Pull bugzilla-5737 into release branch
[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 #include <linux/mutex.h>
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 DEF_FREQUENCY_DOWN_THRESHOLD            (20)
39
40 /* 
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 with CPUFREQ_ETERNAL)
47  * this governor will not work.
48  * All times here are in uS.
49  */
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                  (MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10))
54 #define MIN_SAMPLING_RATE                       (def_sampling_rate / MIN_SAMPLING_RATE_RATIO)
55 #define MAX_SAMPLING_RATE                       (500 * def_sampling_rate)
56 #define DEF_SAMPLING_RATE_LATENCY_MULTIPLIER    (1000)
57 #define DEF_SAMPLING_DOWN_FACTOR                (1)
58 #define MAX_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         unsigned int            down_skip;
69         unsigned int            requested_freq;
70 };
71 static DEFINE_PER_CPU(struct cpu_dbs_info_s, cpu_dbs_info);
72
73 static unsigned int dbs_enable; /* number of CPUs using this policy */
74
75 static DEFINE_MUTEX     (dbs_mutex);
76 static DECLARE_WORK     (dbs_work, do_dbs_timer, NULL);
77
78 struct dbs_tuners {
79         unsigned int            sampling_rate;
80         unsigned int            sampling_down_factor;
81         unsigned int            up_threshold;
82         unsigned int            down_threshold;
83         unsigned int            ignore_nice;
84         unsigned int            freq_step;
85 };
86
87 static struct dbs_tuners dbs_tuners_ins = {
88         .up_threshold           = DEF_FREQUENCY_UP_THRESHOLD,
89         .down_threshold         = DEF_FREQUENCY_DOWN_THRESHOLD,
90         .sampling_down_factor   = DEF_SAMPLING_DOWN_FACTOR,
91         .ignore_nice            = 0,
92         .freq_step              = 5,
93 };
94
95 static inline unsigned int get_cpu_idle_time(unsigned int cpu)
96 {
97         return  kstat_cpu(cpu).cpustat.idle +
98                 kstat_cpu(cpu).cpustat.iowait +
99                 ( dbs_tuners_ins.ignore_nice ?
100                   kstat_cpu(cpu).cpustat.nice :
101                   0);
102 }
103
104 /************************** sysfs interface ************************/
105 static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf)
106 {
107         return sprintf (buf, "%u\n", MAX_SAMPLING_RATE);
108 }
109
110 static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf)
111 {
112         return sprintf (buf, "%u\n", MIN_SAMPLING_RATE);
113 }
114
115 #define define_one_ro(_name)                                    \
116 static struct freq_attr _name =                                 \
117 __ATTR(_name, 0444, show_##_name, NULL)
118
119 define_one_ro(sampling_rate_max);
120 define_one_ro(sampling_rate_min);
121
122 /* cpufreq_conservative Governor Tunables */
123 #define show_one(file_name, object)                                     \
124 static ssize_t show_##file_name                                         \
125 (struct cpufreq_policy *unused, char *buf)                              \
126 {                                                                       \
127         return sprintf(buf, "%u\n", dbs_tuners_ins.object);             \
128 }
129 show_one(sampling_rate, sampling_rate);
130 show_one(sampling_down_factor, sampling_down_factor);
131 show_one(up_threshold, up_threshold);
132 show_one(down_threshold, down_threshold);
133 show_one(ignore_nice_load, ignore_nice);
134 show_one(freq_step, freq_step);
135
136 static ssize_t store_sampling_down_factor(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         if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
