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 
47 static 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 
55 static void do_dbs_timer(struct work_struct *work);
56 
57 struct 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 };
74 static DEFINE_PER_CPU(struct cpu_dbs_info_s, cs_cpu_dbs_info);
75 
76 static unsigned int dbs_enable;	/* number of CPUs using this policy */
77 
78 /*
79  * dbs_mutex protects dbs_enable in governor start/stop.
80  */
81 static DEFINE_MUTEX(dbs_mutex);
82 
83 static 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 
98 static inline cputime64_t get_cpu_idle_time_jiffy(unsigned int cpu,
99 							cputime64_t *wall)
100 {
101 	cputime64_t idle_time;
102 	cputime64_t cur_wall_time;
103 	cputime64_t busy_time;
104 
105 	cur_wall_time = jiffies64_to_cputime64(get_jiffies_64());
106 	busy_time = cputime64_add(kstat_cpu(cpu).cpustat.user,
107 			kstat_cpu(cpu).cpustat.system);
108 
109 	busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.irq);
110 	busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.softirq);
111 	busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.steal);
112 	busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.nice);
113 
114 	idle_time = cputime64_sub(cur_wall_time, busy_time);
115 	if (wall)
116 		*wall = (cputime64_t)jiffies_to_usecs(cur_wall_time);
117 
118 	return (cputime64_t)jiffies_to_usecs(idle_time);
119 }
120 
121 static inline cputime64_t get_cpu_idle_time(unsigned int cpu, cputime64_t *wall)
122 {
123 	u64 idle_time = get_cpu_idle_time_us(cpu, NULL);
124 
125 	if (idle_time == -1ULL)
126 		return get_cpu_idle_time_jiffy(cpu, wall);
127 	else
128 		idle_time += get_cpu_iowait_time_us(cpu, wall);
129 
130 	return idle_time;
131 }
132 
133 /* keep track of frequency transitions */
134 static int
135 dbs_cpufreq_notifier(struct notifier_block *nb, unsigned long val,
136 		     void *data)
137 {
138 	struct cpufreq_freqs *freq = data;
139 	struct cpu_dbs_info_s *this_dbs_info = &per_cpu(cs_cpu_dbs_info,
140 							freq->cpu);
141 
142 	struct cpufreq_policy *policy;
143 
144 	if (!this_dbs_info->enable)
145 		return 0;
146 
147 	policy = this_dbs_info->cur_policy;
148 
149 	/*
150 	 * we only care if our internally tracked freq moves outside
151 	 * the 'valid' ranges of freqency available to us otherwise
152 	 * we do not change it
153 	*/
154 	if (this_dbs_info->requested_freq > policy->max
155 			|| this_dbs_info->requested_freq < policy->min)
156 		this_dbs_info->requested_freq = freq->new;
157 
158 	return 0;
159 }
160 
161 static struct notifier_block dbs_cpufreq_notifier_block = {
162 	.notifier_call = dbs_cpufreq_notifier
163 };
164 
165 /************************** sysfs interface ************************/
166 static ssize_t show_sampling_rate_min(struct kobject *kobj,
167 				      struct attribute *attr, char *buf)
168 {
169 	return sprintf(buf, "%u\n", min_sampling_rate);
170 }
171 
172 define_one_global_ro(sampling_rate_min);
173 
174 /* cpufreq_conservative Governor Tunables */
175 #define show_one(file_name, object)					\
176 static ssize_t show_##file_name						\
177 (struct kobject *kobj, struct attribute *attr, char *buf)		\
178 {									\
179 	return sprintf(buf, "%u\n", dbs_tuners_ins.object);		\
180 }
181 show_one(sampling_rate, sampling_rate);
182 show_one(sampling_down_factor, sampling_down_factor);
183 show_one(up_threshold, up_threshold);
184 show_one(down_threshold, down_threshold);
185 show_one(ignore_nice_load, ignore_nice);
186 show_one(freq_step, freq_step);
187 
188 static ssize_t store_sampling_down_factor(struct kobject *a,
189 					  struct attribute *b,
190 					  const char *buf, size_t count)
191 {
192 	unsigned int input;
193 	int ret;
194 	ret = sscanf(buf, "%u", &input);
195 
196 	if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
197 		return -EINVAL;
198 
199 	dbs_tuners_ins.sampling_down_factor = input;
200 	return count;
