blob: e741c339df768e77e6f155658b16ef2757897afd [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
Mathieu Desnoyersb14893a2009-05-17 10:30:45 -0400101 * DEADLOCK ALERT! (2) : do_dbs_timer() must not take the dbs_mutex, because it
102 * would deadlock with cancel_delayed_work_sync(), which is needed for proper
103 * raceless workqueue teardown.
Venkatesh Pallipadi4ec223d2006-06-21 15:18:34 -0700104 */
Venkatesh Pallipadiffac80e2006-06-28 13:52:18 -0700105static DEFINE_MUTEX(dbs_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700106
Venkatesh Pallipadi2f8a8352006-06-28 13:51:19 -0700107static struct workqueue_struct *kondemand_wq;
Andi Kleen6810b542006-05-08 15:17:31 +0200108
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400109static struct dbs_tuners {
Dave Jones32ee8c32006-02-28 00:43:23 -0500110 unsigned int sampling_rate;
Dave Jones32ee8c32006-02-28 00:43:23 -0500111 unsigned int up_threshold;
venkatesh.pallipadi@intel.come9d95bf2008-08-04 11:59:10 -0700112 unsigned int down_differential;
Dave Jones32ee8c32006-02-28 00:43:23 -0500113 unsigned int ignore_nice;
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400114 unsigned int powersave_bias;
115} dbs_tuners_ins = {
Dave Jones32ee8c32006-02-28 00:43:23 -0500116 .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
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)
140 *wall = cur_wall_time;
141
Venki Pallipadiea487612007-06-20 14:26:24 -0700142 return 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{
147 u64 idle_time = get_cpu_idle_time_us(cpu, wall);
148
149 if (idle_time == -1ULL)
150 return get_cpu_idle_time_jiffy(cpu, wall);
151
venkatesh.pallipadi@intel.com80800912008-08-04 11:59:12 -0700152 return idle_time;
153}
154
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400155/*
156 * Find right freq to be set now with powersave_bias on.
157 * Returns the freq_hi to be used right now and will set freq_hi_jiffies,
158 * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs.
159 */
Adrian Bunkb5ecf602006-08-13 23:00:08 +0200160static unsigned int powersave_bias_target(struct cpufreq_policy *policy,
161 unsigned int freq_next,
162 unsigned int relation)
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400163{
164 unsigned int freq_req, freq_reduc, freq_avg;
165 unsigned int freq_hi, freq_lo;
166 unsigned int index = 0;
167 unsigned int jiffies_total, jiffies_hi, jiffies_lo;
168 struct cpu_dbs_info_s *dbs_info = &per_cpu(cpu_dbs_info, policy->cpu);
169
170 if (!dbs_info->freq_table) {
171 dbs_info->freq_lo = 0;
172 dbs_info->freq_lo_jiffies = 0;
173 return freq_next;
174 }
175
176 cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next,
177 relation, &index);
178 freq_req = dbs_info->freq_table[index].frequency;
179 freq_reduc = freq_req * dbs_tuners_ins.powersave_bias / 1000;
180 freq_avg = freq_req - freq_reduc;
181
182 /* Find freq bounds for freq_avg in freq_table */
183 index = 0;
184 cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
185 CPUFREQ_RELATION_H, &index);
186 freq_lo = dbs_info->freq_table[index].frequency;
187 index = 0;
188 cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
189 CPUFREQ_RELATION_L, &index);
190 freq_hi = dbs_info->freq_table[index].frequency;
191
192 /* Find out how long we have to be in hi and lo freqs */
193 if (freq_hi == freq_lo) {
194 dbs_info->freq_lo = 0;
195 dbs_info->freq_lo_jiffies = 0;
196 return freq_lo;
197 }
198 jiffies_total = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
199 jiffies_hi = (freq_avg - freq_lo) * jiffies_total;
