Merge branch 'x86-cpu-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git...
[linux-2.6.git] / 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 #include <linux/hrtimer.h>
22 #include <linux/tick.h>
23 #include <linux/ktime.h>
24 #include <linux/sched.h>
25
26 /*
27  * dbs is used in this file as a shortform for demandbased switching
28  * It helps to keep variable names smaller, simpler
29  */
30
31 #define DEF_FREQUENCY_DOWN_DIFFERENTIAL         (10)
32 #define DEF_FREQUENCY_UP_THRESHOLD              (80)
33 #define MICRO_FREQUENCY_DOWN_DIFFERENTIAL       (3)
34 #define MICRO_FREQUENCY_UP_THRESHOLD            (95)
35 #define MICRO_FREQUENCY_MIN_SAMPLE_RATE         (10000)
36 #define MIN_FREQUENCY_UP_THRESHOLD              (11)
37 #define MAX_FREQUENCY_UP_THRESHOLD              (100)
38
39 /*
40  * The polling frequency of this governor depends on the capability of
41  * the processor. Default polling frequency is 1000 times the transition
42  * latency of the processor. The governor will work on any processor with
43  * transition latency <= 10mS, using appropriate sampling
44  * rate.
45  * For CPUs with transition latency > 10mS (mostly drivers with CPUFREQ_ETERNAL)
46  * this governor will not work.
47  * All times here are in uS.
48  */
49 #define MIN_SAMPLING_RATE_RATIO                 (2)
50
51 static unsigned int min_sampling_rate;
52
53 #define LATENCY_MULTIPLIER                      (1000)
54 #define MIN_LATENCY_MULTIPLIER                  (100)
55 #define TRANSITION_LATENCY_LIMIT                (10 * 1000 * 1000)
56
57 static void do_dbs_timer(struct work_struct *work);
58 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
59                                 unsigned int event);
60
61 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
62 static
63 #endif
64 struct cpufreq_governor cpufreq_gov_ondemand = {
65        .name                   = "ondemand",
66        .governor               = cpufreq_governor_dbs,
67        .max_transition_latency = TRANSITION_LATENCY_LIMIT,
68        .owner                  = THIS_MODULE,
69 };
70
71 /* Sampling types */
72 enum {DBS_NORMAL_SAMPLE, DBS_SUB_SAMPLE};
73
74 struct cpu_dbs_info_s {
75         cputime64_t prev_cpu_idle;
76         cputime64_t prev_cpu_iowait;
77         cputime64_t prev_cpu_wall;
78         cputime64_t prev_cpu_nice;
79         struct cpufreq_policy *cur_policy;
80         struct delayed_work work;
81         struct cpufreq_frequency_table *freq_table;
82         unsigned int freq_lo;
83         unsigned int freq_lo_jiffies;
84         unsigned int freq_hi_jiffies;
85         int cpu;
86         unsigned int sample_type:1;
87         /*
88          * percpu mutex that serializes governor limit change with
89          * do_dbs_timer invocation. We do not want do_dbs_timer to run
90          * when user is changing the governor or limits.
91          */
92         struct mutex timer_mutex;
93 };
94 static DEFINE_PER_CPU(struct cpu_dbs_info_s, od_cpu_dbs_info);
95
96 static unsigned int dbs_enable; /* number of CPUs using this policy */
97
98 /*
99  * dbs_mutex protects data in dbs_tuners_ins from concurrent changes on
100  * different CPUs. It protects dbs_enable in governor start/stop.