201 }
202 
203 static ssize_t store_sampling_rate(struct kobject *a, struct attribute *b,
204 				   const char *buf, size_t count)
205 {
206 	unsigned int input;
207 	int ret;
208 	ret = sscanf(buf, "%u", &input);
209 
210 	if (ret != 1)
211 		return -EINVAL;
212 
213 	dbs_tuners_ins.sampling_rate = max(input, min_sampling_rate);
214 	return count;
215 }
216 
217 static ssize_t store_up_threshold(struct kobject *a, struct attribute *b,
218 				  const char *buf, size_t count)
219 {
220 	unsigned int input;
221 	int ret;
222 	ret = sscanf(buf, "%u", &input);
223 
224 	if (ret != 1 || input > 100 ||
225 			input <= dbs_tuners_ins.down_threshold)
226 		return -EINVAL;
227 
228 	dbs_tuners_ins.up_threshold = input;
229 	return count;
230 }
231 
232 static ssize_t store_down_threshold(struct kobject *a, struct attribute *b,
233 				    const char *buf, size_t count)
234 {
235 	unsigned int input;
236 	int ret;
237 	ret = sscanf(buf, "%u", &input);
238 
239 	/* cannot be lower than 11 otherwise freq will not fall */
240 	if (ret != 1 || input < 11 || input > 100 ||
241 			input >= dbs_tuners_ins.up_threshold)
242 		return -EINVAL;
243 
244 	dbs_tuners_ins.down_threshold = input;
245 	return count;
246 }
247 
248 static ssize_t store_ignore_nice_load(struct kobject *a, struct attribute *b,
249 				      const char *buf, size_t count)
250 {
251 	unsigned int input;
252 	int ret;
253 
254 	unsigned int j;
255 
256 	ret = sscanf(buf, "%u", &input);
257 	if (ret != 1)
258 		return -EINVAL;
259 
260 	if (input > 1)
261 		input = 1;
262 
263 	if (input == dbs_tuners_ins.ignore_nice) /* nothing to do */
264 		return count;
265 
266 	dbs_tuners_ins.ignore_nice = input;
267 
268 	/* we need to re-evaluate prev_cpu_idle */
269 	for_each_online_cpu(j) {
270 		struct cpu_dbs_info_s *dbs_info;
271 		dbs_info = &per_cpu(cs_cpu_dbs_info, j);
272 		dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
273 						&dbs_info->prev_cpu_wall);
274 		if (dbs_tuners_ins.ignore_nice)
275 			dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
276 	}
277 	return count;
278 }
279 
280 static ssize_t store_freq_step(struct kobject *a, struct attribute *b,
281 			       const char *buf, size_t count)
282 {
283 	unsigned int input;
284 	int ret;
285 	ret = sscanf(buf, "%u", &input);
286 
287 	if (ret != 1)
288 		return -EINVAL;
289 
290 	if (input > 100)
291 		input = 100;
292 
293 	/* no need to test here if freq_step is zero as the user might actually
294 	 * want this, they would be crazy though :) */
295 	dbs_tuners_ins.freq_step = input;
296 	return count;
297 }
298 
299 define_one_global_rw(sampling_rate);
300 define_one_global_rw(sampling_down_factor);
301 define_one_global_rw(up_threshold);
302 define_one_global_rw(down_threshold);
303 define_one_global_rw(ignore_nice_load);
304 define_one_global_rw(freq_step);
305 
306 static struct attribute *dbs_attributes[] = {
307 	&sampling_rate_min.attr,
308 	&sampling_rate.attr,
309 	&sampling_down_factor.attr,
310 	&up_threshold.attr,
311 	&down_threshold.attr,
312 	&ignore_nice_load.attr,
313 	&freq_step.attr,
314 	NULL
315 };
316 
317 static struct attribute_group dbs_attr_group = {
318 	.attrs = dbs_attributes,
319 	.name = "conservative",
320 };
321 
322 /************************** sysfs end ************************/
323 
324 static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
325 {
326 	unsigned int load = 0;
327 	unsigned int max_load = 0;
328 	unsigned int freq_target;
329 
330 	struct cpufreq_policy *policy;
331 	unsigned int j;
332 
333 	policy = this_dbs_info->cur_policy;
334 
335 	/*
336 	 * Every sampling_rate, we check, if current idle time is less
337 	 * than 20% (default), then we try to increase frequency
338 	 * Every sampling_rate*sampling_down_factor, we check, if current
339 	 * idle time is more than 80%, then we try to decrease frequency
340 	 *
341 	 * Any frequency increase takes it to the maximum frequency.
