blob: 836e9b062e5ec4a2e08c935d4c49f17a79ef124b [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
Glauber Costa3292beb2011-11-28 14:45:17 -0200122static inline u64 get_cpu_idle_time_jiffy(unsigned int cpu, u64 *wall)
Dave Jonesdac1c1a2005-05-31 19:03:49 -0700123{
Glauber Costa3292beb2011-11-28 14:45:17 -0200124 u64 idle_time;
Martin Schwidefsky612ef282011-12-19 19:23:15 +0100125 u64 cur_wall_time;
Glauber Costa3292beb2011-11-28 14:45:17 -0200126 u64 busy_time;
Venkatesh Pallipadiccb2fe22006-06-28 13:49:52 -0700127
venkatesh.pallipadi@intel.com34305022008-08-04 11:59:09 -0700128 cur_wall_time = jiffies64_to_cputime64(get_jiffies_64());
Venkatesh Pallipadiccb2fe22006-06-28 13:49:52 -0700129
Martin Schwidefsky612ef282011-12-19 19:23:15 +0100130 busy_time = kcpustat_cpu(cpu).cpustat[CPUTIME_USER];
131 busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_SYSTEM];
Glauber Costa3292beb2011-11-28 14:45:17 -0200132 busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_IRQ];
133 busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_SOFTIRQ];
134 busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_STEAL];
135 busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_NICE];
Venkatesh Pallipadiccb2fe22006-06-28 13:49:52 -0700136
Martin Schwidefsky64861632011-12-15 14:56:09 +0100137 idle_time = cur_wall_time - busy_time;
venkatesh.pallipadi@intel.com34305022008-08-04 11:59:09 -0700138 if (wall)
Glauber Costa3292beb2011-11-28 14:45:17 -0200139 *wall = jiffies_to_usecs(cur_wall_time);
venkatesh.pallipadi@intel.com34305022008-08-04 11:59:09 -0700140
Glauber Costa3292beb2011-11-28 14:45:17 -0200141 return jiffies_to_usecs(idle_time);
Dave Jonesdac1c1a2005-05-31 19:03:49 -0700142}
143
venkatesh.pallipadi@intel.com80800912008-08-04 11:59:12 -0700144static inline cputime64_t get_cpu_idle_time(unsigned int cpu, cputime64_t *wall)
145{
Michal Hocko6beea0c2011-08-24 09:37:48 +0200146 u64 idle_time = get_cpu_idle_time_us(cpu, NULL);
venkatesh.pallipadi@intel.com80800912008-08-04 11:59:12 -0700147
148 if (idle_time == -1ULL)
149 return get_cpu_idle_time_jiffy(cpu, wall);
Michal Hocko6beea0c2011-08-24 09:37:48 +0200150 else
151 idle_time += get_cpu_iowait_time_us(cpu, wall);
venkatesh.pallipadi@intel.com80800912008-08-04 11:59:12 -0700152
venkatesh.pallipadi@intel.com80800912008-08-04 11:59:12 -0700153 return idle_time;
154}
155
Arjan van de Ven6b8fcd92010-05-09 08:26:06 -0700156static inline cputime64_t get_cpu_iowait_time(unsigned int cpu, cputime64_t *wall)
157{
158 u64 iowait_time = get_cpu_iowait_time_us(cpu, wall);
159
160 if (iowait_time == -1ULL)
161 return 0;
162
163 return iowait_time;
164}
165
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400166/*
167 * Find right freq to be set now with powersave_bias on.
168 * Returns the freq_hi to be used right now and will set freq_hi_jiffies,
169 * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs.
170 */
Adrian Bunkb5ecf602006-08-13 23:00:08 +0200171static unsigned int powersave_bias_target(struct cpufreq_policy *policy,
172 unsigned int freq_next,
173 unsigned int relation)
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400174{
175 unsigned int freq_req, freq_reduc, freq_avg;
176 unsigned int freq_hi, freq_lo;
177 unsigned int index = 0;
178 unsigned int jiffies_total, jiffies_hi, jiffies_lo;
Tejun Heo245b2e72009-06-24 15:13:48 +0900179 struct cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
180 policy->cpu);
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400181
182 if (!dbs_info->freq_table) {
183 dbs_info->freq_lo = 0;
184 dbs_info->freq_lo_jiffies = 0;
185 return freq_next;
186 }
187
188 cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next,
189 relation, &index);
190 freq_req = dbs_info->freq_table[index].frequency;
191 freq_reduc = freq_req * dbs_tuners_ins.powersave_bias / 1000;
192 freq_avg = freq_req - freq_reduc;
193
194 /* Find freq bounds for freq_avg in freq_table */
195 index = 0;
