cpufreq_conservative.c 20 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752
  1. /*
  2. * drivers/cpufreq/cpufreq_conservative.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. * (C) 2009 Alexander Clouter <alex@digriz.org.uk>
  8. *
  9. * This program is free software; you can redistribute it and/or modify
  10. * it under the terms of the GNU General Public License version 2 as
  11. * published by the Free Software Foundation.
  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. * dbs is used in this file as a shortform for demandbased switching
  27. * It helps to keep variable names smaller, simpler
  28. */
  29. #define DEF_FREQUENCY_UP_THRESHOLD (80)
  30. #define DEF_FREQUENCY_DOWN_THRESHOLD (20)
  31. /*
  32. * The polling frequency of this governor depends on the capability of
  33. * the processor. Default polling frequency is 1000 times the transition
  34. * latency of the processor. The governor will work on any processor with
  35. * transition latency <= 10mS, using appropriate sampling
  36. * rate.
  37. * For CPUs with transition latency > 10mS (mostly drivers with CPUFREQ_ETERNAL)
  38. * this governor will not work.
  39. * All times here are in uS.
  40. */
  41. #define MIN_SAMPLING_RATE_RATIO (2)
  42. static unsigned int min_sampling_rate;
  43. #define LATENCY_MULTIPLIER (1000)
  44. #define MIN_LATENCY_MULTIPLIER (100)
  45. #define DEF_SAMPLING_DOWN_FACTOR (1)
  46. #define MAX_SAMPLING_DOWN_FACTOR (10)
  47. #define TRANSITION_LATENCY_LIMIT (10 * 1000 * 1000)
  48. static void do_dbs_timer(struct work_struct *work);
  49. struct cpu_dbs_info_s {
  50. cputime64_t prev_cpu_idle;
  51. cputime64_t prev_cpu_wall;
  52. cputime64_t prev_cpu_nice;
  53. struct cpufreq_policy *cur_policy;
  54. struct delayed_work work;
  55. unsigned int down_skip;
  56. unsigned int requested_freq;
  57. int cpu;
  58. unsigned int enable:1;
  59. /*
  60. * percpu mutex that serializes governor limit change with
  61. * do_dbs_timer invocation. We do not want do_dbs_timer to run
  62. * when user is changing the governor or limits.
  63. */
  64. struct mutex timer_mutex;
  65. };
  66. static DEFINE_PER_CPU(struct cpu_dbs_info_s, cs_cpu_dbs_info);
  67. static unsigned int dbs_enable; /* number of CPUs using this policy */
  68. /*
  69. * dbs_mutex protects data in dbs_tuners_ins from concurrent changes on
  70. * different CPUs. It protects dbs_enable in governor start/stop.
  71. */
  72. static DEFINE_MUTEX(dbs_mutex);
  73. static struct workqueue_struct *kconservative_wq;
  74. static struct dbs_tuners {
  75. unsigned int sampling_rate;
  76. unsigned int sampling_down_factor;
  77. unsigned int up_threshold;
  78. unsigned int down_threshold;
  79. unsigned int ignore_nice;
  80. unsigned int freq_step;
  81. } dbs_tuners_ins = {
  82. .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
  83. .down_threshold = DEF_FREQUENCY_DOWN_THRESHOLD,
  84. .sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR,
  85. .ignore_nice = 0,
  86. .freq_step = 5,
  87. };
  88. static inline cputime64_t get_cpu_idle_time_jiffy(unsigned int cpu,
  89. cputime64_t *wall)
  90. {
  91. cputime64_t idle_time;
  92. cputime64_t cur_wall_time;
  93. cputime64_t busy_time;
  94. cur_wall_time = jiffies64_to_cputime64(get_jiffies_64());
  95. busy_time = cputime64_add(kstat_cpu(cpu).cpustat.user,
  96. kstat_cpu(cpu).cpustat.system);
  97. busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.irq);
  98. busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.softirq);
  99. busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.steal);
  100. busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.nice);
  101. idle_time = cputime64_sub(cur_wall_time, busy_time);
  102. if (wall)
  103. *wall = (cputime64_t)jiffies_to_usecs(cur_wall_time);
  104. return (cputime64_t)jiffies_to_usecs(idle_time);;
  105. }
  106. static inline cputime64_t get_cpu_idle_time(unsigned int cpu, cputime64_t *wall)
  107. {
  108. u64 idle_time = get_cpu_idle_time_us(cpu, wall);
  109. if (idle_time == -1ULL)
