1// SPDX-License-Identifier: GPL-2.0
  2/*
  3 * CPUFreq governor based on scheduler-provided CPU utilization data.
  4 *
  5 * Copyright (C) 2016, Intel Corporation
  6 * Author: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
  7 */
  8
  9#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
 10
 11#include "sched.h"
 12
 13#include <linux/sched/cpufreq.h>
 14#include <trace/events/power.h>
 15
 16#define IOWAIT_BOOST_MIN	(SCHED_CAPACITY_SCALE / 8)
 17
 18struct sugov_tunables {
 19	struct gov_attr_set	attr_set;
 20	unsigned int		rate_limit_us;
 21};
 22
 23struct sugov_policy {
 24	struct cpufreq_policy	*policy;
 25
 26	struct sugov_tunables	*tunables;
 27	struct list_head	tunables_hook;
 28
 29	raw_spinlock_t		update_lock;	/* For shared policies */
 30	u64			last_freq_update_time;
 31	s64			freq_update_delay_ns;
 32	unsigned int		next_freq;
 33	unsigned int		cached_raw_freq;
 34
 35	/* The next fields are only needed if fast switch cannot be used: */
 36	struct			irq_work irq_work;
 37	struct			kthread_work work;
 38	struct			mutex work_lock;
 39	struct			kthread_worker worker;
 40	struct task_struct	*thread;
 41	bool			work_in_progress;
 42
 43	bool			limits_changed;
 44	bool			need_freq_update;
 45};
 46
 47struct sugov_cpu {
 48	struct update_util_data	update_util;
 49	struct sugov_policy	*sg_policy;
 50	unsigned int		cpu;
 51
 52	bool			iowait_boost_pending;
 53	unsigned int		iowait_boost;
 54	u64			last_update;
 55
 56	unsigned long		bw_dl;
 57	unsigned long		max;
 58
 59	/* The field below is for single-CPU policies only: */
 60#ifdef CONFIG_NO_HZ_COMMON
 61	unsigned long		saved_idle_calls;
 62#endif
 63};
 64
 65static DEFINE_PER_CPU(struct sugov_cpu, sugov_cpu);
 66
 67/************************ Governor internals ***********************/
 68
 69static bool sugov_should_update_freq(struct sugov_policy *sg_policy, u64 time)
 70{
 71	s64 delta_ns;
 72
 73	/*
 74	 * Since cpufreq_update_util() is called with rq->lock held for
 75	 * the @target_cpu, our per-CPU data is fully serialized.
 76	 *
 77	 * However, drivers cannot in general deal with cross-CPU
 78	 * requests, so while get_next_freq() will work, our
 79	 * sugov_update_commit() call may not for the fast switching platforms.
 80	 *
 81	 * Hence stop here for remote requests if they aren't supported
 82	 * by the hardware, as calculating the frequency is pointless if
 83	 * we cannot in fact act on it.
 84	 *
 85	 * For the slow switching platforms, the kthread is always scheduled on
 86	 * the right set of CPUs and any CPU can find the next frequency and
 87	 * schedule the kthread.
 88	 */
 89	if (sg_policy->policy->fast_switch_enabled &&
 90	    !cpufreq_this_cpu_can_update(sg_policy->policy))
 91		return false;
 92
 93	if (unlikely(sg_policy->limits_changed)) {
 94		sg_policy->limits_changed = false;
 95		sg_policy->need_freq_update = true;
 96		return true;
 97	}
 98
 99	delta_ns = time - sg_policy->last_freq_update_time;
100
101	return delta_ns >= sg_policy->freq_update_delay_ns;
102}
103
104static bool sugov_update_next_freq(struct sugov_policy *sg_policy, u64 time,
105				   unsigned int next_freq)
106{
107	if (sg_policy->next_freq == next_freq)
108		return false;
109
110	sg_policy->next_freq = next_freq;
111	sg_policy->last_freq_update_time = time;
112
113	return true;
114}
115
116static void sugov_fast_switch(struct sugov_policy *sg_policy, u64 time,
117			      unsigned int next_freq)
118{
119	struct cpufreq_policy *policy = sg_policy->policy;
120	int cpu;
121
122	if (!sugov_update_next_freq(sg_policy, time, next_freq))
123		return;
124
125	next_freq = cpufreq_driver_fast_switch(policy, next_freq);
126	if (!next_freq)
127		return;
128
129	policy->cur = next_freq;
130
131	if (trace_cpu_frequency_enabled()) {
132		for_each_cpu(cpu, policy->cpus)
133			trace_cpu_frequency(next_freq, cpu);
134	}
135}
136
137static void sugov_deferred_update(struct sugov_policy *sg_policy, u64 time,
138				  unsigned int next_freq)
139{
140	if (!sugov_update_next_freq(sg_policy, time, next_freq))
141		return;
142
143	if (!sg_policy->work_in_progress) {
144		sg_policy->work_in_progress = true;
145		irq_work_queue(&sg_policy->irq_work);
146	}
147}
148
149/**
150 * get_next_freq - Compute a new frequency for a given cpufreq policy.
