[CPUFREQ] ondemand,conservative governor store the idle ticks for all cpus
[linux-2.6] / drivers / cpufreq / cpufreq_ondemand.c
1 /*
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  */
54 static 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
61 static void do_dbs_timer(void *data);
62
63 struct cpu_dbs_info_s {
64         struct cpufreq_policy   *cur_policy;
65         unsigned int            prev_cpu_idle_up;
66         unsigned int            prev_cpu_idle_down;
67         unsigned int            enable;
68 };
69 static DEFINE_PER_CPU(struct cpu_dbs_info_s, cpu_dbs_info);
70
71 static unsigned int dbs_enable; /* number of CPUs using this policy */
72
73 static DECLARE_MUTEX    (dbs_sem);
74 static DECLARE_WORK     (dbs_work, do_dbs_timer, NULL);
75
76 struct dbs_tuners {
77         unsigned int            sampling_rate;
78         unsigned int            sampling_down_factor;
79         unsigned int            up_threshold;
80         unsigned int            down_threshold;
81         unsigned int            ignore_nice;
82         unsigned int            freq_step;
83 };
84
85 static struct dbs_tuners dbs_tuners_ins = {
86         .up_threshold           = DEF_FREQUENCY_UP_THRESHOLD,
87         .down_threshold         = DEF_FREQUENCY_DOWN_THRESHOLD,
88         .sampling_down_factor   = DEF_SAMPLING_DOWN_FACTOR,
89 };
90
91 static inline unsigned int get_cpu_idle_time(unsigned int cpu)
92 {
93         return  kstat_cpu(cpu).cpustat.idle +
94                 kstat_cpu(cpu).cpustat.iowait +
95                 ( !dbs_tuners_ins.ignore_nice ? 
96                   kstat_cpu(cpu).cpustat.nice :
97                   0);
98 }
99
100 /************************** sysfs interface ************************/
101 static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf)
102 {
103         return sprintf (buf, "%u\n", MAX_SAMPLING_RATE);
104 }
105
106 static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf)
107 {
108         return sprintf (buf, "%u\n", MIN_SAMPLING_RATE);
109 }
110
111 #define define_one_ro(_name)                                    \
112 static struct freq_attr _name =                                 \
113 __ATTR(_name, 0444, show_##_name, NULL)
114
115 define_one_ro(sampling_rate_max);
116 define_one_ro(sampling_rate_min);
117
118 /* cpufreq_ondemand Governor Tunables */
119 #define show_one(file_name, object)                                     \
120 static ssize_t show_##file_name                                         \
121 (struct cpufreq_policy *unused, char *buf)                              \
122 {                                                                       \
123         return sprintf(buf, "%u\n", dbs_tuners_ins.object);             \
124 }
125 show_one(sampling_rate, sampling_rate);
126 show_one(sampling_down_factor, sampling_down_factor);
127 show_one(up_threshold, up_threshold);
128 show_one(down_threshold, down_threshold);
129 show_one(ignore_nice, ignore_nice);
130 show_one(freq_step, freq_step);
131
132 static ssize_t store_sampling_down_factor(struct cpufreq_policy *unused, 
133                 const char *buf, size_t count)
134 {
135         unsigned int input;
136         int ret;
137         ret = sscanf (buf, "%u", &input);
138         if (ret != 1 )
139                 return -EINVAL;
140
141         down(&dbs_sem);
142         dbs_tuners_ins.sampling_down_factor = input;
143         up(&dbs_sem);
144
145         return count;
146 }
147
148 static ssize_t store_sampling_rate(struct cpufreq_policy *unused, 
149                 const char *buf, size_t count)
150 {
151         unsigned int input;
152         int ret;
153         ret = sscanf (buf, "%u", &input);
154
155         down(&dbs_sem);
156         if (ret != 1 || input > MAX_SAMPLING_RATE || input < MIN_SAMPLING_RATE) {
157                 up(&dbs_sem);
158                 return -EINVAL;
159         }
160
161         dbs_tuners_ins.sampling_rate = input;
162         up(&dbs_sem);
163
164         return count;
165 }
166
167 static ssize_t store_up_threshold(struct cpufreq_policy *unused, 
168                 const char *buf, size_t count)
169 {
170         unsigned int input;
171         int ret;
172         ret = sscanf (buf, "%u", &input);
173
174         down(&dbs_sem);
175         if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD || 
176                         input < MIN_FREQUENCY_UP_THRESHOLD ||
177                         input <= dbs_tuners_ins.down_threshold) {
178                 up(&dbs_sem);
179                 return -EINVAL;
180         }
181
