Merge branch 'next' of git://git.kernel.org/pub/scm/linux/kernel/git/davej/cpufreq
[~shefty/rdma-dev.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 DEF_SAMPLING_DOWN_FACTOR                (1)
34 #define MAX_SAMPLING_DOWN_FACTOR                (100000)
35 #define MICRO_FREQUENCY_DOWN_DIFFERENTIAL       (3)
36 #define MICRO_FREQUENCY_UP_THRESHOLD            (95)
37 #define MICRO_FREQUENCY_MIN_SAMPLE_RATE         (10000)
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 #define MIN_SAMPLING_RATE_RATIO                 (2)
52
53 static unsigned int min_sampling_rate;
54
55 #define LATENCY_MULTIPLIER                      (1000)
56 #define MIN_LATENCY_MULTIPLIER                  (100)
57 #define TRANSITION_LATENCY_LIMIT                (10 * 1000 * 1000)
58
59 static void do_dbs_timer(struct work_struct *work);
60 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
61                                 unsigned int event);
62
63 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
64 static
65 #endif
66 struct cpufreq_governor cpufreq_gov_ondemand = {
67        .name                   = "ondemand",
68        .governor               = cpufreq_governor_dbs,
69        .max_transition_latency = TRANSITION_LATENCY_LIMIT,
70        .owner                  = THIS_MODULE,
71 };
72
73 /* Sampling types */
74 enum {DBS_NORMAL_SAMPLE, DBS_SUB_SAMPLE};
75
76 struct cpu_dbs_info_s {
77         cputime64_t prev_cpu_idle;
78         cputime64_t prev_cpu_iowait;
79         cputime64_t prev_cpu_wall;
80         cputime64_t prev_cpu_nice;
81         struct cpufreq_policy *cur_policy;
82         struct delayed_work work;
83         struct cpufreq_frequency_table *freq_table;
84         unsigned int freq_lo;
85         unsigned int freq_lo_jiffies;
86         unsigned int freq_hi_jiffies;
87         unsigned int rate_mult;
88         int cpu;
89         unsigned int sample_type:1;
90         /*
91          * percpu mutex that serializes governor limit change with
92          * do_dbs_timer invocation. We do not want do_dbs_timer to run
93          * when user is changing the governor or limits.
94          */
95         struct mutex timer_mutex;
96 };
97 static DEFINE_PER_CPU(struct cpu_dbs_info_s, od_cpu_dbs_info);
98
99 static unsigned int dbs_enable; /* number of CPUs using this policy */
100
101 /*
102  * dbs_mutex protects dbs_enable in governor start/stop.
103  */
104 static DEFINE_MUTEX(dbs_mutex);
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 sampling_down_factor;
112         unsigned int powersave_bias;
113         unsigned int io_is_busy;
114 } dbs_tuners_ins = {
115         .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
116         .sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR,
117         .down_differential = DEF_FREQUENCY_DOWN_DIFFERENTIAL,
118         .ignore_nice = 0,
119         .powersave_bias = 0,
120 };
121
122 static inline u64 get_cpu_idle_time_jiffy(unsigned int cpu, u64 *wall)
123 {
124         u64 idle_time;
125         u64 cur_wall_time;
126         u64 busy_time;
127
128         cur_wall_time = jiffies64_to_cputime64(get_jiffies_64());
129
130         busy_time  = kcpustat_cpu(cpu).cpustat[CPUTIME_USER];
131         busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_SYSTEM];
132         busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_IRQ];
133         busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_SOFTIRQ];
134         busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_STEAL];
135         busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_NICE];
136
137         idle_time = cur_wall_time - busy_time;
138         if (wall)
139                 *wall = jiffies_to_usecs(cur_wall_time);
140
141         return jiffies_to_usecs(idle_time);
142 }
143
144 static inline cputime64_t get_cpu_idle_time(unsigned int cpu, cputime64_t *wall)
145 {
146         u64 idle_time = get_cpu_idle_time_us(cpu, NULL);
147
148         if (idle_time == -1ULL)
149                 return get_cpu_idle_time_jiffy(cpu, wall);
150         else
151                 idle_time += get_cpu_iowait_time_us(cpu, wall);
152
153         return idle_time;
154 }
155
156 static inline cputime64_t get_cpu_iowait_time(unsigned int cpu, cputime64_t *wall)
157 {
158         u64 iowait_time = get_cpu_iowait_time_us(cpu, wall);
159
160         if (iowait_time == -1ULL)
161                 return 0;
162
163         return iowait_time;
164 }
165
166 /*
167  * Find right freq to be set now with powersave_bias on.