200 jiffies_hi += ((freq_hi - freq_lo) / 2);
201 jiffies_hi /= (freq_hi - freq_lo);
202 jiffies_lo = jiffies_total - jiffies_hi;
203 dbs_info->freq_lo = freq_lo;
204 dbs_info->freq_lo_jiffies = jiffies_lo;
205 dbs_info->freq_hi_jiffies = jiffies_hi;
206 return freq_hi;
207}
208
209static void ondemand_powersave_bias_init(void)
210{
211 int i;
212 for_each_online_cpu(i) {
213 struct cpu_dbs_info_s *dbs_info = &per_cpu(cpu_dbs_info, i);
214 dbs_info->freq_table = cpufreq_frequency_get_table(i);
215 dbs_info->freq_lo = 0;
216 }
217}
218
Linus Torvalds1da177e2005-04-16 15:20:36 -0700219/************************** sysfs interface ************************/
220static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf)
221{
Thomas Renninger9411b4e2009-02-04 11:54:04 +0100222 static int print_once;
223
224 if (!print_once) {
225 printk(KERN_INFO "CPUFREQ: ondemand sampling_rate_max "
226 "sysfs file is deprecated - used by: %s\n",
227 current->comm);
228 print_once = 1;
229 }
Dave Jones2b03f892009-01-18 01:43:44 -0500230 return sprintf(buf, "%u\n", MAX_SAMPLING_RATE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700231}
232
233static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf)
234{
Thomas Renninger9411b4e2009-02-04 11:54:04 +0100235 static int print_once;
236
237 if (!print_once) {
238 printk(KERN_INFO "CPUFREQ: ondemand sampling_rate_min "
239 "sysfs file is deprecated - used by: %s\n",
240 current->comm);
241 print_once = 1;
242 }
Dave Jones2b03f892009-01-18 01:43:44 -0500243 return sprintf(buf, "%u\n", MIN_SAMPLING_RATE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700244}
245
Dave Jones32ee8c32006-02-28 00:43:23 -0500246#define define_one_ro(_name) \
247static struct freq_attr _name = \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700248__ATTR(_name, 0444, show_##_name, NULL)
249
250define_one_ro(sampling_rate_max);
251define_one_ro(sampling_rate_min);
252
253/* cpufreq_ondemand Governor Tunables */
254#define show_one(file_name, object) \
255static ssize_t show_##file_name \
256(struct cpufreq_policy *unused, char *buf) \
257{ \
258 return sprintf(buf, "%u\n", dbs_tuners_ins.object); \
259}
260show_one(sampling_rate, sampling_rate);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700261show_one(up_threshold, up_threshold);
Alexander Clouter001893c2005-12-01 01:09:25 -0800262show_one(ignore_nice_load, ignore_nice);
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400263show_one(powersave_bias, powersave_bias);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700264
Dave Jones32ee8c32006-02-28 00:43:23 -0500265static ssize_t store_sampling_rate(struct cpufreq_policy *unused,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700266 const char *buf, size_t count)
267{
268 unsigned int input;
269 int ret;
Venkatesh Pallipadiffac80e2006-06-28 13:52:18 -0700270 ret = sscanf(buf, "%u", &input);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700271
akpm@osdl.org3fc54d32006-01-13 15:54:22 -0800272 mutex_lock(&dbs_mutex);
Thomas Renninger112124a2009-02-04 11:55:12 +0100273 if (ret != 1) {
akpm@osdl.org3fc54d32006-01-13 15:54:22 -0800274 mutex_unlock(&dbs_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700275 return -EINVAL;
276 }
Thomas Renninger112124a2009-02-04 11:55:12 +0100277 dbs_tuners_ins.sampling_rate = max(input, minimum_sampling_rate());
akpm@osdl.org3fc54d32006-01-13 15:54:22 -0800278 mutex_unlock(&dbs_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700279
280 return count;
281}
282
