[CPUFREQ] Conservative cpufreq governer
[pandora-kernel.git] / drivers / cpufreq / cpufreq_conservative.c
1 /*
2  *  drivers/cpufreq/cpufreq_conservative.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  *            (C)  2004 Alexander Clouter <alex-kernel@digriz.org.uk>
8  *
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.
12  */
13
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/sched.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
32 /*
33  * dbs is used in this file as a shortform for demandbased switching
34  * It helps to keep variable names smaller, simpler
35  */
36
37 #define DEF_FREQUENCY_UP_THRESHOLD              (80)
38 #define MIN_FREQUENCY_UP_THRESHOLD              (0)
39 #define MAX_FREQUENCY_UP_THRESHOLD              (100)
40
41 #define DEF_FREQUENCY_DOWN_THRESHOLD            (20)
42 #define MIN_FREQUENCY_DOWN_THRESHOLD            (0)
43 #define MAX_FREQUENCY_DOWN_THRESHOLD            (100)
44
45 /* 
46  * The polling frequency of this governor depends on the capability of 
47  * the processor. Default polling frequency is 1000 times the transition
48  * latency of the processor. The governor will work on any processor with 
49  * transition latency <= 10mS, using appropriate sampling 
50  * rate.
51  * For CPUs with transition latency > 10mS (mostly drivers with CPUFREQ_ETERNAL)
52  * this governor will not work.
53  * All times here are in uS.
54  */
55 static unsigned int                             def_sampling_rate;
56 #define MIN_SAMPLING_RATE                       (def_sampling_rate / 2)
57 #define MAX_SAMPLING_RATE                       (500 * def_sampling_rate)
58 #define DEF_SAMPLING_RATE_LATENCY_MULTIPLIER    (100000)
59 #define DEF_SAMPLING_DOWN_FACTOR                (5)
60 #define TRANSITION_LATENCY_LIMIT                (10 * 1000)
61
62 static void do_dbs_timer(void *data);
63
64 struct 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 };
70 static DEFINE_PER_CPU(struct cpu_dbs_info_s, cpu_dbs_info);
71
72 static unsigned int dbs_enable; /* number of CPUs using this policy */
73
74 static DECLARE_MUTEX    (dbs_sem);
75 static DECLARE_WORK     (dbs_work, do_dbs_timer, NULL);
76
77 struct dbs_tuners {
78         unsigned int            sampling_rate;
79         unsigned int            sampling_down_factor;
80         unsigned int            up_threshold;
81         unsigned int            down_threshold;
82         unsigned int            ignore_nice;
83         unsigned int            freq_step;
84 };
85
86 static struct dbs_tuners dbs_tuners_ins = {
87         .up_threshold           = DEF_FREQUENCY_UP_THRESHOLD,
88         .down_threshold         = DEF_FREQUENCY_DOWN_THRESHOLD,
89         .sampling_down_factor   = DEF_SAMPLING_DOWN_FACTOR,
90 };
91
92 /************************** sysfs interface ************************/
93 static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf)
94 {
95         return sprintf (buf, "%u\n", MAX_SAMPLING_RATE);
96 }
97
98 static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf)
99 {
100         return sprintf (buf, "%u\n", MIN_SAMPLING_RATE);
101 }
102
103 #define define_one_ro(_name)                                    \
104 static struct freq_attr _name =                                 \
105 __ATTR(_name, 0444, show_##_name, NULL)
106
107 define_one_ro(sampling_rate_max);
108 define_one_ro(sampling_rate_min);
109
110 /* cpufreq_conservative Governor Tunables */
111 #define show_one(file_name, object)                                     \
112 static ssize_t show_##file_name                                         \
113 (struct cpufreq_policy *unused, char *buf)                              \
114 {                                                                       \
115         return sprintf(buf, "%u\n", dbs_tuners_ins.object);             \
116 }
117 show_one(sampling_rate, sampling_rate);
118 show_one(sampling_down_factor, sampling_down_factor);
119 show_one(up_threshold, up_threshold);
120 show_one(down_threshold, down_threshold);
121 show_one(ignore_nice, ignore_nice);
122 show_one(freq_step, freq_step);
123
124 static ssize_t store_sampling_down_factor(struct cpufreq_policy *unused, 
125                 const char *buf, size_t count)
126 {
127         unsigned int input;
128         int ret;
129         ret = sscanf (buf, "%u", &input);
130         if (ret != 1 )
131                 return -EINVAL;
132
133         down(&dbs_sem);
134         dbs_tuners_ins.sampling_down_factor = input;
135         up(&dbs_sem);
136
137         return count;
138 }
139