168  * Returns the freq_hi to be used right now and will set freq_hi_jiffies,
169  * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs.
170  */
171 static unsigned int powersave_bias_target(struct cpufreq_policy *policy,
172                                           unsigned int freq_next,
173                                           unsigned int relation)
174 {
175         unsigned int freq_req, freq_reduc, freq_avg;
176         unsigned int freq_hi, freq_lo;
177         unsigned int index = 0;
178         unsigned int jiffies_total, jiffies_hi, jiffies_lo;
179         struct cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
180                                                    policy->cpu);
181
182         if (!dbs_info->freq_table) {
183                 dbs_info->freq_lo = 0;
184                 dbs_info->freq_lo_jiffies = 0;
185                 return freq_next;
186         }
187
188         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next,
189                         relation, &index);
190         freq_req = dbs_info->freq_table[index].frequency;
191         freq_reduc = freq_req * dbs_tuners_ins.powersave_bias / 1000;
192         freq_avg = freq_req - freq_reduc;
193
194         /* Find freq bounds for freq_avg in freq_table */
195         index = 0;
196         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
197                         CPUFREQ_RELATION_H, &index);
198         freq_lo = dbs_info->freq_table[index].frequency;
199         index = 0;
200         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
201                         CPUFREQ_RELATION_L, &index);
202         freq_hi = dbs_info->freq_table[index].frequency;
203
204         /* Find out how long we have to be in hi and lo freqs */
205         if (freq_hi == freq_lo) {
206                 dbs_info->freq_lo = 0;
207                 dbs_info->freq_lo_jiffies = 0;
208                 return freq_lo;
209         }
210         jiffies_total = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
211         jiffies_hi = (freq_avg - freq_lo) * jiffies_total;
212         jiffies_hi += ((freq_hi - freq_lo) / 2);
213         jiffies_hi /= (freq_hi - freq_lo);
214         jiffies_lo = jiffies_total - jiffies_hi;
215         dbs_info->freq_lo = freq_lo;
216         dbs_info->freq_lo_jiffies = jiffies_lo;
217         dbs_info->freq_hi_jiffies = jiffies_hi;
218         return freq_hi;
219 }
220
221 static void ondemand_powersave_bias_init_cpu(int cpu)
222 {
223         struct cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
224         dbs_info->freq_table = cpufreq_frequency_get_table(cpu);
225         dbs_info->freq_lo = 0;
226 }
227
228 static void ondemand_powersave_bias_init(void)
229 {
230         int i;
231         for_each_online_cpu(i) {
232                 ondemand_powersave_bias_init_cpu(i);
233         }
234 }
235
236 /************************** sysfs interface ************************/
237
238 static ssize_t show_sampling_rate_min(struct kobject *kobj,
239                                       struct attribute *attr, char *buf)
240 {
241         return sprintf(buf, "%u\n", min_sampling_rate);
242 }
243
244 define_one_global_ro(sampling_rate_min);
245
246 /* cpufreq_ondemand Governor Tunables */
247 #define show_one(file_name, object)                                     \
248 static ssize_t show_##file_name                                         \
249 (struct kobject *kobj, struct attribute *attr, char *buf)              \
250 {                                                                       \
251         return sprintf(buf, "%u\n", dbs_tuners_ins.object);             \
252 }
253 show_one(sampling_rate, sampling_rate);
254 show_one(io_is_busy, io_is_busy);
255 show_one(up_threshold, up_threshold);
256 show_one(sampling_down_factor, sampling_down_factor);
257 show_one(ignore_nice_load, ignore_nice);
258 show_one(powersave_bias, powersave_bias);
259
260 static ssize_t store_sampling_rate(struct kobject *a, struct attribute *b,
261                                    const char *buf, size_t count)
262 {
263         unsigned int input;
264         int ret;
265         ret = sscanf(buf, "%u", &input);
266         if (ret != 1)
267                 return -EINVAL;
268         dbs_tuners_ins.sampling_rate = max(input, min_sampling_rate);
269         return count;
270 }
271
272 static ssize_t store_io_is_busy(struct kobject *a, struct attribute *b,