Dave Jones32ee8c32006-02-28 00:43:23 -0500283static ssize_t store_up_threshold(struct cpufreq_policy *unused,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700284 const char *buf, size_t count)
285{
286 unsigned int input;
287 int ret;
Venkatesh Pallipadiffac80e2006-06-28 13:52:18 -0700288 ret = sscanf(buf, "%u", &input);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700289
akpm@osdl.org3fc54d32006-01-13 15:54:22 -0800290 mutex_lock(&dbs_mutex);
Dave Jones32ee8c32006-02-28 00:43:23 -0500291 if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
Dave Jonesc29f1402005-05-31 19:03:50 -0700292 input < MIN_FREQUENCY_UP_THRESHOLD) {
akpm@osdl.org3fc54d32006-01-13 15:54:22 -0800293 mutex_unlock(&dbs_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700294 return -EINVAL;
295 }
296
297 dbs_tuners_ins.up_threshold = input;
akpm@osdl.org3fc54d32006-01-13 15:54:22 -0800298 mutex_unlock(&dbs_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700299
300 return count;
301}
302
Alexander Clouter001893c2005-12-01 01:09:25 -0800303static ssize_t store_ignore_nice_load(struct cpufreq_policy *policy,
Dave Jones3d5ee9e2005-05-31 19:03:47 -0700304 const char *buf, size_t count)
305{
306 unsigned int input;
307 int ret;
308
309 unsigned int j;
Dave Jones32ee8c32006-02-28 00:43:23 -0500310
Venkatesh Pallipadiffac80e2006-06-28 13:52:18 -0700311 ret = sscanf(buf, "%u", &input);
Dave Jones2b03f892009-01-18 01:43:44 -0500312 if (ret != 1)
Dave Jones3d5ee9e2005-05-31 19:03:47 -0700313 return -EINVAL;
314
Dave Jones2b03f892009-01-18 01:43:44 -0500315 if (input > 1)
Dave Jones3d5ee9e2005-05-31 19:03:47 -0700316 input = 1;
Dave Jones32ee8c32006-02-28 00:43:23 -0500317
akpm@osdl.org3fc54d32006-01-13 15:54:22 -0800318 mutex_lock(&dbs_mutex);
Dave Jones2b03f892009-01-18 01:43:44 -0500319 if (input == dbs_tuners_ins.ignore_nice) { /* nothing to do */
akpm@osdl.org3fc54d32006-01-13 15:54:22 -0800320 mutex_unlock(&dbs_mutex);
Dave Jones3d5ee9e2005-05-31 19:03:47 -0700321 return count;
322 }
323 dbs_tuners_ins.ignore_nice = input;
324
Venkatesh Pallipadiccb2fe22006-06-28 13:49:52 -0700325 /* we need to re-evaluate prev_cpu_idle */
Dave Jonesdac1c1a2005-05-31 19:03:49 -0700326 for_each_online_cpu(j) {
Venkatesh Pallipadiccb2fe22006-06-28 13:49:52 -0700327 struct cpu_dbs_info_s *dbs_info;
328 dbs_info = &per_cpu(cpu_dbs_info, j);
venkatesh.pallipadi@intel.com34305022008-08-04 11:59:09 -0700329 dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
330 &dbs_info->prev_cpu_wall);
Venkatesh Pallipadi1ca3abd2009-01-23 09:25:02 -0500331 if (dbs_tuners_ins.ignore_nice)
332 dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
333
Dave Jones3d5ee9e2005-05-31 19:03:47 -0700334 }
akpm@osdl.org3fc54d32006-01-13 15:54:22 -0800335 mutex_unlock(&dbs_mutex);
Dave Jones3d5ee9e2005-05-31 19:03:47 -0700336
337 return count;
338}
339
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400340static ssize_t store_powersave_bias(struct cpufreq_policy *unused,
341 const char *buf, size_t count)
342{
343 unsigned int input;
344 int ret;
345 ret = sscanf(buf, "%u", &input);
346
347 if (ret != 1)
348 return -EINVAL;
349
350 if (input > 1000)
351 input = 1000;
352
353 mutex_lock(&dbs_mutex);
354 dbs_tuners_ins.powersave_bias = input;
355 ondemand_powersave_bias_init();
356 mutex_unlock(&dbs_mutex);
357
358 return count;
359}
360
Linus Torvalds1da177e2005-04-16 15:20:36 -0700361#define define_one_rw(_name) \
362static struct freq_attr _name = \
363__ATTR(_name, 0644, show_##_name, store_##_name)
364
365define_one_rw(sampling_rate);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700366define_one_rw(up_threshold);