182         dbs_tuners_ins.up_threshold = input;
183         up(&dbs_sem);
184
185         return count;
186 }
187
188 static ssize_t store_down_threshold(struct cpufreq_policy *unused, 
189                 const char *buf, size_t count)
190 {
191         unsigned int input;
192         int ret;
193         ret = sscanf (buf, "%u", &input);
194
195         down(&dbs_sem);
196         if (ret != 1 || input > MAX_FREQUENCY_DOWN_THRESHOLD || 
197                         input < MIN_FREQUENCY_DOWN_THRESHOLD ||
198                         input >= dbs_tuners_ins.up_threshold) {
199                 up(&dbs_sem);
200                 return -EINVAL;
201         }
202
203         dbs_tuners_ins.down_threshold = input;
204         up(&dbs_sem);
205
206         return count;
207 }
208
209 static ssize_t store_ignore_nice(struct cpufreq_policy *policy,
210                 const char *buf, size_t count)
211 {
212         unsigned int input;
213         int ret;
214
215         unsigned int j;
216         
217         ret = sscanf (buf, "%u", &input);
218         if ( ret != 1 )
219                 return -EINVAL;
220
221         if ( input > 1 )
222                 input = 1;
223         
224         down(&dbs_sem);
225         if ( input == dbs_tuners_ins.ignore_nice ) { /* nothing to do */
226                 up(&dbs_sem);
227                 return count;
228         }
229         dbs_tuners_ins.ignore_nice = input;
230
231         /* we need to re-evaluate prev_cpu_idle_up and prev_cpu_idle_down */
232         for_each_online_cpu(j) {
233                 struct cpu_dbs_info_s *j_dbs_info;
234                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
235                 j_dbs_info->prev_cpu_idle_up = get_cpu_idle_time(j);
236                 j_dbs_info->prev_cpu_idle_down = j_dbs_info->prev_cpu_idle_up;
237         }
238         up(&dbs_sem);
239
240         return count;
241 }
242
243 static ssize_t store_freq_step(struct cpufreq_policy *policy,
244                 const char *buf, size_t count)
245 {
246         unsigned int input;
247         int ret;
248
249         ret = sscanf (buf, "%u", &input);
250
251         if ( ret != 1 )
252                 return -EINVAL;
253
254         if ( input > 100 )
255                 input = 100;
256         
257         /* no need to test here if freq_step is zero as the user might actually
258          * want this, they would be crazy though :) */
259         down(&dbs_sem);
260         dbs_tuners_ins.freq_step = input;
261         up(&dbs_sem);
262
263         return count;
264 }
265
266 #define define_one_rw(_name) \
267 static struct freq_attr _name = \
268 __ATTR(_name, 0644, show_##_name, store_##_name)
269
270 define_one_rw(sampling_rate);
271 define_one_rw(sampling_down_factor);
272 define_one_rw(up_threshold);
273 define_one_rw(down_threshold);
274 define_one_rw(ignore_nice);
275 define_one_rw(freq_step);
276
277 static struct attribute * dbs_attributes[] = {
278         &sampling_rate_max.attr,
279         &sampling_rate_min.attr,
280         &sampling_rate.attr,
281         &sampling_down_factor.attr,
282         &up_threshold.attr,
283         &down_threshold.attr,
284         &ignore_nice.attr,
285         &freq_step.attr,
286         NULL
287 };
288
289 static struct attribute_group dbs_attr_group = {
290         .attrs = dbs_attributes,
291         .name = "ondemand",
292 };
293
294 /************************** sysfs end ************************/
295
296 static void dbs_check_cpu(int cpu)
297 {
298         unsigned int idle_ticks, up_idle_ticks, down_idle_ticks;
299         unsigned int total_idle_ticks;
300         unsigned int freq_down_step;
301         unsigned int freq_down_sampling_rate;
302         static int down_skip[NR_CPUS];
303         struct cpu_dbs_info_s *this_dbs_info;
304
305         struct cpufreq_policy *policy;
306         unsigned int j;
307
308         this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
309         if (!this_dbs_info->enable)
310                 return;
311
312         policy = this_dbs_info->cur_policy;
313         /* 
314          * The default safe range is 20% to 80% 
315          * Every sampling_rate, we check
316          *      - If current idle time is less than 20%, then we try to 
317          *        increase frequency
318          * Every sampling_rate*sampling_down_factor, we check
319          *      - If current idle time is more than 80%, then we try to
320          *        decrease frequency
321          *
322          * Any frequency increase takes it to the maximum frequency. 
