Merge branch 'x86-cleanups-for-linus' of git://git.kernel.org/pub/scm/linux/kernel...
[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 define_one_ro(_name)            \
250 static struct global_attr _name =       \
251 __ATTR(_name, 0444, show_##_name, NULL)
252
253 define_one_ro(sampling_rate_max);
254 define_one_ro(sampling_rate_min);
255
256 /* cpufreq_ondemand Governor Tunables */
257 #define show_one(file_name, object)                                     \
258 static ssize_t show_##file_name                                         \
259 (struct kobject *kobj, struct attribute *attr, char *buf)              \
260 {                                                                       \
261         return sprintf(buf, "%u\n", dbs_tuners_ins.object);             \
262 }
263 show_one(sampling_rate, sampling_rate);
264 show_one(io_is_busy, io_is_busy);
265 show_one(up_threshold, up_threshold);
266 show_one(ignore_nice_load, ignore_nice);
267 show_one(powersave_bias, powersave_bias);
268
269 /*** delete after deprecation time ***/
270
271 #define DEPRECATION_MSG(file_name)                                      \
272         printk_once(KERN_INFO "CPUFREQ: Per core ondemand sysfs "       \
273                     "interface is deprecated - " #file_name "\n");
274
275 #define show_one_old(file_name)                                         \
276 static ssize_t show_##file_name##_old                                   \
277 (struct cpufreq_policy *unused, char *buf)                              \
278 {                                                                       \
279         printk_once(KERN_INFO "CPUFREQ: Per core ondemand sysfs "       \
280                     "interface is deprecated - " #file_name "\n");      \
281         return show_##file_name(NULL, NULL, buf);                       \
282 }
283 show_one_old(sampling_rate);
284 show_one_old(up_threshold);
285 show_one_old(ignore_nice_load);
286 show_one_old(powersave_bias);
287 show_one_old(sampling_rate_min);
288 show_one_old(sampling_rate_max);
289
290 #define define_one_ro_old(object, _name)       \
291 static struct freq_attr object =               \
292 __ATTR(_name, 0444, show_##_name##_old, NULL)
293
294 define_one_ro_old(sampling_rate_min_old, sampling_rate_min);
295 define_one_ro_old(sampling_rate_max_old, sampling_rate_max);
296
297 /*** delete after deprecation time ***/
298
299 static ssize_t store_sampling_rate(struct kobject *a, struct attribute *b,
300                                    const char *buf, size_t count)
301 {
302         unsigned int input;
303         int ret;
304         ret = sscanf(buf, "%u", &input);
305         if (ret != 1)
306                 return -EINVAL;
307
308         mutex_lock(&dbs_mutex);
309         dbs_tuners_ins.sampling_rate = max(input, min_sampling_rate);
310         mutex_unlock(&dbs_mutex);
311
312         return count;
313 }
314
315 static ssize_t store_io_is_busy(struct kobject *a, struct attribute *b,
316                                    const char *buf, size_t count)
317 {
318         unsigned int input;
319         int ret;
320
321         ret = sscanf(buf, "%u", &input);
322         if (ret != 1)
323                 return -EINVAL;
324
325         mutex_lock(&dbs_mutex);
326         dbs_tuners_ins.io_is_busy = !!input;
327         mutex_unlock(&dbs_mutex);
328
329         return count;
330 }
331
332 static ssize_t store_up_threshold(struct kobject *a, struct attribute *b,
333                                   const char *buf, size_t count)
334 {
335         unsigned int input;
336         int ret;
337         ret = sscanf(buf, "%u", &input);
338
339         if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
340                         input < MIN_FREQUENCY_UP_THRESHOLD) {
341                 return -EINVAL;
342         }
343
344         mutex_lock(&dbs_mutex);
345         dbs_tuners_ins.up_threshold = input;
346         mutex_unlock(&dbs_mutex);
347
348         return count;
349 }
350
351 static ssize_t store_ignore_nice_load(struct kobject *a, struct attribute *b,
352                                       const char *buf, size_t count)
353 {
354         unsigned int input;
355         int ret;
356
357         unsigned int j;
358
359         ret = sscanf(buf, "%u", &input);
360         if (ret != 1)
361                 return -EINVAL;
362
363         if (input > 1)
364                 input = 1;
365
366         mutex_lock(&dbs_mutex);
367         if (input == dbs_tuners_ins.ignore_nice) { /* nothing to do */
368                 mutex_unlock(&dbs_mutex);