273                                    const char *buf, size_t count)
274 {
275         unsigned int input;
276         int ret;
277
278         ret = sscanf(buf, "%u", &input);
279         if (ret != 1)
280                 return -EINVAL;
281         dbs_tuners_ins.io_is_busy = !!input;
282         return count;
283 }
284
285 static ssize_t store_up_threshold(struct kobject *a, struct attribute *b,
286                                   const char *buf, size_t count)
287 {
288         unsigned int input;
289         int ret;
290         ret = sscanf(buf, "%u", &input);
291
292         if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
293                         input < MIN_FREQUENCY_UP_THRESHOLD) {
294                 return -EINVAL;
295         }
296         dbs_tuners_ins.up_threshold = input;
297         return count;
298 }
299
300 static ssize_t store_sampling_down_factor(struct kobject *a,
301                         struct attribute *b, const char *buf, size_t count)
302 {
303         unsigned int input, j;
304         int ret;
305         ret = sscanf(buf, "%u", &input);
306
307         if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
308                 return -EINVAL;
309         dbs_tuners_ins.sampling_down_factor = input;
310
311         /* Reset down sampling multiplier in case it was active */
312         for_each_online_cpu(j) {
313                 struct cpu_dbs_info_s *dbs_info;
314                 dbs_info = &per_cpu(od_cpu_dbs_info, j);
315                 dbs_info->rate_mult = 1;
316         }
317         return count;
318 }
319
320 static ssize_t store_ignore_nice_load(struct kobject *a, struct attribute *b,
321                                       const char *buf, size_t count)
322 {
323         unsigned int input;
324         int ret;
325
326         unsigned int j;
327
328         ret = sscanf(buf, "%u", &input);
329         if (ret != 1)
330                 return -EINVAL;
331
332         if (input > 1)
333                 input = 1;
334
335         if (input == dbs_tuners_ins.ignore_nice) { /* nothing to do */
336                 return count;
337         }
338         dbs_tuners_ins.ignore_nice = input;
339
340         /* we need to re-evaluate prev_cpu_idle */
341         for_each_online_cpu(j) {
342                 struct cpu_dbs_info_s *dbs_info;
343                 dbs_info = &per_cpu(od_cpu_dbs_info, j);
344                 dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
345                                                 &dbs_info->prev_cpu_wall);
346                 if (dbs_tuners_ins.ignore_nice)
347                         dbs_info->prev_cpu_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE];
348
349         }
350         return count;
351 }
352
353 static ssize_t store_powersave_bias(struct kobject *a, struct attribute *b,
354                                     const char *buf, size_t count)
355 {
356         unsigned int input;
357         int ret;
358         ret = sscanf(buf, "%u", &input);
359
360         if (ret != 1)
361                 return -EINVAL;
362
363         if (input > 1000)
364                 input = 1000;
365
366         dbs_tuners_ins.powersave_bias = input;
367         ondemand_powersave_bias_init();
368         return count;
369 }
370
371 define_one_global_rw(sampling_rate);
372 define_one_global_rw(io_is_busy);
373 define_one_global_rw(up_threshold);
374 define_one_global_rw(sampling_down_factor);
375 define_one_global_rw(ignore_nice_load);
376 define_one_global_rw(powersave_bias);
377
378 static struct attribute *dbs_attributes[] = {
379         &sampling_rate_min.attr,
380         &sampling_rate.attr,
381         &up_threshold.attr,
382         &sampling_down_factor.attr,
383         &ignore_nice_load.attr,
384         &powersave_bias.attr,
385         &io_is_busy.attr,
386         NULL
387 };
388
389 static struct attribute_group dbs_attr_group = {
390         .attrs = dbs_attributes,
391         .name = "ondemand",
392 };
393
394 /************************** sysfs end ************************/
395
396 static void dbs_freq_increase(struct cpufreq_policy *p, unsigned int freq)
397 {
398         if (dbs_tuners_ins.powersave_bias)
399                 freq = powersave_bias_target(p, freq, CPUFREQ_RELATION_H);
400         else if (p->cur == p->max)
401                 return;
402
403         __cpufreq_driver_target(p, freq, dbs_tuners_ins.powersave_bias ?