196 cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
197 CPUFREQ_RELATION_H, &index);
198 freq_lo = dbs_info->freq_table[index].frequency;
199 index = 0;
200 cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
201 CPUFREQ_RELATION_L, &index);
202 freq_hi = dbs_info->freq_table[index].frequency;
203
204 /* Find out how long we have to be in hi and lo freqs */
205 if (freq_hi == freq_lo) {
206 dbs_info->freq_lo = 0;
207 dbs_info->freq_lo_jiffies = 0;
208 return freq_lo;
209 }
210 jiffies_total = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
211 jiffies_hi = (freq_avg - freq_lo) * jiffies_total;
212 jiffies_hi += ((freq_hi - freq_lo) / 2);
213 jiffies_hi /= (freq_hi - freq_lo);
214 jiffies_lo = jiffies_total - jiffies_hi;
215 dbs_info->freq_lo = freq_lo;
216 dbs_info->freq_lo_jiffies = jiffies_lo;
217 dbs_info->freq_hi_jiffies = jiffies_hi;
218 return freq_hi;
219}
220
venkatesh.pallipadi@intel.com5a75c822009-07-02 17:08:32 -0700221static void ondemand_powersave_bias_init_cpu(int cpu)
222{
Tejun Heo384be2b2009-08-14 14:41:02 +0900223 struct cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
venkatesh.pallipadi@intel.com5a75c822009-07-02 17:08:32 -0700224 dbs_info->freq_table = cpufreq_frequency_get_table(cpu);
225 dbs_info->freq_lo = 0;
226}
227
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400228static void ondemand_powersave_bias_init(void)
229{
230 int i;
231 for_each_online_cpu(i) {
venkatesh.pallipadi@intel.com5a75c822009-07-02 17:08:32 -0700232 ondemand_powersave_bias_init_cpu(i);
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400233 }
234}
235
Linus Torvalds1da177e2005-04-16 15:20:36 -0700236/************************** sysfs interface ************************/
Thomas Renninger0e625ac2009-07-24 15:25:06 +0200237
Thomas Renninger0e625ac2009-07-24 15:25:06 +0200238static ssize_t show_sampling_rate_min(struct kobject *kobj,
239 struct attribute *attr, char *buf)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700240{
Thomas Renningercef96152009-04-22 13:48:29 +0200241 return sprintf(buf, "%u\n", min_sampling_rate);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700242}
243
Borislav Petkov6dad2a22010-03-31 21:56:46 +0200244define_one_global_ro(sampling_rate_min);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700245
246/* cpufreq_ondemand Governor Tunables */
247#define show_one(file_name, object) \
248static ssize_t show_##file_name \
Thomas Renninger0e625ac2009-07-24 15:25:06 +0200249(struct kobject *kobj, struct attribute *attr, char *buf) \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700250{ \
251 return sprintf(buf, "%u\n", dbs_tuners_ins.object); \
252}
253show_one(sampling_rate, sampling_rate);
Arjan van de Ven19379b12010-05-09 08:26:51 -0700254show_one(io_is_busy, io_is_busy);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700255show_one(up_threshold, up_threshold);
David C Niemi3f78a9f2010-10-06 16:54:24 -0400256show_one(sampling_down_factor, sampling_down_factor);
Alexander Clouter001893c2005-12-01 01:09:25 -0800257show_one(ignore_nice_load, ignore_nice);
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400258show_one(powersave_bias, powersave_bias);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700259
MyungJoo Hamfd0ef7a2012-02-29 17:54:41 +0900260/**
261 * update_sampling_rate - update sampling rate effective immediately if needed.
262 * @new_rate: new sampling rate
263 *
264 * If new rate is smaller than the old, simply updaing
265 * dbs_tuners_int.sampling_rate might not be appropriate. For example,
266 * if the original sampling_rate was 1 second and the requested new sampling
267 * rate is 10 ms because the user needs immediate reaction from ondemand
268 * governor, but not sure if higher frequency will be required or not,
269 * then, the governor may change the sampling rate too late; up to 1 second
270 * later. Thus, if we are reducing the sampling rate, we need to make the
271 * new value effective immediately.
