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)
60 #define sampling_rate_in_HZ(x) (((x * HZ) < (1000 * 1000))?1:((x * HZ) / (1000 * 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;
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);
120 static ssize_t store_sampling_down_factor(struct cpufreq_policy *unused,
121 const char *buf, size_t count)
125 ret = sscanf (buf, "%u", &input);
130 dbs_tuners_ins.sampling_down_factor = input;
136 static ssize_t store_sampling_rate(struct cpufreq_policy *unused,
137 const char *buf, size_t count)
141 ret = sscanf (buf, "%u", &input);
144 if (ret != 1 || input > MAX_SAMPLING_RATE || input < MIN_SAMPLING_RATE) {
149 dbs_tuners_ins.sampling_rate = input;
155 static ssize_t store_up_threshold(struct cpufreq_policy *unused,
156 const char *buf, size_t count)
160 ret = sscanf (buf, "%u", &input);
163 if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
164 input < MIN_FREQUENCY_UP_THRESHOLD ||
165 input <= dbs_tuners_ins.down_threshold) {
170 dbs_tuners_ins.up_threshold = input;
176 static ssize_t store_down_threshold(struct cpufreq_policy *unused,
177 const char *buf, size_t count)
181 ret = sscanf (buf, "%u", &input);
184 if (ret != 1 || input > MAX_FREQUENCY_DOWN_THRESHOLD ||
185 input < MIN_FREQUENCY_DOWN_THRESHOLD ||
186 input >= dbs_tuners_ins.up_threshold) {
191 dbs_tuners_ins.down_threshold = input;
197 #define define_one_rw(_name) \
198 static struct freq_attr _name = \
199 __ATTR(_name, 0644, show_##_name, store_##_name)
201 define_one_rw(sampling_rate);
202 define_one_rw(sampling_down_factor);
203 define_one_rw(up_threshold);
204 define_one_rw(down_threshold);
206 static struct attribute * dbs_attributes[] = {
207 &sampling_rate_max.attr,
208 &sampling_rate_min.attr,
210 &sampling_down_factor.attr,
212 &down_threshold.attr,
216 static struct attribute_group dbs_attr_group = {
217 .attrs = dbs_attributes,
221 /************************** sysfs end ************************/
223 static void dbs_check_cpu(int cpu)
225 unsigned int idle_ticks, up_idle_ticks, down_idle_ticks;
226 unsigned int total_idle_ticks;
227 unsigned int freq_down_step;
228 unsigned int freq_down_sampling_rate;
229 static int down_skip[NR_CPUS];
230 struct cpu_dbs_info_s *this_dbs_info;
232 this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
233 if (!this_dbs_info->enable)
237 * The default safe range is 20% to 80%
238 * Every sampling_rate, we check
239 * - If current idle time is less than 20%, then we try to
241 * Every sampling_rate*sampling_down_factor, we check
242 * - If current idle time is more than 80%, then we try to
245 * Any frequency increase takes it to the maximum frequency.
246 * Frequency reduction happens at minimum steps of
247 * 5% of max_frequency
249 /* Check for frequency increase */
250 total_idle_ticks = kstat_cpu(cpu).cpustat.idle +
251 kstat_cpu(cpu).cpustat.iowait;
252 idle_ticks = total_idle_ticks -
253 this_dbs_info->prev_cpu_idle_up;
254 this_dbs_info->prev_cpu_idle_up = total_idle_ticks;
256 /* Scale idle ticks by 100 and compare with up and down ticks */
258 up_idle_ticks = (100 - dbs_tuners_ins.up_threshold) *
259 sampling_rate_in_HZ(dbs_tuners_ins.sampling_rate);
261 if (idle_ticks < up_idle_ticks) {
262 __cpufreq_driver_target(this_dbs_info->cur_policy,
263 this_dbs_info->cur_policy->max,
266 this_dbs_info->prev_cpu_idle_down = total_idle_ticks;
270 /* Check for frequency decrease */
272 if (down_skip[cpu] < dbs_tuners_ins.sampling_down_factor)
275 idle_ticks = total_idle_ticks -