342 	 * Frequency reduction happens at minimum steps of
343 	 * 5% (default) of maximum frequency
344 	 */
345 
346 	/* Get Absolute Load */
347 	for_each_cpu(j, policy->cpus) {
348 		struct cpu_dbs_info_s *j_dbs_info;
349 		cputime64_t cur_wall_time, cur_idle_time;
350 		unsigned int idle_time, wall_time;
351 
352 		j_dbs_info = &per_cpu(cs_cpu_dbs_info, j);
353 
354 		cur_idle_time = get_cpu_idle_time(j, &cur_wall_time);
355 
356 		wall_time = (unsigned int) cputime64_sub(cur_wall_time,
357 				j_dbs_info->prev_cpu_wall);
358 		j_dbs_info->prev_cpu_wall = cur_wall_time;
359 
360 		idle_time = (unsigned int) cputime64_sub(cur_idle_time,
361 				j_dbs_info->prev_cpu_idle);
362 		j_dbs_info->prev_cpu_idle = cur_idle_time;
363 
364 		if (dbs_tuners_ins.ignore_nice) {
365 			cputime64_t cur_nice;
366 			unsigned long cur_nice_jiffies;
367 
368 			cur_nice = cputime64_sub(kstat_cpu(j).cpustat.nice,
369 					 j_dbs_info->prev_cpu_nice);
370 			/*
371 			 * Assumption: nice time between sampling periods will
372 			 * be less than 2^32 jiffies for 32 bit sys
373 			 */
374 			cur_nice_jiffies = (unsigned long)
375 					cputime64_to_jiffies64(cur_nice);
376 
377 			j_dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
378 			idle_time += jiffies_to_usecs(cur_nice_jiffies);
379 		}
380 
381 		if (unlikely(!wall_time || wall_time < idle_time))
382 			continue;
383 
384 		load = 100 * (wall_time - idle_time) / wall_time;
385 
386 		if (load > max_load)
387 			max_load = load;
388 	}
389 
390 	/*
391 	 * break out if we 'cannot' reduce the speed as the user might
392 	 * want freq_step to be zero
393 	 */
394 	if (dbs_tuners_ins.freq_step == 0)
395 		return;
396 
397 	/* Check for frequency increase */
398 	if (max_load > dbs_tuners_ins.up_threshold) {
399 		this_dbs_info->down_skip = 0;
400 
401 		/* if we are already at full speed then break out early */
402 		if (this_dbs_info->requested_freq == policy->max)
403 			return;
404 
405 		freq_target = (dbs_tuners_ins.freq_step * policy->max) / 100;
406 
407 		/* max freq cannot be less than 100. But who knows.... */
408 		if (unlikely(freq_target == 0))
409 			freq_target = 5;
410 
411 		this_dbs_info->requested_freq += freq_target;
412 		if (this_dbs_info->requested_freq > policy->max)
413 			this_dbs_info->requested_freq = policy->max;
414 
415 		__cpufreq_driver_target(policy, this_dbs_info->requested_freq,
416 			CPUFREQ_RELATION_H);
417 		return;
418 	}
419 
420 	/*
421 	 * The optimal frequency is the frequency that is the lowest that
422 	 * can support the current CPU usage without triggering the up
423 	 * policy. To be safe, we focus 10 points under the threshold.
