drivers/cpufreq/cpufreq_conservative.c

   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)  2004 Alexander Clouter <alex-kernel@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
  14 #include <linux/kernel.h>
  15 #include <linux/module.h>
  16 #include <linux/smp.h>
  17 #include <linux/init.h>
  18 #include <linux/interrupt.h>
  19 #include <linux/ctype.h>
  20 #include <linux/cpufreq.h>
  21 #include <linux/sysctl.h>
  22 #include <linux/types.h>
  23 #include <linux/fs.h>
  24 #include <linux/sysfs.h>
  25 #include <linux/sched.h>
  26 #include <linux/kmod.h>
  27 #include <linux/workqueue.h>
  28 #include <linux/jiffies.h>
  29 #include <linux/kernel_stat.h>
  30 #include <linux/percpu.h>
  31 #include <linux/mutex.h>
  32 /*
  33  * dbs is used in this file as a shortform for demandbased switching
  34  * It helps to keep variable names smaller, simpler
  35  */
  36
  37 #define DEF_FREQUENCY_UP_THRESHOLD              (80)
  38 #define DEF_FREQUENCY_DOWN_THRESHOLD            (20)
  39
  40 /*
  41  * The polling frequency of this governor depends on the capability of
  42  * the processor. Default polling frequency is 1000 times the transition
  43  * latency of the processor. The governor will work on any processor with
  44  * transition latency <= 10mS, using appropriate sampling
  45  * rate.
  46  * For CPUs with transition latency > 10mS (mostly drivers with CPUFREQ_ETERNAL)
  47  * this governor will not work.
  48  * All times here are in uS.
  49  */
  50 static unsigned int                             def_sampling_rate;
  51 #define MIN_SAMPLING_RATE_RATIO                 (2)
  52 /* for correct statistics, we need at least 10 ticks between each measure */
  53 #define MIN_STAT_SAMPLING_RATE                  (MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10))
  54 #define MIN_SAMPLING_RATE                       (def_sampling_rate / MIN_SAMPLING_RATE_RATIO)
  55 #define MAX_SAMPLING_RATE                       (500 * def_sampling_rate)
  56 #define DEF_SAMPLING_RATE_LATENCY_MULTIPLIER    (1000)
  57 #define DEF_SAMPLING_DOWN_FACTOR                (1)
  58 #define MAX_SAMPLING_DOWN_FACTOR                (10)
  59 #define TRANSITION_LATENCY_LIMIT                (10 * 1000)
  60
  61 static void do_dbs_timer(void *data);
  62
  63 struct cpu_dbs_info_s {
  64         struct cpufreq_policy   *cur_policy;
  65         unsigned int            prev_cpu_idle_up;
  66         unsigned int            prev_cpu_idle_down;
  67         unsigned int            enable;
  68         unsigned int            down_skip;
  69         unsigned int            requested_freq;
  70 };
  71 static DEFINE_PER_CPU(struct cpu_dbs_info_s, cpu_dbs_info);
  72
  73 static unsigned int dbs_enable; /* number of CPUs using this policy */
  74
  75 /*
  76  * DEADLOCK ALERT! There is a ordering requirement between cpu_hotplug
  77  * lock and dbs_mutex. cpu_hotplug lock should always be held before
  78  * dbs_mutex. If any function that can potentially take cpu_hotplug lock
  79  * (like __cpufreq_driver_target()) is being called with dbs_mutex taken, then
  80  * cpu_hotplug lock should be taken before that. Note that cpu_hotplug lock
  81  * is recursive for the same process. -Venki
  82  */
  83 static DEFINE_MUTEX     (dbs_mutex);
  84 static DECLARE_WORK     (dbs_work, do_dbs_timer, NULL);
  85
  86 struct dbs_tuners {
  87         unsigned int            sampling_rate;
  88         unsigned int            sampling_down_factor;
  89         unsigned int            up_threshold;
  90         unsigned int            down_threshold;
  91         unsigned int            ignore_nice;
  92         unsigned int            freq_step;
  93 };
  94
  95 static struct dbs_tuners dbs_tuners_ins = {
  96         .up_threshold           = DEF_FREQUENCY_UP_THRESHOLD,
  97         .down_threshold         = DEF_FREQUENCY_DOWN_THRESHOLD,
  98         .sampling_down_factor   = DEF_SAMPLING_DOWN_FACTOR,
  99         .ignore_nice            = 0,
 100         .freq_step              = 5,
 101 };
 102
 103 static inline unsigned int get_cpu_idle_time(unsigned int cpu)
 104 {
 105         return  kstat_cpu(cpu).cpustat.idle +
 106                 kstat_cpu(cpu).cpustat.iowait +
 107                 ( dbs_tuners_ins.ignore_nice ?
