blob: fa8af4ebb1d6b74d1ff73f64e665e22c87b839ee [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * drivers/cpufreq/cpufreq_ondemand.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 *
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.
11 */
12
13#include <linux/kernel.h>
14#include <linux/module.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070015#include <linux/init.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070016#include <linux/cpufreq.h>
Andrew Morton138a01282006-06-23 03:31:19 -070017#include <linux/cpu.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070018#include <linux/jiffies.h>
19#include <linux/kernel_stat.h>
akpm@osdl.org3fc54d32006-01-13 15:54:22 -080020#include <linux/mutex.h>
venkatesh.pallipadi@intel.com80800912008-08-04 11:59:12 -070021#include <linux/hrtimer.h>
22#include <linux/tick.h>
23#include <linux/ktime.h>
Thomas Renninger9411b4e2009-02-04 11:54:04 +010024#include <linux/sched.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070025
26/*
27 * dbs is used in this file as a shortform for demandbased switching
28 * It helps to keep variable names smaller, simpler
29 */
30
venkatesh.pallipadi@intel.come9d95bf2008-08-04 11:59:10 -070031#define DEF_FREQUENCY_DOWN_DIFFERENTIAL (10)
Linus Torvalds1da177e2005-04-16 15:20:36 -070032#define DEF_FREQUENCY_UP_THRESHOLD (80)
David C Niemi3f78a9f2010-10-06 16:54:24 -040033#define DEF_SAMPLING_DOWN_FACTOR (1)
34#define MAX_SAMPLING_DOWN_FACTOR (100000)
venkatesh.pallipadi@intel.com80800912008-08-04 11:59:12 -070035#define MICRO_FREQUENCY_DOWN_DIFFERENTIAL (3)
36#define MICRO_FREQUENCY_UP_THRESHOLD (95)
Thomas Renningercef96152009-04-22 13:48:29 +020037#define MICRO_FREQUENCY_MIN_SAMPLE_RATE (10000)
Dave Jonesc29f1402005-05-31 19:03:50 -070038#define MIN_FREQUENCY_UP_THRESHOLD (11)
Linus Torvalds1da177e2005-04-16 15:20:36 -070039#define MAX_FREQUENCY_UP_THRESHOLD (100)
40
Dave Jones32ee8c32006-02-28 00:43:23 -050041/*
42 * The polling frequency of this governor depends on the capability of
Linus Torvalds1da177e2005-04-16 15:20:36 -070043 * the processor. Default polling frequency is 1000 times the transition
Dave Jones32ee8c32006-02-28 00:43:23 -050044 * latency of the processor. The governor will work on any processor with
45 * transition latency <= 10mS, using appropriate sampling
Linus Torvalds1da177e2005-04-16 15:20:36 -070046 * rate.
47 * For CPUs with transition latency > 10mS (mostly drivers with CPUFREQ_ETERNAL)
48 * this governor will not work.
49 * All times here are in uS.
50 */
Dave Jonesdf8b59b2005-09-20 12:39:35 -070051#define MIN_SAMPLING_RATE_RATIO (2)
Thomas Renninger112124a2009-02-04 11:55:12 +010052
Thomas Renningercef96152009-04-22 13:48:29 +020053static unsigned int min_sampling_rate;
54
Thomas Renninger112124a2009-02-04 11:55:12 +010055#define LATENCY_MULTIPLIER (1000)
Thomas Renningercef96152009-04-22 13:48:29 +020056#define MIN_LATENCY_MULTIPLIER (100)
Thomas Renninger1c256242007-10-02 13:28:12 -070057#define TRANSITION_LATENCY_LIMIT (10 * 1000 * 1000)
Linus Torvalds1da177e2005-04-16 15:20:36 -070058
David Howellsc4028952006-11-22 14:57:56 +000059static void do_dbs_timer(struct work_struct *work);
Thomas Renninger0e625ac2009-07-24 15:25:06 +020060static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
61 unsigned int event);
62
63#ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
64static
65#endif
66struct cpufreq_governor cpufreq_gov_ondemand = {
67 .name = "ondemand",
68 .governor = cpufreq_governor_dbs,
69 .max_transition_latency = TRANSITION_LATENCY_LIMIT,
70 .owner = THIS_MODULE,
71};
David Howellsc4028952006-11-22 14:57:56 +000072
73/* Sampling types */
Venkatesh Pallipadi529af7a2007-02-05 16:12:44 -080074enum {DBS_NORMAL_SAMPLE, DBS_SUB_SAMPLE};
Linus Torvalds1da177e2005-04-16 15:20:36 -070075
76struct cpu_dbs_info_s {
Venkatesh Pallipadiccb2fe22006-06-28 13:49:52 -070077 cputime64_t prev_cpu_idle;
Arjan van de Ven6b8fcd92010-05-09 08:26:06 -070078 cputime64_t prev_cpu_iowait;
Venkatesh Pallipadiccb2fe22006-06-28 13:49:52 -070079 cputime64_t prev_cpu_wall;
venkatesh.pallipadi@intel.com80800912008-08-04 11:59:12 -070080 cputime64_t prev_cpu_nice;
Dave Jones32ee8c32006-02-28 00:43:23 -050081 struct cpufreq_policy *cur_policy;
Dave Jones2b03f892009-01-18 01:43:44 -050082 struct delayed_work work;
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +040083 struct cpufreq_frequency_table *freq_table;
84 unsigned int freq_lo;
85 unsigned int freq_lo_jiffies;
86 unsigned int freq_hi_jiffies;
David C Niemi3f78a9f2010-10-06 16:54:24 -040087 unsigned int rate_mult;
Venkatesh Pallipadi529af7a2007-02-05 16:12:44 -080088 int cpu;
venkatesh.pallipadi@intel.com5a75c822009-07-02 17:08:32 -070089 unsigned int sample_type:1;
90 /*
91 * percpu mutex that serializes governor limit change with
92 * do_dbs_timer invocation. We do not want do_dbs_timer to run
93 * when user is changing the governor or limits.