151 * @sg_policy: schedutil policy object to compute the new frequency for.
152 * @util: Current CPU utilization.
153 * @max: CPU capacity.
154 *
155 * If the utilization is frequency-invariant, choose the new frequency to be
156 * proportional to it, that is
157 *
158 * next_freq = C * max_freq * util / max
159 *
160 * Otherwise, approximate the would-be frequency-invariant utilization by
161 * util_raw * (curr_freq / max_freq) which leads to
162 *
163 * next_freq = C * curr_freq * util_raw / max
164 *
165 * Take C = 1.25 for the frequency tipping point at (util / max) = 0.8.
166 *
167 * The lowest driver-supported frequency which is equal or greater than the raw
168 * next_freq (as calculated above) is returned, subject to policy min/max and
169 * cpufreq driver limitations.
170 */
171static unsigned int get_next_freq(struct sugov_policy *sg_policy,
172				  unsigned long util, unsigned long max)
173{
174	struct cpufreq_policy *policy = sg_policy->policy;
175	unsigned int freq = arch_scale_freq_invariant() ?
176				policy->cpuinfo.max_freq : policy->cur;
177
178	freq = map_util_freq(util, freq, max);
179
180	if (freq == sg_policy->cached_raw_freq && !sg_policy->need_freq_update)
181		return sg_policy->next_freq;
182
183	sg_policy->need_freq_update = false;
184	sg_policy->cached_raw_freq = freq;
185	return cpufreq_driver_resolve_freq(policy, freq);
186}
187
188/*
189 * This function computes an effective utilization for the given CPU, to be
190 * used for frequency selection given the linear relation: f = u * f_max.
191 *
192 * The scheduler tracks the following metrics:
193 *
194 *   cpu_util_{cfs,rt,dl,irq}()
195 *   cpu_bw_dl()
196 *
197 * Where the cfs,rt and dl util numbers are tracked with the same metric and
198 * synchronized windows and are thus directly comparable.
199 *
200 * The cfs,rt,dl utilization are the running times measured with rq->clock_task
201 * which excludes things like IRQ and steal-time. These latter are then accrued
202 * in the irq utilization.
203 *
204 * The DL bandwidth number otoh is not a measured metric but a value computed
205 * based on the task model parameters and gives the minimal utilization
206 * required to meet deadlines.
207 */
208unsigned long schedutil_cpu_util(int cpu, unsigned long util_cfs,
209				 unsigned long max, enum schedutil_type type,
210				 struct task_struct *p)
211{
212	unsigned long dl_util, util, irq;
213	struct rq *rq = cpu_rq(cpu);
214
215	if (!IS_BUILTIN(CONFIG_UCLAMP_TASK) &&
216	    type == FREQUENCY_UTIL && rt_rq_is_runnable(&rq->rt)) {
217		return max;
218	}
219
220	/*
221	 * Early check to see if IRQ/steal time saturates the CPU, can be
222	 * because of inaccuracies in how we track these -- see
223	 * update_irq_load_avg().
224	 */
225	irq = cpu_util_irq(rq);
226	if (unlikely(irq >= max))
227		return max;
228
229	/*
230	 * Because the time spend on RT/DL tasks is visible as 'lost' time to
231	 * CFS tasks and we use the same metric to track the effective
232	 * utilization (PELT windows are synchronized) we can directly add them
233	 * to obtain the CPU's actual utilization.
234	 *
235	 * CFS and RT utilization can be boosted or capped, depending on
236	 * utilization clamp constraints requested by currently RUNNABLE
237	 * tasks.
