Pull bugzilla-5452 into release branch
[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 static struct workqueue_struct *dbs_workq;
78
79 struct dbs_tuners {
80         unsigned int sampling_rate;
81         unsigned int sampling_down_factor;
82         unsigned int up_threshold;
83         unsigned int ignore_nice;
84 };
85
86 static struct dbs_tuners dbs_tuners_ins = {
87         .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
88         .sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR,
89         .ignore_nice = 0,
90 };
91
92 static inline unsigned int get_cpu_idle_time(unsigned int cpu)
93 {
94         return  kstat_cpu(cpu).cpustat.idle +
95                 kstat_cpu(cpu).cpustat.iowait +
96                 ( dbs_tuners_ins.ignore_nice ?
97                   kstat_cpu(cpu).cpustat.nice :
98                   0);
99 }
100
101 /************************** sysfs interface ************************/
102 static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf)
103 {
104         return sprintf (buf, "%u\n", MAX_SAMPLING_RATE);
105 }
106
107 static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf)
108 {
109         return sprintf (buf, "%u\n", MIN_SAMPLING_RATE);
110 }
111
112 #define define_one_ro(_name)            \
113 static struct freq_attr _name =         \
114 __ATTR(_name, 0444, show_##_name, NULL)
115
116 define_one_ro(sampling_rate_max);
117 define_one_ro(sampling_rate_min);
118
119 /* cpufreq_ondemand Governor Tunables */
120 #define show_one(file_name, object)                                     \
121 static ssize_t show_##file_name                                         \
122 (struct cpufreq_policy *unused, char *buf)                              \
123 {                                                                       \
124         return sprintf(buf, "%u\n", dbs_tuners_ins.object);             \
125 }
126 show_one(sampling_rate, sampling_rate);
127 show_one(sampling_down_factor, sampling_down_factor);
128 show_one(up_threshold, up_threshold);
129 show_one(ignore_nice_load, ignore_nice);
130
131 static ssize_t store_sampling_down_factor(struct cpufreq_policy *unused,
132                 const char *buf, size_t count)
133 {
134         unsigned int input;
135         int ret;
136         ret = sscanf (buf, "%u", &input);
137         if (ret != 1 )
138                 return -EINVAL;
139
140         if (input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
141                 return -EINVAL;
142
143         mutex_lock(&dbs_mutex);
144         dbs_tuners_ins.sampling_down_factor = input;
145         mutex_unlock(&dbs_mutex);
146
147         return count;
148 }
149
150 static ssize_t store_sampling_rate(struct cpufreq_policy *unused,
151                 const char *buf, size_t count)
152 {
153         unsigned int input;
154         int ret;
155         ret = sscanf (buf, "%u", &input);
156
157         mutex_lock(&dbs_mutex);
158         if (ret != 1 || input > MAX_SAMPLING_RATE || input < MIN_SAMPLING_RATE) {
159                 mutex_unlock(&dbs_mutex);
160                 return -EINVAL;
161         }
162
163         dbs_tuners_ins.sampling_rate = input;
164         mutex_unlock(&dbs_mutex);
165
166         return count;
167 }
168
169 static ssize_t store_up_threshold(struct cpufreq_policy *unused,
170                 const char *buf, size_t count)
171 {
172         unsigned int input;
173         int ret;
174         ret = sscanf (buf, "%u", &input);
175
176         mutex_lock(&dbs_mutex);
177         if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
178                         input < MIN_FREQUENCY_UP_THRESHOLD) {
179                 mutex_unlock(&dbs_mutex);
180                 return -EINVAL;
181         }
182
183         dbs_tuners_ins.up_threshold = input;
184         mutex_unlock(&dbs_mutex);
185
186         return count;
187 }
188
189 static ssize_t store_ignore_nice_load(struct cpufreq_policy *policy,
190                 const char *buf, size_t count)
191 {
192         unsigned int input;
193         int ret;
194
195         unsigned int j;
196
197         ret = sscanf (buf, "%u", &input);
198         if ( ret != 1 )
199                 return -EINVAL;
200
201         if ( input > 1 )
