Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/roland...
[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/init.h>
16 #include <linux/cpufreq.h>
17 #include <linux/cpu.h>
18 #include <linux/jiffies.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/mutex.h>
21
22 /*
23  * dbs is used in this file as a shortform for demandbased switching
24  * It helps to keep variable names smaller, simpler
25  */
26
27 #define DEF_FREQUENCY_UP_THRESHOLD              (80)
28 #define MIN_FREQUENCY_UP_THRESHOLD              (11)
29 #define MAX_FREQUENCY_UP_THRESHOLD              (100)
30
31 /*
32  * The polling frequency of this governor depends on the capability of
33  * the processor. Default polling frequency is 1000 times the transition
34  * latency of the processor. The governor will work on any processor with
35  * transition latency <= 10mS, using appropriate sampling
36  * rate.
37  * For CPUs with transition latency > 10mS (mostly drivers with CPUFREQ_ETERNAL)
38  * this governor will not work.
39  * All times here are in uS.
40  */
41 static unsigned int def_sampling_rate;
42 #define MIN_SAMPLING_RATE_RATIO                 (2)
43 /* for correct statistics, we need at least 10 ticks between each measure */
44 #define MIN_STAT_SAMPLING_RATE                  (MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10))
45 #define MIN_SAMPLING_RATE                       (def_sampling_rate / MIN_SAMPLING_RATE_RATIO)
46 #define MAX_SAMPLING_RATE                       (500 * def_sampling_rate)
47 #define DEF_SAMPLING_RATE_LATENCY_MULTIPLIER    (1000)
48 #define TRANSITION_LATENCY_LIMIT                (10 * 1000)
49
50 static void do_dbs_timer(void *data);
51
52 struct cpu_dbs_info_s {
53         cputime64_t prev_cpu_idle;
54         cputime64_t prev_cpu_wall;
55         struct cpufreq_policy *cur_policy;
56         struct work_struct work;
57         unsigned int enable;
58         struct cpufreq_frequency_table *freq_table;
59         unsigned int freq_lo;
60         unsigned int freq_lo_jiffies;
61         unsigned int freq_hi_jiffies;
62 };
63 static DEFINE_PER_CPU(struct cpu_dbs_info_s, cpu_dbs_info);
64
65 static unsigned int dbs_enable; /* number of CPUs using this policy */
66
67 /*
68  * DEADLOCK ALERT! There is a ordering requirement between cpu_hotplug
69  * lock and dbs_mutex. cpu_hotplug lock should always be held before
70  * dbs_mutex. If any function that can potentially take cpu_hotplug lock
71  * (like __cpufreq_driver_target()) is being called with dbs_mutex taken, then
72  * cpu_hotplug lock should be taken before that. Note that cpu_hotplug lock
73  * is recursive for the same process. -Venki
74  */
75 static DEFINE_MUTEX(dbs_mutex);
76
77 static struct workqueue_struct  *kondemand_wq;
78
79 static struct dbs_tuners {
80         unsigned int sampling_rate;
81         unsigned int up_threshold;
82         unsigned int ignore_nice;
83         unsigned int powersave_bias;
84 } dbs_tuners_ins = {
85         .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
86         .ignore_nice = 0,
87         .powersave_bias = 0,
88 };
89
90 static inline cputime64_t get_cpu_idle_time(unsigned int cpu)
91 {
92         cputime64_t retval;
93
94         retval = cputime64_add(kstat_cpu(cpu).cpustat.idle,
95                         kstat_cpu(cpu).cpustat.iowait);
96
97         if (dbs_tuners_ins.ignore_nice)
98                 retval = cputime64_add(retval, kstat_cpu(cpu).cpustat.nice);
99
100         return retval;
101 }
102
103 /*
104  * Find right freq to be set now with powersave_bias on.
105  * Returns the freq_hi to be used right now and will set freq_hi_jiffies,
106  * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs.
