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 #ifdef CONFIG_ARCH_RK29
  34 #define MICRO_FREQUENCY_DOWN_DIFFERENTIAL       (10)
  35 #define MICRO_FREQUENCY_UP_THRESHOLD            (80)
  36 #else
  37 #define MICRO_FREQUENCY_DOWN_DIFFERENTIAL       (3)
  38 #define MICRO_FREQUENCY_UP_THRESHOLD            (95)
  39 #endif
  40 #define MICRO_FREQUENCY_MIN_SAMPLE_RATE         (10000)
  41 #define MIN_FREQUENCY_UP_THRESHOLD              (11)
  42 #define MAX_FREQUENCY_UP_THRESHOLD              (100)
  43
  44 /*
  45  * The polling frequency of this governor depends on the capability of
  46  * the processor. Default polling frequency is 1000 times the transition
  47  * latency of the processor. The governor will work on any processor with
  48  * transition latency <= 10mS, using appropriate sampling
  49  * rate.
  50  * For CPUs with transition latency > 10mS (mostly drivers with CPUFREQ_ETERNAL)
  51  * this governor will not work.
  52  * All times here are in uS.
  53  */
  54 #define MIN_SAMPLING_RATE_RATIO                 (2)
  55
  56 static unsigned int min_sampling_rate;
  57
  58 #define LATENCY_MULTIPLIER                      (1000)
  59 #define MIN_LATENCY_MULTIPLIER                  (100)
  60 #define TRANSITION_LATENCY_LIMIT                (10 * 1000 * 1000)
  61
  62 static void do_dbs_timer(struct work_struct *work);
  63 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
  64                                 unsigned int event);
  65
  66 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
  67 static
  68 #endif
  69 struct cpufreq_governor cpufreq_gov_ondemand = {
  70        .name                   = "ondemand",
  71        .governor               = cpufreq_governor_dbs,
  72        .max_transition_latency = TRANSITION_LATENCY_LIMIT,
  73        .owner                  = THIS_MODULE,
  74 };
  75
  76 /* Sampling types */
  77 enum {DBS_NORMAL_SAMPLE, DBS_SUB_SAMPLE};
  78
  79 struct cpu_dbs_info_s {
  80         cputime64_t prev_cpu_idle;
  81         cputime64_t prev_cpu_iowait;
  82         cputime64_t prev_cpu_wall;
  83         cputime64_t prev_cpu_nice;
  84         struct cpufreq_policy *cur_policy;
  85         struct delayed_work work;
  86         struct cpufreq_frequency_table *freq_table;
  87         unsigned int freq_lo;
  88         unsigned int freq_lo_jiffies;
  89         unsigned int freq_hi_jiffies;
  90         int cpu;
  91         unsigned int sample_type:1;
  92         /*
  93          * percpu mutex that serializes governor limit change with
  94          * do_dbs_timer invocation. We do not want do_dbs_timer to run
  95          * when user is changing the governor or limits.
  96          */
  97         struct mutex timer_mutex;
  98 };
  99 static DEFINE_PER_CPU(struct cpu_dbs_info_s, od_cpu_dbs_info);
 100
 101 static unsigned int dbs_enable; /* number of CPUs using this policy */
 102
 103 /*
 104  * dbs_mutex protects data in dbs_tuners_ins from concurrent changes on
 105  * different CPUs. It protects dbs_enable in governor start/stop.
