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/smp.h>
  16 #include <linux/init.h>
  17 #include <linux/interrupt.h>
  18 #include <linux/ctype.h>
  19 #include <linux/cpufreq.h>
  20 #include <linux/sysctl.h>
  21 #include <linux/types.h>
  22 #include <linux/fs.h>
  23 #include <linux/sysfs.h>
  24 #include <linux/sched.h>
  25 #include <linux/kmod.h>
  26 #include <linux/workqueue.h>
  27 #include <linux/jiffies.h>
  28 #include <linux/kernel_stat.h>
  29 #include <linux/percpu.h>
  30 #include <linux/mutex.h>
  31
  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 MIN_FREQUENCY_UP_THRESHOLD              (11)
  39 #define MAX_FREQUENCY_UP_THRESHOLD              (100)
  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 with CPUFREQ_ETERNAL)
  48  * 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                  (MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10))
  55 #define MIN_SAMPLING_RATE                       (def_sampling_rate / MIN_SAMPLING_RATE_RATIO)
  56 #define MAX_SAMPLING_RATE                       (500 * def_sampling_rate)
  57 #define DEF_SAMPLING_RATE_LATENCY_MULTIPLIER    (1000)
  58 #define DEF_SAMPLING_DOWN_FACTOR                (1)
  59 #define MAX_SAMPLING_DOWN_FACTOR                (10)
  60 #define TRANSITION_LATENCY_LIMIT                (10 * 1000)
  61
  62 static void do_dbs_timer(void *data);
  63
  64 struct cpu_dbs_info_s {
  65         struct cpufreq_policy *cur_policy;
  66         unsigned int prev_cpu_idle_up;
  67         unsigned int prev_cpu_idle_down;
  68         unsigned int enable;
  69 };
  70 static DEFINE_PER_CPU(struct cpu_dbs_info_s, cpu_dbs_info);
  71
  72 static unsigned int dbs_enable; /* number of CPUs using this policy */
  73
  74 static DEFINE_MUTEX (dbs_mutex);
  75 static DECLARE_WORK     (dbs_work, do_dbs_timer, NULL);
  76
  77 struct dbs_tuners {
  78         unsigned int sampling_rate;
  79         unsigned int sampling_down_factor;
  80         unsigned int up_threshold;
  81         unsigned int ignore_nice;
  82 };
  83
  84 static struct dbs_tuners dbs_tuners_ins = {
  85         .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
  86         .sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR,
  87         .ignore_nice = 0,
  88 };
  89
  90 static inline unsigned int get_cpu_idle_time(unsigned int cpu)
  91 {
  92         return  kstat_cpu(cpu).cpustat.idle +
  93                 kstat_cpu(cpu).cpustat.iowait +
  94                 ( dbs_tuners_ins.ignore_nice ?
  95                   kstat_cpu(cpu).cpustat.nice :
  96                   0);
  97 }
  98
  99 /************************** sysfs interface ************************/
 100 static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf)
 101 {
 102         return sprintf (buf, "%u\n", MAX_SAMPLING_RATE);
 103 }
 104
 105 static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf)
 106 {
 107         return sprintf (buf, "%u\n", MIN_SAMPLING_RATE);
 108 }
 109
 110 #define define_one_ro(_name)            \
 111 static struct freq_attr _name =         \
 112 __ATTR(_name, 0444, show_##_name, NULL)
 113
 114 define_one_ro(sampling_rate_max);
 115 define_one_ro(sampling_rate_min);
 116
 117 /* cpufreq_ondemand Governor Tunables */
 118 #define show_one(file_name, object)                                     \
 119 static ssize_t show_##file_name                                         \
 120 (struct cpufreq_policy *unused, char *buf)                              \
 121 {                                                                       \
 122         return sprintf(buf, "%u\n", dbs_tuners_ins.object);             \
 123 }
 124 show_one(sampling_rate, sampling_rate);
