blob: f3eb26cd848fa7afe495d09db9f6ad166c3501b6 [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
Viresh Kumar4471a342012-10-26 00:47:42 +020013#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
14
Linus Torvalds1da177e2005-04-16 15:20:36 -070015#include <linux/cpufreq.h>
Viresh Kumar4471a342012-10-26 00:47:42 +020016#include <linux/init.h>
17#include <linux/kernel.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070018#include <linux/kernel_stat.h>
Viresh Kumar4471a342012-10-26 00:47:42 +020019#include <linux/kobject.h>
20#include <linux/module.h>
akpm@osdl.org3fc54d32006-01-13 15:54:22 -080021#include <linux/mutex.h>
Viresh Kumar4471a342012-10-26 00:47:42 +020022#include <linux/percpu-defs.h>
23#include <linux/sysfs.h>
venkatesh.pallipadi@intel.com80800912008-08-04 11:59:12 -070024#include <linux/tick.h>
Viresh Kumar4471a342012-10-26 00:47:42 +020025#include <linux/types.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070026
Viresh Kumar4471a342012-10-26 00:47:42 +020027#include "cpufreq_governor.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070028
Stratos Karafotis06eb09d2013-02-08 17:24:18 +000029/* On-demand governor macros */
venkatesh.pallipadi@intel.come9d95bf2008-08-04 11:59:10 -070030#define DEF_FREQUENCY_DOWN_DIFFERENTIAL (10)
Linus Torvalds1da177e2005-04-16 15:20:36 -070031#define DEF_FREQUENCY_UP_THRESHOLD (80)
David C Niemi3f78a9f2010-10-06 16:54:24 -040032#define DEF_SAMPLING_DOWN_FACTOR (1)
33#define MAX_SAMPLING_DOWN_FACTOR (100000)
venkatesh.pallipadi@intel.com80800912008-08-04 11:59:12 -070034#define MICRO_FREQUENCY_DOWN_DIFFERENTIAL (3)
35#define MICRO_FREQUENCY_UP_THRESHOLD (95)
Thomas Renningercef96152009-04-22 13:48:29 +020036#define MICRO_FREQUENCY_MIN_SAMPLE_RATE (10000)
Dave Jonesc29f1402005-05-31 19:03:50 -070037#define MIN_FREQUENCY_UP_THRESHOLD (11)
Linus Torvalds1da177e2005-04-16 15:20:36 -070038#define MAX_FREQUENCY_UP_THRESHOLD (100)
39
Viresh Kumar4471a342012-10-26 00:47:42 +020040static struct dbs_data od_dbs_data;
41static DEFINE_PER_CPU(struct od_cpu_dbs_info_s, od_cpu_dbs_info);
Thomas Renninger112124a2009-02-04 11:55:12 +010042
Fabio Baltieri3e33ee92012-11-26 18:10:12 +000043#ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
44static struct cpufreq_governor cpufreq_gov_ondemand;
45#endif
46
Viresh Kumar4471a342012-10-26 00:47:42 +020047static struct od_dbs_tuners od_tuners = {
Dave Jones32ee8c32006-02-28 00:43:23 -050048 .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
David C Niemi3f78a9f2010-10-06 16:54:24 -040049 .sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR,
Stratos Karafotis4bd4e422013-02-06 13:34:00 +010050 .adj_up_threshold = DEF_FREQUENCY_UP_THRESHOLD -
51 DEF_FREQUENCY_DOWN_DIFFERENTIAL,
Eric Piel9cbad612006-03-10 11:35:27 +020052 .ignore_nice = 0,
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +040053 .powersave_bias = 0,
Linus Torvalds1da177e2005-04-16 15:20:36 -070054};
55
Viresh Kumar4471a342012-10-26 00:47:42 +020056static void ondemand_powersave_bias_init_cpu(int cpu)
Arjan van de Ven6b8fcd92010-05-09 08:26:06 -070057{
Viresh Kumar4471a342012-10-26 00:47:42 +020058 struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
Arjan van de Ven6b8fcd92010-05-09 08:26:06 -070059
Viresh Kumar4471a342012-10-26 00:47:42 +020060 dbs_info->freq_table = cpufreq_frequency_get_table(cpu);
61 dbs_info->freq_lo = 0;
62}
Arjan van de Ven6b8fcd92010-05-09 08:26:06 -070063
Viresh Kumar4471a342012-10-26 00:47:42 +020064/*
65 * Not all CPUs want IO time to be accounted as busy; this depends on how
66 * efficient idling at a higher frequency/voltage is.