101  */
102 static DEFINE_MUTEX(dbs_mutex);
103
104 static struct workqueue_struct  *kondemand_wq;
105
106 static struct dbs_tuners {
107         unsigned int sampling_rate;
108         unsigned int up_threshold;
109         unsigned int down_differential;
110         unsigned int ignore_nice;
111         unsigned int powersave_bias;
112         unsigned int io_is_busy;
113 } dbs_tuners_ins = {
114         .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
115         .down_differential = DEF_FREQUENCY_DOWN_DIFFERENTIAL,
116         .ignore_nice = 0,
117         .powersave_bias = 0,
118 };
119
120 static inline cputime64_t get_cpu_idle_time_jiffy(unsigned int cpu,
121                                                         cputime64_t *wall)
122 {
123         cputime64_t idle_time;
124         cputime64_t cur_wall_time;
125         cputime64_t busy_time;
126
127         cur_wall_time = jiffies64_to_cputime64(get_jiffies_64());
128         busy_time = cputime64_add(kstat_cpu(cpu).cpustat.user,
129                         kstat_cpu(cpu).cpustat.system);
130
131         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.irq);
132         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.softirq);
133         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.steal);
134         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.nice);
135
136         idle_time = cputime64_sub(cur_wall_time, busy_time);
137         if (wall)
138                 *wall = (cputime64_t)jiffies_to_usecs(cur_wall_time);
139
140         return (cputime64_t)jiffies_to_usecs(idle_time);
141 }
142
143 static inline cputime64_t get_cpu_idle_time(unsigned int cpu, cputime64_t *wall)
144 {
145         u64 idle_time = get_cpu_idle_time_us(cpu, wall);
146
147         if (idle_time == -1ULL)
148                 return get_cpu_idle_time_jiffy(cpu, wall);
149
150         return idle_time;
151 }
152
153 static inline cputime64_t get_cpu_iowait_time(unsigned int cpu, cputime64_t *wall)
154 {
155         u64 iowait_time = get_cpu_iowait_time_us(cpu, wall);
156
157         if (iowait_time == -1ULL)
158                 return 0;
159
160         return iowait_time;
161 }
162
163 /*
164  * Find right freq to be set now with powersave_bias on.
165  * Returns the freq_hi to be used right now and will set freq_hi_jiffies,
166  * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs.
167  */
168 static unsigned int powersave_bias_target(struct cpufreq_policy *policy,
169                                           unsigned int freq_next,
170                                           unsigned int relation)
171 {
172         unsigned int freq_req, freq_reduc, freq_avg;
173         unsigned int freq_hi, freq_lo;
174         unsigned int index = 0;
175         unsigned int jiffies_total, jiffies_hi, jiffies_lo;
176         struct cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
177                                                    policy->cpu);
178
179         if (!dbs_info->freq_table) {
180                 dbs_info->freq_lo = 0;
181                 dbs_info->freq_lo_jiffies = 0;
182                 return freq_next;
183         }
184
185         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next,
186                         relation, &index);
187         freq_req = dbs_info->freq_table[index].frequency;
188         freq_reduc = freq_req * dbs_tuners_ins.powersave_bias / 1000;
189         freq_avg = freq_req - freq_reduc;
190
191         /* Find freq bounds for freq_avg in freq_table */
192         index = 0;
193         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
194                         CPUFREQ_RELATION_H, &index);
195         freq_lo = dbs_info->freq_table[index].frequency;
196         index = 0;
197         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
198                         CPUFREQ_RELATION_L, &index);
199         freq_hi = dbs_info->freq_table[index].frequency;
200
201         /* Find out how long we have to be in hi and lo freqs */
202         if (freq_hi == freq_lo) {
203                 dbs_info->freq_lo = 0;
204                 dbs_info->freq_lo_jiffies = 0;
205                 return freq_lo;
206         }
207         jiffies_total = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
208         jiffies_hi = (freq_avg - freq_lo) * jiffies_total;
209         jiffies_hi += ((freq_hi - freq_lo) / 2);
210         jiffies_hi /= (freq_hi - freq_lo);
211         jiffies_lo = jiffies_total - jiffies_hi;
212         dbs_info->freq_lo = freq_lo;
213         dbs_info->freq_lo_jiffies = jiffies_lo;
214         dbs_info->freq_hi_jiffies = jiffies_hi;
215         return freq_hi;
216 }
217
