->cached_raw_freq = 0;
314	}
315
316	sugov_update_commit(sg_policy, time, next_f);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
317}
318
319static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu, u64 time)
320{
321	struct sugov_policy *sg_policy = sg_cpu->sg_policy;
322	struct cpufreq_policy *policy = sg_policy->policy;
323	unsigned long util = 0, max = 1;
324	unsigned int j;
325
326	for_each_cpu(j, policy->cpus) {
327		struct sugov_cpu *j_sg_cpu = &per_cpu(sugov_cpu, j);
328		unsigned long j_util, j_max;
329		s64 delta_ns;
330
331		sugov_get_util(j_sg_cpu);
 
 
 
332
333		/*
334		 * If the CFS CPU utilization was last updated before the
335		 * previous frequency update and the time elapsed between the
336		 * last update of the CPU utilization and the last frequency
337		 * update is long enough, reset iowait_boost and util_cfs, as
338		 * they are now probably stale. However, still consider the
339		 * CPU contribution if it has some DEADLINE utilization
340		 * (util_dl).
341		 */
342		delta_ns = time - j_sg_cpu->last_update;
343		if (delta_ns > TICK_NSEC) {
344			j_sg_cpu->iowait_boost = 0;
345			j_sg_cpu->iowait_boost_pending = false;
346		}
347
348		j_max = j_sg_cpu->max;
349		j_util = sugov_aggregate_util(j_sg_cpu);
350		sugov_iowait_boost(j_sg_cpu, &j_util, &j_max);
351		if (j_util * max > j_max * util) {
352			util = j_util;
353			max = j_max;
354		}
355	}
356
357	return get_next_freq(sg_policy, util, max);
358}
359
360static void
361sugov_update_shared(struct update_util_data *hook, u64 time, unsigned int flags)
362{
363	struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
364	struct sugov_policy *sg_policy = sg_cpu->sg_policy;
365	unsigned int next_f;
366
367	raw_spin_lock(&sg_policy->update_lock);
368
369	sugov_set_iowait_boost(sg_cpu, time, flags);
370	sg_cpu->last_update = time;
371
372	ignore_dl_rate_limit(sg_cpu, sg_policy);
373
374	if (sugov_should_update_freq(sg_policy, time)) {
375		next_f = sugov_next_freq_shared(sg_cpu, time);
376		sugov_update_commit(sg_policy, time, next_f);
 
 
 
 
 
 
 
377	}
378
379	raw_spin_unlock(&sg_policy->update_lock);
380}
381
382static void sugov_work(struct kthread_work *work)
383{
384	struct sugov_policy *sg_policy = container_of(work, struct sugov_policy, work);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
385
386	mutex_lock(&sg_policy->work_lock);
387	__cpufreq_driver_target(sg_policy->policy, sg_policy->next_freq,
388				CPUFREQ_RELATION_L);
389	mutex_unlock(&sg_policy->work_lock);
390
391	sg_policy->work_in_progress = false;
392}
393
394static void sugov_irq_work(struct irq_work *irq_work)
395{
396	struct sugov_policy *sg_policy;
397
398	sg_policy = container_of(irq_work, struct sugov_policy, irq_work);
399
400	kthread_queue_work(&sg_policy->worker, &sg_policy->work);
401}
402
403/************************** sysfs interface ************************/
404
405static struct sugov_tunables *global_tunables;
406static DEFINE_MUTEX(global_tunables_lock);
407
408static inline struct sugov_tunables *to_sugov_tunables(struct gov_attr_set *attr_set)
409{
410	return container_of(attr_set, struct sugov_tunables, attr_set);
411}
412
413static ssize_t rate_limit_us_show(struct gov_attr_set *attr_set, char *buf)
414{
415	struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
416
417	return sprintf(buf, "%u\n", tunables->rate_limit_us);
418}
419
420static ssize_t
421rate_limit_us_store(struct gov_attr_set *attr_set, const char *buf, size_t count)
422{
423	struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
424	struct sugov_policy *sg_policy;
425	unsigned int rate_limit_us;
426
427	if (kstrtouint(buf, 10, &rate_limit_us))
428		return -EINVAL;
429
430	tunables->rate_limit_us = rate_limit_us;
431
432	list_for_each_entry(sg_policy, &attr_set->policy_list, tunables_hook)
433		sg_policy->freq_update_delay_ns = rate_limit_us * NSEC_PER_USEC;
434
435	return count;
436}
437
438static struct governor_attr rate_limit_us = __ATTR_RW(rate_limit_us);
439
440static struct attribute *sugov_attributes[] = {
441	&rate_limit_us.attr,
442	NULL
443};
 
 
 
 
 
 
 
 
444
445static struct kobj_type sugov_tunables_ktype = {
446	.default_attrs = sugov_attributes,
447	.sysfs_ops = &governor_sysfs_ops,
 