143                 return -EINVAL;
144
145         mutex_lock(&dbs_mutex);
146         dbs_tuners_ins.sampling_down_factor = input;
147         mutex_unlock(&dbs_mutex);
148
149         return count;
150 }
151
152 static ssize_t store_sampling_rate(struct cpufreq_policy *unused, 
153                 const char *buf, size_t count)
154 {
155         unsigned int input;
156         int ret;
157         ret = sscanf (buf, "%u", &input);
158
159         mutex_lock(&dbs_mutex);
160         if (ret != 1 || input > MAX_SAMPLING_RATE || input < MIN_SAMPLING_RATE) {
161                 mutex_unlock(&dbs_mutex);
162                 return -EINVAL;
163         }
164
165         dbs_tuners_ins.sampling_rate = input;
166         mutex_unlock(&dbs_mutex);
167
168         return count;
169 }
170
171 static ssize_t store_up_threshold(struct cpufreq_policy *unused, 
172                 const char *buf, size_t count)
173 {
174         unsigned int input;
175         int ret;
176         ret = sscanf (buf, "%u", &input);
177
178         mutex_lock(&dbs_mutex);
179         if (ret != 1 || input > 100 || input <= dbs_tuners_ins.down_threshold) {
180                 mutex_unlock(&dbs_mutex);
181                 return -EINVAL;
182         }
183
184         dbs_tuners_ins.up_threshold = input;
185         mutex_unlock(&dbs_mutex);
186
187         return count;
188 }
189
190 static ssize_t store_down_threshold(struct cpufreq_policy *unused, 
191                 const char *buf, size_t count)
192 {
193         unsigned int input;
194         int ret;
195         ret = sscanf (buf, "%u", &input);
196
197         mutex_lock(&dbs_mutex);
198         if (ret != 1 || input > 100 || input >= dbs_tuners_ins.up_threshold) {
199                 mutex_unlock(&dbs_mutex);
200                 return -EINVAL;
201         }
202
203         dbs_tuners_ins.down_threshold = input;
204         mutex_unlock(&dbs_mutex);
205
206         return count;
207 }
208
209 static ssize_t store_ignore_nice_load(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         mutex_lock(&dbs_mutex);
225         if ( input == dbs_tuners_ins.ignore_nice ) { /* nothing to do */
226                 mutex_unlock(&dbs_mutex);
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         mutex_unlock(&dbs_mutex);
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         mutex_lock(&dbs_mutex);
260         dbs_tuners_ins.freq_step = input;
261         mutex_unlock(&dbs_mutex);
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_load);
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_load.attr,
285         &freq_step.attr,
286         NULL
287 };
288
289 static struct attribute_group dbs_attr_group = {
290         .attrs = dbs_attributes,
291         .name = "conservative",
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 tmp_idle_ticks, total_idle_ticks;
300         unsigned int freq_step;
301         unsigned int freq_down_sampling_rate;
302         struct cpu_dbs_info_s *this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
303         struct cpufreq_policy *policy;
304
305         if (!this_dbs_info->enable)
306                 return;
307
308         policy = this_dbs_info->cur_policy;
309
310         /* 
311          * The default safe range is 20% to 80% 
312          * Every sampling_rate, we check
313          *      - If current idle time is less than 20%, then we try to 
314          *        increase frequency
315          * Every sampling_rate*sampling_down_factor, we check
316          *      - If current idle time is more than 80%, then we try to
317          *        decrease frequency
318          *
319          * Any frequency increase takes it to the maximum frequency. 