67 * Pavel Machek says this is not so for various generations of AMD and old
68 * Intel systems.
Stratos Karafotis06eb09d2013-02-08 17:24:18 +000069 * Mike Chan (android.com) claims this is also not true for ARM.
Viresh Kumar4471a342012-10-26 00:47:42 +020070 * Because of this, whitelist specific known (series) of CPUs by default, and
71 * leave all others up to the user.
72 */
73static int should_io_be_busy(void)
74{
75#if defined(CONFIG_X86)
76 /*
Stratos Karafotis06eb09d2013-02-08 17:24:18 +000077 * For Intel, Core 2 (model 15) and later have an efficient idle.
Viresh Kumar4471a342012-10-26 00:47:42 +020078 */
79 if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
80 boot_cpu_data.x86 == 6 &&
81 boot_cpu_data.x86_model >= 15)
82 return 1;
83#endif
84 return 0;
Arjan van de Ven6b8fcd92010-05-09 08:26:06 -070085}
86
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +040087/*
88 * Find right freq to be set now with powersave_bias on.
89 * Returns the freq_hi to be used right now and will set freq_hi_jiffies,
90 * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs.
91 */
Adrian Bunkb5ecf602006-08-13 23:00:08 +020092static unsigned int powersave_bias_target(struct cpufreq_policy *policy,
Viresh Kumar4471a342012-10-26 00:47:42 +020093 unsigned int freq_next, unsigned int relation)
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +040094{
95 unsigned int freq_req, freq_reduc, freq_avg;
96 unsigned int freq_hi, freq_lo;
97 unsigned int index = 0;
98 unsigned int jiffies_total, jiffies_hi, jiffies_lo;
Viresh Kumar4471a342012-10-26 00:47:42 +020099 struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
Tejun Heo245b2e72009-06-24 15:13:48 +0900100 policy->cpu);
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400101
102 if (!dbs_info->freq_table) {
103 dbs_info->freq_lo = 0;
104 dbs_info->freq_lo_jiffies = 0;
105 return freq_next;
106 }
107
108 cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next,
109 relation, &index);
110 freq_req = dbs_info->freq_table[index].frequency;
Viresh Kumar4471a342012-10-26 00:47:42 +0200111 freq_reduc = freq_req * od_tuners.powersave_bias / 1000;
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400112 freq_avg = freq_req - freq_reduc;
113
114 /* Find freq bounds for freq_avg in freq_table */
115 index = 0;
116 cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
117 CPUFREQ_RELATION_H, &index);
118 freq_lo = dbs_info->freq_table[index].frequency;
119 index = 0;
120 cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
121 CPUFREQ_RELATION_L, &index);
122 freq_hi = dbs_info->freq_table[index].frequency;
123
124 /* Find out how long we have to be in hi and lo freqs */
125 if (freq_hi == freq_lo) {
126 dbs_info->freq_lo = 0;
127 dbs_info->freq_lo_jiffies = 0;
128 return freq_lo;
129 }
Viresh Kumar4471a342012-10-26 00:47:42 +0200130 jiffies_total = usecs_to_jiffies(od_tuners.sampling_rate);
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400131 jiffies_hi = (freq_avg - freq_lo) * jiffies_total;
132 jiffies_hi += ((freq_hi - freq_lo) / 2);
133 jiffies_hi /= (freq_hi - freq_lo);
134 jiffies_lo = jiffies_total - jiffies_hi;
135 dbs_info->freq_lo = freq_lo;
136 dbs_info->freq_lo_jiffies = jiffies_lo;
137 dbs_info->freq_hi_jiffies = jiffies_hi;
138 return freq_hi;
139}
140
141static void ondemand_powersave_bias_init(void)
142{
143 int i;
144 for_each_online_cpu(i) {
venkatesh.pallipadi@intel.com5a75c822009-07-02 17:08:32 -0700145 ondemand_powersave_bias_init_cpu(i);
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400146 }
147}
148
Viresh Kumar4471a342012-10-26 00:47:42 +0200149static void dbs_freq_increase(struct cpufreq_policy *p, unsigned int freq)
150{
151 if (od_tuners.powersave_bias)
152 freq = powersave_bias_target(p, freq, CPUFREQ_RELATION_H);
153 else if (p->cur == p->max)
154 return;
155
156 __cpufreq_driver_target(p, freq, od_tuners.powersave_bias ?
