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/cpu.h>
  26 #include <linux/sched.h>
  27 #include <linux/kmod.h>
  28 #include <linux/workqueue.h>
  29 #include <linux/jiffies.h>
  30 #include <linux/kernel_stat.h>
  31 #include <linux/percpu.h>
  32 #include <linux/mutex.h>
  33 /*
  34  * dbs is used in this file as a shortform for demandbased switching
  35  * It helps to keep variable names smaller, simpler
  36  */
  37
  38 #define DEF_FREQUENCY_UP_THRESHOLD              (80)
  39 #define DEF_FREQUENCY_DOWN_THRESHOLD            (20)
  40
  41 /*
  42  * The polling frequency of this governor depends on the capability of
  43  * the processor. Default polling frequency is 1000 times the transition
  44  * latency of the processor. The governor will work on any processor with
  45  * transition latency <= 10mS, using appropriate sampling
  46  * rate.
  47  * For CPUs with transition latency > 10mS (mostly drivers
  48  * with CPUFREQ_ETERNAL), this governor will not work.
  49  * All times here are in uS.
  50  */
  51 static unsigned int                             def_sampling_rate;
  52 #define MIN_SAMPLING_RATE_RATIO                 (2)
  53 /* for correct statistics, we need at least 10 ticks between each measure */
  54 #define MIN_STAT_SAMPLING_RATE                  \
  55                         (MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10))
  56 #define MIN_SAMPLING_RATE                       \
  57                         (def_sampling_rate / MIN_SAMPLING_RATE_RATIO)
  58 #define MAX_SAMPLING_RATE                       (500 * def_sampling_rate)
  59 #define DEF_SAMPLING_RATE_LATENCY_MULTIPLIER    (1000)
  60 #define DEF_SAMPLING_DOWN_FACTOR                (1)
  61 #define MAX_SAMPLING_DOWN_FACTOR                (10)
  62 #define TRANSITION_LATENCY_LIMIT                (10 * 1000)
  63
  64 static void do_dbs_timer(struct work_struct *work);
  65
  66 struct cpu_dbs_info_s {
  67         struct cpufreq_policy   *cur_policy;
  68         unsigned int            prev_cpu_idle_up;
  69         unsigned int            prev_cpu_idle_down;
  70         unsigned int            enable;
  71         unsigned int            down_skip;
  72         unsigned int            requested_freq;
  73 };
  74 static DEFINE_PER_CPU(struct cpu_dbs_info_s, cpu_dbs_info);
  75
  76 static unsigned int dbs_enable; /* number of CPUs using this policy */
  77
  78 /*
  79  * DEADLOCK ALERT! There is a ordering requirement between cpu_hotplug
  80  * lock and dbs_mutex. cpu_hotplug lock should always be held before
  81  * dbs_mutex. If any function that can potentially take cpu_hotplug lock
  82  * (like __cpufreq_driver_target()) is being called with dbs_mutex taken, then
  83  * cpu_hotplug lock should be taken before that. Note that cpu_hotplug lock
  84  * is recursive for the same process. -Venki
  85  */
  86 static DEFINE_MUTEX     (dbs_mutex);
  87 static DECLARE_DELAYED_WORK(dbs_work, do_dbs_timer);
  88
  89 struct dbs_tuners {
  90         unsigned int            sampling_rate;
  91         unsigned int            sampling_down_factor;
  92         unsigned int            up_threshold;
  93         unsigned int            down_threshold;
  94         unsigned int            ignore_nice;
  95         unsigned int            freq_step;
  96 };
  97
  98 static struct dbs_tuners dbs_tuners_ins = {
  99         .up_threshold           = DEF_FREQUENCY_UP_THRESHOLD,
 100         .down_threshold         = DEF_FREQUENCY_DOWN_THRESHOLD,
 101         .sampling_down_factor   = DEF_SAMPLING_DOWN_FACTOR,
 102         .ignore_nice            = 0,
 103         .freq_step              = 5,
 104 };
 105
 106 static inline unsigned int get_cpu_idle_time(unsigned int cpu)
 107 {
 108         unsigned int add_nice = 0, ret;
 109
 110         if (dbs_tuners_ins.ignore_nice)
 111                 add_nice = kstat_cpu(cpu).cpustat.nice;
