2 * drivers/cpufreq/cpufreq_ondemand.c
4 * Copyright (C) 2001 Russell King
5 * (C) 2003 Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>.
6 * Jun Nakajima <jun.nakajima@intel.com>
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 as
10 * published by the Free Software Foundation.
13 #include <linux/kernel.h>
14 #include <linux/module.h>
15 #include <linux/smp.h>
16 #include <linux/init.h>
17 #include <linux/interrupt.h>
18 #include <linux/ctype.h>
19 #include <linux/cpufreq.h>
20 #include <linux/sysctl.h>
21 #include <linux/types.h>
23 #include <linux/sysfs.h>
24 #include <linux/sched.h>
25 #include <linux/kmod.h>
26 #include <linux/workqueue.h>
27 #include <linux/jiffies.h>
28 #include <linux/kernel_stat.h>
29 #include <linux/percpu.h>
32 * dbs is used in this file as a shortform for demandbased switching
33 * It helps to keep variable names smaller, simpler
36 #define DEF_FREQUENCY_UP_THRESHOLD (80)
37 #define MIN_FREQUENCY_UP_THRESHOLD (0)
38 #define MAX_FREQUENCY_UP_THRESHOLD (100)
40 #define DEF_FREQUENCY_DOWN_THRESHOLD (20)
41 #define MIN_FREQUENCY_DOWN_THRESHOLD (0)
42 #define MAX_FREQUENCY_DOWN_THRESHOLD (100)
45 * The polling frequency of this governor depends on the capability of
46 * the processor. Default polling frequency is 1000 times the transition
47 * latency of the processor. The governor will work on any processor with
48 * transition latency <= 10mS, using appropriate sampling
50 * For CPUs with transition latency > 10mS (mostly drivers with CPUFREQ_ETERNAL)
51 * this governor will not work.
52 * All times here are in uS.
54 static unsigned int def_sampling_rate;
55 #define MIN_SAMPLING_RATE (def_sampling_rate / 2)
56 #define MAX_SAMPLING_RATE (500 * def_sampling_rate)
57 #define DEF_SAMPLING_RATE_LATENCY_MULTIPLIER (1000)
58 #define DEF_SAMPLING_DOWN_FACTOR (10)
59 #define TRANSITION_LATENCY_LIMIT (10 * 1000)
61 static void do_dbs_timer(void *data);
63 struct cpu_dbs_info_s {
64 struct cpufreq_policy *cur_policy;
65 unsigned int prev_cpu_idle_up;
66 unsigned int prev_cpu_idle_down;
69 static DEFINE_PER_CPU(struct cpu_dbs_info_s, cpu_dbs_info);
71 static unsigned int dbs_enable; /* number of CPUs using this policy */
73 static DECLARE_MUTEX (dbs_sem);
74 static DECLARE_WORK (dbs_work, do_dbs_timer, NULL);
77 unsigned int sampling_rate;
78 unsigned int sampling_down_factor;
79 unsigned int up_threshold;
80 unsigned int down_threshold;
81 unsigned int ignore_nice;
84 static struct dbs_tuners dbs_tuners_ins = {
85 .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
86 .down_threshold = DEF_FREQUENCY_DOWN_THRESHOLD,
87 .sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR,
90 /************************** sysfs interface ************************/
91 static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf)
93 return sprintf (buf, "%u\n", MAX_SAMPLING_RATE);
96 static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf)
98 return sprintf (buf, "%u\n", MIN_SAMPLING_RATE);
101 #define define_one_ro(_name) \
102 static struct freq_attr _name = \
103 __ATTR(_name, 0444, show_##_name, NULL)
105 define_one_ro(sampling_rate_max);
106 define_one_ro(sampling_rate_min);
108 /* cpufreq_ondemand Governor Tunables */
109 #define show_one(file_name, object) \
110 static ssize_t show_##file_name \
111 (struct cpufreq_policy *unused, char *buf) \
113 return sprintf(buf, "%u\n", dbs_tuners_ins.object); \
115 show_one(sampling_rate, sampling_rate);