 108                   kstat_cpu(cpu).cpustat.nice :
 109                   0);
 110 }
 111
 112 /************************** sysfs interface ************************/
 113 static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf)
 114 {
 115         return sprintf (buf, "%u\n", MAX_SAMPLING_RATE);
 116 }
 117
 118 static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf)
 119 {
 120         return sprintf (buf, "%u\n", MIN_SAMPLING_RATE);
 121 }
 122
 123 #define define_one_ro(_name)                                    \
 124 static struct freq_attr _name =                                 \
 125 __ATTR(_name, 0444, show_##_name, NULL)
 126
 127 define_one_ro(sampling_rate_max);
 128 define_one_ro(sampling_rate_min);
 129
 130 /* cpufreq_conservative Governor Tunables */
 131 #define show_one(file_name, object)                                     \
 132 static ssize_t show_##file_name                                         \
 133 (struct cpufreq_policy *unused, char *buf)                              \
 134 {                                                                       \
 135         return sprintf(buf, "%u\n", dbs_tuners_ins.object);             \
 136 }
 137 show_one(sampling_rate, sampling_rate);
 138 show_one(sampling_down_factor, sampling_down_factor);
 139 show_one(up_threshold, up_threshold);
 140 show_one(down_threshold, down_threshold);
 141 show_one(ignore_nice_load, ignore_nice);
 142 show_one(freq_step, freq_step);
 143
 144 static ssize_t store_sampling_down_factor(struct cpufreq_policy *unused,
 145                 const char *buf, size_t count)
 146 {
 147         unsigned int input;
 148         int ret;
 149         ret = sscanf (buf, "%u", &input);
 150         if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
 151                 return -EINVAL;
 152
 153         mutex_lock(&dbs_mutex);
 154         dbs_tuners_ins.sampling_down_factor = input;
 155         mutex_unlock(&dbs_mutex);
 156
 157         return count;
 158 }
 159
 160 static ssize_t store_sampling_rate(struct cpufreq_policy *unused,
 161                 const char *buf, size_t count)
 162 {
 163         unsigned int input;
 164         int ret;
 165         ret = sscanf (buf, "%u", &input);
 166
 167         mutex_lock(&dbs_mutex);
 168         if (ret != 1 || input > MAX_SAMPLING_RATE || input < MIN_SAMPLING_RATE) {
 169                 mutex_unlock(&dbs_mutex);
 170                 return -EINVAL;
 171         }
 172
 173         dbs_tuners_ins.sampling_rate = input;
 174         mutex_unlock(&dbs_mutex);
 175
 176         return count;
 177 }
 178
 179 static ssize_t store_up_threshold(struct cpufreq_policy *unused,
 180                 const char *buf, size_t count)
 181 {
 182         unsigned int input;
 183         int ret;
 184         ret = sscanf (buf, "%u", &input);
 185
 186         mutex_lock(&dbs_mutex);
 187         if (ret != 1 || input > 100 || input <= dbs_tuners_ins.down_threshold) {
 188                 mutex_unlock(&dbs_mutex);
 189                 return -EINVAL;
 190         }
 191
 192         dbs_tuners_ins.up_threshold = input;
 193         mutex_unlock(&dbs_mutex);
 194
 195         return count;
 196 }
 197
 198 static ssize_t store_down_threshold(struct cpufreq_policy *unused,
 199                 const char *buf, size_t count)
 200 {
 201         unsigned int input;
 202         int ret;
 203         ret = sscanf (buf, "%u", &input);
 204
 205         mutex_lock(&dbs_mutex);
 206         if (ret != 1 || input > 100 || input >= dbs_tuners_ins.up_threshold) {
 207                 mutex_unlock(&dbs_mutex);
 208                 return -EINVAL;
 209         }
 210
 211         dbs_tuners_ins.down_threshold = input;
 212         mutex_unlock(&dbs_mutex);
 213
 214         return count;
 215 }
 216
