blob: 338f428a15b765ceaef4686225715d5d2f9f93e3 [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)
venkatesh.pallipadi@intel.com80800912008-08-04 11:59:12 -070033#define MICRO_FREQUENCY_DOWN_DIFFERENTIAL (3)
34#define MICRO_FREQUENCY_UP_THRESHOLD (95)
Dave Jonesc29f1402005-05-31 19:03:50 -070035#define MIN_FREQUENCY_UP_THRESHOLD (11)
Linus Torvalds1da177e2005-04-16 15:20:36 -070036#define MAX_FREQUENCY_UP_THRESHOLD (100)
37
Dave Jones32ee8c32006-02-28 00:43:23 -050038/*
39 * The polling frequency of this governor depends on the capability of
Linus Torvalds1da177e2005-04-16 15:20:36 -070040 * the processor. Default polling frequency is 1000 times the transition
Dave Jones32ee8c32006-02-28 00:43:23 -050041 * latency of the processor. The governor will work on any processor with
42 * transition latency <= 10mS, using appropriate sampling
Linus Torvalds1da177e2005-04-16 15:20:36 -070043 * rate.
44 * For CPUs with transition latency > 10mS (mostly drivers with CPUFREQ_ETERNAL)
45 * this governor will not work.
46 * All times here are in uS.
47 */
Dave Jones32ee8c32006-02-28 00:43:23 -050048static unsigned int def_sampling_rate;
Dave Jonesdf8b59b2005-09-20 12:39:35 -070049#define MIN_SAMPLING_RATE_RATIO (2)
50/* for correct statistics, we need at least 10 ticks between each measure */
Gautham R Shenoye08f5f52006-10-26 16:20:58 +053051#define MIN_STAT_SAMPLING_RATE \
52 (MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10))
53#define MIN_SAMPLING_RATE \
54 (def_sampling_rate / MIN_SAMPLING_RATE_RATIO)
Thomas Renninger112124a2009-02-04 11:55:12 +010055/* Above MIN_SAMPLING_RATE will vanish with its sysfs file soon
56 * Define the minimal settable sampling rate to the greater of:
57 * - "HW transition latency" * 100 (same as default sampling / 10)
58 * - MIN_STAT_SAMPLING_RATE
59 * To avoid that userspace shoots itself.
60*/
61static unsigned int minimum_sampling_rate(void)
62{
63 return max(def_sampling_rate / 10, MIN_STAT_SAMPLING_RATE);
64}
65
66/* This will also vanish soon with removing sampling_rate_max */
Linus Torvalds1da177e2005-04-16 15:20:36 -070067#define MAX_SAMPLING_RATE (500 * def_sampling_rate)
Thomas Renninger112124a2009-02-04 11:55:12 +010068#define LATENCY_MULTIPLIER (1000)
Thomas Renninger1c256242007-10-02 13:28:12 -070069#define TRANSITION_LATENCY_LIMIT (10 * 1000 * 1000)
Linus Torvalds1da177e2005-04-16 15:20:36 -070070
David Howellsc4028952006-11-22 14:57:56 +000071static void do_dbs_timer(struct work_struct *work);
72
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;
78 cputime64_t prev_cpu_wall;
venkatesh.pallipadi@intel.com80800912008-08-04 11:59:12 -070079 cputime64_t prev_cpu_nice;
Dave Jones32ee8c32006-02-28 00:43:23 -050080 struct cpufreq_policy *cur_policy;
Dave Jones2b03f892009-01-18 01:43:44 -050081 struct delayed_work work;
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +040082 struct cpufreq_frequency_table *freq_table;
83 unsigned int freq_lo;
84 unsigned int freq_lo_jiffies;
85 unsigned int freq_hi_jiffies;
Venkatesh Pallipadi529af7a2007-02-05 16:12:44 -080086 int cpu;
87 unsigned int enable:1,
Dave Jones2b03f892009-01-18 01:43:44 -050088 sample_type:1;
Linus Torvalds1da177e2005-04-16 15:20:36 -070089};
90static DEFINE_PER_CPU(struct cpu_dbs_info_s, cpu_dbs_info);
91
92static unsigned int dbs_enable; /* number of CPUs using this policy */
93
Venkatesh Pallipadi4ec223d2006-06-21 15:18:34 -070094/*
95 * DEADLOCK ALERT! There is a ordering requirement between cpu_hotplug
96 * lock and dbs_mutex. cpu_hotplug lock should always be held before
97 * dbs_mutex. If any function that can potentially take cpu_hotplug lock
98 * (like __cpufreq_driver_target()) is being called with dbs_mutex taken, then
99 * cpu_hotplug lock should be taken before that. Note that cpu_hotplug lock
100 * is recursive for the same process. -Venki
101 */
Venkatesh Pallipadiffac80e2006-06-28 13:52:18 -0700102static DEFINE_MUTEX(dbs_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700103
Venkatesh Pallipadi2f8a8352006-06-28 13:51:19 -0700104static struct workqueue_struct *kondemand_wq;
Andi Kleen6810b542006-05-08 15:17:31 +0200105
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;
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400111 unsigned int powersave_bias;
112} dbs_tuners_ins = {
Dave Jones32ee8c32006-02-28 00:43:23 -0500113 .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
venkatesh.pallipadi@intel.come9d95bf2008-08-04 11:59:10 -0700114 .down_differential = DEF_FREQUENCY_DOWN_DIFFERENTIAL,
