SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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 } dbs_tuners_ins = {
 113         .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
 114         .down_differential = DEF_FREQUENCY_DOWN_DIFFERENTIAL,
 115         .ignore_nice = 0,
 116         .powersave_bias = 0,
 117 };
 118
 119 static inline cputime64_t get_cpu_idle_time_jiffy(unsigned int cpu,
 120                                                         cputime64_t *wall)
 121 {
 122         cputime64_t idle_time;
 123         cputime64_t cur_wall_time;
 124         cputime64_t busy_time;
 125
 126         cur_wall_time = jiffies64_to_cputime64(get_jiffies_64());
 127         busy_time = cputime64_add(kstat_cpu(cpu).cpustat.user,
 128                         kstat_cpu(cpu).cpustat.system);
 129
 130         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.irq);
 131         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.softirq);
 132         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.steal);
 133         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.nice);
 134
 135         idle_time = cputime64_sub(cur_wall_time, busy_time);
 136         if (wall)
 137                 *wall = (cputime64_t)jiffies_to_usecs(cur_wall_time);
 138
 139         return (cputime64_t)jiffies_to_usecs(idle_time);
 140 }
 141
 142 static inline cputime64_t get_cpu_idle_time(unsigned int cpu, cputime64_t *wall)
 143 {
 144         u64 idle_time = get_cpu_idle_time_us(cpu, wall);
 145
 146         if (idle_time == -1ULL)
 147                 return get_cpu_idle_time_jiffy(cpu, wall);
 148
 149         return idle_time;
 150 }
 151
 152 static inline cputime64_t get_cpu_iowait_time(unsigned int cpu, cputime64_t *wall)
 153 {
 154         u64 iowait_time = get_cpu_iowait_time_us(cpu, wall);
 155
 156         if (iowait_time == -1ULL)
 157                 return 0;
 158
 159         return iowait_time;
 160 }
 161
 162 /*
 163  * Find right freq to be set now with powersave_bias on.
 164  * Returns the freq_hi to be used right now and will set freq_hi_jiffies,
 165  * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs.
 166  */
 167 static unsigned int powersave_bias_target(struct cpufreq_policy *policy,
 168                                           unsigned int freq_next,
 169                                           unsigned int relation)
 170 {
 171         unsigned int freq_req, freq_reduc, freq_avg;
 172         unsigned int freq_hi, freq_lo;
 173         unsigned int index = 0;
 174         unsigned int jiffies_total, jiffies_hi, jiffies_lo;
 175         struct cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
 176                                                    policy->cpu);
 177
 178         if (!dbs_info->freq_table) {
 179                 dbs_info->freq_lo = 0;
 180                 dbs_info->freq_lo_jiffies = 0;
 181                 return freq_next;
 182         }
 183
 184         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next,
 185                         relation, &index);
 186         freq_req = dbs_info->freq_table[index].frequency;
 187         freq_reduc = freq_req * dbs_tuners_ins.powersave_bias / 1000;
 188         freq_avg = freq_req - freq_reduc;
 189
 190         /* Find freq bounds for freq_avg in freq_table */
 191         index = 0;
 192         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
 193                         CPUFREQ_RELATION_H, &index);
 194         freq_lo = dbs_info->freq_table[index].frequency;
 195         index = 0;
 196         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
 197                         CPUFREQ_RELATION_L, &index);
 198         freq_hi = dbs_info->freq_table[index].frequency;
 199
 200         /* Find out how long we have to be in hi and lo freqs */
 201         if (freq_hi == freq_lo) {
 202                 dbs_info->freq_lo = 0;
 203                 dbs_info->freq_lo_jiffies = 0;
 204                 return freq_lo;
 205         }
 206         jiffies_total = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 207         jiffies_hi = (freq_avg - freq_lo) * jiffies_total;
 208         jiffies_hi += ((freq_hi - freq_lo) / 2);