140 static ssize_t store_sampling_rate(struct cpufreq_policy *unused, 
141                 const char *buf, size_t count)
142 {
143         unsigned int input;
144         int ret;
145         ret = sscanf (buf, "%u", &input);
146
147         down(&dbs_sem);
148         if (ret != 1 || input > MAX_SAMPLING_RATE || input < MIN_SAMPLING_RATE) {
149                 up(&dbs_sem);
150                 return -EINVAL;
151         }
152
153         dbs_tuners_ins.sampling_rate = input;
154         up(&dbs_sem);
155
156         return count;
157 }
158
159 static ssize_t store_up_threshold(struct cpufreq_policy *unused, 
160                 const char *buf, size_t count)
161 {
162         unsigned int input;
163         int ret;
164         ret = sscanf (buf, "%u", &input);
165
166         down(&dbs_sem);
167         if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD || 
168                         input < MIN_FREQUENCY_UP_THRESHOLD ||
169                         input <= dbs_tuners_ins.down_threshold) {
170                 up(&dbs_sem);
171                 return -EINVAL;
172         }
173
174         dbs_tuners_ins.up_threshold = input;
175         up(&dbs_sem);
176
177         return count;
178 }
179
180 static ssize_t store_down_threshold(struct cpufreq_policy *unused, 
181                 const char *buf, size_t count)
182 {
183         unsigned int input;
184         int ret;
185         ret = sscanf (buf, "%u", &input);
186
187         down(&dbs_sem);
188         if (ret != 1 || input > MAX_FREQUENCY_DOWN_THRESHOLD || 
189                         input < MIN_FREQUENCY_DOWN_THRESHOLD ||
190                         input >= dbs_tuners_ins.up_threshold) {
191                 up(&dbs_sem);
192                 return -EINVAL;
193         }
194
195         dbs_tuners_ins.down_threshold = input;
196         up(&dbs_sem);
197
198         return count;
199 }
200
201 static ssize_t store_ignore_nice(struct cpufreq_policy *policy,
202                 const char *buf, size_t count)
203 {
204         unsigned int input;
205         int ret;
206
207         unsigned int j;
208         
209         ret = sscanf (buf, "%u", &input);
210         if ( ret != 1 )
211                 return -EINVAL;
212
213         if ( input > 1 )
214                 input = 1;
215         
216         down(&dbs_sem);
217         if ( input == dbs_tuners_ins.ignore_nice ) { /* nothing to do */
218                 up(&dbs_sem);
219                 return count;
220         }
221         dbs_tuners_ins.ignore_nice = input;
222
223         /* we need to re-evaluate prev_cpu_idle_up and prev_cpu_idle_down */
224         for_each_cpu_mask(j, policy->cpus) {
225                 struct cpu_dbs_info_s *j_dbs_info;
226                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
227                 j_dbs_info->cur_policy = policy;
228
229                 j_dbs_info->prev_cpu_idle_up =
230                         kstat_cpu(j).cpustat.idle +
231                         kstat_cpu(j).cpustat.iowait +
232                         ( !dbs_tuners_ins.ignore_nice
233                           ? kstat_cpu(j).cpustat.nice : 0 );
234                 j_dbs_info->prev_cpu_idle_down = j_dbs_info->prev_cpu_idle_up;
235         }
236         up(&dbs_sem);
237
238         return count;
239 }
240
241 static ssize_t store_freq_step(struct cpufreq_policy *policy,
242                 const char *buf, size_t count)
243 {
244         unsigned int input;
245         int ret;
246
247         ret = sscanf (buf, "%u", &input);
248
249         if ( ret != 1 )
250                 return -EINVAL;
251
252         if ( input > 100 )
253                 input = 100;
254         
255         /* no need to test here if freq_step is zero as the user might actually
256          * want this, they would be crazy though :) */
257         down(&dbs_sem);
258         dbs_tuners_ins.freq_step = input;
259         up(&dbs_sem);
260
261         return count;
262 }
263
264 #define define_one_rw(_name) \
265 static struct freq_attr _name = \
266 __ATTR(_name, 0644, show_##_name, store_##_name)
267
268 define_one_rw(sampling_rate);
269 define_one_rw(sampling_down_factor);
270 define_one_rw(up_threshold);
271 define_one_rw(down_threshold);
272 define_one_rw(ignore_nice);
273 define_one_rw(freq_step);
274
275 static struct attribute * dbs_attributes[] = {
276         &sampling_rate_max.attr,
277         &sampling_rate_min.attr,
278         &sampling_rate.attr,
279         &sampling_down_factor.attr,
280         &up_threshold.attr,
281         &down_threshold.attr,