Eric Piel9cbad612006-03-10 11:35:27 +0200115 .ignore_nice = 0,
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400116 .powersave_bias = 0,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700117};
118
venkatesh.pallipadi@intel.com80800912008-08-04 11:59:12 -0700119static inline cputime64_t get_cpu_idle_time_jiffy(unsigned int cpu,
120 cputime64_t *wall)
Dave Jonesdac1c1a2005-05-31 19:03:49 -0700121{
Venki Pallipadiea487612007-06-20 14:26:24 -0700122 cputime64_t idle_time;
venkatesh.pallipadi@intel.com34305022008-08-04 11:59:09 -0700123 cputime64_t cur_wall_time;
Venki Pallipadiea487612007-06-20 14:26:24 -0700124 cputime64_t busy_time;
Venkatesh Pallipadiccb2fe22006-06-28 13:49:52 -0700125
venkatesh.pallipadi@intel.com34305022008-08-04 11:59:09 -0700126 cur_wall_time = jiffies64_to_cputime64(get_jiffies_64());
Venki Pallipadiea487612007-06-20 14:26:24 -0700127 busy_time = cputime64_add(kstat_cpu(cpu).cpustat.user,
128 kstat_cpu(cpu).cpustat.system);
Venkatesh Pallipadiccb2fe22006-06-28 13:49:52 -0700129
Venki Pallipadiea487612007-06-20 14:26:24 -0700130 busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.irq);
131 busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.softirq);
132 busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.steal);
Venkatesh Pallipadi1ca3abd2009-01-23 09:25:02 -0500133 busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.nice);
Venki Pallipadiea487612007-06-20 14:26:24 -0700134
venkatesh.pallipadi@intel.com34305022008-08-04 11:59:09 -0700135 idle_time = cputime64_sub(cur_wall_time, busy_time);
136 if (wall)
137 *wall = cur_wall_time;
138
Venki Pallipadiea487612007-06-20 14:26:24 -0700139 return idle_time;
Dave Jonesdac1c1a2005-05-31 19:03:49 -0700140}
141
venkatesh.pallipadi@intel.com80800912008-08-04 11:59:12 -0700142static inline cputime64_t get_cpu_idle_time(unsigned int cpu, cputime64_t *wall)
143{
144 u64 idle_time = get_cpu_idle_time_us(cpu, wall);
145
146 if (idle_time == -1ULL)
147 return get_cpu_idle_time_jiffy(cpu, wall);
148
venkatesh.pallipadi@intel.com80800912008-08-04 11:59:12 -0700149 return idle_time;
150}
151
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400152/*
153 * Find right freq to be set now with powersave_bias on.
154 * Returns the freq_hi to be used right now and will set freq_hi_jiffies,
155 * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs.
156 */
Adrian Bunkb5ecf602006-08-13 23:00:08 +0200157static unsigned int powersave_bias_target(struct cpufreq_policy *policy,
158 unsigned int freq_next,
159 unsigned int relation)
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400160{
161 unsigned int freq_req, freq_reduc, freq_avg;
162 unsigned int freq_hi, freq_lo;
163 unsigned int index = 0;
164 unsigned int jiffies_total, jiffies_hi, jiffies_lo;
165 struct cpu_dbs_info_s *dbs_info = &per_cpu(cpu_dbs_info, policy->cpu);
166
167 if (!dbs_info->freq_table) {
168 dbs_info->freq_lo = 0;
169 dbs_info->freq_lo_jiffies = 0;
170 return freq_next;
171 }
172
173 cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next,
174 relation, &index);
175 freq_req = dbs_info->freq_table[index].frequency;
176 freq_reduc = freq_req * dbs_tuners_ins.powersave_bias / 1000;
177 freq_avg = freq_req - freq_reduc;
178
179 /* Find freq bounds for freq_avg in freq_table */
180 index = 0;
181 cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
182 CPUFREQ_RELATION_H, &index);
183 freq_lo = dbs_info->freq_table[index].frequency;
184 index = 0;
185 cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
186 CPUFREQ_RELATION_L, &index);
187 freq_hi = dbs_info->freq_table[index].frequency;
188
189 /* Find out how long we have to be in hi and lo freqs */
190 if (freq_hi == freq_lo) {
191 dbs_info->freq_lo = 0;
192 dbs_info->freq_lo_jiffies = 0;
193 return freq_lo;
194 }
195 jiffies_total = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
196 jiffies_hi = (freq_avg - freq_lo) * jiffies_total;
197 jiffies_hi += ((freq_hi - freq_lo) / 2);
198 jiffies_hi /= (freq_hi - freq_lo);
199 jiffies_lo = jiffies_total - jiffies_hi;
200 dbs_info->freq_lo = freq_lo;
201 dbs_info->freq_lo_jiffies = jiffies_lo;
202 dbs_info->freq_hi_jiffies = jiffies_hi;
203 return freq_hi;
204}
205
206static void ondemand_powersave_bias_init(void)
207{
208 int i;
209 for_each_online_cpu(i) {
210 struct cpu_dbs_info_s *dbs_info = &per_cpu(cpu_dbs_info, i);
211 dbs_info->freq_table = cpufreq_frequency_get_table(i);
