blob: 8d83a21c6477f3e7a3cd223a9bffa4e8b1af1710 [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>
15#include <linux/smp.h>
16#include <linux/init.h>
17#include <linux/interrupt.h>
18#include <linux/ctype.h>
19#include <linux/cpufreq.h>
20#include <linux/sysctl.h>
21#include <linux/types.h>
22#include <linux/fs.h>
23#include <linux/sysfs.h>
24#include <linux/sched.h>
25#include <linux/kmod.h>
26#include <linux/workqueue.h>
27#include <linux/jiffies.h>
28#include <linux/kernel_stat.h>
29#include <linux/percpu.h>
30
31/*
32 * dbs is used in this file as a shortform for demandbased switching
33 * It helps to keep variable names smaller, simpler
34 */
35
36#define DEF_FREQUENCY_UP_THRESHOLD (80)
37#define MIN_FREQUENCY_UP_THRESHOLD (0)
38#define MAX_FREQUENCY_UP_THRESHOLD (100)
39
40#define DEF_FREQUENCY_DOWN_THRESHOLD (20)
41#define MIN_FREQUENCY_DOWN_THRESHOLD (0)
42#define MAX_FREQUENCY_DOWN_THRESHOLD (100)
43
44/*
45 * The polling frequency of this governor depends on the capability of
46 * the processor. Default polling frequency is 1000 times the transition
47 * latency of the processor. The governor will work on any processor with
48 * transition latency <= 10mS, using appropriate sampling
49 * rate.
50 * For CPUs with transition latency > 10mS (mostly drivers with CPUFREQ_ETERNAL)
51 * this governor will not work.
52 * All times here are in uS.
53 */
54static unsigned int def_sampling_rate;
55#define MIN_SAMPLING_RATE (def_sampling_rate / 2)
56#define MAX_SAMPLING_RATE (500 * def_sampling_rate)
57#define DEF_SAMPLING_RATE_LATENCY_MULTIPLIER (1000)
58#define DEF_SAMPLING_DOWN_FACTOR (10)
59#define TRANSITION_LATENCY_LIMIT (10 * 1000)
60#define sampling_rate_in_HZ(x) (((x * HZ) < (1000 * 1000))?1:((x * HZ) / (1000 * 1000)))
61
62static void do_dbs_timer(void *data);
63
64struct cpu_dbs_info_s {
65 struct cpufreq_policy *cur_policy;
66 unsigned int prev_cpu_idle_up;
67 unsigned int prev_cpu_idle_down;
68 unsigned int enable;
69};
70static DEFINE_PER_CPU(struct cpu_dbs_info_s, cpu_dbs_info);
71
72static unsigned int dbs_enable; /* number of CPUs using this policy */
73
74static DECLARE_MUTEX (dbs_sem);
75static DECLARE_WORK (dbs_work, do_dbs_timer, NULL);
76
77struct dbs_tuners {
78 unsigned int sampling_rate;
79 unsigned int sampling_down_factor;
80 unsigned int up_threshold;
81 unsigned int down_threshold;
82};
83
84static struct dbs_tuners dbs_tuners_ins = {
85 .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
86 .down_threshold = DEF_FREQUENCY_DOWN_THRESHOLD,
87 .sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR,
88};
89
90/************************** sysfs interface ************************/
91static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf)
92{
93 return sprintf (buf, "%u\n", MAX_SAMPLING_RATE);
94}
95
96static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf)
97{
98 return sprintf (buf, "%u\n", MIN_SAMPLING_RATE);
99}
100
101#define define_one_ro(_name) \
102static struct freq_attr _name = \
103__ATTR(_name, 0444, show_##_name, NULL)
104
105define_one_ro(sampling_rate_max);
106define_one_ro(sampling_rate_min);
107
108/* cpufreq_ondemand Governor Tunables */
109#define show_one(file_name, object) \
110static ssize_t show_##file_name \
111(struct cpufreq_policy *unused, char *buf) \
112{ \
113 return sprintf(buf, "%u\n", dbs_tuners_ins.object); \
114}
115show_one(sampling_rate, sampling_rate);
116show_one(sampling_down_factor, sampling_down_factor);