323          * Frequency reduction happens at minimum steps of 
324          * 5% (default) of max_frequency 
325          */
326
327         /* Check for frequency increase */
328         total_idle_ticks = get_cpu_idle_time(cpu);
329         idle_ticks = total_idle_ticks -
330                 this_dbs_info->prev_cpu_idle_up;
331         this_dbs_info->prev_cpu_idle_up = total_idle_ticks;
332
333         for_each_cpu_mask(j, policy->cpus) {
334                 unsigned int tmp_idle_ticks;
335                 struct cpu_dbs_info_s *j_dbs_info;
336
337                 if (j == cpu)
338                         continue;
339
340                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
341                 /* Check for frequency increase */
342                 total_idle_ticks = get_cpu_idle_time(j);
343                 tmp_idle_ticks = total_idle_ticks -
344                         j_dbs_info->prev_cpu_idle_up;
345                 j_dbs_info->prev_cpu_idle_up = total_idle_ticks;
346
347                 if (tmp_idle_ticks < idle_ticks)
348                         idle_ticks = tmp_idle_ticks;
349         }
350
351         /* Scale idle ticks by 100 and compare with up and down ticks */
352         idle_ticks *= 100;
353         up_idle_ticks = (100 - dbs_tuners_ins.up_threshold) *
354                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
355
356         if (idle_ticks < up_idle_ticks) {
357                 down_skip[cpu] = 0;
358                 for_each_cpu_mask(j, policy->cpus) {
359                         struct cpu_dbs_info_s *j_dbs_info;
360
361                         j_dbs_info = &per_cpu(cpu_dbs_info, j);
362                         j_dbs_info->prev_cpu_idle_down = 
363                                         j_dbs_info->prev_cpu_idle_up;
364                 }
365                 /* if we are already at full speed then break out early */
366                 if (policy->cur == policy->max)
367                         return;
368                 
369                 __cpufreq_driver_target(policy, policy->max, 
370                         CPUFREQ_RELATION_H);
371                 return;
372         }
373
374         /* Check for frequency decrease */
375         down_skip[cpu]++;
376         if (down_skip[cpu] < dbs_tuners_ins.sampling_down_factor)
377                 return;
378
379         total_idle_ticks = this_dbs_info->prev_cpu_idle_up;
380         idle_ticks = total_idle_ticks -
381                 this_dbs_info->prev_cpu_idle_down;
382         this_dbs_info->prev_cpu_idle_down = total_idle_ticks;
383
384         for_each_cpu_mask(j, policy->cpus) {
385                 unsigned int tmp_idle_ticks;
386                 struct cpu_dbs_info_s *j_dbs_info;
387
388                 if (j == cpu)
389                         continue;
390
391                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
392                 /* Check for frequency decrease */
393                 total_idle_ticks = j_dbs_info->prev_cpu_idle_up;
394                 tmp_idle_ticks = total_idle_ticks -
395                         j_dbs_info->prev_cpu_idle_down;
396                 j_dbs_info->prev_cpu_idle_down = total_idle_ticks;
397
398                 if (tmp_idle_ticks < idle_ticks)
399                         idle_ticks = tmp_idle_ticks;
400         }
401
402         /* Scale idle ticks by 100 and compare with up and down ticks */
403         idle_ticks *= 100;
404         down_skip[cpu] = 0;
405
406         freq_down_sampling_rate = dbs_tuners_ins.sampling_rate *
407                 dbs_tuners_ins.sampling_down_factor;
408         down_idle_ticks = (100 - dbs_tuners_ins.down_threshold) *
409                 usecs_to_jiffies(freq_down_sampling_rate);
410
411         if (idle_ticks > down_idle_ticks ) {
412                 /* if we are already at the lowest speed then break out early
413                  * or if we 'cannot' reduce the speed as the user might want
414                  * freq_step to be zero */
415                 if (policy->cur == policy->min || dbs_tuners_ins.freq_step == 0)
416                         return;
417
418                 freq_down_step = (dbs_tuners_ins.freq_step * policy->max) / 100;
419
420                 /* max freq cannot be less than 100. But who knows.... */
421                 if (unlikely(freq_down_step == 0))
422                         freq_down_step = 5;
423
424                 __cpufreq_driver_target(policy,
425                         policy->cur - freq_down_step,
426                         CPUFREQ_RELATION_H);
427                 return;
428         }
429 }
430
431 static void do_dbs_timer(void *data)
432
433         int i;
434         down(&dbs_sem);
435         for_each_online_cpu(i)
436                 dbs_check_cpu(i);
437         schedule_delayed_work(&dbs_work, 
438                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
439         up(&dbs_sem);
440
441
442 static inline void dbs_timer_init(void)
443 {
444         INIT_WORK(&dbs_work, do_dbs_timer, NULL);