212 dbs_info->freq_lo = 0;
213 }
214}
215
Linus Torvalds1da177e2005-04-16 15:20:36 -0700216/************************** sysfs interface ************************/
217static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf)
218{
Thomas Renninger9411b4e2009-02-04 11:54:04 +0100219 static int print_once;
220
221 if (!print_once) {
222 printk(KERN_INFO "CPUFREQ: ondemand sampling_rate_max "
223 "sysfs file is deprecated - used by: %s\n",
224 current->comm);
225 print_once = 1;
226 }
Dave Jones2b03f892009-01-18 01:43:44 -0500227 return sprintf(buf, "%u\n", MAX_SAMPLING_RATE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700228}
229
230static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf)
231{
Thomas Renninger9411b4e2009-02-04 11:54:04 +0100232 static int print_once;
233
234 if (!print_once) {
235 printk(KERN_INFO "CPUFREQ: ondemand sampling_rate_min "
236 "sysfs file is deprecated - used by: %s\n",
237 current->comm);
238 print_once = 1;
239 }
Dave Jones2b03f892009-01-18 01:43:44 -0500240 return sprintf(buf, "%u\n", MIN_SAMPLING_RATE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700241}
242
Dave Jones32ee8c32006-02-28 00:43:23 -0500243#define define_one_ro(_name) \
244static struct freq_attr _name = \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700245__ATTR(_name, 0444, show_##_name, NULL)
246
247define_one_ro(sampling_rate_max);
248define_one_ro(sampling_rate_min);
249
250/* cpufreq_ondemand Governor Tunables */
251#define show_one(file_name, object) \
252static ssize_t show_##file_name \
253(struct cpufreq_policy *unused, char *buf) \
254{ \
255 return sprintf(buf, "%u\n", dbs_tuners_ins.object); \
256}
257show_one(sampling_rate, sampling_rate);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700258show_one(up_threshold, up_threshold);
Alexander Clouter001893c2005-12-01 01:09:25 -0800259show_one(ignore_nice_load, ignore_nice);
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400260show_one(powersave_bias, powersave_bias);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700261
Dave Jones32ee8c32006-02-28 00:43:23 -0500262static ssize_t store_sampling_rate(struct cpufreq_policy *unused,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700263 const char *buf, size_t count)
264{
265 unsigned int input;
266 int ret;
Venkatesh Pallipadiffac80e2006-06-28 13:52:18 -0700267 ret = sscanf(buf, "%u", &input);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700268
akpm@osdl.org3fc54d32006-01-13 15:54:22 -0800269 mutex_lock(&dbs_mutex);
Thomas Renninger112124a2009-02-04 11:55:12 +0100270 if (ret != 1) {
akpm@osdl.org3fc54d32006-01-13 15:54:22 -0800271 mutex_unlock(&dbs_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700272 return -EINVAL;
273 }
Thomas Renninger112124a2009-02-04 11:55:12 +0100274 dbs_tuners_ins.sampling_rate = max(input, minimum_sampling_rate());
akpm@osdl.org3fc54d32006-01-13 15:54:22 -0800275 mutex_unlock(&dbs_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700276
277 return count;
278}
279
Dave Jones32ee8c32006-02-28 00:43:23 -0500280static ssize_t store_up_threshold(struct cpufreq_policy *unused,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700281 const char *buf, size_t count)
282{
283 unsigned int input;
284 int ret;
Venkatesh Pallipadiffac80e2006-06-28 13:52:18 -0700285 ret = sscanf(buf, "%u", &input);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700286
akpm@osdl.org3fc54d32006-01-13 15:54:22 -0800287 mutex_lock(&dbs_mutex);
Dave Jones32ee8c32006-02-28 00:43:23 -0500288 if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
Dave Jonesc29f1402005-05-31 19:03:50 -0700289 input < MIN_FREQUENCY_UP_THRESHOLD) {
akpm@osdl.org3fc54d32006-01-13 15:54:22 -0800290 mutex_unlock(&dbs_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700291 return -EINVAL;
292 }
293
294 dbs_tuners_ins.up_threshold = input;
akpm@osdl.org3fc54d32006-01-13 15:54:22 -0800295 mutex_unlock(&dbs_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700296
297 return count;
298}
299
Alexander Clouter001893c2005-12-01 01:09:25 -0800300static ssize_t store_ignore_nice_load(struct cpufreq_policy *policy,
Dave Jones3d5ee9e2005-05-31 19:03:47 -0700301 const char *buf, size_t count)
302{
303 unsigned int input;
304 int ret;
305
306 unsigned int j;
Dave Jones32ee8c32006-02-28 00:43:23 -0500307
Venkatesh Pallipadiffac80e2006-06-28 13:52:18 -0700308 ret = sscanf(buf, "%u", &input);
Dave Jones2b03f892009-01-18 01:43:44 -0500309 if (ret != 1)