320          * Frequency reduction happens at minimum steps of 
321          * 5% (default) of max_frequency 
322          */
323
324         /* Check for frequency increase */
325         idle_ticks = UINT_MAX;
326
327         /* Check for frequency increase */
328         total_idle_ticks = get_cpu_idle_time(cpu);
329         tmp_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         if (tmp_idle_ticks < idle_ticks)
334                 idle_ticks = tmp_idle_ticks;
335
336         /* Scale idle ticks by 100 and compare with up and down ticks */
337         idle_ticks *= 100;
338         up_idle_ticks = (100 - dbs_tuners_ins.up_threshold) *
339                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
340
341         if (idle_ticks < up_idle_ticks) {
342                 this_dbs_info->down_skip = 0;
343                 this_dbs_info->prev_cpu_idle_down =
344                         this_dbs_info->prev_cpu_idle_up;
345
346                 /* if we are already at full speed then break out early */
347                 if (this_dbs_info->requested_freq == policy->max)
348                         return;
349                 
350                 freq_step = (dbs_tuners_ins.freq_step * policy->max) / 100;
351
352                 /* max freq cannot be less than 100. But who knows.... */
353                 if (unlikely(freq_step == 0))
354                         freq_step = 5;
355                 
356                 this_dbs_info->requested_freq += freq_step;
357                 if (this_dbs_info->requested_freq > policy->max)
358                         this_dbs_info->requested_freq = policy->max;
359
360                 __cpufreq_driver_target(policy, this_dbs_info->requested_freq,
361                         CPUFREQ_RELATION_H);
362                 return;
363         }
364
365         /* Check for frequency decrease */
366         this_dbs_info->down_skip++;
367         if (this_dbs_info->down_skip < dbs_tuners_ins.sampling_down_factor)
368                 return;
369
370         /* Check for frequency decrease */
371         total_idle_ticks = this_dbs_info->prev_cpu_idle_up;
372         tmp_idle_ticks = total_idle_ticks -
373                 this_dbs_info->prev_cpu_idle_down;
374         this_dbs_info->prev_cpu_idle_down = total_idle_ticks;
375
376         if (tmp_idle_ticks < idle_ticks)
377                 idle_ticks = tmp_idle_ticks;
378
379         /* Scale idle ticks by 100 and compare with up and down ticks */
380         idle_ticks *= 100;
381         this_dbs_info->down_skip = 0;
382
383         freq_down_sampling_rate = dbs_tuners_ins.sampling_rate *
384                 dbs_tuners_ins.sampling_down_factor;
385         down_idle_ticks = (100 - dbs_tuners_ins.down_threshold) *
386                 usecs_to_jiffies(freq_down_sampling_rate);
387
388         if (idle_ticks > down_idle_ticks) {
389                 /*
390                  * if we are already at the lowest speed then break out early
391                  * or if we 'cannot' reduce the speed as the user might want
392                  * freq_step to be zero
393                  */
394                 if (this_dbs_info->requested_freq == policy->min
395                                 || dbs_tuners_ins.freq_step == 0)
396                         return;
397
398                 freq_step = (dbs_tuners_ins.freq_step * policy->max) / 100;
399
400                 /* max freq cannot be less than 100. But who knows.... */
401                 if (unlikely(freq_step == 0))
402                         freq_step = 5;
403
404                 this_dbs_info->requested_freq -= freq_step;
405                 if (this_dbs_info->requested_freq < policy->min)
406                         this_dbs_info->requested_freq = policy->min;
407
408                 __cpufreq_driver_target(policy, this_dbs_info->requested_freq,
409                                 CPUFREQ_RELATION_H);
410                 return;
411         }
412 }
413
414 static void do_dbs_timer(void *data)
415
416         int i;
417         mutex_lock(&dbs_mutex);
418         for_each_online_cpu(i)
419                 dbs_check_cpu(i);
420         schedule_delayed_work(&dbs_work, 
421                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
422         mutex_unlock(&dbs_mutex);
423
424
425 static inline void dbs_timer_init(void)
426 {
427         INIT_WORK(&dbs_work, do_dbs_timer, NULL);
428         schedule_delayed_work(&dbs_work,
429                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
430         return;
431 }
432
433 static inline void dbs_timer_exit(void)
434 {
435         cancel_delayed_work(&dbs_work);
436         return;
437 }
438
439 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
440                                    unsigned int event)
441 {
442         unsigned int cpu = policy->cpu;