Alexander Clouter001893c2005-12-01 01:09:25 -0800367define_one_rw(ignore_nice_load);
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400368define_one_rw(powersave_bias);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700369
Dave Jones2b03f892009-01-18 01:43:44 -0500370static struct attribute *dbs_attributes[] = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700371 &sampling_rate_max.attr,
372 &sampling_rate_min.attr,
373 &sampling_rate.attr,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700374 &up_threshold.attr,
Alexander Clouter001893c2005-12-01 01:09:25 -0800375 &ignore_nice_load.attr,
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400376 &powersave_bias.attr,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700377 NULL
378};
379
380static struct attribute_group dbs_attr_group = {
381 .attrs = dbs_attributes,
382 .name = "ondemand",
383};
384
385/************************** sysfs end ************************/
386
Venkatesh Pallipadi2f8a8352006-06-28 13:51:19 -0700387static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700388{
venkatesh.pallipadi@intel.comc43aa3b2008-08-04 11:59:08 -0700389 unsigned int max_load_freq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700390
391 struct cpufreq_policy *policy;
392 unsigned int j;
393
Linus Torvalds1da177e2005-04-16 15:20:36 -0700394 if (!this_dbs_info->enable)
395 return;
396
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400397 this_dbs_info->freq_lo = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700398 policy = this_dbs_info->cur_policy;
Venki Pallipadiea487612007-06-20 14:26:24 -0700399
Dave Jones32ee8c32006-02-28 00:43:23 -0500400 /*
Dave Jonesc29f1402005-05-31 19:03:50 -0700401 * Every sampling_rate, we check, if current idle time is less
402 * than 20% (default), then we try to increase frequency
Venkatesh Pallipadiccb2fe22006-06-28 13:49:52 -0700403 * Every sampling_rate, we look for a the lowest
Dave Jonesc29f1402005-05-31 19:03:50 -0700404 * frequency which can sustain the load while keeping idle time over
405 * 30%. If such a frequency exist, we try to decrease to this frequency.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700406 *
Dave Jones32ee8c32006-02-28 00:43:23 -0500407 * Any frequency increase takes it to the maximum frequency.
408 * Frequency reduction happens at minimum steps of
409 * 5% (default) of current frequency
Linus Torvalds1da177e2005-04-16 15:20:36 -0700410 */
411
venkatesh.pallipadi@intel.comc43aa3b2008-08-04 11:59:08 -0700412 /* Get Absolute Load - in terms of freq */
413 max_load_freq = 0;
414
Rusty Russell835481d2009-01-04 05:18:06 -0800415 for_each_cpu(j, policy->cpus) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416 struct cpu_dbs_info_s *j_dbs_info;
venkatesh.pallipadi@intel.comc43aa3b2008-08-04 11:59:08 -0700417 cputime64_t cur_wall_time, cur_idle_time;
418 unsigned int idle_time, wall_time;
419 unsigned int load, load_freq;
420 int freq_avg;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700421
Linus Torvalds1da177e2005-04-16 15:20:36 -0700422 j_dbs_info = &per_cpu(cpu_dbs_info, j);
venkatesh.pallipadi@intel.com34305022008-08-04 11:59:09 -0700423
424 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time);
425
venkatesh.pallipadi@intel.comc43aa3b2008-08-04 11:59:08 -0700426 wall_time = (unsigned int) cputime64_sub(cur_wall_time,
427 j_dbs_info->prev_cpu_wall);
428 j_dbs_info->prev_cpu_wall = cur_wall_time;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700429
venkatesh.pallipadi@intel.comc43aa3b2008-08-04 11:59:08 -0700430 idle_time = (unsigned int) cputime64_sub(cur_idle_time,