 106  */
 107 static DEFINE_MUTEX(dbs_mutex);
 108
 109 static struct workqueue_struct  *kondemand_wq;
 110
 111 static struct dbs_tuners {
 112         unsigned int sampling_rate;
 113         unsigned int up_threshold;
 114         unsigned int down_differential;
 115         unsigned int ignore_nice;
 116         unsigned int powersave_bias;
 117         unsigned int io_is_busy;
 118 } dbs_tuners_ins = {
 119         .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
 120         .down_differential = DEF_FREQUENCY_DOWN_DIFFERENTIAL,
 121         .ignore_nice = 0,
 122         .powersave_bias = 0,
 123 };
 124
 125 static inline cputime64_t get_cpu_idle_time_jiffy(unsigned int cpu,
 126                                                         cputime64_t *wall)
 127 {
 128         cputime64_t idle_time;
 129         cputime64_t cur_wall_time;
 130         cputime64_t busy_time;
 131
 132         cur_wall_time = jiffies64_to_cputime64(get_jiffies_64());
 133         busy_time = cputime64_add(kstat_cpu(cpu).cpustat.user,
 134                         kstat_cpu(cpu).cpustat.system);
 135
 136         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.irq);
 137         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.softirq);
 138         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.steal);
 139         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.nice);
 140
 141         idle_time = cputime64_sub(cur_wall_time, busy_time);
 142         if (wall)
 143                 *wall = (cputime64_t)jiffies_to_usecs(cur_wall_time);
 144
 145         return (cputime64_t)jiffies_to_usecs(idle_time);
 146 }
 147
 148 static inline cputime64_t get_cpu_idle_time(unsigned int cpu, cputime64_t *wall)
 149 {
 150         u64 idle_time = get_cpu_idle_time_us(cpu, wall);
 151
 152         if (idle_time == -1ULL)
 153                 return get_cpu_idle_time_jiffy(cpu, wall);
 154
 155         return idle_time;
 156 }
 157
 158 static inline cputime64_t get_cpu_iowait_time(unsigned int cpu, cputime64_t *wall)
 159 {
 160         u64 iowait_time = get_cpu_iowait_time_us(cpu, wall);
 161
 162         if (iowait_time == -1ULL)
 163                 return 0;
 164
 165         return iowait_time;
 166 }
 167
 168 /*
 169  * Find right freq to be set now with powersave_bias on.
 170  * Returns the freq_hi to be used right now and will set freq_hi_jiffies,
 171  * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs.
 172  */
 173 static unsigned int powersave_bias_target(struct cpufreq_policy *policy,
 174                                           unsigned int freq_next,
 175                                           unsigned int relation)
 176 {
 177         unsigned int freq_req, freq_reduc, freq_avg;
 178         unsigned int freq_hi, freq_lo;
 179         unsigned int index = 0;
 180         unsigned int jiffies_total, jiffies_hi, jiffies_lo;
 181         struct cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
 182                                                    policy->cpu);
 183
 184         if (!dbs_info->freq_table) {
 185                 dbs_info->freq_lo = 0;
 186                 dbs_info->freq_lo_jiffies = 0;
 187                 return freq_next;
 188         }
 189
 190         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next,
 191                         relation, &index);
 192         freq_req = dbs_info->freq_table[index].frequency;
 193         freq_reduc = freq_req * dbs_tuners_ins.powersave_bias / 1000;
 194         freq_avg = freq_req - freq_reduc;
 195
 196         /* Find freq bounds for freq_avg in freq_table */
 197         index = 0;
 198         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
 199                         CPUFREQ_RELATION_H, &index);
 200         freq_lo = dbs_info->freq_table[index].frequency;
 201         index = 0;
 202         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
 203                         CPUFREQ_RELATION_L, &index);
 204         freq_hi = dbs_info->freq_table[index].frequency;
 205
 206         /* Find out how long we have to be in hi and lo freqs */
 207         if (freq_hi == freq_lo) {
 208                 dbs_info->freq_lo = 0;
 209                 dbs_info->freq_lo_jiffies = 0;
 210                 return freq_lo;
 211         }
 212         jiffies_total = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 213         jiffies_hi = (freq_avg - freq_lo) * jiffies_total;
 214         jiffies_hi += ((freq_hi - freq_lo) / 2);