107  */
108 static unsigned int powersave_bias_target(struct cpufreq_policy *policy,
109                                           unsigned int freq_next,
110                                           unsigned int relation)
111 {
112         unsigned int freq_req, freq_reduc, freq_avg;
113         unsigned int freq_hi, freq_lo;
114         unsigned int index = 0;
115         unsigned int jiffies_total, jiffies_hi, jiffies_lo;
116         struct cpu_dbs_info_s *dbs_info = &per_cpu(cpu_dbs_info, policy->cpu);
117
118         if (!dbs_info->freq_table) {
119                 dbs_info->freq_lo = 0;
120                 dbs_info->freq_lo_jiffies = 0;
121                 return freq_next;
122         }
123
124         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next,
125                         relation, &index);
126         freq_req = dbs_info->freq_table[index].frequency;
127         freq_reduc = freq_req * dbs_tuners_ins.powersave_bias / 1000;
128         freq_avg = freq_req - freq_reduc;
129
130         /* Find freq bounds for freq_avg in freq_table */
131         index = 0;
132         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
133                         CPUFREQ_RELATION_H, &index);
134         freq_lo = dbs_info->freq_table[index].frequency;
135         index = 0;
136         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
137                         CPUFREQ_RELATION_L, &index);
138         freq_hi = dbs_info->freq_table[index].frequency;
139
140         /* Find out how long we have to be in hi and lo freqs */
141         if (freq_hi == freq_lo) {
142                 dbs_info->freq_lo = 0;
143                 dbs_info->freq_lo_jiffies = 0;
144                 return freq_lo;
145         }
146         jiffies_total = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
147         jiffies_hi = (freq_avg - freq_lo) * jiffies_total;
148         jiffies_hi += ((freq_hi - freq_lo) / 2);
149         jiffies_hi /= (freq_hi - freq_lo);
150         jiffies_lo = jiffies_total - jiffies_hi;
151         dbs_info->freq_lo = freq_lo;
152         dbs_info->freq_lo_jiffies = jiffies_lo;
153         dbs_info->freq_hi_jiffies = jiffies_hi;
154         return freq_hi;
155 }
156
157 static void ondemand_powersave_bias_init(void)
158 {
159         int i;
160         for_each_online_cpu(i) {
161                 struct cpu_dbs_info_s *dbs_info = &per_cpu(cpu_dbs_info, i);
162                 dbs_info->freq_table = cpufreq_frequency_get_table(i);
163                 dbs_info->freq_lo = 0;
164         }
165 }
166
167 /************************** sysfs interface ************************/
168 static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf)
169 {
170         return sprintf (buf, "%u\n", MAX_SAMPLING_RATE);
171 }
172
173 static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf)
174 {
175         return sprintf (buf, "%u\n", MIN_SAMPLING_RATE);
176 }
177
178 #define define_one_ro(_name)            \
179 static struct freq_attr _name =         \
180 __ATTR(_name, 0444, show_##_name, NULL)
181
182 define_one_ro(sampling_rate_max);
183 define_one_ro(sampling_rate_min);
184
185 /* cpufreq_ondemand Governor Tunables */
186 #define show_one(file_name, object)                                     \
187 static ssize_t show_##file_name                                         \
188 (struct cpufreq_policy *unused, char *buf)                              \
189 {                                                                       \
190         return sprintf(buf, "%u\n", dbs_tuners_ins.object);             \
191 }
192 show_one(sampling_rate, sampling_rate);
193 show_one(up_threshold, up_threshold);
194 show_one(ignore_nice_load, ignore_nice);
195 show_one(powersave_bias, powersave_bias);
196
197 static ssize_t store_sampling_rate(struct cpufreq_policy *unused,
198                 const char *buf, size_t count)
199 {
200         unsigned int input;
201         int ret;
202         ret = sscanf(buf, "%u", &input);
203
204         mutex_lock(&dbs_mutex);
205         if (ret != 1 || input > MAX_SAMPLING_RATE || input < MIN_SAMPLING_RATE) {
206                 mutex_unlock(&dbs_mutex);
207                 return -EINVAL;
208         }
209
210         dbs_tuners_ins.sampling_rate = input;
211         mutex_unlock(&dbs_mutex);