431 j_dbs_info->prev_cpu_idle);
432 j_dbs_info->prev_cpu_idle = cur_idle_time;
433
Venkatesh Pallipadi1ca3abd2009-01-23 09:25:02 -0500434 if (dbs_tuners_ins.ignore_nice) {
435 cputime64_t cur_nice;
436 unsigned long cur_nice_jiffies;
437
438 cur_nice = cputime64_sub(kstat_cpu(j).cpustat.nice,
439 j_dbs_info->prev_cpu_nice);
440 /*
441 * Assumption: nice time between sampling periods will
442 * be less than 2^32 jiffies for 32 bit sys
443 */
444 cur_nice_jiffies = (unsigned long)
445 cputime64_to_jiffies64(cur_nice);
446
447 j_dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
448 idle_time += jiffies_to_usecs(cur_nice_jiffies);
449 }
450
venkatesh.pallipadi@intel.com34305022008-08-04 11:59:09 -0700451 if (unlikely(!wall_time || wall_time < idle_time))
venkatesh.pallipadi@intel.comc43aa3b2008-08-04 11:59:08 -0700452 continue;
venkatesh.pallipadi@intel.comc43aa3b2008-08-04 11:59:08 -0700453
454 load = 100 * (wall_time - idle_time) / wall_time;
455
456 freq_avg = __cpufreq_driver_getavg(policy, j);
457 if (freq_avg <= 0)
458 freq_avg = policy->cur;
459
460 load_freq = load * freq_avg;
461 if (load_freq > max_load_freq)
462 max_load_freq = load_freq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463 }
464
Venkatesh Pallipadiccb2fe22006-06-28 13:49:52 -0700465 /* Check for frequency increase */
venkatesh.pallipadi@intel.comc43aa3b2008-08-04 11:59:08 -0700466 if (max_load_freq > dbs_tuners_ins.up_threshold * policy->cur) {
Dave Jonesc11420a2005-05-31 19:03:48 -0700467 /* if we are already at full speed then break out early */
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400468 if (!dbs_tuners_ins.powersave_bias) {
469 if (policy->cur == policy->max)
470 return;
Dave Jones32ee8c32006-02-28 00:43:23 -0500471
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400472 __cpufreq_driver_target(policy, policy->max,
473 CPUFREQ_RELATION_H);
474 } else {
475 int freq = powersave_bias_target(policy, policy->max,
476 CPUFREQ_RELATION_H);
477 __cpufreq_driver_target(policy, freq,
478 CPUFREQ_RELATION_L);
479 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700480 return;
481 }
482
483 /* Check for frequency decrease */
Dave Jonesc29f1402005-05-31 19:03:50 -0700484 /* if we cannot reduce the frequency anymore, break out early */
485 if (policy->cur == policy->min)
486 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700487
Dave Jonesc29f1402005-05-31 19:03:50 -0700488 /*
489 * The optimal frequency is the frequency that is the lowest that
490 * can support the current CPU usage without triggering the up
491 * policy. To be safe, we focus 10 points under the threshold.
492 */
venkatesh.pallipadi@intel.come9d95bf2008-08-04 11:59:10 -0700493 if (max_load_freq <
494 (dbs_tuners_ins.up_threshold - dbs_tuners_ins.down_differential) *
495 policy->cur) {
venkatesh.pallipadi@intel.comc43aa3b2008-08-04 11:59:08 -0700496 unsigned int freq_next;
venkatesh.pallipadi@intel.come9d95bf2008-08-04 11:59:10 -0700497 freq_next = max_load_freq /
498 (dbs_tuners_ins.up_threshold -
499 dbs_tuners_ins.down_differential);
Venkatesh Pallipadidfde5d62006-10-03 12:38:45 -0700500
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400501 if (!dbs_tuners_ins.powersave_bias) {
502 __cpufreq_driver_target(policy, freq_next,
503 CPUFREQ_RELATION_L);
504 } else {
505 int freq = powersave_bias_target(policy, freq_next,
506 CPUFREQ_RELATION_L);
507 __cpufreq_driver_target(policy, freq,
508 CPUFREQ_RELATION_L);
509 }
Venkatesh Pallipadiccb2fe22006-06-28 13:49:52 -0700510 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700511}