218 static void ondemand_powersave_bias_init_cpu(int cpu)
219 {
220         struct cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
221         dbs_info->freq_table = cpufreq_frequency_get_table(cpu);
222         dbs_info->freq_lo = 0;
223 }
224
225 static void ondemand_powersave_bias_init(void)
226 {
227         int i;
228         for_each_online_cpu(i) {
229                 ondemand_powersave_bias_init_cpu(i);
230         }
231 }
232
233 /************************** sysfs interface ************************/
234
235 static ssize_t show_sampling_rate_max(struct kobject *kobj,
236                                       struct attribute *attr, char *buf)
237 {
238         printk_once(KERN_INFO "CPUFREQ: ondemand sampling_rate_max "
239                "sysfs file is deprecated - used by: %s\n", current->comm);
240         return sprintf(buf, "%u\n", -1U);
241 }
242
243 static ssize_t show_sampling_rate_min(struct kobject *kobj,
244                                       struct attribute *attr, char *buf)
245 {
246         return sprintf(buf, "%u\n", min_sampling_rate);
247 }
248
249 define_one_global_ro(sampling_rate_max);
250 define_one_global_ro(sampling_rate_min);
251
252 /* cpufreq_ondemand Governor Tunables */
253 #define show_one(file_name, object)                                     \
254 static ssize_t show_##file_name                                         \
255 (struct kobject *kobj, struct attribute *attr, char *buf)              \
256 {                                                                       \
257         return sprintf(buf, "%u\n", dbs_tuners_ins.object);             \
258 }
259 show_one(sampling_rate, sampling_rate);
260 show_one(io_is_busy, io_is_busy);
261 show_one(up_threshold, up_threshold);
262 show_one(ignore_nice_load, ignore_nice);
263 show_one(powersave_bias, powersave_bias);
264
265 /*** delete after deprecation time ***/
266
267 #define DEPRECATION_MSG(file_name)                                      \
268         printk_once(KERN_INFO "CPUFREQ: Per core ondemand sysfs "       \
269                     "interface is deprecated - " #file_name "\n");
270
271 #define show_one_old(file_name)                                         \
272 static ssize_t show_##file_name##_old                                   \
273 (struct cpufreq_policy *unused, char *buf)                              \
274 {                                                                       \
275         printk_once(KERN_INFO "CPUFREQ: Per core ondemand sysfs "       \
276                     "interface is deprecated - " #file_name "\n");      \
277         return show_##file_name(NULL, NULL, buf);                       \
278 }
279 show_one_old(sampling_rate);
280 show_one_old(up_threshold);
281 show_one_old(ignore_nice_load);
282 show_one_old(powersave_bias);
283 show_one_old(sampling_rate_min);
284 show_one_old(sampling_rate_max);
285
286 cpufreq_freq_attr_ro_old(sampling_rate_min);
287 cpufreq_freq_attr_ro_old(sampling_rate_max);
288
289 /*** delete after deprecation time ***/
290
291 static ssize_t store_sampling_rate(struct kobject *a, struct attribute *b,
292                                    const char *buf, size_t count)
293 {
294         unsigned int input;
295         int ret;
296         ret = sscanf(buf, "%u", &input);
297         if (ret != 1)
298                 return -EINVAL;
299
300         mutex_lock(&dbs_mutex);
301         dbs_tuners_ins.sampling_rate = max(input, min_sampling_rate);
302         mutex_unlock(&dbs_mutex);
303
304         return count;
305 }
306
307 static ssize_t store_io_is_busy(struct kobject *a, struct attribute *b,
308                                    const char *buf, size_t count)
309 {
310         unsigned int input;
311         int ret;
312
313         ret = sscanf(buf, "%u", &input);
314         if (ret != 1)
315                 return -EINVAL;
316
317         mutex_lock(&dbs_mutex);
318         dbs_tuners_ins.io_is_busy = !!input;
319         mutex_unlock(&dbs_mutex);
320
321         return count;
322 }
323
324 static ssize_t store_up_threshold(struct kobject *a, struct attribute *b,
325                                   const char *buf, size_t count)
326 {
327         unsigned int input;
328         int ret;
329         ret = sscanf(buf, "%u", &input);
330
331         if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
332                         input < MIN_FREQUENCY_UP_THRESHOLD) {
333                 return -EINVAL;
334         }
335
336         mutex_lock(&dbs_mutex);
337         dbs_tuners_ins.up_threshold = input;
338         mutex_unlock(&dbs_mutex);
339
340         return count;
341 }
342
343 static ssize_t store_ignore_nice_load(struct kobject *a, struct attribute *b,