117show_one(up_threshold, up_threshold);
118show_one(down_threshold, down_threshold);
119
120static ssize_t store_sampling_down_factor(struct cpufreq_policy *unused,
121 const char *buf, size_t count)
122{
123 unsigned int input;
124 int ret;
125 ret = sscanf (buf, "%u", &input);
126 if (ret != 1 )
127 return -EINVAL;
128
129 down(&dbs_sem);
130 dbs_tuners_ins.sampling_down_factor = input;
131 up(&dbs_sem);
132
133 return count;
134}
135
136static ssize_t store_sampling_rate(struct cpufreq_policy *unused,
137 const char *buf, size_t count)
138{
139 unsigned int input;
140 int ret;
141 ret = sscanf (buf, "%u", &input);
142
143 down(&dbs_sem);
144 if (ret != 1 || input > MAX_SAMPLING_RATE || input < MIN_SAMPLING_RATE) {
145 up(&dbs_sem);
146 return -EINVAL;
147 }
148
149 dbs_tuners_ins.sampling_rate = input;
150 up(&dbs_sem);
151
152 return count;
153}
154
155static ssize_t store_up_threshold(struct cpufreq_policy *unused,
156 const char *buf, size_t count)
157{
158 unsigned int input;
159 int ret;
160 ret = sscanf (buf, "%u", &input);
161
162 down(&dbs_sem);
163 if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
164 input < MIN_FREQUENCY_UP_THRESHOLD ||
165 input <= dbs_tuners_ins.down_threshold) {
166 up(&dbs_sem);
167 return -EINVAL;
168 }
169
170 dbs_tuners_ins.up_threshold = input;
171 up(&dbs_sem);
172
173 return count;
174}
175
176static ssize_t store_down_threshold(struct cpufreq_policy *unused,
177 const char *buf, size_t count)
178{
179 unsigned int input;
180 int ret;
181 ret = sscanf (buf, "%u", &input);
182
183 down(&dbs_sem);
184 if (ret != 1 || input > MAX_FREQUENCY_DOWN_THRESHOLD ||
185 input < MIN_FREQUENCY_DOWN_THRESHOLD ||
186 input >= dbs_tuners_ins.up_threshold) {
187 up(&dbs_sem);
188 return -EINVAL;
189 }
190
191 dbs_tuners_ins.down_threshold = input;
192 up(&dbs_sem);
193
194 return count;
195}
196
197#define define_one_rw(_name) \
198static struct freq_attr _name = \
199__ATTR(_name, 0644, show_##_name, store_##_name)
200
201define_one_rw(sampling_rate);
202define_one_rw(sampling_down_factor);
203define_one_rw(up_threshold);
204define_one_rw(down_threshold);
205
206static struct attribute * dbs_attributes[] = {
207 &sampling_rate_max.attr,
208 &sampling_rate_min.attr,
209 &sampling_rate.attr,
210 &sampling_down_factor.attr,
211 &up_threshold.attr,
212 &down_threshold.attr,
213 NULL
214};
215
216static struct attribute_group dbs_attr_group = {
217 .attrs = dbs_attributes,
218 .name = "ondemand",
219};
220
221/************************** sysfs end ************************/
222
223static void dbs_check_cpu(int cpu)
224{
225 unsigned int idle_ticks, up_idle_ticks, down_idle_ticks;
226 unsigned int total_idle_ticks;
227 unsigned int freq_down_step;
228 unsigned int freq_down_sampling_rate;
229 static int down_skip[NR_CPUS];
230 struct cpu_dbs_info_s *this_dbs_info;
231
232 struct cpufreq_policy *policy;
233 unsigned int j;
234
235 this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
236 if (!this_dbs_info->enable)
237 return;
238
239 policy = this_dbs_info->cur_policy;
240 /*
241 * The default safe range is 20% to 80%
242 * Every sampling_rate, we check
243 * - If current idle time is less than 20%, then we try to
244 * increase frequency
245 * Every sampling_rate*sampling_down_factor, we check
246 * - If current idle time is more than 80%, then we try to
247 * decrease frequency
248 *
249 * Any frequency increase takes it to the maximum frequency.