157 CPUFREQ_RELATION_L : CPUFREQ_RELATION_H);
158}
159
160/*
161 * Every sampling_rate, we check, if current idle time is less than 20%
Stratos Karafotis06eb09d2013-02-08 17:24:18 +0000162 * (default), then we try to increase frequency. Every sampling_rate, we look
163 * for the lowest frequency which can sustain the load while keeping idle time
Viresh Kumar4471a342012-10-26 00:47:42 +0200164 * over 30%. If such a frequency exist, we try to decrease to this frequency.
165 *
166 * Any frequency increase takes it to the maximum frequency. Frequency reduction
167 * happens at minimum steps of 5% (default) of current frequency
168 */
169static void od_check_cpu(int cpu, unsigned int load_freq)
170{
171 struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
172 struct cpufreq_policy *policy = dbs_info->cdbs.cur_policy;
173
174 dbs_info->freq_lo = 0;
175
176 /* Check for frequency increase */
177 if (load_freq > od_tuners.up_threshold * policy->cur) {
178 /* If switching to max speed, apply sampling_down_factor */
179 if (policy->cur < policy->max)
180 dbs_info->rate_mult =
181 od_tuners.sampling_down_factor;
182 dbs_freq_increase(policy, policy->max);
183 return;
184 }
185
186 /* Check for frequency decrease */
187 /* if we cannot reduce the frequency anymore, break out early */
188 if (policy->cur == policy->min)
189 return;
190
191 /*
192 * The optimal frequency is the frequency that is the lowest that can
193 * support the current CPU usage without triggering the up policy. To be
194 * safe, we focus 10 points under the threshold.
195 */
Stratos Karafotis4bd4e422013-02-06 13:34:00 +0100196 if (load_freq < od_tuners.adj_up_threshold * policy->cur) {
Viresh Kumar4471a342012-10-26 00:47:42 +0200197 unsigned int freq_next;
Stratos Karafotis4bd4e422013-02-06 13:34:00 +0100198 freq_next = load_freq / od_tuners.adj_up_threshold;
Viresh Kumar4471a342012-10-26 00:47:42 +0200199
200 /* No longer fully busy, reset rate_mult */
201 dbs_info->rate_mult = 1;
202
203 if (freq_next < policy->min)
204 freq_next = policy->min;
205
206 if (!od_tuners.powersave_bias) {
207 __cpufreq_driver_target(policy, freq_next,
208 CPUFREQ_RELATION_L);
209 } else {
210 int freq = powersave_bias_target(policy, freq_next,
211 CPUFREQ_RELATION_L);
212 __cpufreq_driver_target(policy, freq,
213 CPUFREQ_RELATION_L);
214 }
215 }
216}
217
Fabio Baltierida53d612012-12-27 14:55:40 +0000218static void od_dbs_timer(struct work_struct *work)
219{
220 struct delayed_work *dw = to_delayed_work(work);
221 struct od_cpu_dbs_info_s *dbs_info =
222 container_of(work, struct od_cpu_dbs_info_s, cdbs.work.work);
Viresh Kumar44472662013-01-31 17:28:02 +0000223 unsigned int cpu = dbs_info->cdbs.cur_policy->cpu;