 215         jiffies_hi /= (freq_hi - freq_lo);
 216         jiffies_lo = jiffies_total - jiffies_hi;
 217         dbs_info->freq_lo = freq_lo;
 218         dbs_info->freq_lo_jiffies = jiffies_lo;
 219         dbs_info->freq_hi_jiffies = jiffies_hi;
 220         return freq_hi;
 221 }
 222
 223 static void ondemand_powersave_bias_init_cpu(int cpu)
 224 {
 225         struct cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
 226         dbs_info->freq_table = cpufreq_frequency_get_table(cpu);
 227         dbs_info->freq_lo = 0;
 228 }
 229
 230 static void ondemand_powersave_bias_init(void)
 231 {
 232         int i;
 233         for_each_online_cpu(i) {
 234                 ondemand_powersave_bias_init_cpu(i);
 235         }
 236 }
 237
 238 /************************** sysfs interface ************************/
 239
 240 static ssize_t show_sampling_rate_max(struct kobject *kobj,
 241                                       struct attribute *attr, char *buf)
 242 {
 243         printk_once(KERN_INFO "CPUFREQ: ondemand sampling_rate_max "
 244                "sysfs file is deprecated - used by: %s\n", current->comm);
 245         return sprintf(buf, "%u\n", -1U);
 246 }
 247
 248 static ssize_t show_sampling_rate_min(struct kobject *kobj,
 249                                       struct attribute *attr, char *buf)
 250 {
 251         return sprintf(buf, "%u\n", min_sampling_rate);
 252 }
 253
 254 define_one_global_ro(sampling_rate_max);
 255 define_one_global_ro(sampling_rate_min);
 256
 257 /* cpufreq_ondemand Governor Tunables */
 258 #define show_one(file_name, object)                                     \
 259 static ssize_t show_##file_name                                         \
 260 (struct kobject *kobj, struct attribute *attr, char *buf)              \
 261 {                                                                       \
 262         return sprintf(buf, "%u\n", dbs_tuners_ins.object);             \
 263 }
 264 show_one(sampling_rate, sampling_rate);
 265 show_one(io_is_busy, io_is_busy);
 266 show_one(up_threshold, up_threshold);
 267 show_one(ignore_nice_load, ignore_nice);
 268 show_one(powersave_bias, powersave_bias);
 269
 270 /*** delete after deprecation time ***/
 271
 272 #define DEPRECATION_MSG(file_name)                                      \
 273         printk_once(KERN_INFO "CPUFREQ: Per core ondemand sysfs "       \
 274                     "interface is deprecated - " #file_name "\n");
 275
 276 #define show_one_old(file_name)                                         \
 277 static ssize_t show_##file_name##_old                                   \
 278 (struct cpufreq_policy *unused, char *buf)                              \
 279 {                                                                       \
 280         printk_once(KERN_INFO "CPUFREQ: Per core ondemand sysfs "       \
 281                     "interface is deprecated - " #file_name "\n");      \
 282         return show_##file_name(NULL, NULL, buf);                       \
 283 }
 284 show_one_old(sampling_rate);
 285 show_one_old(up_threshold);
 286 show_one_old(ignore_nice_load);
 287 show_one_old(powersave_bias);
 288 show_one_old(sampling_rate_min);
 289 show_one_old(sampling_rate_max);
 290
 291 cpufreq_freq_attr_ro_old(sampling_rate_min);
 292 cpufreq_freq_attr_ro_old(sampling_rate_max);
 293
 294 /*** delete after deprecation time ***/
 295
 296 static ssize_t store_sampling_rate(struct kobject *a, struct attribute *b,
 297                                    const char *buf, size_t count)
 298 {
 299         unsigned int input;
 300         int ret;
 301         ret = sscanf(buf, "%u", &input);
 302         if (ret != 1)
 303                 return -EINVAL;
 304
 305         mutex_lock(&dbs_mutex);
 306         dbs_tuners_ins.sampling_rate = max(input, min_sampling_rate);
 307         mutex_unlock(&dbs_mutex);
 308
 309         return count;
 310 }
 311
 312 static ssize_t store_io_is_busy(struct kobject *a, struct attribute *b,
 313                                    const char *buf, size_t count)
 314 {
 315         unsigned int input;
 316         int ret;
 317
 318         ret = sscanf(buf, "%u", &input);
 319         if (ret != 1)
 320                 return -EINVAL;
 321
 322         mutex_lock(&dbs_mutex);
 323         dbs_tuners_ins.io_is_busy = !!input;
 324         mutex_unlock(&dbs_mutex);
 325
 326         return count;
 327 }
 328
 329 static ssize_t store_up_threshold(struct kobject *a, struct attribute *b,
 330                                   const char *buf, size_t count)