238	 * When there are no CFS RUNNABLE tasks, clamps are released and
239	 * frequency will be gracefully reduced with the utilization decay.
240	 */
241	util = util_cfs + cpu_util_rt(rq);
242	if (type == FREQUENCY_UTIL)
243		util = uclamp_util_with(rq, util, p);
244
245	dl_util = cpu_util_dl(rq);
246
247	/*
248	 * For frequency selection we do not make cpu_util_dl() a permanent part
249	 * of this sum because we want to use cpu_bw_dl() later on, but we need
250	 * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
251	 * that we select f_max when there is no idle time.
252	 *
253	 * NOTE: numerical errors or stop class might cause us to not quite hit
254	 * saturation when we should -- something for later.
255	 */
256	if (util + dl_util >= max)
257		return max;
258
259	/*
260	 * OTOH, for energy computation we need the estimated running time, so
261	 * include util_dl and ignore dl_bw.
262	 */
263	if (type == ENERGY_UTIL)
264		util += dl_util;
265
266	/*
267	 * There is still idle time; further improve the number by using the
268	 * irq metric. Because IRQ/steal time is hidden from the task clock we
269	 * need to scale the task numbers:
270	 *
271	 *              max - irq
272	 *   U' = irq + --------- * U
273	 *                 max
274	 */
275	util = scale_irq_capacity(util, irq, max);
276	util += irq;
277
278	/*
279	 * Bandwidth required by DEADLINE must always be granted while, for
280	 * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
281	 * to gracefully reduce the frequency when no tasks show up for longer
282	 * periods of time.
283	 *
284	 * Ideally we would like to set bw_dl as min/guaranteed freq and util +
285	 * bw_dl as requested freq. However, cpufreq is not yet ready for such
286	 * an interface. So, we only do the latter for now.
287	 */
288	if (type == FREQUENCY_UTIL)
289		util += cpu_bw_dl(rq);
290
291	return min(max, util);
292}
293
294static unsigned long sugov_get_util(struct sugov_cpu *sg_cpu)
295{
296	struct rq *rq = cpu_rq(sg_cpu->cpu);
297	unsigned long util = cpu_util_cfs(rq);
298	unsigned long max = arch_scale_cpu_capacity(sg_cpu->cpu);
299
300	sg_cpu->max = max;
301	sg_cpu->bw_dl = cpu_bw_dl(rq);
302
303	return schedutil_cpu_util(sg_cpu->cpu, util, max, FREQUENCY_UTIL, NULL);
304}
305
306/**
307 * sugov_iowait_reset() - Reset the IO boost status of a CPU.
308 * @sg_cpu: the sugov data for the CPU to boost
309 * @time: the update time from the caller
310 * @set_iowait_boost: true if an IO boost has been requested
311 *
312 * The IO wait boost of a task is disabled after a tick since the last update
313 * of a CPU. If a new IO wait boost is requested after more then a tick, then
314 * we enable the boost starting from IOWAIT_BOOST_MIN, which improves energy
315 * efficiency by ignoring sporadic wakeups from IO.
316 */
317static bool sugov_iowait_reset(struct sugov_cpu *sg_cpu, u64 time,
318			       bool set_iowait_boost)
319{
320	s64 delta_ns = time - sg_cpu->last_update;
321
322	/* Reset boost only if a tick has elapsed since last request */
323	if (delta_ns <= TICK_NSEC)
324		return false;
325
326	sg_cpu->iowait_boost = set_iowait_boost ? IOWAIT_BOOST_MIN : 0;
327	sg_cpu->iowait_boost_pending = set_iowait_boost;
328
329	return true;
330}
331
332/**
333 * sugov_iowait_boost() - Updates the IO boost status of a CPU.
334 * @sg_cpu: the sugov data for the CPU to boost
335 * @time: the update time from the caller
336 * @flags: SCHED_CPUFREQ_IOWAIT if the task is waking up after an IO wait
337 *
338 * Each time a task wakes up after an IO operation, the CPU utilization can be
339 * boosted to a certain utilization which doubles at each "frequent and
340 * successive" wakeup from IO, ranging from IOWAIT_BOOST_MIN to the utilization
341 * of the maximum OPP.
