[CPUFREQ] cpufreq_conservative: keep ignore_nice_load and freq_step values when resel...
[linux-2.6.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 < 0 ||
180                         input <= dbs_tuners_ins.down_threshold) {
181                 mutex_unlock(&dbs_mutex);
182                 return -EINVAL;
183         }
184
185         dbs_tuners_ins.up_threshold = input;
186         mutex_unlock(&dbs_mutex);
187
188         return count;
189 }
190
191 static ssize_t store_down_threshold(struct cpufreq_policy *unused, 
192                 const char *buf, size_t count)
193 {
194         unsigned int input;
195         int ret;
196         ret = sscanf (buf, "%u", &input);
197
198         mutex_lock(&dbs_mutex);
199         if (ret != 1 || input > 100 || input < 0 ||
200                         input >= dbs_tuners_ins.up_threshold) {
201                 mutex_unlock(&dbs_mutex);
202                 return -EINVAL;
203         }
204
205         dbs_tuners_ins.down_threshold = input;
206         mutex_unlock(&dbs_mutex);
207
208         return count;
209 }
210
211 static ssize_t store_ignore_nice_load(struct cpufreq_policy *policy,
212                 const char *buf, size_t count)
213 {
214         unsigned int input;
215         int ret;
216
217         unsigned int j;
218         
219         ret = sscanf (buf, "%u", &input);
220         if ( ret != 1 )
221                 return -EINVAL;
222
223         if ( input > 1 )
224                 input = 1;
225         
226         mutex_lock(&dbs_mutex);
227         if ( input == dbs_tuners_ins.ignore_nice ) { /* nothing to do */
228                 mutex_unlock(&dbs_mutex);
229                 return count;
230         }
231         dbs_tuners_ins.ignore_nice = input;
232
233         /* we need to re-evaluate prev_cpu_idle_up and prev_cpu_idle_down */
234         for_each_online_cpu(j) {
235                 struct cpu_dbs_info_s *j_dbs_info;
236                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
237                 j_dbs_info->prev_cpu_idle_up = get_cpu_idle_time(j);
238                 j_dbs_info->prev_cpu_idle_down = j_dbs_info->prev_cpu_idle_up;
239         }
240         mutex_unlock(&dbs_mutex);
241
242         return count;
243 }
244
245 static ssize_t store_freq_step(struct cpufreq_policy *policy,
246                 const char *buf, size_t count)
247 {
248         unsigned int input;
249         int ret;
250
251         ret = sscanf (buf, "%u", &input);
252
253         if ( ret != 1 )
254                 return -EINVAL;
255
256         if ( input > 100 )
257                 input = 100;
258         
259         /* no need to test here if freq_step is zero as the user might actually
260          * want this, they would be crazy though :) */
261         mutex_lock(&dbs_mutex);
262         dbs_tuners_ins.freq_step = input;
263         mutex_unlock(&dbs_mutex);
264
265         return count;
266 }
267
268 #define define_one_rw(_name) \
269 static struct freq_attr _name = \
270 __ATTR(_name, 0644, show_##_name, store_##_name)
271
272 define_one_rw(sampling_rate);
273 define_one_rw(sampling_down_factor);
274 define_one_rw(up_threshold);
275 define_one_rw(down_threshold);
276 define_one_rw(ignore_nice_load);
277 define_one_rw(freq_step);
278
279 static struct attribute * dbs_attributes[] = {
280         &sampling_rate_max.attr,
281         &sampling_rate_min.attr,
282         &sampling_rate.attr,
283         &sampling_down_factor.attr,
284         &up_threshold.attr,
285         &down_threshold.attr,
286         &ignore_nice_load.attr,
287         &freq_step.attr,
288         NULL
289 };
290
291 static struct attribute_group dbs_attr_group = {
292         .attrs = dbs_attributes,
293         .name = "conservative",
294 };
295
296 /************************** sysfs end ************************/
297
298 static void dbs_check_cpu(int cpu)
299 {
300         unsigned int idle_ticks, up_idle_ticks, down_idle_ticks;
301         unsigned int tmp_idle_ticks, total_idle_ticks;
302         unsigned int freq_step;
303         unsigned int freq_down_sampling_rate;
304         struct cpu_dbs_info_s *this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
305         struct cpufreq_policy *policy;
306
307         if (!this_dbs_info->enable)
308                 return;
309
310         policy = this_dbs_info->cur_policy;
311
312         /* 
313          * The default safe range is 20% to 80% 
314          * Every sampling_rate, we check
315          *      - If current idle time is less than 20%, then we try to 
316          *        increase frequency
317          * Every sampling_rate*sampling_down_factor, we check
318          *      - If current idle time is more than 80%, then we try to
319          *        decrease frequency
320          *
321          * Any frequency increase takes it to the maximum frequency. 
