Merge with /pub/scm/linux/kernel/git/torvalds/linux-2.6.git
[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 #include <linux/mutex.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              (11)
39 #define MAX_FREQUENCY_UP_THRESHOLD              (100)
40
41 /* 
42  * The polling frequency of this governor depends on the capability of 
43  * the processor. Default polling frequency is 1000 times the transition
44  * latency of the processor. The governor will work on any processor with 
45  * transition latency <= 10mS, using appropriate sampling 
46  * rate.
47  * For CPUs with transition latency > 10mS (mostly drivers with CPUFREQ_ETERNAL)
48  * this governor will not work.
49  * All times here are in uS.
50  */
51 static unsigned int                             def_sampling_rate;
52 #define MIN_SAMPLING_RATE_RATIO                 (2)
53 /* for correct statistics, we need at least 10 ticks between each measure */
54 #define MIN_STAT_SAMPLING_RATE                  (MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10))
55 #define MIN_SAMPLING_RATE                       (def_sampling_rate / MIN_SAMPLING_RATE_RATIO)
56 #define MAX_SAMPLING_RATE                       (500 * def_sampling_rate)
57 #define DEF_SAMPLING_RATE_LATENCY_MULTIPLIER    (1000)
58 #define DEF_SAMPLING_DOWN_FACTOR                (1)
59 #define MAX_SAMPLING_DOWN_FACTOR                (10)
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 DEFINE_MUTEX     (dbs_mutex);
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            ignore_nice;
82 };
83
84 static struct dbs_tuners dbs_tuners_ins = {
85         .up_threshold           = DEF_FREQUENCY_UP_THRESHOLD,
86         .sampling_down_factor   = DEF_SAMPLING_DOWN_FACTOR,
87 };
88
89 static inline unsigned int get_cpu_idle_time(unsigned int cpu)
90 {
91         return  kstat_cpu(cpu).cpustat.idle +
92                 kstat_cpu(cpu).cpustat.iowait +
93                 ( dbs_tuners_ins.ignore_nice ?
94                   kstat_cpu(cpu).cpustat.nice :
95                   0);
96 }
97
98 /************************** sysfs interface ************************/
99 static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf)
100 {
101         return sprintf (buf, "%u\n", MAX_SAMPLING_RATE);
102 }
103
104 static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf)
105 {
106         return sprintf (buf, "%u\n", MIN_SAMPLING_RATE);
107 }
108
109 #define define_one_ro(_name)                                    \
110 static struct freq_attr _name =                                 \
111 __ATTR(_name, 0444, show_##_name, NULL)
112
113 define_one_ro(sampling_rate_max);
114 define_one_ro(sampling_rate_min);
115
116 /* cpufreq_ondemand Governor Tunables */
117 #define show_one(file_name, object)                                     \
118 static ssize_t show_##file_name                                         \
119 (struct cpufreq_policy *unused, char *buf)                              \
120 {                                                                       \
121         return sprintf(buf, "%u\n", dbs_tuners_ins.object);             \
122 }
123 show_one(sampling_rate, sampling_rate);
124 show_one(sampling_down_factor, sampling_down_factor);
125 show_one(up_threshold, up_threshold);
126 show_one(ignore_nice_load, ignore_nice);
127
128 static ssize_t store_sampling_down_factor(struct cpufreq_policy *unused, 
129                 const char *buf, size_t count)
130 {
131         unsigned int input;
132         int ret;
133         ret = sscanf (buf, "%u", &input);
134         if (ret != 1 )
135                 return -EINVAL;
136
137         if (input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
138                 return -EINVAL;
139
140         mutex_lock(&dbs_mutex);
141         dbs_tuners_ins.sampling_down_factor = input;
142         mutex_unlock(&dbs_mutex);
143
144         return count;
145 }
146
147 static ssize_t store_sampling_rate(struct cpufreq_policy *unused, 
148                 const char *buf, size_t count)
149 {
150         unsigned int input;