404                         CPUFREQ_RELATION_L : CPUFREQ_RELATION_H);
405 }
406
407 static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
408 {
409         unsigned int max_load_freq;
410
411         struct cpufreq_policy *policy;
412         unsigned int j;
413
414         this_dbs_info->freq_lo = 0;
415         policy = this_dbs_info->cur_policy;
416
417         /*
418          * Every sampling_rate, we check, if current idle time is less
419          * than 20% (default), then we try to increase frequency
420          * Every sampling_rate, we look for a the lowest
421          * frequency which can sustain the load while keeping idle time over
422          * 30%. If such a frequency exist, we try to decrease to this frequency.
423          *
424          * Any frequency increase takes it to the maximum frequency.
425          * Frequency reduction happens at minimum steps of
426          * 5% (default) of current frequency
427          */
428
429         /* Get Absolute Load - in terms of freq */
430         max_load_freq = 0;
431
432         for_each_cpu(j, policy->cpus) {
433                 struct cpu_dbs_info_s *j_dbs_info;
434                 cputime64_t cur_wall_time, cur_idle_time, cur_iowait_time;
435                 unsigned int idle_time, wall_time, iowait_time;
436                 unsigned int load, load_freq;
437                 int freq_avg;
438
439                 j_dbs_info = &per_cpu(od_cpu_dbs_info, j);
440
441                 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time);
442                 cur_iowait_time = get_cpu_iowait_time(j, &cur_wall_time);
443
444                 wall_time = (unsigned int)
445                         (cur_wall_time - j_dbs_info->prev_cpu_wall);
446                 j_dbs_info->prev_cpu_wall = cur_wall_time;
447
448                 idle_time = (unsigned int)
449                         (cur_idle_time - j_dbs_info->prev_cpu_idle);
450                 j_dbs_info->prev_cpu_idle = cur_idle_time;
451
452                 iowait_time = (unsigned int)
453                         (cur_iowait_time - j_dbs_info->prev_cpu_iowait);
454                 j_dbs_info->prev_cpu_iowait = cur_iowait_time;
455
456                 if (dbs_tuners_ins.ignore_nice) {
457                         u64 cur_nice;
458                         unsigned long cur_nice_jiffies;
459
460                         cur_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE] -
461                                          j_dbs_info->prev_cpu_nice;
462                         /*
463                          * Assumption: nice time between sampling periods will
464                          * be less than 2^32 jiffies for 32 bit sys
465                          */
466                         cur_nice_jiffies = (unsigned long)
467                                         cputime64_to_jiffies64(cur_nice);
468
469                         j_dbs_info->prev_cpu_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE];
470                         idle_time += jiffies_to_usecs(cur_nice_jiffies);
471                 }
472
473                 /*
474                  * For the purpose of ondemand, waiting for disk IO is an
475                  * indication that you're performance critical, and not that
476                  * the system is actually idle. So subtract the iowait time
477                  * from the cpu idle time.
478                  */
479
480                 if (dbs_tuners_ins.io_is_busy && idle_time >= iowait_time)
481                         idle_time -= iowait_time;
482
483                 if (unlikely(!wall_time || wall_time < idle_time))
484                         continue;
485
486                 load = 100 * (wall_time - idle_time) / wall_time;
487
488                 freq_avg = __cpufreq_driver_getavg(policy, j);
489                 if (freq_avg <= 0)
490                         freq_avg = policy->cur;
491
492                 load_freq = load * freq_avg;
493                 if (load_freq > max_load_freq)
494                         max_load_freq = load_freq;
495         }
496
497         /* Check for frequency increase */
498         if (max_load_freq > dbs_tuners_ins.up_threshold * policy->cur) {
499                 /* If switching to max speed, apply sampling_down_factor */
500                 if (policy->cur < policy->max)
501                         this_dbs_info->rate_mult =
502                                 dbs_tuners_ins.sampling_down_factor;
503                 dbs_freq_increase(policy, policy->max);
504                 return;
505         }
506
507         /* Check for frequency decrease */
508         /* if we cannot reduce the frequency anymore, break out early */
509         if (policy->cur == policy->min)
510                 return;
511
512         /*
513          * The optimal frequency is the frequency that is the lowest that
514          * can support the current CPU usage without triggering the up
515          * policy. To be safe, we focus 10 points under the threshold.
