2 * drivers/cpufreq/cpufreq_conservative.c
4 * Copyright (C) 2001 Russell King
5 * (C) 2003 Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>.
6 * Jun Nakajima <jun.nakajima@intel.com>
7 * (C) 2004 Alexander Clouter <alex-kernel@digriz.org.uk>
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.
14 #include <linux/kernel.h>
15 #include <linux/module.h>
16 #include <linux/smp.h>
17 #include <linux/init.h>
18 #include <linux/interrupt.h>
19 #include <linux/ctype.h>
20 #include <linux/cpufreq.h>
21 #include <linux/sysctl.h>
22 #include <linux/types.h>
24 #include <linux/sysfs.h>
25 #include <linux/sched.h>
26 #include <linux/kmod.h>
27 #include <linux/workqueue.h>
28 #include <linux/jiffies.h>
29 #include <linux/kernel_stat.h>
30 #include <linux/percpu.h>
33 * dbs is used in this file as a shortform for demandbased switching
34 * It helps to keep variable names smaller, simpler
37 #define DEF_FREQUENCY_UP_THRESHOLD (80)
38 #define MIN_FREQUENCY_UP_THRESHOLD (0)
39 #define MAX_FREQUENCY_UP_THRESHOLD (100)
41 #define DEF_FREQUENCY_DOWN_THRESHOLD (20)
42 #define MIN_FREQUENCY_DOWN_THRESHOLD (0)
43 #define MAX_FREQUENCY_DOWN_THRESHOLD (100)
46 * The polling frequency of this governor depends on the capability of
47 * the processor. Default polling frequency is 1000 times the transition
48 * latency of the processor. The governor will work on any processor with
49 * transition latency <= 10mS, using appropriate sampling
51 * For CPUs with transition latency > 10mS (mostly drivers with CPUFREQ_ETERNAL)
52 * this governor will not work.
53 * All times here are in uS.
55 static unsigned int def_sampling_rate;
56 #define MIN_SAMPLING_RATE (def_sampling_rate / 2)
57 #define MAX_SAMPLING_RATE (500 * def_sampling_rate)
58 #define DEF_SAMPLING_RATE_LATENCY_MULTIPLIER (100000)
59 #define DEF_SAMPLING_DOWN_FACTOR (5)
60 #define TRANSITION_LATENCY_LIMIT (10 * 1000)
62 static void do_dbs_timer(void *data);
64 struct cpu_dbs_info_s {
65 struct cpufreq_policy *cur_policy;
66 unsigned int prev_cpu_idle_up;
67 unsigned int prev_cpu_idle_down;
70 static DEFINE_PER_CPU(struct cpu_dbs_info_s, cpu_dbs_info);
72 static unsigned int dbs_enable; /* number of CPUs using this policy */
74 static DECLARE_MUTEX (dbs_sem);
75 static DECLARE_WORK (dbs_work, do_dbs_timer, NULL);
78 unsigned int sampling_rate;
79 unsigned int sampling_down_factor;
80 unsigned int up_threshold;
81 unsigned int down_threshold;
82 unsigned int ignore_nice;
83 unsigned int freq_step;
86 static struct dbs_tuners dbs_tuners_ins = {
87 .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
88 .down_threshold = DEF_FREQUENCY_DOWN_THRESHOLD,
89 .sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR,
92 /************************** sysfs interface ************************/
93 static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf)
95 return sprintf (buf, "%u\n", MAX_SAMPLING_RATE);
98 static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf)
100 return sprintf (buf, "%u\n", MIN_SAMPLING_RATE);
103 #define define_one_ro(_name) \
104 static struct freq_attr _name = \
105 __ATTR(_name, 0444, show_##_name, NULL)
107 define_one_ro(sampling_rate_max);
108 define_one_ro(sampling_rate_min);
110 /* cpufreq_conservative Governor Tunables */
111 #define show_one(file_name, object) \
112 static ssize_t show_##file_name \
113 (struct cpufreq_policy *unused, char *buf) \
115 return sprintf(buf, "%u\n", dbs_tuners_ins.object); \
117 show_one(sampling_rate, sampling_rate);
118 show_one(sampling_down_factor, sampling_down_factor);
119 show_one(up_threshold, up_threshold);
120 show_one(down_threshold, down_threshold);
121 show_one(ignore_nice, ignore_nice);
122 show_one(freq_step, freq_step);
124 static ssize_t store_sampling_down_factor(struct cpufreq_policy *unused,
125 const char *buf, size_t count)
