Merge branch 'reset-seq' of master.kernel.org:/pub/scm/linux/kernel/git/jgarzik/libat...
[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                  \
45                         (MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10))
46 #define MIN_SAMPLING_RATE                       \
47                         (def_sampling_rate / MIN_SAMPLING_RATE_RATIO)
48 #define MAX_SAMPLING_RATE                       (500 * def_sampling_rate)
49 #define DEF_SAMPLING_RATE_LATENCY_MULTIPLIER    (1000)
50 #define TRANSITION_LATENCY_LIMIT                (10 * 1000)
51
52 static void do_dbs_timer(struct work_struct *work);
53
54 /* Sampling types */
55 enum {DBS_NORMAL_SAMPLE, DBS_SUB_SAMPLE};
56
57 struct cpu_dbs_info_s {
58         cputime64_t prev_cpu_idle;
59         cputime64_t prev_cpu_wall;
60         struct cpufreq_policy *cur_policy;
61         struct delayed_work work;
62         struct cpufreq_frequency_table *freq_table;
63         unsigned int freq_lo;
64         unsigned int freq_lo_jiffies;
65         unsigned int freq_hi_jiffies;
66         int cpu;
67         unsigned int enable:1,
68                      sample_type:1;
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 /*
75  * DEADLOCK ALERT! There is a ordering requirement between cpu_hotplug
76  * lock and dbs_mutex. cpu_hotplug lock should always be held before
77  * dbs_mutex. If any function that can potentially take cpu_hotplug lock
78  * (like __cpufreq_driver_target()) is being called with dbs_mutex taken, then
79  * cpu_hotplug lock should be taken before that. Note that cpu_hotplug lock
80  * is recursive for the same process. -Venki
81  */
82 static DEFINE_MUTEX(dbs_mutex);
83
84 static struct workqueue_struct  *kondemand_wq;
85
86 static struct dbs_tuners {
87         unsigned int sampling_rate;
88         unsigned int up_threshold;
89         unsigned int ignore_nice;
90         unsigned int powersave_bias;
91 } dbs_tuners_ins = {
92         .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
93         .ignore_nice = 0,
94         .powersave_bias = 0,
95 };
96
97 static inline cputime64_t get_cpu_idle_time(unsigned int cpu)
98 {
99         cputime64_t retval;
100
101         retval = cputime64_add(kstat_cpu(cpu).cpustat.idle,
102                         kstat_cpu(cpu).cpustat.iowait);
103
104         if (dbs_tuners_ins.ignore_nice)
105                 retval = cputime64_add(retval, kstat_cpu(cpu).cpustat.nice);
106
107         return retval;
108 }
109
110 /*
111  * Find right freq to be set now with powersave_bias on.
112  * Returns the freq_hi to be used right now and will set freq_hi_jiffies,
113  * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs.
114  */
115 static unsigned int powersave_bias_target(struct cpufreq_policy *policy,
116                                           unsigned int freq_next,
117                                           unsigned int relation)
118 {
119         unsigned int freq_req, freq_reduc, freq_avg;
120         unsigned int freq_hi, freq_lo;
121         unsigned int index = 0;
122         unsigned int jiffies_total, jiffies_hi, jiffies_lo;
123         struct cpu_dbs_info_s *dbs_info = &per_cpu(cpu_dbs_info, policy->cpu);
124
125         if (!dbs_info->freq_table) {
126                 dbs_info->freq_lo = 0;
127                 dbs_info->freq_lo_jiffies = 0;
128                 return freq_next;
129         }
130
131         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next,
132                         relation, &index);
133         freq_req = dbs_info->freq_table[index].frequency;
134         freq_reduc = freq_req * dbs_tuners_ins.powersave_bias / 1000;
135         freq_avg = freq_req - freq_reduc;
136
137         /* Find freq bounds for freq_avg in freq_table */
138         index = 0;
139         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
140                         CPUFREQ_RELATION_H, &index);
141         freq_lo = dbs_info->freq_table[index].frequency;
142         index = 0;
143         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
144                         CPUFREQ_RELATION_L, &index);
145         freq_hi = dbs_info->freq_table[index].frequency;
146
147         /* Find out how long we have to be in hi and lo freqs */
148         if (freq_hi == freq_lo) {
149                 dbs_info->freq_lo = 0;
150                 dbs_info->freq_lo_jiffies = 0;
151                 return freq_lo;
152         }
153         jiffies_total = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