424 	 */
425 	if (max_load < (dbs_tuners_ins.down_threshold - 10)) {
426 		freq_target = (dbs_tuners_ins.freq_step * policy->max) / 100;
427 
428 		this_dbs_info->requested_freq -= freq_target;
429 		if (this_dbs_info->requested_freq < policy->min)
430 			this_dbs_info->requested_freq = policy->min;
431 
432 		/*
433 		 * if we cannot reduce the frequency anymore, break out early
434 		 */
435 		if (policy->cur == policy->min)
436 			return;
437 
438 		__cpufreq_driver_target(policy, this_dbs_info->requested_freq,
439 				CPUFREQ_RELATION_H);
440 		return;
441 	}
442 }
443 
444 static void do_dbs_timer(struct work_struct *work)
445 {
446 	struct cpu_dbs_info_s *dbs_info =
447 		container_of(work, struct cpu_dbs_info_s, work.work);
448 	unsigned int cpu = dbs_info->cpu;
449 
450 	/* We want all CPUs to do sampling nearly on same jiffy */
451 	int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
452 
453 	delay -= jiffies % delay;
454 
455 	mutex_lock(&dbs_info->timer_mutex);
456 
457 	dbs_check_cpu(dbs_info);
458 
459 	schedule_delayed_work_on(cpu, &dbs_info->work, delay);
460 	mutex_unlock(&dbs_info->timer_mutex);
461 }
462 
463 static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
464 {
465 	/* We want all CPUs to do sampling nearly on same jiffy */
466 	int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
467 	delay -= jiffies % delay;
468 
469 	dbs_info->enable = 1;
470 	INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer);
471 	schedule_delayed_work_on(dbs_info->cpu, &dbs_info->work, delay);
472 }
473 
474 static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
475 {
476 	dbs_info->enable = 0;
477 	cancel_delayed_work_sync(&dbs_info->work);
478 }
479 
480 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
481 				   unsigned int event)
482 {
483 	unsigned int cpu = policy->cpu;
484 	struct cpu_dbs_info_s *this_dbs_info;
485 	unsigned int j;
486 	int rc;
487 
488 	this_dbs_info = &per_cpu(cs_cpu_dbs_info, cpu);
489 
490 	switch (event) {
491 	case CPUFREQ_GOV_START:
492 		if ((!cpu_online(cpu)) || (!policy->cur))
493 			return -EINVAL;
494 
495 		mutex_lock(&dbs_mutex);
496 
497 		for_each_cpu(j, policy->cpus) {
498 			struct cpu_dbs_info_s *j_dbs_info;
499 			j_dbs_info = &per_cpu(cs_cpu_dbs_info, j);
500 			j_dbs_info->cur_policy = policy;
501 
502 			j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
503 						&j_dbs_info->prev_cpu_wall);
504 			if (dbs_tuners_ins.ignore_nice) {
505 				j_dbs_info->prev_cpu_nice =
506 						kstat_cpu(j).cpustat.nice;
507 			}
508 		}
509 		this_dbs_info->down_skip = 0;
510 		this_dbs_info->requested_freq = policy->cur;
511 
512 		mutex_init(&this_dbs_info->timer_mutex);
513 		dbs_enable++;
514 		/*
515 		 * Start the timerschedule work, when this governor
516 		 * is used for first time
517 		 */
518 		if (dbs_enable == 1) {
519 			unsigned int latency;
520 			/* policy latency is in nS. Convert it to uS first */
521 			latency = policy->cpuinfo.transition_latency / 1000;
522 			if (latency == 0)
523 				latency = 1;
524 
