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