202                 input = 1;
203
204         mutex_lock(&dbs_mutex);
205         if ( input == dbs_tuners_ins.ignore_nice ) { /* nothing to do */
206                 mutex_unlock(&dbs_mutex);
207                 return count;
208         }
209         dbs_tuners_ins.ignore_nice = input;
210
211         /* we need to re-evaluate prev_cpu_idle_up and prev_cpu_idle_down */
212         for_each_online_cpu(j) {
213                 struct cpu_dbs_info_s *j_dbs_info;
214                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
215                 j_dbs_info->prev_cpu_idle_up = get_cpu_idle_time(j);
216                 j_dbs_info->prev_cpu_idle_down = j_dbs_info->prev_cpu_idle_up;
217         }
218         mutex_unlock(&dbs_mutex);
219
220         return count;
221 }
222
223 #define define_one_rw(_name) \
224 static struct freq_attr _name = \
225 __ATTR(_name, 0644, show_##_name, store_##_name)
226
227 define_one_rw(sampling_rate);
228 define_one_rw(sampling_down_factor);
229 define_one_rw(up_threshold);
230 define_one_rw(ignore_nice_load);
231
232 static struct attribute * dbs_attributes[] = {
233         &sampling_rate_max.attr,
234         &sampling_rate_min.attr,
235         &sampling_rate.attr,
236         &sampling_down_factor.attr,
237         &up_threshold.attr,
238         &ignore_nice_load.attr,
239         NULL
240 };
241
242 static struct attribute_group dbs_attr_group = {
243         .attrs = dbs_attributes,
244         .name = "ondemand",
245 };
246
247 /************************** sysfs end ************************/
248
249 static void dbs_check_cpu(int cpu)
250 {
251         unsigned int idle_ticks, up_idle_ticks, total_ticks;
252         unsigned int freq_next;
253         unsigned int freq_down_sampling_rate;
254         static int down_skip[NR_CPUS];
255         struct cpu_dbs_info_s *this_dbs_info;
256
257         struct cpufreq_policy *policy;
258         unsigned int j;
259
260         this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
261         if (!this_dbs_info->enable)
262                 return;
263
264         policy = this_dbs_info->cur_policy;
265         /*
266          * Every sampling_rate, we check, if current idle time is less
267          * than 20% (default), then we try to increase frequency
268          * Every sampling_rate*sampling_down_factor, we look for a the lowest
269          * frequency which can sustain the load while keeping idle time over
270          * 30%. If such a frequency exist, we try to decrease to this frequency.
271          *
272          * Any frequency increase takes it to the maximum frequency.
273          * Frequency reduction happens at minimum steps of
274          * 5% (default) of current frequency
275          */
276
277         /* Check for frequency increase */
278         idle_ticks = UINT_MAX;
279         for_each_cpu_mask(j, policy->cpus) {
280                 unsigned int tmp_idle_ticks, total_idle_ticks;
281                 struct cpu_dbs_info_s *j_dbs_info;
282
283                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
284                 total_idle_ticks = get_cpu_idle_time(j);
285                 tmp_idle_ticks = total_idle_ticks -
286                         j_dbs_info->prev_cpu_idle_up;
287                 j_dbs_info->prev_cpu_idle_up = total_idle_ticks;
288
289                 if (tmp_idle_ticks < idle_ticks)
290                         idle_ticks = tmp_idle_ticks;
291         }
292
293         /* Scale idle ticks by 100 and compare with up and down ticks */
294         idle_ticks *= 100;
295         up_idle_ticks = (100 - dbs_tuners_ins.up_threshold) *
296                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
297
298         if (idle_ticks < up_idle_ticks) {
299                 down_skip[cpu] = 0;
300                 for_each_cpu_mask(j, policy->cpus) {
301                         struct cpu_dbs_info_s *j_dbs_info;
302
303                         j_dbs_info = &per_cpu(cpu_dbs_info, j);
304                         j_dbs_info->prev_cpu_idle_down =
305                                         j_dbs_info->prev_cpu_idle_up;
306                 }
307                 /* if we are already at full speed then break out early */