151         int ret;
152         ret = sscanf (buf, "%u", &input);
153
154         mutex_lock(&dbs_mutex);
155         if (ret != 1 || input > MAX_SAMPLING_RATE || input < MIN_SAMPLING_RATE) {
156                 mutex_unlock(&dbs_mutex);
157                 return -EINVAL;
158         }
159
160         dbs_tuners_ins.sampling_rate = input;
161         mutex_unlock(&dbs_mutex);
162
163         return count;
164 }
165
166 static ssize_t store_up_threshold(struct cpufreq_policy *unused, 
167                 const char *buf, size_t count)
168 {
169         unsigned int input;
170         int ret;
171         ret = sscanf (buf, "%u", &input);
172
173         mutex_lock(&dbs_mutex);
174         if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD || 
175                         input < MIN_FREQUENCY_UP_THRESHOLD) {
176                 mutex_unlock(&dbs_mutex);
177                 return -EINVAL;
178         }
179
180         dbs_tuners_ins.up_threshold = input;
181         mutex_unlock(&dbs_mutex);
182
183         return count;
184 }
185
186 static ssize_t store_ignore_nice_load(struct cpufreq_policy *policy,
187                 const char *buf, size_t count)
188 {
189         unsigned int input;
190         int ret;
191
192         unsigned int j;
193         
194         ret = sscanf (buf, "%u", &input);
195         if ( ret != 1 )
196                 return -EINVAL;
197
198         if ( input > 1 )
199                 input = 1;
200         
201         mutex_lock(&dbs_mutex);
202         if ( input == dbs_tuners_ins.ignore_nice ) { /* nothing to do */
203                 mutex_unlock(&dbs_mutex);
204                 return count;
205         }
206         dbs_tuners_ins.ignore_nice = input;
207
208         /* we need to re-evaluate prev_cpu_idle_up and prev_cpu_idle_down */
209         for_each_online_cpu(j) {
210                 struct cpu_dbs_info_s *j_dbs_info;
211                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
212                 j_dbs_info->prev_cpu_idle_up = get_cpu_idle_time(j);
213                 j_dbs_info->prev_cpu_idle_down = j_dbs_info->prev_cpu_idle_up;
214         }
215         mutex_unlock(&dbs_mutex);
216
217         return count;
218 }
219
220 #define define_one_rw(_name) \
221 static struct freq_attr _name = \
222 __ATTR(_name, 0644, show_##_name, store_##_name)
223
224 define_one_rw(sampling_rate);
225 define_one_rw(sampling_down_factor);
226 define_one_rw(up_threshold);
227 define_one_rw(ignore_nice_load);
228
229 static struct attribute * dbs_attributes[] = {
230         &sampling_rate_max.attr,
231         &sampling_rate_min.attr,
232         &sampling_rate.attr,
233         &sampling_down_factor.attr,
234         &up_threshold.attr,
235         &ignore_nice_load.attr,
236         NULL
237 };
238
239 static struct attribute_group dbs_attr_group = {
240         .attrs = dbs_attributes,
241         .name = "ondemand",
242 };
243
244 /************************** sysfs end ************************/
245
246 static void dbs_check_cpu(int cpu)
247 {
248         unsigned int idle_ticks, up_idle_ticks, total_ticks;
249         unsigned int freq_next;
250         unsigned int freq_down_sampling_rate;
251         static int down_skip[NR_CPUS];
252         struct cpu_dbs_info_s *this_dbs_info;
253
254         struct cpufreq_policy *policy;
255         unsigned int j;
256
257         this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
258         if (!this_dbs_info->enable)
259                 return;
260
261         policy = this_dbs_info->cur_policy;
262         /* 
263          * Every sampling_rate, we check, if current idle time is less
264          * than 20% (default), then we try to increase frequency
265          * Every sampling_rate*sampling_down_factor, we look for a the lowest
266          * frequency which can sustain the load while keeping idle time over
267          * 30%. If such a frequency exist, we try to decrease to this frequency.
268          *
269          * Any frequency increase takes it to the maximum frequency. 
