ia64/kvm: compilation fix. export account_system_vtime.
[pv_ops_mirror.git] / drivers / cpufreq / cpufreq_conservative.c
blob5d3a04ba6ad2c38aaf18d443b508511206bfac0c
1 /*
2 * drivers/cpufreq/cpufreq_conservative.c
4 * Copyright (C) 2001 Russell King
5 * (C) 2003 Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>.
6 * Jun Nakajima <jun.nakajima@intel.com>
7 * (C) 2004 Alexander Clouter <alex-kernel@digriz.org.uk>
9 * This program is free software; you can redistribute it and/or modify
10 * it under the terms of the GNU General Public License version 2 as
11 * published by the Free Software Foundation.
14 #include <linux/kernel.h>
15 #include <linux/module.h>
16 #include <linux/smp.h>
17 #include <linux/init.h>
18 #include <linux/interrupt.h>
19 #include <linux/ctype.h>
20 #include <linux/cpufreq.h>
21 #include <linux/sysctl.h>
22 #include <linux/types.h>
23 #include <linux/fs.h>
24 #include <linux/sysfs.h>
25 #include <linux/cpu.h>
26 #include <linux/kmod.h>
27 #include <linux/workqueue.h>
28 #include <linux/jiffies.h>
29 #include <linux/kernel_stat.h>
30 #include <linux/percpu.h>
31 #include <linux/mutex.h>
33 * dbs is used in this file as a shortform for demandbased switching
34 * It helps to keep variable names smaller, simpler
37 #define DEF_FREQUENCY_UP_THRESHOLD (80)
38 #define DEF_FREQUENCY_DOWN_THRESHOLD (20)
41 * The polling frequency of this governor depends on the capability of
42 * the processor. Default polling frequency is 1000 times the transition
43 * latency of the processor. The governor will work on any processor with
44 * transition latency <= 10mS, using appropriate sampling
45 * rate.
46 * For CPUs with transition latency > 10mS (mostly drivers
47 * with CPUFREQ_ETERNAL), this governor will not work.
48 * All times here are in uS.
50 static unsigned int def_sampling_rate;
51 #define MIN_SAMPLING_RATE_RATIO (2)
52 /* for correct statistics, we need at least 10 ticks between each measure */
53 #define MIN_STAT_SAMPLING_RATE \
54 (MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10))
55 #define MIN_SAMPLING_RATE \
56 (def_sampling_rate / MIN_SAMPLING_RATE_RATIO)
57 #define MAX_SAMPLING_RATE (500 * def_sampling_rate)
58 #define DEF_SAMPLING_RATE_LATENCY_MULTIPLIER (1000)
59 #define DEF_SAMPLING_DOWN_FACTOR (1)
60 #define MAX_SAMPLING_DOWN_FACTOR (10)
61 #define TRANSITION_LATENCY_LIMIT (10 * 1000 * 1000)
63 static void do_dbs_timer(struct work_struct *work);
65 struct cpu_dbs_info_s {
66 struct cpufreq_policy *cur_policy;
67 unsigned int prev_cpu_idle_up;
68 unsigned int prev_cpu_idle_down;
69 unsigned int enable;
70 unsigned int down_skip;
71 unsigned int requested_freq;
73 static DEFINE_PER_CPU(struct cpu_dbs_info_s, cpu_dbs_info);
75 static unsigned int dbs_enable; /* number of CPUs using this policy */
78 * DEADLOCK ALERT! There is a ordering requirement between cpu_hotplug
79 * lock and dbs_mutex. cpu_hotplug lock should always be held before
80 * dbs_mutex. If any function that can potentially take cpu_hotplug lock
81 * (like __cpufreq_driver_target()) is being called with dbs_mutex taken, then
82 * cpu_hotplug lock should be taken before that. Note that cpu_hotplug lock
83 * is recursive for the same process. -Venki
85 static DEFINE_MUTEX (dbs_mutex);
86 static DECLARE_DELAYED_WORK(dbs_work, do_dbs_timer);
88 struct dbs_tuners {
89 unsigned int sampling_rate;