308                 if (policy->cur == policy->max)
309                         return;
310
311                 __cpufreq_driver_target(policy, policy->max,
312                         CPUFREQ_RELATION_H);
313                 return;
314         }
315
316         /* Check for frequency decrease */
317         down_skip[cpu]++;
318         if (down_skip[cpu] < dbs_tuners_ins.sampling_down_factor)
319                 return;
320
321         idle_ticks = UINT_MAX;
322         for_each_cpu_mask(j, policy->cpus) {
323                 unsigned int tmp_idle_ticks, total_idle_ticks;
324                 struct cpu_dbs_info_s *j_dbs_info;
325
326                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
327                 /* Check for frequency decrease */
328                 total_idle_ticks = j_dbs_info->prev_cpu_idle_up;
329                 tmp_idle_ticks = total_idle_ticks -
330                         j_dbs_info->prev_cpu_idle_down;
331                 j_dbs_info->prev_cpu_idle_down = total_idle_ticks;
332
333                 if (tmp_idle_ticks < idle_ticks)
334                         idle_ticks = tmp_idle_ticks;
335         }
336
337         down_skip[cpu] = 0;
338         /* if we cannot reduce the frequency anymore, break out early */
339         if (policy->cur == policy->min)
340                 return;
341
342         /* Compute how many ticks there are between two measurements */
343         freq_down_sampling_rate = dbs_tuners_ins.sampling_rate *
344                 dbs_tuners_ins.sampling_down_factor;
345         total_ticks = usecs_to_jiffies(freq_down_sampling_rate);
346
347         /*
348          * The optimal frequency is the frequency that is the lowest that
349          * can support the current CPU usage without triggering the up
350          * policy. To be safe, we focus 10 points under the threshold.
351          */
352         freq_next = ((total_ticks - idle_ticks) * 100) / total_ticks;
353         freq_next = (freq_next * policy->cur) /
354                         (dbs_tuners_ins.up_threshold - 10);
355
356         if (freq_next < policy->min)
357                 freq_next = policy->min;
358
359         if (freq_next <= ((policy->cur * 95) / 100))
360                 __cpufreq_driver_target(policy, freq_next, CPUFREQ_RELATION_L);
361 }
362
363 static void do_dbs_timer(void *data)
364 {
365         int i;
366         mutex_lock(&dbs_mutex);
367         for_each_online_cpu(i)
368                 dbs_check_cpu(i);
369         queue_delayed_work(dbs_workq, &dbs_work,
370                            usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
371         mutex_unlock(&dbs_mutex);
372 }
373
374 static inline void dbs_timer_init(void)
375 {
376         INIT_WORK(&dbs_work, do_dbs_timer, NULL);
377         if (!dbs_workq)
378                 dbs_workq = create_singlethread_workqueue("ondemand");
379         if (!dbs_workq) {
380                 printk(KERN_ERR "ondemand: Cannot initialize kernel thread\n");
381                 return;
382         }
383         queue_delayed_work(dbs_workq, &dbs_work,
384                            usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
385         return;
386 }
387
388 static inline void dbs_timer_exit(void)
389 {
390         if (dbs_workq)
391                 cancel_rearming_delayed_workqueue(dbs_workq, &dbs_work);
392 }
393
394 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
395                                    unsigned int event)
396 {
397         unsigned int cpu = policy->cpu;
398         struct cpu_dbs_info_s *this_dbs_info;
399         unsigned int j;
400
401         this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
402
403         switch (event) {
404         case CPUFREQ_GOV_START:
405                 if ((!cpu_online(cpu)) ||
406                     (!policy->cur))
407                         return -EINVAL;
408
409                 if (policy->cpuinfo.transition_latency >
410                                 (TRANSITION_LATENCY_LIMIT * 1000)) {
411                         printk(KERN_WARNING "ondemand governor failed to load "
412                                "due to too long transition latency\n");
413                         return -EINVAL;
414                 }