Dave Jones3d5ee9e2005-05-31 19:03:47 -0700310 return -EINVAL;
311
Dave Jones2b03f892009-01-18 01:43:44 -0500312 if (input > 1)
Dave Jones3d5ee9e2005-05-31 19:03:47 -0700313 input = 1;
Dave Jones32ee8c32006-02-28 00:43:23 -0500314
akpm@osdl.org3fc54d32006-01-13 15:54:22 -0800315 mutex_lock(&dbs_mutex);
Dave Jones2b03f892009-01-18 01:43:44 -0500316 if (input == dbs_tuners_ins.ignore_nice) { /* nothing to do */
akpm@osdl.org3fc54d32006-01-13 15:54:22 -0800317 mutex_unlock(&dbs_mutex);
Dave Jones3d5ee9e2005-05-31 19:03:47 -0700318 return count;
319 }
320 dbs_tuners_ins.ignore_nice = input;
321
Venkatesh Pallipadiccb2fe22006-06-28 13:49:52 -0700322 /* we need to re-evaluate prev_cpu_idle */
Dave Jonesdac1c1a2005-05-31 19:03:49 -0700323 for_each_online_cpu(j) {
Venkatesh Pallipadiccb2fe22006-06-28 13:49:52 -0700324 struct cpu_dbs_info_s *dbs_info;
325 dbs_info = &per_cpu(cpu_dbs_info, j);
venkatesh.pallipadi@intel.com34305022008-08-04 11:59:09 -0700326 dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
327 &dbs_info->prev_cpu_wall);
Venkatesh Pallipadi1ca3abd2009-01-23 09:25:02 -0500328 if (dbs_tuners_ins.ignore_nice)
329 dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
330
Dave Jones3d5ee9e2005-05-31 19:03:47 -0700331 }
akpm@osdl.org3fc54d32006-01-13 15:54:22 -0800332 mutex_unlock(&dbs_mutex);
Dave Jones3d5ee9e2005-05-31 19:03:47 -0700333
334 return count;
335}
336
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400337static ssize_t store_powersave_bias(struct cpufreq_policy *unused,
338 const char *buf, size_t count)
339{
340 unsigned int input;
341 int ret;
342 ret = sscanf(buf, "%u", &input);
343
344 if (ret != 1)
345 return -EINVAL;
346
347 if (input > 1000)
348 input = 1000;
349
350 mutex_lock(&dbs_mutex);
351 dbs_tuners_ins.powersave_bias = input;
352 ondemand_powersave_bias_init();
353 mutex_unlock(&dbs_mutex);
354
355 return count;
356}
357
Linus Torvalds1da177e2005-04-16 15:20:36 -0700358#define define_one_rw(_name) \
359static struct freq_attr _name = \
360__ATTR(_name, 0644, show_##_name, store_##_name)
361
362define_one_rw(sampling_rate);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700363define_one_rw(up_threshold);
Alexander Clouter001893c2005-12-01 01:09:25 -0800364define_one_rw(ignore_nice_load);
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400365define_one_rw(powersave_bias);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700366
Dave Jones2b03f892009-01-18 01:43:44 -0500367static struct attribute *dbs_attributes[] = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700368 &sampling_rate_max.attr,
369 &sampling_rate_min.attr,
370 &sampling_rate.attr,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700371 &up_threshold.attr,
Alexander Clouter001893c2005-12-01 01:09:25 -0800372 &ignore_nice_load.attr,
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400373 &powersave_bias.attr,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700374 NULL
375};
376
377static struct attribute_group dbs_attr_group = {
378 .attrs = dbs_attributes,
379 .name = "ondemand",
380};
381
382/************************** sysfs end ************************/
383
Venkatesh Pallipadi2f8a8352006-06-28 13:51:19 -0700384static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700385{
venkatesh.pallipadi@intel.comc43aa3b2008-08-04 11:59:08 -0700386 unsigned int max_load_freq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700387
388 struct cpufreq_policy *policy;
389 unsigned int j;
390
Linus Torvalds1da177e2005-04-16 15:20:36 -0700391 if (!this_dbs_info->enable)
392 return;
393
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400394 this_dbs_info->freq_lo = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700395 policy = this_dbs_info->cur_policy;
Venki Pallipadiea487612007-06-20 14:26:24 -0700396
Dave Jones32ee8c32006-02-28 00:43:23 -0500397 /*
Dave Jonesc29f1402005-05-31 19:03:50 -0700398 * Every sampling_rate, we check, if current idle time is less
399 * than 20% (default), then we try to increase frequency
Venkatesh Pallipadiccb2fe22006-06-28 13:49:52 -0700400 * Every sampling_rate, we look for a the lowest
Dave Jonesc29f1402005-05-31 19:03:50 -0700401 * frequency which can sustain the load while keeping idle time over
402 * 30%. If such a frequency exist, we try to decrease to this frequency.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700403 *
Dave Jones32ee8c32006-02-28 00:43:23 -0500404 * Any frequency increase takes it to the maximum frequency.