443         struct cpu_dbs_info_s *this_dbs_info;
444         unsigned int j;
445
446         this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
447
448         switch (event) {
449         case CPUFREQ_GOV_START:
450                 if ((!cpu_online(cpu)) || 
451                     (!policy->cur))
452                         return -EINVAL;
453
454                 if (policy->cpuinfo.transition_latency >
455                                 (TRANSITION_LATENCY_LIMIT * 1000))
456                         return -EINVAL;
457                 if (this_dbs_info->enable) /* Already enabled */
458                         break;
459                  
460                 mutex_lock(&dbs_mutex);
461                 for_each_cpu_mask(j, policy->cpus) {
462                         struct cpu_dbs_info_s *j_dbs_info;
463                         j_dbs_info = &per_cpu(cpu_dbs_info, j);
464                         j_dbs_info->cur_policy = policy;
465                 
466                         j_dbs_info->prev_cpu_idle_up = get_cpu_idle_time(cpu);
467                         j_dbs_info->prev_cpu_idle_down
468                                 = j_dbs_info->prev_cpu_idle_up;
469                 }
470                 this_dbs_info->enable = 1;
471                 this_dbs_info->down_skip = 0;
472                 this_dbs_info->requested_freq = policy->cur;
473                 sysfs_create_group(&policy->kobj, &dbs_attr_group);
474                 dbs_enable++;
475                 /*
476                  * Start the timerschedule work, when this governor
477                  * is used for first time
478                  */
479                 if (dbs_enable == 1) {
480                         unsigned int latency;
481                         /* policy latency is in nS. Convert it to uS first */
482                         latency = policy->cpuinfo.transition_latency / 1000;
483                         if (latency == 0)
484                                 latency = 1;
485
486                         def_sampling_rate = 10 * latency *
487                                         DEF_SAMPLING_RATE_LATENCY_MULTIPLIER;
488
489                         if (def_sampling_rate < MIN_STAT_SAMPLING_RATE)
490                                 def_sampling_rate = MIN_STAT_SAMPLING_RATE;
491
492                         dbs_tuners_ins.sampling_rate = def_sampling_rate;
493
494                         dbs_timer_init();
495                 }
496                 
497                 mutex_unlock(&dbs_mutex);
498                 break;
499
500         case CPUFREQ_GOV_STOP:
501                 mutex_lock(&dbs_mutex);
502                 this_dbs_info->enable = 0;
503                 sysfs_remove_group(&policy->kobj, &dbs_attr_group);
504                 dbs_enable--;
505                 /*
506                  * Stop the timerschedule work, when this governor
507                  * is used for first time
508                  */
509                 if (dbs_enable == 0) 
510                         dbs_timer_exit();
511                 
512                 mutex_unlock(&dbs_mutex);
513
514                 break;
515
516         case CPUFREQ_GOV_LIMITS:
517                 mutex_lock(&dbs_mutex);
518                 if (policy->max < this_dbs_info->cur_policy->cur)
519                         __cpufreq_driver_target(
520                                         this_dbs_info->cur_policy,
521                                         policy->max, CPUFREQ_RELATION_H);
522                 else if (policy->min > this_dbs_info->cur_policy->cur)
523                         __cpufreq_driver_target(
524                                         this_dbs_info->cur_policy,
525                                         policy->min, CPUFREQ_RELATION_L);
526                 mutex_unlock(&dbs_mutex);
527                 break;
528         }
529         return 0;
530 }
531
532 static struct cpufreq_governor cpufreq_gov_dbs = {
533         .name           = "conservative",
534         .governor       = cpufreq_governor_dbs,
535         .owner          = THIS_MODULE,
536 };
537
538 static int __init cpufreq_gov_dbs_init(void)
539 {
540         return cpufreq_register_governor(&cpufreq_gov_dbs);
541 }
542
543 static void __exit cpufreq_gov_dbs_exit(void)
544 {
545         /* Make sure that the scheduled work is indeed not running */
546         flush_scheduled_work();
547
548         cpufreq_unregister_governor(&cpufreq_gov_dbs);
549 }
550
551
552 MODULE_AUTHOR ("Alexander Clouter <alex-kernel@digriz.org.uk>");
553 MODULE_DESCRIPTION ("'cpufreq_conservative' - A dynamic cpufreq governor for "
554                 "Low Latency Frequency Transition capable processors "
555                 "optimised for use in a battery environment");
556 MODULE_LICENSE ("GPL");
557
558 module_init(cpufreq_gov_dbs_init);
559 module_exit(cpufreq_gov_dbs_exit);