322          * Frequency reduction happens at minimum steps of 
323          * 5% (default) of max_frequency 
324          */
325
326         /* Check for frequency increase */
327         idle_ticks = UINT_MAX;
328
329         /* Check for frequency increase */
330         total_idle_ticks = get_cpu_idle_time(cpu);
331         tmp_idle_ticks = total_idle_ticks -
332                 this_dbs_info->prev_cpu_idle_up;
333         this_dbs_info->prev_cpu_idle_up = total_idle_ticks;
334
335         if (tmp_idle_ticks < idle_ticks)
336                 idle_ticks = tmp_idle_ticks;
337
338         /* Scale idle ticks by 100 and compare with up and down ticks */
339         idle_ticks *= 100;
340         up_idle_ticks = (100 - dbs_tuners_ins.up_threshold) *
341                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
342
343         if (idle_ticks < up_idle_ticks) {
344                 this_dbs_info->down_skip = 0;
345                 this_dbs_info->prev_cpu_idle_down =
346                         this_dbs_info->prev_cpu_idle_up;
347
348                 /* if we are already at full speed then break out early */
349                 if (this_dbs_info->requested_freq == policy->max)
350                         return;
351                 
352                 freq_step = (dbs_tuners_ins.freq_step * policy->max) / 100;
353
354                 /* max freq cannot be less than 100. But who knows.... */
355                 if (unlikely(freq_step == 0))
356                         freq_step = 5;
357                 
358                 this_dbs_info->requested_freq += freq_step;
359                 if (this_dbs_info->requested_freq > policy->max)
360                         this_dbs_info->requested_freq = policy->max;
361
362                 __cpufreq_driver_target(policy, this_dbs_info->requested_freq,
363                         CPUFREQ_RELATION_H);
364                 return;
365         }
366
367         /* Check for frequency decrease */
368         this_dbs_info->down_skip++;
369         if (this_dbs_info->down_skip < dbs_tuners_ins.sampling_down_factor)
370                 return;
371
372         /* Check for frequency decrease */
373         total_idle_ticks = this_dbs_info->prev_cpu_idle_up;
374         tmp_idle_ticks = total_idle_ticks -
375                 this_dbs_info->prev_cpu_idle_down;
376         this_dbs_info->prev_cpu_idle_down = total_idle_ticks;
377
378         if (tmp_idle_ticks < idle_ticks)
379                 idle_ticks = tmp_idle_ticks;
380
381         /* Scale idle ticks by 100 and compare with up and down ticks */
382         idle_ticks *= 100;
383         this_dbs_info->down_skip = 0;
384
385         freq_down_sampling_rate = dbs_tuners_ins.sampling_rate *
386                 dbs_tuners_ins.sampling_down_factor;
387         down_idle_ticks = (100 - dbs_tuners_ins.down_threshold) *
388                 usecs_to_jiffies(freq_down_sampling_rate);
389
390         if (idle_ticks > down_idle_ticks) {
391                 /*
392                  * if we are already at the lowest speed then break out early
393                  * or if we 'cannot' reduce the speed as the user might want
394                  * freq_step to be zero
395                  */
396                 if (this_dbs_info->requested_freq == policy->min
397                                 || dbs_tuners_ins.freq_step == 0)
398                         return;
399
400                 freq_step = (dbs_tuners_ins.freq_step * policy->max) / 100;
401
402                 /* max freq cannot be less than 100. But who knows.... */
403                 if (unlikely(freq_step == 0))
404                         freq_step = 5;
405
406                 this_dbs_info->requested_freq -= freq_step;
407                 if (this_dbs_info->requested_freq < policy->min)
408                         this_dbs_info->requested_freq = policy->min;
409
410                 __cpufreq_driver_target(policy, this_dbs_info->requested_freq,
411                                 CPUFREQ_RELATION_H);
412                 return;
413         }
414 }
415
416 static void do_dbs_timer(void *data)
417
418         int i;
419         mutex_lock(&dbs_mutex);
420         for_each_online_cpu(i)
421                 dbs_check_cpu(i);
422         schedule_delayed_work(&dbs_work, 
423                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
424         mutex_unlock(&dbs_mutex);
425
426
427 static inline void dbs_timer_init(void)
428 {
429         INIT_WORK(&dbs_work, do_dbs_timer, NULL);
430         schedule_delayed_work(&dbs_work,
431                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
432         return;
433 }
434
435 static inline void dbs_timer_exit(void)
436 {
437         cancel_delayed_work(&dbs_work);
438         return;
439 }
440
441 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
442                                    unsigned int event)
443 {