94 */
95 struct mutex timer_mutex;
Linus Torvalds1da177e2005-04-16 15:20:36 -070096};
Tejun Heo245b2e72009-06-24 15:13:48 +090097static DEFINE_PER_CPU(struct cpu_dbs_info_s, od_cpu_dbs_info);
Linus Torvalds1da177e2005-04-16 15:20:36 -070098
99static unsigned int dbs_enable; /* number of CPUs using this policy */
100
Venkatesh Pallipadi4ec223d2006-06-21 15:18:34 -0700101/*
Thomas Renninger326c86d2011-03-03 21:31:27 +0100102 * dbs_mutex protects dbs_enable in governor start/stop.
Venkatesh Pallipadi4ec223d2006-06-21 15:18:34 -0700103 */
Venkatesh Pallipadiffac80e2006-06-28 13:52:18 -0700104static DEFINE_MUTEX(dbs_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700105
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400106static struct dbs_tuners {
Dave Jones32ee8c32006-02-28 00:43:23 -0500107 unsigned int sampling_rate;
Dave Jones32ee8c32006-02-28 00:43:23 -0500108 unsigned int up_threshold;
venkatesh.pallipadi@intel.come9d95bf2008-08-04 11:59:10 -0700109 unsigned int down_differential;
Dave Jones32ee8c32006-02-28 00:43:23 -0500110 unsigned int ignore_nice;
David C Niemi3f78a9f2010-10-06 16:54:24 -0400111 unsigned int sampling_down_factor;
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400112 unsigned int powersave_bias;
Arjan van de Ven19379b12010-05-09 08:26:51 -0700113 unsigned int io_is_busy;
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400114} dbs_tuners_ins = {
Dave Jones32ee8c32006-02-28 00:43:23 -0500115 .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
David C Niemi3f78a9f2010-10-06 16:54:24 -0400116 .sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR,
venkatesh.pallipadi@intel.come9d95bf2008-08-04 11:59:10 -0700117 .down_differential = DEF_FREQUENCY_DOWN_DIFFERENTIAL,
Eric Piel9cbad612006-03-10 11:35:27 +0200118 .ignore_nice = 0,
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400119 .powersave_bias = 0,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700120};
121
venkatesh.pallipadi@intel.com80800912008-08-04 11:59:12 -0700122static inline cputime64_t get_cpu_idle_time_jiffy(unsigned int cpu,
123 cputime64_t *wall)
Dave Jonesdac1c1a2005-05-31 19:03:49 -0700124{
Venki Pallipadiea487612007-06-20 14:26:24 -0700125 cputime64_t idle_time;
venkatesh.pallipadi@intel.com34305022008-08-04 11:59:09 -0700126 cputime64_t cur_wall_time;
Venki Pallipadiea487612007-06-20 14:26:24 -0700127 cputime64_t busy_time;
Venkatesh Pallipadiccb2fe22006-06-28 13:49:52 -0700128
venkatesh.pallipadi@intel.com34305022008-08-04 11:59:09 -0700129 cur_wall_time = jiffies64_to_cputime64(get_jiffies_64());
Venki Pallipadiea487612007-06-20 14:26:24 -0700130 busy_time = cputime64_add(kstat_cpu(cpu).cpustat.user,
131 kstat_cpu(cpu).cpustat.system);
Venkatesh Pallipadiccb2fe22006-06-28 13:49:52 -0700132
Venki Pallipadiea487612007-06-20 14:26:24 -0700133 busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.irq);
134 busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.softirq);
135 busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.steal);
Venkatesh Pallipadi1ca3abd2009-01-23 09:25:02 -0500136 busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.nice);
Venki Pallipadiea487612007-06-20 14:26:24 -0700137
venkatesh.pallipadi@intel.com34305022008-08-04 11:59:09 -0700138 idle_time = cputime64_sub(cur_wall_time, busy_time);
139 if (wall)
Pallipadi, Venkatesh54c9a352009-11-11 16:50:29 -0800140 *wall = (cputime64_t)jiffies_to_usecs(cur_wall_time);
venkatesh.pallipadi@intel.com34305022008-08-04 11:59:09 -0700141
Pallipadi, Venkatesh54c9a352009-11-11 16:50:29 -0800142 return (cputime64_t)jiffies_to_usecs(idle_time);
Dave Jonesdac1c1a2005-05-31 19:03:49 -0700143}
144
venkatesh.pallipadi@intel.com80800912008-08-04 11:59:12 -0700145static inline cputime64_t get_cpu_idle_time(unsigned int cpu, cputime64_t *wall)
146{
Michal Hocko6beea0c2011-08-24 09:37:48 +0200147 u64 idle_time = get_cpu_idle_time_us(cpu, NULL);
venkatesh.pallipadi@intel.com80800912008-08-04 11:59:12 -0700148
149 if (idle_time == -1ULL)
150 return get_cpu_idle_time_jiffy(cpu, wall);
Michal Hocko6beea0c2011-08-24 09:37:48 +0200151 else
152 idle_time += get_cpu_iowait_time_us(cpu, wall);
venkatesh.pallipadi@intel.com80800912008-08-04 11:59:12 -0700153
venkatesh.pallipadi@intel.com80800912008-08-04 11:59:12 -0700154 return idle_time;
155}
156
Arjan van de Ven6b8fcd92010-05-09 08:26:06 -0700157static inline cputime64_t get_cpu_iowait_time(unsigned int cpu, cputime64_t *wall)
158{
159 u64 iowait_time = get_cpu_iowait_time_us(cpu, wall);
160
161 if (iowait_time == -1ULL)
162 return 0;
163
164 return iowait_time;
165}
166
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400167/*
168 * Find right freq to be set now with powersave_bias on.