 125 show_one(sampling_down_factor, sampling_down_factor);
 126 show_one(up_threshold, up_threshold);
 127 show_one(ignore_nice_load, ignore_nice);
 128
 129 static ssize_t store_sampling_down_factor(struct cpufreq_policy *unused,
 130                 const char *buf, size_t count)
 131 {
 132         unsigned int input;
 133         int ret;
 134         ret = sscanf (buf, "%u", &input);
 135         if (ret != 1 )
 136                 return -EINVAL;
 137
 138         if (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 > MAX_FREQUENCY_UP_THRESHOLD ||
 176                         input < MIN_FREQUENCY_UP_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_ignore_nice_load(struct cpufreq_policy *policy,
 188                 const char *buf, size_t count)
 189 {
 190         unsigned int input;
 191         int ret;
 192
 193         unsigned int j;
 194
 195         ret = sscanf (buf, "%u", &input);
 196         if ( ret != 1 )
 197                 return -EINVAL;
 198
 199         if ( input > 1 )
 200                 input = 1;
 201
 202         mutex_lock(&dbs_mutex);
 203         if ( input == dbs_tuners_ins.ignore_nice ) { /* nothing to do */
 204                 mutex_unlock(&dbs_mutex);
 205                 return count;
 206         }
 207         dbs_tuners_ins.ignore_nice = input;
 208
 209         /* we need to re-evaluate prev_cpu_idle_up and prev_cpu_idle_down */
 210         for_each_online_cpu(j) {
 211                 struct cpu_dbs_info_s *j_dbs_info;
 212                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
 213                 j_dbs_info->prev_cpu_idle_up = get_cpu_idle_time(j);
 214                 j_dbs_info->prev_cpu_idle_down = j_dbs_info->prev_cpu_idle_up;
 215         }
 216         mutex_unlock(&dbs_mutex);
 217
 218         return count;
 219 }
 220
 221 #define define_one_rw(_name) \
 222 static struct freq_attr _name = \
 223 __ATTR(_name, 0644, show_##_name, store_##_name)
 224
 225 define_one_rw(sampling_rate);
 226 define_one_rw(sampling_down_factor);
 227 define_one_rw(up_threshold);
 228 define_one_rw(ignore_nice_load);
 229
 230 static struct attribute * dbs_attributes[] = {
 231         &sampling_rate_max.attr,
 232         &sampling_rate_min.attr,
 233         &sampling_rate.attr,
 234         &sampling_down_factor.attr,
 235         &up_threshold.attr,
 236         &ignore_nice_load.attr,
 237         NULL
 238 };
 239
 240 static struct attribute_group dbs_attr_group = {
 241         .attrs = dbs_attributes,
 242         .name = "ondemand",
 243 };
 244
 245 /************************** sysfs end ************************/
 246
 247 static void dbs_check_cpu(int cpu)
 248 {
 249         unsigned int idle_ticks, up_idle_ticks, total_ticks;
 250         unsigned int freq_next;
 251         unsigned int freq_down_sampling_rate;
 252         static int down_skip[NR_CPUS];
 253         struct cpu_dbs_info_s *this_dbs_info;
 254
 255         struct cpufreq_policy *policy;
 256         unsigned int j;
 257
 258         this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
 259         if (!this_dbs_info->enable)
 260                 return;
 261
 262         policy = this_dbs_info->cur_policy;
 263         /*
 264          * Every sampling_rate, we check, if current idle time is less
 265          * than 20% (default), then we try to increase frequency
 266          * Every sampling_rate*sampling_down_factor, we look for a the lowest
 267          * frequency which can sustain the load while keeping idle time over
 268          * 30%. If such a frequency exist, we try to decrease to this frequency.
 269          *
 270          * Any frequency increase takes it to the maximum frequency.