 209         jiffies_hi /= (freq_hi - freq_lo);
 210         jiffies_lo = jiffies_total - jiffies_hi;
 211         dbs_info->freq_lo = freq_lo;
 212         dbs_info->freq_lo_jiffies = jiffies_lo;
 213         dbs_info->freq_hi_jiffies = jiffies_hi;
 214         return freq_hi;
 215 }
 216
 217 static void ondemand_powersave_bias_init_cpu(int cpu)
 218 {
 219         struct cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
 220         dbs_info->freq_table = cpufreq_frequency_get_table(cpu);
 221         dbs_info->freq_lo = 0;
 222 }
 223
 224 static void ondemand_powersave_bias_init(void)
 225 {
 226         int i;
 227         for_each_online_cpu(i) {
 228                 ondemand_powersave_bias_init_cpu(i);
 229         }
 230 }
 231
 232 /************************** sysfs interface ************************/
 233
 234 static ssize_t show_sampling_rate_max(struct kobject *kobj,
 235                                       struct attribute *attr, char *buf)
 236 {
 237         printk_once(KERN_INFO "CPUFREQ: ondemand sampling_rate_max "
 238                "sysfs file is deprecated - used by: %s\n", current->comm);
 239         return sprintf(buf, "%u\n", -1U);
 240 }
 241
 242 static ssize_t show_sampling_rate_min(struct kobject *kobj,
 243                                       struct attribute *attr, char *buf)
 244 {
 245         return sprintf(buf, "%u\n", min_sampling_rate);
 246 }
 247
 248 #define define_one_ro(_name)            \
 249 static struct global_attr _name =       \
 250 __ATTR(_name, 0444, show_##_name, NULL)
 251
 252 define_one_ro(sampling_rate_max);
 253 define_one_ro(sampling_rate_min);
 254
 255 /* cpufreq_ondemand Governor Tunables */
 256 #define show_one(file_name, object)                                     \
 257 static ssize_t show_##file_name                                         \
 258 (struct kobject *kobj, struct attribute *attr, char *buf)              \
 259 {                                                                       \
 260         return sprintf(buf, "%u\n", dbs_tuners_ins.object);             \
 261 }
 262 show_one(sampling_rate, sampling_rate);
 263 show_one(up_threshold, up_threshold);
 264 show_one(ignore_nice_load, ignore_nice);
 265 show_one(powersave_bias, powersave_bias);
 266
 267 /*** delete after deprecation time ***/
 268
 269 #define DEPRECATION_MSG(file_name)                                      \
 270         printk_once(KERN_INFO "CPUFREQ: Per core ondemand sysfs "       \
 271                     "interface is deprecated - " #file_name "\n");
 272
 273 #define show_one_old(file_name)                                         \
 274 static ssize_t show_##file_name##_old                                   \
 275 (struct cpufreq_policy *unused, char *buf)                              \
 276 {                                                                       \
 277         printk_once(KERN_INFO "CPUFREQ: Per core ondemand sysfs "       \
 278                     "interface is deprecated - " #file_name "\n");      \
 279         return show_##file_name(NULL, NULL, buf);                       \
 280 }
 281 show_one_old(sampling_rate);
 282 show_one_old(up_threshold);
 283 show_one_old(ignore_nice_load);
 284 show_one_old(powersave_bias);
 285 show_one_old(sampling_rate_min);
 286 show_one_old(sampling_rate_max);
 287
 288 #define define_one_ro_old(object, _name)       \
 289 static struct freq_attr object =               \
 290 __ATTR(_name, 0444, show_##_name##_old, NULL)
 291
 292 define_one_ro_old(sampling_rate_min_old, sampling_rate_min);
 293 define_one_ro_old(sampling_rate_max_old, sampling_rate_max);
 294
 295 /*** delete after deprecation time ***/
 296
 297 static ssize_t store_sampling_rate(struct kobject *a, struct attribute *b,
 298                                    const char *buf, size_t count)
 299 {
 300         unsigned int input;
 301         int ret;
 302         ret = sscanf(buf, "%u", &input);
 303         if (ret != 1)
 304                 return -EINVAL;
 305
 306         mutex_lock(&dbs_mutex);
 307         dbs_tuners_ins.sampling_rate = max(input, min_sampling_rate);
 308         mutex_unlock(&dbs_mutex);
 309
 310         return count;
 311 }
 312
