1/*
2 *  drivers/cpufreq/cpufreq_conservative.c
3 *
4 *  Copyright (C)  2001 Russell King
5 *            (C)  2003 Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>.
6 *                      Jun Nakajima <jun.nakajima@intel.com>
7 *            (C)  2009 Alexander Clouter <alex@digriz.org.uk>
8 *
9 * This program is free software; you can redistribute it and/or modify
10 * it under the terms of the GNU General Public License version 2 as
11 * published by the Free Software Foundation.
12 */
13
14#include <linux/kernel.h>
15#include <linux/module.h>
16#include <linux/init.h>
17#include <linux/cpufreq.h>
18#include <linux/cpu.h>
19#include <linux/jiffies.h>
20#include <linux/kernel_stat.h>
21#include <linux/mutex.h>
22#include <linux/hrtimer.h>
23#include <linux/tick.h>
24#include <linux/ktime.h>
25#include <linux/sched.h>
26
27/*
28 * dbs is used in this file as a shortform for demandbased switching
29 * It helps to keep variable names smaller, simpler
30 */
31
32#define DEF_FREQUENCY_UP_THRESHOLD		(80)
33#define DEF_FREQUENCY_DOWN_THRESHOLD		(20)
34
35/*
36 * The polling frequency of this governor depends on the capability of
37 * the processor. Default polling frequency is 1000 times the transition
38 * latency of the processor. The governor will work on any processor with
39 * transition latency <= 10mS, using appropriate sampling
40 * rate.
41 * For CPUs with transition latency > 10mS (mostly drivers with CPUFREQ_ETERNAL)
42 * this governor will not work.
43 * All times here are in uS.
44 */
45#define MIN_SAMPLING_RATE_RATIO			(2)
46
47static unsigned int min_sampling_rate;
48
49#define LATENCY_MULTIPLIER			(1000)
50#define MIN_LATENCY_MULTIPLIER			(100)
51#define DEF_SAMPLING_DOWN_FACTOR		(1)
52#define MAX_SAMPLING_DOWN_FACTOR		(10)
53#define TRANSITION_LATENCY_LIMIT		(10 * 1000 * 1000)
54
55static void do_dbs_timer(struct work_struct *work);
56
57struct cpu_dbs_info_s {
58	cputime64_t prev_cpu_idle;
59	cputime64_t prev_cpu_wall;
60	cputime64_t prev_cpu_nice;
61	struct cpufreq_policy *cur_policy;
62	struct delayed_work work;
63	unsigned int down_skip;
64	unsigned int requested_freq;
65	int cpu;
66	unsigned int enable:1;
67	/*
68	 * percpu mutex that serializes governor limit change with
69	 * do_dbs_timer invocation. We do not want do_dbs_timer to run
70	 * when user is changing the governor or limits.
71	 */
72	struct mutex timer_mutex;
73};
74static DEFINE_PER_CPU(struct cpu_dbs_info_s, cs_cpu_dbs_info);
75
76static unsigned int dbs_enable;	/* number of CPUs using this policy */
77
78/*
79 * dbs_mutex protects dbs_enable in governor start/stop.