224 struct od_cpu_dbs_info_s *core_dbs_info = &per_cpu(od_cpu_dbs_info,
225 cpu);
226 int delay, sample_type = core_dbs_info->sample_type;
227 bool eval_load;
Fabio Baltierida53d612012-12-27 14:55:40 +0000228
Viresh Kumar44472662013-01-31 17:28:02 +0000229 mutex_lock(&core_dbs_info->cdbs.timer_mutex);
230 eval_load = need_load_eval(&core_dbs_info->cdbs,
231 od_tuners.sampling_rate);
232
233 /* Common NORMAL_SAMPLE setup */
234 core_dbs_info->sample_type = OD_NORMAL_SAMPLE;
235 if (sample_type == OD_SUB_SAMPLE) {
236 delay = core_dbs_info->freq_lo_jiffies;
237 if (eval_load)
238 __cpufreq_driver_target(core_dbs_info->cdbs.cur_policy,
239 core_dbs_info->freq_lo,
240 CPUFREQ_RELATION_H);
Fabio Baltierida53d612012-12-27 14:55:40 +0000241 } else {
Viresh Kumar44472662013-01-31 17:28:02 +0000242 if (eval_load)
243 dbs_check_cpu(&od_dbs_data, cpu);
244 if (core_dbs_info->freq_lo) {
245 /* Setup timer for SUB_SAMPLE */
246 core_dbs_info->sample_type = OD_SUB_SAMPLE;
247 delay = core_dbs_info->freq_hi_jiffies;
248 } else {
249 delay = delay_for_sampling_rate(od_tuners.sampling_rate
250 * core_dbs_info->rate_mult);
251 }
Fabio Baltierida53d612012-12-27 14:55:40 +0000252 }
Viresh Kumar44472662013-01-31 17:28:02 +0000253
254 schedule_delayed_work_on(smp_processor_id(), dw, delay);
255 mutex_unlock(&core_dbs_info->cdbs.timer_mutex);
Fabio Baltierida53d612012-12-27 14:55:40 +0000256}
257
Linus Torvalds1da177e2005-04-16 15:20:36 -0700258/************************** sysfs interface ************************/
Thomas Renninger0e625ac2009-07-24 15:25:06 +0200259
Thomas Renninger0e625ac2009-07-24 15:25:06 +0200260static ssize_t show_sampling_rate_min(struct kobject *kobj,
261 struct attribute *attr, char *buf)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700262{
Viresh Kumar4471a342012-10-26 00:47:42 +0200263 return sprintf(buf, "%u\n", od_dbs_data.min_sampling_rate);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700264}
265
MyungJoo Hamfd0ef7a2012-02-29 17:54:41 +0900266/**
267 * update_sampling_rate - update sampling rate effective immediately if needed.
268 * @new_rate: new sampling rate
269 *
Stratos Karafotis06eb09d2013-02-08 17:24:18 +0000270 * If new rate is smaller than the old, simply updating
Viresh Kumar4471a342012-10-26 00:47:42 +0200271 * dbs_tuners_int.sampling_rate might not be appropriate. For example, if the
272 * original sampling_rate was 1 second and the requested new sampling rate is 10
273 * ms because the user needs immediate reaction from ondemand governor, but not
274 * sure if higher frequency will be required or not, then, the governor may
275 * change the sampling rate too late; up to 1 second later. Thus, if we are
276 * reducing the sampling rate, we need to make the new value effective
277 * immediately.