344                                       const char *buf, size_t count)
345 {
346         unsigned int input;
347         int ret;
348
349         unsigned int j;
350
351         ret = sscanf(buf, "%u", &input);
352         if (ret != 1)
353                 return -EINVAL;
354
355         if (input > 1)
356                 input = 1;
357
358         mutex_lock(&dbs_mutex);
359         if (input == dbs_tuners_ins.ignore_nice) { /* nothing to do */
360                 mutex_unlock(&dbs_mutex);
361                 return count;
362         }
363         dbs_tuners_ins.ignore_nice = input;
364
365         /* we need to re-evaluate prev_cpu_idle */
366         for_each_online_cpu(j) {
367                 struct cpu_dbs_info_s *dbs_info;
368                 dbs_info = &per_cpu(od_cpu_dbs_info, j);
369                 dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
370                                                 &dbs_info->prev_cpu_wall);
371                 if (dbs_tuners_ins.ignore_nice)
372                         dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
373
374         }
375         mutex_unlock(&dbs_mutex);
376
377         return count;
378 }
379
380 static ssize_t store_powersave_bias(struct kobject *a, struct attribute *b,
381                                     const char *buf, size_t count)
382 {
383         unsigned int input;
384         int ret;
385         ret = sscanf(buf, "%u", &input);
386
387         if (ret != 1)
388                 return -EINVAL;
389
390         if (input > 1000)
391                 input = 1000;
392
393         mutex_lock(&dbs_mutex);
394         dbs_tuners_ins.powersave_bias = input;
395         ondemand_powersave_bias_init();
396         mutex_unlock(&dbs_mutex);
397
398         return count;
399 }
400
401 define_one_global_rw(sampling_rate);
402 define_one_global_rw(io_is_busy);
403 define_one_global_rw(up_threshold);
404 define_one_global_rw(ignore_nice_load);
405 define_one_global_rw(powersave_bias);
406
407 static struct attribute *dbs_attributes[] = {
408         &sampling_rate_max.attr,
409         &sampling_rate_min.attr,
410         &sampling_rate.attr,
411         &up_threshold.attr,
412         &ignore_nice_load.attr,
413         &powersave_bias.attr,
414         &io_is_busy.attr,
415         NULL
416 };
417
418 static struct attribute_group dbs_attr_group = {
419         .attrs = dbs_attributes,
420         .name = "ondemand",
421 };
422
423 /*** delete after deprecation time ***/
424
425 #define write_one_old(file_name)                                        \
426 static ssize_t store_##file_name##_old                                  \
427 (struct cpufreq_policy *unused, const char *buf, size_t count)          \
428 {                                                                       \
429        printk_once(KERN_INFO "CPUFREQ: Per core ondemand sysfs "        \
430                    "interface is deprecated - " #file_name "\n");       \
431        return store_##file_name(NULL, NULL, buf, count);                \
432 }
433 write_one_old(sampling_rate);
434 write_one_old(up_threshold);
435 write_one_old(ignore_nice_load);
436 write_one_old(powersave_bias);
437
438 cpufreq_freq_attr_rw_old(sampling_rate);
439 cpufreq_freq_attr_rw_old(up_threshold);
440 cpufreq_freq_attr_rw_old(ignore_nice_load);
441 cpufreq_freq_attr_rw_old(powersave_bias);
442
443 static struct attribute *dbs_attributes_old[] = {
444        &sampling_rate_max_old.attr,
445        &sampling_rate_min_old.attr,
446        &sampling_rate_old.attr,
447        &up_threshold_old.attr,
448        &ignore_nice_load_old.attr,
449        &powersave_bias_old.attr,
450        NULL
451 };
452
453 static struct attribute_group dbs_attr_group_old = {
454        .attrs = dbs_attributes_old,
455        .name = "ondemand",
456 };
457
458 /*** delete after deprecation time ***/
459
460 /************************** sysfs end ************************/
461
462 static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
463 {
464         unsigned int max_load_freq;
465
466         struct cpufreq_policy *policy;
467         unsigned int j;
468
469         this_dbs_info->freq_lo = 0;
470         policy = this_dbs_info->cur_policy;
471
472         /*
473          * Every sampling_rate, we check, if current idle time is less
474          * than 20% (default), then we try to increase frequency
475          * Every sampling_rate, we look for a the lowest
476          * frequency which can sustain the load while keeping idle time over
477          * 30%. If such a frequency exist, we try to decrease to this frequency.
478          *
479          * Any frequency increase takes it to the maximum frequency.