80 */
81static DEFINE_MUTEX(dbs_mutex);
82
83static struct dbs_tuners {
84	unsigned int sampling_rate;
85	unsigned int sampling_down_factor;
86	unsigned int up_threshold;
87	unsigned int down_threshold;
88	unsigned int ignore_nice;
89	unsigned int freq_step;
90} dbs_tuners_ins = {
91	.up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
92	.down_threshold = DEF_FREQUENCY_DOWN_THRESHOLD,
93	.sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR,
94	.ignore_nice = 0,
95	.freq_step = 5,
96};
97
98static inline u64 get_cpu_idle_time_jiffy(unsigned int cpu, u64 *wall)
99{
100	u64 idle_time;
101	u64 cur_wall_time;
102	u64 busy_time;
103
104	cur_wall_time = jiffies64_to_cputime64(get_jiffies_64());
105
106	busy_time  = kcpustat_cpu(cpu).cpustat[CPUTIME_USER];
107	busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_SYSTEM];
108	busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_IRQ];
109	busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_SOFTIRQ];
110	busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_STEAL];
111	busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_NICE];
112
113	idle_time = cur_wall_time - busy_time;
114	if (wall)
115		*wall = jiffies_to_usecs(cur_wall_time);
116
117	return jiffies_to_usecs(idle_time);
118}
119
120static inline cputime64_t get_cpu_idle_time(unsigned int cpu, cputime64_t *wall)
121{
122	u64 idle_time = get_cpu_idle_time_us(cpu, NULL);
123
124	if (idle_time == -1ULL)
125		return get_cpu_idle_time_jiffy(cpu, wall);
126	else
127		idle_time += get_cpu_iowait_time_us(cpu, wall);
128
129	return idle_time;
130}
131
132/* keep track of frequency transitions */
133static int
134dbs_cpufreq_notifier(struct notifier_block *nb, unsigned long val,
135		     void *data)
136{
137	struct cpufreq_freqs *freq = data;
138	struct cpu_dbs_info_s *this_dbs_info = &per_cpu(cs_cpu_dbs_info,
139							freq->cpu);
140
141	struct cpufreq_policy *policy;
142
143	if (!this_dbs_info->enable)
144		return 0;
145
146	policy = this_dbs_info->cur_policy;
147
148	/*
149	 * we only care if our internally tracked freq moves outside
150	 * the 'valid' ranges of freqency available to us otherwise
151	 * we do not change it
152	*/
153	if (this_dbs_info->requested_freq > policy->max
154			|| this_dbs_info->requested_freq < policy->min)
155		this_dbs_info->requested_freq = freq->new;
156
157	return 0;
158}
159
160static struct notifier_block dbs_cpufreq_notifier_block = {
161	.notifier_call = dbs_cpufreq_notifier
162};
163
164/************************** sysfs interface ************************/
165static ssize_t show_sampling_rate_min(struct kobject *kobj,
166				      struct attribute *attr, char *buf)
167{
168	return sprintf(buf, "%u\n", min_sampling_rate);
169}
170
171define_one_global_ro(sampling_rate_min);
172
173/* cpufreq_conservative Governor Tunables */
174#define show_one(file_name, object)					\
175static ssize_t show_##file_name						\
176(struct kobject *kobj, struct attribute *attr, char *buf)		\
177{									\
178	return sprintf(buf, "%u\n", dbs_tuners_ins.object);		\
179}
180show_one(sampling_rate, sampling_rate);
181show_one(sampling_down_factor, sampling_down_factor);
182show_one(up_threshold, up_threshold);
183show_one(down_threshold, down_threshold);
184show_one(ignore_nice_load, ignore_nice);
185show_one(freq_step, freq_step);
186
187static ssize_t store_sampling_down_factor(struct kobject *a,
188					  struct attribute *b,
189					  const char *buf, size_t count)
190{
191	unsigned int input;
192	int ret;
193	ret = sscanf(buf, "%u", &input);
194
195	if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
196		return -EINVAL;
197
198	dbs_tuners_ins.sampling_down_factor = input;
199	return count;
200}
201
202static ssize_t store_sampling_rate(struct kobject *a, struct attribute *b,
203				   const char *buf, size_t count)
204{
205	unsigned int input;
206	int ret;
207	ret = sscanf(buf, "%u", &input);
208
209	if (ret != 1)
210		return -EINVAL;
211
212	dbs_tuners_ins.sampling_rate = max(input, min_sampling_rate);