MyungJoo Hamfd0ef7a2012-02-29 17:54:41 +0900278 */
279static void update_sampling_rate(unsigned int new_rate)
280{
281 int cpu;
282
Viresh Kumar4471a342012-10-26 00:47:42 +0200283 od_tuners.sampling_rate = new_rate = max(new_rate,
284 od_dbs_data.min_sampling_rate);
MyungJoo Hamfd0ef7a2012-02-29 17:54:41 +0900285
286 for_each_online_cpu(cpu) {
287 struct cpufreq_policy *policy;
Viresh Kumar4471a342012-10-26 00:47:42 +0200288 struct od_cpu_dbs_info_s *dbs_info;
MyungJoo Hamfd0ef7a2012-02-29 17:54:41 +0900289 unsigned long next_sampling, appointed_at;
290
291 policy = cpufreq_cpu_get(cpu);
292 if (!policy)
293 continue;
Fabio Baltieri3e33ee92012-11-26 18:10:12 +0000294 if (policy->governor != &cpufreq_gov_ondemand) {
295 cpufreq_cpu_put(policy);
296 continue;
297 }
Fabio Baltieri8ee2ec52012-12-27 14:55:42 +0000298 dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
MyungJoo Hamfd0ef7a2012-02-29 17:54:41 +0900299 cpufreq_cpu_put(policy);
300
Viresh Kumar4471a342012-10-26 00:47:42 +0200301 mutex_lock(&dbs_info->cdbs.timer_mutex);
MyungJoo Hamfd0ef7a2012-02-29 17:54:41 +0900302
Viresh Kumar4471a342012-10-26 00:47:42 +0200303 if (!delayed_work_pending(&dbs_info->cdbs.work)) {
304 mutex_unlock(&dbs_info->cdbs.timer_mutex);
MyungJoo Hamfd0ef7a2012-02-29 17:54:41 +0900305 continue;
306 }
307
Viresh Kumar4471a342012-10-26 00:47:42 +0200308 next_sampling = jiffies + usecs_to_jiffies(new_rate);
309 appointed_at = dbs_info->cdbs.work.timer.expires;
MyungJoo Hamfd0ef7a2012-02-29 17:54:41 +0900310
311 if (time_before(next_sampling, appointed_at)) {
312
Viresh Kumar4471a342012-10-26 00:47:42 +0200313 mutex_unlock(&dbs_info->cdbs.timer_mutex);
314 cancel_delayed_work_sync(&dbs_info->cdbs.work);
315 mutex_lock(&dbs_info->cdbs.timer_mutex);
MyungJoo Hamfd0ef7a2012-02-29 17:54:41 +0900316
Fabio Baltieri8ee2ec52012-12-27 14:55:42 +0000317 schedule_delayed_work_on(cpu, &dbs_info->cdbs.work,
Viresh Kumar4471a342012-10-26 00:47:42 +0200318 usecs_to_jiffies(new_rate));
MyungJoo Hamfd0ef7a2012-02-29 17:54:41 +0900319
320 }
Viresh Kumar4471a342012-10-26 00:47:42 +0200321 mutex_unlock(&dbs_info->cdbs.timer_mutex);
MyungJoo Hamfd0ef7a2012-02-29 17:54:41 +0900322 }
323}
324
Thomas Renninger0e625ac2009-07-24 15:25:06 +0200325static ssize_t store_sampling_rate(struct kobject *a, struct attribute *b,
326 const char *buf, size_t count)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700327{
328 unsigned int input;
329 int ret;
Venkatesh Pallipadiffac80e2006-06-28 13:52:18 -0700330 ret = sscanf(buf, "%u", &input);
venkatesh.pallipadi@intel.com5a75c822009-07-02 17:08:32 -0700331 if (ret != 1)
332 return -EINVAL;
MyungJoo Hamfd0ef7a2012-02-29 17:54:41 +0900333 update_sampling_rate(input);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700334 return count;
335}
336
Arjan van de Ven19379b12010-05-09 08:26:51 -0700337static ssize_t store_io_is_busy(struct kobject *a, struct attribute *b,
338 const char *buf, size_t count)
339{
340 unsigned int input;
341 int ret;
342
343 ret = sscanf(buf, "%u", &input);
344 if (ret != 1)
345 return -EINVAL;
Viresh Kumar4471a342012-10-26 00:47:42 +0200346 od_tuners.io_is_busy = !!input;
Arjan van de Ven19379b12010-05-09 08:26:51 -0700347 return count;
348}
349
Thomas Renninger0e625ac2009-07-24 15:25:06 +0200350static ssize_t store_up_threshold(struct kobject *a, struct attribute *b,
351 const char *buf, size_t count)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700352{