  110. return get_cpu_idle_time_jiffy(cpu, wall);
  111. return idle_time;
  112. }
  113. /* keep track of frequency transitions */
  114. static int
  115. dbs_cpufreq_notifier(struct notifier_block *nb, unsigned long val,
  116. void *data)
  117. {
  118. struct cpufreq_freqs *freq = data;
  119. struct cpu_dbs_info_s *this_dbs_info = &per_cpu(cs_cpu_dbs_info,
  120. freq->cpu);
  121. struct cpufreq_policy *policy;
  122. if (!this_dbs_info->enable)
  123. return 0;
  124. policy = this_dbs_info->cur_policy;
  125. /*
  126. * we only care if our internally tracked freq moves outside
  127. * the 'valid' ranges of freqency available to us otherwise
  128. * we do not change it
  129. */
  130. if (this_dbs_info->requested_freq > policy->max
  131. || this_dbs_info->requested_freq < policy->min)
  132. this_dbs_info->requested_freq = freq->new;
  133. return 0;
  134. }
  135. static struct notifier_block dbs_cpufreq_notifier_block = {
  136. .notifier_call = dbs_cpufreq_notifier
  137. };
  138. /************************** sysfs interface ************************/
  139. static ssize_t show_sampling_rate_max(struct kobject *kobj,
  140. struct attribute *attr, char *buf)
  141. {
  142. printk_once(KERN_INFO "CPUFREQ: conservative sampling_rate_max "
  143. "sysfs file is deprecated - used by: %s\n", current->comm);
  144. return sprintf(buf, "%u\n", -1U);
  145. }
  146. static ssize_t show_sampling_rate_min(struct kobject *kobj,
  147. struct attribute *attr, char *buf)
  148. {
  149. return sprintf(buf, "%u\n", min_sampling_rate);
  150. }
  151. define_one_global_ro(sampling_rate_max);
  152. define_one_global_ro(sampling_rate_min);
  153. /* cpufreq_conservative Governor Tunables */
  154. #define show_one(file_name, object) \
  155. static ssize_t show_##file_name \
  156. (struct kobject *kobj, struct attribute *attr, char *buf) \
  157. { \
  158. return sprintf(buf, "%u\n", dbs_tuners_ins.object); \
  159. }
  160. show_one(sampling_rate, sampling_rate);
  161. show_one(sampling_down_factor, sampling_down_factor);
  162. show_one(up_threshold, up_threshold);
  163. show_one(down_threshold, down_threshold);
  164. show_one(ignore_nice_load, ignore_nice);
  165. show_one(freq_step, freq_step);
  166. /*** delete after deprecation time ***/
  167. #define DEPRECATION_MSG(file_name) \
  168. printk_once(KERN_INFO "CPUFREQ: Per core conservative sysfs " \
  169. "interface is deprecated - " #file_name "\n");
  170. #define show_one_old(file_name) \
  171. static ssize_t show_##file_name##_old \
  172. (struct cpufreq_policy *unused, char *buf) \
  173. { \
  174. printk_once(KERN_INFO "CPUFREQ: Per core conservative sysfs " \
  175. "interface is deprecated - " #file_name "\n"); \
  176. return show_##file_name(NULL, NULL, buf); \
  177. }
  178. show_one_old(sampling_rate);
  179. show_one_old(sampling_down_factor);
  180. show_one_old(up_threshold);
  181. show_one_old(down_threshold);
  182. show_one_old(ignore_nice_load);
  183. show_one_old(freq_step);
  184. show_one_old(sampling_rate_min);
  185. show_one_old(sampling_rate_max);
  186. cpufreq_freq_attr_ro_old(sampling_rate_min);
  187. cpufreq_freq_attr_ro_old(sampling_rate_max);
  188. /*** delete after deprecation time ***/
  189. static ssize_t store_sampling_down_factor(struct kobject *a,
  190. struct attribute *b,
  191. const char *buf, size_t count)
  192. {
  193. unsigned int input;
  194. int ret;
  195. ret = sscanf(buf, "%u", &input);
  196. if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
  197. return -EINVAL;
  198. mutex_lock(&dbs_mutex);
  199. dbs_tuners_ins.sampling_down_factor = input;
  200. mutex_unlock(&dbs_mutex);
  201. return count;
  202. }
  203. static ssize_t store_sampling_rate(struct kobject *a, struct attribute *b,
  204. const char *buf, size_t count)
  205. {
  206. unsigned int input;
  207. int ret;
  208. ret = sscanf(buf, "%u", &input);
  209. if (ret != 1)
  210. return -EINVAL;
  211. mutex_lock(&dbs_mutex);
  212. dbs_tuners_ins.sampling_rate = max(input, min_sampling_rate);