213	return count;
214}
215
216static ssize_t store_up_threshold(struct kobject *a, struct attribute *b,
217				  const char *buf, size_t count)
218{
219	unsigned int input;
220	int ret;
221	ret = sscanf(buf, "%u", &input);
222
223	if (ret != 1 || input > 100 ||
224			input <= dbs_tuners_ins.down_threshold)
225		return -EINVAL;
226
227	dbs_tuners_ins.up_threshold = input;
228	return count;
229}
230
231static ssize_t store_down_threshold(struct kobject *a, struct attribute *b,
232				    const char *buf, size_t count)
233{
234	unsigned int input;
235	int ret;
236	ret = sscanf(buf, "%u", &input);
237
238	/* cannot be lower than 11 otherwise freq will not fall */
239	if (ret != 1 || input < 11 || input > 100 ||
240			input >= dbs_tuners_ins.up_threshold)
241		return -EINVAL;
242
243	dbs_tuners_ins.down_threshold = input;
244	return count;
245}
246
247static ssize_t store_ignore_nice_load(struct kobject *a, struct attribute *b,
248				      const char *buf, size_t count)
249{
250	unsigned int input;
251	int ret;
252
253	unsigned int j;
254
255	ret = sscanf(buf, "%u", &input);
256	if (ret != 1)
257		return -EINVAL;
258
259	if (input > 1)
260		input = 1;
261
262	if (input == dbs_tuners_ins.ignore_nice) /* nothing to do */
263		return count;
264
265	dbs_tuners_ins.ignore_nice = input;
266
267	/* we need to re-evaluate prev_cpu_idle */
268	for_each_online_cpu(j) {
269		struct cpu_dbs_info_s *dbs_info;
270		dbs_info = &per_cpu(cs_cpu_dbs_info, j);
271		dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
272						&dbs_info->prev_cpu_wall);
273		if (dbs_tuners_ins.ignore_nice)
274			dbs_info->prev_cpu_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE];
275	}
276	return count;
277}
278
279static ssize_t store_freq_step(struct kobject *a, struct attribute *b,
280			       const char *buf, size_t count)
281{
282	unsigned int input;
283	int ret;
284	ret = sscanf(buf, "%u", &input);
285
286	if (ret != 1)
287		return -EINVAL;
288
289	if (input > 100)
290		input = 100;
291
292	/* no need to test here if freq_step is zero as the user might actually
293	 * want this, they would be crazy though :) */
294	dbs_tuners_ins.freq_step = input;
295	return count;
296}
297
298define_one_global_rw(sampling_rate);
299define_one_global_rw(sampling_down_factor);
300define_one_global_rw(up_threshold);
301define_one_global_rw(down_threshold);
302define_one_global_rw(ignore_nice_load);
303define_one_global_rw(freq_step);
304
305static struct attribute *dbs_attributes[] = {
306	&sampling_rate_min.attr,
307	&sampling_rate.attr,
308	&sampling_down_factor.attr,
309	&up_threshold.attr,
310	&down_threshold.attr,
311	&ignore_nice_load.attr,
312	&freq_step.attr,
313	NULL
314};
315
316static struct attribute_group dbs_attr_group = {
317	.attrs = dbs_attributes,
318	.name = "conservative",
319};
320
321/************************** sysfs end ************************/
322
323static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
324{
325	unsigned int load = 0;
326	unsigned int max_load = 0;
327	unsigned int freq_target;
328
329	struct cpufreq_policy *policy;
330	unsigned int j;
331
332	policy = this_dbs_info->cur_policy;
333
334	/*
335	 * Every sampling_rate, we check, if current idle time is less
336	 * than 20% (default), then we try to increase frequency
337	 * Every sampling_rate*sampling_down_factor, we check, if current
338	 * idle time is more than 80%, then we try to decrease frequency
339	 *
340	 * Any frequency increase takes it to the maximum frequency.
341	 * Frequency reduction happens at minimum steps of
342	 * 5% (default) of maximum frequency
343	 */
344
345	/* Get Absolute Load */
346	for_each_cpu(j, policy->cpus) {
347		struct cpu_dbs_info_s *j_dbs_info;
348		cputime64_t cur_wall_time, cur_idle_time;
349		unsigned int idle_time, wall_time;
350
351		j_dbs_info = &per_cpu(cs_cpu_dbs_info, j);
352
353		cur_idle_time = get_cpu_idle_time(j, &cur_wall_time);
354
355		wall_time = (unsigned int)
356			(cur_wall_time - j_dbs_info->prev_cpu_wall);