516          */
517         if (max_load_freq <
518             (dbs_tuners_ins.up_threshold - dbs_tuners_ins.down_differential) *
519              policy->cur) {
520                 unsigned int freq_next;
521                 freq_next = max_load_freq /
522                                 (dbs_tuners_ins.up_threshold -
523                                  dbs_tuners_ins.down_differential);
524
525                 /* No longer fully busy, reset rate_mult */
526                 this_dbs_info->rate_mult = 1;
527
528                 if (freq_next < policy->min)
529                         freq_next = policy->min;
530
531                 if (!dbs_tuners_ins.powersave_bias) {
532                         __cpufreq_driver_target(policy, freq_next,
533                                         CPUFREQ_RELATION_L);
534                 } else {
535                         int freq = powersave_bias_target(policy, freq_next,
536                                         CPUFREQ_RELATION_L);
537                         __cpufreq_driver_target(policy, freq,
538                                 CPUFREQ_RELATION_L);
539                 }
540         }
541 }
542
543 static void do_dbs_timer(struct work_struct *work)
544 {
545         struct cpu_dbs_info_s *dbs_info =
546                 container_of(work, struct cpu_dbs_info_s, work.work);
547         unsigned int cpu = dbs_info->cpu;
548         int sample_type = dbs_info->sample_type;
549
550         int delay;
551
552         mutex_lock(&dbs_info->timer_mutex);
553
554         /* Common NORMAL_SAMPLE setup */
555         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
556         if (!dbs_tuners_ins.powersave_bias ||
557             sample_type == DBS_NORMAL_SAMPLE) {
558                 dbs_check_cpu(dbs_info);
559                 if (dbs_info->freq_lo) {
560                         /* Setup timer for SUB_SAMPLE */
561                         dbs_info->sample_type = DBS_SUB_SAMPLE;
562                         delay = dbs_info->freq_hi_jiffies;
563                 } else {
564                         /* We want all CPUs to do sampling nearly on
565                          * same jiffy
566                          */
567                         delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate
568                                 * dbs_info->rate_mult);
569
570                         if (num_online_cpus() > 1)
571                                 delay -= jiffies % delay;
572                 }
573         } else {
574                 __cpufreq_driver_target(dbs_info->cur_policy,
575                         dbs_info->freq_lo, CPUFREQ_RELATION_H);
576                 delay = dbs_info->freq_lo_jiffies;
577         }
578         schedule_delayed_work_on(cpu, &dbs_info->work, delay);
579         mutex_unlock(&dbs_info->timer_mutex);
580 }
581
582 static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
583 {
584         /* We want all CPUs to do sampling nearly on same jiffy */
585         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
586
587         if (num_online_cpus() > 1)
588                 delay -= jiffies % delay;
589
590         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
591         INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer);
592         schedule_delayed_work_on(dbs_info->cpu, &dbs_info->work, delay);
593 }
594
595 static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
596 {
597         cancel_delayed_work_sync(&dbs_info->work);
598 }
599
600 /*
601  * Not all CPUs want IO time to be accounted as busy; this dependson how
602  * efficient idling at a higher frequency/voltage is.
603  * Pavel Machek says this is not so for various generations of AMD and old
604  * Intel systems.
605  * Mike Chan (androidlcom) calis this is also not true for ARM.
606  * Because of this, whitelist specific known (series) of CPUs by default, and
607  * leave all others up to the user.
608  */
609 static int should_io_be_busy(void)
610 {
611 #if defined(CONFIG_X86)
612         /*
613          * For Intel, Core 2 (model 15) andl later have an efficient idle.