 313 static ssize_t store_up_threshold(struct kobject *a, struct attribute *b,
 314                                   const char *buf, size_t count)
 315 {
 316         unsigned int input;
 317         int ret;
 318         ret = sscanf(buf, "%u", &input);
 319
 320         if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
 321                         input < MIN_FREQUENCY_UP_THRESHOLD) {
 322                 return -EINVAL;
 323         }
 324
 325         mutex_lock(&dbs_mutex);
 326         dbs_tuners_ins.up_threshold = input;
 327         mutex_unlock(&dbs_mutex);
 328
 329         return count;
 330 }
 331
 332 static ssize_t store_ignore_nice_load(struct kobject *a, struct attribute *b,
 333                                       const char *buf, size_t count)
 334 {
 335         unsigned int input;
 336         int ret;
 337
 338         unsigned int j;
 339
 340         ret = sscanf(buf, "%u", &input);
 341         if (ret != 1)
 342                 return -EINVAL;
 343
 344         if (input > 1)
 345                 input = 1;
 346
 347         mutex_lock(&dbs_mutex);
 348         if (input == dbs_tuners_ins.ignore_nice) { /* nothing to do */
 349                 mutex_unlock(&dbs_mutex);
 350                 return count;
 351         }
 352         dbs_tuners_ins.ignore_nice = input;
 353
 354         /* we need to re-evaluate prev_cpu_idle */
 355         for_each_online_cpu(j) {
 356                 struct cpu_dbs_info_s *dbs_info;
 357                 dbs_info = &per_cpu(od_cpu_dbs_info, j);
 358                 dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
 359                                                 &dbs_info->prev_cpu_wall);
 360                 if (dbs_tuners_ins.ignore_nice)
 361                         dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
 362
 363         }
 364         mutex_unlock(&dbs_mutex);
 365
 366         return count;
 367 }
 368
 369 static ssize_t store_powersave_bias(struct kobject *a, struct attribute *b,
 370                                     const char *buf, size_t count)
 371 {
 372         unsigned int input;
 373         int ret;
 374         ret = sscanf(buf, "%u", &input);
 375
 376         if (ret != 1)
 377                 return -EINVAL;
 378
 379         if (input > 1000)
 380                 input = 1000;
 381
 382         mutex_lock(&dbs_mutex);
 383         dbs_tuners_ins.powersave_bias = input;
 384         ondemand_powersave_bias_init();
 385         mutex_unlock(&dbs_mutex);
 386
 387         return count;
 388 }
 389
 390 #define define_one_rw(_name) \
 391 static struct global_attr _name = \
 392 __ATTR(_name, 0644, show_##_name, store_##_name)
 393
 394 define_one_rw(sampling_rate);
 395 define_one_rw(up_threshold);
 396 define_one_rw(ignore_nice_load);
 397 define_one_rw(powersave_bias);
 398
 399 static struct attribute *dbs_attributes[] = {
 400         &sampling_rate_max.attr,
 401         &sampling_rate_min.attr,
 402         &sampling_rate.attr,
 403         &up_threshold.attr,
 404         &ignore_nice_load.attr,
 405         &powersave_bias.attr,
 406         NULL
 407 };
 408
 409 static struct attribute_group dbs_attr_group = {
 410         .attrs = dbs_attributes,
 411         .name = "ondemand",
 412 };
 413
 414 /*** delete after deprecation time ***/
 415
 416 #define write_one_old(file_name)                                        \
 417 static ssize_t store_##file_name##_old                                  \
 418 (struct cpufreq_policy *unused, const char *buf, size_t count)          \
 419 {                                                                       \
 420        printk_once(KERN_INFO "CPUFREQ: Per core ondemand sysfs "        \
 421                    "interface is deprecated - " #file_name "\n");       \
 422        return store_##file_name(NULL, NULL, buf, count);                \
 423 }
 424 write_one_old(sampling_rate);
 425 write_one_old(up_threshold);
 426 write_one_old(ignore_nice_load);
 427 write_one_old(powersave_bias);
 428
 429 #define define_one_rw_old(object, _name)       \
 430 static struct freq_attr object =               \
 431 __ATTR(_name, 0644, show_##_name##_old, store_##_name##_old)
 432