369                 return count;
370         }
371         dbs_tuners_ins.ignore_nice = input;
372
373         /* we need to re-evaluate prev_cpu_idle */
374         for_each_online_cpu(j) {
375                 struct cpu_dbs_info_s *dbs_info;
376                 dbs_info = &per_cpu(od_cpu_dbs_info, j);
377                 dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
378                                                 &dbs_info->prev_cpu_wall);
379                 if (dbs_tuners_ins.ignore_nice)
380                         dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
381
382         }
383         mutex_unlock(&dbs_mutex);
384
385         return count;
386 }
387
388 static ssize_t store_powersave_bias(struct kobject *a, struct attribute *b,
389                                     const char *buf, size_t count)
390 {
391         unsigned int input;
392         int ret;
393         ret = sscanf(buf, "%u", &input);
394
395         if (ret != 1)
396                 return -EINVAL;
397
398         if (input > 1000)
399                 input = 1000;
400
401         mutex_lock(&dbs_mutex);
402         dbs_tuners_ins.powersave_bias = input;
403         ondemand_powersave_bias_init();
404         mutex_unlock(&dbs_mutex);
405
406         return count;
407 }
408
409 #define define_one_rw(_name) \
410 static struct global_attr _name = \
411 __ATTR(_name, 0644, show_##_name, store_##_name)
412
413 define_one_rw(sampling_rate);
414 define_one_rw(io_is_busy);
415 define_one_rw(up_threshold);
416 define_one_rw(ignore_nice_load);
417 define_one_rw(powersave_bias);
418
419 static struct attribute *dbs_attributes[] = {
420         &sampling_rate_max.attr,
421         &sampling_rate_min.attr,
422         &sampling_rate.attr,
423         &up_threshold.attr,
424         &ignore_nice_load.attr,
425         &powersave_bias.attr,
426         &io_is_busy.attr,
427         NULL
428 };
429
430 static struct attribute_group dbs_attr_group = {
431         .attrs = dbs_attributes,
432         .name = "ondemand",
433 };
434
435 /*** delete after deprecation time ***/
436
437 #define write_one_old(file_name)                                        \
438 static ssize_t store_##file_name##_old                                  \
439 (struct cpufreq_policy *unused, const char *buf, size_t count)          \
440 {                                                                       \
441        printk_once(KERN_INFO "CPUFREQ: Per core ondemand sysfs "        \
442                    "interface is deprecated - " #file_name "\n");       \
443        return store_##file_name(NULL, NULL, buf, count);                \
444 }
445 write_one_old(sampling_rate);
446 write_one_old(up_threshold);
447 write_one_old(ignore_nice_load);
448 write_one_old(powersave_bias);
449
450 #define define_one_rw_old(object, _name)       \
451 static struct freq_attr object =               \
452 __ATTR(_name, 0644, show_##_name##_old, store_##_name##_old)
453
454 define_one_rw_old(sampling_rate_old, sampling_rate);
455 define_one_rw_old(up_threshold_old, up_threshold);
456 define_one_rw_old(ignore_nice_load_old, ignore_nice_load);
457 define_one_rw_old(powersave_bias_old, powersave_bias);
458
459 static struct attribute *dbs_attributes_old[] = {
460        &sampling_rate_max_old.attr,
461        &sampling_rate_min_old.attr,
462        &sampling_rate_old.attr,
463        &up_threshold_old.attr,
464        &ignore_nice_load_old.attr,
465        &powersave_bias_old.attr,
466        NULL
467 };
468
469 static struct attribute_group dbs_attr_group_old = {
470        .attrs = dbs_attributes_old,
471        .name = "ondemand",
472 };
473
474 /*** delete after deprecation time ***/
475
476 /************************** sysfs end ************************/
477
478 static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
479 {
480         unsigned int max_load_freq;
481
482         struct cpufreq_policy *policy;
483         unsigned int j;
484
485         this_dbs_info->freq_lo = 0;
486         policy = this_dbs_info->cur_policy;
487
488         /*
489          * Every sampling_rate, we check, if current idle time is less
490          * than 20% (default), then we try to increase frequency
491          * Every sampling_rate, we look for a the lowest
492          * frequency which can sustain the load while keeping idle time over
493          * 30%. If such a frequency exist, we try to decrease to this frequency.
494          *
495          * Any frequency increase takes it to the maximum frequency.