 217 static ssize_t store_ignore_nice_load(struct cpufreq_policy *policy,
 218                 const char *buf, size_t count)
 219 {
 220         unsigned int input;
 221         int ret;
 222
 223         unsigned int j;
 224
 225         ret = sscanf (buf, "%u", &input);
 226         if ( ret != 1 )
 227                 return -EINVAL;
 228
 229         if ( input > 1 )
 230                 input = 1;
 231
 232         mutex_lock(&dbs_mutex);
 233         if ( input == dbs_tuners_ins.ignore_nice ) { /* nothing to do */
 234                 mutex_unlock(&dbs_mutex);
 235                 return count;
 236         }
 237         dbs_tuners_ins.ignore_nice = input;
 238
 239         /* we need to re-evaluate prev_cpu_idle_up and prev_cpu_idle_down */
 240         for_each_online_cpu(j) {
 241                 struct cpu_dbs_info_s *j_dbs_info;
 242                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
 243                 j_dbs_info->prev_cpu_idle_up = get_cpu_idle_time(j);
 244                 j_dbs_info->prev_cpu_idle_down = j_dbs_info->prev_cpu_idle_up;
 245         }
 246         mutex_unlock(&dbs_mutex);
 247
 248         return count;
 249 }
 250
 251 static ssize_t store_freq_step(struct cpufreq_policy *policy,
 252                 const char *buf, size_t count)
 253 {
 254         unsigned int input;
 255         int ret;
 256
 257         ret = sscanf (buf, "%u", &input);
 258
 259         if ( ret != 1 )
 260                 return -EINVAL;
 261
 262         if ( input > 100 )
 263                 input = 100;
 264
 265         /* no need to test here if freq_step is zero as the user might actually
 266          * want this, they would be crazy though :) */
 267         mutex_lock(&dbs_mutex);
 268         dbs_tuners_ins.freq_step = input;
 269         mutex_unlock(&dbs_mutex);
 270
 271         return count;
 272 }
 273
 274 #define define_one_rw(_name) \
 275 static struct freq_attr _name = \
 276 __ATTR(_name, 0644, show_##_name, store_##_name)
 277
 278 define_one_rw(sampling_rate);
 279 define_one_rw(sampling_down_factor);
 280 define_one_rw(up_threshold);
 281 define_one_rw(down_threshold);
 282 define_one_rw(ignore_nice_load);
 283 define_one_rw(freq_step);
 284
 285 static struct attribute * dbs_attributes[] = {
 286         &sampling_rate_max.attr,
 287         &sampling_rate_min.attr,
 288         &sampling_rate.attr,
 289         &sampling_down_factor.attr,
 290         &up_threshold.attr,
 291         &down_threshold.attr,
 292         &ignore_nice_load.attr,
 293         &freq_step.attr,
 294         NULL
 295 };
 296
 297 static struct attribute_group dbs_attr_group = {
 298         .attrs = dbs_attributes,
 299         .name = "conservative",
 300 };
 301
 302 /************************** sysfs end ************************/
 303
 304 static void dbs_check_cpu(int cpu)
 305 {
 306         unsigned int idle_ticks, up_idle_ticks, down_idle_ticks;
 307         unsigned int tmp_idle_ticks, total_idle_ticks;
 308         unsigned int freq_step;
 309         unsigned int freq_down_sampling_rate;
 310         struct cpu_dbs_info_s *this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
 311         struct cpufreq_policy *policy;
 312
 313         if (!this_dbs_info->enable)
 314                 return;
 315
 316         policy = this_dbs_info->cur_policy;
 317
 318         /*
 319          * The default safe range is 20% to 80%
 320          * Every sampling_rate, we check
 321          *      - If current idle time is less than 20%, then we try to
 322          *        increase frequency
 323          * Every sampling_rate*sampling_down_factor, we check
 324          *      - If current idle time is more than 80%, then we try to
 325          *        decrease frequency
 326          *
 327          * Any frequency increase takes it to the maximum frequency.