405 * Frequency reduction happens at minimum steps of
406 * 5% (default) of current frequency
Linus Torvalds1da177e2005-04-16 15:20:36 -0700407 */
408
venkatesh.pallipadi@intel.comc43aa3b2008-08-04 11:59:08 -0700409 /* Get Absolute Load - in terms of freq */
410 max_load_freq = 0;
411
Rusty Russell835481d2009-01-04 05:18:06 -0800412 for_each_cpu(j, policy->cpus) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700413 struct cpu_dbs_info_s *j_dbs_info;
venkatesh.pallipadi@intel.comc43aa3b2008-08-04 11:59:08 -0700414 cputime64_t cur_wall_time, cur_idle_time;
415 unsigned int idle_time, wall_time;
416 unsigned int load, load_freq;
417 int freq_avg;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700418
Linus Torvalds1da177e2005-04-16 15:20:36 -0700419 j_dbs_info = &per_cpu(cpu_dbs_info, j);
venkatesh.pallipadi@intel.com34305022008-08-04 11:59:09 -0700420
421 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time);
422
venkatesh.pallipadi@intel.comc43aa3b2008-08-04 11:59:08 -0700423 wall_time = (unsigned int) cputime64_sub(cur_wall_time,
424 j_dbs_info->prev_cpu_wall);
425 j_dbs_info->prev_cpu_wall = cur_wall_time;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700426
venkatesh.pallipadi@intel.comc43aa3b2008-08-04 11:59:08 -0700427 idle_time = (unsigned int) cputime64_sub(cur_idle_time,
428 j_dbs_info->prev_cpu_idle);
429 j_dbs_info->prev_cpu_idle = cur_idle_time;
430
Venkatesh Pallipadi1ca3abd2009-01-23 09:25:02 -0500431 if (dbs_tuners_ins.ignore_nice) {
432 cputime64_t cur_nice;
433 unsigned long cur_nice_jiffies;
434
435 cur_nice = cputime64_sub(kstat_cpu(j).cpustat.nice,
436 j_dbs_info->prev_cpu_nice);
437 /*
438 * Assumption: nice time between sampling periods will
439 * be less than 2^32 jiffies for 32 bit sys
440 */
441 cur_nice_jiffies = (unsigned long)
442 cputime64_to_jiffies64(cur_nice);
443
444 j_dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
445 idle_time += jiffies_to_usecs(cur_nice_jiffies);
446 }
447
venkatesh.pallipadi@intel.com34305022008-08-04 11:59:09 -0700448 if (unlikely(!wall_time || wall_time < idle_time))
venkatesh.pallipadi@intel.comc43aa3b2008-08-04 11:59:08 -0700449 continue;
venkatesh.pallipadi@intel.comc43aa3b2008-08-04 11:59:08 -0700450
451 load = 100 * (wall_time - idle_time) / wall_time;
452
453 freq_avg = __cpufreq_driver_getavg(policy, j);
454 if (freq_avg <= 0)
455 freq_avg = policy->cur;
456
457 load_freq = load * freq_avg;
458 if (load_freq > max_load_freq)
459 max_load_freq = load_freq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700460 }
461
Venkatesh Pallipadiccb2fe22006-06-28 13:49:52 -0700462 /* Check for frequency increase */
venkatesh.pallipadi@intel.comc43aa3b2008-08-04 11:59:08 -0700463 if (max_load_freq > dbs_tuners_ins.up_threshold * policy->cur) {
Dave Jonesc11420a2005-05-31 19:03:48 -0700464 /* if we are already at full speed then break out early */
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400465 if (!dbs_tuners_ins.powersave_bias) {
466 if (policy->cur == policy->max)
467 return;
Dave Jones32ee8c32006-02-28 00:43:23 -0500468
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400469 __cpufreq_driver_target(policy, policy->max,
470 CPUFREQ_RELATION_H);
471 } else {
472 int freq = powersave_bias_target(policy, policy->max,
473 CPUFREQ_RELATION_H);
474 __cpufreq_driver_target(policy, freq,
475 CPUFREQ_RELATION_L);
476 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700477 return;
478 }
479
480 /* Check for frequency decrease */
Dave Jonesc29f1402005-05-31 19:03:50 -0700481 /* if we cannot reduce the frequency anymore, break out early */
482 if (policy->cur == policy->min)
483 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700484
Dave Jonesc29f1402005-05-31 19:03:50 -0700485 /*
486 * The optimal frequency is the frequency that is the lowest that
487 * can support the current CPU usage without triggering the up
488 * policy. To be safe, we focus 10 points under the threshold.