116 show_one(sampling_down_factor, sampling_down_factor);
117 show_one(up_threshold, up_threshold);
118 show_one(down_threshold, down_threshold);
119 show_one(ignore_nice, ignore_nice);
121 static ssize_t store_sampling_down_factor(struct cpufreq_policy *unused,
122 const char *buf, size_t count)
126 ret = sscanf (buf, "%u", &input);
131 dbs_tuners_ins.sampling_down_factor = input;
137 static ssize_t store_sampling_rate(struct cpufreq_policy *unused,
138 const char *buf, size_t count)
142 ret = sscanf (buf, "%u", &input);
145 if (ret != 1 || input > MAX_SAMPLING_RATE || input < MIN_SAMPLING_RATE) {
150 dbs_tuners_ins.sampling_rate = input;
156 static ssize_t store_up_threshold(struct cpufreq_policy *unused,
157 const char *buf, size_t count)
161 ret = sscanf (buf, "%u", &input);
164 if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
165 input < MIN_FREQUENCY_UP_THRESHOLD ||
166 input <= dbs_tuners_ins.down_threshold) {
171 dbs_tuners_ins.up_threshold = input;
177 static ssize_t store_down_threshold(struct cpufreq_policy *unused,
178 const char *buf, size_t count)
182 ret = sscanf (buf, "%u", &input);
185 if (ret != 1 || input > MAX_FREQUENCY_DOWN_THRESHOLD ||
186 input < MIN_FREQUENCY_DOWN_THRESHOLD ||
187 input >= dbs_tuners_ins.up_threshold) {
192 dbs_tuners_ins.down_threshold = input;
198 static ssize_t store_ignore_nice(struct cpufreq_policy *policy,
199 const char *buf, size_t count)
206 ret = sscanf (buf, "%u", &input);
214 if ( input == dbs_tuners_ins.ignore_nice ) { /* nothing to do */
218 dbs_tuners_ins.ignore_nice = input;
220 /* we need to re-evaluate prev_cpu_idle_up and prev_cpu_idle_down */
221 for_each_cpu_mask(j, policy->cpus) {
222 struct cpu_dbs_info_s *j_dbs_info;
223 j_dbs_info = &per_cpu(cpu_dbs_info, j);
224 j_dbs_info->cur_policy = policy;
226 j_dbs_info->prev_cpu_idle_up =
227 kstat_cpu(j).cpustat.idle +
228 kstat_cpu(j).cpustat.iowait +
229 ( !dbs_tuners_ins.ignore_nice
230 ? kstat_cpu(j).cpustat.nice : 0 );
231 j_dbs_info->prev_cpu_idle_down = j_dbs_info->prev_cpu_idle_up;
238 #define define_one_rw(_name) \
239 static struct freq_attr _name = \
240 __ATTR(_name, 0644, show_##_name, store_##_name)
242 define_one_rw(sampling_rate);
243 define_one_rw(sampling_down_factor);
244 define_one_rw(up_threshold);
245 define_one_rw(down_threshold);
246 define_one_rw(ignore_nice);
248 static struct attribute * dbs_attributes[] = {
249 &sampling_rate_max.attr,
250 &sampling_rate_min.attr,
252 &sampling_down_factor.attr,
254 &down_threshold.attr,
259 static struct attribute_group dbs_attr_group = {
260 .attrs = dbs_attributes,
264 /************************** sysfs end ************************/
266 static void dbs_check_cpu(int cpu)
268 unsigned int idle_ticks, up_idle_ticks, down_idle_ticks;
269 unsigned int total_idle_ticks;
270 unsigned int freq_down_step;
271 unsigned int freq_down_sampling_rate;
272 static int down_skip[NR_CPUS];
273 struct cpu_dbs_info_s *this_dbs_info;
275 struct cpufreq_policy *policy;
278 this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
279 if (!this_dbs_info->enable)
282 policy = this_dbs_info->cur_policy;
284 * The default safe range is 20% to 80%
285 * Every sampling_rate, we check
286 * - If current idle time is less than 20%, then we try to
288 * Every sampling_rate*sampling_down_factor, we check
289 * - If current idle time is more than 80%, then we try to
292 * Any frequency increase takes it to the maximum frequency.