 112
 113         ret =   kstat_cpu(cpu).cpustat.idle +
 114                 kstat_cpu(cpu).cpustat.iowait +
 115                 add_nice;
 116
 117         return ret;
 118 }
 119
 120 /************************** sysfs interface ************************/
 121 static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf)
 122 {
 123         return sprintf (buf, "%u\n", MAX_SAMPLING_RATE);
 124 }
 125
 126 static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf)
 127 {
 128         return sprintf (buf, "%u\n", MIN_SAMPLING_RATE);
 129 }
 130
 131 #define define_one_ro(_name)                                    \
 132 static struct freq_attr _name =                                 \
 133 __ATTR(_name, 0444, show_##_name, NULL)
 134
 135 define_one_ro(sampling_rate_max);
 136 define_one_ro(sampling_rate_min);
 137
 138 /* cpufreq_conservative Governor Tunables */
 139 #define show_one(file_name, object)                                     \
 140 static ssize_t show_##file_name                                         \
 141 (struct cpufreq_policy *unused, char *buf)                              \
 142 {                                                                       \
 143         return sprintf(buf, "%u\n", dbs_tuners_ins.object);             \
 144 }
 145 show_one(sampling_rate, sampling_rate);
 146 show_one(sampling_down_factor, sampling_down_factor);
 147 show_one(up_threshold, up_threshold);
 148 show_one(down_threshold, down_threshold);
 149 show_one(ignore_nice_load, ignore_nice);
 150 show_one(freq_step, freq_step);
 151
 152 static ssize_t store_sampling_down_factor(struct cpufreq_policy *unused,
 153                 const char *buf, size_t count)
 154 {
 155         unsigned int input;
 156         int ret;
 157         ret = sscanf (buf, "%u", &input);
 158         if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
 159                 return -EINVAL;
 160
 161         mutex_lock(&dbs_mutex);
 162         dbs_tuners_ins.sampling_down_factor = input;
 163         mutex_unlock(&dbs_mutex);
 164
 165         return count;
 166 }
 167
 168 static ssize_t store_sampling_rate(struct cpufreq_policy *unused,
 169                 const char *buf, size_t count)
 170 {
 171         unsigned int input;
 172         int ret;
 173         ret = sscanf (buf, "%u", &input);
 174
 175         mutex_lock(&dbs_mutex);
 176         if (ret != 1 || input > MAX_SAMPLING_RATE || input < MIN_SAMPLING_RATE) {
 177                 mutex_unlock(&dbs_mutex);
 178                 return -EINVAL;
 179         }
 180
 181         dbs_tuners_ins.sampling_rate = input;
 182         mutex_unlock(&dbs_mutex);
 183
 184         return count;
 185 }
 186
 187 static ssize_t store_up_threshold(struct cpufreq_policy *unused,
 188                 const char *buf, size_t count)
 189 {
 190         unsigned int input;
 191         int ret;
 192         ret = sscanf (buf, "%u", &input);
 193
 194         mutex_lock(&dbs_mutex);
 195         if (ret != 1 || input > 100 || input <= dbs_tuners_ins.down_threshold) {
 196                 mutex_unlock(&dbs_mutex);
 197                 return -EINVAL;
 198         }
 199
 200         dbs_tuners_ins.up_threshold = input;
 201         mutex_unlock(&dbs_mutex);
 202
 203         return count;
 204 }
 205
 206 static ssize_t store_down_threshold(struct cpufreq_policy *unused,
 207                 const char *buf, size_t count)
 208 {
 209         unsigned int input;
 210         int ret;
 211         ret = sscanf (buf, "%u", &input);
 212
 213         mutex_lock(&dbs_mutex);
 214         if (ret != 1 || input > 100 || input >= dbs_tuners_ins.up_threshold) {
 215                 mutex_unlock(&dbs_mutex);
 216                 return -EINVAL;
 217         }
 218
 219         dbs_tuners_ins.down_threshold = input;
 220         mutex_unlock(&dbs_mutex);
 221
 222         return count;