272 */
273static void update_sampling_rate(unsigned int new_rate)
274{
275 int cpu;
276
277 dbs_tuners_ins.sampling_rate = new_rate
278 = max(new_rate, min_sampling_rate);
279
280 for_each_online_cpu(cpu) {
281 struct cpufreq_policy *policy;
282 struct cpu_dbs_info_s *dbs_info;
283 unsigned long next_sampling, appointed_at;
284
285 policy = cpufreq_cpu_get(cpu);
286 if (!policy)
287 continue;
288 dbs_info = &per_cpu(od_cpu_dbs_info, policy->cpu);
289 cpufreq_cpu_put(policy);
290
291 mutex_lock(&dbs_info->timer_mutex);
292
293 if (!delayed_work_pending(&dbs_info->work)) {
294 mutex_unlock(&dbs_info->timer_mutex);
295 continue;
296 }
297
298 next_sampling = jiffies + usecs_to_jiffies(new_rate);
299 appointed_at = dbs_info->work.timer.expires;
300
301
302 if (time_before(next_sampling, appointed_at)) {
303
304 mutex_unlock(&dbs_info->timer_mutex);
305 cancel_delayed_work_sync(&dbs_info->work);
306 mutex_lock(&dbs_info->timer_mutex);
307
308 schedule_delayed_work_on(dbs_info->cpu, &dbs_info->work,
309 usecs_to_jiffies(new_rate));
310
311 }
312 mutex_unlock(&dbs_info->timer_mutex);
313 }
314}
315
Thomas Renninger0e625ac2009-07-24 15:25:06 +0200316static ssize_t store_sampling_rate(struct kobject *a, struct attribute *b,
317 const char *buf, size_t count)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700318{
319 unsigned int input;
320 int ret;
Venkatesh Pallipadiffac80e2006-06-28 13:52:18 -0700321 ret = sscanf(buf, "%u", &input);
venkatesh.pallipadi@intel.com5a75c822009-07-02 17:08:32 -0700322 if (ret != 1)
323 return -EINVAL;
MyungJoo Hamfd0ef7a2012-02-29 17:54:41 +0900324 update_sampling_rate(input);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700325 return count;
326}
327
Arjan van de Ven19379b12010-05-09 08:26:51 -0700328static ssize_t store_io_is_busy(struct kobject *a, struct attribute *b,
329 const char *buf, size_t count)
330{
331 unsigned int input;
332 int ret;
333
334 ret = sscanf(buf, "%u", &input);
335 if (ret != 1)
336 return -EINVAL;
Arjan van de Ven19379b12010-05-09 08:26:51 -0700337 dbs_tuners_ins.io_is_busy = !!input;
Arjan van de Ven19379b12010-05-09 08:26:51 -0700338 return count;
339}
340
Thomas Renninger0e625ac2009-07-24 15:25:06 +0200341static ssize_t store_up_threshold(struct kobject *a, struct attribute *b,
342 const char *buf, size_t count)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700343{
344 unsigned int input;
345 int ret;
Venkatesh Pallipadiffac80e2006-06-28 13:52:18 -0700346 ret = sscanf(buf, "%u", &input);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700347
Dave Jones32ee8c32006-02-28 00:43:23 -0500348 if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
Dave Jonesc29f1402005-05-31 19:03:50 -0700349 input < MIN_FREQUENCY_UP_THRESHOLD) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700350 return -EINVAL;
351 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700352 dbs_tuners_ins.up_threshold = input;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700353 return count;
354}
355
David C Niemi3f78a9f2010-10-06 16:54:24 -0400356static ssize_t store_sampling_down_factor(struct kobject *a,
357 struct attribute *b, const char *buf, size_t count)
358{
359 unsigned int input, j;
360 int ret;
361 ret = sscanf(buf, "%u", &input);
362
363 if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
364 return -EINVAL;
David C Niemi3f78a9f2010-10-06 16:54:24 -0400365 dbs_tuners_ins.sampling_down_factor = input;
366
367 /* Reset down sampling multiplier in case it was active */
368 for_each_online_cpu(j) {
369 struct cpu_dbs_info_s *dbs_info;
370 dbs_info = &per_cpu(od_cpu_dbs_info, j);
371 dbs_info->rate_mult = 1;
372 }
David C Niemi3f78a9f2010-10-06 16:54:24 -0400373 return count;
374}
375
Thomas Renninger0e625ac2009-07-24 15:25:06 +0200376static ssize_t store_ignore_nice_load(struct kobject *a, struct attribute *b,
377 const char *buf, size_t count)
Dave Jones3d5ee9e2005-05-31 19:03:47 -0700378{
379 unsigned int input;
380 int ret;
381
382 unsigned int j;
Dave Jones32ee8c32006-02-28 00:43:23 -0500383
Venkatesh Pallipadiffac80e2006-06-28 13:52:18 -0700384 ret = sscanf(buf, "%u", &input);
Dave Jones2b03f892009-01-18 01:43:44 -0500385 if (ret != 1)
Dave Jones3d5ee9e2005-05-31 19:03:47 -0700386 return -EINVAL;
387
Dave Jones2b03f892009-01-18 01:43:44 -0500388 if (input > 1)
Dave Jones3d5ee9e2005-05-31 19:03:47 -0700389 input = 1;
Dave Jones32ee8c32006-02-28 00:43:23 -0500390
Dave Jones2b03f892009-01-18 01:43:44 -0500391 if (input == dbs_tuners_ins.ignore_nice) { /* nothing to do */
Dave Jones3d5ee9e2005-05-31 19:03:47 -0700392 return count;