 433 define_one_rw_old(sampling_rate_old, sampling_rate);
 434 define_one_rw_old(up_threshold_old, up_threshold);
 435 define_one_rw_old(ignore_nice_load_old, ignore_nice_load);
 436 define_one_rw_old(powersave_bias_old, powersave_bias);
 437
 438 static struct attribute *dbs_attributes_old[] = {
 439        &sampling_rate_max_old.attr,
 440        &sampling_rate_min_old.attr,
 441        &sampling_rate_old.attr,
 442        &up_threshold_old.attr,
 443        &ignore_nice_load_old.attr,
 444        &powersave_bias_old.attr,
 445        NULL
 446 };
 447
 448 static struct attribute_group dbs_attr_group_old = {
 449        .attrs = dbs_attributes_old,
 450        .name = "ondemand",
 451 };
 452
 453 /*** delete after deprecation time ***/
 454
 455 /************************** sysfs end ************************/
 456
 457 static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
 458 {
 459         unsigned int max_load_freq;
 460
 461         struct cpufreq_policy *policy;
 462         unsigned int j;
 463
 464         this_dbs_info->freq_lo = 0;
 465         policy = this_dbs_info->cur_policy;
 466
 467         /*
 468          * Every sampling_rate, we check, if current idle time is less
 469          * than 20% (default), then we try to increase frequency
 470          * Every sampling_rate, we look for a the lowest
 471          * frequency which can sustain the load while keeping idle time over
 472          * 30%. If such a frequency exist, we try to decrease to this frequency.
 473          *
 474          * Any frequency increase takes it to the maximum frequency.
 475          * Frequency reduction happens at minimum steps of
 476          * 5% (default) of current frequency
 477          */
 478
 479         /* Get Absolute Load - in terms of freq */
 480         max_load_freq = 0;
 481
 482         for_each_cpu(j, policy->cpus) {
 483                 struct cpu_dbs_info_s *j_dbs_info;
 484                 cputime64_t cur_wall_time, cur_idle_time, cur_iowait_time;
 485                 unsigned int idle_time, wall_time, iowait_time;
 486                 unsigned int load, load_freq;
 487                 int freq_avg;
 488
 489                 j_dbs_info = &per_cpu(od_cpu_dbs_info, j);
 490
 491                 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time);
 492                 cur_iowait_time = get_cpu_iowait_time(j, &cur_wall_time);
 493
 494                 wall_time = (unsigned int) cputime64_sub(cur_wall_time,
 495                                 j_dbs_info->prev_cpu_wall);
 496                 j_dbs_info->prev_cpu_wall = cur_wall_time;
 497
 498                 idle_time = (unsigned int) cputime64_sub(cur_idle_time,
 499                                 j_dbs_info->prev_cpu_idle);
 500                 j_dbs_info->prev_cpu_idle = cur_idle_time;
 501
 502                 iowait_time = (unsigned int) cputime64_sub(cur_iowait_time,
 503                                 j_dbs_info->prev_cpu_iowait);
 504                 j_dbs_info->prev_cpu_iowait = cur_iowait_time;
 505
 506                 if (dbs_tuners_ins.ignore_nice) {
 507                         cputime64_t cur_nice;
 508                         unsigned long cur_nice_jiffies;
 509
 510                         cur_nice = cputime64_sub(kstat_cpu(j).cpustat.nice,
 511                                          j_dbs_info->prev_cpu_nice);
 512                         /*
 513                          * Assumption: nice time between sampling periods will
 514                          * be less than 2^32 jiffies for 32 bit sys
 515                          */
 516                         cur_nice_jiffies = (unsigned long)
 517                                         cputime64_to_jiffies64(cur_nice);
 518
 519                         j_dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
 520                         idle_time += jiffies_to_usecs(cur_nice_jiffies);
 521                 }
 522
 523                 /*
 524                  * For the purpose of ondemand, waiting for disk IO is an
 525                  * indication that you're performance critical, and not that
 526                  * the system is actually idle. So subtract the iowait time
 527                  * from the cpu idle time.