212
213         return count;
214 }
215
216 static ssize_t store_up_threshold(struct cpufreq_policy *unused,
217                 const char *buf, size_t count)
218 {
219         unsigned int input;
220         int ret;
221         ret = sscanf(buf, "%u", &input);
222
223         mutex_lock(&dbs_mutex);
224         if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
225                         input < MIN_FREQUENCY_UP_THRESHOLD) {
226                 mutex_unlock(&dbs_mutex);
227                 return -EINVAL;
228         }
229
230         dbs_tuners_ins.up_threshold = input;
231         mutex_unlock(&dbs_mutex);
232
233         return count;
234 }
235
236 static ssize_t store_ignore_nice_load(struct cpufreq_policy *policy,
237                 const char *buf, size_t count)
238 {
239         unsigned int input;
240         int ret;
241
242         unsigned int j;
243
244         ret = sscanf(buf, "%u", &input);
245         if ( ret != 1 )
246                 return -EINVAL;
247
248         if ( input > 1 )
249                 input = 1;
250
251         mutex_lock(&dbs_mutex);
252         if ( input == dbs_tuners_ins.ignore_nice ) { /* nothing to do */
253                 mutex_unlock(&dbs_mutex);
254                 return count;
255         }
256         dbs_tuners_ins.ignore_nice = input;
257
258         /* we need to re-evaluate prev_cpu_idle */
259         for_each_online_cpu(j) {
260                 struct cpu_dbs_info_s *dbs_info;
261                 dbs_info = &per_cpu(cpu_dbs_info, j);
262                 dbs_info->prev_cpu_idle = get_cpu_idle_time(j);
263                 dbs_info->prev_cpu_wall = get_jiffies_64();
264         }
265         mutex_unlock(&dbs_mutex);
266
267         return count;
268 }
269
270 static ssize_t store_powersave_bias(struct cpufreq_policy *unused,
271                 const char *buf, size_t count)
272 {
273         unsigned int input;
274         int ret;
275         ret = sscanf(buf, "%u", &input);
276
277         if (ret != 1)
278                 return -EINVAL;
279
280         if (input > 1000)
281                 input = 1000;
282
283         mutex_lock(&dbs_mutex);
284         dbs_tuners_ins.powersave_bias = input;
285         ondemand_powersave_bias_init();
286         mutex_unlock(&dbs_mutex);
287
288         return count;
289 }
290
291 #define define_one_rw(_name) \
292 static struct freq_attr _name = \
293 __ATTR(_name, 0644, show_##_name, store_##_name)
294
295 define_one_rw(sampling_rate);
296 define_one_rw(up_threshold);
297 define_one_rw(ignore_nice_load);
298 define_one_rw(powersave_bias);
299
300 static struct attribute * dbs_attributes[] = {
301         &sampling_rate_max.attr,
302         &sampling_rate_min.attr,
303         &sampling_rate.attr,
304         &up_threshold.attr,
305         &ignore_nice_load.attr,
306         &powersave_bias.attr,
307         NULL
308 };
309
310 static struct attribute_group dbs_attr_group = {
311         .attrs = dbs_attributes,
312         .name = "ondemand",
313 };
314
315 /************************** sysfs end ************************/
316
317 static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
318 {
319         unsigned int idle_ticks, total_ticks;
320         unsigned int load;
321         cputime64_t cur_jiffies;
322
323         struct cpufreq_policy *policy;
324         unsigned int j;
325
326         if (!this_dbs_info->enable)
327                 return;
328
329         this_dbs_info->freq_lo = 0;
330         policy = this_dbs_info->cur_policy;
331         cur_jiffies = jiffies64_to_cputime64(get_jiffies_64());
332         total_ticks = (unsigned int) cputime64_sub(cur_jiffies,
333                         this_dbs_info->prev_cpu_wall);
334         this_dbs_info->prev_cpu_wall = cur_jiffies;
335         if (!total_ticks)
336                 return;
337         /*
338          * Every sampling_rate, we check, if current idle time is less
339          * than 20% (default), then we try to increase frequency
340          * Every sampling_rate, we look for a the lowest
341          * frequency which can sustain the load while keeping idle time over
342          * 30%. If such a frequency exist, we try to decrease to this frequency.
343          *
344          * Any frequency increase takes it to the maximum frequency.