489 */
venkatesh.pallipadi@intel.come9d95bf2008-08-04 11:59:10 -0700490 if (max_load_freq <
491 (dbs_tuners_ins.up_threshold - dbs_tuners_ins.down_differential) *
492 policy->cur) {
venkatesh.pallipadi@intel.comc43aa3b2008-08-04 11:59:08 -0700493 unsigned int freq_next;
venkatesh.pallipadi@intel.come9d95bf2008-08-04 11:59:10 -0700494 freq_next = max_load_freq /
495 (dbs_tuners_ins.up_threshold -
496 dbs_tuners_ins.down_differential);
Venkatesh Pallipadidfde5d62006-10-03 12:38:45 -0700497
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400498 if (!dbs_tuners_ins.powersave_bias) {
499 __cpufreq_driver_target(policy, freq_next,
500 CPUFREQ_RELATION_L);
501 } else {
502 int freq = powersave_bias_target(policy, freq_next,
503 CPUFREQ_RELATION_L);
504 __cpufreq_driver_target(policy, freq,
505 CPUFREQ_RELATION_L);
506 }
Venkatesh Pallipadiccb2fe22006-06-28 13:49:52 -0700507 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700508}
509
David Howellsc4028952006-11-22 14:57:56 +0000510static void do_dbs_timer(struct work_struct *work)
Dave Jones32ee8c32006-02-28 00:43:23 -0500511{
Venkatesh Pallipadi529af7a2007-02-05 16:12:44 -0800512 struct cpu_dbs_info_s *dbs_info =
513 container_of(work, struct cpu_dbs_info_s, work.work);
514 unsigned int cpu = dbs_info->cpu;
515 int sample_type = dbs_info->sample_type;
516
Alexey Starikovskiy1ce28d62006-07-31 22:25:20 +0400517 /* We want all CPUs to do sampling nearly on same jiffy */
518 int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
David Howellsc4028952006-11-22 14:57:56 +0000519
Alexey Starikovskiy1ce28d62006-07-31 22:25:20 +0400520 delay -= jiffies % delay;
Venkatesh Pallipadi2f8a8352006-06-28 13:51:19 -0700521
Venkatesh Pallipadi56463b72007-02-05 16:12:45 -0800522 if (lock_policy_rwsem_write(cpu) < 0)
Linus Torvalds2cd7cbd2006-07-23 12:05:00 -0700523 return;
Venkatesh Pallipadi56463b72007-02-05 16:12:45 -0800524
525 if (!dbs_info->enable) {
526 unlock_policy_rwsem_write(cpu);
527 return;
528 }
529
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400530 /* Common NORMAL_SAMPLE setup */
David Howellsc4028952006-11-22 14:57:56 +0000531 dbs_info->sample_type = DBS_NORMAL_SAMPLE;
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400532 if (!dbs_tuners_ins.powersave_bias ||
David Howellsc4028952006-11-22 14:57:56 +0000533 sample_type == DBS_NORMAL_SAMPLE) {
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400534 dbs_check_cpu(dbs_info);
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400535 if (dbs_info->freq_lo) {
536 /* Setup timer for SUB_SAMPLE */
David Howellsc4028952006-11-22 14:57:56 +0000537 dbs_info->sample_type = DBS_SUB_SAMPLE;
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400538 delay = dbs_info->freq_hi_jiffies;
539 }
540 } else {
541 __cpufreq_driver_target(dbs_info->cur_policy,
Dave Jones2b03f892009-01-18 01:43:44 -0500542 dbs_info->freq_lo, CPUFREQ_RELATION_H);
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400543 }
Alexey Starikovskiy1ce28d62006-07-31 22:25:20 +0400544 queue_delayed_work_on(cpu, kondemand_wq, &dbs_info->work, delay);
Venkatesh Pallipadi56463b72007-02-05 16:12:45 -0800545 unlock_policy_rwsem_write(cpu);
Dave Jones32ee8c32006-02-28 00:43:23 -0500546}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700547
Venkatesh Pallipadi529af7a2007-02-05 16:12:44 -0800548static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700549{
Alexey Starikovskiy1ce28d62006-07-31 22:25:20 +0400550 /* We want all CPUs to do sampling nearly on same jiffy */
551 int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
552 delay -= jiffies % delay;
Venkatesh Pallipadi2f8a8352006-06-28 13:51:19 -0700553
Dave Jonesc18a1482007-02-10 20:03:51 -0500554 dbs_info->enable = 1;
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400555 ondemand_powersave_bias_init();
David Howellsc4028952006-11-22 14:57:56 +0000556 dbs_info->sample_type = DBS_NORMAL_SAMPLE;