270          * Frequency reduction happens at minimum steps of 
271          * 5% (default) of current frequency 
272          */
273
274         /* Check for frequency increase */
275         idle_ticks = UINT_MAX;
276         for_each_cpu_mask(j, policy->cpus) {
277                 unsigned int tmp_idle_ticks, total_idle_ticks;
278                 struct cpu_dbs_info_s *j_dbs_info;
279
280                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
281                 total_idle_ticks = get_cpu_idle_time(j);
282                 tmp_idle_ticks = total_idle_ticks -
283                         j_dbs_info->prev_cpu_idle_up;
284                 j_dbs_info->prev_cpu_idle_up = total_idle_ticks;
285
286                 if (tmp_idle_ticks < idle_ticks)
287                         idle_ticks = tmp_idle_ticks;
288         }
289
290         /* Scale idle ticks by 100 and compare with up and down ticks */
291         idle_ticks *= 100;
292         up_idle_ticks = (100 - dbs_tuners_ins.up_threshold) *
293                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
294
295         if (idle_ticks < up_idle_ticks) {
296                 down_skip[cpu] = 0;
297                 for_each_cpu_mask(j, policy->cpus) {
298                         struct cpu_dbs_info_s *j_dbs_info;
299
300                         j_dbs_info = &per_cpu(cpu_dbs_info, j);
301                         j_dbs_info->prev_cpu_idle_down = 
302                                         j_dbs_info->prev_cpu_idle_up;
303                 }
304                 /* if we are already at full speed then break out early */
305                 if (policy->cur == policy->max)
306                         return;
307                 
308                 __cpufreq_driver_target(policy, policy->max, 
309                         CPUFREQ_RELATION_H);
310                 return;
311         }
312
313         /* Check for frequency decrease */
314         down_skip[cpu]++;
315         if (down_skip[cpu] < dbs_tuners_ins.sampling_down_factor)
316                 return;
317
318         idle_ticks = UINT_MAX;
319         for_each_cpu_mask(j, policy->cpus) {
320                 unsigned int tmp_idle_ticks, total_idle_ticks;
321                 struct cpu_dbs_info_s *j_dbs_info;
322
323                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
324                 /* Check for frequency decrease */
325                 total_idle_ticks = j_dbs_info->prev_cpu_idle_up;
326                 tmp_idle_ticks = total_idle_ticks -
327                         j_dbs_info->prev_cpu_idle_down;
328                 j_dbs_info->prev_cpu_idle_down = total_idle_ticks;
329
330                 if (tmp_idle_ticks < idle_ticks)
331                         idle_ticks = tmp_idle_ticks;
332         }
333
334         down_skip[cpu] = 0;
335         /* if we cannot reduce the frequency anymore, break out early */
336         if (policy->cur == policy->min)
337                 return;
338
339         /* Compute how many ticks there are between two measurements */
340         freq_down_sampling_rate = dbs_tuners_ins.sampling_rate *
341                 dbs_tuners_ins.sampling_down_factor;
342         total_ticks = usecs_to_jiffies(freq_down_sampling_rate);
343
344         /*
345          * The optimal frequency is the frequency that is the lowest that
346          * can support the current CPU usage without triggering the up
347          * policy. To be safe, we focus 10 points under the threshold.
348          */
349         freq_next = ((total_ticks - idle_ticks) * 100) / total_ticks;
350         freq_next = (freq_next * policy->cur) / 
351                         (dbs_tuners_ins.up_threshold - 10);
352
353         if (freq_next <= ((policy->cur * 95) / 100))
354                 __cpufreq_driver_target(policy, freq_next, CPUFREQ_RELATION_L);
355 }
356
357 static void do_dbs_timer(void *data)
358
359         int i;
360         mutex_lock(&dbs_mutex);
361         for_each_online_cpu(i)
362                 dbs_check_cpu(i);
363         schedule_delayed_work(&dbs_work, 
364                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
365         mutex_unlock(&dbs_mutex);
366
367
368 static inline void dbs_timer_init(void)
369 {
370         INIT_WORK(&dbs_work, do_dbs_timer, NULL);
371         schedule_delayed_work(&dbs_work,
372                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
373         return;
374 }
375
376 static inline void dbs_timer_exit(void)
377 {
378         cancel_delayed_work(&dbs_work);
379         return;
380 }
381
382 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
383                                    unsigned int event)
384 {
385         unsigned int cpu = policy->cpu;