393 }
394 dbs_tuners_ins.ignore_nice = input;
395
Venkatesh Pallipadiccb2fe22006-06-28 13:49:52 -0700396 /* we need to re-evaluate prev_cpu_idle */
Dave Jonesdac1c1a2005-05-31 19:03:49 -0700397 for_each_online_cpu(j) {
Venkatesh Pallipadiccb2fe22006-06-28 13:49:52 -0700398 struct cpu_dbs_info_s *dbs_info;
Tejun Heo245b2e72009-06-24 15:13:48 +0900399 dbs_info = &per_cpu(od_cpu_dbs_info, j);
venkatesh.pallipadi@intel.com34305022008-08-04 11:59:09 -0700400 dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
401 &dbs_info->prev_cpu_wall);
Venkatesh Pallipadi1ca3abd2009-01-23 09:25:02 -0500402 if (dbs_tuners_ins.ignore_nice)
Glauber Costa3292beb2011-11-28 14:45:17 -0200403 dbs_info->prev_cpu_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE];
Venkatesh Pallipadi1ca3abd2009-01-23 09:25:02 -0500404
Dave Jones3d5ee9e2005-05-31 19:03:47 -0700405 }
Dave Jones3d5ee9e2005-05-31 19:03:47 -0700406 return count;
407}
408
Thomas Renninger0e625ac2009-07-24 15:25:06 +0200409static ssize_t store_powersave_bias(struct kobject *a, struct attribute *b,
410 const char *buf, size_t count)
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400411{
412 unsigned int input;
413 int ret;
414 ret = sscanf(buf, "%u", &input);
415
416 if (ret != 1)
417 return -EINVAL;
418
419 if (input > 1000)
420 input = 1000;
421
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400422 dbs_tuners_ins.powersave_bias = input;
423 ondemand_powersave_bias_init();
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400424 return count;
425}
426
Borislav Petkov6dad2a22010-03-31 21:56:46 +0200427define_one_global_rw(sampling_rate);
Linus Torvalds07d77752010-05-18 08:49:13 -0700428define_one_global_rw(io_is_busy);
Borislav Petkov6dad2a22010-03-31 21:56:46 +0200429define_one_global_rw(up_threshold);
David C Niemi3f78a9f2010-10-06 16:54:24 -0400430define_one_global_rw(sampling_down_factor);
Borislav Petkov6dad2a22010-03-31 21:56:46 +0200431define_one_global_rw(ignore_nice_load);
432define_one_global_rw(powersave_bias);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700433
Dave Jones2b03f892009-01-18 01:43:44 -0500434static struct attribute *dbs_attributes[] = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700435 &sampling_rate_min.attr,
436 &sampling_rate.attr,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700437 &up_threshold.attr,
David C Niemi3f78a9f2010-10-06 16:54:24 -0400438 &sampling_down_factor.attr,
Alexander Clouter001893c2005-12-01 01:09:25 -0800439 &ignore_nice_load.attr,
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400440 &powersave_bias.attr,
Arjan van de Ven19379b12010-05-09 08:26:51 -0700441 &io_is_busy.attr,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700442 NULL
443};
444
445static struct attribute_group dbs_attr_group = {
446 .attrs = dbs_attributes,
447 .name = "ondemand",
448};
449
450/************************** sysfs end ************************/
451
Mike Chan00e299f2010-01-26 17:06:47 -0800452static void dbs_freq_increase(struct cpufreq_policy *p, unsigned int freq)
453{
454 if (dbs_tuners_ins.powersave_bias)
455 freq = powersave_bias_target(p, freq, CPUFREQ_RELATION_H);
456 else if (p->cur == p->max)
457 return;
458
459 __cpufreq_driver_target(p, freq, dbs_tuners_ins.powersave_bias ?
460 CPUFREQ_RELATION_L : CPUFREQ_RELATION_H);
461}
462
Venkatesh Pallipadi2f8a8352006-06-28 13:51:19 -0700463static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700464{
venkatesh.pallipadi@intel.comc43aa3b2008-08-04 11:59:08 -0700465 unsigned int max_load_freq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700466
467 struct cpufreq_policy *policy;
468 unsigned int j;
469
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400470 this_dbs_info->freq_lo = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700471 policy = this_dbs_info->cur_policy;
Venki Pallipadiea487612007-06-20 14:26:24 -0700472
Dave Jones32ee8c32006-02-28 00:43:23 -0500473 /*
Dave Jonesc29f1402005-05-31 19:03:50 -0700474 * Every sampling_rate, we check, if current idle time is less
475 * than 20% (default), then we try to increase frequency
Venkatesh Pallipadiccb2fe22006-06-28 13:49:52 -0700476 * Every sampling_rate, we look for a the lowest
Dave Jonesc29f1402005-05-31 19:03:50 -0700477 * frequency which can sustain the load while keeping idle time over
478 * 30%. If such a frequency exist, we try to decrease to this frequency.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700479 *
Dave Jones32ee8c32006-02-28 00:43:23 -0500480 * Any frequency increase takes it to the maximum frequency.