Venki Pallipadi28287032007-05-08 00:27:47 -0700557 INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer);
Venkatesh Pallipadi529af7a2007-02-05 16:12:44 -0800558 queue_delayed_work_on(dbs_info->cpu, kondemand_wq, &dbs_info->work,
Dave Jones2b03f892009-01-18 01:43:44 -0500559 delay);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700560}
561
Linus Torvalds2cd7cbd2006-07-23 12:05:00 -0700562static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700563{
Linus Torvalds2cd7cbd2006-07-23 12:05:00 -0700564 dbs_info->enable = 0;
565 cancel_delayed_work(&dbs_info->work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700566}
567
568static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
569 unsigned int event)
570{
571 unsigned int cpu = policy->cpu;
572 struct cpu_dbs_info_s *this_dbs_info;
573 unsigned int j;
Jeff Garzik914f7c32006-10-20 14:31:00 -0700574 int rc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700575
576 this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
577
578 switch (event) {
579 case CPUFREQ_GOV_START:
Venkatesh Pallipadiffac80e2006-06-28 13:52:18 -0700580 if ((!cpu_online(cpu)) || (!policy->cur))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700581 return -EINVAL;
582
Linus Torvalds1da177e2005-04-16 15:20:36 -0700583 if (this_dbs_info->enable) /* Already enabled */
584 break;
Dave Jones32ee8c32006-02-28 00:43:23 -0500585
akpm@osdl.org3fc54d32006-01-13 15:54:22 -0800586 mutex_lock(&dbs_mutex);
Venkatesh Pallipadi2f8a8352006-06-28 13:51:19 -0700587 dbs_enable++;
Jeff Garzik914f7c32006-10-20 14:31:00 -0700588
589 rc = sysfs_create_group(&policy->kobj, &dbs_attr_group);
590 if (rc) {
Jeff Garzik914f7c32006-10-20 14:31:00 -0700591 dbs_enable--;
592 mutex_unlock(&dbs_mutex);
593 return rc;
594 }
595
Rusty Russell835481d2009-01-04 05:18:06 -0800596 for_each_cpu(j, policy->cpus) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700597 struct cpu_dbs_info_s *j_dbs_info;
598 j_dbs_info = &per_cpu(cpu_dbs_info, j);
599 j_dbs_info->cur_policy = policy;
Dave Jones32ee8c32006-02-28 00:43:23 -0500600
venkatesh.pallipadi@intel.com34305022008-08-04 11:59:09 -0700601 j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
602 &j_dbs_info->prev_cpu_wall);
Venkatesh Pallipadi1ca3abd2009-01-23 09:25:02 -0500603 if (dbs_tuners_ins.ignore_nice) {
604 j_dbs_info->prev_cpu_nice =
605 kstat_cpu(j).cpustat.nice;
606 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700607 }
Venkatesh Pallipadi529af7a2007-02-05 16:12:44 -0800608 this_dbs_info->cpu = cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700609 /*
610 * Start the timerschedule work, when this governor
611 * is used for first time
612 */
613 if (dbs_enable == 1) {
614 unsigned int latency;
615 /* policy latency is in nS. Convert it to uS first */
Dave Jonesdf8b59b2005-09-20 12:39:35 -0700616 latency = policy->cpuinfo.transition_latency / 1000;
617 if (latency == 0)
618 latency = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700619
Thomas Renninger112124a2009-02-04 11:55:12 +0100620 def_sampling_rate =
621 max(latency * LATENCY_MULTIPLIER,
622 MIN_STAT_SAMPLING_RATE);
Dave Jonesdf8b59b2005-09-20 12:39:35 -0700623
Linus Torvalds1da177e2005-04-16 15:20:36 -0700624 dbs_tuners_ins.sampling_rate = def_sampling_rate;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700625 }
Venkatesh Pallipadi529af7a2007-02-05 16:12:44 -0800626 dbs_timer_init(this_dbs_info);
Dave Jones32ee8c32006-02-28 00:43:23 -0500627
akpm@osdl.org3fc54d32006-01-13 15:54:22 -0800628 mutex_unlock(&dbs_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700629 break;
630
631 case CPUFREQ_GOV_STOP:
akpm@osdl.org3fc54d32006-01-13 15:54:22 -0800632 mutex_lock(&dbs_mutex);
Linus Torvalds2cd7cbd2006-07-23 12:05:00 -0700633 dbs_timer_exit(this_dbs_info);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700634 sysfs_remove_group(&policy->kobj, &dbs_attr_group);