129 ret = sscanf (buf, "%u", &input);
134 dbs_tuners_ins.sampling_down_factor = input;
140 static ssize_t store_sampling_rate(struct cpufreq_policy *unused,
141 const char *buf, size_t count)
145 ret = sscanf (buf, "%u", &input);
148 if (ret != 1 || input > MAX_SAMPLING_RATE || input < MIN_SAMPLING_RATE) {
153 dbs_tuners_ins.sampling_rate = input;
159 static ssize_t store_up_threshold(struct cpufreq_policy *unused,
160 const char *buf, size_t count)
164 ret = sscanf (buf, "%u", &input);
167 if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
168 input < MIN_FREQUENCY_UP_THRESHOLD ||
169 input <= dbs_tuners_ins.down_threshold) {
174 dbs_tuners_ins.up_threshold = input;
180 static ssize_t store_down_threshold(struct cpufreq_policy *unused,
181 const char *buf, size_t count)
185 ret = sscanf (buf, "%u", &input);
188 if (ret != 1 || input > MAX_FREQUENCY_DOWN_THRESHOLD ||
189 input < MIN_FREQUENCY_DOWN_THRESHOLD ||
190 input >= dbs_tuners_ins.up_threshold) {
195 dbs_tuners_ins.down_threshold = input;
201 static ssize_t store_ignore_nice(struct cpufreq_policy *policy,
202 const char *buf, size_t count)
209 ret = sscanf (buf, "%u", &input);
217 if ( input == dbs_tuners_ins.ignore_nice ) { /* nothing to do */
221 dbs_tuners_ins.ignore_nice = input;
223 /* we need to re-evaluate prev_cpu_idle_up and prev_cpu_idle_down */
224 for_each_cpu_mask(j, policy->cpus) {
225 struct cpu_dbs_info_s *j_dbs_info;
226 j_dbs_info = &per_cpu(cpu_dbs_info, j);
227 j_dbs_info->cur_policy = policy;
229 j_dbs_info->prev_cpu_idle_up =
230 kstat_cpu(j).cpustat.idle +
231 kstat_cpu(j).cpustat.iowait +
232 ( !dbs_tuners_ins.ignore_nice
233 ? kstat_cpu(j).cpustat.nice : 0 );
234 j_dbs_info->prev_cpu_idle_down = j_dbs_info->prev_cpu_idle_up;
241 static ssize_t store_freq_step(struct cpufreq_policy *policy,
242 const char *buf, size_t count)
247 ret = sscanf (buf, "%u", &input);
255 /* no need to test here if freq_step is zero as the user might actually
256 * want this, they would be crazy though :) */
258 dbs_tuners_ins.freq_step = input;
264 #define define_one_rw(_name) \
265 static struct freq_attr _name = \
266 __ATTR(_name, 0644, show_##_name, store_##_name)
268 define_one_rw(sampling_rate);
269 define_one_rw(sampling_down_factor);
270 define_one_rw(up_threshold);
271 define_one_rw(down_threshold);
272 define_one_rw(ignore_nice);
273 define_one_rw(freq_step);
275 static struct attribute * dbs_attributes[] = {
276 &sampling_rate_max.attr,
277 &sampling_rate_min.attr,
279 &sampling_down_factor.attr,
281 &down_threshold.attr,
287 static struct attribute_group dbs_attr_group = {
288 .attrs = dbs_attributes,
289 .name = "conservative",
292 /************************** sysfs end ************************/
294 static void dbs_check_cpu(int cpu)
296 unsigned int idle_ticks, up_idle_ticks, down_idle_ticks;
297 unsigned int total_idle_ticks;
298 unsigned int freq_step;
299 unsigned int freq_down_sampling_rate;
300 static int down_skip[NR_CPUS];
301 static int requested_freq[NR_CPUS];
302 static unsigned short init_flag = 0;
303 struct cpu_dbs_info_s *this_dbs_info;
304 struct cpu_dbs_info_s *dbs_info;
306 struct cpufreq_policy *policy;
309 this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
310 if (!this_dbs_info->enable)
313 policy = this_dbs_info->cur_policy;
315 if ( init_flag == 0 ) {
316 for ( /* NULL */; init_flag < NR_CPUS; init_flag++ ) {
317 dbs_info = &per_cpu(cpu_dbs_info, init_flag);
318 requested_freq[cpu] = dbs_info->cur_policy->cur;
324 * The default safe range is 20% to 80%
325 * Every sampling_rate, we check
326 * - If current idle time is less than 20%, then we try to
328 * Every sampling_rate*sampling_down_factor, we check
329 * - If current idle time is more than 80%, then we try to
332 * Any frequency increase takes it to the maximum frequency.