169 * Returns the freq_hi to be used right now and will set freq_hi_jiffies,
170 * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs.
171 */
Adrian Bunkb5ecf602006-08-13 23:00:08 +0200172static unsigned int powersave_bias_target(struct cpufreq_policy *policy,
173 unsigned int freq_next,
174 unsigned int relation)
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400175{
176 unsigned int freq_req, freq_reduc, freq_avg;
177 unsigned int freq_hi, freq_lo;
178 unsigned int index = 0;
179 unsigned int jiffies_total, jiffies_hi, jiffies_lo;
Tejun Heo245b2e72009-06-24 15:13:48 +0900180 struct cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
181 policy->cpu);
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400182
183 if (!dbs_info->freq_table) {
184 dbs_info->freq_lo = 0;
185 dbs_info->freq_lo_jiffies = 0;
186 return freq_next;
187 }
188
189 cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next,
190 relation, &index);
191 freq_req = dbs_info->freq_table[index].frequency;
192 freq_reduc = freq_req * dbs_tuners_ins.powersave_bias / 1000;
193 freq_avg = freq_req - freq_reduc;
194
195 /* Find freq bounds for freq_avg in freq_table */
196 index = 0;
197 cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
198 CPUFREQ_RELATION_H, &index);
199 freq_lo = dbs_info->freq_table[index].frequency;
200 index = 0;
201 cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
202 CPUFREQ_RELATION_L, &index);
203 freq_hi = dbs_info->freq_table[index].frequency;
204
205 /* Find out how long we have to be in hi and lo freqs */
206 if (freq_hi == freq_lo) {
207 dbs_info->freq_lo = 0;
208 dbs_info->freq_lo_jiffies = 0;
209 return freq_lo;
210 }
211 jiffies_total = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
212 jiffies_hi = (freq_avg - freq_lo) * jiffies_total;
213 jiffies_hi += ((freq_hi - freq_lo) / 2);
214 jiffies_hi /= (freq_hi - freq_lo);
215 jiffies_lo = jiffies_total - jiffies_hi;
216 dbs_info->freq_lo = freq_lo;
217 dbs_info->freq_lo_jiffies = jiffies_lo;
218 dbs_info->freq_hi_jiffies = jiffies_hi;
219 return freq_hi;
220}
221
venkatesh.pallipadi@intel.com5a75c822009-07-02 17:08:32 -0700222static void ondemand_powersave_bias_init_cpu(int cpu)
223{
Tejun Heo384be2b2009-08-14 14:41:02 +0900224 struct cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
venkatesh.pallipadi@intel.com5a75c822009-07-02 17:08:32 -0700225 dbs_info->freq_table = cpufreq_frequency_get_table(cpu);
226 dbs_info->freq_lo = 0;
227}
228
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400229static void ondemand_powersave_bias_init(void)
230{
231 int i;
232 for_each_online_cpu(i) {
venkatesh.pallipadi@intel.com5a75c822009-07-02 17:08:32 -0700233 ondemand_powersave_bias_init_cpu(i);
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400234 }
235}
236
Linus Torvalds1da177e2005-04-16 15:20:36 -0700237/************************** sysfs interface ************************/
Thomas Renninger0e625ac2009-07-24 15:25:06 +0200238
Thomas Renninger0e625ac2009-07-24 15:25:06 +0200239static ssize_t show_sampling_rate_min(struct kobject *kobj,
240 struct attribute *attr, char *buf)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700241{
Thomas Renningercef96152009-04-22 13:48:29 +0200242 return sprintf(buf, "%u\n", min_sampling_rate);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700243}
244
Borislav Petkov6dad2a22010-03-31 21:56:46 +0200245define_one_global_ro(sampling_rate_min);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700246
247/* cpufreq_ondemand Governor Tunables */
248#define show_one(file_name, object) \
249static ssize_t show_##file_name \
Thomas Renninger0e625ac2009-07-24 15:25:06 +0200250(struct kobject *kobj, struct attribute *attr, char *buf) \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700251{ \
252 return sprintf(buf, "%u\n", dbs_tuners_ins.object); \
253}
254show_one(sampling_rate, sampling_rate);
Arjan van de Ven19379b12010-05-09 08:26:51 -0700255show_one(io_is_busy, io_is_busy);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700256show_one(up_threshold, up_threshold);
David C Niemi3f78a9f2010-10-06 16:54:24 -0400257show_one(sampling_down_factor, sampling_down_factor);
Alexander Clouter001893c2005-12-01 01:09:25 -0800258show_one(ignore_nice_load, ignore_nice);
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400259show_one(powersave_bias, powersave_bias);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700260
Thomas Renninger0e625ac2009-07-24 15:25:06 +0200261static ssize_t store_sampling_rate(struct kobject *a, struct attribute *b,
262 const char *buf, size_t count)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700263{
264 unsigned int input;
265 int ret;
Venkatesh Pallipadiffac80e2006-06-28 13:52:18 -0700266 ret = sscanf(buf, "%u", &input);
venkatesh.pallipadi@intel.com5a75c822009-07-02 17:08:32 -0700267 if (ret != 1)
268 return -EINVAL;
Thomas Renningercef96152009-04-22 13:48:29 +0200269 dbs_tuners_ins.sampling_rate = max(input, min_sampling_rate);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700270 return count;
271}
272
Arjan van de Ven19379b12010-05-09 08:26:51 -0700273static ssize_t store_io_is_busy(struct kobject *a, struct attribute *b,
274 const char *buf, size_t count)
275{
276 unsigned int input;
277 int ret;
278
279 ret = sscanf(buf, "%u", &input);
280 if (ret != 1)
281 return -EINVAL;
Arjan van de Ven19379b12010-05-09 08:26:51 -0700282 dbs_tuners_ins.io_is_busy = !!input;
Arjan van de Ven19379b12010-05-09 08:26:51 -0700283 return count;
284}
285