 331 {
 332         unsigned int input;
 333         int ret;
 334         ret = sscanf(buf, "%u", &input);
 335
 336         if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
 337                         input < MIN_FREQUENCY_UP_THRESHOLD) {
 338                 return -EINVAL;
 339         }
 340
 341         mutex_lock(&dbs_mutex);
 342         dbs_tuners_ins.up_threshold = input;
 343         mutex_unlock(&dbs_mutex);
 344
 345         return count;
 346 }
 347
 348 static ssize_t store_ignore_nice_load(struct kobject *a, struct attribute *b,
 349                                       const char *buf, size_t count)
 350 {
 351         unsigned int input;
 352         int ret;
 353
 354         unsigned int j;
 355
 356         ret = sscanf(buf, "%u", &input);
 357         if (ret != 1)
 358                 return -EINVAL;
 359
 360         if (input > 1)
 361                 input = 1;
 362
 363         mutex_lock(&dbs_mutex);
 364         if (input == dbs_tuners_ins.ignore_nice) { /* nothing to do */
 365                 mutex_unlock(&dbs_mutex);
 366                 return count;
 367         }
 368         dbs_tuners_ins.ignore_nice = input;
 369
 370         /* we need to re-evaluate prev_cpu_idle */
 371         for_each_online_cpu(j) {
 372                 struct cpu_dbs_info_s *dbs_info;
 373                 dbs_info = &per_cpu(od_cpu_dbs_info, j);
 374                 dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
 375                                                 &dbs_info->prev_cpu_wall);
 376                 if (dbs_tuners_ins.ignore_nice)
 377                         dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
 378
 379         }
 380         mutex_unlock(&dbs_mutex);
 381
 382         return count;
 383 }
 384
 385 static ssize_t store_powersave_bias(struct kobject *a, struct attribute *b,
 386                                     const char *buf, size_t count)
 387 {
 388         unsigned int input;
 389         int ret;
 390         ret = sscanf(buf, "%u", &input);
 391
 392         if (ret != 1)
 393                 return -EINVAL;
 394
 395         if (input > 1000)
 396                 input = 1000;
 397
 398         mutex_lock(&dbs_mutex);
 399         dbs_tuners_ins.powersave_bias = input;
 400         ondemand_powersave_bias_init();
 401         mutex_unlock(&dbs_mutex);
 402
 403         return count;
 404 }
 405
 406 define_one_global_rw(sampling_rate);
 407 define_one_global_rw(io_is_busy);
 408 define_one_global_rw(up_threshold);
 409 define_one_global_rw(ignore_nice_load);
 410 define_one_global_rw(powersave_bias);
 411
 412 static struct attribute *dbs_attributes[] = {
 413         &sampling_rate_max.attr,
 414         &sampling_rate_min.attr,
 415         &sampling_rate.attr,
 416         &up_threshold.attr,
 417         &ignore_nice_load.attr,
 418         &powersave_bias.attr,
 419         &io_is_busy.attr,
 420         NULL
 421 };
 422
 423 static struct attribute_group dbs_attr_group = {
 424         .attrs = dbs_attributes,
 425         .name = "ondemand",
 426 };
 427
 428 /*** delete after deprecation time ***/
 429
 430 #define write_one_old(file_name)                                        \
 431 static ssize_t store_##file_name##_old                                  \
 432 (struct cpufreq_policy *unused, const char *buf, size_t count)          \
 433 {                                                                       \
 434        printk_once(KERN_INFO "CPUFREQ: Per core ondemand sysfs "        \
 435                    "interface is deprecated - " #file_name "\n");       \
 436        return store_##file_name(NULL, NULL, buf, count);                \
 437 }
 438 write_one_old(sampling_rate);
 439 write_one_old(up_threshold);
 440 write_one_old(ignore_nice_load);
 441 write_one_old(powersave_bias);
 442
 443 cpufreq_freq_attr_rw_old(sampling_rate);
 444 cpufreq_freq_attr_rw_old(up_threshold);
 445 cpufreq_freq_attr_rw_old(ignore_nice_load);
 446 cpufreq_freq_attr_rw_old(powersave_bias);
 447
 448 static struct attribute *dbs_attributes_old[] = {
 449        &sampling_rate_max_old.attr,
 450        &sampling_rate_min_old.attr,
 451        &sampling_rate_old.attr,
 452        &up_threshold_old.attr,
 453        &ignore_nice_load_old.attr,
 454        &powersave_bias_old.attr,
 455        NULL
 456 };
 457
 458 static struct attribute_group dbs_attr_group_old = {
 459        .attrs = dbs_attributes_old,
 460        .name = "ondemand",
 461 };