345          * Frequency reduction happens at minimum steps of
346          * 5% (default) of current frequency
347          */
348
349         /* Get Idle Time */
350         idle_ticks = UINT_MAX;
351         for_each_cpu_mask(j, policy->cpus) {
352                 cputime64_t total_idle_ticks;
353                 unsigned int tmp_idle_ticks;
354                 struct cpu_dbs_info_s *j_dbs_info;
355
356                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
357                 total_idle_ticks = get_cpu_idle_time(j);
358                 tmp_idle_ticks = (unsigned int) cputime64_sub(total_idle_ticks,
359                                 j_dbs_info->prev_cpu_idle);
360                 j_dbs_info->prev_cpu_idle = total_idle_ticks;
361
362                 if (tmp_idle_ticks < idle_ticks)
363                         idle_ticks = tmp_idle_ticks;
364         }
365         load = (100 * (total_ticks - idle_ticks)) / total_ticks;
366
367         /* Check for frequency increase */
368         if (load > dbs_tuners_ins.up_threshold) {
369                 /* if we are already at full speed then break out early */
370                 if (!dbs_tuners_ins.powersave_bias) {
371                         if (policy->cur == policy->max)
372                                 return;
373
374                         __cpufreq_driver_target(policy, policy->max,
375                                 CPUFREQ_RELATION_H);
376                 } else {
377                         int freq = powersave_bias_target(policy, policy->max,
378                                         CPUFREQ_RELATION_H);
379                         __cpufreq_driver_target(policy, freq,
380                                 CPUFREQ_RELATION_L);
381                 }
382                 return;
383         }
384
385         /* Check for frequency decrease */
386         /* if we cannot reduce the frequency anymore, break out early */
387         if (policy->cur == policy->min)
388                 return;
389
390         /*
391          * The optimal frequency is the frequency that is the lowest that
392          * can support the current CPU usage without triggering the up
393          * policy. To be safe, we focus 10 points under the threshold.
394          */
395         if (load < (dbs_tuners_ins.up_threshold - 10)) {
396                 unsigned int freq_next = (policy->cur * load) /
397                         (dbs_tuners_ins.up_threshold - 10);
398                 if (!dbs_tuners_ins.powersave_bias) {
399                         __cpufreq_driver_target(policy, freq_next,
400                                         CPUFREQ_RELATION_L);
401                 } else {
402                         int freq = powersave_bias_target(policy, freq_next,
403                                         CPUFREQ_RELATION_L);
404                         __cpufreq_driver_target(policy, freq,
405                                 CPUFREQ_RELATION_L);
406                 }
407         }
408 }
409
410 /* Sampling types */
411 enum {DBS_NORMAL_SAMPLE, DBS_SUB_SAMPLE};
412
413 static void do_dbs_timer(void *data)
414 {
415         unsigned int cpu = smp_processor_id();
416         struct cpu_dbs_info_s *dbs_info = &per_cpu(cpu_dbs_info, cpu);
417         /* We want all CPUs to do sampling nearly on same jiffy */
418         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
419         delay -= jiffies % delay;
420
421         if (!dbs_info->enable)
422                 return;
423         /* Common NORMAL_SAMPLE setup */
424         INIT_WORK(&dbs_info->work, do_dbs_timer, (void *)DBS_NORMAL_SAMPLE);
425         if (!dbs_tuners_ins.powersave_bias ||
426             (unsigned long) data == DBS_NORMAL_SAMPLE) {
427                 lock_cpu_hotplug();
428                 dbs_check_cpu(dbs_info);
429                 unlock_cpu_hotplug();
430                 if (dbs_info->freq_lo) {
431                         /* Setup timer for SUB_SAMPLE */
432                         INIT_WORK(&dbs_info->work, do_dbs_timer,
433                                         (void *)DBS_SUB_SAMPLE);
434                         delay = dbs_info->freq_hi_jiffies;
435                 }
436         } else {
437                 __cpufreq_driver_target(dbs_info->cur_policy,
438                                         dbs_info->freq_lo,
439                                         CPUFREQ_RELATION_H);
440         }
441         queue_delayed_work_on(cpu, kondemand_wq, &dbs_info->work, delay);
442 }
443
444 static inline void dbs_timer_init(unsigned int cpu)
445 {
446         struct cpu_dbs_info_s *dbs_info = &per_cpu(cpu_dbs_info, cpu);
447         /* We want all CPUs to do sampling nearly on same jiffy */
448         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
449         delay -= jiffies % delay;
450
451         ondemand_powersave_bias_init();
452         INIT_WORK(&dbs_info->work, do_dbs_timer, NULL);
453         queue_delayed_work_on(cpu, kondemand_wq, &dbs_info->work, delay);
454 }
455
456 static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
457 {
458         dbs_info->enable = 0;
459         cancel_delayed_work(&dbs_info->work);