  213. mutex_unlock(&dbs_mutex);
  214. return count;
  215. }
  216. static ssize_t store_up_threshold(struct kobject *a, struct attribute *b,
  217. const char *buf, size_t count)
  218. {
  219. unsigned int input;
  220. int ret;
  221. ret = sscanf(buf, "%u", &input);
  222. mutex_lock(&dbs_mutex);
  223. if (ret != 1 || input > 100 ||
  224. input <= dbs_tuners_ins.down_threshold) {
  225. mutex_unlock(&dbs_mutex);
  226. return -EINVAL;
  227. }
  228. dbs_tuners_ins.up_threshold = input;
  229. mutex_unlock(&dbs_mutex);
  230. return count;
  231. }
  232. static ssize_t store_down_threshold(struct kobject *a, struct attribute *b,
  233. const char *buf, size_t count)
  234. {
  235. unsigned int input;
  236. int ret;
  237. ret = sscanf(buf, "%u", &input);
  238. mutex_lock(&dbs_mutex);
  239. /* cannot be lower than 11 otherwise freq will not fall */
  240. if (ret != 1 || input < 11 || input > 100 ||
  241. input >= dbs_tuners_ins.up_threshold) {
  242. mutex_unlock(&dbs_mutex);
  243. return -EINVAL;
  244. }
  245. dbs_tuners_ins.down_threshold = input;
  246. mutex_unlock(&dbs_mutex);
  247. return count;
  248. }
  249. static ssize_t store_ignore_nice_load(struct kobject *a, struct attribute *b,
  250. const char *buf, size_t count)
  251. {
  252. unsigned int input;
  253. int ret;
  254. unsigned int j;
  255. ret = sscanf(buf, "%u", &input);
  256. if (ret != 1)
  257. return -EINVAL;
  258. if (input > 1)
  259. input = 1;
  260. mutex_lock(&dbs_mutex);
  261. if (input == dbs_tuners_ins.ignore_nice) { /* nothing to do */
  262. mutex_unlock(&dbs_mutex);
  263. return count;
  264. }
  265. dbs_tuners_ins.ignore_nice = input;
  266. /* we need to re-evaluate prev_cpu_idle */
  267. for_each_online_cpu(j) {
  268. struct cpu_dbs_info_s *dbs_info;
  269. dbs_info = &per_cpu(cs_cpu_dbs_info, j);
  270. dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
  271. &dbs_info->prev_cpu_wall);
  272. if (dbs_tuners_ins.ignore_nice)
  273. dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
  274. }
  275. mutex_unlock(&dbs_mutex);
  276. return count;
  277. }
  278. static ssize_t store_freq_step(struct kobject *a, struct attribute *b,
  279. const char *buf, size_t count)
  280. {
  281. unsigned int input;
  282. int ret;
  283. ret = sscanf(buf, "%u", &input);
  284. if (ret != 1)
  285. return -EINVAL;
  286. if (input > 100)
  287. input = 100;
  288. /* no need to test here if freq_step is zero as the user might actually
  289. * want this, they would be crazy though :) */
  290. mutex_lock(&dbs_mutex);
  291. dbs_tuners_ins.freq_step = input;
  292. mutex_unlock(&dbs_mutex);
  293. return count;
  294. }
  295. define_one_global_rw(sampling_rate);
  296. define_one_global_rw(sampling_down_factor);
  297. define_one_global_rw(up_threshold);
  298. define_one_global_rw(down_threshold);
  299. define_one_global_rw(ignore_nice_load);
  300. define_one_global_rw(freq_step);
  301. static struct attribute *dbs_attributes[] = {
  302. &sampling_rate_max.attr,
  303. &sampling_rate_min.attr,
  304. &sampling_rate.attr,
  305. &sampling_down_factor.attr,
  306. &up_threshold.attr,
  307. &down_threshold.attr,
  308. &ignore_nice_load.attr,
  309. &freq_step.attr,
  310. NULL
  311. };
  312. static struct attribute_group dbs_attr_group = {
  313. .attrs = dbs_attributes,
  314. .name = "conservative",
  315. };
  316. /*** delete after deprecation time ***/
  317. #define write_one_old(file_name) \
  318. static ssize_t store_##file_name##_old \
  319. (struct cpufreq_policy *unused, const char *buf, size_t count) \
  320. { \
  321. printk_once(KERN_INFO "CPUFREQ: Per core conservative sysfs " \
  322. "interface is deprecated - " #file_name "\n"); \
  323. return store_##file_name(NULL, NULL, buf, count); \
  324. }
  325. write_one_old(sampling_rate);
  326. write_one_old(sampling_down_factor);
  327. write_one_old(up_threshold);
  328. write_one_old(down_threshold);
  329. write_one_old(ignore_nice_load);
  330. write_one_old(freq_step);
  331. cpufreq_freq_attr_rw_old(sampling_rate);