 462
 463 /*** delete after deprecation time ***/
 464
 465 /************************** sysfs end ************************/
 466
 467 static void dbs_freq_increase(struct cpufreq_policy *p, unsigned int freq)
 468 {
 469         if (dbs_tuners_ins.powersave_bias)
 470                 freq = powersave_bias_target(p, freq, CPUFREQ_RELATION_H);
 471         else if (p->cur == p->max)
 472                 return;
 473
 474         __cpufreq_driver_target(p, freq, dbs_tuners_ins.powersave_bias ?
 475                         CPUFREQ_RELATION_L : CPUFREQ_RELATION_H);
 476 }
 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                 dbs_freq_increase(policy, policy->max);
 571                 return;
 572         }
 573
 574         /* Check for frequency decrease */
 575         /* if we cannot reduce the frequency anymore, break out early */
 576         if (policy->cur == policy->min)
 577                 return;
 578
 579         /*
 580          * The optimal frequency is the frequency that is the lowest that
 581          * can support the current CPU usage without triggering the up
 582          * policy. To be safe, we focus 10 points under the threshold.
 583          */
 584         if (max_load_freq <
 585             (dbs_tuners_ins.up_threshold - dbs_tuners_ins.down_differential) *
 586              policy->cur) {
 587                 unsigned int freq_next;
 588                 freq_next = max_load_freq /
 589                                 (dbs_tuners_ins.up_threshold -
 590                                  dbs_tuners_ins.down_differential);
 591
 592                 if (freq_next < policy->min)
 593                         freq_next = policy->min;
 594
 595                 if (!dbs_tuners_ins.powersave_bias) {
 596                         __cpufreq_driver_target(policy, freq_next,
 597                                         CPUFREQ_RELATION_L);
 598                 } else {
 599                         int freq = powersave_bias_target(policy, freq_next,
 600                                         CPUFREQ_RELATION_L);
 601                         __cpufreq_driver_target(policy, freq,
 602                                 CPUFREQ_RELATION_L);
 603                 }
 604         }
 605 }
 606
 607 static void do_dbs_timer(struct work_struct *work)
 608 {
 609         struct cpu_dbs_info_s *dbs_info =
 610                 container_of(work, struct cpu_dbs_info_s, work.work);
 611         unsigned int cpu = dbs_info->cpu;
 612         int sample_type = dbs_info->sample_type;
 613
 614         /* We want all CPUs to do sampling nearly on same jiffy */
 615         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 616
 617         if (num_online_cpus() > 1)
 618                 delay -= jiffies % delay;
 619
 620         mutex_lock(&dbs_info->timer_mutex);
 621
 622         /* Common NORMAL_SAMPLE setup */
 623         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
 624         if (!dbs_tuners_ins.powersave_bias ||
 625             sample_type == DBS_NORMAL_SAMPLE) {
 626                 dbs_check_cpu(dbs_info);
 627                 if (dbs_info->freq_lo) {
 628                         /* Setup timer for SUB_SAMPLE */
 629                         dbs_info->sample_type = DBS_SUB_SAMPLE;
 630                         delay = dbs_info->freq_hi_jiffies;
 631                 }
 632         } else {
 633                 __cpufreq_driver_target(dbs_info->cur_policy,
 634                         dbs_info->freq_lo, CPUFREQ_RELATION_H);
 635         }
 636         queue_delayed_work_on(cpu, kondemand_wq, &dbs_info->work, delay);
 637         mutex_unlock(&dbs_info->timer_mutex);
 638 }
 639
 640 static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
 641 {
 642         /* We want all CPUs to do sampling nearly on same jiffy */
 643         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 644
 645         if (num_online_cpus() > 1)
 646                 delay -= jiffies % delay;
 647
 648         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
 649         INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer);
 650         queue_delayed_work_on(dbs_info->cpu, kondemand_wq, &dbs_info->work,
 651                 delay);
 652 }
 653
 654 static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
 655 {
 656         cancel_delayed_work_sync(&dbs_info->work);
 657 }
 658
 659 /*
 660  * Not all CPUs want IO time to be accounted as busy; this dependson how
 661  * efficient idling at a higher frequency/voltage is.