282         &ignore_nice.attr,
283         &freq_step.attr,
284         NULL
285 };
286
287 static struct attribute_group dbs_attr_group = {
288         .attrs = dbs_attributes,
289         .name = "conservative",
290 };
291
292 /************************** sysfs end ************************/
293
294 static void dbs_check_cpu(int cpu)
295 {
296         unsigned int idle_ticks, up_idle_ticks, down_idle_ticks;
297         unsigned int total_idle_ticks;
298         unsigned int freq_step;
299         unsigned int freq_down_sampling_rate;
300         static int down_skip[NR_CPUS];
301         static int requested_freq[NR_CPUS];
302         static unsigned short init_flag = 0;
303         struct cpu_dbs_info_s *this_dbs_info;
304         struct cpu_dbs_info_s *dbs_info;
305
306         struct cpufreq_policy *policy;
307         unsigned int j;
308
309         this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
310         if (!this_dbs_info->enable)
311                 return;
312
313         policy = this_dbs_info->cur_policy;
314
315         if ( init_flag == 0 ) {
316                 for ( /* NULL */; init_flag < NR_CPUS; init_flag++ ) {
317                         dbs_info = &per_cpu(cpu_dbs_info, init_flag);
318                         requested_freq[cpu] = dbs_info->cur_policy->cur;
319                 }
320                 init_flag = 1;
321         }
322         
323         /* 
324          * The default safe range is 20% to 80% 
325          * Every sampling_rate, we check
326          *      - If current idle time is less than 20%, then we try to 
327          *        increase frequency
328          * Every sampling_rate*sampling_down_factor, we check
329          *      - If current idle time is more than 80%, then we try to
330          *        decrease frequency
331          *
332          * Any frequency increase takes it to the maximum frequency. 
333          * Frequency reduction happens at minimum steps of 
334          * 5% (default) of max_frequency 
335          */
336
337         /* Check for frequency increase */
338         total_idle_ticks = kstat_cpu(cpu).cpustat.idle +
339                 kstat_cpu(cpu).cpustat.iowait;
340           /* consider 'nice' tasks as 'idle' time too if required */
341           if (dbs_tuners_ins.ignore_nice == 0)
342                 total_idle_ticks += kstat_cpu(cpu).cpustat.nice;
343         idle_ticks = total_idle_ticks -
344                 this_dbs_info->prev_cpu_idle_up;
345         this_dbs_info->prev_cpu_idle_up = total_idle_ticks;
346         
347
348         for_each_cpu_mask(j, policy->cpus) {
349                 unsigned int tmp_idle_ticks;
350                 struct cpu_dbs_info_s *j_dbs_info;
351
352                 if (j == cpu)
353                         continue;
354
355                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
356                 /* Check for frequency increase */
357                 total_idle_ticks = kstat_cpu(j).cpustat.idle +
358                         kstat_cpu(j).cpustat.iowait;
359                   /* consider 'nice' too? */
360                   if (dbs_tuners_ins.ignore_nice == 0)
361                            total_idle_ticks += kstat_cpu(j).cpustat.nice;
362                 tmp_idle_ticks = total_idle_ticks -
363                         j_dbs_info->prev_cpu_idle_up;
364                 j_dbs_info->prev_cpu_idle_up = total_idle_ticks;
365
366                 if (tmp_idle_ticks < idle_ticks)
367                         idle_ticks = tmp_idle_ticks;
368         }
369
370         /* Scale idle ticks by 100 and compare with up and down ticks */
371         idle_ticks *= 100;
372         up_idle_ticks = (100 - dbs_tuners_ins.up_threshold) *
373                 usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
374
375         if (idle_ticks < up_idle_ticks) {
376                 /* if we are already at full speed then break out early */
377                 if (requested_freq[cpu] == policy->max)
378                         return;
379                 
380                 freq_step = (dbs_tuners_ins.freq_step * policy->max) / 100;
381
382                 /* max freq cannot be less than 100. But who knows.... */
383                 if (unlikely(freq_step == 0))
384                         freq_step = 5;
385                 
386                 requested_freq[cpu] += freq_step;
387                 if (requested_freq[cpu] > policy->max)
388                         requested_freq[cpu] = policy->max;
389
390                 __cpufreq_driver_target(policy, requested_freq[cpu], 