293 * Frequency reduction happens at minimum steps of
294 * 5% of max_frequency
297 /* Check for frequency increase */
298 total_idle_ticks = kstat_cpu(cpu).cpustat.idle +
299 kstat_cpu(cpu).cpustat.iowait;
300 /* consider 'nice' tasks as 'idle' time too if required */
301 if (dbs_tuners_ins.ignore_nice == 0)
302 total_idle_ticks += kstat_cpu(cpu).cpustat.nice;
303 idle_ticks = total_idle_ticks -
304 this_dbs_info->prev_cpu_idle_up;
305 this_dbs_info->prev_cpu_idle_up = total_idle_ticks;
308 for_each_cpu_mask(j, policy->cpus) {
309 unsigned int tmp_idle_ticks;
310 struct cpu_dbs_info_s *j_dbs_info;
315 j_dbs_info = &per_cpu(cpu_dbs_info, j);
316 /* Check for frequency increase */
317 total_idle_ticks = kstat_cpu(j).cpustat.idle +
318 kstat_cpu(j).cpustat.iowait;
319 /* consider 'nice' too? */
320 if (dbs_tuners_ins.ignore_nice == 0)
321 total_idle_ticks += kstat_cpu(j).cpustat.nice;
322 tmp_idle_ticks = total_idle_ticks -
323 j_dbs_info->prev_cpu_idle_up;
324 j_dbs_info->prev_cpu_idle_up = total_idle_ticks;
326 if (tmp_idle_ticks < idle_ticks)
327 idle_ticks = tmp_idle_ticks;
330 /* Scale idle ticks by 100 and compare with up and down ticks */
332 up_idle_ticks = (100 - dbs_tuners_ins.up_threshold) *
333 usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
335 if (idle_ticks < up_idle_ticks) {
336 /* if we are already at full speed then break out early */
337 if (policy->cur == policy->max)
340 __cpufreq_driver_target(policy, policy->max,
343 this_dbs_info->prev_cpu_idle_down = total_idle_ticks;
347 /* Check for frequency decrease */
349 if (down_skip[cpu] < dbs_tuners_ins.sampling_down_factor)
352 total_idle_ticks = kstat_cpu(cpu).cpustat.idle +
353 kstat_cpu(cpu).cpustat.iowait;
354 /* consider 'nice' too? */
355 if (dbs_tuners_ins.ignore_nice == 0)
356 total_idle_ticks += kstat_cpu(cpu).cpustat.nice;
357 idle_ticks = total_idle_ticks -
358 this_dbs_info->prev_cpu_idle_down;
359 this_dbs_info->prev_cpu_idle_down = total_idle_ticks;
361 for_each_cpu_mask(j, policy->cpus) {
362 unsigned int tmp_idle_ticks;
363 struct cpu_dbs_info_s *j_dbs_info;
368 j_dbs_info = &per_cpu(cpu_dbs_info, j);
369 /* Check for frequency increase */
370 total_idle_ticks = kstat_cpu(j).cpustat.idle +
371 kstat_cpu(j).cpustat.iowait;
372 /* consider 'nice' too? */
373 if (dbs_tuners_ins.ignore_nice == 0)
374 total_idle_ticks += kstat_cpu(j).cpustat.nice;
375 tmp_idle_ticks = total_idle_ticks -
376 j_dbs_info->prev_cpu_idle_down;
377 j_dbs_info->prev_cpu_idle_down = total_idle_ticks;
379 if (tmp_idle_ticks < idle_ticks)
380 idle_ticks = tmp_idle_ticks;
383 /* Scale idle ticks by 100 and compare with up and down ticks */
387 freq_down_sampling_rate = dbs_tuners_ins.sampling_rate *
388 dbs_tuners_ins.sampling_down_factor;
389 down_idle_ticks = (100 - dbs_tuners_ins.down_threshold) *
390 usecs_to_jiffies(freq_down_sampling_rate);
392 if (idle_ticks > down_idle_ticks ) {
393 /* if we are already at the lowest speed then break out early */
394 if (policy->cur == policy->min)
397 freq_down_step = (5 * policy->max) / 100;
399 /* max freq cannot be less than 100. But who knows.... */