 662  * Pavel Machek says this is not so for various generations of AMD and old
 663  * Intel systems.
 664  * Mike Chan (androidlcom) calis this is also not true for ARM.
 665  * Because of this, whitelist specific known (series) of CPUs by default, and
 666  * leave all others up to the user.
 667  */
 668 static int should_io_be_busy(void)
 669 {
 670 #if defined(CONFIG_X86)
 671         /*
 672          * For Intel, Core 2 (model 15) andl later have an efficient idle.
 673          */
 674         if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
 675             boot_cpu_data.x86 == 6 &&
 676             boot_cpu_data.x86_model >= 15)
 677                 return 1;
 678 #endif
 679         return 0;
 680 }
 681
 682 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
 683                                    unsigned int event)
 684 {
 685         unsigned int cpu = policy->cpu;
 686         struct cpu_dbs_info_s *this_dbs_info;
 687         unsigned int j;
 688         int rc;
 689
 690         this_dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
 691
 692         switch (event) {
 693         case CPUFREQ_GOV_START:
 694                 if ((!cpu_online(cpu)) || (!policy->cur))
 695                         return -EINVAL;
 696
 697                 mutex_lock(&dbs_mutex);
 698
 699                 rc = sysfs_create_group(&policy->kobj, &dbs_attr_group_old);
 700                 if (rc) {
 701                         mutex_unlock(&dbs_mutex);
 702                         return rc;
 703                 }
 704
 705                 dbs_enable++;
 706                 for_each_cpu(j, policy->cpus) {
 707                         struct cpu_dbs_info_s *j_dbs_info;
 708                         j_dbs_info = &per_cpu(od_cpu_dbs_info, j);
 709                         j_dbs_info->cur_policy = policy;
 710
 711                         j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
 712                                                 &j_dbs_info->prev_cpu_wall);
 713                         if (dbs_tuners_ins.ignore_nice) {
 714                                 j_dbs_info->prev_cpu_nice =
 715                                                 kstat_cpu(j).cpustat.nice;
 716                         }
 717                 }
 718                 this_dbs_info->cpu = cpu;
 719                 ondemand_powersave_bias_init_cpu(cpu);
 720                 /*
 721                  * Start the timerschedule work, when this governor
 722                  * is used for first time
 723                  */
 724                 if (dbs_enable == 1) {
 725                         unsigned int latency;
 726
 727                         rc = sysfs_create_group(cpufreq_global_kobject,
 728                                                 &dbs_attr_group);
 729                         if (rc) {
 730                                 mutex_unlock(&dbs_mutex);
 731                                 return rc;
 732                         }
 733
 734                         /* policy latency is in nS. Convert it to uS first */
 735                         latency = policy->cpuinfo.transition_latency / 1000;
 736                         if (latency == 0)
 737                                 latency = 1;
 738                         /* Bring kernel and HW constraints together */
 739                         min_sampling_rate = max(min_sampling_rate,
 740                                         MIN_LATENCY_MULTIPLIER * latency);
 741                         dbs_tuners_ins.sampling_rate =
 742                                 max(min_sampling_rate,
 743                                     latency * LATENCY_MULTIPLIER);
 744                         dbs_tuners_ins.io_is_busy = should_io_be_busy();
 745                 }