342 *
343 * To keep doubling, an IO boost has to be requested at least once per tick,
344 * otherwise we restart from the utilization of the minimum OPP.
345 */
346static void sugov_iowait_boost(struct sugov_cpu *sg_cpu, u64 time,
347			       unsigned int flags)
348{
349	bool set_iowait_boost = flags & SCHED_CPUFREQ_IOWAIT;
350
351	/* Reset boost if the CPU appears to have been idle enough */
352	if (sg_cpu->iowait_boost &&
353	    sugov_iowait_reset(sg_cpu, time, set_iowait_boost))
354		return;
355
356	/* Boost only tasks waking up after IO */
357	if (!set_iowait_boost)
358		return;
359
360	/* Ensure boost doubles only one time at each request */
361	if (sg_cpu->iowait_boost_pending)
362		return;
363	sg_cpu->iowait_boost_pending = true;
364
365	/* Double the boost at each request */
366	if (sg_cpu->iowait_boost) {
367		sg_cpu->iowait_boost =
368			min_t(unsigned int, sg_cpu->iowait_boost << 1, SCHED_CAPACITY_SCALE);
369		return;
370	}
371
372	/* First wakeup after IO: start with minimum boost */
373	sg_cpu->iowait_boost = IOWAIT_BOOST_MIN;
374}
375
376/**
377 * sugov_iowait_apply() - Apply the IO boost to a CPU.
378 * @sg_cpu: the sugov data for the cpu to boost
379 * @time: the update time from the caller
380 * @util: the utilization to (eventually) boost
381 * @max: the maximum value the utilization can be boosted to
382 *
383 * A CPU running a task which woken up after an IO operation can have its
384 * utilization boosted to speed up the completion of those IO operations.
385 * The IO boost value is increased each time a task wakes up from IO, in
386 * sugov_iowait_apply(), and it's instead decreased by this function,
387 * each time an increase has not been requested (!iowait_boost_pending).
388 *
389 * A CPU which also appears to have been idle for at least one tick has also
390 * its IO boost utilization reset.
391 *
392 * This mechanism is designed to boost high frequently IO waiting tasks, while
393 * being more conservative on tasks which does sporadic IO operations.
394 */
395static unsigned long sugov_iowait_apply(struct sugov_cpu *sg_cpu, u64 time,
396					unsigned long util, unsigned long max)
397{
398	unsigned long boost;
399
400	/* No boost currently required */
401	if (!sg_cpu->iowait_boost)
402		return util;
403
404	/* Reset boost if the CPU appears to have been idle enough */
405	if (sugov_iowait_reset(sg_cpu, time, false))
406		return util;
407
408	if (!sg_cpu->iowait_boost_pending) {
409		/*
410		 * No boost pending; reduce the boost value.
411		 */
412		sg_cpu->iowait_boost >>= 1;
413		if (sg_cpu->iowait_boost < IOWAIT_BOOST_MIN) {
414			sg_cpu->iowait_boost = 0;
415			return util;
416		}
417	}
418
419	sg_cpu->iowait_boost_pending = false;
420
421	/*
422	 * @util is already in capacity scale; convert iowait_boost
423	 * into the same scale so we can compare.
424	 */
425	boost = (sg_cpu->iowait_boost * max) >> SCHED_CAPACITY_SHIFT;
426	return max(boost, util);
427}
428
429#ifdef CONFIG_NO_HZ_COMMON
430static bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu)
431{
432	unsigned long idle_calls = tick_nohz_get_idle_calls_cpu(sg_cpu->cpu);
433	bool ret = idle_calls == sg_cpu->saved_idle_calls;
434
435	sg_cpu->saved_idle_calls = idle_calls;
436	return ret;
437}
438#else
439static inline bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu) { return false; }
440#endif /* CONFIG_NO_HZ_COMMON */
441
442/*
443 * Make sugov_should_update_freq() ignore the rate limit when DL
444 * has increased the utilization.