154         jiffies_hi = (freq_avg - freq_lo) * jiffies_total;
155         jiffies_hi += ((freq_hi - freq_lo) / 2);
156         jiffies_hi /= (freq_hi - freq_lo);
157         jiffies_lo = jiffies_total - jiffies_hi;
158         dbs_info->freq_lo = freq_lo;
159         dbs_info->freq_lo_jiffies = jiffies_lo;
160         dbs_info->freq_hi_jiffies = jiffies_hi;
161         return freq_hi;
162 }
163
164 static void ondemand_powersave_bias_init(void)
165 {
166         int i;
167         for_each_online_cpu(i) {
168                 struct cpu_dbs_info_s *dbs_info = &per_cpu(cpu_dbs_info, i);
169                 dbs_info->freq_table = cpufreq_frequency_get_table(i);
170                 dbs_info->freq_lo = 0;
171         }
172 }
173
174 /************************** sysfs interface ************************/
175 static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf)
176 {
177         return sprintf (buf, "%u\n", MAX_SAMPLING_RATE);
178 }
179
180 static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf)
181 {
182         return sprintf (buf, "%u\n", MIN_SAMPLING_RATE);
183 }
184
185 #define define_one_ro(_name)            \
186 static struct freq_attr _name =         \
187 __ATTR(_name, 0444, show_##_name, NULL)
188
189 define_one_ro(sampling_rate_max);
190 define_one_ro(sampling_rate_min);
191
192 /* cpufreq_ondemand Governor Tunables */
193 #define show_one(file_name, object)                                     \
194 static ssize_t show_##file_name                                         \
195 (struct cpufreq_policy *unused, char *buf)                              \
196 {                                                                       \
197         return sprintf(buf, "%u\n", dbs_tuners_ins.object);             \
198 }
199 show_one(sampling_rate, sampling_rate);
200 show_one(up_threshold, up_threshold);
201 show_one(ignore_nice_load, ignore_nice);
202 show_one(powersave_bias, powersave_bias);
203
204 static ssize_t store_sampling_rate(struct cpufreq_policy *unused,
205                 const char *buf, size_t count)
206 {
207         unsigned int input;
208         int ret;
209         ret = sscanf(buf, "%u", &input);
210
211         mutex_lock(&dbs_mutex);
212         if (ret != 1 || input > MAX_SAMPLING_RATE
213                      || input < MIN_SAMPLING_RATE) {
214                 mutex_unlock(&dbs_mutex);
215                 return -EINVAL;
216         }
217
218         dbs_tuners_ins.sampling_rate = input;
219         mutex_unlock(&dbs_mutex);
220
221         return count;
222 }
223
224 static ssize_t store_up_threshold(struct cpufreq_policy *unused,
225                 const char *buf, size_t count)
226 {
227         unsigned int input;
228         int ret;
229         ret = sscanf(buf, "%u", &input);
230
231         mutex_lock(&dbs_mutex);
232         if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
233                         input < MIN_FREQUENCY_UP_THRESHOLD) {
234                 mutex_unlock(&dbs_mutex);
235                 return -EINVAL;
236         }
237
238         dbs_tuners_ins.up_threshold = input;
239         mutex_unlock(&dbs_mutex);
240
241         return count;
242 }
243
244 static ssize_t store_ignore_nice_load(struct cpufreq_policy *policy,
245                 const char *buf, size_t count)
246 {
247         unsigned int input;
248         int ret;
249
250         unsigned int j;
251
252         ret = sscanf(buf, "%u", &input);
253         if ( ret != 1 )
254                 return -EINVAL;
255
256         if ( input > 1 )
257                 input = 1;
258
259         mutex_lock(&dbs_mutex);
260         if ( input == dbs_tuners_ins.ignore_nice ) { /* nothing to do */
261                 mutex_unlock(&dbs_mutex);
262                 return count;
263         }
264         dbs_tuners_ins.ignore_nice = input;
265
266         /* we need to re-evaluate prev_cpu_idle */
267         for_each_online_cpu(j) {
268                 struct cpu_dbs_info_s *dbs_info;
269                 dbs_info = &per_cpu(cpu_dbs_info, j);
270                 dbs_info->prev_cpu_idle = get_cpu_idle_time(j);
271                 dbs_info->prev_cpu_wall = get_jiffies_64();
272         }
273         mutex_unlock(&dbs_mutex);
274
275         return count;
276 }
277
278 static ssize_t store_powersave_bias(struct cpufreq_policy *unused,
279                 const char *buf, size_t count)
280 {
281         unsigned int input;
282         int ret;
283         ret = sscanf(buf, "%u", &input);
284
285         if (ret != 1)