Thomas Renninger0e625ac2009-07-24 15:25:06 +0200286static ssize_t store_up_threshold(struct kobject *a, struct attribute *b,
287 const char *buf, size_t count)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700288{
289 unsigned int input;
290 int ret;
Venkatesh Pallipadiffac80e2006-06-28 13:52:18 -0700291 ret = sscanf(buf, "%u", &input);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700292
Dave Jones32ee8c32006-02-28 00:43:23 -0500293 if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
Dave Jonesc29f1402005-05-31 19:03:50 -0700294 input < MIN_FREQUENCY_UP_THRESHOLD) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700295 return -EINVAL;
296 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700297 dbs_tuners_ins.up_threshold = input;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700298 return count;
299}
300
David C Niemi3f78a9f2010-10-06 16:54:24 -0400301static ssize_t store_sampling_down_factor(struct kobject *a,
302 struct attribute *b, const char *buf, size_t count)
303{
304 unsigned int input, j;
305 int ret;
306 ret = sscanf(buf, "%u", &input);
307
308 if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
309 return -EINVAL;
David C Niemi3f78a9f2010-10-06 16:54:24 -0400310 dbs_tuners_ins.sampling_down_factor = input;
311
312 /* Reset down sampling multiplier in case it was active */
313 for_each_online_cpu(j) {
314 struct cpu_dbs_info_s *dbs_info;
315 dbs_info = &per_cpu(od_cpu_dbs_info, j);
316 dbs_info->rate_mult = 1;
317 }
David C Niemi3f78a9f2010-10-06 16:54:24 -0400318 return count;
319}
320
Thomas Renninger0e625ac2009-07-24 15:25:06 +0200321static ssize_t store_ignore_nice_load(struct kobject *a, struct attribute *b,
322 const char *buf, size_t count)
Dave Jones3d5ee9e2005-05-31 19:03:47 -0700323{
324 unsigned int input;
325 int ret;
326
327 unsigned int j;
Dave Jones32ee8c32006-02-28 00:43:23 -0500328
Venkatesh Pallipadiffac80e2006-06-28 13:52:18 -0700329 ret = sscanf(buf, "%u", &input);
Dave Jones2b03f892009-01-18 01:43:44 -0500330 if (ret != 1)
Dave Jones3d5ee9e2005-05-31 19:03:47 -0700331 return -EINVAL;
332
Dave Jones2b03f892009-01-18 01:43:44 -0500333 if (input > 1)
Dave Jones3d5ee9e2005-05-31 19:03:47 -0700334 input = 1;
Dave Jones32ee8c32006-02-28 00:43:23 -0500335
Dave Jones2b03f892009-01-18 01:43:44 -0500336 if (input == dbs_tuners_ins.ignore_nice) { /* nothing to do */
Dave Jones3d5ee9e2005-05-31 19:03:47 -0700337 return count;
338 }
339 dbs_tuners_ins.ignore_nice = input;
340
Venkatesh Pallipadiccb2fe22006-06-28 13:49:52 -0700341 /* we need to re-evaluate prev_cpu_idle */
Dave Jonesdac1c1a2005-05-31 19:03:49 -0700342 for_each_online_cpu(j) {
Venkatesh Pallipadiccb2fe22006-06-28 13:49:52 -0700343 struct cpu_dbs_info_s *dbs_info;
Tejun Heo245b2e72009-06-24 15:13:48 +0900344 dbs_info = &per_cpu(od_cpu_dbs_info, j);
venkatesh.pallipadi@intel.com34305022008-08-04 11:59:09 -0700345 dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
346 &dbs_info->prev_cpu_wall);
Venkatesh Pallipadi1ca3abd2009-01-23 09:25:02 -0500347 if (dbs_tuners_ins.ignore_nice)
348 dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
349
Dave Jones3d5ee9e2005-05-31 19:03:47 -0700350 }
Dave Jones3d5ee9e2005-05-31 19:03:47 -0700351 return count;
352}
353
Thomas Renninger0e625ac2009-07-24 15:25:06 +0200354static ssize_t store_powersave_bias(struct kobject *a, struct attribute *b,
355 const char *buf, size_t count)
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400356{
357 unsigned int input;
358 int ret;
359 ret = sscanf(buf, "%u", &input);
360
361 if (ret != 1)
362 return -EINVAL;
363
364 if (input > 1000)
365 input = 1000;
366
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400367 dbs_tuners_ins.powersave_bias = input;
368 ondemand_powersave_bias_init();
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400369 return count;
370}
371
Borislav Petkov6dad2a22010-03-31 21:56:46 +0200372define_one_global_rw(sampling_rate);
Linus Torvalds07d77752010-05-18 08:49:13 -0700373define_one_global_rw(io_is_busy);
Borislav Petkov6dad2a22010-03-31 21:56:46 +0200374define_one_global_rw(up_threshold);
David C Niemi3f78a9f2010-10-06 16:54:24 -0400375define_one_global_rw(sampling_down_factor);
Borislav Petkov6dad2a22010-03-31 21:56:46 +0200376define_one_global_rw(ignore_nice_load);
377define_one_global_rw(powersave_bias);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700378
Dave Jones2b03f892009-01-18 01:43:44 -0500379static struct attribute *dbs_attributes[] = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700380 &sampling_rate_min.attr,
381 &sampling_rate.attr,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700382 &up_threshold.attr,
David C Niemi3f78a9f2010-10-06 16:54:24 -0400383 &sampling_down_factor.attr,
Alexander Clouter001893c2005-12-01 01:09:25 -0800384 &ignore_nice_load.attr,
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400385 &powersave_bias.attr,
Arjan van de Ven19379b12010-05-09 08:26:51 -0700386 &io_is_busy.attr,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700387 NULL
388};
389
390static struct attribute_group dbs_attr_group = {
391 .attrs = dbs_attributes,
392 .name = "ondemand",
393};
394
395/************************** sysfs end ************************/
396
Mike Chan00e299f2010-01-26 17:06:47 -0800397static void dbs_freq_increase(struct cpufreq_policy *p, unsigned int freq)
398{
399 if (dbs_tuners_ins.powersave_bias)
400 freq = powersave_bias_target(p, freq, CPUFREQ_RELATION_H);
401 else if (p->cur == p->max)
402 return;
403
404 __cpufreq_driver_target(p, freq, dbs_tuners_ins.powersave_bias ?