391                         CPUFREQ_RELATION_H);
392                 down_skip[cpu] = 0;
393                 this_dbs_info->prev_cpu_idle_down = total_idle_ticks;
394                 return;
395         }
396
397         /* Check for frequency decrease */
398         down_skip[cpu]++;
399         if (down_skip[cpu] < dbs_tuners_ins.sampling_down_factor)
400                 return;
401
402         total_idle_ticks = kstat_cpu(cpu).cpustat.idle +
403                 kstat_cpu(cpu).cpustat.iowait;
404           /* consider 'nice' too? */
405           if (dbs_tuners_ins.ignore_nice == 0)
406                   total_idle_ticks += kstat_cpu(cpu).cpustat.nice;
407         idle_ticks = total_idle_ticks -
408                 this_dbs_info->prev_cpu_idle_down;
409         this_dbs_info->prev_cpu_idle_down = total_idle_ticks;
410
411         for_each_cpu_mask(j, policy->cpus) {
412                 unsigned int tmp_idle_ticks;
413                 struct cpu_dbs_info_s *j_dbs_info;
414
415                 if (j == cpu)
416                         continue;
417
418                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
419                 /* Check for frequency increase */
420                 total_idle_ticks = kstat_cpu(j).cpustat.idle +
421                         kstat_cpu(j).cpustat.iowait;
422                   /* consider 'nice' too? */
423                   if (dbs_tuners_ins.ignore_nice == 0)
424                         total_idle_ticks += kstat_cpu(j).cpustat.nice;
425                 tmp_idle_ticks = total_idle_ticks -
426                         j_dbs_info->prev_cpu_idle_down;
427                 j_dbs_info->prev_cpu_idle_down = total_idle_ticks;
428
429                 if (tmp_idle_ticks < idle_ticks)
430                         idle_ticks = tmp_idle_ticks;
431         }
432
433         /* Scale idle ticks by 100 and compare with up and down ticks */
434         idle_ticks *= 100;
435         down_skip[cpu] = 0;
436
437         freq_down_sampling_rate = dbs_tuners_ins.sampling_rate *
438                 dbs_tuners_ins.sampling_down_factor;
439         down_idle_ticks = (100 - dbs_tuners_ins.down_threshold) *
440                         usecs_to_jiffies(freq_down_sampling_rate);
441
442         if (idle_ticks > down_idle_ticks ) {
443                 /* if we are already at the lowest speed then break out early
444                  * or if we 'cannot' reduce the speed as the user might want
445                  * freq_step to be zero */
446                 if (requested_freq[cpu] == policy->min
447                                 || dbs_tuners_ins.freq_step == 0)
448                         return;
449
450                 freq_step = (dbs_tuners_ins.freq_step * policy->max) / 100;
451
452                 /* max freq cannot be less than 100. But who knows.... */
453                 if (unlikely(freq_step == 0))
454                         freq_step = 5;
455
456                 requested_freq[cpu] -= freq_step;
457                 if (requested_freq[cpu] < policy->min)
458                         requested_freq[cpu] = policy->min;
459
460                 __cpufreq_driver_target(policy,
461                         requested_freq[cpu],
462                         CPUFREQ_RELATION_H);
463                 return;
464         }
465 }
466
467 static void do_dbs_timer(void *data)
468
469         int i;
470         down(&dbs_sem);
471         for_each_online_cpu(i)
472                 dbs_check_cpu(i);
473         schedule_delayed_work(&dbs_work, 
474                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
475         up(&dbs_sem);
476
477
478 static inline void dbs_timer_init(void)
479 {
480         INIT_WORK(&dbs_work, do_dbs_timer, NULL);
481         schedule_delayed_work(&dbs_work,
482                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
483         return;
484 }
485
486 static inline void dbs_timer_exit(void)
487 {
488         cancel_delayed_work(&dbs_work);
489         return;
490 }
491
492 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
493                                    unsigned int event)
494 {
495         unsigned int cpu = policy->cpu;
496         struct cpu_dbs_info_s *this_dbs_info;
497         unsigned int j;
498
499         this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
500
501         switch (event) {
502         case CPUFREQ_GOV_START:
503                 if ((!cpu_online(cpu)) || 
504                     (!policy->cur))
505                         return -EINVAL;