444         unsigned int cpu = policy->cpu;
445         struct cpu_dbs_info_s *this_dbs_info;
446         unsigned int j;
447
448         this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
449
450         switch (event) {
451         case CPUFREQ_GOV_START:
452                 if ((!cpu_online(cpu)) || 
453                     (!policy->cur))
454                         return -EINVAL;
455
456                 if (policy->cpuinfo.transition_latency >
457                                 (TRANSITION_LATENCY_LIMIT * 1000))
458                         return -EINVAL;
459                 if (this_dbs_info->enable) /* Already enabled */
460                         break;
461                  
462                 mutex_lock(&dbs_mutex);
463                 for_each_cpu_mask(j, policy->cpus) {
464                         struct cpu_dbs_info_s *j_dbs_info;
465                         j_dbs_info = &per_cpu(cpu_dbs_info, j);
466                         j_dbs_info->cur_policy = policy;
467                 
468                         j_dbs_info->prev_cpu_idle_up = get_cpu_idle_time(cpu);
469                         j_dbs_info->prev_cpu_idle_down
470                                 = j_dbs_info->prev_cpu_idle_up;
471                 }
472                 this_dbs_info->enable = 1;
473                 this_dbs_info->down_skip = 0;
474                 this_dbs_info->requested_freq = policy->cur;
475                 sysfs_create_group(&policy->kobj, &dbs_attr_group);
476                 dbs_enable++;
477                 /*
478                  * Start the timerschedule work, when this governor
479                  * is used for first time
480                  */
481                 if (dbs_enable == 1) {
482                         unsigned int latency;
483                         /* policy latency is in nS. Convert it to uS first */
484                         latency = policy->cpuinfo.transition_latency / 1000;
485                         if (latency == 0)
486                                 latency = 1;
487
488                         def_sampling_rate = 10 * latency *
489                                         DEF_SAMPLING_RATE_LATENCY_MULTIPLIER;
490
491                         if (def_sampling_rate < MIN_STAT_SAMPLING_RATE)
492                                 def_sampling_rate = MIN_STAT_SAMPLING_RATE;
493
494                         dbs_tuners_ins.sampling_rate = def_sampling_rate;
495
496                         dbs_timer_init();
497                 }
498                 
499                 mutex_unlock(&dbs_mutex);
500                 break;
501
502         case CPUFREQ_GOV_STOP:
503                 mutex_lock(&dbs_mutex);
504                 this_dbs_info->enable = 0;
505                 sysfs_remove_group(&policy->kobj, &dbs_attr_group);
506                 dbs_enable--;
507                 /*
508                  * Stop the timerschedule work, when this governor
509                  * is used for first time
510                  */
511                 if (dbs_enable == 0) 
512                         dbs_timer_exit();
513                 
514                 mutex_unlock(&dbs_mutex);
515
516                 break;
517
518         case CPUFREQ_GOV_LIMITS:
519                 mutex_lock(&dbs_mutex);
520                 if (policy->max < this_dbs_info->cur_policy->cur)
521                         __cpufreq_driver_target(
522                                         this_dbs_info->cur_policy,
523                                         policy->max, CPUFREQ_RELATION_H);
524                 else if (policy->min > this_dbs_info->cur_policy->cur)
525                         __cpufreq_driver_target(
526                                         this_dbs_info->cur_policy,
527                                         policy->min, CPUFREQ_RELATION_L);
528                 mutex_unlock(&dbs_mutex);
529                 break;
530         }
531         return 0;
532 }
533
534 static struct cpufreq_governor cpufreq_gov_dbs = {
535         .name           = "conservative",
536         .governor       = cpufreq_governor_dbs,
537         .owner          = THIS_MODULE,
538 };
539
540 static int __init cpufreq_gov_dbs_init(void)
541 {
542         return cpufreq_register_governor(&cpufreq_gov_dbs);
543 }
544
545 static void __exit cpufreq_gov_dbs_exit(void)
546 {
547         /* Make sure that the scheduled work is indeed not running */
548         flush_scheduled_work();
549
550         cpufreq_unregister_governor(&cpufreq_gov_dbs);
551 }
552
553
554 MODULE_AUTHOR ("Alexander Clouter <alex-kernel@digriz.org.uk>");
555 MODULE_DESCRIPTION ("'cpufreq_conservative' - A dynamic cpufreq governor for "
556                 "Low Latency Frequency Transition capable processors "
557                 "optimised for use in a battery environment");
558 MODULE_LICENSE ("GPL");
559
560 module_init(cpufreq_gov_dbs_init);
561 module_exit(cpufreq_gov_dbs_exit);