276 this_dbs_info->prev_cpu_idle_down;
277 /* Scale idle ticks by 100 and compare with up and down ticks */
280 this_dbs_info->prev_cpu_idle_down = total_idle_ticks;
282 freq_down_sampling_rate = dbs_tuners_ins.sampling_rate *
283 dbs_tuners_ins.sampling_down_factor;
284 down_idle_ticks = (100 - dbs_tuners_ins.down_threshold) *
285 sampling_rate_in_HZ(freq_down_sampling_rate);
287 if (idle_ticks > down_idle_ticks ) {
288 freq_down_step = (5 * this_dbs_info->cur_policy->max) / 100;
290 /* max freq cannot be less than 100. But who knows.... */
291 if (unlikely(freq_down_step == 0))
294 __cpufreq_driver_target(this_dbs_info->cur_policy,
295 this_dbs_info->cur_policy->cur - freq_down_step,
301 static void do_dbs_timer(void *data)
305 for (i = 0; i < NR_CPUS; i++)
308 schedule_delayed_work(&dbs_work,
309 sampling_rate_in_HZ(dbs_tuners_ins.sampling_rate));
313 static inline void dbs_timer_init(void)
315 INIT_WORK(&dbs_work, do_dbs_timer, NULL);
316 schedule_work(&dbs_work);
320 static inline void dbs_timer_exit(void)
322 cancel_delayed_work(&dbs_work);
326 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
329 unsigned int cpu = policy->cpu;
330 struct cpu_dbs_info_s *this_dbs_info;
332 this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
335 case CPUFREQ_GOV_START:
336 if ((!cpu_online(cpu)) ||
340 if (policy->cpuinfo.transition_latency >
341 (TRANSITION_LATENCY_LIMIT * 1000))
343 if (this_dbs_info->enable) /* Already enabled */
347 this_dbs_info->cur_policy = policy;
349 this_dbs_info->prev_cpu_idle_up =
350 kstat_cpu(cpu).cpustat.idle +
351 kstat_cpu(cpu).cpustat.iowait;
352 this_dbs_info->prev_cpu_idle_down =
353 kstat_cpu(cpu).cpustat.idle +
354 kstat_cpu(cpu).cpustat.iowait;
355 this_dbs_info->enable = 1;
356 sysfs_create_group(&policy->kobj, &dbs_attr_group);
359 * Start the timerschedule work, when this governor
360 * is used for first time
362 if (dbs_enable == 1) {
363 unsigned int latency;
364 /* policy latency is in nS. Convert it to uS first */
366 latency = policy->cpuinfo.transition_latency;
370 def_sampling_rate = (latency / 1000) *
371 DEF_SAMPLING_RATE_LATENCY_MULTIPLIER;
372 dbs_tuners_ins.sampling_rate = def_sampling_rate;
380 case CPUFREQ_GOV_STOP:
382 this_dbs_info->enable = 0;
383 sysfs_remove_group(&policy->kobj, &dbs_attr_group);
386 * Stop the timerschedule work, when this governor
387 * is used for first time
396 case CPUFREQ_GOV_LIMITS:
398 if (policy->max < this_dbs_info->cur_policy->cur)
399 __cpufreq_driver_target(
400 this_dbs_info->cur_policy,
401 policy->max, CPUFREQ_RELATION_H);
402 else if (policy->min > this_dbs_info->cur_policy->cur)
403 __cpufreq_driver_target(
404 this_dbs_info->cur_policy,
405 policy->min, CPUFREQ_RELATION_L);
412 struct cpufreq_governor cpufreq_gov_dbs = {
414 .governor = cpufreq_governor_dbs,
415 .owner = THIS_MODULE,
417 EXPORT_SYMBOL(cpufreq_gov_dbs);
419 static int __init cpufreq_gov_dbs_init(void)
421 return cpufreq_register_governor(&cpufreq_gov_dbs);
424 static void __exit cpufreq_gov_dbs_exit(void)
426 /* Make sure that the scheduled work is indeed not running */
427 flush_scheduled_work();
429 cpufreq_unregister_governor(&cpufreq_gov_dbs);
433 MODULE_AUTHOR ("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
434 MODULE_DESCRIPTION ("'cpufreq_ondemand' - A dynamic cpufreq governor for "
435 "Low Latency Frequency Transition capable processors");
436 MODULE_LICENSE ("GPL");
438 module_init(cpufreq_gov_dbs_init);
439 module_exit(cpufreq_gov_dbs_exit);