448};
449
450/********************** cpufreq governor interface *********************/
451
452static struct cpufreq_governor schedutil_gov;
453
454static struct sugov_policy *sugov_policy_alloc(struct cpufreq_policy *policy)
455{
456	struct sugov_policy *sg_policy;
457
458	sg_policy = kzalloc(sizeof(*sg_policy), GFP_KERNEL);
459	if (!sg_policy)
460		return NULL;
461
462	sg_policy->policy = policy;
463	raw_spin_lock_init(&sg_policy->update_lock);
464	return sg_policy;
465}
466
467static void sugov_policy_free(struct sugov_policy *sg_policy)
468{
469	kfree(sg_policy);
470}
471
472static int sugov_kthread_create(struct sugov_policy *sg_policy)
473{
474	struct task_struct *thread;
475	struct sched_attr attr = {
476		.size		= sizeof(struct sched_attr),
477		.sched_policy	= SCHED_DEADLINE,
478		.sched_flags	= SCHED_FLAG_SUGOV,
479		.sched_nice	= 0,
480		.sched_priority	= 0,
481		/*
482		 * Fake (unused) bandwidth; workaround to "fix"
483		 * priority inheritance.
484		 */
485		.sched_runtime	=  1000000,
486		.sched_deadline = 10000000,
487		.sched_period	= 10000000,
488	};
489	struct cpufreq_policy *policy = sg_policy->policy;
490	int ret;
491
492	/* kthread only required for slow path */
493	if (policy->fast_switch_enabled)
494		return 0;
495
496	kthread_init_work(&sg_policy->work, sugov_work);
497	kthread_init_worker(&sg_policy->worker);
498	thread = kthread_create(kthread_worker_fn, &sg_policy->worker,
499				"sugov:%d",
500				cpumask_first(policy->related_cpus));
501	if (IS_ERR(thread)) {
502		pr_err("failed to create sugov thread: %ld\n", PTR_ERR(thread));
503		return PTR_ERR(thread);
504	}
505
506	ret = sched_setattr_nocheck(thread, &attr);
507	if (ret) {
508		kthread_stop(thread);
509		pr_warn("%s: failed to set SCHED_DEADLINE\n", __func__);
510		return ret;
511	}
512
513	sg_policy->thread = thread;
514
515	/* Kthread is bound to all CPUs by default */
516	if (!policy->dvfs_possible_from_any_cpu)
517		kthread_bind_mask(thread, policy->related_cpus);
518
519	init_irq_work(&sg_policy->irq_work, sugov_irq_work);
520	mutex_init(&sg_policy->work_lock);
521
522	wake_up_process(thread);
523
524	return 0;
525}
526
527static void sugov_kthread_stop(struct sugov_policy *sg_policy)
528{
529	/* kthread only required for slow path */
530	if (sg_policy->policy->fast_switch_enabled)
531		return;
532
533	kthread_flush_worker(&sg_policy->worker);
534	kthread_stop(sg_policy->thread);
535	mutex_destroy(&sg_policy->work_lock);
536}
537
538static struct sugov_tunables *sugov_tunables_alloc(struct sugov_policy *sg_policy)
539{
540	struct sugov_tunables *tunables;
541
542	tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
543	if (tunables) {
544		gov_attr_set_init(&tunables->attr_set, &sg_policy->tunables_hook);
545		if (!have_governor_per_policy())
546			global_tunables = tunables;
547	}
548	return tunables;
549}
550
551static void sugov_tunables_free(struct sugov_tunables *tunables)
552{
553	if (!have_governor_per_policy())
554		global_tunables = NULL;
555
556	kfree(tunables);
557}
558
559static int sugov_init(struct cpufreq_policy *policy)
560{
561	struct sugov_policy *sg_policy;
562	struct sugov_tunables *tunables;
563	int ret = 0;
564
565	/* State should be equivalent to EXIT */
566	if (policy->governor_data)
567		return -EBUSY;
568
569	cpufreq_enable_fast_switch(policy);
570
571	sg_policy = sugov_policy_alloc(policy);
572	if (!sg_policy) {
573		ret = -ENOMEM;
574		goto disable_fast_switch;
575	}
576
577	ret = sugov_kthread_create(sg_policy);
578	if (ret)
579		goto free_sg_policy;
580
581	mutex_lock(&global_tunables_lock);
582
583	if (global_tunables) {
584		if (WARN_ON(have_governor_per_policy())) {
585			ret = -EINVAL;
586			goto stop_kthread;
587		}
588		policy->governor_data = sg_policy;
589		sg_policy->tunables = global_tunables;
590
591		gov_attr_set_get(&global_tunables->attr_set, &sg_policy->tunables_hook);
592		goto out;
593	}
594
595	tunables = sugov_tunables_alloc(sg_policy);
596	if (!tunables) {
597		ret = -ENOMEM;
598		goto stop_kthread;
599	}
600
601	tunables->rate_limit_us = cpufreq_policy_transition_delay_us(policy);
602
603	policy->governor_data = sg_policy;
604	sg_policy->tunables = tunables;
605
606	ret = kobject_init_and_add(&tunables->attr_set.kobj, &sugov_tunables_ktype,
607				   get_governor_parent_kobj(policy), "%s",
608				   schedutil_gov.name);
609	if (ret)
610		goto fail;
611
612out:
613	mutex_unlock(&global_tunables_lock);
614	return 0;
615
616fail:
 