481 * Frequency reduction happens at minimum steps of
482 * 5% (default) of current frequency
Linus Torvalds1da177e2005-04-16 15:20:36 -0700483 */
484
venkatesh.pallipadi@intel.comc43aa3b2008-08-04 11:59:08 -0700485 /* Get Absolute Load - in terms of freq */
486 max_load_freq = 0;
487
Rusty Russell835481d2009-01-04 05:18:06 -0800488 for_each_cpu(j, policy->cpus) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700489 struct cpu_dbs_info_s *j_dbs_info;
Arjan van de Ven6b8fcd92010-05-09 08:26:06 -0700490 cputime64_t cur_wall_time, cur_idle_time, cur_iowait_time;
491 unsigned int idle_time, wall_time, iowait_time;
venkatesh.pallipadi@intel.comc43aa3b2008-08-04 11:59:08 -0700492 unsigned int load, load_freq;
493 int freq_avg;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700494
Tejun Heo245b2e72009-06-24 15:13:48 +0900495 j_dbs_info = &per_cpu(od_cpu_dbs_info, j);
venkatesh.pallipadi@intel.com34305022008-08-04 11:59:09 -0700496
497 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time);
Arjan van de Ven6b8fcd92010-05-09 08:26:06 -0700498 cur_iowait_time = get_cpu_iowait_time(j, &cur_wall_time);
venkatesh.pallipadi@intel.com34305022008-08-04 11:59:09 -0700499
Martin Schwidefsky64861632011-12-15 14:56:09 +0100500 wall_time = (unsigned int)
501 (cur_wall_time - j_dbs_info->prev_cpu_wall);
venkatesh.pallipadi@intel.comc43aa3b2008-08-04 11:59:08 -0700502 j_dbs_info->prev_cpu_wall = cur_wall_time;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700503
Martin Schwidefsky64861632011-12-15 14:56:09 +0100504 idle_time = (unsigned int)
505 (cur_idle_time - j_dbs_info->prev_cpu_idle);
venkatesh.pallipadi@intel.comc43aa3b2008-08-04 11:59:08 -0700506 j_dbs_info->prev_cpu_idle = cur_idle_time;
507
Martin Schwidefsky64861632011-12-15 14:56:09 +0100508 iowait_time = (unsigned int)
509 (cur_iowait_time - j_dbs_info->prev_cpu_iowait);
Arjan van de Ven6b8fcd92010-05-09 08:26:06 -0700510 j_dbs_info->prev_cpu_iowait = cur_iowait_time;
511
Venkatesh Pallipadi1ca3abd2009-01-23 09:25:02 -0500512 if (dbs_tuners_ins.ignore_nice) {
Glauber Costa3292beb2011-11-28 14:45:17 -0200513 u64 cur_nice;
Venkatesh Pallipadi1ca3abd2009-01-23 09:25:02 -0500514 unsigned long cur_nice_jiffies;
515
Glauber Costa3292beb2011-11-28 14:45:17 -0200516 cur_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE] -
517 j_dbs_info->prev_cpu_nice;
Venkatesh Pallipadi1ca3abd2009-01-23 09:25:02 -0500518 /*
519 * Assumption: nice time between sampling periods will
520 * be less than 2^32 jiffies for 32 bit sys
521 */
522 cur_nice_jiffies = (unsigned long)
523 cputime64_to_jiffies64(cur_nice);
524
Glauber Costa3292beb2011-11-28 14:45:17 -0200525 j_dbs_info->prev_cpu_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE];
Venkatesh Pallipadi1ca3abd2009-01-23 09:25:02 -0500526 idle_time += jiffies_to_usecs(cur_nice_jiffies);
527 }
528
Arjan van de Ven6b8fcd92010-05-09 08:26:06 -0700529 /*
530 * For the purpose of ondemand, waiting for disk IO is an
531 * indication that you're performance critical, and not that
532 * the system is actually idle. So subtract the iowait time
533 * from the cpu idle time.
534 */
535
Arjan van de Ven19379b12010-05-09 08:26:51 -0700536 if (dbs_tuners_ins.io_is_busy && idle_time >= iowait_time)
Arjan van de Ven6b8fcd92010-05-09 08:26:06 -0700537 idle_time -= iowait_time;
538
venkatesh.pallipadi@intel.com34305022008-08-04 11:59:09 -0700539 if (unlikely(!wall_time || wall_time < idle_time))
venkatesh.pallipadi@intel.comc43aa3b2008-08-04 11:59:08 -0700540 continue;
venkatesh.pallipadi@intel.comc43aa3b2008-08-04 11:59:08 -0700541
542 load = 100 * (wall_time - idle_time) / wall_time;
543
544 freq_avg = __cpufreq_driver_getavg(policy, j);
545 if (freq_avg <= 0)
546 freq_avg = policy->cur;
547
548 load_freq = load * freq_avg;
549 if (load_freq > max_load_freq)
550 max_load_freq = load_freq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700551 }
552
Venkatesh Pallipadiccb2fe22006-06-28 13:49:52 -0700553 /* Check for frequency increase */
venkatesh.pallipadi@intel.comc43aa3b2008-08-04 11:59:08 -0700554 if (max_load_freq > dbs_tuners_ins.up_threshold * policy->cur) {
David C Niemi3f78a9f2010-10-06 16:54:24 -0400555 /* If switching to max speed, apply sampling_down_factor */
556 if (policy->cur < policy->max)
557 this_dbs_info->rate_mult =
558 dbs_tuners_ins.sampling_down_factor;
Mike Chan00e299f2010-01-26 17:06:47 -0800559 dbs_freq_increase(policy, policy->max);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700560 return;
561 }
562
563 /* Check for frequency decrease */
Dave Jonesc29f1402005-05-31 19:03:50 -0700564 /* if we cannot reduce the frequency anymore, break out early */
565 if (policy->cur == policy->min)
566 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700567
Dave Jonesc29f1402005-05-31 19:03:50 -0700568 /*
569 * The optimal frequency is the frequency that is the lowest that
570 * can support the current CPU usage without triggering the up
571 * policy. To be safe, we focus 10 points under the threshold.