445 */
446static inline void ignore_dl_rate_limit(struct sugov_cpu *sg_cpu, struct sugov_policy *sg_policy)
447{
448	if (cpu_bw_dl(cpu_rq(sg_cpu->cpu)) > sg_cpu->bw_dl)
449		sg_policy->limits_changed = true;
450}
451
452static void sugov_update_single(struct update_util_data *hook, u64 time,
453				unsigned int flags)
454{
455	struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
456	struct sugov_policy *sg_policy = sg_cpu->sg_policy;
457	unsigned long util, max;
458	unsigned int next_f;
459	bool busy;
460
461	sugov_iowait_boost(sg_cpu, time, flags);
462	sg_cpu->last_update = time;
463
464	ignore_dl_rate_limit(sg_cpu, sg_policy);
465
466	if (!sugov_should_update_freq(sg_policy, time))
467		return;
468
469	/* Limits may have changed, don't skip frequency update */
470	busy = !sg_policy->need_freq_update && sugov_cpu_is_busy(sg_cpu);
471
472	util = sugov_get_util(sg_cpu);
473	max = sg_cpu->max;
474	util = sugov_iowait_apply(sg_cpu, time, util, max);
475	next_f = get_next_freq(sg_policy, util, max);
476	/*
477	 * Do not reduce the frequency if the CPU has not been idle
478	 * recently, as the reduction is likely to be premature then.
479	 */
480	if (busy && next_f < sg_policy->next_freq) {
481		next_f = sg_policy->next_freq;
482
483		/* Reset cached freq as next_freq has changed */
484		sg_policy->cached_raw_freq = 0;
485	}
486
487	/*
488	 * This code runs under rq->lock for the target CPU, so it won't run
489	 * concurrently on two different CPUs for the same target and it is not
490	 * necessary to acquire the lock in the fast switch case.
491	 */
492	if (sg_policy->policy->fast_switch_enabled) {
493		sugov_fast_switch(sg_policy, time, next_f);
494	} else {
495		raw_spin_lock(&sg_policy->update_lock);
496		sugov_deferred_update(sg_policy, time, next_f);
497		raw_spin_unlock(&sg_policy->update_lock);
498	}
499}
500
501static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu, u64 time)
502{
503	struct sugov_policy *sg_policy = sg_cpu->sg_policy;
504	struct cpufreq_policy *policy = sg_policy->policy;
505	unsigned long util = 0, max = 1;
506	unsigned int j;
507
508	for_each_cpu(j, policy->cpus) {
509		struct sugov_cpu *j_sg_cpu = &per_cpu(sugov_cpu, j);
510		unsigned long j_util, j_max;
511
512		j_util = sugov_get_util(j_sg_cpu);
513		j_max = j_sg_cpu->max;
514		j_util = sugov_iowait_apply(j_sg_cpu, time, j_util, j_max);
515
516		if (j_util * max > j_max * util) {
517			util = j_util;
518			max = j_max;
519		}
520	}
521
522	return get_next_freq(sg_policy, util, max);
523}
524
525static void
526sugov_update_shared(struct update_util_data *hook, u64 time, unsigned int flags)
527{
528	struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
529	struct sugov_policy *sg_policy = sg_cpu->sg_policy;
530	unsigned int next_f;
531
532	raw_spin_lock(&sg_policy->update_lock);
533
534	sugov_iowait_boost(sg_cpu, time, flags);
535	sg_cpu->last_update = time;
536
537	ignore_dl_rate_limit(sg_cpu, sg_policy);
538
539	if (sugov_should_update_freq(sg_policy, time)) {
540		next_f = sugov_next_freq_shared(sg_cpu, time);
541
542		if (sg_policy->policy->fast_switch_enabled)
543			sugov_fast_switch(sg_policy, time, next_f);
544		else
545			sugov_deferred_update(sg_policy, time, next_f);
546	}
547
548	raw_spin_unlock(&sg_policy->update_lock);
549}
550
551static void sugov_work(struct kthread_work *work)
552{
553	struct sugov_policy *sg_policy = container_of(work, struct sugov_policy, work);
554	unsigned int freq;
555	unsigned long flags;
556
557	/*
558	 * Hold sg_policy->update_lock shortly to handle the case where:
559	 * incase sg_policy->next_freq is read here, and then updated by
560	 * sugov_deferred_update() just before work_in_progress is set to false
561	 * here, we may miss queueing the new update.
562	 *
563	 * Note: If a work was queued after the update_lock is released,
564	 * sugov_work() will just be called again by kthread_work code; and the
565	 * request will be proceed before the sugov thread sleeps.