 528                  */
 529
 530                 if (idle_time >= iowait_time)
 531                         idle_time -= iowait_time;
 532
 533                 if (unlikely(!wall_time || wall_time < idle_time))
 534                         continue;
 535
 536                 load = 100 * (wall_time - idle_time) / wall_time;
 537
 538                 freq_avg = __cpufreq_driver_getavg(policy, j);
 539                 if (freq_avg <= 0)
 540                         freq_avg = policy->cur;
 541
 542                 load_freq = load * freq_avg;
 543                 if (load_freq > max_load_freq)
 544                         max_load_freq = load_freq;
 545         }
 546
 547         /* Check for frequency increase */
 548         if (max_load_freq > dbs_tuners_ins.up_threshold * policy->cur) {
 549                 /* if we are already at full speed then break out early */
 550                 if (!dbs_tuners_ins.powersave_bias) {
 551                         if (policy->cur == policy->max)
 552                                 return;
 553
 554                         __cpufreq_driver_target(policy, policy->max,
 555                                 CPUFREQ_RELATION_H);
 556                 } else {
 557                         int freq = powersave_bias_target(policy, policy->max,
 558                                         CPUFREQ_RELATION_H);
 559                         __cpufreq_driver_target(policy, freq,
 560                                 CPUFREQ_RELATION_L);
 561                 }
 562                 return;
 563         }
 564
 565         /* Check for frequency decrease */
 566         /* if we cannot reduce the frequency anymore, break out early */
 567         if (policy->cur == policy->min)
 568                 return;
 569
 570         /*
 571          * The optimal frequency is the frequency that is the lowest that
 572          * can support the current CPU usage without triggering the up
 573          * policy. To be safe, we focus 10 points under the threshold.
 574          */
 575         if (max_load_freq <
 576             (dbs_tuners_ins.up_threshold - dbs_tuners_ins.down_differential) *
 577              policy->cur) {
 578                 unsigned int freq_next;
 579                 freq_next = max_load_freq /
 580                                 (dbs_tuners_ins.up_threshold -
 581                                  dbs_tuners_ins.down_differential);
 582
 583                 if (freq_next < policy->min)
 584                         freq_next = policy->min;
 585
 586                 if (!dbs_tuners_ins.powersave_bias) {
 587                         __cpufreq_driver_target(policy, freq_next,
 588                                         CPUFREQ_RELATION_L);
 589                 } else {
 590                         int freq = powersave_bias_target(policy, freq_next,
 591                                         CPUFREQ_RELATION_L);
 592                         __cpufreq_driver_target(policy, freq,
 593                                 CPUFREQ_RELATION_L);
 594                 }
 595         }
 596 }
 597
 598 static void do_dbs_timer(struct work_struct *work)
 599 {
 600         struct cpu_dbs_info_s *dbs_info =
 601                 container_of(work, struct cpu_dbs_info_s, work.work);
 602         unsigned int cpu = dbs_info->cpu;
 603         int sample_type = dbs_info->sample_type;
 604
 605         /* We want all CPUs to do sampling nearly on same jiffy */
 606         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 607
 608         delay -= jiffies % delay;
 609         mutex_lock(&dbs_info->timer_mutex);
 610
 611         /* Common NORMAL_SAMPLE setup */
 612         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
 613         if (!dbs_tuners_ins.powersave_bias ||
 614             sample_type == DBS_NORMAL_SAMPLE) {
 615                 dbs_check_cpu(dbs_info);
 616                 if (dbs_info->freq_lo) {