 746                 mutex_unlock(&dbs_mutex);
 747
 748                 mutex_init(&this_dbs_info->timer_mutex);
 749                 dbs_timer_init(this_dbs_info);
 750                 break;
 751
 752         case CPUFREQ_GOV_STOP:
 753                 dbs_timer_exit(this_dbs_info);
 754
 755                 mutex_lock(&dbs_mutex);
 756                 sysfs_remove_group(&policy->kobj, &dbs_attr_group_old);
 757                 mutex_destroy(&this_dbs_info->timer_mutex);
 758                 dbs_enable--;
 759                 mutex_unlock(&dbs_mutex);
 760                 if (!dbs_enable)
 761                         sysfs_remove_group(cpufreq_global_kobject,
 762                                            &dbs_attr_group);
 763
 764                 break;
 765
 766         case CPUFREQ_GOV_LIMITS:
 767                 mutex_lock(&this_dbs_info->timer_mutex);
 768                 if (policy->max < this_dbs_info->cur_policy->cur)
 769                         __cpufreq_driver_target(this_dbs_info->cur_policy,
 770                                 policy->max, CPUFREQ_RELATION_H);
 771                 else if (policy->min > this_dbs_info->cur_policy->cur)
 772                         __cpufreq_driver_target(this_dbs_info->cur_policy,
 773                                 policy->min, CPUFREQ_RELATION_L);
 774                 mutex_unlock(&this_dbs_info->timer_mutex);
 775                 break;
 776         }
 777         return 0;
 778 }
 779
 780 static int __init cpufreq_gov_dbs_init(void)
 781 {
 782         int err;
 783         cputime64_t wall;
 784         u64 idle_time;
 785         int cpu = get_cpu();
 786
 787         idle_time = get_cpu_idle_time_us(cpu, &wall);
 788         put_cpu();
 789         if (idle_time != -1ULL) {
 790                 /* Idle micro accounting is supported. Use finer thresholds */
 791                 dbs_tuners_ins.up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
 792                 dbs_tuners_ins.down_differential =
 793                                         MICRO_FREQUENCY_DOWN_DIFFERENTIAL;
 794                 /*
 795                  * In no_hz/micro accounting case we set the minimum frequency
 796                  * not depending on HZ, but fixed (very low). The deferred
 797                  * timer might skip some samples if idle/sleeping as needed.
 798                 */
 799                 min_sampling_rate = MICRO_FREQUENCY_MIN_SAMPLE_RATE;
 800         } else {
 801                 /* For correct statistics, we need 10 ticks for each measure */
 802                 min_sampling_rate =
 803                         MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10);
 804         }
 805
 806         kondemand_wq = create_workqueue("kondemand");
 807         if (!kondemand_wq) {
 808                 printk(KERN_ERR "Creation of kondemand failed\n");
 809                 return -EFAULT;
 810         }
 811         err = cpufreq_register_governor(&cpufreq_gov_ondemand);
 812         if (err)
 813                 destroy_workqueue(kondemand_wq);
 814
 815         return err;
 816 }
 817
 818 static void __exit cpufreq_gov_dbs_exit(void)
 819 {
 820         cpufreq_unregister_governor(&cpufreq_gov_ondemand);
 821         destroy_workqueue(kondemand_wq);
 822 }
 823
 824
 825 MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
 826 MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
 827 MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
 828         "Low Latency Frequency Transition capable processors");
 829 MODULE_LICENSE("GPL");
 830
 831 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
 832 fs_initcall(cpufreq_gov_dbs_init);
 833 #else
 834 module_init(cpufreq_gov_dbs_init);
 835 #endif
 836 module_exit(cpufreq_gov_dbs_exit);