415                 if (this_dbs_info->enable) /* Already enabled */
416                         break;
417
418                 mutex_lock(&dbs_mutex);
419                 for_each_cpu_mask(j, policy->cpus) {
420                         struct cpu_dbs_info_s *j_dbs_info;
421                         j_dbs_info = &per_cpu(cpu_dbs_info, j);
422                         j_dbs_info->cur_policy = policy;
423
424                         j_dbs_info->prev_cpu_idle_up = get_cpu_idle_time(j);
425                         j_dbs_info->prev_cpu_idle_down
426                                 = j_dbs_info->prev_cpu_idle_up;
427                 }
428                 this_dbs_info->enable = 1;
429                 sysfs_create_group(&policy->kobj, &dbs_attr_group);
430                 dbs_enable++;
431                 /*
432                  * Start the timerschedule work, when this governor
433                  * is used for first time
434                  */
435                 if (dbs_enable == 1) {
436                         unsigned int latency;
437                         /* policy latency is in nS. Convert it to uS first */
438                         latency = policy->cpuinfo.transition_latency / 1000;
439                         if (latency == 0)
440                                 latency = 1;
441
442                         def_sampling_rate = latency *
443                                         DEF_SAMPLING_RATE_LATENCY_MULTIPLIER;
444
445                         if (def_sampling_rate < MIN_STAT_SAMPLING_RATE)
446                                 def_sampling_rate = MIN_STAT_SAMPLING_RATE;
447
448                         dbs_tuners_ins.sampling_rate = def_sampling_rate;
449                         dbs_timer_init();
450                 }
451
452                 mutex_unlock(&dbs_mutex);
453                 break;
454
455         case CPUFREQ_GOV_STOP:
456                 mutex_lock(&dbs_mutex);
457                 this_dbs_info->enable = 0;
458                 sysfs_remove_group(&policy->kobj, &dbs_attr_group);
459                 dbs_enable--;
460                 /*
461                  * Stop the timerschedule work, when this governor
462                  * is used for first time
463                  */
464                 if (dbs_enable == 0)
465                         dbs_timer_exit();
466
467                 mutex_unlock(&dbs_mutex);
468
469                 break;
470
471         case CPUFREQ_GOV_LIMITS:
472                 mutex_lock(&dbs_mutex);
473                 if (policy->max < this_dbs_info->cur_policy->cur)
474                         __cpufreq_driver_target(
475                                         this_dbs_info->cur_policy,
476                                         policy->max, CPUFREQ_RELATION_H);
477                 else if (policy->min > this_dbs_info->cur_policy->cur)
478                         __cpufreq_driver_target(
479                                         this_dbs_info->cur_policy,
480                                         policy->min, CPUFREQ_RELATION_L);
481                 mutex_unlock(&dbs_mutex);
482                 break;
483         }
484         return 0;
485 }
486
487 static struct cpufreq_governor cpufreq_gov_dbs = {
488         .name           = "ondemand",
489         .governor       = cpufreq_governor_dbs,
490         .owner          = THIS_MODULE,
491 };
492
493 static int __init cpufreq_gov_dbs_init(void)
494 {
495         return cpufreq_register_governor(&cpufreq_gov_dbs);
496 }
497
498 static void __exit cpufreq_gov_dbs_exit(void)
499 {
500         /* Make sure that the scheduled work is indeed not running.
501            Assumes the timer has been cancelled first. */
502         if (dbs_workq) {
503                 flush_workqueue(dbs_workq);
504                 destroy_workqueue(dbs_workq);
505         }
506
507         cpufreq_unregister_governor(&cpufreq_gov_dbs);
508 }
509
510
511 MODULE_AUTHOR ("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
512 MODULE_DESCRIPTION ("'cpufreq_ondemand' - A dynamic cpufreq governor for "
513                 "Low Latency Frequency Transition capable processors");
514 MODULE_LICENSE ("GPL");
515
516 module_init(cpufreq_gov_dbs_init);
517 module_exit(cpufreq_gov_dbs_exit);