566	 */
567	raw_spin_lock_irqsave(&sg_policy->update_lock, flags);
568	freq = sg_policy->next_freq;
569	sg_policy->work_in_progress = false;
570	raw_spin_unlock_irqrestore(&sg_policy->update_lock, flags);
571
572	mutex_lock(&sg_policy->work_lock);
573	__cpufreq_driver_target(sg_policy->policy, freq, CPUFREQ_RELATION_L);
574	mutex_unlock(&sg_policy->work_lock);
575}
576
577static void sugov_irq_work(struct irq_work *irq_work)
578{
579	struct sugov_policy *sg_policy;
580
581	sg_policy = container_of(irq_work, struct sugov_policy, irq_work);
582
583	kthread_queue_work(&sg_policy->worker, &sg_policy->work);
584}
585
586/************************** sysfs interface ************************/
587
588static struct sugov_tunables *global_tunables;
589static DEFINE_MUTEX(global_tunables_lock);
590
591static inline struct sugov_tunables *to_sugov_tunables(struct gov_attr_set *attr_set)
592{
593	return container_of(attr_set, struct sugov_tunables, attr_set);
594}
595
596static ssize_t rate_limit_us_show(struct gov_attr_set *attr_set, char *buf)
597{
598	struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
599
600	return sprintf(buf, "%u\n", tunables->rate_limit_us);
601}
602
603static ssize_t
604rate_limit_us_store(struct gov_attr_set *attr_set, const char *buf, size_t count)
605{
606	struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
607	struct sugov_policy *sg_policy;
608	unsigned int rate_limit_us;
609
610	if (kstrtouint(buf, 10, &rate_limit_us))
611		return -EINVAL;
612
613	tunables->rate_limit_us = rate_limit_us;
614
615	list_for_each_entry(sg_policy, &attr_set->policy_list, tunables_hook)
616		sg_policy->freq_update_delay_ns = rate_limit_us * NSEC_PER_USEC;
617
618	return count;
619}
620
621static struct governor_attr rate_limit_us = __ATTR_RW(rate_limit_us);
622
623static struct attribute *sugov_attrs[] = {
624	&rate_limit_us.attr,
625	NULL
626};
627ATTRIBUTE_GROUPS(sugov);
628
629static struct kobj_type sugov_tunables_ktype = {
630	.default_groups = sugov_groups,
631	.sysfs_ops = &governor_sysfs_ops,
632};
633
634/********************** cpufreq governor interface *********************/
635
636struct cpufreq_governor schedutil_gov;
637
638static struct sugov_policy *sugov_policy_alloc(struct cpufreq_policy *policy)
639{
640	struct sugov_policy *sg_policy;
641
642	sg_policy = kzalloc(sizeof(*sg_policy), GFP_KERNEL);
643	if (!sg_policy)
644		return NULL;
645
646	sg_policy->policy = policy;
647	raw_spin_lock_init(&sg_policy->update_lock);
648	return sg_policy;
649}
650
651static void sugov_policy_free(struct sugov_policy *sg_policy)
652{
653	kfree(sg_policy);
654}
655
656static int sugov_kthread_create(struct sugov_policy *sg_policy)
657{
658	struct task_struct *thread;
659	struct sched_attr attr = {
660		.size		= sizeof(struct sched_attr),
661		.sched_policy	= SCHED_DEADLINE,
662		.sched_flags	= SCHED_FLAG_SUGOV,
663		.sched_nice	= 0,
664		.sched_priority	= 0,
665		/*
666		 * Fake (unused) bandwidth; workaround to "fix"
667		 * priority inheritance.