333 * Frequency reduction happens at minimum steps of
334 * 5% (default) of max_frequency
337 /* Check for frequency increase */
338 total_idle_ticks = kstat_cpu(cpu).cpustat.idle +
339 kstat_cpu(cpu).cpustat.iowait;
340 /* consider 'nice' tasks as 'idle' time too if required */
341 if (dbs_tuners_ins.ignore_nice == 0)
342 total_idle_ticks += kstat_cpu(cpu).cpustat.nice;
343 idle_ticks = total_idle_ticks -
344 this_dbs_info->prev_cpu_idle_up;
345 this_dbs_info->prev_cpu_idle_up = total_idle_ticks;
348 for_each_cpu_mask(j, policy->cpus) {
349 unsigned int tmp_idle_ticks;
350 struct cpu_dbs_info_s *j_dbs_info;
355 j_dbs_info = &per_cpu(cpu_dbs_info, j);
356 /* Check for frequency increase */
357 total_idle_ticks = kstat_cpu(j).cpustat.idle +
358 kstat_cpu(j).cpustat.iowait;
359 /* consider 'nice' too? */
360 if (dbs_tuners_ins.ignore_nice == 0)
361 total_idle_ticks += kstat_cpu(j).cpustat.nice;
362 tmp_idle_ticks = total_idle_ticks -
363 j_dbs_info->prev_cpu_idle_up;
364 j_dbs_info->prev_cpu_idle_up = total_idle_ticks;
366 if (tmp_idle_ticks < idle_ticks)
367 idle_ticks = tmp_idle_ticks;
370 /* Scale idle ticks by 100 and compare with up and down ticks */
372 up_idle_ticks = (100 - dbs_tuners_ins.up_threshold) *
373 usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
375 if (idle_ticks < up_idle_ticks) {
376 /* if we are already at full speed then break out early */
377 if (requested_freq[cpu] == policy->max)
380 freq_step = (dbs_tuners_ins.freq_step * policy->max) / 100;
382 /* max freq cannot be less than 100. But who knows.... */
383 if (unlikely(freq_step == 0))
386 requested_freq[cpu] += freq_step;
387 if (requested_freq[cpu] > policy->max)
388 requested_freq[cpu] = policy->max;
390 __cpufreq_driver_target(policy, requested_freq[cpu],
393 this_dbs_info->prev_cpu_idle_down = total_idle_ticks;
397 /* Check for frequency decrease */
399 if (down_skip[cpu] < dbs_tuners_ins.sampling_down_factor)
402 total_idle_ticks = kstat_cpu(cpu).cpustat.idle +
403 kstat_cpu(cpu).cpustat.iowait;
404 /* consider 'nice' too? */
405 if (dbs_tuners_ins.ignore_nice == 0)
406 total_idle_ticks += kstat_cpu(cpu).cpustat.nice;
407 idle_ticks = total_idle_ticks -
408 this_dbs_info->prev_cpu_idle_down;
409 this_dbs_info->prev_cpu_idle_down = total_idle_ticks;
411 for_each_cpu_mask(j, policy->cpus) {
412 unsigned int tmp_idle_ticks;
413 struct cpu_dbs_info_s *j_dbs_info;
418 j_dbs_info = &per_cpu(cpu_dbs_info, j);
419 /* Check for frequency increase */
420 total_idle_ticks = kstat_cpu(j).cpustat.idle +
421 kstat_cpu(j).cpustat.iowait;
422 /* consider 'nice' too? */
423 if (dbs_tuners_ins.ignore_nice == 0)
424 total_idle_ticks += kstat_cpu(j).cpustat.nice;
425 tmp_idle_ticks = total_idle_ticks -
426 j_dbs_info->prev_cpu_idle_down;
427 j_dbs_info->prev_cpu_idle_down = total_idle_ticks;
429 if (tmp_idle_ticks < idle_ticks)
430 idle_ticks = tmp_idle_ticks;
433 /* Scale idle ticks by 100 and compare with up and down ticks */
437 freq_down_sampling_rate = dbs_tuners_ins.sampling_rate *
438 dbs_tuners_ins.sampling_down_factor;
439 down_idle_ticks = (100 - dbs_tuners_ins.down_threshold) *
440 usecs_to_jiffies(freq_down_sampling_rate);
442 if (idle_ticks > down_idle_ticks ) {
443 /* if we are already at the lowest speed then break out early
444 * or if we 'cannot' reduce the speed as the user might want
445 * freq_step to be zero */
446 if (requested_freq[cpu] == policy->min
447 || dbs_tuners_ins.freq_step == 0)