496          * Frequency reduction happens at minimum steps of
497          * 5% (default) of current frequency
498          */
499
500         /* Get Absolute Load - in terms of freq */
501         max_load_freq = 0;
502
503         for_each_cpu(j, policy->cpus) {
504                 struct cpu_dbs_info_s *j_dbs_info;
505                 cputime64_t cur_wall_time, cur_idle_time, cur_iowait_time;
506                 unsigned int idle_time, wall_time, iowait_time;
507                 unsigned int load, load_freq;
508                 int freq_avg;
509
510                 j_dbs_info = &per_cpu(od_cpu_dbs_info, j);
511
512                 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time);
513                 cur_iowait_time = get_cpu_iowait_time(j, &cur_wall_time);
514
515                 wall_time = (unsigned int) cputime64_sub(cur_wall_time,
516                                 j_dbs_info->prev_cpu_wall);
517                 j_dbs_info->prev_cpu_wall = cur_wall_time;
518
519                 idle_time = (unsigned int) cputime64_sub(cur_idle_time,
520                                 j_dbs_info->prev_cpu_idle);
521                 j_dbs_info->prev_cpu_idle = cur_idle_time;
522
523                 iowait_time = (unsigned int) cputime64_sub(cur_iowait_time,
524                                 j_dbs_info->prev_cpu_iowait);
525                 j_dbs_info->prev_cpu_iowait = cur_iowait_time;
526
527                 if (dbs_tuners_ins.ignore_nice) {
528                         cputime64_t cur_nice;
529                         unsigned long cur_nice_jiffies;
530
531                         cur_nice = cputime64_sub(kstat_cpu(j).cpustat.nice,
532                                          j_dbs_info->prev_cpu_nice);
533                         /*
534                          * Assumption: nice time between sampling periods will
535                          * be less than 2^32 jiffies for 32 bit sys
536                          */
537                         cur_nice_jiffies = (unsigned long)
538                                         cputime64_to_jiffies64(cur_nice);
539
540                         j_dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
541                         idle_time += jiffies_to_usecs(cur_nice_jiffies);
542                 }
543
544                 /*
545                  * For the purpose of ondemand, waiting for disk IO is an
546                  * indication that you're performance critical, and not that
547                  * the system is actually idle. So subtract the iowait time
548                  * from the cpu idle time.
549                  */
550
551                 if (dbs_tuners_ins.io_is_busy && idle_time >= iowait_time)
552                         idle_time -= iowait_time;
553
554                 if (unlikely(!wall_time || wall_time < idle_time))
555                         continue;
556
557                 load = 100 * (wall_time - idle_time) / wall_time;
558
559                 freq_avg = __cpufreq_driver_getavg(policy, j);
560                 if (freq_avg <= 0)
561                         freq_avg = policy->cur;
562
563                 load_freq = load * freq_avg;
564                 if (load_freq > max_load_freq)
565                         max_load_freq = load_freq;
566         }
567
568         /* Check for frequency increase */
569         if (max_load_freq > dbs_tuners_ins.up_threshold * policy->cur) {
570                 /* if we are already at full speed then break out early */
571                 if (!dbs_tuners_ins.powersave_bias) {
572                         if (policy->cur == policy->max)
573                                 return;
574
575                         __cpufreq_driver_target(policy, policy->max,
576                                 CPUFREQ_RELATION_H);
577                 } else {
578                         int freq = powersave_bias_target(policy, policy->max,
579                                         CPUFREQ_RELATION_H);
580                         __cpufreq_driver_target(policy, freq,
581                                 CPUFREQ_RELATION_L);
582                 }
583                 return;
584         }
585
586         /* Check for frequency decrease */
587         /* if we cannot reduce the frequency anymore, break out early */
588         if (policy->cur == policy->min)
589                 return;
590
591         /*
592          * The optimal frequency is the frequency that is the lowest that
593          * can support the current CPU usage without triggering the up
594          * policy. To be safe, we focus 10 points under the threshold.