572 */
venkatesh.pallipadi@intel.come9d95bf2008-08-04 11:59:10 -0700573 if (max_load_freq <
574 (dbs_tuners_ins.up_threshold - dbs_tuners_ins.down_differential) *
575 policy->cur) {
venkatesh.pallipadi@intel.comc43aa3b2008-08-04 11:59:08 -0700576 unsigned int freq_next;
venkatesh.pallipadi@intel.come9d95bf2008-08-04 11:59:10 -0700577 freq_next = max_load_freq /
578 (dbs_tuners_ins.up_threshold -
579 dbs_tuners_ins.down_differential);
Venkatesh Pallipadidfde5d62006-10-03 12:38:45 -0700580
David C Niemi3f78a9f2010-10-06 16:54:24 -0400581 /* No longer fully busy, reset rate_mult */
582 this_dbs_info->rate_mult = 1;
583
Nagananda.Chumbalkar@hp.com1dbf5882009-12-21 23:40:52 +0100584 if (freq_next < policy->min)
585 freq_next = policy->min;
586
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400587 if (!dbs_tuners_ins.powersave_bias) {
588 __cpufreq_driver_target(policy, freq_next,
589 CPUFREQ_RELATION_L);
590 } else {
591 int freq = powersave_bias_target(policy, freq_next,
592 CPUFREQ_RELATION_L);
593 __cpufreq_driver_target(policy, freq,
594 CPUFREQ_RELATION_L);
595 }
Venkatesh Pallipadiccb2fe22006-06-28 13:49:52 -0700596 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700597}
598
David Howellsc4028952006-11-22 14:57:56 +0000599static void do_dbs_timer(struct work_struct *work)
Dave Jones32ee8c32006-02-28 00:43:23 -0500600{
Venkatesh Pallipadi529af7a2007-02-05 16:12:44 -0800601 struct cpu_dbs_info_s *dbs_info =
602 container_of(work, struct cpu_dbs_info_s, work.work);
603 unsigned int cpu = dbs_info->cpu;
604 int sample_type = dbs_info->sample_type;
605
Vincent Guittot5cb2c3b2011-02-07 17:14:25 +0100606 int delay;
Jocelyn Falempea665df92010-03-11 14:01:11 -0800607
venkatesh.pallipadi@intel.com5a75c822009-07-02 17:08:32 -0700608 mutex_lock(&dbs_info->timer_mutex);
Venkatesh Pallipadi56463b72007-02-05 16:12:45 -0800609
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400610 /* Common NORMAL_SAMPLE setup */
David Howellsc4028952006-11-22 14:57:56 +0000611 dbs_info->sample_type = DBS_NORMAL_SAMPLE;
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400612 if (!dbs_tuners_ins.powersave_bias ||
David Howellsc4028952006-11-22 14:57:56 +0000613 sample_type == DBS_NORMAL_SAMPLE) {
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400614 dbs_check_cpu(dbs_info);
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400615 if (dbs_info->freq_lo) {
616 /* Setup timer for SUB_SAMPLE */
David Howellsc4028952006-11-22 14:57:56 +0000617 dbs_info->sample_type = DBS_SUB_SAMPLE;
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400618 delay = dbs_info->freq_hi_jiffies;
Vincent Guittot5cb2c3b2011-02-07 17:14:25 +0100619 } else {
620 /* We want all CPUs to do sampling nearly on
621 * same jiffy
622 */
623 delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate
624 * dbs_info->rate_mult);
625
626 if (num_online_cpus() > 1)
627 delay -= jiffies % delay;
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400628 }
629 } else {
630 __cpufreq_driver_target(dbs_info->cur_policy,
Dave Jones2b03f892009-01-18 01:43:44 -0500631 dbs_info->freq_lo, CPUFREQ_RELATION_H);
Vincent Guittot5cb2c3b2011-02-07 17:14:25 +0100632 delay = dbs_info->freq_lo_jiffies;
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400633 }
Tejun Heo57df5572011-01-26 12:12:50 +0100634 schedule_delayed_work_on(cpu, &dbs_info->work, delay);
venkatesh.pallipadi@intel.com5a75c822009-07-02 17:08:32 -0700635 mutex_unlock(&dbs_info->timer_mutex);
Dave Jones32ee8c32006-02-28 00:43:23 -0500636}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700637
Venkatesh Pallipadi529af7a2007-02-05 16:12:44 -0800638static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700639{
Alexey Starikovskiy1ce28d62006-07-31 22:25:20 +0400640 /* We want all CPUs to do sampling nearly on same jiffy */
641 int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
Jocelyn Falempea665df92010-03-11 14:01:11 -0800642
643 if (num_online_cpus() > 1)
644 delay -= jiffies % delay;
Venkatesh Pallipadi2f8a8352006-06-28 13:51:19 -0700645
David Howellsc4028952006-11-22 14:57:56 +0000646 dbs_info->sample_type = DBS_NORMAL_SAMPLE;
Venki Pallipadi28287032007-05-08 00:27:47 -0700647 INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer);
Tejun Heo57df5572011-01-26 12:12:50 +0100648 schedule_delayed_work_on(dbs_info->cpu, &dbs_info->work, delay);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700649}
650
Linus Torvalds2cd7cbd2006-07-23 12:05:00 -0700651static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700652{
Mathieu Desnoyersb14893a2009-05-17 10:30:45 -0400653 cancel_delayed_work_sync(&dbs_info->work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700654}
655
Arjan van de Ven19379b12010-05-09 08:26:51 -0700656/*
657 * Not all CPUs want IO time to be accounted as busy; this dependson how
658 * efficient idling at a higher frequency/voltage is.