90 unsigned int sampling_down_factor;
91 unsigned int up_threshold;
92 unsigned int down_threshold;
93 unsigned int ignore_nice;
94 unsigned int freq_step;
97 static struct dbs_tuners dbs_tuners_ins = {
98 .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
99 .down_threshold = DEF_FREQUENCY_DOWN_THRESHOLD,
100 .sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR,
101 .ignore_nice = 0,
102 .freq_step = 5,
105 static inline unsigned int get_cpu_idle_time(unsigned int cpu)
107 unsigned int add_nice = 0, ret;
109 if (dbs_tuners_ins.ignore_nice)
110 add_nice = kstat_cpu(cpu).cpustat.nice;
112 ret = kstat_cpu(cpu).cpustat.idle +
113 kstat_cpu(cpu).cpustat.iowait +
114 add_nice;
116 return ret;
119 /* keep track of frequency transitions */
120 static int
121 dbs_cpufreq_notifier(struct notifier_block *nb, unsigned long val,
122 void *data)
124 struct cpufreq_freqs *freq = data;
125 struct cpu_dbs_info_s *this_dbs_info = &per_cpu(cpu_dbs_info,
126 freq->cpu);
128 if (!this_dbs_info->enable)
129 return 0;
131 this_dbs_info->requested_freq = freq->new;
133 return 0;
136 static struct notifier_block dbs_cpufreq_notifier_block = {
137 .notifier_call = dbs_cpufreq_notifier
140 /************************** sysfs interface ************************/
141 static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf)
143 return sprintf (buf, "%u\n", MAX_SAMPLING_RATE);
146 static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf)
148 return sprintf (buf, "%u\n", MIN_SAMPLING_RATE);
151 #define define_one_ro(_name) \
152 static struct freq_attr _name = \
153 __ATTR(_name, 0444, show_##_name, NULL)
155 define_one_ro(sampling_rate_max);
156 define_one_ro(sampling_rate_min);
158 /* cpufreq_conservative Governor Tunables */
159 #define show_one(file_name, object) \
160 static ssize_t show_##file_name \
161 (struct cpufreq_policy *unused, char *buf) \
163 return sprintf(buf, "%u\n", dbs_tuners_ins.object); \
165 show_one(sampling_rate, sampling_rate);
166 show_one(sampling_down_factor, sampling_down_factor);
167 show_one(up_threshold, up_threshold);
168 show_one(down_threshold, down_threshold);
169 show_one(ignore_nice_load, ignore_nice);
170 show_one(freq_step, freq_step);
172 static ssize_t store_sampling_down_factor(struct cpufreq_policy *unused,
173 const char *buf, size_t count)
175 unsigned int input;
176 int ret;
177 ret = sscanf (buf, "%u", &input);
178 if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
179 return -EINVAL;
181 mutex_lock(&dbs_mutex);
182 dbs_tuners_ins.sampling_down_factor = input;
183 mutex_unlock(&dbs_mutex);
185 return count;
188 static ssize_t store_sampling_rate(struct cpufreq_policy *unused,
189 const char *buf, size_t count)
191 unsigned int input;
192 int ret;
193 ret = sscanf (buf, "%u", &input);
195 mutex_lock(&dbs_mutex);
196 if (ret != 1 || input > MAX_SAMPLING_RATE || input < MIN_SAMPLING_RATE) {
197 mutex_unlock(&dbs_mutex);
198 return -EINVAL;
201 dbs_tuners_ins.sampling_rate = input;
202 mutex_unlock(&dbs_mutex);
204 return count;
207 static ssize_t store_up_threshold(struct cpufreq_policy *unused,
208 const char *buf, size_t count)
210 unsigned int input;
211 int ret;
212 ret = sscanf (buf, "%u", &input);
214 mutex_lock(&dbs_mutex);
215 if (ret != 1 || input > 100 || input <= dbs_tuners_ins.down_threshold) {
216 mutex_unlock(&dbs_mutex);
217 return -EINVAL;
220 dbs_tuners_ins.up_threshold = input;