460         flush_workqueue(kondemand_wq);
461 }
462
463 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
464                                    unsigned int event)
465 {
466         unsigned int cpu = policy->cpu;
467         struct cpu_dbs_info_s *this_dbs_info;
468         unsigned int j;
469
470         this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
471
472         switch (event) {
473         case CPUFREQ_GOV_START:
474                 if ((!cpu_online(cpu)) || (!policy->cur))
475                         return -EINVAL;
476
477                 if (policy->cpuinfo.transition_latency >
478                                 (TRANSITION_LATENCY_LIMIT * 1000)) {
479                         printk(KERN_WARNING "ondemand governor failed to load "
480                                "due to too long transition latency\n");
481                         return -EINVAL;
482                 }
483                 if (this_dbs_info->enable) /* Already enabled */
484                         break;
485
486                 mutex_lock(&dbs_mutex);
487                 dbs_enable++;
488                 if (dbs_enable == 1) {
489                         kondemand_wq = create_workqueue("kondemand");
490                         if (!kondemand_wq) {
491                                 printk(KERN_ERR "Creation of kondemand failed\n");
492                                 dbs_enable--;
493                                 mutex_unlock(&dbs_mutex);
494                                 return -ENOSPC;
495                         }
496                 }
497                 for_each_cpu_mask(j, policy->cpus) {
498                         struct cpu_dbs_info_s *j_dbs_info;
499                         j_dbs_info = &per_cpu(cpu_dbs_info, j);
500                         j_dbs_info->cur_policy = policy;
501
502                         j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j);
503                         j_dbs_info->prev_cpu_wall = get_jiffies_64();
504                 }
505                 this_dbs_info->enable = 1;
506                 sysfs_create_group(&policy->kobj, &dbs_attr_group);
507                 /*
508                  * Start the timerschedule work, when this governor
509                  * is used for first time
510                  */
511                 if (dbs_enable == 1) {
512                         unsigned int latency;
513                         /* policy latency is in nS. Convert it to uS first */
514                         latency = policy->cpuinfo.transition_latency / 1000;
515                         if (latency == 0)
516                                 latency = 1;
517
518                         def_sampling_rate = latency *
519                                         DEF_SAMPLING_RATE_LATENCY_MULTIPLIER;
520
521                         if (def_sampling_rate < MIN_STAT_SAMPLING_RATE)
522                                 def_sampling_rate = MIN_STAT_SAMPLING_RATE;
523
524                         dbs_tuners_ins.sampling_rate = def_sampling_rate;
525                 }
526                 dbs_timer_init(policy->cpu);
527
528                 mutex_unlock(&dbs_mutex);
529                 break;
530
531         case CPUFREQ_GOV_STOP:
532                 mutex_lock(&dbs_mutex);
533                 dbs_timer_exit(this_dbs_info);
534                 sysfs_remove_group(&policy->kobj, &dbs_attr_group);
535                 dbs_enable--;
536                 if (dbs_enable == 0)
537                         destroy_workqueue(kondemand_wq);
538
539                 mutex_unlock(&dbs_mutex);
540
541                 break;
542
543         case CPUFREQ_GOV_LIMITS:
544                 mutex_lock(&dbs_mutex);
545                 if (policy->max < this_dbs_info->cur_policy->cur)
546                         __cpufreq_driver_target(this_dbs_info->cur_policy,
547                                                 policy->max,
548                                                 CPUFREQ_RELATION_H);
549                 else if (policy->min > this_dbs_info->cur_policy->cur)
550                         __cpufreq_driver_target(this_dbs_info->cur_policy,
551                                                 policy->min,
552                                                 CPUFREQ_RELATION_L);
553                 mutex_unlock(&dbs_mutex);
554                 break;
555         }
556         return 0;
557 }
558
559 static struct cpufreq_governor cpufreq_gov_dbs = {
560         .name = "ondemand",
561         .governor = cpufreq_governor_dbs,
562         .owner = THIS_MODULE,
563 };
564
565 static int __init cpufreq_gov_dbs_init(void)
566 {
567         return cpufreq_register_governor(&cpufreq_gov_dbs);
568 }
569
570 static void __exit cpufreq_gov_dbs_exit(void)
571 {
572         cpufreq_unregister_governor(&cpufreq_gov_dbs);
573 }
574
575
576 MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
577 MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
578 MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
579                    "Low Latency Frequency Transition capable processors");
580 MODULE_LICENSE("GPL");
581
582 module_init(cpufreq_gov_dbs_init);
583 module_exit(cpufreq_gov_dbs_exit);