250 * Frequency reduction happens at minimum steps of
251 * 5% of max_frequency
252 */
253
254 /* Check for frequency increase */
255 total_idle_ticks = kstat_cpu(cpu).cpustat.idle +
256 kstat_cpu(cpu).cpustat.iowait;
257 idle_ticks = total_idle_ticks -
258 this_dbs_info->prev_cpu_idle_up;
259 this_dbs_info->prev_cpu_idle_up = total_idle_ticks;
260
261
262 for_each_cpu_mask(j, policy->cpus) {
263 unsigned int tmp_idle_ticks;
264 struct cpu_dbs_info_s *j_dbs_info;
265
266 if (j == cpu)
267 continue;
268
269 j_dbs_info = &per_cpu(cpu_dbs_info, j);
270 /* Check for frequency increase */
271 total_idle_ticks = kstat_cpu(j).cpustat.idle +
272 kstat_cpu(j).cpustat.iowait;
273 tmp_idle_ticks = total_idle_ticks -
274 j_dbs_info->prev_cpu_idle_up;
275 j_dbs_info->prev_cpu_idle_up = total_idle_ticks;
276
277 if (tmp_idle_ticks < idle_ticks)
278 idle_ticks = tmp_idle_ticks;
279 }
280
281 /* Scale idle ticks by 100 and compare with up and down ticks */
282 idle_ticks *= 100;
283 up_idle_ticks = (100 - dbs_tuners_ins.up_threshold) *
284 sampling_rate_in_HZ(dbs_tuners_ins.sampling_rate);
285
286 if (idle_ticks < up_idle_ticks) {
287 __cpufreq_driver_target(policy, policy->max,
288 CPUFREQ_RELATION_H);
289 down_skip[cpu] = 0;
290 this_dbs_info->prev_cpu_idle_down = total_idle_ticks;
291 return;
292 }
293
294 /* Check for frequency decrease */
295 down_skip[cpu]++;
296 if (down_skip[cpu] < dbs_tuners_ins.sampling_down_factor)
297 return;
298
299 total_idle_ticks = kstat_cpu(cpu).cpustat.idle +
300 kstat_cpu(cpu).cpustat.iowait;
301 idle_ticks = total_idle_ticks -
302 this_dbs_info->prev_cpu_idle_down;
303 this_dbs_info->prev_cpu_idle_down = total_idle_ticks;
304
305 for_each_cpu_mask(j, policy->cpus) {
306 unsigned int tmp_idle_ticks;
307 struct cpu_dbs_info_s *j_dbs_info;
308
309 if (j == cpu)
310 continue;
311
312 j_dbs_info = &per_cpu(cpu_dbs_info, j);
313 /* Check for frequency increase */
314 total_idle_ticks = kstat_cpu(j).cpustat.idle +
315 kstat_cpu(j).cpustat.iowait;
316 tmp_idle_ticks = total_idle_ticks -
317 j_dbs_info->prev_cpu_idle_down;
318 j_dbs_info->prev_cpu_idle_down = total_idle_ticks;
319
320 if (tmp_idle_ticks < idle_ticks)
321 idle_ticks = tmp_idle_ticks;
322 }
323
324 /* Scale idle ticks by 100 and compare with up and down ticks */
325 idle_ticks *= 100;
326 down_skip[cpu] = 0;
327
328 freq_down_sampling_rate = dbs_tuners_ins.sampling_rate *
329 dbs_tuners_ins.sampling_down_factor;
330 down_idle_ticks = (100 - dbs_tuners_ins.down_threshold) *
331 sampling_rate_in_HZ(freq_down_sampling_rate);
332
333 if (idle_ticks > down_idle_ticks ) {
334 freq_down_step = (5 * policy->max) / 100;
335
336 /* max freq cannot be less than 100. But who knows.... */
337 if (unlikely(freq_down_step == 0))
338 freq_down_step = 5;
339
340 __cpufreq_driver_target(policy,
341 policy->cur - freq_down_step,
342 CPUFREQ_RELATION_H);
343 return;
344 }
345}
346
347static void do_dbs_timer(void *data)
348{
349 int i;
350 down(&dbs_sem);
351 for (i = 0; i < NR_CPUS; i++)
352 if (cpu_online(i))
353 dbs_check_cpu(i);
354 schedule_delayed_work(&dbs_work,
355 sampling_rate_in_HZ(dbs_tuners_ins.sampling_rate));
356 up(&dbs_sem);
357}
358
359static inline void dbs_timer_init(void)
360{
361 INIT_WORK(&dbs_work, do_dbs_timer, NULL);
362 schedule_delayed_work(&dbs_work,