405 CPUFREQ_RELATION_L : CPUFREQ_RELATION_H);
406}
407
Venkatesh Pallipadi2f8a8352006-06-28 13:51:19 -0700408static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700409{
venkatesh.pallipadi@intel.comc43aa3b2008-08-04 11:59:08 -0700410 unsigned int max_load_freq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700411
412 struct cpufreq_policy *policy;
413 unsigned int j;
414
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400415 this_dbs_info->freq_lo = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416 policy = this_dbs_info->cur_policy;
Venki Pallipadiea487612007-06-20 14:26:24 -0700417
Dave Jones32ee8c32006-02-28 00:43:23 -0500418 /*
Dave Jonesc29f1402005-05-31 19:03:50 -0700419 * Every sampling_rate, we check, if current idle time is less
420 * than 20% (default), then we try to increase frequency
Venkatesh Pallipadiccb2fe22006-06-28 13:49:52 -0700421 * Every sampling_rate, we look for a the lowest
Dave Jonesc29f1402005-05-31 19:03:50 -0700422 * frequency which can sustain the load while keeping idle time over
423 * 30%. If such a frequency exist, we try to decrease to this frequency.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700424 *
Dave Jones32ee8c32006-02-28 00:43:23 -0500425 * Any frequency increase takes it to the maximum frequency.
426 * Frequency reduction happens at minimum steps of
427 * 5% (default) of current frequency
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428 */
429
venkatesh.pallipadi@intel.comc43aa3b2008-08-04 11:59:08 -0700430 /* Get Absolute Load - in terms of freq */
431 max_load_freq = 0;
432
Rusty Russell835481d2009-01-04 05:18:06 -0800433 for_each_cpu(j, policy->cpus) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700434 struct cpu_dbs_info_s *j_dbs_info;
Arjan van de Ven6b8fcd92010-05-09 08:26:06 -0700435 cputime64_t cur_wall_time, cur_idle_time, cur_iowait_time;
436 unsigned int idle_time, wall_time, iowait_time;
venkatesh.pallipadi@intel.comc43aa3b2008-08-04 11:59:08 -0700437 unsigned int load, load_freq;
438 int freq_avg;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700439
Tejun Heo245b2e72009-06-24 15:13:48 +0900440 j_dbs_info = &per_cpu(od_cpu_dbs_info, j);
venkatesh.pallipadi@intel.com34305022008-08-04 11:59:09 -0700441
442 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time);
Arjan van de Ven6b8fcd92010-05-09 08:26:06 -0700443 cur_iowait_time = get_cpu_iowait_time(j, &cur_wall_time);
venkatesh.pallipadi@intel.com34305022008-08-04 11:59:09 -0700444
venkatesh.pallipadi@intel.comc43aa3b2008-08-04 11:59:08 -0700445 wall_time = (unsigned int) cputime64_sub(cur_wall_time,
446 j_dbs_info->prev_cpu_wall);
447 j_dbs_info->prev_cpu_wall = cur_wall_time;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700448
venkatesh.pallipadi@intel.comc43aa3b2008-08-04 11:59:08 -0700449 idle_time = (unsigned int) cputime64_sub(cur_idle_time,
450 j_dbs_info->prev_cpu_idle);
451 j_dbs_info->prev_cpu_idle = cur_idle_time;
452
Arjan van de Ven6b8fcd92010-05-09 08:26:06 -0700453 iowait_time = (unsigned int) cputime64_sub(cur_iowait_time,
454 j_dbs_info->prev_cpu_iowait);
455 j_dbs_info->prev_cpu_iowait = cur_iowait_time;
456
Venkatesh Pallipadi1ca3abd2009-01-23 09:25:02 -0500457 if (dbs_tuners_ins.ignore_nice) {
458 cputime64_t cur_nice;
459 unsigned long cur_nice_jiffies;
460
461 cur_nice = cputime64_sub(kstat_cpu(j).cpustat.nice,
462 j_dbs_info->prev_cpu_nice);
463 /*
464 * Assumption: nice time between sampling periods will
465 * be less than 2^32 jiffies for 32 bit sys
466 */
467 cur_nice_jiffies = (unsigned long)
468 cputime64_to_jiffies64(cur_nice);
469
470 j_dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
471 idle_time += jiffies_to_usecs(cur_nice_jiffies);
472 }
473
Arjan van de Ven6b8fcd92010-05-09 08:26:06 -0700474 /*
475 * For the purpose of ondemand, waiting for disk IO is an
476 * indication that you're performance critical, and not that
477 * the system is actually idle. So subtract the iowait time
478 * from the cpu idle time.