512
David Howellsc4028952006-11-22 14:57:56 +0000513static void do_dbs_timer(struct work_struct *work)
Dave Jones32ee8c32006-02-28 00:43:23 -0500514{
Venkatesh Pallipadi529af7a2007-02-05 16:12:44 -0800515 struct cpu_dbs_info_s *dbs_info =
516 container_of(work, struct cpu_dbs_info_s, work.work);
517 unsigned int cpu = dbs_info->cpu;
518 int sample_type = dbs_info->sample_type;
519
Alexey Starikovskiy1ce28d62006-07-31 22:25:20 +0400520 /* We want all CPUs to do sampling nearly on same jiffy */
521 int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
David Howellsc4028952006-11-22 14:57:56 +0000522
Alexey Starikovskiy1ce28d62006-07-31 22:25:20 +0400523 delay -= jiffies % delay;
Venkatesh Pallipadi2f8a8352006-06-28 13:51:19 -0700524
Venkatesh Pallipadi56463b72007-02-05 16:12:45 -0800525 if (lock_policy_rwsem_write(cpu) < 0)
Linus Torvalds2cd7cbd2006-07-23 12:05:00 -0700526 return;
Venkatesh Pallipadi56463b72007-02-05 16:12:45 -0800527
528 if (!dbs_info->enable) {
529 unlock_policy_rwsem_write(cpu);
530 return;
531 }
532
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400533 /* Common NORMAL_SAMPLE setup */
David Howellsc4028952006-11-22 14:57:56 +0000534 dbs_info->sample_type = DBS_NORMAL_SAMPLE;
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400535 if (!dbs_tuners_ins.powersave_bias ||
David Howellsc4028952006-11-22 14:57:56 +0000536 sample_type == DBS_NORMAL_SAMPLE) {
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400537 dbs_check_cpu(dbs_info);
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400538 if (dbs_info->freq_lo) {
539 /* Setup timer for SUB_SAMPLE */
David Howellsc4028952006-11-22 14:57:56 +0000540 dbs_info->sample_type = DBS_SUB_SAMPLE;
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400541 delay = dbs_info->freq_hi_jiffies;
542 }
543 } else {
544 __cpufreq_driver_target(dbs_info->cur_policy,
Dave Jones2b03f892009-01-18 01:43:44 -0500545 dbs_info->freq_lo, CPUFREQ_RELATION_H);
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400546 }
Alexey Starikovskiy1ce28d62006-07-31 22:25:20 +0400547 queue_delayed_work_on(cpu, kondemand_wq, &dbs_info->work, delay);
Venkatesh Pallipadi56463b72007-02-05 16:12:45 -0800548 unlock_policy_rwsem_write(cpu);
Dave Jones32ee8c32006-02-28 00:43:23 -0500549}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700550
Venkatesh Pallipadi529af7a2007-02-05 16:12:44 -0800551static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700552{
Alexey Starikovskiy1ce28d62006-07-31 22:25:20 +0400553 /* We want all CPUs to do sampling nearly on same jiffy */
554 int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
555 delay -= jiffies % delay;
Venkatesh Pallipadi2f8a8352006-06-28 13:51:19 -0700556
Dave Jonesc18a1482007-02-10 20:03:51 -0500557 dbs_info->enable = 1;
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400558 ondemand_powersave_bias_init();
David Howellsc4028952006-11-22 14:57:56 +0000559 dbs_info->sample_type = DBS_NORMAL_SAMPLE;
Venki Pallipadi28287032007-05-08 00:27:47 -0700560 INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer);
Venkatesh Pallipadi529af7a2007-02-05 16:12:44 -0800561 queue_delayed_work_on(dbs_info->cpu, kondemand_wq, &dbs_info->work,
Dave Jones2b03f892009-01-18 01:43:44 -0500562 delay);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700563}
564
Linus Torvalds2cd7cbd2006-07-23 12:05:00 -0700565static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700566{
Linus Torvalds2cd7cbd2006-07-23 12:05:00 -0700567 dbs_info->enable = 0;