  332. cpufreq_freq_attr_rw_old(sampling_down_factor);
  333. cpufreq_freq_attr_rw_old(up_threshold);
  334. cpufreq_freq_attr_rw_old(down_threshold);
  335. cpufreq_freq_attr_rw_old(ignore_nice_load);
  336. cpufreq_freq_attr_rw_old(freq_step);
  337. static struct attribute *dbs_attributes_old[] = {
  338. &sampling_rate_max_old.attr,
  339. &sampling_rate_min_old.attr,
  340. &sampling_rate_old.attr,
  341. &sampling_down_factor_old.attr,
  342. &up_threshold_old.attr,
  343. &down_threshold_old.attr,
  344. &ignore_nice_load_old.attr,
  345. &freq_step_old.attr,
  346. NULL
  347. };
  348. static struct attribute_group dbs_attr_group_old = {
  349. .attrs = dbs_attributes_old,
  350. .name = "conservative",
  351. };
  352. /*** delete after deprecation time ***/
  353. /************************** sysfs end ************************/
  354. static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
  355. {
  356. unsigned int load = 0;
  357. unsigned int max_load = 0;
  358. unsigned int freq_target;
  359. struct cpufreq_policy *policy;
  360. unsigned int j;
  361. policy = this_dbs_info->cur_policy;
  362. /*
  363. * Every sampling_rate, we check, if current idle time is less
  364. * than 20% (default), then we try to increase frequency
  365. * Every sampling_rate*sampling_down_factor, we check, if current
  366. * idle time is more than 80%, then we try to decrease frequency
  367. *
  368. * Any frequency increase takes it to the maximum frequency.
  369. * Frequency reduction happens at minimum steps of
  370. * 5% (default) of maximum frequency
  371. */
  372. /* Get Absolute Load */
  373. for_each_cpu(j, policy->cpus) {
  374. struct cpu_dbs_info_s *j_dbs_info;
  375. cputime64_t cur_wall_time, cur_idle_time;
  376. unsigned int idle_time, wall_time;
  377. j_dbs_info = &per_cpu(cs_cpu_dbs_info, j);
  378. cur_idle_time = get_cpu_idle_time(j, &cur_wall_time);
  379. wall_time = (unsigned int) cputime64_sub(cur_wall_time,
  380. j_dbs_info->prev_cpu_wall);
  381. j_dbs_info->prev_cpu_wall = cur_wall_time;
  382. idle_time = (unsigned int) cputime64_sub(cur_idle_time,
  383. j_dbs_info->prev_cpu_idle);
  384. j_dbs_info->prev_cpu_idle = cur_idle_time;
  385. if (dbs_tuners_ins.ignore_nice) {
  386. cputime64_t cur_nice;
  387. unsigned long cur_nice_jiffies;
  388. cur_nice = cputime64_sub(kstat_cpu(j).cpustat.nice,
  389. j_dbs_info->prev_cpu_nice);
  390. /*
  391. * Assumption: nice time between sampling periods will
  392. * be less than 2^32 jiffies for 32 bit sys
  393. */
  394. cur_nice_jiffies = (unsigned long)
  395. cputime64_to_jiffies64(cur_nice);
  396. j_dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
  397. idle_time += jiffies_to_usecs(cur_nice_jiffies);
  398. }
  399. if (unlikely(!wall_time || wall_time < idle_time))
  400. continue;
  401. load = 100 * (wall_time - idle_time) / wall_time;
  402. if (load > max_load)
  403. max_load = load;
  404. }
  405. /*
  406. * break out if we 'cannot' reduce the speed as the user might
  407. * want freq_step to be zero
  408. */
  409. if (dbs_tuners_ins.freq_step == 0)
  410. return;
  411. /* Check for frequency increase */
  412. if (max_load > dbs_tuners_ins.up_threshold) {
  413. this_dbs_info->down_skip = 0;
  414. /* if we are already at full speed then break out early */
  415. if (this_dbs_info->requested_freq == policy->max)
  416. return;
  417. freq_target = (dbs_tuners_ins.freq_step * policy->max) / 100;
  418. /* max freq cannot be less than 100. But who knows.... */
  419. if (unlikely(freq_target == 0))
  420. freq_target = 5;
  421. this_dbs_info->requested_freq += freq_target;
  422. if (this_dbs_info->requested_freq > policy->max)
  423. this_dbs_info->requested_freq = policy->max;
  424. __cpufreq_driver_target(policy, this_dbs_info->requested_freq,
  425. CPUFREQ_RELATION_H);
  426. return;
  427. }
  428. /*
  429. * The optimal frequency is the frequency that is the lowest that
  430. * can support the current CPU usage without triggering the up
  431. * policy. To be safe, we focus 10 points under the threshold.