480          * Frequency reduction happens at minimum steps of
481          * 5% (default) of current frequency
482          */
483
484         /* Get Absolute Load - in terms of freq */
485         max_load_freq = 0;
486
487         for_each_cpu(j, policy->cpus) {
488                 struct cpu_dbs_info_s *j_dbs_info;
489                 cputime64_t cur_wall_time, cur_idle_time, cur_iowait_time;
490                 unsigned int idle_time, wall_time, iowait_time;
491                 unsigned int load, load_freq;
492                 int freq_avg;
493
494                 j_dbs_info = &per_cpu(od_cpu_dbs_info, j);
495
496                 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time);
497                 cur_iowait_time = get_cpu_iowait_time(j, &cur_wall_time);
498
499                 wall_time = (unsigned int) cputime64_sub(cur_wall_time,
500                                 j_dbs_info->prev_cpu_wall);
501                 j_dbs_info->prev_cpu_wall = cur_wall_time;
502
503                 idle_time = (unsigned int) cputime64_sub(cur_idle_time,
504                                 j_dbs_info->prev_cpu_idle);
505                 j_dbs_info->prev_cpu_idle = cur_idle_time;
506
507                 iowait_time = (unsigned int) cputime64_sub(cur_iowait_time,
508                                 j_dbs_info->prev_cpu_iowait);
509                 j_dbs_info->prev_cpu_iowait = cur_iowait_time;
510
511                 if (dbs_tuners_ins.ignore_nice) {
512                         cputime64_t cur_nice;
513                         unsigned long cur_nice_jiffies;
514
515                         cur_nice = cputime64_sub(kstat_cpu(j).cpustat.nice,
516                                          j_dbs_info->prev_cpu_nice);
517                         /*
518                          * Assumption: nice time between sampling periods will
519                          * be less than 2^32 jiffies for 32 bit sys
520                          */
521                         cur_nice_jiffies = (unsigned long)
522                                         cputime64_to_jiffies64(cur_nice);
523
524                         j_dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
525                         idle_time += jiffies_to_usecs(cur_nice_jiffies);
526                 }
527
528                 /*
529                  * For the purpose of ondemand, waiting for disk IO is an
530                  * indication that you're performance critical, and not that
531                  * the system is actually idle. So subtract the iowait time
532                  * from the cpu idle time.
533                  */
534
535                 if (dbs_tuners_ins.io_is_busy && idle_time >= iowait_time)
536                         idle_time -= iowait_time;
537
538                 if (unlikely(!wall_time || wall_time < idle_time))
539                         continue;
540
541                 load = 100 * (wall_time - idle_time) / wall_time;
542
543                 freq_avg = __cpufreq_driver_getavg(policy, j);
544                 if (freq_avg <= 0)
545                         freq_avg = policy->cur;
546
547                 load_freq = load * freq_avg;
548                 if (load_freq > max_load_freq)
549                         max_load_freq = load_freq;
550         }
551
552         /* Check for frequency increase */
553         if (max_load_freq > dbs_tuners_ins.up_threshold * policy->cur) {
554                 /* if we are already at full speed then break out early */
555                 if (!dbs_tuners_ins.powersave_bias) {
556                         if (policy->cur == policy->max)
557                                 return;
558
559                         __cpufreq_driver_target(policy, policy->max,
560                                 CPUFREQ_RELATION_H);
561                 } else {
562                         int freq = powersave_bias_target(policy, policy->max,
563                                         CPUFREQ_RELATION_H);
564                         __cpufreq_driver_target(policy, freq,
565                                 CPUFREQ_RELATION_L);
566                 }
567                 return;
568         }
569
570         /* Check for frequency decrease */
571         /* if we cannot reduce the frequency anymore, break out early */
572         if (policy->cur == policy->min)
573                 return;
574
575         /*
576          * The optimal frequency is the frequency that is the lowest that
577          * can support the current CPU usage without triggering the up
578          * policy. To be safe, we focus 10 points under the threshold.