353 unsigned int input;
354 int ret;
Venkatesh Pallipadiffac80e2006-06-28 13:52:18 -0700355 ret = sscanf(buf, "%u", &input);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700356
Dave Jones32ee8c32006-02-28 00:43:23 -0500357 if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
Dave Jonesc29f1402005-05-31 19:03:50 -0700358 input < MIN_FREQUENCY_UP_THRESHOLD) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700359 return -EINVAL;
360 }
Stratos Karafotis4bd4e422013-02-06 13:34:00 +0100361 /* Calculate the new adj_up_threshold */
362 od_tuners.adj_up_threshold += input;
363 od_tuners.adj_up_threshold -= od_tuners.up_threshold;
364
Viresh Kumar4471a342012-10-26 00:47:42 +0200365 od_tuners.up_threshold = input;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700366 return count;
367}
368
David C Niemi3f78a9f2010-10-06 16:54:24 -0400369static ssize_t store_sampling_down_factor(struct kobject *a,
370 struct attribute *b, const char *buf, size_t count)
371{
372 unsigned int input, j;
373 int ret;
374 ret = sscanf(buf, "%u", &input);
375
376 if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
377 return -EINVAL;
Viresh Kumar4471a342012-10-26 00:47:42 +0200378 od_tuners.sampling_down_factor = input;
David C Niemi3f78a9f2010-10-06 16:54:24 -0400379
380 /* Reset down sampling multiplier in case it was active */
381 for_each_online_cpu(j) {
Viresh Kumar4471a342012-10-26 00:47:42 +0200382 struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
383 j);
David C Niemi3f78a9f2010-10-06 16:54:24 -0400384 dbs_info->rate_mult = 1;
385 }
David C Niemi3f78a9f2010-10-06 16:54:24 -0400386 return count;
387}
388
Thomas Renninger0e625ac2009-07-24 15:25:06 +0200389static ssize_t store_ignore_nice_load(struct kobject *a, struct attribute *b,
390 const char *buf, size_t count)
Dave Jones3d5ee9e2005-05-31 19:03:47 -0700391{
392 unsigned int input;
393 int ret;
394
395 unsigned int j;
Dave Jones32ee8c32006-02-28 00:43:23 -0500396
Venkatesh Pallipadiffac80e2006-06-28 13:52:18 -0700397 ret = sscanf(buf, "%u", &input);
Dave Jones2b03f892009-01-18 01:43:44 -0500398 if (ret != 1)
Dave Jones3d5ee9e2005-05-31 19:03:47 -0700399 return -EINVAL;
400
Dave Jones2b03f892009-01-18 01:43:44 -0500401 if (input > 1)
Dave Jones3d5ee9e2005-05-31 19:03:47 -0700402 input = 1;
Dave Jones32ee8c32006-02-28 00:43:23 -0500403
Viresh Kumar4471a342012-10-26 00:47:42 +0200404 if (input == od_tuners.ignore_nice) { /* nothing to do */
Dave Jones3d5ee9e2005-05-31 19:03:47 -0700405 return count;
406 }
Viresh Kumar4471a342012-10-26 00:47:42 +0200407 od_tuners.ignore_nice = input;
Dave Jones3d5ee9e2005-05-31 19:03:47 -0700408
Venkatesh Pallipadiccb2fe22006-06-28 13:49:52 -0700409 /* we need to re-evaluate prev_cpu_idle */
Dave Jonesdac1c1a2005-05-31 19:03:49 -0700410 for_each_online_cpu(j) {
Viresh Kumar4471a342012-10-26 00:47:42 +0200411 struct od_cpu_dbs_info_s *dbs_info;
Tejun Heo245b2e72009-06-24 15:13:48 +0900412 dbs_info = &per_cpu(od_cpu_dbs_info, j);
Viresh Kumar4471a342012-10-26 00:47:42 +0200413 dbs_info->cdbs.prev_cpu_idle = get_cpu_idle_time(j,
414 &dbs_info->cdbs.prev_cpu_wall);
415 if (od_tuners.ignore_nice)
416 dbs_info->cdbs.prev_cpu_nice =
417 kcpustat_cpu(j).cpustat[CPUTIME_NICE];
Venkatesh Pallipadi1ca3abd2009-01-23 09:25:02 -0500418
Dave Jones3d5ee9e2005-05-31 19:03:47 -0700419 }
Dave Jones3d5ee9e2005-05-31 19:03:47 -0700420 return count;
421}
422