221 mutex_unlock(&dbs_mutex);
223 return count;
226 static ssize_t store_down_threshold(struct cpufreq_policy *unused,
227 const char *buf, size_t count)
229 unsigned int input;
230 int ret;
231 ret = sscanf (buf, "%u", &input);
233 mutex_lock(&dbs_mutex);
234 if (ret != 1 || input > 100 || input >= dbs_tuners_ins.up_threshold) {
235 mutex_unlock(&dbs_mutex);
236 return -EINVAL;
239 dbs_tuners_ins.down_threshold = input;
240 mutex_unlock(&dbs_mutex);
242 return count;
245 static ssize_t store_ignore_nice_load(struct cpufreq_policy *policy,
246 const char *buf, size_t count)
248 unsigned int input;
249 int ret;
251 unsigned int j;
253 ret = sscanf(buf, "%u", &input);
254 if (ret != 1)
255 return -EINVAL;
257 if (input > 1)
258 input = 1;
260 mutex_lock(&dbs_mutex);
261 if (input == dbs_tuners_ins.ignore_nice) { /* nothing to do */
262 mutex_unlock(&dbs_mutex);
263 return count;
265 dbs_tuners_ins.ignore_nice = input;
267 /* we need to re-evaluate prev_cpu_idle_up and prev_cpu_idle_down */
268 for_each_online_cpu(j) {
269 struct cpu_dbs_info_s *j_dbs_info;
270 j_dbs_info = &per_cpu(cpu_dbs_info, j);
271 j_dbs_info->prev_cpu_idle_up = get_cpu_idle_time(j);
272 j_dbs_info->prev_cpu_idle_down = j_dbs_info->prev_cpu_idle_up;
274 mutex_unlock(&dbs_mutex);
276 return count;
279 static ssize_t store_freq_step(struct cpufreq_policy *policy,
280 const char *buf, size_t count)
282 unsigned int input;
283 int ret;
285 ret = sscanf(buf, "%u", &input);
287 if (ret != 1)
288 return -EINVAL;
290 if (input > 100)
291 input = 100;
293 /* no need to test here if freq_step is zero as the user might actually
294 * want this, they would be crazy though :) */
295 mutex_lock(&dbs_mutex);
296 dbs_tuners_ins.freq_step = input;
297 mutex_unlock(&dbs_mutex);
299 return count;
302 #define define_one_rw(_name) \
303 static struct freq_attr _name = \
304 __ATTR(_name, 0644, show_##_name, store_##_name)
306 define_one_rw(sampling_rate);
307 define_one_rw(sampling_down_factor);
308 define_one_rw(up_threshold);
309 define_one_rw(down_threshold);
310 define_one_rw(ignore_nice_load);
311 define_one_rw(freq_step);
313 static struct attribute * dbs_attributes[] = {
314 &sampling_rate_max.attr,
315 &sampling_rate_min.attr,
316 &sampling_rate.attr,
317 &sampling_down_factor.attr,
318 &up_threshold.attr,
319 &down_threshold.attr,
320 &ignore_nice_load.attr,
321 &freq_step.attr,
322 NULL
325 static struct attribute_group dbs_attr_group = {
326 .attrs = dbs_attributes,
327 .name = "conservative",
330 /************************** sysfs end ************************/
332 static void dbs_check_cpu(int cpu)
334 unsigned int idle_ticks, up_idle_ticks, down_idle_ticks;
335 unsigned int tmp_idle_ticks, total_idle_ticks;
336 unsigned int freq_step;
337 unsigned int freq_down_sampling_rate;
338 struct cpu_dbs_info_s *this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
339 struct cpufreq_policy *policy;
341 if (!this_dbs_info->enable)
342 return;
344 policy = this_dbs_info->cur_policy;
347 * The default safe range is 20% to 80%
348 * Every sampling_rate, we check
349 * - If current idle time is less than 20%, then we try to
350 * increase frequency
351 * Every sampling_rate*sampling_down_factor, we check
352 * - If current idle time is more than 80%, then we try to
353 * decrease frequency
355 * Any frequency increase takes it to the maximum frequency.