614          */
615         if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
616             boot_cpu_data.x86 == 6 &&
617             boot_cpu_data.x86_model >= 15)
618                 return 1;
619 #endif
620         return 0;
621 }
622
623 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
624                                    unsigned int event)
625 {
626         unsigned int cpu = policy->cpu;
627         struct cpu_dbs_info_s *this_dbs_info;
628         unsigned int j;
629         int rc;
630
631         this_dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
632
633         switch (event) {
634         case CPUFREQ_GOV_START:
635                 if ((!cpu_online(cpu)) || (!policy->cur))
636                         return -EINVAL;
637
638                 mutex_lock(&dbs_mutex);
639
640                 dbs_enable++;
641                 for_each_cpu(j, policy->cpus) {
642                         struct cpu_dbs_info_s *j_dbs_info;
643                         j_dbs_info = &per_cpu(od_cpu_dbs_info, j);
644                         j_dbs_info->cur_policy = policy;
645
646                         j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
647                                                 &j_dbs_info->prev_cpu_wall);
648                         if (dbs_tuners_ins.ignore_nice)
649                                 j_dbs_info->prev_cpu_nice =
650                                                 kcpustat_cpu(j).cpustat[CPUTIME_NICE];
651                 }
652                 this_dbs_info->cpu = cpu;
653                 this_dbs_info->rate_mult = 1;
654                 ondemand_powersave_bias_init_cpu(cpu);
655                 /*
656                  * Start the timerschedule work, when this governor
657                  * is used for first time
658                  */
659                 if (dbs_enable == 1) {
660                         unsigned int latency;
661
662                         rc = sysfs_create_group(cpufreq_global_kobject,
663                                                 &dbs_attr_group);
664                         if (rc) {
665                                 mutex_unlock(&dbs_mutex);
666                                 return rc;
667                         }
668
669                         /* policy latency is in nS. Convert it to uS first */
670                         latency = policy->cpuinfo.transition_latency / 1000;
671                         if (latency == 0)
672                                 latency = 1;
673                         /* Bring kernel and HW constraints together */
674                         min_sampling_rate = max(min_sampling_rate,
675                                         MIN_LATENCY_MULTIPLIER * latency);
676                         dbs_tuners_ins.sampling_rate =
677                                 max(min_sampling_rate,
678                                     latency * LATENCY_MULTIPLIER);
679                         dbs_tuners_ins.io_is_busy = should_io_be_busy();
680                 }
681                 mutex_unlock(&dbs_mutex);
682
683                 mutex_init(&this_dbs_info->timer_mutex);
684                 dbs_timer_init(this_dbs_info);
685                 break;
686
687         case CPUFREQ_GOV_STOP:
688                 dbs_timer_exit(this_dbs_info);
689
690                 mutex_lock(&dbs_mutex);
691                 mutex_destroy(&this_dbs_info->timer_mutex);
692                 dbs_enable--;
693                 mutex_unlock(&dbs_mutex);
694                 if (!dbs_enable)
695                         sysfs_remove_group(cpufreq_global_kobject,
696                                            &dbs_attr_group);
697
698                 break;
699
700         case CPUFREQ_GOV_LIMITS:
701                 mutex_lock(&this_dbs_info->timer_mutex);
702                 if (policy->max < this_dbs_info->cur_policy->cur)
703                         __cpufreq_driver_target(this_dbs_info->cur_policy,
704                                 policy->max, CPUFREQ_RELATION_H);
705                 else if (policy->min > this_dbs_info->cur_policy->cur)
706                         __cpufreq_driver_target(this_dbs_info->cur_policy,
707                                 policy->min, CPUFREQ_RELATION_L);
708                 mutex_unlock(&this_dbs_info->timer_mutex);
709                 break;
710         }
711         return 0;
712 }
713
714 static int __init cpufreq_gov_dbs_init(void)
715 {
716         u64 idle_time;
717         int cpu = get_cpu();
718
719         idle_time = get_cpu_idle_time_us(cpu, NULL);
720         put_cpu();
721         if (idle_time != -1ULL) {
722                 /* Idle micro accounting is supported. Use finer thresholds */
723                 dbs_tuners_ins.up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
724                 dbs_tuners_ins.down_differential =
725                                         MICRO_FREQUENCY_DOWN_DIFFERENTIAL;
726                 /*
727                  * In nohz/micro accounting case we set the minimum frequency
728                  * not depending on HZ, but fixed (very low). The deferred
729                  * timer might skip some samples if idle/sleeping as needed.
730                 */
731                 min_sampling_rate = MICRO_FREQUENCY_MIN_SAMPLE_RATE;
732         } else {
733                 /* For correct statistics, we need 10 ticks for each measure */
734                 min_sampling_rate =
735                         MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10);
736         }
737
738         return cpufreq_register_governor(&cpufreq_gov_ondemand);
739 }
740
741 static void __exit cpufreq_gov_dbs_exit(void)
742 {
743         cpufreq_unregister_governor(&cpufreq_gov_ondemand);
744 }
745
746
747 MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
748 MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
749 MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
750         "Low Latency Frequency Transition capable processors");
751 MODULE_LICENSE("GPL");
752
753 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
754 fs_initcall(cpufreq_gov_dbs_init);
755 #else
756 module_init(cpufreq_gov_dbs_init);
757 #endif
758 module_exit(cpufreq_gov_dbs_exit);