Thomas Renninger0e625ac2009-07-24 15:25:06 +0200423static ssize_t store_powersave_bias(struct kobject *a, struct attribute *b,
424 const char *buf, size_t count)
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400425{
426 unsigned int input;
427 int ret;
428 ret = sscanf(buf, "%u", &input);
429
430 if (ret != 1)
431 return -EINVAL;
432
433 if (input > 1000)
434 input = 1000;
435
Viresh Kumar4471a342012-10-26 00:47:42 +0200436 od_tuners.powersave_bias = input;
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400437 ondemand_powersave_bias_init();
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400438 return count;
439}
440
Viresh Kumar4471a342012-10-26 00:47:42 +0200441show_one(od, sampling_rate, sampling_rate);
442show_one(od, io_is_busy, io_is_busy);
443show_one(od, up_threshold, up_threshold);
444show_one(od, sampling_down_factor, sampling_down_factor);
445show_one(od, ignore_nice_load, ignore_nice);
446show_one(od, powersave_bias, powersave_bias);
447
Borislav Petkov6dad2a22010-03-31 21:56:46 +0200448define_one_global_rw(sampling_rate);
Linus Torvalds07d77752010-05-18 08:49:13 -0700449define_one_global_rw(io_is_busy);
Borislav Petkov6dad2a22010-03-31 21:56:46 +0200450define_one_global_rw(up_threshold);
David C Niemi3f78a9f2010-10-06 16:54:24 -0400451define_one_global_rw(sampling_down_factor);
Borislav Petkov6dad2a22010-03-31 21:56:46 +0200452define_one_global_rw(ignore_nice_load);
453define_one_global_rw(powersave_bias);
Viresh Kumar4471a342012-10-26 00:47:42 +0200454define_one_global_ro(sampling_rate_min);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700455
Dave Jones2b03f892009-01-18 01:43:44 -0500456static struct attribute *dbs_attributes[] = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457 &sampling_rate_min.attr,
458 &sampling_rate.attr,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700459 &up_threshold.attr,
David C Niemi3f78a9f2010-10-06 16:54:24 -0400460 &sampling_down_factor.attr,
Alexander Clouter001893c2005-12-01 01:09:25 -0800461 &ignore_nice_load.attr,
Alexey Starikovskiy05ca0352006-07-31 22:28:12 +0400462 &powersave_bias.attr,
Arjan van de Ven19379b12010-05-09 08:26:51 -0700463 &io_is_busy.attr,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700464 NULL
465};
466
Viresh Kumar4471a342012-10-26 00:47:42 +0200467static struct attribute_group od_attr_group = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700468 .attrs = dbs_attributes,
469 .name = "ondemand",
470};
471
472/************************** sysfs end ************************/
473
Viresh Kumar4471a342012-10-26 00:47:42 +0200474define_get_cpu_dbs_routines(od_cpu_dbs_info);
Mike Chan00e299f2010-01-26 17:06:47 -0800475
Viresh Kumar4471a342012-10-26 00:47:42 +0200476static struct od_ops od_ops = {
477 .io_busy = should_io_be_busy,
478 .powersave_bias_init_cpu = ondemand_powersave_bias_init_cpu,
479 .powersave_bias_target = powersave_bias_target,
480 .freq_increase = dbs_freq_increase,
481};
482
483static struct dbs_data od_dbs_data = {
484 .governor = GOV_ONDEMAND,
485 .attr_group = &od_attr_group,
486 .tuners = &od_tuners,
487 .get_cpu_cdbs = get_cpu_cdbs,
488 .get_cpu_dbs_info_s = get_cpu_dbs_info_s,
489 .gov_dbs_timer = od_dbs_timer,
490 .gov_check_cpu = od_check_cpu,
491 .gov_ops = &od_ops,
492};
493
494static int od_cpufreq_governor_dbs(struct cpufreq_policy *policy,
495 unsigned int event)
496{
497 return cpufreq_governor_dbs(&od_dbs_data, policy, event);
Mike Chan00e299f2010-01-26 17:06:47 -0800498}
499