286                 return -EINVAL;
287
288         if (input > 1000)
289                 input = 1000;
290
291         mutex_lock(&dbs_mutex);
292         dbs_tuners_ins.powersave_bias = input;
293         ondemand_powersave_bias_init();
294         mutex_unlock(&dbs_mutex);
295
296         return count;
297 }
298
299 #define define_one_rw(_name) \
300 static struct freq_attr _name = \
301 __ATTR(_name, 0644, show_##_name, store_##_name)
302
303 define_one_rw(sampling_rate);
304 define_one_rw(up_threshold);
305 define_one_rw(ignore_nice_load);
306 define_one_rw(powersave_bias);
307
308 static struct attribute * dbs_attributes[] = {
309         &sampling_rate_max.attr,
310         &sampling_rate_min.attr,
311         &sampling_rate.attr,
312         &up_threshold.attr,
313         &ignore_nice_load.attr,
314         &powersave_bias.attr,
315         NULL
316 };
317
318 static struct attribute_group dbs_attr_group = {
319         .attrs = dbs_attributes,
320         .name = "ondemand",
321 };
322
323 /************************** sysfs end ************************/
324
325 static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
326 {
327         unsigned int idle_ticks, total_ticks;
328         unsigned int load;
329         cputime64_t cur_jiffies;
330
331         struct cpufreq_policy *policy;
332         unsigned int j;
333
334         if (!this_dbs_info->enable)
335                 return;
336
337         this_dbs_info->freq_lo = 0;
338         policy = this_dbs_info->cur_policy;
339         cur_jiffies = jiffies64_to_cputime64(get_jiffies_64());
340         total_ticks = (unsigned int) cputime64_sub(cur_jiffies,
341                         this_dbs_info->prev_cpu_wall);
342         this_dbs_info->prev_cpu_wall = cur_jiffies;
343         if (!total_ticks)
344                 return;
345         /*
346          * Every sampling_rate, we check, if current idle time is less
347          * than 20% (default), then we try to increase frequency
348          * Every sampling_rate, we look for a the lowest
349          * frequency which can sustain the load while keeping idle time over
350          * 30%. If such a frequency exist, we try to decrease to this frequency.
351          *
352          * Any frequency increase takes it to the maximum frequency.
353          * Frequency reduction happens at minimum steps of
354          * 5% (default) of current frequency
355          */
356
357         /* Get Idle Time */
358         idle_ticks = UINT_MAX;
359         for_each_cpu_mask(j, policy->cpus) {
360                 cputime64_t total_idle_ticks;
361                 unsigned int tmp_idle_ticks;
362                 struct cpu_dbs_info_s *j_dbs_info;
363
364                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
365                 total_idle_ticks = get_cpu_idle_time(j);
366                 tmp_idle_ticks = (unsigned int) cputime64_sub(total_idle_ticks,
367                                 j_dbs_info->prev_cpu_idle);
368                 j_dbs_info->prev_cpu_idle = total_idle_ticks;
369
370                 if (tmp_idle_ticks < idle_ticks)
371                         idle_ticks = tmp_idle_ticks;
372         }
373         load = (100 * (total_ticks - idle_ticks)) / total_ticks;
374
375         /* Check for frequency increase */
376         if (load > dbs_tuners_ins.up_threshold) {
377                 /* if we are already at full speed then break out early */
378                 if (!dbs_tuners_ins.powersave_bias) {
379                         if (policy->cur == policy->max)
380                                 return;
381
382                         __cpufreq_driver_target(policy, policy->max,
383                                 CPUFREQ_RELATION_H);
384                 } else {
385                         int freq = powersave_bias_target(policy, policy->max,
386                                         CPUFREQ_RELATION_H);
387                         __cpufreq_driver_target(policy, freq,
388                                 CPUFREQ_RELATION_L);
389                 }
390                 return;
391         }
392
393         /* Check for frequency decrease */
394         /* if we cannot reduce the frequency anymore, break out early */
395         if (policy->cur == policy->min)
396                 return;
397
398         /*
399          * The optimal frequency is the frequency that is the lowest that
400          * can support the current CPU usage without triggering the up
401          * policy. To be safe, we focus 10 points under the threshold.