595          */
596         if (max_load_freq <
597             (dbs_tuners_ins.up_threshold - dbs_tuners_ins.down_differential) *
598              policy->cur) {
599                 unsigned int freq_next;
600                 freq_next = max_load_freq /
601                                 (dbs_tuners_ins.up_threshold -
602                                  dbs_tuners_ins.down_differential);
603
604                 if (freq_next < policy->min)
605                         freq_next = policy->min;
606
607                 if (!dbs_tuners_ins.powersave_bias) {
608                         __cpufreq_driver_target(policy, freq_next,
609                                         CPUFREQ_RELATION_L);
610                 } else {
611                         int freq = powersave_bias_target(policy, freq_next,
612                                         CPUFREQ_RELATION_L);
613                         __cpufreq_driver_target(policy, freq,
614                                 CPUFREQ_RELATION_L);
615                 }
616         }
617 }
618
619 static void do_dbs_timer(struct work_struct *work)
620 {
621         struct cpu_dbs_info_s *dbs_info =
622                 container_of(work, struct cpu_dbs_info_s, work.work);
623         unsigned int cpu = dbs_info->cpu;
624         int sample_type = dbs_info->sample_type;
625
626         /* We want all CPUs to do sampling nearly on same jiffy */
627         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
628
629         delay -= jiffies % delay;
630         mutex_lock(&dbs_info->timer_mutex);
631
632         /* Common NORMAL_SAMPLE setup */
633         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
634         if (!dbs_tuners_ins.powersave_bias ||
635             sample_type == DBS_NORMAL_SAMPLE) {
636                 dbs_check_cpu(dbs_info);
637                 if (dbs_info->freq_lo) {
638                         /* Setup timer for SUB_SAMPLE */
639                         dbs_info->sample_type = DBS_SUB_SAMPLE;
640                         delay = dbs_info->freq_hi_jiffies;
641                 }
642         } else {
643                 __cpufreq_driver_target(dbs_info->cur_policy,
644                         dbs_info->freq_lo, CPUFREQ_RELATION_H);
645         }
646         queue_delayed_work_on(cpu, kondemand_wq, &dbs_info->work, delay);
647         mutex_unlock(&dbs_info->timer_mutex);
648 }
649
650 static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
651 {
652         /* We want all CPUs to do sampling nearly on same jiffy */
653         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
654         delay -= jiffies % delay;
655
656         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
657         INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer);
658         queue_delayed_work_on(dbs_info->cpu, kondemand_wq, &dbs_info->work,
659                 delay);
660 }
661
662 static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
663 {
664         cancel_delayed_work_sync(&dbs_info->work);
665 }
666
667 /*
668  * Not all CPUs want IO time to be accounted as busy; this dependson how
669  * efficient idling at a higher frequency/voltage is.
670  * Pavel Machek says this is not so for various generations of AMD and old
671  * Intel systems.
672  * Mike Chan (androidlcom) calis this is also not true for ARM.
673  * Because of this, whitelist specific known (series) of CPUs by default, and
674  * leave all others up to the user.
675  */
676 static int should_io_be_busy(void)
677 {
678 #if defined(CONFIG_X86)
679         /*
680          * For Intel, Core 2 (model 15) andl later have an efficient idle.
681          */
682         if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
683             boot_cpu_data.x86 == 6 &&
684             boot_cpu_data.x86_model >= 15)
685                 return 1;
686 #endif
687         return 0;
688 }
689
690 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
691                                    unsigned int event)
692 {
693         unsigned int cpu = policy->cpu;
694         struct cpu_dbs_info_s *this_dbs_info;
695         unsigned int j;
696         int rc;
697
698         this_dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
699
700         switch (event) {
701         case CPUFREQ_GOV_START:
702                 if ((!cpu_online(cpu)) || (!policy->cur))
703                         return -EINVAL;
704
705                 mutex_lock(&dbs_mutex);
706
707                 rc = sysfs_create_group(&policy->kobj, &dbs_attr_group_old);
708                 if (rc) {
709                         mutex_unlock(&dbs_mutex);
710                         return rc;
711                 }
712
713                 dbs_enable++;
714                 for_each_cpu(j, policy->cpus) {
715                         struct cpu_dbs_info_s *j_dbs_info;
716                         j_dbs_info = &per_cpu(od_cpu_dbs_info, j);
717                         j_dbs_info->cur_policy = policy;
718
719                         j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
720                                                 &j_dbs_info->prev_cpu_wall);
721                         if (dbs_tuners_ins.ignore_nice) {
722                                 j_dbs_info->prev_cpu_nice =