450 freq_step = (dbs_tuners_ins.freq_step * policy->max) / 100;
452 /* max freq cannot be less than 100. But who knows.... */
453 if (unlikely(freq_step == 0))
456 requested_freq[cpu] -= freq_step;
457 if (requested_freq[cpu] < policy->min)
458 requested_freq[cpu] = policy->min;
460 __cpufreq_driver_target(policy,
467 static void do_dbs_timer(void *data)
471 for_each_online_cpu(i)
473 schedule_delayed_work(&dbs_work,
474 usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
478 static inline void dbs_timer_init(void)
480 INIT_WORK(&dbs_work, do_dbs_timer, NULL);
481 schedule_delayed_work(&dbs_work,
482 usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
486 static inline void dbs_timer_exit(void)
488 cancel_delayed_work(&dbs_work);
492 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
495 unsigned int cpu = policy->cpu;
496 struct cpu_dbs_info_s *this_dbs_info;
499 this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
502 case CPUFREQ_GOV_START:
503 if ((!cpu_online(cpu)) ||
507 if (policy->cpuinfo.transition_latency >
508 (TRANSITION_LATENCY_LIMIT * 1000))
510 if (this_dbs_info->enable) /* Already enabled */
514 for_each_cpu_mask(j, policy->cpus) {
515 struct cpu_dbs_info_s *j_dbs_info;
516 j_dbs_info = &per_cpu(cpu_dbs_info, j);
517 j_dbs_info->cur_policy = policy;
519 j_dbs_info->prev_cpu_idle_up =
520 kstat_cpu(j).cpustat.idle +
521 kstat_cpu(j).cpustat.iowait +
522 ( !dbs_tuners_ins.ignore_nice
523 ? kstat_cpu(j).cpustat.nice : 0 );
524 j_dbs_info->prev_cpu_idle_down
525 = j_dbs_info->prev_cpu_idle_up;
527 this_dbs_info->enable = 1;
528 sysfs_create_group(&policy->kobj, &dbs_attr_group);
531 * Start the timerschedule work, when this governor
532 * is used for first time
534 if (dbs_enable == 1) {
535 unsigned int latency;
536 /* policy latency is in nS. Convert it to uS first */
538 latency = policy->cpuinfo.transition_latency;
542 def_sampling_rate = (latency / 1000) *
543 DEF_SAMPLING_RATE_LATENCY_MULTIPLIER;
544 dbs_tuners_ins.sampling_rate = def_sampling_rate;
545 dbs_tuners_ins.ignore_nice = 0;
546 dbs_tuners_ins.freq_step = 5;
554 case CPUFREQ_GOV_STOP:
556 this_dbs_info->enable = 0;
557 sysfs_remove_group(&policy->kobj, &dbs_attr_group);
560 * Stop the timerschedule work, when this governor
561 * is used for first time
570 case CPUFREQ_GOV_LIMITS:
572 if (policy->max < this_dbs_info->cur_policy->cur)
573 __cpufreq_driver_target(
574 this_dbs_info->cur_policy,
575 policy->max, CPUFREQ_RELATION_H);
576 else if (policy->min > this_dbs_info->cur_policy->cur)
577 __cpufreq_driver_target(
578 this_dbs_info->cur_policy,
579 policy->min, CPUFREQ_RELATION_L);
586 static struct cpufreq_governor cpufreq_gov_dbs = {
587 .name = "conservative",
588 .governor = cpufreq_governor_dbs,
589 .owner = THIS_MODULE,
592 static int __init cpufreq_gov_dbs_init(void)
594 return cpufreq_register_governor(&cpufreq_gov_dbs);
597 static void __exit cpufreq_gov_dbs_exit(void)
599 /* Make sure that the scheduled work is indeed not running */
600 flush_scheduled_work();
602 cpufreq_unregister_governor(&cpufreq_gov_dbs);
606 MODULE_AUTHOR ("Alexander Clouter <alex-kernel@digriz.org.uk>");
607 MODULE_DESCRIPTION ("'cpufreq_conservative' - A dynamic cpufreq governor for "
608 "Low Latency Frequency Transition capable processors "
609 "optimised for use in a battery environment");
610 MODULE_LICENSE ("GPL");
612 module_init(cpufreq_gov_dbs_init);
613 module_exit(cpufreq_gov_dbs_exit);