402          */
403         if (load < (dbs_tuners_ins.up_threshold - 10)) {
404                 unsigned int freq_next, freq_cur;
405
406                 freq_cur = __cpufreq_driver_getavg(policy);
407                 if (!freq_cur)
408                         freq_cur = policy->cur;
409
410                 freq_next = (freq_cur * load) /
411                         (dbs_tuners_ins.up_threshold - 10);
412
413                 if (!dbs_tuners_ins.powersave_bias) {
414                         __cpufreq_driver_target(policy, freq_next,
415                                         CPUFREQ_RELATION_L);
416                 } else {
417                         int freq = powersave_bias_target(policy, freq_next,
418                                         CPUFREQ_RELATION_L);
419                         __cpufreq_driver_target(policy, freq,
420                                 CPUFREQ_RELATION_L);
421                 }
422         }
423 }
424
425 static void do_dbs_timer(struct work_struct *work)
426 {
427         struct cpu_dbs_info_s *dbs_info =
428                 container_of(work, struct cpu_dbs_info_s, work.work);
429         unsigned int cpu = dbs_info->cpu;
430         int sample_type = dbs_info->sample_type;
431
432         /* We want all CPUs to do sampling nearly on same jiffy */
433         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
434
435         delay -= jiffies % delay;
436
437         if (lock_policy_rwsem_write(cpu) < 0)
438                 return;
439
440         if (!dbs_info->enable) {
441                 unlock_policy_rwsem_write(cpu);
442                 return;
443         }
444
445         /* Common NORMAL_SAMPLE setup */
446         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
447         if (!dbs_tuners_ins.powersave_bias ||
448             sample_type == DBS_NORMAL_SAMPLE) {
449                 dbs_check_cpu(dbs_info);
450                 if (dbs_info->freq_lo) {
451                         /* Setup timer for SUB_SAMPLE */
452                         dbs_info->sample_type = DBS_SUB_SAMPLE;
453                         delay = dbs_info->freq_hi_jiffies;
454                 }
455         } else {
456                 __cpufreq_driver_target(dbs_info->cur_policy,
457                                         dbs_info->freq_lo,
458                                         CPUFREQ_RELATION_H);
459         }
460         queue_delayed_work_on(cpu, kondemand_wq, &dbs_info->work, delay);
461         unlock_policy_rwsem_write(cpu);
462 }
463
464 static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
465 {
466         /* We want all CPUs to do sampling nearly on same jiffy */
467         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
468         delay -= jiffies % delay;
469
470         dbs_info->enable = 1;
471         ondemand_powersave_bias_init();
472         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
473         INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer);
474         queue_delayed_work_on(dbs_info->cpu, kondemand_wq, &dbs_info->work,
475                               delay);
476 }
477
478 static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
479 {
480         dbs_info->enable = 0;
481         cancel_delayed_work(&dbs_info->work);
482 }
483
484 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
485                                    unsigned int event)
486 {
487         unsigned int cpu = policy->cpu;
488         struct cpu_dbs_info_s *this_dbs_info;
489         unsigned int j;
490         int rc;
491
492         this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
493
494         switch (event) {
495         case CPUFREQ_GOV_START:
496                 if ((!cpu_online(cpu)) || (!policy->cur))
497                         return -EINVAL;
498
499                 if (policy->cpuinfo.transition_latency >
500                                 (TRANSITION_LATENCY_LIMIT * 1000)) {
501                         printk(KERN_WARNING "ondemand governor failed to load "
502                                "due to too long transition latency\n");
503                         return -EINVAL;
504                 }