479 */
480
Arjan van de Ven19379b12010-05-09 08:26:51 -0700481 if (dbs_tuners_ins.io_is_busy && idle_time >= iowait_time)
Arjan van de Ven6b8fcd92010-05-09 08:26:06 -0700482 idle_time -= iowait_time;
483
venkatesh.pallipadi@intel.com34305022008-08-04 11:59:09 -0700484 if (unlikely(!wall_time || wall_time < idle_time))
venkatesh.pallipadi@intel.comc43aa3b2008-08-04 11:59:08 -0700485 continue;
venkatesh.pallipadi@intel.comc43aa3b2008-08-04 11:59:08 -0700486
487 load = 100 * (wall_time - idle_time) / wall_time;
488
489 freq_avg = __cpufreq_driver_getavg(policy, j);
490 if (freq_avg <= 0)
491 freq_avg = policy->cur;
492
493 load_freq = load * freq_avg;
494 if (load_freq > max_load_freq)
495 max_load_freq = load_freq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700496 }
497
Venkatesh Pallipadiccb2fe22006-06-28 13:49:52 -0700498 /* Check for frequency increase */
venkatesh.pallipadi@intel.comc43aa3b2008-08-04 11:59:08 -0700499 if (max_load_freq > dbs_tuners_ins.up_threshold * policy->cur) {
David C Niemi3f78a9f2010-10-06 16:54:24 -0400500 /* If switching to max speed, apply sampling_down_factor */
501 if (policy->cur < policy->max)
502 this_dbs_info->rate_mult =
503 dbs_tuners_ins.sampling_down_factor;
Mike Chan00e299f2010-01-26 17:06:47 -0800504 dbs_freq_increase(policy, policy->max);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700505 return;
506 }
507
508 /* Check for frequency decrease */
Dave Jonesc29f1402005-05-31 19:03:50 -0700509 /* if we cannot reduce the frequency anymore, break out early */
510 if (policy->cur == policy->min)
511 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700512
Dave Jonesc29f1402005-05-31 19:03:50 -0700513 /*
514 * The optimal frequency is the frequency that is the lowest that
515 * can support the current CPU usage without triggering the up
516 * policy. To be safe, we focus 10 points under the threshold.
517 */
venkatesh.pallipadi@intel.come9d95bf2008-08-04 11:59:10 -0700518 if (max_load_freq <
519 (dbs_tuners_ins.up_threshold - dbs_tuners_ins.down_differential) *
520 policy->cur) {
venkatesh.pallipadi@intel.comc43aa3b2008-08-04 11:59:08 -0700521 unsigned int freq_next;
venkatesh.pallipadi@intel.come9d95bf2008-08-04 11:59:10 -0700522 freq_next = max_load_freq /
523 (dbs_tuners_ins.up_threshold -
524 dbs_tuners_ins.down_differential);
Venkatesh Pallipadidfde5d62006-10-03 12:38:45 -0700525
David C Niemi3f78a9f2010-10-06 16:54:24 -0400526 /* No longer fully busy, reset rate_mult */
527 this_dbs_info->rate_mult = 1;
528
Nagananda.Chumbalkar@hp.com1dbf5882009-12-21 23:40:52 +0100529 if (freq_next < policy->min)
530 freq_next = policy->min;
531
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400532 if (!dbs_tuners_ins.powersave_bias) {
533 __cpufreq_driver_target(policy, freq_next,
534 CPUFREQ_RELATION_L);
535 } else {
536 int freq = powersave_bias_target(policy, freq_next,
537 CPUFREQ_RELATION_L);
538 __cpufreq_driver_target(policy, freq,
539 CPUFREQ_RELATION_L);
540 }
Venkatesh Pallipadiccb2fe22006-06-28 13:49:52 -0700541 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700542}
543
David Howellsc4028952006-11-22 14:57:56 +0000544static void do_dbs_timer(struct work_struct *work)
Dave Jones32ee8c32006-02-28 00:43:23 -0500545{
Venkatesh Pallipadi529af7a2007-02-05 16:12:44 -0800546 struct cpu_dbs_info_s *dbs_info =
547 container_of(work, struct cpu_dbs_info_s, work.work);
548 unsigned int cpu = dbs_info->cpu;
549 int sample_type = dbs_info->sample_type;
550
Vincent Guittot5cb2c3b2011-02-07 17:14:25 +0100551 int delay;
Jocelyn Falempea665df92010-03-11 14:01:11 -0800552
venkatesh.pallipadi@intel.com5a75c822009-07-02 17:08:32 -0700553 mutex_lock(&dbs_info->timer_mutex);
Venkatesh Pallipadi56463b72007-02-05 16:12:45 -0800554
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400555 /* Common NORMAL_SAMPLE setup */
David Howellsc4028952006-11-22 14:57:56 +0000556 dbs_info->sample_type = DBS_NORMAL_SAMPLE;
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400557 if (!dbs_tuners_ins.powersave_bias ||
David Howellsc4028952006-11-22 14:57:56 +0000558 sample_type == DBS_NORMAL_SAMPLE) {
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400559 dbs_check_cpu(dbs_info);
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400560 if (dbs_info->freq_lo) {
561 /* Setup timer for SUB_SAMPLE */
David Howellsc4028952006-11-22 14:57:56 +0000562 dbs_info->sample_type = DBS_SUB_SAMPLE;
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400563 delay = dbs_info->freq_hi_jiffies;
Vincent Guittot5cb2c3b2011-02-07 17:14:25 +0100564 } else {
565 /* We want all CPUs to do sampling nearly on
566 * same jiffy
567 */
568 delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate
569 * dbs_info->rate_mult);
570
571 if (num_online_cpus() > 1)
572 delay -= jiffies % delay;
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400573 }
574 } else {
575 __cpufreq_driver_target(dbs_info->cur_policy,
Dave Jones2b03f892009-01-18 01:43:44 -0500576 dbs_info->freq_lo, CPUFREQ_RELATION_H);