579          */
580         if (max_load_freq <
581             (dbs_tuners_ins.up_threshold - dbs_tuners_ins.down_differential) *
582              policy->cur) {
583                 unsigned int freq_next;
584                 freq_next = max_load_freq /
585                                 (dbs_tuners_ins.up_threshold -
586                                  dbs_tuners_ins.down_differential);
587
588                 if (freq_next < policy->min)
589                         freq_next = policy->min;
590
591                 if (!dbs_tuners_ins.powersave_bias) {
592                         __cpufreq_driver_target(policy, freq_next,
593                                         CPUFREQ_RELATION_L);
594                 } else {
595                         int freq = powersave_bias_target(policy, freq_next,
596                                         CPUFREQ_RELATION_L);
597                         __cpufreq_driver_target(policy, freq,
598                                 CPUFREQ_RELATION_L);
599                 }
600         }
601 }
602
603 static void do_dbs_timer(struct work_struct *work)
604 {
605         struct cpu_dbs_info_s *dbs_info =
606                 container_of(work, struct cpu_dbs_info_s, work.work);
607         unsigned int cpu = dbs_info->cpu;
608         int sample_type = dbs_info->sample_type;
609
610         /* We want all CPUs to do sampling nearly on same jiffy */
611         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
612
613         delay -= jiffies % delay;
614         mutex_lock(&dbs_info->timer_mutex);
615
616         /* Common NORMAL_SAMPLE setup */
617         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
618         if (!dbs_tuners_ins.powersave_bias ||
619             sample_type == DBS_NORMAL_SAMPLE) {
620                 dbs_check_cpu(dbs_info);
621                 if (dbs_info->freq_lo) {
622                         /* Setup timer for SUB_SAMPLE */
623                         dbs_info->sample_type = DBS_SUB_SAMPLE;
624                         delay = dbs_info->freq_hi_jiffies;
625                 }
626         } else {
627                 __cpufreq_driver_target(dbs_info->cur_policy,
628                         dbs_info->freq_lo, CPUFREQ_RELATION_H);
629         }
630         queue_delayed_work_on(cpu, kondemand_wq, &dbs_info->work, delay);
631         mutex_unlock(&dbs_info->timer_mutex);
632 }
633
634 static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
635 {
636         /* We want all CPUs to do sampling nearly on same jiffy */
637         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
638         delay -= jiffies % delay;
639
640         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
641         INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer);
642         queue_delayed_work_on(dbs_info->cpu, kondemand_wq, &dbs_info->work,
643                 delay);
644 }
645
646 static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
647 {
648         cancel_delayed_work_sync(&dbs_info->work);
649 }
650
651 /*
652  * Not all CPUs want IO time to be accounted as busy; this dependson how
653  * efficient idling at a higher frequency/voltage is.
654  * Pavel Machek says this is not so for various generations of AMD and old
655  * Intel systems.
656  * Mike Chan (androidlcom) calis this is also not true for ARM.
657  * Because of this, whitelist specific known (series) of CPUs by default, and
658  * leave all others up to the user.
659  */
660 static int should_io_be_busy(void)
661 {
662 #if defined(CONFIG_X86)
663         /*
664          * For Intel, Core 2 (model 15) andl later have an efficient idle.
665          */
666         if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
667             boot_cpu_data.x86 == 6 &&
668             boot_cpu_data.x86_model >= 15)
669                 return 1;
670 #endif
671         return 0;
672 }
673
674 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
675                                    unsigned int event)
676 {
677         unsigned int cpu = policy->cpu;
678         struct cpu_dbs_info_s *this_dbs_info;
679         unsigned int j;
680         int rc;
681
682         this_dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
683
684         switch (event) {
685         case CPUFREQ_GOV_START:
686                 if ((!cpu_online(cpu)) || (!policy->cur))
687                         return -EINVAL;
688
689                 mutex_lock(&dbs_mutex);
690
691                 rc = sysfs_create_group(&policy->kobj, &dbs_attr_group_old);
692                 if (rc) {
693                         mutex_unlock(&dbs_mutex);
694                         return rc;
695                 }
696
697                 dbs_enable++;
698                 for_each_cpu(j, policy->cpus) {
699                         struct cpu_dbs_info_s *j_dbs_info;
700                         j_dbs_info = &per_cpu(od_cpu_dbs_info, j);
701                         j_dbs_info->cur_policy = policy;
702
703                         j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
704                                                 &j_dbs_info->prev_cpu_wall);
705                         if (dbs_tuners_ins.ignore_nice) {
706                                 j_dbs_info->prev_cpu_nice =