386         struct cpu_dbs_info_s *this_dbs_info;
387         unsigned int j;
388
389         this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
390
391         switch (event) {
392         case CPUFREQ_GOV_START:
393                 if ((!cpu_online(cpu)) || 
394                     (!policy->cur))
395                         return -EINVAL;
396
397                 if (policy->cpuinfo.transition_latency >
398                                 (TRANSITION_LATENCY_LIMIT * 1000))
399                         return -EINVAL;
400                 if (this_dbs_info->enable) /* Already enabled */
401                         break;
402                  
403                 mutex_lock(&dbs_mutex);
404                 for_each_cpu_mask(j, policy->cpus) {
405                         struct cpu_dbs_info_s *j_dbs_info;
406                         j_dbs_info = &per_cpu(cpu_dbs_info, j);
407                         j_dbs_info->cur_policy = policy;
408                 
409                         j_dbs_info->prev_cpu_idle_up = get_cpu_idle_time(j);
410                         j_dbs_info->prev_cpu_idle_down
411                                 = j_dbs_info->prev_cpu_idle_up;
412                 }
413                 this_dbs_info->enable = 1;
414                 sysfs_create_group(&policy->kobj, &dbs_attr_group);
415                 dbs_enable++;
416                 /*
417                  * Start the timerschedule work, when this governor
418                  * is used for first time
419                  */
420                 if (dbs_enable == 1) {
421                         unsigned int latency;
422                         /* policy latency is in nS. Convert it to uS first */
423                         latency = policy->cpuinfo.transition_latency / 1000;
424                         if (latency == 0)
425                                 latency = 1;
426
427                         def_sampling_rate = latency *
428                                         DEF_SAMPLING_RATE_LATENCY_MULTIPLIER;
429
430                         if (def_sampling_rate < MIN_STAT_SAMPLING_RATE)
431                                 def_sampling_rate = MIN_STAT_SAMPLING_RATE;
432
433                         dbs_tuners_ins.sampling_rate = def_sampling_rate;
434                         dbs_tuners_ins.ignore_nice = 0;
435
436                         dbs_timer_init();
437                 }
438                 
439                 mutex_unlock(&dbs_mutex);
440                 break;
441
442         case CPUFREQ_GOV_STOP:
443                 mutex_lock(&dbs_mutex);
444                 this_dbs_info->enable = 0;
445                 sysfs_remove_group(&policy->kobj, &dbs_attr_group);
446                 dbs_enable--;
447                 /*
448                  * Stop the timerschedule work, when this governor
449                  * is used for first time
450                  */
451                 if (dbs_enable == 0) 
452                         dbs_timer_exit();
453                 
454                 mutex_unlock(&dbs_mutex);
455
456                 break;
457
458         case CPUFREQ_GOV_LIMITS:
459                 mutex_lock(&dbs_mutex);
460                 if (policy->max < this_dbs_info->cur_policy->cur)
461                         __cpufreq_driver_target(
462                                         this_dbs_info->cur_policy,
463                                         policy->max, CPUFREQ_RELATION_H);
464                 else if (policy->min > this_dbs_info->cur_policy->cur)
465                         __cpufreq_driver_target(
466                                         this_dbs_info->cur_policy,
467                                         policy->min, CPUFREQ_RELATION_L);
468                 mutex_unlock(&dbs_mutex);
469                 break;
470         }
471         return 0;
472 }
473
474 static struct cpufreq_governor cpufreq_gov_dbs = {
475         .name           = "ondemand",
476         .governor       = cpufreq_governor_dbs,
477         .owner          = THIS_MODULE,
478 };
479
480 static int __init cpufreq_gov_dbs_init(void)
481 {
482         return cpufreq_register_governor(&cpufreq_gov_dbs);
483 }
484
485 static void __exit cpufreq_gov_dbs_exit(void)
486 {
487         /* Make sure that the scheduled work is indeed not running */
488         flush_scheduled_work();
489
490         cpufreq_unregister_governor(&cpufreq_gov_dbs);
491 }
492
493
494 MODULE_AUTHOR ("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
495 MODULE_DESCRIPTION ("'cpufreq_ondemand' - A dynamic cpufreq governor for "
496                 "Low Latency Frequency Transition capable processors");
497 MODULE_LICENSE ("GPL");
498
499 module_init(cpufreq_gov_dbs_init);
500 module_exit(cpufreq_gov_dbs_exit);