723                                                 kstat_cpu(j).cpustat.nice;
724                         }
725                 }
726                 this_dbs_info->cpu = cpu;
727                 ondemand_powersave_bias_init_cpu(cpu);
728                 /*
729                  * Start the timerschedule work, when this governor
730                  * is used for first time
731                  */
732                 if (dbs_enable == 1) {
733                         unsigned int latency;
734
735                         rc = sysfs_create_group(cpufreq_global_kobject,
736                                                 &dbs_attr_group);
737                         if (rc) {
738                                 mutex_unlock(&dbs_mutex);
739                                 return rc;
740                         }
741
742                         /* policy latency is in nS. Convert it to uS first */
743                         latency = policy->cpuinfo.transition_latency / 1000;
744                         if (latency == 0)
745                                 latency = 1;
746                         /* Bring kernel and HW constraints together */
747                         min_sampling_rate = max(min_sampling_rate,
748                                         MIN_LATENCY_MULTIPLIER * latency);
749                         dbs_tuners_ins.sampling_rate =
750                                 max(min_sampling_rate,
751                                     latency * LATENCY_MULTIPLIER);
752                         dbs_tuners_ins.io_is_busy = should_io_be_busy();
753                 }
754                 mutex_unlock(&dbs_mutex);
755
756                 mutex_init(&this_dbs_info->timer_mutex);
757                 dbs_timer_init(this_dbs_info);
758                 break;
759
760         case CPUFREQ_GOV_STOP:
761                 dbs_timer_exit(this_dbs_info);
762
763                 mutex_lock(&dbs_mutex);
764                 sysfs_remove_group(&policy->kobj, &dbs_attr_group_old);
765                 mutex_destroy(&this_dbs_info->timer_mutex);
766                 dbs_enable--;
767                 mutex_unlock(&dbs_mutex);
768                 if (!dbs_enable)
769                         sysfs_remove_group(cpufreq_global_kobject,
770                                            &dbs_attr_group);
771
772                 break;
773
774         case CPUFREQ_GOV_LIMITS:
775                 mutex_lock(&this_dbs_info->timer_mutex);
776                 if (policy->max < this_dbs_info->cur_policy->cur)
777                         __cpufreq_driver_target(this_dbs_info->cur_policy,
778                                 policy->max, CPUFREQ_RELATION_H);
779                 else if (policy->min > this_dbs_info->cur_policy->cur)
780                         __cpufreq_driver_target(this_dbs_info->cur_policy,
781                                 policy->min, CPUFREQ_RELATION_L);
782                 mutex_unlock(&this_dbs_info->timer_mutex);
783                 break;
784         }
785         return 0;
786 }
787
788 static int __init cpufreq_gov_dbs_init(void)
789 {
790         int err;
791         cputime64_t wall;
792         u64 idle_time;
793         int cpu = get_cpu();
794
795         idle_time = get_cpu_idle_time_us(cpu, &wall);
796         put_cpu();
797         if (idle_time != -1ULL) {
798                 /* Idle micro accounting is supported. Use finer thresholds */
799                 dbs_tuners_ins.up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
800                 dbs_tuners_ins.down_differential =
801                                         MICRO_FREQUENCY_DOWN_DIFFERENTIAL;
802                 /*
803                  * In no_hz/micro accounting case we set the minimum frequency
804                  * not depending on HZ, but fixed (very low). The deferred
805                  * timer might skip some samples if idle/sleeping as needed.
806                 */
807                 min_sampling_rate = MICRO_FREQUENCY_MIN_SAMPLE_RATE;
808         } else {
809                 /* For correct statistics, we need 10 ticks for each measure */
810                 min_sampling_rate =
811                         MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10);
812         }
813
814         kondemand_wq = create_workqueue("kondemand");
815         if (!kondemand_wq) {
816                 printk(KERN_ERR "Creation of kondemand failed\n");
817                 return -EFAULT;
818         }
819         err = cpufreq_register_governor(&cpufreq_gov_ondemand);
820         if (err)
821                 destroy_workqueue(kondemand_wq);
822
823         return err;
824 }
825
826 static void __exit cpufreq_gov_dbs_exit(void)
827 {
828         cpufreq_unregister_governor(&cpufreq_gov_ondemand);
829         destroy_workqueue(kondemand_wq);
830 }
831
832
833 MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
834 MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
835 MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
836         "Low Latency Frequency Transition capable processors");
837 MODULE_LICENSE("GPL");
838
839 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
840 fs_initcall(cpufreq_gov_dbs_init);
841 #else
842 module_init(cpufreq_gov_dbs_init);
843 #endif
844 module_exit(cpufreq_gov_dbs_exit);