668		 */
669		.sched_runtime	=  1000000,
670		.sched_deadline = 10000000,
671		.sched_period	= 10000000,
672	};
673	struct cpufreq_policy *policy = sg_policy->policy;
674	int ret;
675
676	/* kthread only required for slow path */
677	if (policy->fast_switch_enabled)
678		return 0;
679
680	kthread_init_work(&sg_policy->work, sugov_work);
681	kthread_init_worker(&sg_policy->worker);
682	thread = kthread_create(kthread_worker_fn, &sg_policy->worker,
683				"sugov:%d",
684				cpumask_first(policy->related_cpus));
685	if (IS_ERR(thread)) {
686		pr_err("failed to create sugov thread: %ld\n", PTR_ERR(thread));
687		return PTR_ERR(thread);
688	}
689
690	ret = sched_setattr_nocheck(thread, &attr);
691	if (ret) {
692		kthread_stop(thread);
693		pr_warn("%s: failed to set SCHED_DEADLINE\n", __func__);
694		return ret;
695	}
696
697	sg_policy->thread = thread;
698	kthread_bind_mask(thread, policy->related_cpus);
699	init_irq_work(&sg_policy->irq_work, sugov_irq_work);
700	mutex_init(&sg_policy->work_lock);
701
702	wake_up_process(thread);
703
704	return 0;
705}
706
707static void sugov_kthread_stop(struct sugov_policy *sg_policy)
708{
709	/* kthread only required for slow path */
710	if (sg_policy->policy->fast_switch_enabled)
711		return;
712
713	kthread_flush_worker(&sg_policy->worker);
714	kthread_stop(sg_policy->thread);
715	mutex_destroy(&sg_policy->work_lock);
716}
717
718static struct sugov_tunables *sugov_tunables_alloc(struct sugov_policy *sg_policy)
719{
720	struct sugov_tunables *tunables;
721
722	tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
723	if (tunables) {
724		gov_attr_set_init(&tunables->attr_set, &sg_policy->tunables_hook);
725		if (!have_governor_per_policy())
726			global_tunables = tunables;
727	}
728	return tunables;
729}
730
731static void sugov_tunables_free(struct sugov_tunables *tunables)
732{
733	if (!have_governor_per_policy())
734		global_tunables = NULL;
735
736	kfree(tunables);
737}
738
739static int sugov_init(struct cpufreq_policy *policy)
740{
741	struct sugov_policy *sg_policy;
742	struct sugov_tunables *tunables;
743	int ret = 0;
744
745	/* State should be equivalent to EXIT */
746	if (policy->governor_data)
747		return -EBUSY;
748
749	cpufreq_enable_fast_switch(policy);
750
751	sg_policy = sugov_policy_alloc(policy);
752	if (!sg_policy) {
753		ret = -ENOMEM;
754		goto disable_fast_switch;
755	}
756
757	ret = sugov_kthread_create(sg_policy);
758	if (ret)
759		goto free_sg_policy;
760
761	mutex_lock(&global_tunables_lock);
762
763	if (global_tunables) {
764		if (WARN_ON(have_governor_per_policy())) {
765			ret = -EINVAL;
766			goto stop_kthread;
767		}
768		policy->governor_data = sg_policy;
769		sg_policy->tunables = global_tunables;
770
771		gov_attr_set_get(&global_tunables->attr_set, &sg_policy->tunables_hook);
772		goto out;
773	}
774
775	tunables = sugov_tunables_alloc(sg_policy);
776	if (!tunables) {
777		ret = -ENOMEM;
778		goto stop_kthread;
779	}
780
781	tunables->rate_limit_us = cpufreq_policy_transition_delay_us(policy);
782
783	policy->governor_data = sg_policy;
784	sg_policy->tunables = tunables;
785
786	ret = kobject_init_and_add(&tunables->attr_set.kobj, &sugov_tunables_ktype,
787				   get_governor_parent_kobj(policy), "%s",
788				   schedutil_gov.name);
789	if (ret)
790		goto fail;
791
792out:
793	mutex_unlock(&global_tunables_lock);
794	return 0;
795
796fail:
797	kobject_put(&tunables->attr_set.kobj);
798	policy->governor_data = NULL;
799	sugov_tunables_free(tunables);
800
801stop_kthread:
802	sugov_kthread_stop(sg_policy);
803	mutex_unlock(&global_tunables_lock);
804
805free_sg_policy:
806	sugov_policy_free(sg_policy);
807
808disable_fast_switch:
809	cpufreq_disable_fast_switch(policy);
810
811	pr_err("initialization failed (error %d)\n", ret);
812	return ret;
813}
814
815static void sugov_exit(struct cpufreq_policy *policy)
816{