 271          * Frequency reduction happens at minimum steps of
 272          * 5% (default) of current frequency
 273          */
 274
 275         /* Check for frequency increase */
 276         idle_ticks = UINT_MAX;
 277         for_each_cpu_mask(j, policy->cpus) {
 278                 unsigned int tmp_idle_ticks, total_idle_ticks;
 279                 struct cpu_dbs_info_s *j_dbs_info;
 280
 281                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
 282                 total_idle_ticks = get_cpu_idle_time(j);
 283                 tmp_idle_ticks = total_idle_ticks -
 284                         j_dbs_info->prev_cpu_idle_up;
 285                 j_dbs_info->prev_cpu_idle_up = total_idle_ticks;
 286
 287                 if (tmp_idle_ticks < idle_ticks)
 288                         idle_ticks = tmp_idle_ticks;
 289         }
 290
 291         /* Scale idle ticks by 100 and compare with up and down ticks */
 292         idle_ticks *= 100;
 293         up_idle_ticks = (100 - dbs_tuners_ins.up_threshold) *
 294                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 295
 296         if (idle_ticks < up_idle_ticks) {
 297                 down_skip[cpu] = 0;
 298                 for_each_cpu_mask(j, policy->cpus) {
 299                         struct cpu_dbs_info_s *j_dbs_info;
 300
 301                         j_dbs_info = &per_cpu(cpu_dbs_info, j);
 302                         j_dbs_info->prev_cpu_idle_down =
 303                                         j_dbs_info->prev_cpu_idle_up;
 304                 }
 305                 /* if we are already at full speed then break out early */
 306                 if (policy->cur == policy->max)
 307                         return;
 308
 309                 __cpufreq_driver_target(policy, policy->max,
 310                         CPUFREQ_RELATION_H);
 311                 return;
 312         }
 313
 314         /* Check for frequency decrease */
 315         down_skip[cpu]++;
 316         if (down_skip[cpu] < dbs_tuners_ins.sampling_down_factor)
 317                 return;
 318
 319         idle_ticks = UINT_MAX;
 320         for_each_cpu_mask(j, policy->cpus) {
 321                 unsigned int tmp_idle_ticks, total_idle_ticks;
 322                 struct cpu_dbs_info_s *j_dbs_info;
 323
 324                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
 325                 /* Check for frequency decrease */
 326                 total_idle_ticks = j_dbs_info->prev_cpu_idle_up;
 327                 tmp_idle_ticks = total_idle_ticks -
 328                         j_dbs_info->prev_cpu_idle_down;
 329                 j_dbs_info->prev_cpu_idle_down = total_idle_ticks;
 330
 331                 if (tmp_idle_ticks < idle_ticks)
 332                         idle_ticks = tmp_idle_ticks;
 333         }
 334
 335         down_skip[cpu] = 0;
 336         /* if we cannot reduce the frequency anymore, break out early */
 337         if (policy->cur == policy->min)
 338                 return;
 339
 340         /* Compute how many ticks there are between two measurements */
 341         freq_down_sampling_rate = dbs_tuners_ins.sampling_rate *
 342                 dbs_tuners_ins.sampling_down_factor;
 343         total_ticks = usecs_to_jiffies(freq_down_sampling_rate);
 344
 345         /*
 346          * The optimal frequency is the frequency that is the lowest that
 347          * can support the current CPU usage without triggering the up
 348          * policy. To be safe, we focus 10 points under the threshold.
 349          */
 350         freq_next = ((total_ticks - idle_ticks) * 100) / total_ticks;
 351         freq_next = (freq_next * policy->cur) /
 352                         (dbs_tuners_ins.up_threshold - 10);
 353
 354         if (freq_next <= ((policy->cur * 95) / 100))
 355                 __cpufreq_driver_target(policy, freq_next, CPUFREQ_RELATION_L);
 356 }
 357
 358 static void do_dbs_timer(void *data)
 359 {
 360         int i;
 361         mutex_lock(&dbs_mutex);
 362         for_each_online_cpu(i)
 363                 dbs_check_cpu(i);
 364         schedule_delayed_work(&dbs_work,
 365                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
 366         mutex_unlock(&dbs_mutex);
 367 }
 368
 369 static inline void dbs_timer_init(void)
 370 {
 371         INIT_WORK(&dbs_work, do_dbs_timer, NULL);
 372         schedule_delayed_work(&dbs_work,
 373                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
 374         return;
 375 }
 376
 377 static inline void dbs_timer_exit(void)
 378 {
 379         cancel_delayed_work(&dbs_work);
 380         return;
 381 }
 382
 383 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
 384                                    unsigned int event)
 385 {
 386         unsigned int cpu = policy->cpu;
 387         struct cpu_dbs_info_s *this_dbs_info;