356 * Frequency reduction happens at minimum steps of
357 * 5% (default) of max_frequency
360 /* Check for frequency increase */
361 idle_ticks = UINT_MAX;
363 /* Check for frequency increase */
364 total_idle_ticks = get_cpu_idle_time(cpu);
365 tmp_idle_ticks = total_idle_ticks -
366 this_dbs_info->prev_cpu_idle_up;
367 this_dbs_info->prev_cpu_idle_up = total_idle_ticks;
369 if (tmp_idle_ticks < idle_ticks)
370 idle_ticks = tmp_idle_ticks;
372 /* Scale idle ticks by 100 and compare with up and down ticks */
373 idle_ticks *= 100;
374 up_idle_ticks = (100 - dbs_tuners_ins.up_threshold) *
375 usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
377 if (idle_ticks < up_idle_ticks) {
378 this_dbs_info->down_skip = 0;
379 this_dbs_info->prev_cpu_idle_down =
380 this_dbs_info->prev_cpu_idle_up;
382 /* if we are already at full speed then break out early */
383 if (this_dbs_info->requested_freq == policy->max)
384 return;
386 freq_step = (dbs_tuners_ins.freq_step * policy->max) / 100;
388 /* max freq cannot be less than 100. But who knows.... */
389 if (unlikely(freq_step == 0))
390 freq_step = 5;
392 this_dbs_info->requested_freq += freq_step;
393 if (this_dbs_info->requested_freq > policy->max)
394 this_dbs_info->requested_freq = policy->max;
396 __cpufreq_driver_target(policy, this_dbs_info->requested_freq,
397 CPUFREQ_RELATION_H);
398 return;
401 /* Check for frequency decrease */
402 this_dbs_info->down_skip++;
403 if (this_dbs_info->down_skip < dbs_tuners_ins.sampling_down_factor)
404 return;
406 /* Check for frequency decrease */
407 total_idle_ticks = this_dbs_info->prev_cpu_idle_up;
408 tmp_idle_ticks = total_idle_ticks -
409 this_dbs_info->prev_cpu_idle_down;
410 this_dbs_info->prev_cpu_idle_down = total_idle_ticks;
412 if (tmp_idle_ticks < idle_ticks)
413 idle_ticks = tmp_idle_ticks;
415 /* Scale idle ticks by 100 and compare with up and down ticks */
416 idle_ticks *= 100;
417 this_dbs_info->down_skip = 0;
419 freq_down_sampling_rate = dbs_tuners_ins.sampling_rate *
420 dbs_tuners_ins.sampling_down_factor;
421 down_idle_ticks = (100 - dbs_tuners_ins.down_threshold) *
422 usecs_to_jiffies(freq_down_sampling_rate);
424 if (idle_ticks > down_idle_ticks) {
426 * if we are already at the lowest speed then break out early
427 * or if we 'cannot' reduce the speed as the user might want
428 * freq_step to be zero
430 if (this_dbs_info->requested_freq == policy->min
431 || dbs_tuners_ins.freq_step == 0)
432 return;
434 freq_step = (dbs_tuners_ins.freq_step * policy->max) / 100;
436 /* max freq cannot be less than 100. But who knows.... */
437 if (unlikely(freq_step == 0))
438 freq_step = 5;
440 this_dbs_info->requested_freq -= freq_step;
441 if (this_dbs_info->requested_freq < policy->min)
442 this_dbs_info->requested_freq = policy->min;
444 __cpufreq_driver_target(policy, this_dbs_info->requested_freq,
445 CPUFREQ_RELATION_H);
446 return;
450 static void do_dbs_timer(struct work_struct *work)
452 int i;
453 mutex_lock(&dbs_mutex);
454 for_each_online_cpu(i)
455 dbs_check_cpu(i);
456 schedule_delayed_work(&dbs_work,
457 usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
458 mutex_unlock(&dbs_mutex);
461 static inline void dbs_timer_init(void)
463 schedule_delayed_work(&dbs_work,
464 usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
465 return;
468 static inline void dbs_timer_exit(void)
470 cancel_delayed_work(&dbs_work);
471 return;
474 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
475 unsigned int event)
477 unsigned int cpu = policy->cpu;