Vincent Guittot5cb2c3b2011-02-07 17:14:25 +0100577 delay = dbs_info->freq_lo_jiffies;
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400578 }
Tejun Heo57df5572011-01-26 12:12:50 +0100579 schedule_delayed_work_on(cpu, &dbs_info->work, delay);
venkatesh.pallipadi@intel.com5a75c822009-07-02 17:08:32 -0700580 mutex_unlock(&dbs_info->timer_mutex);
Dave Jones32ee8c32006-02-28 00:43:23 -0500581}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700582
Venkatesh Pallipadi529af7a2007-02-05 16:12:44 -0800583static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700584{
Alexey Starikovskiy1ce28d62006-07-31 22:25:20 +0400585 /* We want all CPUs to do sampling nearly on same jiffy */
586 int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
Jocelyn Falempea665df92010-03-11 14:01:11 -0800587
588 if (num_online_cpus() > 1)
589 delay -= jiffies % delay;
Venkatesh Pallipadi2f8a8352006-06-28 13:51:19 -0700590
David Howellsc4028952006-11-22 14:57:56 +0000591 dbs_info->sample_type = DBS_NORMAL_SAMPLE;
Venki Pallipadi28287032007-05-08 00:27:47 -0700592 INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer);
Tejun Heo57df5572011-01-26 12:12:50 +0100593 schedule_delayed_work_on(dbs_info->cpu, &dbs_info->work, delay);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700594}
595
Linus Torvalds2cd7cbd2006-07-23 12:05:00 -0700596static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700597{
Mathieu Desnoyersb14893a2009-05-17 10:30:45 -0400598 cancel_delayed_work_sync(&dbs_info->work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700599}
600
Arjan van de Ven19379b12010-05-09 08:26:51 -0700601/*
602 * Not all CPUs want IO time to be accounted as busy; this dependson how
603 * efficient idling at a higher frequency/voltage is.
604 * Pavel Machek says this is not so for various generations of AMD and old
605 * Intel systems.
606 * Mike Chan (androidlcom) calis this is also not true for ARM.
607 * Because of this, whitelist specific known (series) of CPUs by default, and
608 * leave all others up to the user.
609 */
610static int should_io_be_busy(void)
611{
612#if defined(CONFIG_X86)
613 /*
614 * For Intel, Core 2 (model 15) andl later have an efficient idle.
615 */
616 if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
617 boot_cpu_data.x86 == 6 &&
618 boot_cpu_data.x86_model >= 15)
619 return 1;
620#endif
621 return 0;
622}
623
Linus Torvalds1da177e2005-04-16 15:20:36 -0700624static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
625 unsigned int event)
626{
627 unsigned int cpu = policy->cpu;
628 struct cpu_dbs_info_s *this_dbs_info;
629 unsigned int j;
Jeff Garzik914f7c32006-10-20 14:31:00 -0700630 int rc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700631
Tejun Heo245b2e72009-06-24 15:13:48 +0900632 this_dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700633
634 switch (event) {
635 case CPUFREQ_GOV_START:
Venkatesh Pallipadiffac80e2006-06-28 13:52:18 -0700636 if ((!cpu_online(cpu)) || (!policy->cur))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700637 return -EINVAL;
638
akpm@osdl.org3fc54d32006-01-13 15:54:22 -0800639 mutex_lock(&dbs_mutex);
Jeff Garzik914f7c32006-10-20 14:31:00 -0700640
venkatesh.pallipadi@intel.com5a75c822009-07-02 17:08:32 -0700641 dbs_enable++;
Rusty Russell835481d2009-01-04 05:18:06 -0800642 for_each_cpu(j, policy->cpus) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700643 struct cpu_dbs_info_s *j_dbs_info;
Tejun Heo245b2e72009-06-24 15:13:48 +0900644 j_dbs_info = &per_cpu(od_cpu_dbs_info, j);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700645 j_dbs_info->cur_policy = policy;
Dave Jones32ee8c32006-02-28 00:43:23 -0500646
venkatesh.pallipadi@intel.com34305022008-08-04 11:59:09 -0700647 j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
648 &j_dbs_info->prev_cpu_wall);
Venkatesh Pallipadi1ca3abd2009-01-23 09:25:02 -0500649 if (dbs_tuners_ins.ignore_nice) {
650 j_dbs_info->prev_cpu_nice =
651 kstat_cpu(j).cpustat.nice;
652 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700653 }
Venkatesh Pallipadi529af7a2007-02-05 16:12:44 -0800654 this_dbs_info->cpu = cpu;
David C Niemi3f78a9f2010-10-06 16:54:24 -0400655 this_dbs_info->rate_mult = 1;
venkatesh.pallipadi@intel.com5a75c822009-07-02 17:08:32 -0700656 ondemand_powersave_bias_init_cpu(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700657 /*
658 * Start the timerschedule work, when this governor
659 * is used for first time
660 */
661 if (dbs_enable == 1) {
662 unsigned int latency;
Thomas Renninger0e625ac2009-07-24 15:25:06 +0200663
664 rc = sysfs_create_group(cpufreq_global_kobject,
665 &dbs_attr_group);
666 if (rc) {
667 mutex_unlock(&dbs_mutex);
668 return rc;
669 }
670
Linus Torvalds1da177e2005-04-16 15:20:36 -0700671 /* policy latency is in nS. Convert it to uS first */
Dave Jonesdf8b59b2005-09-20 12:39:35 -0700672 latency = policy->cpuinfo.transition_latency / 1000;
673 if (latency == 0)
674 latency = 1;
Thomas Renningercef96152009-04-22 13:48:29 +0200675 /* Bring kernel and HW constraints together */