357		j_dbs_info->prev_cpu_wall = cur_wall_time;
358
359		idle_time = (unsigned int)
360			(cur_idle_time - j_dbs_info->prev_cpu_idle);
361		j_dbs_info->prev_cpu_idle = cur_idle_time;
362
363		if (dbs_tuners_ins.ignore_nice) {
364			u64 cur_nice;
365			unsigned long cur_nice_jiffies;
366
367			cur_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE] -
368					 j_dbs_info->prev_cpu_nice;
369			/*
370			 * Assumption: nice time between sampling periods will
371			 * be less than 2^32 jiffies for 32 bit sys
372			 */
373			cur_nice_jiffies = (unsigned long)
374					cputime64_to_jiffies64(cur_nice);
375
376			j_dbs_info->prev_cpu_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE];
377			idle_time += jiffies_to_usecs(cur_nice_jiffies);
378		}
379
380		if (unlikely(!wall_time || wall_time < idle_time))
381			continue;
382
383		load = 100 * (wall_time - idle_time) / wall_time;
384
385		if (load > max_load)
386			max_load = load;
387	}
388
389	/*
390	 * break out if we 'cannot' reduce the speed as the user might
391	 * want freq_step to be zero
392	 */
393	if (dbs_tuners_ins.freq_step == 0)
394		return;
395
396	/* Check for frequency increase */
397	if (max_load > dbs_tuners_ins.up_threshold) {
398		this_dbs_info->down_skip = 0;
399
400		/* if we are already at full speed then break out early */
401		if (this_dbs_info->requested_freq == policy->max)
402			return;
403
404		freq_target = (dbs_tuners_ins.freq_step * policy->max) / 100;
405
406		/* max freq cannot be less than 100. But who knows.... */
407		if (unlikely(freq_target == 0))
408			freq_target = 5;
409
410		this_dbs_info->requested_freq += freq_target;
411		if (this_dbs_info->requested_freq > policy->max)
412			this_dbs_info->requested_freq = policy->max;
413
414		__cpufreq_driver_target(policy, this_dbs_info->requested_freq,
415			CPUFREQ_RELATION_H);
416		return;
417	}
418
419	/*
420	 * The optimal frequency is the frequency that is the lowest that
421	 * can support the current CPU usage without triggering the up
422	 * policy. To be safe, we focus 10 points under the threshold.
423	 */
424	if (max_load < (dbs_tuners_ins.down_threshold - 10)) {
425		freq_target = (dbs_tuners_ins.freq_step * policy->max) / 100;
426
427		this_dbs_info->requested_freq -= freq_target;
428		if (this_dbs_info->requested_freq < policy->min)
429			this_dbs_info->requested_freq = policy->min;
430
431		/*
432		 * if we cannot reduce the frequency anymore, break out early
433		 */
434		if (policy->cur == policy->min)
435			return;
436
437		__cpufreq_driver_target(policy, this_dbs_info->requested_freq,
438				CPUFREQ_RELATION_H);
439		return;
440	}
441}
442
443static void do_dbs_timer(struct work_struct *work)
444{
445	struct cpu_dbs_info_s *dbs_info =
446		container_of(work, struct cpu_dbs_info_s, work.work);
447	unsigned int cpu = dbs_info->cpu;
448
449	/* We want all CPUs to do sampling nearly on same jiffy */
450	int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
451
452	delay -= jiffies % delay;
453
454	mutex_lock(&dbs_info->timer_mutex);
455
456	dbs_check_cpu(dbs_info);
457
458	schedule_delayed_work_on(cpu, &dbs_info->work, delay);
459	mutex_unlock(&dbs_info->timer_mutex);
460}
461
462static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
463{
464	/* We want all CPUs to do sampling nearly on same jiffy */
465	int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
466	delay -= jiffies % delay;
467
468	dbs_info->enable = 1;
469	INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer);
470	schedule_delayed_work_on(dbs_info->cpu, &dbs_info->work, delay);
471}
472
473static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
474{
475	dbs_info->enable = 0;
476	cancel_delayed_work_sync(&dbs_info->work);
477}
478
479static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
480				   unsigned int event)
481{
482	unsigned int cpu = policy->cpu;
483	struct cpu_dbs_info_s *this_dbs_info;
484	unsigned int j;
485	int rc;
486
487	this_dbs_info = &per_cpu(cs_cpu_dbs_info, cpu);