445         schedule_delayed_work(&dbs_work,
446                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
447         return;
448 }
449
450 static inline void dbs_timer_exit(void)
451 {
452         cancel_delayed_work(&dbs_work);
453         return;
454 }
455
456 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
457                                    unsigned int event)
458 {
459         unsigned int cpu = policy->cpu;
460         struct cpu_dbs_info_s *this_dbs_info;
461         unsigned int j;
462
463         this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
464
465         switch (event) {
466         case CPUFREQ_GOV_START:
467                 if ((!cpu_online(cpu)) || 
468                     (!policy->cur))
469                         return -EINVAL;
470
471                 if (policy->cpuinfo.transition_latency >
472                                 (TRANSITION_LATENCY_LIMIT * 1000))
473                         return -EINVAL;
474                 if (this_dbs_info->enable) /* Already enabled */
475                         break;
476                  
477                 down(&dbs_sem);
478                 for_each_cpu_mask(j, policy->cpus) {
479                         struct cpu_dbs_info_s *j_dbs_info;
480                         j_dbs_info = &per_cpu(cpu_dbs_info, j);
481                         j_dbs_info->cur_policy = policy;
482                 
483                         j_dbs_info->prev_cpu_idle_up = get_cpu_idle_time(j);
484                         j_dbs_info->prev_cpu_idle_down
485                                 = j_dbs_info->prev_cpu_idle_up;
486                 }
487                 this_dbs_info->enable = 1;
488                 sysfs_create_group(&policy->kobj, &dbs_attr_group);
489                 dbs_enable++;
490                 /*
491                  * Start the timerschedule work, when this governor
492                  * is used for first time
493                  */
494                 if (dbs_enable == 1) {
495                         unsigned int latency;
496                         /* policy latency is in nS. Convert it to uS first */
497
498                         latency = policy->cpuinfo.transition_latency;
499                         if (latency < 1000)
500                                 latency = 1000;
501
502                         def_sampling_rate = (latency / 1000) *
503                                         DEF_SAMPLING_RATE_LATENCY_MULTIPLIER;
504                         dbs_tuners_ins.sampling_rate = def_sampling_rate;
505                         dbs_tuners_ins.ignore_nice = 0;
506                         dbs_tuners_ins.freq_step = 5;
507
508                         dbs_timer_init();
509                 }
510                 
511                 up(&dbs_sem);
512                 break;
513
514         case CPUFREQ_GOV_STOP:
515                 down(&dbs_sem);
516                 this_dbs_info->enable = 0;
517                 sysfs_remove_group(&policy->kobj, &dbs_attr_group);
518                 dbs_enable--;
519                 /*
520                  * Stop the timerschedule work, when this governor
521                  * is used for first time
522                  */
523                 if (dbs_enable == 0) 
524                         dbs_timer_exit();
525                 
526                 up(&dbs_sem);
527
528                 break;
529
530         case CPUFREQ_GOV_LIMITS:
531                 down(&dbs_sem);
532                 if (policy->max < this_dbs_info->cur_policy->cur)
533                         __cpufreq_driver_target(
534                                         this_dbs_info->cur_policy,
535                                         policy->max, CPUFREQ_RELATION_H);
536                 else if (policy->min > this_dbs_info->cur_policy->cur)
537                         __cpufreq_driver_target(
538                                         this_dbs_info->cur_policy,
539                                         policy->min, CPUFREQ_RELATION_L);
540                 up(&dbs_sem);
541                 break;
542         }
543         return 0;
544 }
545
546 static struct cpufreq_governor cpufreq_gov_dbs = {
547         .name           = "ondemand",
548         .governor       = cpufreq_governor_dbs,
549         .owner          = THIS_MODULE,
550 };
551
552 static int __init cpufreq_gov_dbs_init(void)
553 {
554         return cpufreq_register_governor(&cpufreq_gov_dbs);
555 }
556
557 static void __exit cpufreq_gov_dbs_exit(void)
558 {
559         /* Make sure that the scheduled work is indeed not running */
560         flush_scheduled_work();
561
562         cpufreq_unregister_governor(&cpufreq_gov_dbs);
563 }
564
565
566 MODULE_AUTHOR ("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
567 MODULE_DESCRIPTION ("'cpufreq_ondemand' - A dynamic cpufreq governor for "
568                 "Low Latency Frequency Transition capable processors");
569 MODULE_LICENSE ("GPL");
570
571 module_init(cpufreq_gov_dbs_init);
572 module_exit(cpufreq_gov_dbs_exit);