400 if (unlikely(freq_down_step == 0))
403 __cpufreq_driver_target(policy,
404 policy->cur - freq_down_step,
410 static void do_dbs_timer(void *data)
414 for_each_online_cpu(i)
416 schedule_delayed_work(&dbs_work,
417 usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
421 static inline void dbs_timer_init(void)
423 INIT_WORK(&dbs_work, do_dbs_timer, NULL);
424 schedule_delayed_work(&dbs_work,
425 usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
429 static inline void dbs_timer_exit(void)
431 cancel_delayed_work(&dbs_work);
435 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
438 unsigned int cpu = policy->cpu;
439 struct cpu_dbs_info_s *this_dbs_info;
442 this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
445 case CPUFREQ_GOV_START:
446 if ((!cpu_online(cpu)) ||
450 if (policy->cpuinfo.transition_latency >
451 (TRANSITION_LATENCY_LIMIT * 1000))
453 if (this_dbs_info->enable) /* Already enabled */
457 for_each_cpu_mask(j, policy->cpus) {
458 struct cpu_dbs_info_s *j_dbs_info;
459 j_dbs_info = &per_cpu(cpu_dbs_info, j);
460 j_dbs_info->cur_policy = policy;
462 j_dbs_info->prev_cpu_idle_up =
463 kstat_cpu(j).cpustat.idle +
464 kstat_cpu(j).cpustat.iowait +
465 ( !dbs_tuners_ins.ignore_nice
466 ? kstat_cpu(j).cpustat.nice : 0 );
467 j_dbs_info->prev_cpu_idle_down
468 = j_dbs_info->prev_cpu_idle_up;
470 this_dbs_info->enable = 1;
471 sysfs_create_group(&policy->kobj, &dbs_attr_group);
474 * Start the timerschedule work, when this governor
475 * is used for first time
477 if (dbs_enable == 1) {
478 unsigned int latency;
479 /* policy latency is in nS. Convert it to uS first */
481 latency = policy->cpuinfo.transition_latency;
485 def_sampling_rate = (latency / 1000) *
486 DEF_SAMPLING_RATE_LATENCY_MULTIPLIER;
487 dbs_tuners_ins.sampling_rate = def_sampling_rate;
488 dbs_tuners_ins.ignore_nice = 0;
496 case CPUFREQ_GOV_STOP:
498 this_dbs_info->enable = 0;
499 sysfs_remove_group(&policy->kobj, &dbs_attr_group);
502 * Stop the timerschedule work, when this governor
503 * is used for first time
512 case CPUFREQ_GOV_LIMITS:
514 if (policy->max < this_dbs_info->cur_policy->cur)
515 __cpufreq_driver_target(
516 this_dbs_info->cur_policy,
517 policy->max, CPUFREQ_RELATION_H);
518 else if (policy->min > this_dbs_info->cur_policy->cur)
519 __cpufreq_driver_target(
520 this_dbs_info->cur_policy,
521 policy->min, CPUFREQ_RELATION_L);
528 static struct cpufreq_governor cpufreq_gov_dbs = {
530 .governor = cpufreq_governor_dbs,
531 .owner = THIS_MODULE,
534 static int __init cpufreq_gov_dbs_init(void)
536 return cpufreq_register_governor(&cpufreq_gov_dbs);
539 static void __exit cpufreq_gov_dbs_exit(void)
541 /* Make sure that the scheduled work is indeed not running */
542 flush_scheduled_work();
544 cpufreq_unregister_governor(&cpufreq_gov_dbs);
548 MODULE_AUTHOR ("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
549 MODULE_DESCRIPTION ("'cpufreq_ondemand' - A dynamic cpufreq governor for "
550 "Low Latency Frequency Transition capable processors");
551 MODULE_LICENSE ("GPL");
553 module_init(cpufreq_gov_dbs_init);
554 module_exit(cpufreq_gov_dbs_exit);