525 			rc = sysfs_create_group(cpufreq_global_kobject,
526 						&dbs_attr_group);
527 			if (rc) {
528 				mutex_unlock(&dbs_mutex);
529 				return rc;
530 			}
531 
532 			/*
533 			 * conservative does not implement micro like ondemand
534 			 * governor, thus we are bound to jiffes/HZ
535 			 */
536 			min_sampling_rate =
537 				MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10);
538 			/* Bring kernel and HW constraints together */
539 			min_sampling_rate = max(min_sampling_rate,
540 					MIN_LATENCY_MULTIPLIER * latency);
541 			dbs_tuners_ins.sampling_rate =
542 				max(min_sampling_rate,
543 				    latency * LATENCY_MULTIPLIER);
544 
545 			cpufreq_register_notifier(
546 					&dbs_cpufreq_notifier_block,
547 					CPUFREQ_TRANSITION_NOTIFIER);
548 		}
549 		mutex_unlock(&dbs_mutex);
550 
551 		dbs_timer_init(this_dbs_info);
552 
553 		break;
554 
555 	case CPUFREQ_GOV_STOP:
556 		dbs_timer_exit(this_dbs_info);
557 
558 		mutex_lock(&dbs_mutex);
559 		dbs_enable--;
560 		mutex_destroy(&this_dbs_info->timer_mutex);
561 
562 		/*
563 		 * Stop the timerschedule work, when this governor
564 		 * is used for first time
565 		 */
566 		if (dbs_enable == 0)
567 			cpufreq_unregister_notifier(
568 					&dbs_cpufreq_notifier_block,
569 					CPUFREQ_TRANSITION_NOTIFIER);
570 
571 		mutex_unlock(&dbs_mutex);
572 		if (!dbs_enable)
573 			sysfs_remove_group(cpufreq_global_kobject,
574 					   &dbs_attr_group);
575 
576 		break;
577 
578 	case CPUFREQ_GOV_LIMITS:
579 		mutex_lock(&this_dbs_info->timer_mutex);
580 		if (policy->max < this_dbs_info->cur_policy->cur)
581 			__cpufreq_driver_target(
582 					this_dbs_info->cur_policy,
583 					policy->max, CPUFREQ_RELATION_H);
584 		else if (policy->min > this_dbs_info->cur_policy->cur)
585 			__cpufreq_driver_target(
586 					this_dbs_info->cur_policy,
587 					policy->min, CPUFREQ_RELATION_L);
588 		mutex_unlock(&this_dbs_info->timer_mutex);
589 
590 		break;
591 	}
592 	return 0;
593 }
594 
595 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_CONSERVATIVE
596 static
597 #endif
598 struct cpufreq_governor cpufreq_gov_conservative = {
599 	.name			= "conservative",
600 	.governor		= cpufreq_governor_dbs,
601 	.max_transition_latency	= TRANSITION_LATENCY_LIMIT,
602 	.owner			= THIS_MODULE,
603 };
604 
605 static int __init cpufreq_gov_dbs_init(void)
606 {
607 	return cpufreq_register_governor(&cpufreq_gov_conservative);
608 }
609 
610 static void __exit cpufreq_gov_dbs_exit(void)
611 {
612 	cpufreq_unregister_governor(&cpufreq_gov_conservative);
613 }
614 
615 
616 MODULE_AUTHOR("Alexander Clouter <alex@digriz.org.uk>");
617 MODULE_DESCRIPTION("'cpufreq_conservative' - A dynamic cpufreq governor for "
618 		"Low Latency Frequency Transition capable processors "
619 		"optimised for use in a battery environment");
620 MODULE_LICENSE("GPL");
621 
622 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_CONSERVATIVE
623 fs_initcall(cpufreq_gov_dbs_init);
624 #else
625 module_init(cpufreq_gov_dbs_init);
626 #endif
627 module_exit(cpufreq_gov_dbs_exit);
628