817	struct sugov_policy *sg_policy = policy->governor_data;
818	struct sugov_tunables *tunables = sg_policy->tunables;
819	unsigned int count;
820
821	mutex_lock(&global_tunables_lock);
822
823	count = gov_attr_set_put(&tunables->attr_set, &sg_policy->tunables_hook);
824	policy->governor_data = NULL;
825	if (!count)
826		sugov_tunables_free(tunables);
827
828	mutex_unlock(&global_tunables_lock);
829
830	sugov_kthread_stop(sg_policy);
831	sugov_policy_free(sg_policy);
832	cpufreq_disable_fast_switch(policy);
833}
834
835static int sugov_start(struct cpufreq_policy *policy)
836{
837	struct sugov_policy *sg_policy = policy->governor_data;
838	unsigned int cpu;
839
840	sg_policy->freq_update_delay_ns	= sg_policy->tunables->rate_limit_us * NSEC_PER_USEC;
841	sg_policy->last_freq_update_time	= 0;
842	sg_policy->next_freq			= 0;
843	sg_policy->work_in_progress		= false;
844	sg_policy->limits_changed		= false;
845	sg_policy->need_freq_update		= false;
846	sg_policy->cached_raw_freq		= 0;
847
848	for_each_cpu(cpu, policy->cpus) {
849		struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
850
851		memset(sg_cpu, 0, sizeof(*sg_cpu));
852		sg_cpu->cpu			= cpu;
853		sg_cpu->sg_policy		= sg_policy;
854	}
855
856	for_each_cpu(cpu, policy->cpus) {
857		struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
858
859		cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util,
860					     policy_is_shared(policy) ?
861							sugov_update_shared :
862							sugov_update_single);
863	}
864	return 0;
865}
866
867static void sugov_stop(struct cpufreq_policy *policy)
868{
869	struct sugov_policy *sg_policy = policy->governor_data;
870	unsigned int cpu;
871
872	for_each_cpu(cpu, policy->cpus)
873		cpufreq_remove_update_util_hook(cpu);
874
875	synchronize_rcu();
876
877	if (!policy->fast_switch_enabled) {
878		irq_work_sync(&sg_policy->irq_work);
879		kthread_cancel_work_sync(&sg_policy->work);
880	}
881}
882
883static void sugov_limits(struct cpufreq_policy *policy)
884{
885	struct sugov_policy *sg_policy = policy->governor_data;
886
887	if (!policy->fast_switch_enabled) {
888		mutex_lock(&sg_policy->work_lock);
889		cpufreq_policy_apply_limits(policy);
890		mutex_unlock(&sg_policy->work_lock);
891	}
892
893	sg_policy->limits_changed = true;
894}
895
896struct cpufreq_governor schedutil_gov = {
897	.name			= "schedutil",
898	.owner			= THIS_MODULE,
899	.dynamic_switching	= true,
900	.init			= sugov_init,
901	.exit			= sugov_exit,
902	.start			= sugov_start,
903	.stop			= sugov_stop,
904	.limits			= sugov_limits,
905};
906
907#ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL
908struct cpufreq_governor *cpufreq_default_governor(void)
909{
910	return &schedutil_gov;
911}
912#endif
913
914static int __init sugov_register(void)
915{
916	return cpufreq_register_governor(&schedutil_gov);
917}
918fs_initcall(sugov_register);
919
920#ifdef CONFIG_ENERGY_MODEL
921extern bool sched_energy_update;
922extern struct mutex sched_energy_mutex;
923
924static void rebuild_sd_workfn(struct work_struct *work)
925{
926	mutex_lock(&sched_energy_mutex);
927	sched_energy_update = true;
928	rebuild_sched_domains();
929	sched_energy_update = false;
930	mutex_unlock(&sched_energy_mutex);
931}
932static DECLARE_WORK(rebuild_sd_work, rebuild_sd_workfn);
933
934/*
935 * EAS shouldn't be attempted without sugov, so rebuild the sched_domains
936 * on governor changes to make sure the scheduler knows about it.
937 */
938void sched_cpufreq_governor_change(struct cpufreq_policy *policy,
939				  struct cpufreq_governor *old_gov)
940{
941	if (old_gov == &schedutil_gov || policy->governor == &schedutil_gov) {
942		/*
943		 * When called from the cpufreq_register_driver() path, the
944		 * cpu_hotplug_lock is already held, so use a work item to
945		 * avoid nested locking in rebuild_sched_domains().
946		 */
947		schedule_work(&rebuild_sd_work);
948	}
949
950}
951#endif