 223 }
 224
 225 static ssize_t store_ignore_nice_load(struct cpufreq_policy *policy,
 226                 const char *buf, size_t count)
 227 {
 228         unsigned int input;
 229         int ret;
 230
 231         unsigned int j;
 232
 233         ret = sscanf (buf, "%u", &input);
 234         if ( ret != 1 )
 235                 return -EINVAL;
 236
 237         if ( input > 1 )
 238                 input = 1;
 239
 240         mutex_lock(&dbs_mutex);
 241         if ( input == dbs_tuners_ins.ignore_nice ) { /* nothing to do */
 242                 mutex_unlock(&dbs_mutex);
 243                 return count;
 244         }
 245         dbs_tuners_ins.ignore_nice = input;
 246
 247         /* we need to re-evaluate prev_cpu_idle_up and prev_cpu_idle_down */
 248         for_each_online_cpu(j) {
 249                 struct cpu_dbs_info_s *j_dbs_info;
 250                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
 251                 j_dbs_info->prev_cpu_idle_up = get_cpu_idle_time(j);
 252                 j_dbs_info->prev_cpu_idle_down = j_dbs_info->prev_cpu_idle_up;
 253         }
 254         mutex_unlock(&dbs_mutex);
 255
 256         return count;
 257 }
 258
 259 static ssize_t store_freq_step(struct cpufreq_policy *policy,
 260                 const char *buf, size_t count)
 261 {
 262         unsigned int input;
 263         int ret;
 264
 265         ret = sscanf (buf, "%u", &input);
 266
 267         if ( ret != 1 )
 268                 return -EINVAL;
 269
 270         if ( input > 100 )
 271                 input = 100;
 272
 273         /* no need to test here if freq_step is zero as the user might actually
 274          * want this, they would be crazy though :) */
 275         mutex_lock(&dbs_mutex);
 276         dbs_tuners_ins.freq_step = input;
 277         mutex_unlock(&dbs_mutex);
 278
 279         return count;
 280 }
 281
 282 #define define_one_rw(_name) \
 283 static struct freq_attr _name = \
 284 __ATTR(_name, 0644, show_##_name, store_##_name)
 285
 286 define_one_rw(sampling_rate);
 287 define_one_rw(sampling_down_factor);
 288 define_one_rw(up_threshold);
 289 define_one_rw(down_threshold);
 290 define_one_rw(ignore_nice_load);
 291 define_one_rw(freq_step);
 292
 293 static struct attribute * dbs_attributes[] = {
 294         &sampling_rate_max.attr,
 295         &sampling_rate_min.attr,
 296         &sampling_rate.attr,
 297         &sampling_down_factor.attr,
 298         &up_threshold.attr,
 299         &down_threshold.attr,
 300         &ignore_nice_load.attr,
 301         &freq_step.attr,
 302         NULL
 303 };
 304
 305 static struct attribute_group dbs_attr_group = {
 306         .attrs = dbs_attributes,
 307         .name = "conservative",
 308 };
 309
 310 /************************** sysfs end ************************/
 311
 312 static void dbs_check_cpu(int cpu)
 313 {
 314         unsigned int idle_ticks, up_idle_ticks, down_idle_ticks;
 315         unsigned int tmp_idle_ticks, total_idle_ticks;
 316         unsigned int freq_step;
 317         unsigned int freq_down_sampling_rate;
 318         struct cpu_dbs_info_s *this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
 319         struct cpufreq_policy *policy;
 320
 321         if (!this_dbs_info->enable)
 322                 return;
 323
 324         policy = this_dbs_info->cur_policy;
 325
 326         /*
 327          * The default safe range is 20% to 80%
 328          * Every sampling_rate, we check
 329          *      - If current idle time is less than 20%, then we try to
 330          *        increase frequency
 331          * Every sampling_rate*sampling_down_factor, we check
 332          *      - If current idle time is more than 80%, then we try to
 333          *        decrease frequency
 334          *
 335          * Any frequency increase takes it to the maximum frequency.