617	policy->governor_data = NULL;
618	sugov_tunables_free(tunables);
619
620stop_kthread:
621	sugov_kthread_stop(sg_policy);
622	mutex_unlock(&global_tunables_lock);
623
624free_sg_policy:
625	sugov_policy_free(sg_policy);
626
627disable_fast_switch:
628	cpufreq_disable_fast_switch(policy);
629
630	pr_err("initialization failed (error %d)\n", ret);
631	return ret;
632}
633
634static void sugov_exit(struct cpufreq_policy *policy)
635{
636	struct sugov_policy *sg_policy = policy->governor_data;
637	struct sugov_tunables *tunables = sg_policy->tunables;
638	unsigned int count;
639
640	mutex_lock(&global_tunables_lock);
641
642	count = gov_attr_set_put(&tunables->attr_set, &sg_policy->tunables_hook);
643	policy->governor_data = NULL;
644	if (!count)
645		sugov_tunables_free(tunables);
646
647	mutex_unlock(&global_tunables_lock);
648
649	sugov_kthread_stop(sg_policy);
650	sugov_policy_free(sg_policy);
651	cpufreq_disable_fast_switch(policy);
652}
653
654static int sugov_start(struct cpufreq_policy *policy)
655{
656	struct sugov_policy *sg_policy = policy->governor_data;
 
657	unsigned int cpu;
658
659	sg_policy->freq_update_delay_ns	= sg_policy->tunables->rate_limit_us * NSEC_PER_USEC;
660	sg_policy->last_freq_update_time	= 0;
661	sg_policy->next_freq			= UINT_MAX;
662	sg_policy->work_in_progress		= false;
663	sg_policy->need_freq_update		= false;
664	sg_policy->cached_raw_freq		= 0;
665
 
 
666	for_each_cpu(cpu, policy->cpus) {
667		struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
668
669		memset(sg_cpu, 0, sizeof(*sg_cpu));
670		sg_cpu->cpu			= cpu;
671		sg_cpu->sg_policy		= sg_policy;
672		sg_cpu->iowait_boost_max	= policy->cpuinfo.max_freq;
673	}
674
 
 
 
 
 
 
 
675	for_each_cpu(cpu, policy->cpus) {
676		struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
677
678		cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util,
679					     policy_is_shared(policy) ?
680							sugov_update_shared :
681							sugov_update_single);
682	}
683	return 0;
684}
685
686static void sugov_stop(struct cpufreq_policy *policy)
687{
688	struct sugov_policy *sg_policy = policy->governor_data;
689	unsigned int cpu;
690
691	for_each_cpu(cpu, policy->cpus)
692		cpufreq_remove_update_util_hook(cpu);
693
694	synchronize_sched();
695
696	if (!policy->fast_switch_enabled) {
697		irq_work_sync(&sg_policy->irq_work);
698		kthread_cancel_work_sync(&sg_policy->work);
699	}
700}
701
702static void sugov_limits(struct cpufreq_policy *policy)
703{
704	struct sugov_policy *sg_policy = policy->governor_data;
705
706	if (!policy->fast_switch_enabled) {
707		mutex_lock(&sg_policy->work_lock);
708		cpufreq_policy_apply_limits(policy);
709		mutex_unlock(&sg_policy->work_lock);
710	}
711
712	sg_policy->need_freq_update = true;
713}
714
715static struct cpufreq_governor schedutil_gov = {
716	.name			= "schedutil",
717	.owner			= THIS_MODULE,
718	.dynamic_switching	= true,
719	.init			= sugov_init,
720	.exit			= sugov_exit,
721	.start			= sugov_start,
722	.stop			= sugov_stop,
723	.limits			= sugov_limits,
724};
725
726#ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL
727struct cpufreq_governor *cpufreq_default_governor(void)
728{
729	return &schedutil_gov;
730}
731#endif
732
733static int __init sugov_register(void)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
734{
735	return cpufreq_register_governor(&schedutil_gov);
 
 
 
 
 
 
 
 
736}
737fs_initcall(sugov_register);