 617                         /* Setup timer for SUB_SAMPLE */
 618                         dbs_info->sample_type = DBS_SUB_SAMPLE;
 619                         delay = dbs_info->freq_hi_jiffies;
 620                 }
 621         } else {
 622                 __cpufreq_driver_target(dbs_info->cur_policy,
 623                         dbs_info->freq_lo, CPUFREQ_RELATION_H);
 624         }
 625         queue_delayed_work_on(cpu, kondemand_wq, &dbs_info->work, delay);
 626         mutex_unlock(&dbs_info->timer_mutex);
 627 }
 628
 629 static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
 630 {
 631         /* We want all CPUs to do sampling nearly on same jiffy */
 632         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 633         delay -= jiffies % delay;
 634
 635         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
 636         INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer);
 637         queue_delayed_work_on(dbs_info->cpu, kondemand_wq, &dbs_info->work,
 638                 delay);
 639 }
 640
 641 static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
 642 {
 643         cancel_delayed_work_sync(&dbs_info->work);
 644 }
 645
 646 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
 647                                    unsigned int event)
 648 {
 649         unsigned int cpu = policy->cpu;
 650         struct cpu_dbs_info_s *this_dbs_info;
 651         unsigned int j;
 652         int rc;
 653
 654         this_dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
 655
 656         switch (event) {
 657         case CPUFREQ_GOV_START:
 658                 if ((!cpu_online(cpu)) || (!policy->cur))
 659                         return -EINVAL;
 660
 661                 mutex_lock(&dbs_mutex);
 662
 663                 rc = sysfs_create_group(&policy->kobj, &dbs_attr_group_old);
 664                 if (rc) {
 665                         mutex_unlock(&dbs_mutex);
 666                         return rc;
 667                 }
 668
 669                 dbs_enable++;
 670                 for_each_cpu(j, policy->cpus) {
 671                         struct cpu_dbs_info_s *j_dbs_info;
 672                         j_dbs_info = &per_cpu(od_cpu_dbs_info, j);
 673                         j_dbs_info->cur_policy = policy;
 674
 675                         j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
 676                                                 &j_dbs_info->prev_cpu_wall);
 677                         if (dbs_tuners_ins.ignore_nice) {
 678                                 j_dbs_info->prev_cpu_nice =
 679                                                 kstat_cpu(j).cpustat.nice;
 680                         }
 681                 }
 682                 this_dbs_info->cpu = cpu;
 683                 ondemand_powersave_bias_init_cpu(cpu);
 684                 /*
 685                  * Start the timerschedule work, when this governor
 686                  * is used for first time
 687                  */
 688                 if (dbs_enable == 1) {
 689                         unsigned int latency;
 690
 691                         rc = sysfs_create_group(cpufreq_global_kobject,
 692                                                 &dbs_attr_group);
 693                         if (rc) {
 694                                 mutex_unlock(&dbs_mutex);
 695                                 return rc;
 696                         }
 697
 698                         /* policy latency is in nS. Convert it to uS first */
 699                         latency = policy->cpuinfo.transition_latency / 1000;
 700                         if (latency == 0)
 701                                 latency = 1;
 702                         /* Bring kernel and HW constraints together */
 703                         min_sampling_rate = max(min_sampling_rate,
 704                                         MIN_LATENCY_MULTIPLIER * latency);
 705                         dbs_tuners_ins.sampling_rate =