478 struct cpu_dbs_info_s *this_dbs_info;
479 unsigned int j;
480 int rc;
482 this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
484 switch (event) {
485 case CPUFREQ_GOV_START:
486 if ((!cpu_online(cpu)) || (!policy->cur))
487 return -EINVAL;
489 if (this_dbs_info->enable) /* Already enabled */
490 break;
492 mutex_lock(&dbs_mutex);
494 rc = sysfs_create_group(&policy->kobj, &dbs_attr_group);
495 if (rc) {
496 mutex_unlock(&dbs_mutex);
497 return rc;
500 for_each_cpu_mask(j, policy->cpus) {
501 struct cpu_dbs_info_s *j_dbs_info;
502 j_dbs_info = &per_cpu(cpu_dbs_info, j);
503 j_dbs_info->cur_policy = policy;
505 j_dbs_info->prev_cpu_idle_up = get_cpu_idle_time(cpu);
506 j_dbs_info->prev_cpu_idle_down
507 = j_dbs_info->prev_cpu_idle_up;
509 this_dbs_info->enable = 1;
510 this_dbs_info->down_skip = 0;
511 this_dbs_info->requested_freq = policy->cur;
513 dbs_enable++;
515 * Start the timerschedule work, when this governor
516 * is used for first time
518 if (dbs_enable == 1) {
519 unsigned int latency;
520 /* policy latency is in nS. Convert it to uS first */
521 latency = policy->cpuinfo.transition_latency / 1000;
522 if (latency == 0)
523 latency = 1;
525 def_sampling_rate = 10 * latency *
526 DEF_SAMPLING_RATE_LATENCY_MULTIPLIER;
528 if (def_sampling_rate < MIN_STAT_SAMPLING_RATE)
529 def_sampling_rate = MIN_STAT_SAMPLING_RATE;
531 dbs_tuners_ins.sampling_rate = def_sampling_rate;
533 dbs_timer_init();
534 cpufreq_register_notifier(
535 &dbs_cpufreq_notifier_block,
536 CPUFREQ_TRANSITION_NOTIFIER);
539 mutex_unlock(&dbs_mutex);
540 break;
542 case CPUFREQ_GOV_STOP:
543 mutex_lock(&dbs_mutex);
544 this_dbs_info->enable = 0;
545 sysfs_remove_group(&policy->kobj, &dbs_attr_group);
546 dbs_enable--;
548 * Stop the timerschedule work, when this governor
549 * is used for first time
551 if (dbs_enable == 0) {
552 dbs_timer_exit();
553 cpufreq_unregister_notifier(
554 &dbs_cpufreq_notifier_block,
555 CPUFREQ_TRANSITION_NOTIFIER);
558 mutex_unlock(&dbs_mutex);
560 break;
562 case CPUFREQ_GOV_LIMITS:
563 mutex_lock(&dbs_mutex);
564 if (policy->max < this_dbs_info->cur_policy->cur)
565 __cpufreq_driver_target(
566 this_dbs_info->cur_policy,
567 policy->max, CPUFREQ_RELATION_H);
568 else if (policy->min > this_dbs_info->cur_policy->cur)
569 __cpufreq_driver_target(
570 this_dbs_info->cur_policy,
571 policy->min, CPUFREQ_RELATION_L);
572 mutex_unlock(&dbs_mutex);
573 break;
575 return 0;
578 struct cpufreq_governor cpufreq_gov_conservative = {
579 .name = "conservative",
580 .governor = cpufreq_governor_dbs,
581 .max_transition_latency = TRANSITION_LATENCY_LIMIT,
582 .owner = THIS_MODULE,
584 EXPORT_SYMBOL(cpufreq_gov_conservative);
586 static int __init cpufreq_gov_dbs_init(void)
588 return cpufreq_register_governor(&cpufreq_gov_conservative);
591 static void __exit cpufreq_gov_dbs_exit(void)
593 /* Make sure that the scheduled work is indeed not running */
594 flush_scheduled_work();
596 cpufreq_unregister_governor(&cpufreq_gov_conservative);
600 MODULE_AUTHOR ("Alexander Clouter <alex-kernel@digriz.org.uk>");
601 MODULE_DESCRIPTION ("'cpufreq_conservative' - A dynamic cpufreq governor for "
602 "Low Latency Frequency Transition capable processors "
603 "optimised for use in a battery environment");
604 MODULE_LICENSE ("GPL");
606 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_CONSERVATIVE
607 fs_initcall(cpufreq_gov_dbs_init);
608 #else
609 module_init(cpufreq_gov_dbs_init);
610 #endif
611 module_exit(cpufreq_gov_dbs_exit);