505                 if (this_dbs_info->enable) /* Already enabled */
506                         break;
507
508                 mutex_lock(&dbs_mutex);
509                 dbs_enable++;
510
511                 rc = sysfs_create_group(&policy->kobj, &dbs_attr_group);
512                 if (rc) {
513                         dbs_enable--;
514                         mutex_unlock(&dbs_mutex);
515                         return rc;
516                 }
517
518                 for_each_cpu_mask(j, policy->cpus) {
519                         struct cpu_dbs_info_s *j_dbs_info;
520                         j_dbs_info = &per_cpu(cpu_dbs_info, j);
521                         j_dbs_info->cur_policy = policy;
522
523                         j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j);
524                         j_dbs_info->prev_cpu_wall = get_jiffies_64();
525                 }
526                 this_dbs_info->cpu = cpu;
527                 /*
528                  * Start the timerschedule work, when this governor
529                  * is used for first time
530                  */
531                 if (dbs_enable == 1) {
532                         unsigned int latency;
533                         /* policy latency is in nS. Convert it to uS first */
534                         latency = policy->cpuinfo.transition_latency / 1000;
535                         if (latency == 0)
536                                 latency = 1;
537
538                         def_sampling_rate = latency *
539                                         DEF_SAMPLING_RATE_LATENCY_MULTIPLIER;
540
541                         if (def_sampling_rate < MIN_STAT_SAMPLING_RATE)
542                                 def_sampling_rate = MIN_STAT_SAMPLING_RATE;
543
544                         dbs_tuners_ins.sampling_rate = def_sampling_rate;
545                 }
546                 dbs_timer_init(this_dbs_info);
547
548                 mutex_unlock(&dbs_mutex);
549                 break;
550
551         case CPUFREQ_GOV_STOP:
552                 mutex_lock(&dbs_mutex);
553                 dbs_timer_exit(this_dbs_info);
554                 sysfs_remove_group(&policy->kobj, &dbs_attr_group);
555                 dbs_enable--;
556                 mutex_unlock(&dbs_mutex);
557
558                 break;
559
560         case CPUFREQ_GOV_LIMITS:
561                 mutex_lock(&dbs_mutex);
562                 if (policy->max < this_dbs_info->cur_policy->cur)
563                         __cpufreq_driver_target(this_dbs_info->cur_policy,
564                                                 policy->max,
565                                                 CPUFREQ_RELATION_H);
566                 else if (policy->min > this_dbs_info->cur_policy->cur)
567                         __cpufreq_driver_target(this_dbs_info->cur_policy,
568                                                 policy->min,
569                                                 CPUFREQ_RELATION_L);
570                 mutex_unlock(&dbs_mutex);
571                 break;
572         }
573         return 0;
574 }
575
576 static struct cpufreq_governor cpufreq_gov_dbs = {
577         .name = "ondemand",
578         .governor = cpufreq_governor_dbs,
579         .owner = THIS_MODULE,
580 };
581
582 static int __init cpufreq_gov_dbs_init(void)
583 {
584         kondemand_wq = create_workqueue("kondemand");
585         if (!kondemand_wq) {
586                 printk(KERN_ERR "Creation of kondemand failed\n");
587                 return -EFAULT;
588         }
589         return cpufreq_register_governor(&cpufreq_gov_dbs);
590 }
591
592 static void __exit cpufreq_gov_dbs_exit(void)
593 {
594         cpufreq_unregister_governor(&cpufreq_gov_dbs);
595         destroy_workqueue(kondemand_wq);
596 }
597
598
599 MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
600 MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
601 MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
602                    "Low Latency Frequency Transition capable processors");
603 MODULE_LICENSE("GPL");
604
605 module_init(cpufreq_gov_dbs_init);
606 module_exit(cpufreq_gov_dbs_exit);
607