707                                                 kstat_cpu(j).cpustat.nice;
708                         }
709                 }
710                 this_dbs_info->cpu = cpu;
711                 ondemand_powersave_bias_init_cpu(cpu);
712                 /*
713                  * Start the timerschedule work, when this governor
714                  * is used for first time
715                  */
716                 if (dbs_enable == 1) {
717                         unsigned int latency;
718
719                         rc = sysfs_create_group(cpufreq_global_kobject,
720                                                 &dbs_attr_group);
721                         if (rc) {
722                                 mutex_unlock(&dbs_mutex);
723                                 return rc;
724                         }
725
726                         /* policy latency is in nS. Convert it to uS first */
727                         latency = policy->cpuinfo.transition_latency / 1000;
728                         if (latency == 0)
729                                 latency = 1;
730                         /* Bring kernel and HW constraints together */
731                         min_sampling_rate = max(min_sampling_rate,
732                                         MIN_LATENCY_MULTIPLIER * latency);
733                         dbs_tuners_ins.sampling_rate =
734                                 max(min_sampling_rate,
735                                     latency * LATENCY_MULTIPLIER);
736                         dbs_tuners_ins.io_is_busy = should_io_be_busy();
737                 }
738                 mutex_unlock(&dbs_mutex);
739
740                 mutex_init(&this_dbs_info->timer_mutex);
741                 dbs_timer_init(this_dbs_info);
742                 break;
743
744         case CPUFREQ_GOV_STOP:
745                 dbs_timer_exit(this_dbs_info);
746
747                 mutex_lock(&dbs_mutex);
748                 sysfs_remove_group(&policy->kobj, &dbs_attr_group_old);
749                 mutex_destroy(&this_dbs_info->timer_mutex);
750                 dbs_enable--;
751                 mutex_unlock(&dbs_mutex);
752                 if (!dbs_enable)
753                         sysfs_remove_group(cpufreq_global_kobject,
754                                            &dbs_attr_group);
755
756                 break;
757
758         case CPUFREQ_GOV_LIMITS:
759                 mutex_lock(&this_dbs_info->timer_mutex);
760                 if (policy->max < this_dbs_info->cur_policy->cur)
761                         __cpufreq_driver_target(this_dbs_info->cur_policy,
762                                 policy->max, CPUFREQ_RELATION_H);
763                 else if (policy->min > this_dbs_info->cur_policy->cur)
764                         __cpufreq_driver_target(this_dbs_info->cur_policy,
765                                 policy->min, CPUFREQ_RELATION_L);
766                 mutex_unlock(&this_dbs_info->timer_mutex);
767                 break;
768         }
769         return 0;
770 }
771
772 static int __init cpufreq_gov_dbs_init(void)
773 {
774         int err;
775         cputime64_t wall;
776         u64 idle_time;
777         int cpu = get_cpu();
778
779         idle_time = get_cpu_idle_time_us(cpu, &wall);
780         put_cpu();
781         if (idle_time != -1ULL) {
782                 /* Idle micro accounting is supported. Use finer thresholds */
783                 dbs_tuners_ins.up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
784                 dbs_tuners_ins.down_differential =
785                                         MICRO_FREQUENCY_DOWN_DIFFERENTIAL;
786                 /*
787                  * In no_hz/micro accounting case we set the minimum frequency
788                  * not depending on HZ, but fixed (very low). The deferred
789                  * timer might skip some samples if idle/sleeping as needed.
790                 */
791                 min_sampling_rate = MICRO_FREQUENCY_MIN_SAMPLE_RATE;
792         } else {
793                 /* For correct statistics, we need 10 ticks for each measure */
794                 min_sampling_rate =
795                         MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10);
796         }
797
798         kondemand_wq = create_workqueue("kondemand");
799         if (!kondemand_wq) {
800                 printk(KERN_ERR "Creation of kondemand failed\n");
801                 return -EFAULT;
802         }
803         err = cpufreq_register_governor(&cpufreq_gov_ondemand);
804         if (err)
805                 destroy_workqueue(kondemand_wq);
806
807         return err;
808 }
809
810 static void __exit cpufreq_gov_dbs_exit(void)
811 {
812         cpufreq_unregister_governor(&cpufreq_gov_ondemand);
813         destroy_workqueue(kondemand_wq);
814 }
815
816
817 MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
818 MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
819 MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
820         "Low Latency Frequency Transition capable processors");
821 MODULE_LICENSE("GPL");
822
823 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
824 fs_initcall(cpufreq_gov_dbs_init);
825 #else
826 module_init(cpufreq_gov_dbs_init);
827 #endif
828 module_exit(cpufreq_gov_dbs_exit);