 706                                 max(min_sampling_rate,
 707                                     latency * LATENCY_MULTIPLIER);
 708                 }
 709                 mutex_unlock(&dbs_mutex);
 710
 711                 mutex_init(&this_dbs_info->timer_mutex);
 712                 dbs_timer_init(this_dbs_info);
 713                 break;
 714
 715         case CPUFREQ_GOV_STOP:
 716                 dbs_timer_exit(this_dbs_info);
 717
 718                 mutex_lock(&dbs_mutex);
 719                 sysfs_remove_group(&policy->kobj, &dbs_attr_group_old);
 720                 mutex_destroy(&this_dbs_info->timer_mutex);
 721                 dbs_enable--;
 722                 mutex_unlock(&dbs_mutex);
 723                 if (!dbs_enable)
 724                         sysfs_remove_group(cpufreq_global_kobject,
 725                                            &dbs_attr_group);
 726
 727                 break;
 728
 729         case CPUFREQ_GOV_LIMITS:
 730                 mutex_lock(&this_dbs_info->timer_mutex);
 731                 if (policy->max < this_dbs_info->cur_policy->cur)
 732                         __cpufreq_driver_target(this_dbs_info->cur_policy,
 733                                 policy->max, CPUFREQ_RELATION_H);
 734                 else if (policy->min > this_dbs_info->cur_policy->cur)
 735                         __cpufreq_driver_target(this_dbs_info->cur_policy,
 736                                 policy->min, CPUFREQ_RELATION_L);
 737                 mutex_unlock(&this_dbs_info->timer_mutex);
 738                 break;
 739         }
 740         return 0;
 741 }
 742
 743 static int __init cpufreq_gov_dbs_init(void)
 744 {
 745         int err;
 746         cputime64_t wall;
 747         u64 idle_time;
 748         int cpu = get_cpu();
 749
 750         idle_time = get_cpu_idle_time_us(cpu, &wall);
 751         put_cpu();
 752         if (idle_time != -1ULL) {
 753                 /* Idle micro accounting is supported. Use finer thresholds */
 754                 dbs_tuners_ins.up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
 755                 dbs_tuners_ins.down_differential =
 756                                         MICRO_FREQUENCY_DOWN_DIFFERENTIAL;
 757                 /*
 758                  * In no_hz/micro accounting case we set the minimum frequency
 759                  * not depending on HZ, but fixed (very low). The deferred
 760                  * timer might skip some samples if idle/sleeping as needed.
 761                 */
 762                 min_sampling_rate = MICRO_FREQUENCY_MIN_SAMPLE_RATE;
 763         } else {
 764                 /* For correct statistics, we need 10 ticks for each measure */
 765                 min_sampling_rate =
 766                         MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10);
 767         }
 768
 769         kondemand_wq = create_workqueue("kondemand");
 770         if (!kondemand_wq) {
 771                 printk(KERN_ERR "Creation of kondemand failed\n");
 772                 return -EFAULT;
 773         }
 774         err = cpufreq_register_governor(&cpufreq_gov_ondemand);
 775         if (err)
 776                 destroy_workqueue(kondemand_wq);
 777
 778         return err;
 779 }
 780
 781 static void __exit cpufreq_gov_dbs_exit(void)
 782 {
 783         cpufreq_unregister_governor(&cpufreq_gov_ondemand);
 784         destroy_workqueue(kondemand_wq);
 785 }
 786
 787
 788 MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
 789 MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
 790 MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
 791         "Low Latency Frequency Transition capable processors");
 792 MODULE_LICENSE("GPL");
 793
 794 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
 795 fs_initcall(cpufreq_gov_dbs_init);
 796 #else
 797 module_init(cpufreq_gov_dbs_init);
 798 #endif
 799 module_exit(cpufreq_gov_dbs_exit);