Viresh Kumar4471a342012-10-26 00:47:42 +0200500#ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
501static
Arjan van de Ven19379b12010-05-09 08:26:51 -0700502#endif
Viresh Kumar4471a342012-10-26 00:47:42 +0200503struct cpufreq_governor cpufreq_gov_ondemand = {
504 .name = "ondemand",
505 .governor = od_cpufreq_governor_dbs,
506 .max_transition_latency = TRANSITION_LATENCY_LIMIT,
507 .owner = THIS_MODULE,
508};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700509
Linus Torvalds1da177e2005-04-16 15:20:36 -0700510static int __init cpufreq_gov_dbs_init(void)
511{
Andrea Righi4f6e6b92008-09-18 10:43:40 +0000512 u64 idle_time;
513 int cpu = get_cpu();
venkatesh.pallipadi@intel.com80800912008-08-04 11:59:12 -0700514
Viresh Kumar4471a342012-10-26 00:47:42 +0200515 mutex_init(&od_dbs_data.mutex);
Kamalesh Babulal21f2e3c2011-12-09 16:18:42 +0530516 idle_time = get_cpu_idle_time_us(cpu, NULL);
Andrea Righi4f6e6b92008-09-18 10:43:40 +0000517 put_cpu();
venkatesh.pallipadi@intel.com80800912008-08-04 11:59:12 -0700518 if (idle_time != -1ULL) {
519 /* Idle micro accounting is supported. Use finer thresholds */
Viresh Kumar4471a342012-10-26 00:47:42 +0200520 od_tuners.up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
Stratos Karafotis4bd4e422013-02-06 13:34:00 +0100521 od_tuners.adj_up_threshold = MICRO_FREQUENCY_UP_THRESHOLD -
522 MICRO_FREQUENCY_DOWN_DIFFERENTIAL;
Thomas Renningercef96152009-04-22 13:48:29 +0200523 /*
Paul Bollebd74b322011-08-06 14:33:43 +0200524 * In nohz/micro accounting case we set the minimum frequency
Thomas Renningercef96152009-04-22 13:48:29 +0200525 * not depending on HZ, but fixed (very low). The deferred
526 * timer might skip some samples if idle/sleeping as needed.
527 */
Viresh Kumar4471a342012-10-26 00:47:42 +0200528 od_dbs_data.min_sampling_rate = MICRO_FREQUENCY_MIN_SAMPLE_RATE;
Thomas Renningercef96152009-04-22 13:48:29 +0200529 } else {
530 /* For correct statistics, we need 10 ticks for each measure */
Viresh Kumar4471a342012-10-26 00:47:42 +0200531 od_dbs_data.min_sampling_rate = MIN_SAMPLING_RATE_RATIO *
532 jiffies_to_usecs(10);
venkatesh.pallipadi@intel.com80800912008-08-04 11:59:12 -0700533 }
Akinobu Mita888a7942008-07-14 12:00:45 +0900534
Tejun Heo57df5572011-01-26 12:12:50 +0100535 return cpufreq_register_governor(&cpufreq_gov_ondemand);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700536}
537
538static void __exit cpufreq_gov_dbs_exit(void)
539{
Thomas Renninger1c256242007-10-02 13:28:12 -0700540 cpufreq_unregister_governor(&cpufreq_gov_ondemand);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700541}
542
Venkatesh Pallipadiffac80e2006-06-28 13:52:18 -0700543MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
544MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
545MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
Dave Jones2b03f892009-01-18 01:43:44 -0500546 "Low Latency Frequency Transition capable processors");
Venkatesh Pallipadiffac80e2006-06-28 13:52:18 -0700547MODULE_LICENSE("GPL");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700548
Johannes Weiner69157192008-01-17 15:21:08 -0800549#ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
550fs_initcall(cpufreq_gov_dbs_init);
551#else
Linus Torvalds1da177e2005-04-16 15:20:36 -0700552module_init(cpufreq_gov_dbs_init);
Johannes Weiner69157192008-01-17 15:21:08 -0800553#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700554module_exit(cpufreq_gov_dbs_exit);