635 dbs_enable--;
akpm@osdl.org3fc54d32006-01-13 15:54:22 -0800636 mutex_unlock(&dbs_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700637
638 break;
639
640 case CPUFREQ_GOV_LIMITS:
akpm@osdl.org3fc54d32006-01-13 15:54:22 -0800641 mutex_lock(&dbs_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700642 if (policy->max < this_dbs_info->cur_policy->cur)
Venkatesh Pallipadiffac80e2006-06-28 13:52:18 -0700643 __cpufreq_driver_target(this_dbs_info->cur_policy,
Dave Jones2b03f892009-01-18 01:43:44 -0500644 policy->max, CPUFREQ_RELATION_H);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700645 else if (policy->min > this_dbs_info->cur_policy->cur)
Venkatesh Pallipadiffac80e2006-06-28 13:52:18 -0700646 __cpufreq_driver_target(this_dbs_info->cur_policy,
Dave Jones2b03f892009-01-18 01:43:44 -0500647 policy->min, CPUFREQ_RELATION_L);
akpm@osdl.org3fc54d32006-01-13 15:54:22 -0800648 mutex_unlock(&dbs_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700649 break;
650 }
651 return 0;
652}
653
Sven Wegenerc4d14bc2008-09-20 16:50:08 +0200654#ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
655static
656#endif
Thomas Renninger1c256242007-10-02 13:28:12 -0700657struct cpufreq_governor cpufreq_gov_ondemand = {
658 .name = "ondemand",
659 .governor = cpufreq_governor_dbs,
660 .max_transition_latency = TRANSITION_LATENCY_LIMIT,
661 .owner = THIS_MODULE,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700662};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700663
664static int __init cpufreq_gov_dbs_init(void)
665{
Akinobu Mita888a7942008-07-14 12:00:45 +0900666 int err;
venkatesh.pallipadi@intel.com80800912008-08-04 11:59:12 -0700667 cputime64_t wall;
Andrea Righi4f6e6b92008-09-18 10:43:40 +0000668 u64 idle_time;
669 int cpu = get_cpu();
venkatesh.pallipadi@intel.com80800912008-08-04 11:59:12 -0700670
Andrea Righi4f6e6b92008-09-18 10:43:40 +0000671 idle_time = get_cpu_idle_time_us(cpu, &wall);
672 put_cpu();
venkatesh.pallipadi@intel.com80800912008-08-04 11:59:12 -0700673 if (idle_time != -1ULL) {
674 /* Idle micro accounting is supported. Use finer thresholds */
675 dbs_tuners_ins.up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
676 dbs_tuners_ins.down_differential =
677 MICRO_FREQUENCY_DOWN_DIFFERENTIAL;
678 }
Akinobu Mita888a7942008-07-14 12:00:45 +0900679
Venkatesh Pallipadi56463b72007-02-05 16:12:45 -0800680 kondemand_wq = create_workqueue("kondemand");
681 if (!kondemand_wq) {
682 printk(KERN_ERR "Creation of kondemand failed\n");
683 return -EFAULT;
684 }
Akinobu Mita888a7942008-07-14 12:00:45 +0900685 err = cpufreq_register_governor(&cpufreq_gov_ondemand);
686 if (err)
687 destroy_workqueue(kondemand_wq);
688
689 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700690}
691
692static void __exit cpufreq_gov_dbs_exit(void)
693{
Thomas Renninger1c256242007-10-02 13:28:12 -0700694 cpufreq_unregister_governor(&cpufreq_gov_ondemand);
Venkatesh Pallipadi56463b72007-02-05 16:12:45 -0800695 destroy_workqueue(kondemand_wq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700696}
697
698
Venkatesh Pallipadiffac80e2006-06-28 13:52:18 -0700699MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
700MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
701MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
Dave Jones2b03f892009-01-18 01:43:44 -0500702 "Low Latency Frequency Transition capable processors");
Venkatesh Pallipadiffac80e2006-06-28 13:52:18 -0700703MODULE_LICENSE("GPL");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700704
Johannes Weiner69157192008-01-17 15:21:08 -0800705#ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
706fs_initcall(cpufreq_gov_dbs_init);
707#else
Linus Torvalds1da177e2005-04-16 15:20:36 -0700708module_init(cpufreq_gov_dbs_init);
Johannes Weiner69157192008-01-17 15:21:08 -0800709#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700710module_exit(cpufreq_gov_dbs_exit);