506
507                 if (policy->cpuinfo.transition_latency >
508                                 (TRANSITION_LATENCY_LIMIT * 1000))
509                         return -EINVAL;
510                 if (this_dbs_info->enable) /* Already enabled */
511                         break;
512                  
513                 down(&dbs_sem);
514                 for_each_cpu_mask(j, policy->cpus) {
515                         struct cpu_dbs_info_s *j_dbs_info;
516                         j_dbs_info = &per_cpu(cpu_dbs_info, j);
517                         j_dbs_info->cur_policy = policy;
518                 
519                         j_dbs_info->prev_cpu_idle_up = 
520                                 kstat_cpu(j).cpustat.idle +
521                                 kstat_cpu(j).cpustat.iowait +
522                                 ( !dbs_tuners_ins.ignore_nice
523                                   ? kstat_cpu(j).cpustat.nice : 0 );
524                         j_dbs_info->prev_cpu_idle_down
525                                 = j_dbs_info->prev_cpu_idle_up;
526                 }
527                 this_dbs_info->enable = 1;
528                 sysfs_create_group(&policy->kobj, &dbs_attr_group);
529                 dbs_enable++;
530                 /*
531                  * Start the timerschedule work, when this governor
532                  * is used for first time
533                  */
534                 if (dbs_enable == 1) {
535                         unsigned int latency;
536                         /* policy latency is in nS. Convert it to uS first */
537
538                         latency = policy->cpuinfo.transition_latency;
539                         if (latency < 1000)
540                                 latency = 1000;
541
542                         def_sampling_rate = (latency / 1000) *
543                                         DEF_SAMPLING_RATE_LATENCY_MULTIPLIER;
544                         dbs_tuners_ins.sampling_rate = def_sampling_rate;
545                         dbs_tuners_ins.ignore_nice = 0;
546                         dbs_tuners_ins.freq_step = 5;
547
548                         dbs_timer_init();
549                 }
550                 
551                 up(&dbs_sem);
552                 break;
553
554         case CPUFREQ_GOV_STOP:
555                 down(&dbs_sem);
556                 this_dbs_info->enable = 0;
557                 sysfs_remove_group(&policy->kobj, &dbs_attr_group);
558                 dbs_enable--;
559                 /*
560                  * Stop the timerschedule work, when this governor
561                  * is used for first time
562                  */
563                 if (dbs_enable == 0) 
564                         dbs_timer_exit();
565                 
566                 up(&dbs_sem);
567
568                 break;
569
570         case CPUFREQ_GOV_LIMITS:
571                 down(&dbs_sem);
572                 if (policy->max < this_dbs_info->cur_policy->cur)
573                         __cpufreq_driver_target(
574                                         this_dbs_info->cur_policy,
575                                         policy->max, CPUFREQ_RELATION_H);
576                 else if (policy->min > this_dbs_info->cur_policy->cur)
577                         __cpufreq_driver_target(
578                                         this_dbs_info->cur_policy,
579                                         policy->min, CPUFREQ_RELATION_L);
580                 up(&dbs_sem);
581                 break;
582         }
583         return 0;
584 }
585
586 static struct cpufreq_governor cpufreq_gov_dbs = {
587         .name           = "conservative",
588         .governor       = cpufreq_governor_dbs,
589         .owner          = THIS_MODULE,
590 };
591
592 static int __init cpufreq_gov_dbs_init(void)
593 {
594         return cpufreq_register_governor(&cpufreq_gov_dbs);
595 }
596
597 static void __exit cpufreq_gov_dbs_exit(void)
598 {
599         /* Make sure that the scheduled work is indeed not running */
600         flush_scheduled_work();
601
602         cpufreq_unregister_governor(&cpufreq_gov_dbs);
603 }
604
605
606 MODULE_AUTHOR ("Alexander Clouter <alex-kernel@digriz.org.uk>");
607 MODULE_DESCRIPTION ("'cpufreq_conservative' - A dynamic cpufreq governor for "
608                 "Low Latency Frequency Transition capable processors "
609                 "optimised for use in a battery environment");
610 MODULE_LICENSE ("GPL");
611
612 module_init(cpufreq_gov_dbs_init);
613 module_exit(cpufreq_gov_dbs_exit);