  432. */
  433. if (max_load < (dbs_tuners_ins.down_threshold - 10)) {
  434. freq_target = (dbs_tuners_ins.freq_step * policy->max) / 100;
  435. this_dbs_info->requested_freq -= freq_target;
  436. if (this_dbs_info->requested_freq < policy->min)
  437. this_dbs_info->requested_freq = policy->min;
  438. /*
  439. * if we cannot reduce the frequency anymore, break out early
  440. */
  441. if (policy->cur == policy->min)
  442. return;
  443. __cpufreq_driver_target(policy, this_dbs_info->requested_freq,
  444. CPUFREQ_RELATION_H);
  445. return;
  446. }
  447. }
  448. static void do_dbs_timer(struct work_struct *work)
  449. {
  450. struct cpu_dbs_info_s *dbs_info =
  451. container_of(work, struct cpu_dbs_info_s, work.work);
  452. unsigned int cpu = dbs_info->cpu;
  453. /* We want all CPUs to do sampling nearly on same jiffy */
  454. int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
  455. delay -= jiffies % delay;
  456. mutex_lock(&dbs_info->timer_mutex);
  457. dbs_check_cpu(dbs_info);
  458. queue_delayed_work_on(cpu, kconservative_wq, &dbs_info->work, delay);
  459. mutex_unlock(&dbs_info->timer_mutex);
  460. }
  461. static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
  462. {
  463. /* We want all CPUs to do sampling nearly on same jiffy */
  464. int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
  465. delay -= jiffies % delay;
  466. dbs_info->enable = 1;
  467. INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer);
  468. queue_delayed_work_on(dbs_info->cpu, kconservative_wq, &dbs_info->work,
  469. delay);
  470. }
  471. static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
  472. {
  473. dbs_info->enable = 0;
  474. cancel_delayed_work_sync(&dbs_info->work);
  475. }
  476. static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
  477. unsigned int event)
  478. {
  479. unsigned int cpu = policy->cpu;
  480. struct cpu_dbs_info_s *this_dbs_info;
  481. unsigned int j;
  482. int rc;
  483. this_dbs_info = &per_cpu(cs_cpu_dbs_info, cpu);
  484. switch (event) {
  485. case CPUFREQ_GOV_START:
  486. if ((!cpu_online(cpu)) || (!policy->cur))
  487. return -EINVAL;
  488. mutex_lock(&dbs_mutex);
  489. rc = sysfs_create_group(&policy->kobj, &dbs_attr_group_old);
  490. if (rc) {
  491. mutex_unlock(&dbs_mutex);
  492. return rc;
  493. }
  494. for_each_cpu(j, policy->cpus) {
  495. struct cpu_dbs_info_s *j_dbs_info;
  496. j_dbs_info = &per_cpu(cs_cpu_dbs_info, j);
  497. j_dbs_info->cur_policy = policy;
  498. j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
  499. &j_dbs_info->prev_cpu_wall);
  500. if (dbs_tuners_ins.ignore_nice) {
  501. j_dbs_info->prev_cpu_nice =
  502. kstat_cpu(j).cpustat.nice;
  503. }
  504. }
  505. this_dbs_info->down_skip = 0;
  506. this_dbs_info->requested_freq = policy->cur;
  507. mutex_init(&this_dbs_info->timer_mutex);
  508. dbs_enable++;
  509. /*
  510. * Start the timerschedule work, when this governor
  511. * is used for first time
  512. */
  513. if (dbs_enable == 1) {
  514. unsigned int latency;
  515. /* policy latency is in nS. Convert it to uS first */
  516. latency = policy->cpuinfo.transition_latency / 1000;
  517. if (latency == 0)
  518. latency = 1;
  519. rc = sysfs_create_group(cpufreq_global_kobject,
  520. &dbs_attr_group);
  521. if (rc) {