488
489	switch (event) {
490	case CPUFREQ_GOV_START:
491		if ((!cpu_online(cpu)) || (!policy->cur))
492			return -EINVAL;
493
494		mutex_lock(&dbs_mutex);
495
496		for_each_cpu(j, policy->cpus) {
497			struct cpu_dbs_info_s *j_dbs_info;
498			j_dbs_info = &per_cpu(cs_cpu_dbs_info, j);
499			j_dbs_info->cur_policy = policy;
500
501			j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
502						&j_dbs_info->prev_cpu_wall);
503			if (dbs_tuners_ins.ignore_nice)
504				j_dbs_info->prev_cpu_nice =
505						kcpustat_cpu(j).cpustat[CPUTIME_NICE];
506		}
507		this_dbs_info->down_skip = 0;
508		this_dbs_info->requested_freq = policy->cur;
509
510		mutex_init(&this_dbs_info->timer_mutex);
511		dbs_enable++;
512		/*
513		 * Start the timerschedule work, when this governor
514		 * is used for first time
515		 */
516		if (dbs_enable == 1) {
517			unsigned int latency;
518			/* policy latency is in nS. Convert it to uS first */
519			latency = policy->cpuinfo.transition_latency / 1000;
520			if (latency == 0)
521				latency = 1;
522
523			rc = sysfs_create_group(cpufreq_global_kobject,
524						&dbs_attr_group);
525			if (rc) {
526				mutex_unlock(&dbs_mutex);
527				return rc;
528			}
529
530			/*
531			 * conservative does not implement micro like ondemand
532			 * governor, thus we are bound to jiffes/HZ
533			 */
534			min_sampling_rate =
535				MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10);
536			/* Bring kernel and HW constraints together */
537			min_sampling_rate = max(min_sampling_rate,
538					MIN_LATENCY_MULTIPLIER * latency);
539			dbs_tuners_ins.sampling_rate =
540				max(min_sampling_rate,
541				    latency * LATENCY_MULTIPLIER);
542
543			cpufreq_register_notifier(
544					&dbs_cpufreq_notifier_block,
545					CPUFREQ_TRANSITION_NOTIFIER);
546		}
547		mutex_unlock(&dbs_mutex);
548
549		dbs_timer_init(this_dbs_info);
550
551		break;
552
553	case CPUFREQ_GOV_STOP:
554		dbs_timer_exit(this_dbs_info);
555
556		mutex_lock(&dbs_mutex);
557		dbs_enable--;
558		mutex_destroy(&this_dbs_info->timer_mutex);
559
560		/*
561		 * Stop the timerschedule work, when this governor
562		 * is used for first time
563		 */
564		if (dbs_enable == 0)
565			cpufreq_unregister_notifier(
566					&dbs_cpufreq_notifier_block,
567					CPUFREQ_TRANSITION_NOTIFIER);
568
569		mutex_unlock(&dbs_mutex);
570		if (!dbs_enable)
571			sysfs_remove_group(cpufreq_global_kobject,
572					   &dbs_attr_group);
573
574		break;
575
576	case CPUFREQ_GOV_LIMITS:
577		mutex_lock(&this_dbs_info->timer_mutex);
578		if (policy->max < this_dbs_info->cur_policy->cur)
579			__cpufreq_driver_target(
580					this_dbs_info->cur_policy,
581					policy->max, CPUFREQ_RELATION_H);
582		else if (policy->min > this_dbs_info->cur_policy->cur)
583			__cpufreq_driver_target(
584					this_dbs_info->cur_policy,
585					policy->min, CPUFREQ_RELATION_L);
586		mutex_unlock(&this_dbs_info->timer_mutex);
587
588		break;
589	}
590	return 0;
591}
592
593#ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_CONSERVATIVE
594static
595#endif
596struct cpufreq_governor cpufreq_gov_conservative = {
597	.name			= "conservative",
598	.governor		= cpufreq_governor_dbs,
599	.max_transition_latency	= TRANSITION_LATENCY_LIMIT,
600	.owner			= THIS_MODULE,
601};
602
603static int __init cpufreq_gov_dbs_init(void)
604{
605	return cpufreq_register_governor(&cpufreq_gov_conservative);
606}
607
608static void __exit cpufreq_gov_dbs_exit(void)
609{
610	cpufreq_unregister_governor(&cpufreq_gov_conservative);
611}
612
613
614MODULE_AUTHOR("Alexander Clouter <alex@digriz.org.uk>");
615MODULE_DESCRIPTION("'cpufreq_conservative' - A dynamic cpufreq governor for "
616		"Low Latency Frequency Transition capable processors "
617		"optimised for use in a battery environment");
618MODULE_LICENSE("GPL");
619
620#ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_CONSERVATIVE
621fs_initcall(cpufreq_gov_dbs_init);
622#else
623module_init(cpufreq_gov_dbs_init);
624#endif
625module_exit(cpufreq_gov_dbs_exit);
626