 336          * Frequency reduction happens at minimum steps of
 337          * 5% (default) of max_frequency
 338          */
 339
 340         /* Check for frequency increase */
 341         idle_ticks = UINT_MAX;
 342
 343         /* Check for frequency increase */
 344         total_idle_ticks = get_cpu_idle_time(cpu);
 345         tmp_idle_ticks = total_idle_ticks -
 346                 this_dbs_info->prev_cpu_idle_up;
 347         this_dbs_info->prev_cpu_idle_up = total_idle_ticks;
 348
 349         if (tmp_idle_ticks < idle_ticks)
 350                 idle_ticks = tmp_idle_ticks;
 351
 352         /* Scale idle ticks by 100 and compare with up and down ticks */
 353         idle_ticks *= 100;
 354         up_idle_ticks = (100 - dbs_tuners_ins.up_threshold) *
 355                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 356
 357         if (idle_ticks < up_idle_ticks) {
 358                 this_dbs_info->down_skip = 0;
 359                 this_dbs_info->prev_cpu_idle_down =
 360                         this_dbs_info->prev_cpu_idle_up;
 361
 362                 /* if we are already at full speed then break out early */
 363                 if (this_dbs_info->requested_freq == policy->max)
 364                         return;
 365
 366                 freq_step = (dbs_tuners_ins.freq_step * policy->max) / 100;
 367
 368                 /* max freq cannot be less than 100. But who knows.... */
 369                 if (unlikely(freq_step == 0))
 370                         freq_step = 5;
 371
 372                 this_dbs_info->requested_freq += freq_step;
 373                 if (this_dbs_info->requested_freq > policy->max)
 374                         this_dbs_info->requested_freq = policy->max;
 375
 376                 __cpufreq_driver_target(policy, this_dbs_info->requested_freq,
 377                         CPUFREQ_RELATION_H);
 378                 return;
 379         }
 380
 381         /* Check for frequency decrease */
 382         this_dbs_info->down_skip++;
 383         if (this_dbs_info->down_skip < dbs_tuners_ins.sampling_down_factor)
 384                 return;
 385
 386         /* Check for frequency decrease */
 387         total_idle_ticks = this_dbs_info->prev_cpu_idle_up;
 388         tmp_idle_ticks = total_idle_ticks -
 389                 this_dbs_info->prev_cpu_idle_down;
 390         this_dbs_info->prev_cpu_idle_down = total_idle_ticks;
 391
 392         if (tmp_idle_ticks < idle_ticks)
 393                 idle_ticks = tmp_idle_ticks;
 394
 395         /* Scale idle ticks by 100 and compare with up and down ticks */
 396         idle_ticks *= 100;
 397         this_dbs_info->down_skip = 0;
 398
 399         freq_down_sampling_rate = dbs_tuners_ins.sampling_rate *
 400                 dbs_tuners_ins.sampling_down_factor;
 401         down_idle_ticks = (100 - dbs_tuners_ins.down_threshold) *
 402                 usecs_to_jiffies(freq_down_sampling_rate);
 403
 404         if (idle_ticks > down_idle_ticks) {
 405                 /*
 406                  * if we are already at the lowest speed then break out early
 407                  * or if we 'cannot' reduce the speed as the user might want
 408                  * freq_step to be zero
 409                  */
 410                 if (this_dbs_info->requested_freq == policy->min
 411                                 || dbs_tuners_ins.freq_step == 0)
 412                         return;
 413
 414                 freq_step = (dbs_tuners_ins.freq_step * policy->max) / 100;
 415
 416                 /* max freq cannot be less than 100. But who knows.... */
 417                 if (unlikely(freq_step == 0))
 418                         freq_step = 5;
 419
 420                 this_dbs_info->requested_freq -= freq_step;
 421                 if (this_dbs_info->requested_freq < policy->min)
 422                         this_dbs_info->requested_freq = policy->min;
 423
 424                 __cpufreq_driver_target(policy, this_dbs_info->requested_freq,
 425                                 CPUFREQ_RELATION_H);
 426                 return;
 427         }
 428 }
 429
 430 static void do_dbs_timer(struct work_struct *work)
 431 {
 432         int i;
 433         lock_cpu_hotplug();
 434         mutex_lock(&dbs_mutex);
 435         for_each_online_cpu(i)
 436                 dbs_check_cpu(i);
 437         schedule_delayed_work(&dbs_work,
 438                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
 439         mutex_unlock(&dbs_mutex);
 440         unlock_cpu_hotplug();
 441 }
 442
 443 static inline void dbs_timer_init(void)
 444 {
 445         schedule_delayed_work(&dbs_work,
 446                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
 447         return;
 448 }
 449
 450 static inline void dbs_timer_exit(void)
 451 {
 452         cancel_delayed_work(&dbs_work);
 453         return;
 454 }
 455
 456 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
 457                                    unsigned int event)
 458 {
 459         unsigned int cpu = policy->cpu;