676 min_sampling_rate = max(min_sampling_rate,
677 MIN_LATENCY_MULTIPLIER * latency);
678 dbs_tuners_ins.sampling_rate =
679 max(min_sampling_rate,
680 latency * LATENCY_MULTIPLIER);
Arjan van de Ven19379b12010-05-09 08:26:51 -0700681 dbs_tuners_ins.io_is_busy = should_io_be_busy();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700682 }
akpm@osdl.org3fc54d32006-01-13 15:54:22 -0800683 mutex_unlock(&dbs_mutex);
venkatesh.pallipadi@intel.com7d26e2d2009-07-02 17:08:30 -0700684
Thomas Renninger0e625ac2009-07-24 15:25:06 +0200685 mutex_init(&this_dbs_info->timer_mutex);
venkatesh.pallipadi@intel.com7d26e2d2009-07-02 17:08:30 -0700686 dbs_timer_init(this_dbs_info);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700687 break;
688
689 case CPUFREQ_GOV_STOP:
Linus Torvalds2cd7cbd2006-07-23 12:05:00 -0700690 dbs_timer_exit(this_dbs_info);
venkatesh.pallipadi@intel.com7d26e2d2009-07-02 17:08:30 -0700691
692 mutex_lock(&dbs_mutex);
venkatesh.pallipadi@intel.com5a75c822009-07-02 17:08:32 -0700693 mutex_destroy(&this_dbs_info->timer_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700694 dbs_enable--;
akpm@osdl.org3fc54d32006-01-13 15:54:22 -0800695 mutex_unlock(&dbs_mutex);
Thomas Renninger0e625ac2009-07-24 15:25:06 +0200696 if (!dbs_enable)
697 sysfs_remove_group(cpufreq_global_kobject,
698 &dbs_attr_group);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700699
700 break;
701
702 case CPUFREQ_GOV_LIMITS:
venkatesh.pallipadi@intel.com5a75c822009-07-02 17:08:32 -0700703 mutex_lock(&this_dbs_info->timer_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700704 if (policy->max < this_dbs_info->cur_policy->cur)
Venkatesh Pallipadiffac80e2006-06-28 13:52:18 -0700705 __cpufreq_driver_target(this_dbs_info->cur_policy,
Dave Jones2b03f892009-01-18 01:43:44 -0500706 policy->max, CPUFREQ_RELATION_H);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700707 else if (policy->min > this_dbs_info->cur_policy->cur)
Venkatesh Pallipadiffac80e2006-06-28 13:52:18 -0700708 __cpufreq_driver_target(this_dbs_info->cur_policy,
Dave Jones2b03f892009-01-18 01:43:44 -0500709 policy->min, CPUFREQ_RELATION_L);
venkatesh.pallipadi@intel.com5a75c822009-07-02 17:08:32 -0700710 mutex_unlock(&this_dbs_info->timer_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700711 break;
712 }
713 return 0;
714}
715
Linus Torvalds1da177e2005-04-16 15:20:36 -0700716static int __init cpufreq_gov_dbs_init(void)
717{
venkatesh.pallipadi@intel.com80800912008-08-04 11:59:12 -0700718 cputime64_t wall;
Andrea Righi4f6e6b92008-09-18 10:43:40 +0000719 u64 idle_time;
720 int cpu = get_cpu();
venkatesh.pallipadi@intel.com80800912008-08-04 11:59:12 -0700721
Andrea Righi4f6e6b92008-09-18 10:43:40 +0000722 idle_time = get_cpu_idle_time_us(cpu, &wall);
723 put_cpu();
venkatesh.pallipadi@intel.com80800912008-08-04 11:59:12 -0700724 if (idle_time != -1ULL) {
725 /* Idle micro accounting is supported. Use finer thresholds */
726 dbs_tuners_ins.up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
727 dbs_tuners_ins.down_differential =
728 MICRO_FREQUENCY_DOWN_DIFFERENTIAL;
Thomas Renningercef96152009-04-22 13:48:29 +0200729 /*
Paul Bollebd74b322011-08-06 14:33:43 +0200730 * In nohz/micro accounting case we set the minimum frequency
Thomas Renningercef96152009-04-22 13:48:29 +0200731 * not depending on HZ, but fixed (very low). The deferred
732 * timer might skip some samples if idle/sleeping as needed.
733 */
734 min_sampling_rate = MICRO_FREQUENCY_MIN_SAMPLE_RATE;
735 } else {
736 /* For correct statistics, we need 10 ticks for each measure */
737 min_sampling_rate =
738 MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10);
venkatesh.pallipadi@intel.com80800912008-08-04 11:59:12 -0700739 }
Akinobu Mita888a7942008-07-14 12:00:45 +0900740
Tejun Heo57df5572011-01-26 12:12:50 +0100741 return cpufreq_register_governor(&cpufreq_gov_ondemand);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700742}
743
744static void __exit cpufreq_gov_dbs_exit(void)
745{
Thomas Renninger1c256242007-10-02 13:28:12 -0700746 cpufreq_unregister_governor(&cpufreq_gov_ondemand);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700747}
748
749
Venkatesh Pallipadiffac80e2006-06-28 13:52:18 -0700750MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
751MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
752MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
Dave Jones2b03f892009-01-18 01:43:44 -0500753 "Low Latency Frequency Transition capable processors");
Venkatesh Pallipadiffac80e2006-06-28 13:52:18 -0700754MODULE_LICENSE("GPL");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700755
Johannes Weiner69157192008-01-17 15:21:08 -0800756#ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
757fs_initcall(cpufreq_gov_dbs_init);
758#else
Linus Torvalds1da177e2005-04-16 15:20:36 -0700759module_init(cpufreq_gov_dbs_init);
Johannes Weiner69157192008-01-17 15:21:08 -0800760#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700761module_exit(cpufreq_gov_dbs_exit);