 328          * Frequency reduction happens at minimum steps of
 329          * 5% (default) of max_frequency
 330          */
 331
 332         /* Check for frequency increase */
 333         idle_ticks = UINT_MAX;
 334
 335         /* Check for frequency increase */
 336         total_idle_ticks = get_cpu_idle_time(cpu);
 337         tmp_idle_ticks = total_idle_ticks -
 338                 this_dbs_info->prev_cpu_idle_up;
 339         this_dbs_info->prev_cpu_idle_up = total_idle_ticks;
 340
 341         if (tmp_idle_ticks < idle_ticks)
 342                 idle_ticks = tmp_idle_ticks;
 343
 344         /* Scale idle ticks by 100 and compare with up and down ticks */
 345         idle_ticks *= 100;
 346         up_idle_ticks = (100 - dbs_tuners_ins.up_threshold) *
 347                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 348
 349         if (idle_ticks < up_idle_ticks) {
 350                 this_dbs_info->down_skip = 0;
 351                 this_dbs_info->prev_cpu_idle_down =
 352                         this_dbs_info->prev_cpu_idle_up;
 353
 354                 /* if we are already at full speed then break out early */
 355                 if (this_dbs_info->requested_freq == policy->max)
 356                         return;
 357
 358                 freq_step = (dbs_tuners_ins.freq_step * policy->max) / 100;
 359
 360                 /* max freq cannot be less than 100. But who knows.... */
 361                 if (unlikely(freq_step == 0))
 362                         freq_step = 5;
 363
 364                 this_dbs_info->requested_freq += freq_step;
 365                 if (this_dbs_info->requested_freq > policy->max)
 366                         this_dbs_info->requested_freq = policy->max;
 367
 368                 __cpufreq_driver_target(policy, this_dbs_info->requested_freq,
 369                         CPUFREQ_RELATION_H);
 370                 return;
 371         }
 372
 373         /* Check for frequency decrease */
 374         this_dbs_info->down_skip++;
 375         if (this_dbs_info->down_skip < dbs_tuners_ins.sampling_down_factor)
 376                 return;
 377
 378         /* Check for frequency decrease */
 379         total_idle_ticks = this_dbs_info->prev_cpu_idle_up;
 380         tmp_idle_ticks = total_idle_ticks -
 381                 this_dbs_info->prev_cpu_idle_down;
 382         this_dbs_info->prev_cpu_idle_down = total_idle_ticks;
 383
 384         if (tmp_idle_ticks < idle_ticks)
 385                 idle_ticks = tmp_idle_ticks;
 386
 387         /* Scale idle ticks by 100 and compare with up and down ticks */
 388         idle_ticks *= 100;
 389         this_dbs_info->down_skip = 0;
 390
 391         freq_down_sampling_rate = dbs_tuners_ins.sampling_rate *
 392                 dbs_tuners_ins.sampling_down_factor;
 393         down_idle_ticks = (100 - dbs_tuners_ins.down_threshold) *
 394                 usecs_to_jiffies(freq_down_sampling_rate);
 395
 396         if (idle_ticks > down_idle_ticks) {
 397                 /*
 398                  * if we are already at the lowest speed then break out early
 399                  * or if we 'cannot' reduce the speed as the user might want
 400                  * freq_step to be zero
 401                  */
 402                 if (this_dbs_info->requested_freq == policy->min
 403                                 || dbs_tuners_ins.freq_step == 0)
 404                         return;
 405
 406                 freq_step = (dbs_tuners_ins.freq_step * policy->max) / 100;
 407
 408                 /* max freq cannot be less than 100. But who knows.... */
 409                 if (unlikely(freq_step == 0))
 410                         freq_step = 5;
 411
 412                 this_dbs_info->requested_freq -= freq_step;
 413                 if (this_dbs_info->requested_freq < policy->min)
 414                         this_dbs_info->requested_freq = policy->min;
 415
 416                 __cpufreq_driver_target(policy, this_dbs_info->requested_freq,
 417                                 CPUFREQ_RELATION_H);
 418                 return;
 419         }
 420 }
 421
 422 static void do_dbs_timer(void *data)
 423 {
 424         int i;
 425         lock_cpu_hotplug();
 426         mutex_lock(&dbs_mutex);
 427         for_each_online_cpu(i)
 428                 dbs_check_cpu(i);
 429         schedule_delayed_work(&dbs_work,
 430                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
 431         mutex_unlock(&dbs_mutex);
 432         unlock_cpu_hotplug();
 433 }
 434
 435 static inline void dbs_timer_init(void)
 436 {
 437         INIT_WORK(&dbs_work, do_dbs_timer, NULL);
 438         schedule_delayed_work(&dbs_work,
 439                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
 440         return;
 441 }
 442
 443 static inline void dbs_timer_exit(void)
 444 {
 445         cancel_delayed_work(&dbs_work);
 446         return;
 447 }
 448
 449 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
 450                                    unsigned int event)