659 * Pavel Machek says this is not so for various generations of AMD and old
660 * Intel systems.
661 * Mike Chan (androidlcom) calis this is also not true for ARM.
662 * Because of this, whitelist specific known (series) of CPUs by default, and
663 * leave all others up to the user.
664 */
665static int should_io_be_busy(void)
666{
667#if defined(CONFIG_X86)
668 /*
669 * For Intel, Core 2 (model 15) andl later have an efficient idle.
670 */
671 if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
672 boot_cpu_data.x86 == 6 &&
673 boot_cpu_data.x86_model >= 15)
674 return 1;
675#endif
676 return 0;
677}
678
Linus Torvalds1da177e2005-04-16 15:20:36 -0700679static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
680 unsigned int event)
681{
682 unsigned int cpu = policy->cpu;
683 struct cpu_dbs_info_s *this_dbs_info;
684 unsigned int j;
Jeff Garzik914f7c32006-10-20 14:31:00 -0700685 int rc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700686
Tejun Heo245b2e72009-06-24 15:13:48 +0900687 this_dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700688
689 switch (event) {
690 case CPUFREQ_GOV_START:
Venkatesh Pallipadiffac80e2006-06-28 13:52:18 -0700691 if ((!cpu_online(cpu)) || (!policy->cur))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700692 return -EINVAL;
693
akpm@osdl.org3fc54d32006-01-13 15:54:22 -0800694 mutex_lock(&dbs_mutex);
Jeff Garzik914f7c32006-10-20 14:31:00 -0700695
venkatesh.pallipadi@intel.com5a75c822009-07-02 17:08:32 -0700696 dbs_enable++;
Rusty Russell835481d2009-01-04 05:18:06 -0800697 for_each_cpu(j, policy->cpus) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700698 struct cpu_dbs_info_s *j_dbs_info;
Tejun Heo245b2e72009-06-24 15:13:48 +0900699 j_dbs_info = &per_cpu(od_cpu_dbs_info, j);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700700 j_dbs_info->cur_policy = policy;
Dave Jones32ee8c32006-02-28 00:43:23 -0500701
venkatesh.pallipadi@intel.com34305022008-08-04 11:59:09 -0700702 j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
703 &j_dbs_info->prev_cpu_wall);
Glauber Costa3292beb2011-11-28 14:45:17 -0200704 if (dbs_tuners_ins.ignore_nice)
Venkatesh Pallipadi1ca3abd2009-01-23 09:25:02 -0500705 j_dbs_info->prev_cpu_nice =
Glauber Costa3292beb2011-11-28 14:45:17 -0200706 kcpustat_cpu(j).cpustat[CPUTIME_NICE];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700707 }
Venkatesh Pallipadi529af7a2007-02-05 16:12:44 -0800708 this_dbs_info->cpu = cpu;
David C Niemi3f78a9f2010-10-06 16:54:24 -0400709 this_dbs_info->rate_mult = 1;
venkatesh.pallipadi@intel.com5a75c822009-07-02 17:08:32 -0700710 ondemand_powersave_bias_init_cpu(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700711 /*
712 * Start the timerschedule work, when this governor
713 * is used for first time
714 */
715 if (dbs_enable == 1) {
716 unsigned int latency;
Thomas Renninger0e625ac2009-07-24 15:25:06 +0200717
718 rc = sysfs_create_group(cpufreq_global_kobject,
719 &dbs_attr_group);
720 if (rc) {
721 mutex_unlock(&dbs_mutex);
722 return rc;
723 }
724
Linus Torvalds1da177e2005-04-16 15:20:36 -0700725 /* policy latency is in nS. Convert it to uS first */
Dave Jonesdf8b59b2005-09-20 12:39:35 -0700726 latency = policy->cpuinfo.transition_latency / 1000;
727 if (latency == 0)
728 latency = 1;
Thomas Renningercef96152009-04-22 13:48:29 +0200729 /* Bring kernel and HW constraints together */
730 min_sampling_rate = max(min_sampling_rate,
731 MIN_LATENCY_MULTIPLIER * latency);
732 dbs_tuners_ins.sampling_rate =
733 max(min_sampling_rate,
734 latency * LATENCY_MULTIPLIER);
Arjan van de Ven19379b12010-05-09 08:26:51 -0700735 dbs_tuners_ins.io_is_busy = should_io_be_busy();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700736 }
akpm@osdl.org3fc54d32006-01-13 15:54:22 -0800737 mutex_unlock(&dbs_mutex);
venkatesh.pallipadi@intel.com7d26e2d2009-07-02 17:08:30 -0700738
Thomas Renninger0e625ac2009-07-24 15:25:06 +0200739 mutex_init(&this_dbs_info->timer_mutex);
venkatesh.pallipadi@intel.com7d26e2d2009-07-02 17:08:30 -0700740 dbs_timer_init(this_dbs_info);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700741 break;
742
743 case CPUFREQ_GOV_STOP:
Linus Torvalds2cd7cbd2006-07-23 12:05:00 -0700744 dbs_timer_exit(this_dbs_info);
venkatesh.pallipadi@intel.com7d26e2d2009-07-02 17:08:30 -0700745
746 mutex_lock(&dbs_mutex);
venkatesh.pallipadi@intel.com5a75c822009-07-02 17:08:32 -0700747 mutex_destroy(&this_dbs_info->timer_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700748 dbs_enable--;
akpm@osdl.org3fc54d32006-01-13 15:54:22 -0800749 mutex_unlock(&dbs_mutex);
Thomas Renninger0e625ac2009-07-24 15:25:06 +0200750 if (!dbs_enable)
751 sysfs_remove_group(cpufreq_global_kobject,
752 &dbs_attr_group);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700753
754 break;
755
756 case CPUFREQ_GOV_LIMITS:
venkatesh.pallipadi@intel.com5a75c822009-07-02 17:08:32 -0700757 mutex_lock(&this_dbs_info->timer_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700758 if (policy->max < this_dbs_info->cur_policy->cur)
Venkatesh Pallipadiffac80e2006-06-28 13:52:18 -0700759 __cpufreq_driver_target(this_dbs_info->cur_policy,
Dave Jones2b03f892009-01-18 01:43:44 -0500760 policy->max, CPUFREQ_RELATION_H);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700761 else if (policy->min > this_dbs_info->cur_policy->cur)
Venkatesh Pallipadiffac80e2006-06-28 13:52:18 -0700762 __cpufreq_driver_target(this_dbs_info->cur_policy,
Dave Jones2b03f892009-01-18 01:43:44 -0500763 policy->min, CPUFREQ_RELATION_L);
venkatesh.pallipadi@intel.com5a75c822009-07-02 17:08:32 -0700764 mutex_unlock(&this_dbs_info->timer_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700765 break;
766 }
767 return 0;
768}
769
Linus Torvalds1da177e2005-04-16 15:20:36 -0700770static int __init cpufreq_gov_dbs_init(void)
771{
Andrea Righi4f6e6b92008-09-18 10:43:40 +0000772 u64 idle_time;
773 int cpu = get_cpu();
venkatesh.pallipadi@intel.com80800912008-08-04 11:59:12 -0700774
Kamalesh Babulal21f2e3c2011-12-09 16:18:42 +0530775 idle_time = get_cpu_idle_time_us(cpu, NULL);
Andrea Righi4f6e6b92008-09-18 10:43:40 +0000776 put_cpu();
venkatesh.pallipadi@intel.com80800912008-08-04 11:59:12 -0700777 if (idle_time != -1ULL) {
778 /* Idle micro accounting is supported. Use finer thresholds */
779 dbs_tuners_ins.up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
780 dbs_tuners_ins.down_differential =
781 MICRO_FREQUENCY_DOWN_DIFFERENTIAL;
Thomas Renningercef96152009-04-22 13:48:29 +0200782 /*
Paul Bollebd74b322011-08-06 14:33:43 +0200783 * In nohz/micro accounting case we set the minimum frequency
Thomas Renningercef96152009-04-22 13:48:29 +0200784 * not depending on HZ, but fixed (very low). The deferred
785 * timer might skip some samples if idle/sleeping as needed.
786 */
787 min_sampling_rate = MICRO_FREQUENCY_MIN_SAMPLE_RATE;
788 } else {
789 /* For correct statistics, we need 10 ticks for each measure */
790 min_sampling_rate =
791 MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10);
venkatesh.pallipadi@intel.com80800912008-08-04 11:59:12 -0700792 }
Akinobu Mita888a7942008-07-14 12:00:45 +0900793
Tejun Heo57df5572011-01-26 12:12:50 +0100794 return cpufreq_register_governor(&cpufreq_gov_ondemand);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700795}
796
797static void __exit cpufreq_gov_dbs_exit(void)
798{
Thomas Renninger1c256242007-10-02 13:28:12 -0700799 cpufreq_unregister_governor(&cpufreq_gov_ondemand);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800}
801
802
Venkatesh Pallipadiffac80e2006-06-28 13:52:18 -0700803MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
804MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
805MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
Dave Jones2b03f892009-01-18 01:43:44 -0500806 "Low Latency Frequency Transition capable processors");
Venkatesh Pallipadiffac80e2006-06-28 13:52:18 -0700807MODULE_LICENSE("GPL");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700808
Johannes Weiner69157192008-01-17 15:21:08 -0800809#ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
810fs_initcall(cpufreq_gov_dbs_init);
811#else
Linus Torvalds1da177e2005-04-16 15:20:36 -0700812module_init(cpufreq_gov_dbs_init);
Johannes Weiner69157192008-01-17 15:21:08 -0800813#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700814module_exit(cpufreq_gov_dbs_exit);