363 sampling_rate_in_HZ(dbs_tuners_ins.sampling_rate));
364 return;
365}
366
367static inline void dbs_timer_exit(void)
368{
369 cancel_delayed_work(&dbs_work);
370 return;
371}
372
373static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
374 unsigned int event)
375{
376 unsigned int cpu = policy->cpu;
377 struct cpu_dbs_info_s *this_dbs_info;
378 unsigned int j;
379
380 this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
381
382 switch (event) {
383 case CPUFREQ_GOV_START:
384 if ((!cpu_online(cpu)) ||
385 (!policy->cur))
386 return -EINVAL;
387
388 if (policy->cpuinfo.transition_latency >
389 (TRANSITION_LATENCY_LIMIT * 1000))
390 return -EINVAL;
391 if (this_dbs_info->enable) /* Already enabled */
392 break;
393
394 down(&dbs_sem);
395 for_each_cpu_mask(j, policy->cpus) {
396 struct cpu_dbs_info_s *j_dbs_info;
397 j_dbs_info = &per_cpu(cpu_dbs_info, j);
398 j_dbs_info->cur_policy = policy;
399
400 j_dbs_info->prev_cpu_idle_up =
401 kstat_cpu(j).cpustat.idle +
402 kstat_cpu(j).cpustat.iowait;
403 j_dbs_info->prev_cpu_idle_down =
404 kstat_cpu(j).cpustat.idle +
405 kstat_cpu(j).cpustat.iowait;
406 }
407 this_dbs_info->enable = 1;
408 sysfs_create_group(&policy->kobj, &dbs_attr_group);
409 dbs_enable++;
410 /*
411 * Start the timerschedule work, when this governor
412 * is used for first time
413 */
414 if (dbs_enable == 1) {
415 unsigned int latency;
416 /* policy latency is in nS. Convert it to uS first */
417
418 latency = policy->cpuinfo.transition_latency;
419 if (latency < 1000)
420 latency = 1000;
421
422 def_sampling_rate = (latency / 1000) *
423 DEF_SAMPLING_RATE_LATENCY_MULTIPLIER;
424 dbs_tuners_ins.sampling_rate = def_sampling_rate;
425
426 dbs_timer_init();
427 }
428
429 up(&dbs_sem);
430 break;
431
432 case CPUFREQ_GOV_STOP:
433 down(&dbs_sem);
434 this_dbs_info->enable = 0;
435 sysfs_remove_group(&policy->kobj, &dbs_attr_group);
436 dbs_enable--;
437 /*
438 * Stop the timerschedule work, when this governor
439 * is used for first time
440 */
441 if (dbs_enable == 0)
442 dbs_timer_exit();
443
444 up(&dbs_sem);
445
446 break;
447
448 case CPUFREQ_GOV_LIMITS:
449 down(&dbs_sem);
450 if (policy->max < this_dbs_info->cur_policy->cur)
451 __cpufreq_driver_target(
452 this_dbs_info->cur_policy,
453 policy->max, CPUFREQ_RELATION_H);
454 else if (policy->min > this_dbs_info->cur_policy->cur)
455 __cpufreq_driver_target(
456 this_dbs_info->cur_policy,
457 policy->min, CPUFREQ_RELATION_L);
458 up(&dbs_sem);
459 break;
460 }
461 return 0;
462}
463
464struct cpufreq_governor cpufreq_gov_dbs = {
465 .name = "ondemand",
466 .governor = cpufreq_governor_dbs,
467 .owner = THIS_MODULE,
468};
469EXPORT_SYMBOL(cpufreq_gov_dbs);
470
471static int __init cpufreq_gov_dbs_init(void)
472{
473 return cpufreq_register_governor(&cpufreq_gov_dbs);
474}
475
476static void __exit cpufreq_gov_dbs_exit(void)
477{
478 /* Make sure that the scheduled work is indeed not running */
479 flush_scheduled_work();
480
481 cpufreq_unregister_governor(&cpufreq_gov_dbs);
482}
483
484
485MODULE_AUTHOR ("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
486MODULE_DESCRIPTION ("'cpufreq_ondemand' - A dynamic cpufreq governor for "
487 "Low Latency Frequency Transition capable processors");
488MODULE_LICENSE ("GPL");
489
490module_init(cpufreq_gov_dbs_init);
491module_exit(cpufreq_gov_dbs_exit);