 451 {
 452         unsigned int cpu = policy->cpu;
 453         struct cpu_dbs_info_s *this_dbs_info;
 454         unsigned int j;
 455
 456         this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
 457
 458         switch (event) {
 459         case CPUFREQ_GOV_START:
 460                 if ((!cpu_online(cpu)) ||
 461                     (!policy->cur))
 462                         return -EINVAL;
 463
 464                 if (policy->cpuinfo.transition_latency >
 465                                 (TRANSITION_LATENCY_LIMIT * 1000))
 466                         return -EINVAL;
 467                 if (this_dbs_info->enable) /* Already enabled */
 468                         break;
 469
 470                 mutex_lock(&dbs_mutex);
 471                 for_each_cpu_mask(j, policy->cpus) {
 472                         struct cpu_dbs_info_s *j_dbs_info;
 473                         j_dbs_info = &per_cpu(cpu_dbs_info, j);
 474                         j_dbs_info->cur_policy = policy;
 475
 476                         j_dbs_info->prev_cpu_idle_up = get_cpu_idle_time(cpu);
 477                         j_dbs_info->prev_cpu_idle_down
 478                                 = j_dbs_info->prev_cpu_idle_up;
 479                 }
 480                 this_dbs_info->enable = 1;
 481                 this_dbs_info->down_skip = 0;
 482                 this_dbs_info->requested_freq = policy->cur;
 483                 sysfs_create_group(&policy->kobj, &dbs_attr_group);
 484                 dbs_enable++;
 485                 /*
 486                  * Start the timerschedule work, when this governor
 487                  * is used for first time
 488                  */
 489                 if (dbs_enable == 1) {
 490                         unsigned int latency;
 491                         /* policy latency is in nS. Convert it to uS first */
 492                         latency = policy->cpuinfo.transition_latency / 1000;
 493                         if (latency == 0)
 494                                 latency = 1;
 495
 496                         def_sampling_rate = 10 * latency *
 497                                         DEF_SAMPLING_RATE_LATENCY_MULTIPLIER;
 498
 499                         if (def_sampling_rate < MIN_STAT_SAMPLING_RATE)
 500                                 def_sampling_rate = MIN_STAT_SAMPLING_RATE;
 501
 502                         dbs_tuners_ins.sampling_rate = def_sampling_rate;
 503
 504                         dbs_timer_init();
 505                 }
 506
 507                 mutex_unlock(&dbs_mutex);
 508                 break;
 509
 510         case CPUFREQ_GOV_STOP:
 511                 mutex_lock(&dbs_mutex);
 512                 this_dbs_info->enable = 0;
 513                 sysfs_remove_group(&policy->kobj, &dbs_attr_group);
 514                 dbs_enable--;
 515                 /*
 516                  * Stop the timerschedule work, when this governor
 517                  * is used for first time
 518                  */
 519                 if (dbs_enable == 0)
 520                         dbs_timer_exit();
 521
 522                 mutex_unlock(&dbs_mutex);
 523
 524                 break;
 525
 526         case CPUFREQ_GOV_LIMITS:
 527                 lock_cpu_hotplug();
 528                 mutex_lock(&dbs_mutex);
 529                 if (policy->max < this_dbs_info->cur_policy->cur)
 530                         __cpufreq_driver_target(
 531                                         this_dbs_info->cur_policy,
 532                                         policy->max, CPUFREQ_RELATION_H);
 533                 else if (policy->min > this_dbs_info->cur_policy->cur)
 534                         __cpufreq_driver_target(
 535                                         this_dbs_info->cur_policy,
 536                                         policy->min, CPUFREQ_RELATION_L);
 537                 mutex_unlock(&dbs_mutex);
 538                 unlock_cpu_hotplug();
 539                 break;
 540         }
 541         return 0;
 542 }
 543
 544 static struct cpufreq_governor cpufreq_gov_dbs = {
 545         .name           = "conservative",
 546         .governor       = cpufreq_governor_dbs,
 547         .owner          = THIS_MODULE,
 548 };
 549
 550 static int __init cpufreq_gov_dbs_init(void)
 551 {
 552         return cpufreq_register_governor(&cpufreq_gov_dbs);
 553 }
 554
 555 static void __exit cpufreq_gov_dbs_exit(void)
 556 {
 557         /* Make sure that the scheduled work is indeed not running */
 558         flush_scheduled_work();
 559
 560         cpufreq_unregister_governor(&cpufreq_gov_dbs);
 561 }
 562
 563
 564 MODULE_AUTHOR ("Alexander Clouter <alex-kernel@digriz.org.uk>");
 565 MODULE_DESCRIPTION ("'cpufreq_conservative' - A dynamic cpufreq governor for "
 566                 "Low Latency Frequency Transition capable processors "
 567                 "optimised for use in a battery environment");
 568 MODULE_LICENSE ("GPL");
 569
 570 module_init(cpufreq_gov_dbs_init);
 571 module_exit(cpufreq_gov_dbs_exit);