 460         struct cpu_dbs_info_s *this_dbs_info;
 461         unsigned int j;
 462         int rc;
 463
 464         this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
 465
 466         switch (event) {
 467         case CPUFREQ_GOV_START:
 468                 if ((!cpu_online(cpu)) ||
 469                     (!policy->cur))
 470                         return -EINVAL;
 471
 472                 if (policy->cpuinfo.transition_latency >
 473                                 (TRANSITION_LATENCY_LIMIT * 1000))
 474                         return -EINVAL;
 475                 if (this_dbs_info->enable) /* Already enabled */
 476                         break;
 477
 478                 mutex_lock(&dbs_mutex);
 479
 480                 rc = sysfs_create_group(&policy->kobj, &dbs_attr_group);
 481                 if (rc) {
 482                         mutex_unlock(&dbs_mutex);
 483                         return rc;
 484                 }
 485
 486                 for_each_cpu_mask(j, policy->cpus) {
 487                         struct cpu_dbs_info_s *j_dbs_info;
 488                         j_dbs_info = &per_cpu(cpu_dbs_info, j);
 489                         j_dbs_info->cur_policy = policy;
 490
 491                         j_dbs_info->prev_cpu_idle_up = get_cpu_idle_time(cpu);
 492                         j_dbs_info->prev_cpu_idle_down
 493                                 = j_dbs_info->prev_cpu_idle_up;
 494                 }
 495                 this_dbs_info->enable = 1;
 496                 this_dbs_info->down_skip = 0;
 497                 this_dbs_info->requested_freq = policy->cur;
 498
 499                 dbs_enable++;
 500                 /*
 501                  * Start the timerschedule work, when this governor
 502                  * is used for first time
 503                  */
 504                 if (dbs_enable == 1) {
 505                         unsigned int latency;
 506                         /* policy latency is in nS. Convert it to uS first */
 507                         latency = policy->cpuinfo.transition_latency / 1000;
 508                         if (latency == 0)
 509                                 latency = 1;
 510
 511                         def_sampling_rate = 10 * latency *
 512                                         DEF_SAMPLING_RATE_LATENCY_MULTIPLIER;
 513
 514                         if (def_sampling_rate < MIN_STAT_SAMPLING_RATE)
 515                                 def_sampling_rate = MIN_STAT_SAMPLING_RATE;
 516
 517                         dbs_tuners_ins.sampling_rate = def_sampling_rate;
 518
 519                         dbs_timer_init();
 520                 }
 521
 522                 mutex_unlock(&dbs_mutex);
 523                 break;
 524
 525         case CPUFREQ_GOV_STOP:
 526                 mutex_lock(&dbs_mutex);
 527                 this_dbs_info->enable = 0;
 528                 sysfs_remove_group(&policy->kobj, &dbs_attr_group);
 529                 dbs_enable--;
 530                 /*
 531                  * Stop the timerschedule work, when this governor
 532                  * is used for first time
 533                  */
 534                 if (dbs_enable == 0)
 535                         dbs_timer_exit();
 536
 537                 mutex_unlock(&dbs_mutex);
 538
 539                 break;
 540
 541         case CPUFREQ_GOV_LIMITS:
 542                 mutex_lock(&dbs_mutex);
 543                 if (policy->max < this_dbs_info->cur_policy->cur)
 544                         __cpufreq_driver_target(
 545                                         this_dbs_info->cur_policy,
 546                                         policy->max, CPUFREQ_RELATION_H);
 547                 else if (policy->min > this_dbs_info->cur_policy->cur)
 548                         __cpufreq_driver_target(
 549                                         this_dbs_info->cur_policy,
 550                                         policy->min, CPUFREQ_RELATION_L);
 551                 mutex_unlock(&dbs_mutex);
 552                 break;
 553         }
 554         return 0;
 555 }
 556
 557 static struct cpufreq_governor cpufreq_gov_dbs = {
 558         .name           = "conservative",
 559         .governor       = cpufreq_governor_dbs,
 560         .owner          = THIS_MODULE,
 561 };
 562
 563 static int __init cpufreq_gov_dbs_init(void)
 564 {
 565         return cpufreq_register_governor(&cpufreq_gov_dbs);
 566 }
 567
 568 static void __exit cpufreq_gov_dbs_exit(void)
 569 {
 570         /* Make sure that the scheduled work is indeed not running */
 571         flush_scheduled_work();
 572
 573         cpufreq_unregister_governor(&cpufreq_gov_dbs);
 574 }
 575
 576
 577 MODULE_AUTHOR ("Alexander Clouter <alex-kernel@digriz.org.uk>");
 578 MODULE_DESCRIPTION ("'cpufreq_conservative' - A dynamic cpufreq governor for "
 579                 "Low Latency Frequency Transition capable processors "
 580                 "optimised for use in a battery environment");
 581 MODULE_LICENSE ("GPL");
 582
 583 module_init(cpufreq_gov_dbs_init);
 584 module_exit(cpufreq_gov_dbs_exit);