  522. mutex_unlock(&dbs_mutex);
  523. return rc;
  524. }
  525. /*
  526. * conservative does not implement micro like ondemand
  527. * governor, thus we are bound to jiffes/HZ
  528. */
  529. min_sampling_rate =
  530. MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10);
  531. /* Bring kernel and HW constraints together */
  532. min_sampling_rate = max(min_sampling_rate,
  533. MIN_LATENCY_MULTIPLIER * latency);
  534. dbs_tuners_ins.sampling_rate =
  535. max(min_sampling_rate,
  536. latency * LATENCY_MULTIPLIER);
  537. cpufreq_register_notifier(
  538. &dbs_cpufreq_notifier_block,
  539. CPUFREQ_TRANSITION_NOTIFIER);
  540. }
  541. mutex_unlock(&dbs_mutex);
  542. dbs_timer_init(this_dbs_info);
  543. break;
  544. case CPUFREQ_GOV_STOP:
  545. dbs_timer_exit(this_dbs_info);
  546. mutex_lock(&dbs_mutex);
  547. sysfs_remove_group(&policy->kobj, &dbs_attr_group_old);
  548. dbs_enable--;
  549. mutex_destroy(&this_dbs_info->timer_mutex);
  550. /*
  551. * Stop the timerschedule work, when this governor
  552. * is used for first time
  553. */
  554. if (dbs_enable == 0)
  555. cpufreq_unregister_notifier(
  556. &dbs_cpufreq_notifier_block,
  557. CPUFREQ_TRANSITION_NOTIFIER);
  558. mutex_unlock(&dbs_mutex);
  559. if (!dbs_enable)
  560. sysfs_remove_group(cpufreq_global_kobject,
  561. &dbs_attr_group);
  562. break;
  563. case CPUFREQ_GOV_LIMITS:
  564. mutex_lock(&this_dbs_info->timer_mutex);
  565. if (policy->max < this_dbs_info->cur_policy->cur)
  566. __cpufreq_driver_target(
  567. this_dbs_info->cur_policy,
  568. policy->max, CPUFREQ_RELATION_H);
  569. else if (policy->min > this_dbs_info->cur_policy->cur)
  570. __cpufreq_driver_target(
  571. this_dbs_info->cur_policy,
  572. policy->min, CPUFREQ_RELATION_L);
  573. mutex_unlock(&this_dbs_info->timer_mutex);
  574. break;
  575. }
  576. return 0;
  577. }
  578. #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_CONSERVATIVE
  579. static
  580. #endif
  581. struct cpufreq_governor cpufreq_gov_conservative = {
  582. .name = "conservative",
  583. .governor = cpufreq_governor_dbs,
  584. .max_transition_latency = TRANSITION_LATENCY_LIMIT,
  585. .owner = THIS_MODULE,
  586. };
  587. static int __init cpufreq_gov_dbs_init(void)
  588. {
  589. int err;
  590. kconservative_wq = create_workqueue("kconservative");
  591. if (!kconservative_wq) {
  592. printk(KERN_ERR "Creation of kconservative failed\n");
  593. return -EFAULT;
  594. }
  595. err = cpufreq_register_governor(&cpufreq_gov_conservative);
  596. if (err)
  597. destroy_workqueue(kconservative_wq);
  598. return err;
  599. }
  600. static void __exit cpufreq_gov_dbs_exit(void)
  601. {
  602. cpufreq_unregister_governor(&cpufreq_gov_conservative);
  603. destroy_workqueue(kconservative_wq);
  604. }
  605. MODULE_AUTHOR("Alexander Clouter <alex@digriz.org.uk>");
  606. MODULE_DESCRIPTION("'cpufreq_conservative' - A dynamic cpufreq governor for "
  607. "Low Latency Frequency Transition capable processors "
  608. "optimised for use in a battery environment");
  609. MODULE_LICENSE("GPL");
  610. #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_CONSERVATIVE
  611. fs_initcall(cpufreq_gov_dbs_init);
  612. #else
  613. module_init(cpufreq_gov_dbs_init);
  614. #endif
  615. module_exit(cpufreq_gov_dbs_exit);