 388         unsigned int j;
 389
 390         this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
 391
 392         switch (event) {
 393         case CPUFREQ_GOV_START:
 394                 if ((!cpu_online(cpu)) ||
 395                     (!policy->cur))
 396                         return -EINVAL;
 397
 398                 if (policy->cpuinfo.transition_latency >
 399                                 (TRANSITION_LATENCY_LIMIT * 1000)) {
 400                         printk(KERN_WARNING "ondemand governor failed to load "
 401                                "due to too long transition latency\n");
 402                         return -EINVAL;
 403                 }
 404                 if (this_dbs_info->enable) /* Already enabled */
 405                         break;
 406
 407                 mutex_lock(&dbs_mutex);
 408                 for_each_cpu_mask(j, policy->cpus) {
 409                         struct cpu_dbs_info_s *j_dbs_info;
 410                         j_dbs_info = &per_cpu(cpu_dbs_info, j);
 411                         j_dbs_info->cur_policy = policy;
 412
 413                         j_dbs_info->prev_cpu_idle_up = get_cpu_idle_time(j);
 414                         j_dbs_info->prev_cpu_idle_down
 415                                 = j_dbs_info->prev_cpu_idle_up;
 416                 }
 417                 this_dbs_info->enable = 1;
 418                 sysfs_create_group(&policy->kobj, &dbs_attr_group);
 419                 dbs_enable++;
 420                 /*
 421                  * Start the timerschedule work, when this governor
 422                  * is used for first time
 423                  */
 424                 if (dbs_enable == 1) {
 425                         unsigned int latency;
 426                         /* policy latency is in nS. Convert it to uS first */
 427                         latency = policy->cpuinfo.transition_latency / 1000;
 428                         if (latency == 0)
 429                                 latency = 1;
 430
 431                         def_sampling_rate = latency *
 432                                         DEF_SAMPLING_RATE_LATENCY_MULTIPLIER;
 433
 434                         if (def_sampling_rate < MIN_STAT_SAMPLING_RATE)
 435                                 def_sampling_rate = MIN_STAT_SAMPLING_RATE;
 436
 437                         dbs_tuners_ins.sampling_rate = def_sampling_rate;
 438                         dbs_timer_init();
 439                 }
 440
 441                 mutex_unlock(&dbs_mutex);
 442                 break;
 443
 444         case CPUFREQ_GOV_STOP:
 445                 mutex_lock(&dbs_mutex);
 446                 this_dbs_info->enable = 0;
 447                 sysfs_remove_group(&policy->kobj, &dbs_attr_group);
 448                 dbs_enable--;
 449                 /*
 450                  * Stop the timerschedule work, when this governor
 451                  * is used for first time
 452                  */
 453                 if (dbs_enable == 0)
 454                         dbs_timer_exit();
 455
 456                 mutex_unlock(&dbs_mutex);
 457
 458                 break;
 459
 460         case CPUFREQ_GOV_LIMITS:
 461                 mutex_lock(&dbs_mutex);
 462                 if (policy->max < this_dbs_info->cur_policy->cur)
 463                         __cpufreq_driver_target(
 464                                         this_dbs_info->cur_policy,
 465                                         policy->max, CPUFREQ_RELATION_H);
 466                 else if (policy->min > this_dbs_info->cur_policy->cur)
 467                         __cpufreq_driver_target(
 468                                         this_dbs_info->cur_policy,
 469                                         policy->min, CPUFREQ_RELATION_L);
 470                 mutex_unlock(&dbs_mutex);
 471                 break;
 472         }
 473         return 0;
 474 }
 475
 476 static struct cpufreq_governor cpufreq_gov_dbs = {
 477         .name           = "ondemand",
 478         .governor       = cpufreq_governor_dbs,
 479         .owner          = THIS_MODULE,
 480 };
 481
 482 static int __init cpufreq_gov_dbs_init(void)
 483 {
 484         return cpufreq_register_governor(&cpufreq_gov_dbs);
 485 }
 486
 487 static void __exit cpufreq_gov_dbs_exit(void)
 488 {
 489         /* Make sure that the scheduled work is indeed not running */
 490         flush_scheduled_work();
 491
 492         cpufreq_unregister_governor(&cpufreq_gov_dbs);
 493 }
 494
 495
 496 MODULE_AUTHOR ("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
 497 MODULE_DESCRIPTION ("'cpufreq_ondemand' - A dynamic cpufreq governor for "
 498                 "Low Latency Frequency Transition capable processors");
 499 MODULE_LICENSE ("GPL");
 500
 501 module_init(cpufreq_gov_dbs_init);
 502 module_exit(cpufreq_gov_dbs_exit);