Mathieu Desnoyersb14893a2009-05-17 10:30:45 -0400568 cancel_delayed_work_sync(&dbs_info->work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700569}
570
571static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
572 unsigned int event)
573{
574 unsigned int cpu = policy->cpu;
575 struct cpu_dbs_info_s *this_dbs_info;
576 unsigned int j;
Jeff Garzik914f7c32006-10-20 14:31:00 -0700577 int rc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700578
579 this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
580
581 switch (event) {
582 case CPUFREQ_GOV_START:
Venkatesh Pallipadiffac80e2006-06-28 13:52:18 -0700583 if ((!cpu_online(cpu)) || (!policy->cur))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700584 return -EINVAL;
585
Linus Torvalds1da177e2005-04-16 15:20:36 -0700586 if (this_dbs_info->enable) /* Already enabled */
587 break;
Dave Jones32ee8c32006-02-28 00:43:23 -0500588
akpm@osdl.org3fc54d32006-01-13 15:54:22 -0800589 mutex_lock(&dbs_mutex);
Venkatesh Pallipadi2f8a8352006-06-28 13:51:19 -0700590 dbs_enable++;
Jeff Garzik914f7c32006-10-20 14:31:00 -0700591
592 rc = sysfs_create_group(&policy->kobj, &dbs_attr_group);
593 if (rc) {
Jeff Garzik914f7c32006-10-20 14:31:00 -0700594 dbs_enable--;
595 mutex_unlock(&dbs_mutex);
596 return rc;
597 }
598
Rusty Russell835481d2009-01-04 05:18:06 -0800599 for_each_cpu(j, policy->cpus) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700600 struct cpu_dbs_info_s *j_dbs_info;
601 j_dbs_info = &per_cpu(cpu_dbs_info, j);
602 j_dbs_info->cur_policy = policy;
Dave Jones32ee8c32006-02-28 00:43:23 -0500603
venkatesh.pallipadi@intel.com34305022008-08-04 11:59:09 -0700604 j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
605 &j_dbs_info->prev_cpu_wall);
Venkatesh Pallipadi1ca3abd2009-01-23 09:25:02 -0500606 if (dbs_tuners_ins.ignore_nice) {
607 j_dbs_info->prev_cpu_nice =
608 kstat_cpu(j).cpustat.nice;
609 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700610 }
Venkatesh Pallipadi529af7a2007-02-05 16:12:44 -0800611 this_dbs_info->cpu = cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700612 /*
613 * Start the timerschedule work, when this governor
614 * is used for first time
615 */
616 if (dbs_enable == 1) {
617 unsigned int latency;
618 /* policy latency is in nS. Convert it to uS first */
Dave Jonesdf8b59b2005-09-20 12:39:35 -0700619 latency = policy->cpuinfo.transition_latency / 1000;
620 if (latency == 0)
621 latency = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700622
Thomas Renninger112124a2009-02-04 11:55:12 +0100623 def_sampling_rate =
624 max(latency * LATENCY_MULTIPLIER,
625 MIN_STAT_SAMPLING_RATE);
Dave Jonesdf8b59b2005-09-20 12:39:35 -0700626
Linus Torvalds1da177e2005-04-16 15:20:36 -0700627 dbs_tuners_ins.sampling_rate = def_sampling_rate;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700628 }
Venkatesh Pallipadi529af7a2007-02-05 16:12:44 -0800629 dbs_timer_init(this_dbs_info);
Dave Jones32ee8c32006-02-28 00:43:23 -0500630
akpm@osdl.org3fc54d32006-01-13 15:54:22 -0800631 mutex_unlock(&dbs_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700632 break;
633
634 case CPUFREQ_GOV_STOP:
akpm@osdl.org3fc54d32006-01-13 15:54:22 -0800635 mutex_lock(&dbs_mutex);
Linus Torvalds2cd7cbd2006-07-23 12:05:00 -0700636 dbs_timer_exit(this_dbs_info);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700637 sysfs_remove_group(&policy->kobj, &dbs_attr_group);
638 dbs_enable--;
akpm@osdl.org3fc54d32006-01-13 15:54:22 -0800639 mutex_unlock(&dbs_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700640
641 break;
642
643 case CPUFREQ_GOV_LIMITS:
akpm@osdl.org3fc54d32006-01-13 15:54:22 -0800644 mutex_lock(&dbs_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700645 if (policy->max < 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->max, CPUFREQ_RELATION_H);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700648 else if (policy->min > this_dbs_info->cur_policy->cur)
Venkatesh Pallipadiffac80e2006-06-28 13:52:18 -0700649 __cpufreq_driver_target(this_dbs_info->cur_policy,
Dave Jones2b03f892009-01-18 01:43:44 -0500650 policy->min, CPUFREQ_RELATION_L);
akpm@osdl.org3fc54d32006-01-13 15:54:22 -0800651 mutex_unlock(&dbs_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700652 break;
653 }
654 return 0;
655}
656
Sven Wegenerc4d14bc2008-09-20 16:50:08 +0200657#ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
658static
659#endif
Thomas Renninger1c256242007-10-02 13:28:12 -0700660struct cpufreq_governor cpufreq_gov_ondemand = {
661 .name = "ondemand",
662 .governor = cpufreq_governor_dbs,
663 .max_transition_latency = TRANSITION_LATENCY_LIMIT,
664 .owner = THIS_MODULE,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700665};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700666
667static int __init cpufreq_gov_dbs_init(void)
668{
Akinobu Mita888a7942008-07-14 12:00:45 +0900669 int err;
venkatesh.pallipadi@intel.com80800912008-08-04 11:59:12 -0700670 cputime64_t wall;
Andrea Righi4f6e6b92008-09-18 10:43:40 +0000671 u64 idle_time;
672 int cpu = get_cpu();
venkatesh.pallipadi@intel.com80800912008-08-04 11:59:12 -0700673
Andrea Righi4f6e6b92008-09-18 10:43:40 +0000674 idle_time = get_cpu_idle_time_us(cpu, &wall);
675 put_cpu();
venkatesh.pallipadi@intel.com80800912008-08-04 11:59:12 -0700676 if (idle_time != -1ULL) {
677 /* Idle micro accounting is supported. Use finer thresholds */
678 dbs_tuners_ins.up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
679 dbs_tuners_ins.down_differential =
680 MICRO_FREQUENCY_DOWN_DIFFERENTIAL;
681 }
Akinobu Mita888a7942008-07-14 12:00:45 +0900682
Venkatesh Pallipadi56463b72007-02-05 16:12:45 -0800683 kondemand_wq = create_workqueue("kondemand");
684 if (!kondemand_wq) {
685 printk(KERN_ERR "Creation of kondemand failed\n");
686 return -EFAULT;
687 }
Akinobu Mita888a7942008-07-14 12:00:45 +0900688 err = cpufreq_register_governor(&cpufreq_gov_ondemand);
689 if (err)
690 destroy_workqueue(kondemand_wq);
691
692 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700693}
694
695static void __exit cpufreq_gov_dbs_exit(void)
696{
Thomas Renninger1c256242007-10-02 13:28:12 -0700697 cpufreq_unregister_governor(&cpufreq_gov_ondemand);
Venkatesh Pallipadi56463b72007-02-05 16:12:45 -0800698 destroy_workqueue(kondemand_wq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700699}
700
701
Venkatesh Pallipadiffac80e2006-06-28 13:52:18 -0700702MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
703MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
704MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
Dave Jones2b03f892009-01-18 01:43:44 -0500705 "Low Latency Frequency Transition capable processors");
Venkatesh Pallipadiffac80e2006-06-28 13:52:18 -0700706MODULE_LICENSE("GPL");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700707
Johannes Weiner69157192008-01-17 15:21:08 -0800708#ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
709fs_initcall(cpufreq_gov_dbs_init);
710#else
Linus Torvalds1da177e2005-04-16 15:20:36 -0700711module_init(cpufreq_gov_dbs_init);
Johannes Weiner69157192008-01-17 15:21:08 -0800712#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700713module_exit(cpufreq_gov_dbs_exit);